Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Brain network pattern analysis with positron emission tomography data : application to Parkinson's disease Fu, FangLu Jessie 2020

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Notice for Google Chrome users:
If you are having trouble viewing or searching the PDF with Google Chrome, please download it here instead.

Item Metadata

Download

Media
24-ubc_2020_november_fu_jessiefanglu.pdf [ 13.87MB ]
Metadata
JSON: 24-1.0391842.json
JSON-LD: 24-1.0391842-ld.json
RDF/XML (Pretty): 24-1.0391842-rdf.xml
RDF/JSON: 24-1.0391842-rdf.json
Turtle: 24-1.0391842-turtle.txt
N-Triples: 24-1.0391842-rdf-ntriples.txt
Original Record: 24-1.0391842-source.json
Full Text
24-1.0391842-fulltext.txt
Citation
24-1.0391842.ris

Full Text

Vruin bytwork duttyrn Unulysis withdositron Emission homogruphy DutuNUppliwution to durkinson's DisyusybyJessie FangLu FuB.Sc., The University of British Columbia, 2014M.Sc., The University of British Columbia, 2016A THESIS SUBMITTED IN PARTIAL FULFILLMENT OFTHE REQUIREMENTS FOR THE DEGREE OFDOCTOR OF PHILOSOPHYinThe Faculty of Graduate and Postdoctoral Studies(Physics)THE UNIVERSITY OF BRITISH COLUMBIA(Vancouver)June 2020c© Jessie FangLu Fu 2020The following individuals certify that they have read, and recommend to the Facultyof Graduate and Postdoctoral Studies for acceptance, the dissertation entitled:Brain Network Pattern Analysis with Positron Emission Tomography Data:Application to Parkinson’s Diseasesubmitted by Jessie Fanglu Fu in partial fulfillment of the requirements forthe degree of Doctor of Philosophyin PhysicsZxvmining CommittzzOVesna Sossi, Physics and AstronomySupervisorMartin J McKeown, MedicineCo-supervisorUrs Hafeli, Pharmaceutical ScienceSupervisory Committee MembersShannon Kolind, MedicineSupervisory Committee MembersIngrid Stairs, Physics and AstronomySupervisory Committee MembersAlexander Rauscher, MedicineUniversity ExaminerPurang Abolmaesumi, Electrical and Computer EngineeringUniversity ExamineriiUvstruwtPositron emission tomography (PET) is commonly used to investigate changeswithin the brain due to aging and disease. Because our brain works as an integratedsystem where multiple brain regions work together to perform complex tasks, net-work pattern analyses (a subset of machine-learning methods) are often found toprovide complementary, more sensitive and more robust information compared totraditional univariate analyses, especially in the field of magnetic resonance imaging(MRI). However, network pattern analyses have not been commonly used to studyneurotransmitter changes using PET data. In addition, the emergence of multi-tracer imaging studies highlights the needs to develop novel joint analysis methodsto extract and combine complementary information from each imaging dataset toobtain a complete picture of the complex brain states. This thesis constitutes oneof the first applications of such methods in the PET field.Parkinsons disease (PD) is the second most common neurodegenerative disorder.It has a long prodromal stage, and non-motor symptoms occur alongside or evenbefore motor symptoms. Initially thought to affect predominantly the dopaminergicsystem, PD is now deemed to be associated with alterations in several other non-dopaminergic neurotransmitter systems. Such changes, specific to PD, are some-times difficult to detect, especially in prodromal and early stages of the disease; theinteractions between different disease-related mechanisms also remain largely un-clear. In addition, the disease origin is unknown and there is currently no effectivecure for PD.In this thesis work, we 1) explored deterministic spatial connectivity changesin the serotonergic system that are sensitive for detecting subtle changes in theprodromal and early disease stages; 2) introduced dynamic mode decompositionto extract spatio-temporal patterns of dopaminergic denervation for modeling dis-ease progression; 3) introduced a novel joint pattern analysis approach to extractcomplementary information in the dopaminergic and serotonergic systems and theirrelationships with treatment response and treatment-induced complications. Theseiiinovel methods not only lead to new understanding of PD, but also provide moresensitive and deterministic tools for the analysis of PET data in a variety of clinicalapplications.ivLuy gummuryIn the work presented in this thesis, we developed and validated various networkpattern analysis methods for analyzing multi-tracer positron emission tomography(PET) data. In particular, we applied these methods to study changes related toParkinson’s disease (PD). Our results demonstrated that the disease affects multiplebrain regions following a deterministic network pattern, and such effects can be seenin multiple neurotransmitter systems reflecting different aspects of brain functions.Overall, these novel methods not only provide more sensitive tools for extractingmeaningful information from PET data, but also lead to new understanding of PD.vdryfuwyA version of Chapter 2 has been published as JC [C [uA IC Kl–uzhinA hC aiuAZC hhvhinfvryA cC kvfviA JC bxKznzizA cC czilsonA gC bvwroukA bC VChvxhzliA YC lilzA bC JC bxKzofinA VC JC htozsslA vny kC hossiA InvzstigvBtion of szrotonzrgix evrkinsons yiszvszBrzlvtzy xovvrivnxz pvttzrn usingpFFCrBYVhWDeZiA czuroImvgz ClinCA volC FNA ppC KJGKKEA JvnC GEFM. Iwas responsible for development of the analysis methodology, image preprocessing,statistical analysis, clinical interpretation and manuscript composition. Scanningprocedures were performed by the staff members of the UBC PET imaging group.I. Klyuzhin, R. Mabrouk and M.J. McKeown provided feedbacks on the develop-ment of the analysis methodology. S. Liu, M.A. Sacheli, A.J. Stoessl and V. Sossicontributed to clinical interpretation. E. Shahinfard and N. Vafai contributed toimage pre-processing. J. McKenzie and N. Neilson contributed to recruitment ofstudy participants. A.J. Stoessl and V. Sossi designed the study. V. Sossi was thethe supervisory author involved throughout the project in the concept formationand manuscript preparation.The work described in Chapter 3 represents original unpublished materials. I wasresponsible for the development of the presented algorithms and methods and theirvalidation and testing. V. Sossi was the supervisory author involved throughout theproject.A version of Chapter 4 has been published as JC [C [uA IC hC Kl–uzhinA bCJC bxKzofinA VC JC htozsslA vny kC hossiA covzl yvtvByrivznA zquvtionBfrzz mzthoy xvpturzs spvtioBtzmporvl pvttzrns of nzuroyzgznzrvtion inevrkinsons yiszvszO Vpplixvtion of y–nvmix moyz yzxomposition to eZiAczuroImvgz ClinCA volC GJA pC FEGFJEA JvnC GEGEC. I was responsible for devel-opment of the analysis methodology, statistical analysis, clinical interpretation andmanuscript composition. Scanning procedures were performed by the staff membersof the UBC PET imaging group. I. Klyuzhin contributed to image preprocessing.M.J. McKeown provided feedbacks on the development of the analysis methodol-viogy. A.J. Stoessl and V. Sossi contributed to clinical interpretation. A.J. Stoessland V. Sossi designed the study. V. Sossi was the the supervisory author involvedthroughout the project in the concept formation and manuscript preparation.A version of Chapter 5 has been published as JC [C [uA IC Kl–uzhinA JCbxKznzizA cC czilsonA ZC hhvhinfvryA KC YinzllzA bC JC bxKzofinA VCJC htozsslA vny kC hossiA Joint pvttzrn vnvl–sis vpplizy to eZi YVivny kbViG imvging rzvzvls nzfi insights into evrkinsons yiszvsz inByuxzy przs–nvptix vltzrvtionsA czuroImvgz ClinCA pC FEFMJKA bv– GEFN.I was responsible for development of the analysis methodology, image preprocess-ing, statistical analysis, clinical interpretation and manuscript composition. Scan-ning procedures were performed by the staff members of the UBC PET imaginggroup. K.Dinelle, E. Shahinfard and N. Vafai contributed to image preprocess-ing. J. McKenzie and N. Neilson contributed to recruitment of study participants.I. Klyuzhin and M.J. McKeown contributed to the development of the analysispipeline. A.J. Stoessl and V. Sossi designed the study and contributed to clinicalinterpretation. V. Sossi was the supervisory author involved throughout the projectin the concept formation and manuscript preparation.A version of Chapter 6 has been submitted to a peer-reviewed journal as JC [C[uA bC bvtorvzzoA JC bxKznzizA cC czilsonA cC kvfviA KC YinzllzA VCCC [zxilioA bC JC bxKzofinA VC JC htozsslA vny kC hossiA 7hzrotonzrgixs–stzm vffzxts lzvoyopv rzsponsz vny y–skinzsiv in zvrl– evrkinson7. Iwas responsible for development of the analysis methodology, image preprocessing,statistical analysis, clinical interpretation and manuscript composition. Scanningprocedures were performed by the staff members of the UBC PET imaging group.M. Matarazzo, A.J.Stoessl and V.Sossi contributed to clinical interpretation. N.Vafai contributed to image preprocessing. J. McKenzie and N. Neilson contributedto recruitment of study participants. M.J. McKeown contributed to the developmentof the analysis pipeline. A.C. Felicio contributed to preliminary data collection.A.J. Stoessl and V. Sossi designed the study. V. Sossi was the the supervisoryauthor involved throughout the project in the concept formation and manuscriptpreparation.This study was approved by UBC Research Human Ethics Board, in particularthe Clinical Research Ethics Board, under ’The Evolution of PD’ (certificate number:H12-00843), ’Serotonergic Innervation, Dopamine Release and Complications in PD’(certificate number: H12-01450), ’PET Imaging of [11C]DASB’ (certificate number:viiH11-02620).viiihuvly of ContyntsVwstrvxt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiav– hummvr– . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . verzfvxz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viivwlz of Contznts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixaist of ivwlzs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xivaist of [igurzs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviaist of Vwwrzvivtions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxivVxknofilzygmznts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .xxviiiYzyixvtion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxxF Introyuxtion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Thesis Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Positron Emission Tomography . . . . . . . . . . . . . . . . . . . . . 31.2.1 Radiotracer . . . . . . . . . . . . . . . . . . . . . . . . . . . 41.2.2 Radioisotope Decay . . . . . . . . . . . . . . . . . . . . . . . 51.2.3 Detection System . . . . . . . . . . . . . . . . . . . . . . . . 61.2.4 Image Reconstruction . . . . . . . . . . . . . . . . . . . . . . 91.2.5 Post-Reconstruction Smoothing . . . . . . . . . . . . . . . . 121.3 Kinetic Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121.3.1 Reference Tissue Models . . . . . . . . . . . . . . . . . . . . 131.3.2 Simplified Reference Tissue Model . . . . . . . . . . . . . . . 151.3.3 Logan Graphical Analysis . . . . . . . . . . . . . . . . . . . . 16ix1.3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171.4 Parkinson’s Disease . . . . . . . . . . . . . . . . . . . . . . . . . . . 171.4.1 Symptoms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171.4.2 Leucine-Rich Repeat Kinase 2 (LRRK2) Mutation . . . . . . 181.4.3 Treatment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191.4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191.5 Neurotransmitter Systems and PET Radiotracers . . . . . . . . . . 191.5.1 Dopaminergic System . . . . . . . . . . . . . . . . . . . . . . 201.5.2 Serotonergic System . . . . . . . . . . . . . . . . . . . . . . . 221.5.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241.6 Network Pattern Analyses . . . . . . . . . . . . . . . . . . . . . . . 241.6.1 Linear Models of Spatial Patterns: Principal Component Anal-ysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261.6.2 Network Topology . . . . . . . . . . . . . . . . . . . . . . . . 281.6.3 Spatio-temporal Patterns . . . . . . . . . . . . . . . . . . . . 311.6.4 Multi-tracer Joint Pattern Analysis . . . . . . . . . . . . . . 321.7 Research Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . 34G ainzvr boyzls of hpvtivl evttzrns . . . . . . . . . . . . . . . . . . . 362.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372.3 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . . . 402.3.1 Study Participants . . . . . . . . . . . . . . . . . . . . . . . . 402.3.2 Statistical Analysis and Clinical Data . . . . . . . . . . . . . 432.3.3 Scanning Protocol . . . . . . . . . . . . . . . . . . . . . . . . 432.3.4 Image Processing and Analysis . . . . . . . . . . . . . . . . . 442.3.5 SSM Multivariate Pattern Analysis . . . . . . . . . . . . . . 442.3.6 Univariate Analysis . . . . . . . . . . . . . . . . . . . . . . . 462.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462.4.1 sPD and LRRK2-PD Spatial Covariance Pattern . . . . . . . 462.4.2 LRRK2-NMC Spatial Covariance Pattern . . . . . . . . . . . 492.4.3 Univariate Analysis . . . . . . . . . . . . . . . . . . . . . . . 502.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 512.5.1 Possible Functional Basis for SPDRP and LRRK2-NMC Pat-tern Topography . . . . . . . . . . . . . . . . . . . . . . . . . 52x2.5.2 LRRK2-NMC Pattern Protection or Compensation . . . . . 532.5.3 Comparison with Univariate Analysis . . . . . . . . . . . . . 542.5.4 Comparison with Other Network Analysis Patterns . . . . . 552.5.5 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 552.6 PCA Versus Other Spatial Models . . . . . . . . . . . . . . . . . . . 552.6.1 Dimension Reduction Methods . . . . . . . . . . . . . . . . . 562.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64H cztfiork iopolog– . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 653.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 653.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 653.3 Methods and Materials . . . . . . . . . . . . . . . . . . . . . . . . . 663.3.1 Study Participants . . . . . . . . . . . . . . . . . . . . . . . . 663.3.2 Scanning Protocol . . . . . . . . . . . . . . . . . . . . . . . . 663.3.3 Image Processing and Analysis . . . . . . . . . . . . . . . . . 673.3.4 Graph Theory Analysis . . . . . . . . . . . . . . . . . . . . . 683.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713.4.1 Adjacency Matrices . . . . . . . . . . . . . . . . . . . . . . . 713.4.2 Graph Theory Metrics . . . . . . . . . . . . . . . . . . . . . 733.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 753.5.1 Adjacency Matrices . . . . . . . . . . . . . . . . . . . . . . . 753.5.2 Graph Theory Metrics . . . . . . . . . . . . . . . . . . . . . 763.5.3 Limitations and Considerations . . . . . . . . . . . . . . . . 773.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78I hpvtioBizmporvl evttzrns . . . . . . . . . . . . . . . . . . . . . . . . 794.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 794.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 804.3 Methods and Materials . . . . . . . . . . . . . . . . . . . . . . . . . 834.3.1 Study Participants . . . . . . . . . . . . . . . . . . . . . . . . 834.3.2 Scanning Protocols . . . . . . . . . . . . . . . . . . . . . . . 834.3.3 Imaging Processing . . . . . . . . . . . . . . . . . . . . . . . 844.4 Introducing DMD . . . . . . . . . . . . . . . . . . . . . . . . . . . . 844.4.1 DMD Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 854.4.2 DMD for Tracking Disease Progression . . . . . . . . . . . . 864.4.3 Robustness and Reproducibility of DMD . . . . . . . . . . . 88xi4.4.4 Univariate Analysis . . . . . . . . . . . . . . . . . . . . . . . 894.4.5 Comparison between dynamic mode decomposition (DMD)and principal component analysis (PCA) . . . . . . . . . . . 894.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 904.5.1 Putamen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 914.5.2 Caudate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 934.5.3 Comparison between DMD and PCA . . . . . . . . . . . . . 944.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 954.6.1 DMD Spatio-Temporal Patterns . . . . . . . . . . . . . . . . 964.6.2 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 984.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100J Joint evttzrn Vnvl–sis . . . . . . . . . . . . . . . . . . . . . . . . . . . 1015.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1015.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1025.3 Methods and Materials . . . . . . . . . . . . . . . . . . . . . . . . . 1065.3.1 Study Participants . . . . . . . . . . . . . . . . . . . . . . . . 1065.3.2 Scanning Protocols . . . . . . . . . . . . . . . . . . . . . . . 1065.3.3 Image Processing and Analysis . . . . . . . . . . . . . . . . . 1075.3.4 Univariate Analysis . . . . . . . . . . . . . . . . . . . . . . . 1085.3.5 Joint Pattern Analysis . . . . . . . . . . . . . . . . . . . . . 1085.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1125.4.1 Univariate Analysis . . . . . . . . . . . . . . . . . . . . . . . 1125.4.2 Joint Pattern Analysis . . . . . . . . . . . . . . . . . . . . . 1135.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1185.5.1 Univariate Analysis . . . . . . . . . . . . . . . . . . . . . . . 1185.5.2 Joint Pattern Analysis . . . . . . . . . . . . . . . . . . . . . 1205.5.3 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1235.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123K Joint evttzrn Vnvl–sis { Vpplixvtion . . . . . . . . . . . . . . . . . 1256.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1256.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1266.3 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . . . 1286.3.1 Study Participants . . . . . . . . . . . . . . . . . . . . . . . . 1286.3.2 Scanning Protocols . . . . . . . . . . . . . . . . . . . . . . . 128xii6.3.3 Image Processing and Analysis . . . . . . . . . . . . . . . . . 1306.3.4 Statistical Analysis . . . . . . . . . . . . . . . . . . . . . . . 1316.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1336.4.1 Clinical Follow-up . . . . . . . . . . . . . . . . . . . . . . . . 1336.4.2 Univariate Analysis . . . . . . . . . . . . . . . . . . . . . . . 1346.4.3 Joint Pattern Analysis . . . . . . . . . . . . . . . . . . . . . 1356.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1386.5.1 Univariate Analysis . . . . . . . . . . . . . . . . . . . . . . . 1386.5.2 Joint Pattern Analysis . . . . . . . . . . . . . . . . . . . . . 1396.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142L Conxlusion vny [uturz Yirzxtions . . . . . . . . . . . . . . . . . . . 143Wiwliogrvph– . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148xiiiList of huvlys2.1 Characteristics of study participants. Data are meanstandard devi-ation, unless otherwise indicated. sPD=sporadic Parkinsons diseasesubjects; LRRK2-PD = manifest LRRK2 mutation carriers; LRRK2-NMC = non-manifesting LRRK2 mutation carriers; MDS-UPDRS =Movement Disorder Society Unified Parkinsons Disease Rating Scale;N/A = not applicable. Disease duration was estimated as time fromonset of motor symptoms as reported by the patients. *p values forage, MoCA and BDI were calculated by ANOVA followed by post-hocanalyses. p values for genter ratio were calculated by Fishers exacttest. p values for disease duration and MDS-UPDRS III scores werecalculated by independent two-tailed T-test. p values for Hoehn andYahr stage were calculated by Mann-Whitney test. All p values werefalse discovery rate-corrected for multiple comparisons. Data from8 of 9 healthy controls, 15 of 15 sPD, 5 of 8 LRRK2-PD, and 4 of9 LRRK2-NMC (missing information was due to language barriers).†Data from 15 of 15 sPD, 5 of 8 LRRK2-PD, and 6 of 9 LRRK2-NMC. 422.2 Comparison between PCA and other similar dimension reductionmethods in terms of group discrimination power between healthy con-trols and sPD subjects, variance accounted for, and correlation be-tween the subject scores and disease duration using the DASB PETdata. Overall, PCA seems to be most suitable method for this par-ticular study as discussed in the text below. . . . . . . . . . . . . . . 563.1 ROI names and orders for the correlation matrices shown in Fig.3.1A. 73xiv4.1 Clinical characteristics of all subjects. All numbers are reported asmean± standard deviation. Disease duration estimated as the timefrom onset of motor symptoms as reported by the patients. PD =Parkinson’s disease subjects; MDS-UPDRS = Movement DisorderSociety Unified Parkinson’s Disease Rating Scale; MoCA = MontrealCognitive Assessment; H&Y=Hoehn and Yahr scale. . . . . . . . . . 834.2 DMD output parameters. All numbers are reported as meanstandarddeviation. DMD amplitudes (unitless) and decay constants determinethe intercept and shape of the exponential temporal curves in eachstriatal region. Temporal curves with more negative decay constantsdrop more quickly with increasing disease duration compared to thetemporal curves with decay constants closer to zero. Higher DMDamplitude represents higher expression of the DMD spatial patternat disease onset. The total percentage variance explained were calcu-lated with the first two modes for putamen and with the first modefor caudate. Standard deviation of the DMD parameters were cal-culated from leave-one-out cross-validation. DMD = dynamic modedecomposition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 915.1 Clinical characteristics of all subjects. All numbers are reported asmean± standard deviation. * disease duration estimated as the timefrom onset of motor symptoms as reported by the patients. † diseaseduration estimated as the time of clinical diagnosis. PD=Parkinsonsdisease subjects; MDS-UPDRS=Movement Disorder Society UnifiedParkinsons Disease Rating Scale; MoCA=Montreal Cognitive Assess-ment; BDI=Beck Depression Inventory; LED=Levodopa equivalentdose. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1065.2 Table 2: Correlation strength g2 and significance between each pairof canonical variates. *=significant at p-value = 0.05 . . . . . . . . . 113xvList of Figurys1.1 Overview for the thesis. . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Nucleus decays to emit positron (+) which travels to the site ofannihilation where it annihilates with an electron (z−) producing two511 keV gamma rays () in opposite directions. r is the displacementfrom the parent nucleus to the site of annihilation, whereas p is theactual path of + [1] . . . . . . . . . . . . . . . . . . . . . . . . . . . 61.3 True coincidence detection vs scatter and random coincidence detec-tion in PET [2] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71.4 Systematic diagram of the reference tissue model [3]. Using a refer-ence region, parameter Xp can be eliminated. . . . . . . . . . . . . . 131.5 VMAT2 is labeled by [11X]-DTBZ, and the membrane DAT can belabeled by [11X]-MP. Dopamine D2 receptor availability can be la-beled by [11X]-RAC, which is sensitive to synaptic levels of dopamine[4] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201.6 Neuronal projections of four dopamine systems in human brain [5] . 211.7 DAT binding to healthy controls (A), LRRK2-NMC (B) and LRRK2-PD (C). LRRK2-PD shows most reduced DAT binding throughoutthe striatum (C). Striatal DAT binding is also reduced in LRRK2-NMC compared to healthy controls (B) [6] . . . . . . . . . . . . . . . 221.8 Serotonin projections in the brain. Serotonin is produced in nucleusraphe and projects onto other brain regions [7] . . . . . . . . . . . . 231.9 PCA on example data points in two-dimensional coordinate. . . . . . 271.10 Illustration of some measures of network topology [8] . . . . . . . . . 29xvi1.11 a) random network; b) small-world network with most edges connect-ing neighboring nodes and a few long edges to create short-cuts fordistant sub-networks [9]; c) scale-free network with most nodes havefew connections to other nodes, while several nodes (black connec-tions) are connected to many other nodes. . . . . . . . . . . . . . . . 301.12 Conceptual diagram of MCCA [10] . . . . . . . . . . . . . . . . . . . 322.1 Serotonergic Parkinsons disease-related pattern (SPDRP) identifiedby comparing sporadic Parkinsons disease subjects and healthy con-trols. Regions with significant weights on the averaged SPDRP wereoverlaid onto a T1 MRI image. Blue (red) indicates regions with rel-atively decreased (increased) binding in sporadic Parkinsons diseasesubjects compared to healthy controls. . . . . . . . . . . . . . . . . . 472.2 Subject scores projected onto the serotonergic Parkinsons disease-related pattern (SPDPR) for all four subject groups. The three out-liers were labeled as H1013 and H1079 in the LRRK2-PD group, andH814 in the LRRK2-NMC group. sPD = sporadic Parkinsons dis-ease subjects; LRRK2-PD = manifested LRRK2 mutation carriers;LRRK2-NMC = non-manifesting LRRK2 mutation carriers; * = sig-nificant at P Q0.05 level; **=significant at P Q0.01 level. . . . . . . 482.3 Scatter plot of projected subject scores, denoting the strength of theserotonergic Parkinsons disease-related pattern (SPDRP) expressionas a function of disease durations (left) and as a function of DTBZbinding expressed as fractions to age-matched normal controls (right)for sporadic PD and LRRK2-PD. The two outliers are labeled in thesame way as in Fig.2.2. The best line fit was done without these twosubjects. LRRK2-PD = manifest LRRK2 mutation carriers. . . . . . 482.4 Subject scores projected onto LRRK2-NMC pattern in all four sub-ject groups. sPD=sporadic Parkinsons disease subjects; LRRK2-PD = manifested LRRK2 mutation carriers; LRRK2-NMC = non-manifesting LRRK2 mutation carriers; ** = significant at P Q0.01level. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50xvii2.5 Serotonergic non-manifesting LRRK2 mutation carriers (LRRK2-NMC)pattern identified by comparing LRRK2-NMC and healthy controls.Regions with significant weights on the averaged LRRK2-NMC pat-tern were overlaid onto a T1 MRI image. Blue (red) indicates regionswith relatively decreased (increased) binding in LRRK2-NMC com-pared to healthy controls. LRRK2-NMC = non-manifesting LRRK2mutation carriers; PPN = pedunculopontine nucleus. . . . . . . . . . 502.6 PD-related spatial patterns in the serotonergic system obtained withA) PCA, B) ICA, C) PLS-DA, and D) RPCA methods. Error barswere calculated with leave-one-out cross validation. . . . . . . . . . . 572.7 Expression of the spatial patterns obtained with the PCA approach.Left: Subject scores (z-transformed) for healthy controls (label 0)and sPD subjects (label 1). Right: Scatter plot for disease duration(years) versus subject scores for sPD subjects. . . . . . . . . . . . . . 582.8 Illustration for PCA versus ICA decomposition. . . . . . . . . . . . . 582.9 Expression of the spatial patterns obtained with the ICA approach.Left: Subject scores (z-transformed) for healthy controls (label 0)and sPD subjects (label 1). Right: Scatter plot for disease duration(years) versus subject scores for sPD subjects. . . . . . . . . . . . . . 592.10 Illustration for PCA versus PLS decomposition. The red and purpledots indicate two different groups of subjects (predictors). . . . . . . 602.11 Expression of the spatial patterns obtained with the PLS-DA ap-proach. Left: Subject scores (z-transformed) for healthy controls(label 0) and sPD subjects (label 1). Right: Scatter plot for diseaseduration (years) versus subject scores for sPD subjects. . . . . . . . 612.12 Illustration for PCA versus RPCA decomposition. . . . . . . . . . . 622.13 Expression of the spatial patterns obtained with the RPCA approach.Left: Subject scores (z-transformed) for healthy controls (label 0)and sPD subjects (label 1). Right: Scatter plot for disease duration(years) versus subject scores for sPD subjects. . . . . . . . . . . . . . 63xviii3.1 A) Full correlation matrices (ROI by ROI) for the four subject groupsat threshold = 0.3 (when taking the top 30% of the strongest corre-lations in each subject group) which was used as input for furtheranalyses; ROI names are listed in Table3.1 B) Histograms of the ab-solute values of the full correlation coefficients for the four subjectgroups without thresholding; red vertical lines represent mean valuesfor all correlation coefficients in each subject group. . . . . . . . . . 723.2 Graph theory metrics for the four subject groups measured at dif-ferent sparsity thresholds (20%-50%). Error bars were obtained withleave-one-out cross-validation. . . . . . . . . . . . . . . . . . . . . . . 743.3 Graph theory metrics for the four subject groups measured at 30%sparsity thresholds. Error bars were obtained with leave-one-outcross-validation. * = p-value Q0.05. ** = p-value Q0.01. Signifi-cance was estimated with 100 iterations of random permutation test. 754.1 [11C](+)dihydrotetrabenazine (DTBZ) PET image for a healthy con-trol (left) and a Parkinson’s disease (PD) subject (right). PD subjectshowed characteristic asymmetric tracer uptake in the less and moreaffected sides and a spatio-temporal pattern of dopaminergic loss withthe posterior putamen affected before the anterior putamen and cau-date. PET = Positron Emission Tomography. . . . . . . . . . . . . . 824.2 Schematic diagram for dynamic mode decomposition (DMD) analysispipeline. (A) In the data preparation step, the 3D parametric PETtracer binding image of each subject is stretched into a flattened col-umn vector. Each column vector is then concatenated horizontallyaccording to the disease durations of all subjects. DMD then decom-poses the reshaped PET data into DMD modes (spatial patterns),each associated with an unique temporal dynamic curves. (B) Thereshaped PET data are used to construct mt−1 and mt matrices astime shifted version of each other, which are then used as input toDMD. PET = Positron Emission Tomography. . . . . . . . . . . . . 87xix4.3 DMD modes (spatial patterns) in the less and more affected putamen(A) and caudate (B). DMD was applied to the less and more affectedsides separately. In the putamen, DMD mode 1 showed an anterior-posterior gradient and DMD mode 2 showed a dorsal-ventral gradientin both the less and more affected sides. DMD mode 1 in the caudateshowed a head-tail gradient in both the less and more affected sides.Spatial patterns are displayed as maximum intensity projection ontothe entire region of interest. DMD = dynamic mode decomposition.LM = lateral-medial. AP = anterior-posterior. DV = dorsal-ventral. 924.4 (A) The first DMD temporal curve in the less and more affected puta-men, associated with the anterior-posterior gradient. (B) The secondDMD temporal curve in the less and more affected putamen, associ-ated with the dorsal-ventral gradient. (C) Combined DMD temporalcurves for the first and second DMD modes in the less and more af-fected putamen. (D) Averaged DTBZ activity ratios in the less andmore affected putamen versus disease duration and the best expo-nential fit curve. Error bars were generated from leave-one-out crossvalidation. DMD = dynamic mode decomposition. DTBZ = dihy-drotetrabenazine. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 934.5 (A) The first DMD temporal curve in the less and more affected cau-date, is associated with the head-tail gradient. (B) Averaged DTBZactivity ratios in the less and more affected caudate versus diseaseduration and the best exponential fit curve. Error bars were gen-erated from leave-one-out cross validation. DMD= dynamic modedecomposition. DTBZ = dihydrotetrabenazine. . . . . . . . . . . . . 944.6 Comparison between DMD temporal expression of mode 1 (left) andmode 2 (right) and PCA scores of PCA pattern 1 (left) and PCApattern 2 (right) in the less affected putamen. Both DMD temporalexpressions and PCA scores are Z-transformed. Spatial patterns aredisplayed as maximum intensity projection onto the entire region ofinterest. DMD = dynamic mode decomposition. PCA = principalcomponent analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95xx5.1 [11C]-dihydrotetrabenazine (DTBZ) PET image (left) and [11C]d-threo-methylphenidate (MP) PET image (right) for a Parkinson’sdisease (PD) subject. PD subject showed characteristic asymmetrictracer uptake in the less and more affected hemispheres. PD subjectalso showed spatio-temporal pattern of dopaminergic loss with theposterior putamen (putamen 3) affected before the anterior putamen(putamen 1) and caudate. . . . . . . . . . . . . . . . . . . . . . . . . 1035.2 Illustration of the decomposition and regression step. X and Y arethe whitened input matrices (feature by subject) of non-displaceablebinding potential (WeaW) values obtained from step 2. The trans-formed data (canonical variates) U and V are calculated using CCAin step 3, which contains the most highly correlated subject scoresalong each component (in this case 5). The CCA weights matrices (Aand B) are the regression coefficients from least absolute shrinkageand selection operator (LASSO) in step 4. Xresidual and Yresidualare the regression residuals. CCA = canonical correlation analysis. . 1095.3 Scatter plots for average DTBZ and MP WeaW values in the lessaffected putamen versus disease duration (estimated from the timeof symptoms onset) in months. Both DTBZ (left) and MP (right)WeaW values correlated significantly with disease duration. S15 felloutside the 95% confidence interval. WeaW = non-displaceable bind-ing potential. DTBZ = dihydrotetrabenazine. MP = methylphenidate.1125.4 Common spatial patterns along the first three pairs of canonical vari-ates for DTBZ and MP. Stars indicate the ROIs with significant CCAloadings. ROI = region of interest; CCA = canonical correlation anal-ysis; GP = globus pallidus; VS = ventral striatum; SN=substantia ni-gra; VTA = ventral tegmental area. DTBZ = dihydrotetrabenazine.MP = methylphenidate. . . . . . . . . . . . . . . . . . . . . . . . . . 1145.5 Correlation between subject scores and disease duration as estimatedfrom the time of symptoms onset (months) for DTBZ and MP alongthe third pair of canonical variates. DTBZ = dihydrotetrabenazine.MP = methylphenidate. . . . . . . . . . . . . . . . . . . . . . . . . . 116xxi5.6 Unique spatial patterns along the first three pairs of canonical variatesfor DTBZ (top) and MP (bottom). Stars indicate the ROIs withsignificant CCA loadings. ROI = region of interest; CCA = canonicalcorrelation analysis; GP = globus pallidus; SN = substantia nigra;VS = ventral striatum; VTA = ventral tegmental area. DTBZ =dihydrotetrabenazine. MP = methylphenidate. . . . . . . . . . . . . 1176.1 MRI and four different images (averaged concentration) from PETscans done on an example Parkinsons disease subject. DTBZ = Di-hydrotetrabenazine. MP = Methylphenidate. DASB = 3-amino-4-(2-dimethylaminomethylphenylsulfanyl)-benzonitrile. RAC = Raclopride.1296.2 Schematic diagram for the joint pattern analysis pipeline. Four sets ofinput data included the binding potential values (WeaW) for DTBZ,MP and DASB and dopamine release values in eight striatal regions.The joint pattern analysis was applied onto all tracer datasets. Theoutputs of the joint pattern analysis are defined by highly correlatedsubject scores across all tracers for each set of canonical variates andthe associated spatial patterns for each tracer for each set of canonicalvariates. MCCA = Multi-set canonical correlation analysis. PD =Parkinsons disease. VMAT2 = vesicular monoamine transporter 2.DAT = dopamine transporter. SERT = serotonin transporter. DA= dopamine release. . . . . . . . . . . . . . . . . . . . . . . . . . . . 1326.3 Percentage dopamine release estimated 1 h after levodopa intake inthe eight striatal regions. * indicates P-value Q0.05. ** indicatesP-value Q0.01. P1 = anterior putamen. P2 = middle putamen. P3= posterior putamen. . . . . . . . . . . . . . . . . . . . . . . . . . . 1346.4 Correlation between percentage dopamine release1h and motor re-sponse to levodopa (LD) in the more affected anterior putamen (A)and more affected middle putamen (B). . . . . . . . . . . . . . . . . 1356.5 A) Correlations between the subject scores for DTBZ, MP and DASBspatial patterns and dopamine release1h spatial pattern along the firstset of canonical variates. B) Spatial patterns along the first set ofcanonical variates. DA = dopamine. MCCA = Multi-set canonicalcorrelation analysis. P1 = anterior putamen. P2 = middle putamen.P3 = posterior putamen . . . . . . . . . . . . . . . . . . . . . . . . . 136xxii6.6 Correlations between motor response to levodopa (LD) and subjectscores for the dopamine release1h pattern (A), subject scores for theDTBZ pattern (B) and disease duration (C) along the first set ofcanonical variates. DA = dopamine. LD = levodopa. . . . . . . . . . 1376.7 Projection scores on the spatial patterns along the first set of canon-ical variates for subjects who were stable or developed dyskinesiasyears after the baseline scans and subjects who are currently not duefor the follow-up. ** indicates P-value Q0.01. . . . . . . . . . . . . 1377.1 Overview for the thesis. . . . . . . . . . . . . . . . . . . . . . . . . . 143xxiiiList of UvvryviutionsFM[Byopv 6-[18F]-fluoro-L-dopa. 49VIC Akaike information criterion. 45, 69VcCdkV analysis of covariance. 45, 46VcdkV analysis of variance. 43, 131Vg activity ratio. 84, 86, 88–91, 94, 96, 97, 99WYI Beck depression inventory. 41, 66, 106, 128WbI body mass index. 38, 40, 66, 106, 128We binding potential. 14, 15WeaW non-displaceable binding potential. 15, 16, 44, 46, 50, 51, 66, 68, 77, 99,107, 108, 112, 113, 115, 118, 119, 123, 130, 145CCV canonical correlation analysis. 104, 105, 108, 110–112, 123Ci computed tomography. 4, 9YV dopamine. 102–105, 122YVhW 3-amino-4-(2-dimethylaminomethyl- phenylsulfanyl)-benzonitrile. 22, 24–26, 36–38, 40, 43, 44, 46, 49–56, 61, 64, 66–71, 125, 127–133, 135, 138, 139YVi dopamine transporter. 20, 21, 24, 101, 105, 106, 108, 118–120, 122, 123YbY dynamic mode decomposition. xii, 2, 31, 81, 82, 84, 85, 88–91, 93–100xxivYiWo dihydrotetrabenazine. 20, 21, 24, 36, 40, 43, 46, 47, 49, 51, 64, 79, 81, 83,84, 89, 90, 95, 97, 99, 105–108, 112, 113, 115–125, 127–133, 135, 136, 139–141ZZG electroencephalography. 31, 34, 65, 81, 86, 104, 105ZbG electromyography. 104, 105[YG fluorodeoxyglucose. 25, 28, 30, 39, 40, 55, 66, 77fbgI functional magnetic resonance imaging. 26, 30, 31, 34, 52, 55, 65, 81, 86,104[l]b full-width at half maximum. 12, 43, 67, 83, 107, 129GaVhhd graphical least absolute shrinkage and selection operator. 69, 71]&n Hoehn and Yahr scale. 41, 66, 83]ggi high resolution research tomography. 9, 12, 43, 67, 83, 106, 128ICV independent component analysis. 59, 80, 81, 104jICV joint independent component analysis. 104aVhhd least absolute shrinkage and selection operator. 111, 142adg line of response. 4, 7–11aggKG leucine-rich repeat kinase 2. 18, 36–38, 43, 46, 52, 54, 66, 78, 144aggKGBcbC non-manifesting LRRK2 mutation carriers. xviii, 22, 38–40, 46,47, 49–54, 64–66, 71, 74–78aggKGBeY Parkinson’s patients with LRRK2 mutation. 38–40, 43, 46, 47, 49,51, 53–55, 64, 66, 71, 74, 76bCCV multi-set canonical correlation analysis. 32–34, 131bYhBjeYgh movement disorder society unified Parkinsons disease rating scalepart III. 41, 43, 66, 83, 84, 106, 128, 131, 133, 139, 140xxvbcI Montreal Neurological Institute. 44, 67, 68, 107, 130boCV Montreal cognitive assessment. 41, 66, 83, 106, 128be d-threo-methylphenidate. 20, 21, 24, 105–108, 112, 113, 115, 116, 118–125,127–133, 135bgI magnetic resonance imaging. 4, 34, 43, 65, 67, 83, 84, 104, 106, 107, 128–130deBZb ordinary Poisson expectation maximization. 11deBdhZb ordinary Poisson ordered subset expectation maximization. 11, 43, 67,83, 107, 129dhC orthogonal signal correction. 34, 105, 108, 111eC principal component. 27, 28, 45, 46, 49, 55, 56eCV principal component analysis. xii, 2, 26–28, 31, 32, 36, 39, 45, 55, 56, 59–63,80, 81, 84, 89, 90, 94, 96, 99, 104, 123eY Parkinson’s disease. 2, 17–25, 28, 30, 32, 34–41, 46, 49, 51–56, 61, 63–66, 74,76, 78, 79, 81, 83, 88, 95, 98, 100–106, 119, 120, 122, 123, 126, 127, 143, 144eYCe Parkinson’s disease-related cognitive pattern. 28, 40, 55eYge Parkinson’s disease-related pattern. 28, 39, 55eZi positron emission tomography. 1–6, 9, 12, 15, 19, 22, 24–26, 28, 30, 34, 36,37, 39, 40, 43, 44, 63–68, 77, 79–84, 88, 90, 95, 96, 98, 100, 101, 104–107, 112,119, 125, 127–131, 139, 143, 146eahBYV partial least square discriminative analysis. 60, 61ebis photomultiplier tubes. 7eec pedunculopontine nucleus. 44, 53, 68gVC raclopride. 21, 22, 24, 125, 127–130gdI region of interest. 12, 16, 26, 44, 45, 50, 54, 55, 65, 67–69, 80, 81, 107, 108,112, 115, 118, 123, 130xxvigeCV robust principal component analysis. 62, 63gib reference tissue model. 13, 15hZgi serotonin transporter. 22, 37, 39, 40, 49, 51, 52, 54, 76seY sporadic Parkinson’s disease. 38–40, 43, 46–49, 51, 53, 55, 56, 60, 64, 66, 71,74heYge serotonergic Parkinsons disease-related pattern. 40, 46, 47, 49, 51–55, 64heZCi single-photon emission computed tomography. 80, 81hgib simplified reference tissue model. 15, 16, 44, 68, 130hgibG two-step simplified reference tissue model. 44hhb scaled subprofile model. 28, 39, 40, 44, 46hkY singular value decomposition. 27hlZYY subject without evidence of dopaminergic deficit. 49iVC time activity curve. 12, 16, 44, 68i[Z turbo field echo. 43, 67, 107, 129ieg topographic profile rating. 45jeYgh unified Parkinson’s disease rating Scale. 28, 39, 49kbViG vesicular mono-amine transporter type 2. 20, 21, 40, 81, 99, 101, 105,106, 108, 118–120, 122, 128, 141kh ventral striatum. 103, 107, 115, 118, 120kiV ventral tegmental area. 103, 107, 112, 115, 118, 122xxviiUwknowlydgmyntsFirst of all, I would like to thank my supervisor, Professor Vesna Sossi, for yourcontinuous support and mentorship throughout my B.Sc. to Ph.D. years at UBC.Your attention to details and work ethics inspired me to grow professionally andpersonally. Thank you for leading the way along my academic path for first providingme (as a Biophysics undergraduate student) a summer volunteer position in the laband guiding me all the way to a Ph.D. degree in Medical Physics.Secondly, I express my greatest gratitude to my co-supervisor, Professor MartinJ. McKeown, for the constructive feedbacks on the analysis methods and advicefor my career and personal growth. The valuable time we spent on the blackboarddiscussing algorithms pushed me to better understand the algorithms and becomea better coder. Thank you for always having your office door open and having theanswers ready whenever I need help.Further, I would like to thank Professor A. Jon Stoessl for all the valuable inputon both the technical and clinical aspects of my M.Sc. and Ph.D. projects. Yourenthusiasm about research and your care for the patients helped me to shape myresearch focus and interest as a physicist on better understanding neurodegenerativediseases and the brain.Next, thank you Professor Doris Doudet for all the advice and encouragementon how to become a better researcher. Thank you for always introducing me topeople at the conferences and thank you for always sharing your life with us.I would also like to acknowledge my committee members, Professors Urs Hafeli,Shannon Kolind and Ingrid Stairs for their valuable time and constructive feedback.I also would like to thank the current and past members of the UBC PET imaginggroup and the Pacific Parkinson’s Research Centre for their support: Ivan Klyuzhin,Anna Schildt, Nikolay Shenkov, Nasim Vafai, Kevin Cheng, Elham Shahinfard, LucyAceves, Joyce Lam, Katie Dinelle, Siobhan McCormick, Carolyn English, NicoleNeilson, Jess McKenzie, Michele Matorazzo, Mariya Cherkasova, Connor Beving-ton, Rick Kornelsen, Chenoa Mah, Susan Liu, Christina Rodriguez, Agnes Kwok,xxviiiDevavrat Nene, Julia Mannheim, Tilman Wegener, Jordan Hanania, Soojin Lee,Robert Baumeister, Saurabh Garg, Rostom Mabrouk, Andrew Robertson, GregStortz and Ellen Chen.In addition, I would also like to acknowledge the funding support received fromthe NSERC CREATE (TRIUMF IsoSiM) program. Finally, I would like to expressmy endless gratitude towards all the participants in the study.xxixDydiwutionThis thesis is dedicated to my parents.xxxChuptyr EIntroduwtionEBE hhysis cvyrviywThe work described in this thesis consists of seven chapters and has resulted in threefirst author journal publications [11–13], two second author journal publications[14, 15], one first author manuscript under review, one first author manuscript underpreparation, one second author manuscript under review and a number of conferencepresentations. The following paragraphs summarize the key contents in each chapterand an overview diagram for chapters 2-6 is shown in Fig.1.1.Figure 1.1: Overview for the thesis.Chvptzr FO IntroyuxtionThis chapter contains an overview for the background materials associated with thisthesis work. The chapter first starts with the basic principles of positron emission1tomography (PET) imaging, from data collection to image reconstruction, followedby the application of multi-tracer PET imaging in the study of neurotransmitterchanges in Parkinson’s disease (PD), and finally a brief introduction of the networkpattern analyses methods used in this thesis work in order to more effectively extractdisease-related information from the imaging data.Chvptzr GO ainzvr boyzls of hpvtivl evttzrnsIn the work described in this chapter, a dimension reduction method (principalcomponent analysis (PCA)) was applied to extract spatial patterns in PET dataspecifically targeting the serotonergic system. We compared the results obtainedfrom PCA to results obtained from traditional univariate analyses and other simi-lar dimension reduction techniques. We then discussed the advantages of examin-ing network behaviours in multiple brain regions compared to examining localizedchanges in a single brain region. The higher sensibility and robustness obtained withthe PCA method also provide the rational for the development of network patternanalysis methods for this thesis work.Chvptzr HO cztfiork iopolog–We extended the analysis of spatial distribution of PET to study network topologychanges using graph theory analysis in the work done in this chapter. We discussedthe applicability and limitations of the graph theory analysis method for analyzingPET data.Chvptzr IO hpvtioBizmporvl evttzrnsIn the work described in this chapter, we introduced and validated a novel analysismethod (dynamic mode decomposition (DMD)) that models both disease-relatedspatial and temporal changes in PET data. We compared the results obtained fromDMD to results obtained with PCA (method introduced in Chapter2) and resultsobtained with traditional univariate analysis. We then demonstrated the uniquestrength of the proposed method to model temporal changes related to disease pro-gression using PET data.2Chvptzr JO Joint evttzrn Vnvl–sisWe introduced and validated a novel joint pattern analysis approach for extractingcomplementary information from multi-tracer PET data in the work described inthis chapter. We compared the results obtained with this joint pattern approach toresults obtained with applying traditional univariate analysis to each PET datasetseparately. We demonstrated that the complementary information obtained by ef-fectively combining information multi-tracer PET datasets were more sensitive todisease-induced changes compared to univariate analyses.Chvptzr KO Joint evttzrn Vnvl–sis B VpplixvtionIn the work done in this chapter, we further demonstrated the applicability of theproposed joint pattern analysis approach in a more complex multi-tracer PET studyinvolving five different PET scans targeting different aspects of the brain.Chvptzr LO ConxlusionThis chapter summarizes the key results in this thesis work and potential futureapplications.EBF dositron Emission homogruphyFunctional neuroimaging with PET is among the most sensitive ways to study neu-ropathology of various neurodegenerative diseases in vivo. In this section, basicprinciples behind PET radiotracers, radioisotope decays, signal detection, data cor-rections and image reconstruction required to obtain quantitative measurements ofbrain function will be reviewed.dvzrvizfiPET is a nuclear imaging technique which uses spatial and temporal distributions ofradiotracers to provide functional information in the tissues of interest. Radiotrac-ers are typically made by attaching positron emitting radioisotopes to a biologicalrelevant molecule. Once the radiotracers are injected and carried to the site of inter-est by the blood, the emitted positrons travel a short distance before annihilation.Annihilation between a positron and a nearby electron produces two 511 keV gammarays or annihilation photons, which are emitted in opposite directions (to a good3approximation) along a straight line path. The imaginary line that joins the thedetector pair where the two gamma rays are detected within the coincidence windowis used to assign those matching gamma rays to a specific line of response (LOR).The measured annihilation photon pairs or ’coincidence events’ of radioactive decaysalong each LOR are then used to generate a 3D volumetric image of the object beingstudied via various image reconstruction algorithms. Comparing to other commonneuroimaging modalities, PET has the highest sensitivity and specificity for detect-ing molecular targets, so is ideal for studying changes in neurotransmitter systems.However, PET has relatively low temporal and spatial resolution compared to imag-ing modalities like magnetic resonance imaging (MRI) and computed tomography(CT).EBFBE fudiotruwyrA radiotracer is a biological molecule tagged with a radioisotope. Positron-emittingisotopes commonly used in medical research are often produced by cyclotrons orgenerators. Short-lived isotopes, such as 11X with a half-life of 20O4min, have to beproduced by an in-house or nearby cyclotron. In this project, all tracers are taggedwith 11X radioisotope, which was produced at TRIUMF (TRI-University MesonFacility). TRIUMF is Canada’s particle accelerator centre and the home for thelargest cyclotron in the world.Isotopes suitable for PET imaging often have the following characteristics [16]:1. optimal half-lives to capture the kinetics of the radiotracer2. emitted positrons have relatively low energy and short range (distance trav-eled) before annihilation to reduce the degradation effect on spatial resolutionin the image3. isotope production easily available4. do not change the biochemical properties of the labeled moleculesThe choice of tracers/molecules depends on the tissue to be studied and questionsof interest. In this project, we used various radiotracers to target different aspects ofthe neurotransmitter systems of the brain (more details will be given in Section1.5).Ideally a tracer should have the following properties [16]:1. high specificity to the biological process or site of interest42. minimal radiolabelled metabolites (products that remain after a tracer is bro-ken down) to avoid isotope being carried away from the site of interest if themetabolites are not trapped3. relatively easy synthesis process that takes relatively short-time compared tothe isotope decay time and has a good yieldTo summarize, radiotracers should have the same physiological properties andthus the same tracer kinetics as the unlabeled biological molecules such that theydo not alter the functions of normal tissues. The ’trace’ amount of radiotracersadministrated into the patients should not induce any pharmacological effect norperturb the biological process of interest [17] [16]. In the next section, the mappingof the spatial and temporal distributions of the radiotracers using the measuredcoincidence events will be described.EBFBF fudioisotopy DywuyAn excessive number of protons in a nucleus leads to imbalance between the attrac-tive and repulsive forces to hold the nucleus permanently together (i.e., unstable).The process in which an unstable nucleus emits particles in order to return to thestable state is termed ’radioactive decay’.In PET, a proton in an unstable nucleus (X) is converted into a neutron byreleasing a positron (+) and a neutrino (). After decaying, the nucleus returns tothe stable form (Y) or undergoes further decay:Tmm =Tm−1 n + z+ +  (1.1)The radioactivity (A) of a sample is the number of decays per unit time (in theunits of Becquerel, decays/second). For a given radioisotope, the probability of de-cay is modelled by an exponential function with a decay constant  (the probabilitythat a nucleus will decay per unit time).c(t) = c0z−tV(t) =−dc(t)dt= c(t)i1P2 =ln(2)(1.2)5where N is the number of radioactive particle in the sample and the half-life (i1P2)is the time when activity of sample has halved.The emitted positron loses kinetic energy via Coulomb’s interaction with thesurrounding electrons resulting in a relatively short traveling distance (r) within themedium before colliding with an electron (annihilation) as shown in Fig.1.2. Thissmall distance (r) is called positron range and can degrade the spatial resolution ofthe PET images. One can correct for the effect of positron range using Monte Carlosimulations to model the positron range distribution for a particular isotope [18].The positron then collides with an electron in a process called annihilation afterlosing most, if not all, of its kinetic energy. Since both the positron and electronhave a rest energy of 511 keV, the annihilated particles are replaced by two gammaray photons with equal energy traveling at opposite directions due to energy andmomentum conservation[1].Figure 1.2: Nucleus decays to emit positron (+) which travels to the site of anni-hilation where it annihilates with an electron (z−) producing two 511 keV gammarays () in opposite directions. r is the displacement from the parent nucleus to thesite of annihilation, whereas p is the actual path of + [1]EBFBG Dytywtion gystymThe two gamma photons originating from the annihilation event traveling in oppo-site directions are detected in coincidence (within a few nanoseconds) by the PETdetectors. The scintillation detectors are composed of crystals that can detect thegamma rays and convert them into light. The incident photon interacts with thecrystals and an electron is emitted by either Compton scatter or photoelectric effect.6This electron loses energy as it travels through the crystal, exciting other electronson its way. The excited electrons then release energy in the form of visible light or’scintillation photons’ and return to the ground state [19].The scintillation photons then pass through the photomultiplier tubes (PMTs),and this process results in a short electrical pulse. This electrical pulse is furtheramplified by the PMTs. When two signals from opposing detectors are detected incoincidence, the two detected signals form a LOR. This information is then sent to acomputer [20]. Ideally, two gamma photons reaching the detectors at the same timeare detected as a coincidence event, however, different times of arrival or temporalmismatches may occur due to the finite timing resolution of the scintillation crystalsand the processing time of PMTs [16]. The temporal mismatches are taken intoaccount by the coincidence time window, usually in the order of 6-10 ns [21]. Inaddition, when annihilation event happens at location closer to one detector thanthe other, there is a slight time delay for one photon compared to the other one.On the other hand, multiple annihilations can happen simultaneously, and the finitecoincidence time window can register photons with different annihilation origins ascoincidence events by accident. These ’accidental’ and random coincidences will bedescribed shortly.Figure 1.3: True coincidence detection vs scatter and random coincidence detectionin PET [2]Coinxiyznxzs YztzxtionNot all coincidence events detected are true annihilation events that can give mean-ingful information to the true state of brain function, some coincidence events resultin degradation in contrast and resolution in the PET images. There are three mainkinds of coincidence events as shown in Fig.1.3.The first type is true coincidence event which occurs when two 511 keV photonsfrom the annihilation event are detected by a pair of scintillators simultaneously7without undergoing any interactions with the surrounding tissues [21]. The secondtype is random coincidence event which occurs when two photons detected withinthe same coincidence window are generated from different annihilation origins. Thethird type is scatter coincidence event which occurs when the direction of the trav-eling photons is altered due to Compton scatter. This results in incorrect LORassignment that does not reflect the true location of the annihilation event. Ran-dom and scattered coincidence events contaminate the PET signal and decreasethe image contrast. Random coincidence events are often corrected with delayedwindow method.ehoton Intzrvxtions vny VttznuvtionThe primary type of interaction for 511 keV photon in tissue is Compton scattering,in which a photon interacts with an outer shell electron. This interaction causes theincident photon to change direction (scattered) and lose energy. The lost energyis transferred to a recoil electron [22]. The energy of the scattered photon afterinteraction is given by [22]:Z′ =Z1 + (ZRm0x2)(1− cos ) (1.3)where E’ is the energy of the scattered photon, E is the energy of incident photon,m0x2 is the rest mass of an electron, and  is the scattering angle.Photoelectric absorption, another type of photon interaction, is less likely tooccur in human tissues for 511 keV photons [23]. In photoelectric absorption, theincident photon is completely absorbed by an atom which then ejects an energeticelectron from its inner shell [22].Together, these interactions lead to photon attenuation, which reduces the num-ber of photons reaching the detector exponentially as the length of the mediumincreases. The probability that a photon will reach the detector is given by:e = exp−∫ x0 (x)wx (1.4)where P is the probability of a photon reaching the detector at distance x throughsome attenuating material, and  is the linear attenuation coefficient.Because the interaction probabilities for two photons originated from a coinci-dence event are independent of each other, the total probability that both photons8will reach the detector and been recorded as a coincidence event is given by [23]:ev = exp− ∫ L0 (x)wx (1.5)where L is the distance between two detectors.Vttznuvtion vny hxvttzr Corrzxtion Attenuation correction is essential foraccurate estimation of PET tracer uptake. The amount of attenuation depends onthe density of the surrounding medium and the distance traveled by the photon pairas in Eq.1.5. Consequently, the effect of attenuation causes fewer annihilations tobe detected near the centre of the object than those near the edge of the object,which introduces a bias in image contrast.To correct for attenuation or recover the loss of the incident photons (mainlydue to Compton scattering within tissue), a transmission scan by a rotating externalpositron source with and without (i.e., blank scan) the attenuating object inside thescanner field of view is performed on the PET scanner to obtain the attenuationmap from all LOR [24]. In combined PET/CT scanner, CT images can be used forattenuation correction [25]. For the high resolution research tomography (HRRT)scanner used in this thesis work, a rotating 137Cs point source is used for attenuationcorrection.Scatter coincidence events can be corrected using complex simulation methods[2]. Note that the scatter correction removes the erroneous events due to Comp-ton scatter, whereas the attenuation correction recovers the loss of events due toCompton scatter.cormvlizvtion Corrzxtion PET scintillators may have varying sensitivity andefficiency for detecting incoming photons due to variabilities in photon incidentangle, imperfections in scintillation crystals and other electronics. A normalizationscan is typically done using a rotating rod source for the HRRT scanner. A detectionnormalization correction factor is assigned to each LOR to account for the detectionvariability.EBFB4 Imugy fywonstruwtionAs mentioned before, the detected/measured number of positron decays (coincidenceevents or counts) along each LOR is a combination of true, random and scattered9coincidence events. During the image reconstruction step, we want to estimate theamount and location of the true events and reconstruct activity counts into a 3Dimage representation of the tracer distribution. The measured number of coincidenceevents along the ith LOR is represented by ni (projections), then:ni =J∑j=1pijj = Fei(j) (1.6)where j the mean activity in voxel j, pij is the system matrix element that describesthe probability of radioactivity originating in voxel j to be detected along LORi. Feiis the forward projection or sum of voxel activity along LORi. The system matrixp may also contain the correction factors for normalization and attenuation. Thegoal of image reconstruction is to transform the measured coincidence (projection)data ni to the volumetric activity image. The number of detected coincidence eventsfollows a Poisson distribution, which models the probability of a number of eventsper time period. Therefore, the noise in the measured PET signal also follows aPoisson distribution. Low number of detected counts can lead to increasing noise inthe data.There are two commonly used types of reconstruction algorithms for PET im-ages, the analytical and iterative reconstruction methods [1][26][16].Vnvl–tixvl gzxonstruxtionRadon transform is the line-integral transformation of the number of coincidenceevents to a projection. In an analytical reconstruction model, one can use the inverseof the discrete Radon transform to directly find an estimation of  (radiotraceractivity image) from the projections n. This direct estimation of  allows for fastcomputation. Prior to reconstruction, the input projections need to be correctedfor attenuation, scatter and random coincidence and detector sensitivities. Theanalytical reconstruction approach does not account for noise within reconstruction.This can lead to increasing noise in the image [27].Itzrvtivz gzxonstruxtionIn reality, noise in the projection data is stochastic and follows a Poisson-like distri-bution as mentioned before. In addition, there is no exact solution for . Therefore,the more commonly used reconstruction method in research currently is the iterative10reconstruction algorithm. Instead of solving for  directly, one estimates the realimage  iteratively using a statistical model.With an iterative maximum likelihood expectation maximization algorithm [28],one starts with an initial guess for the true image, 0 (often an uniformly con-stant image), then forward projects 0 to the projection or LOR space. Theseestimated projections are then compared to the measured projections by taking theratio between them, and an update factor for each image voxel is computed byback-projecting the ratio using all the LORs passing through the given voxel. Theupdate factor is then applied to each activity image voxel and normalized by thesensitivity image voxel (i.e. back-projection of projection value of 1 everywhere) togenerate the next iteration of the activity image estimate.The expectation maximization algorithm for estimating  is shown below:m+1j =mj∑Ii=1 pijI∑i=1(pijniFei(mi )) (1.7)where m+1j is the activity in voxel j in the mth iteration,∑Ii=1(pij) is the knownsystem matrix containing correction factors for normalization and attenuation.To correct for scatter and random coincidence events within reconstruction, theordinary Poisson expectation maximization (OP-EM) algorithm is often used:m+1j =mj∑Ii=1 pijI∑i=1(pijniFei(mi ) + ri + si) (1.8)where ri is the number of estimated random coincidence events and si is the numberof estimated scattered coincidence events along the ith LOR and the algorithmpreserves the Poisson nature of the measured data.As the number of iterations (m) increases, a more accurate representation of thereal image is obtained at the expense of longer computational time and higher noisesince the measured data are noisy [27][26]. In practice, a pre-determined (relativelylow) number of iteration is used followed by post-filtering to control the noise leveland improve the signal-to-noise ratio in the image.OP-EM, however, can be computationally slow when the number of projectionsis large. A modified version of OP-EM, the ordinary Poisson ordered subset expec-tation maximization (OP-OSEM), can reduce the reconstruction time by dividingthe projection data into subsets. OP-EM algorithm is used to reconstruct a subset11of projection data to obtain an image estimate, then the image estimate is passedon as the input for reconstructing the next subset of data until all subsets havebeen processed. Instead of updating the image estimate once after going throughall the projection or LOR data, OP-OSEM updates the image estimate after goingthrough each subset of the projection data. A complete iteration thus consists ofmultiple image updates. The use of multiple subsets to increase accuracy of imageestimate improves the speed of the reconstruction algorithm at the expense of slightincrease in variance [27]. All PET images used in the work described in this thesiswere constructed using OP-OSEM with 16 subsets and 6 iterations (i.e., a total of96 updates).EBFBI dostAfywonstruwtion gmoothingThe noise in the detected coincidence events follows a Poisson-like distribution and isthe dominant source of noise in the PET signal, however there is additional Gaussiannoise from other sources such as electronics. After reconstruction, a Gaussian filteris often applied to improve signal-to-noise ratios in the image. We used a 2 or 3mmfull-width at half maximum (FWHM) Gaussian filter to the images obtained withthe HRRT scanner (with spatial resolution of 2O5mm3) to reduce noise withoutsignificantly degrading the resolution (i.e., FWHM of the filter is similar to theresolution of the scanner).EBG Kinytiw aodylingIn dynamic PET imaging, the measured data are divided into a number of shortframes, i.e. radioactivity is estimated at different time points. A time activitycurve (TAC) is a curve describing the radioactivity concentration over time for eachimage voxel or a specific region of interest (ROI). TAC reflects the dynamic natureof the radiotracer and contains information about various physiological processes,such as tissue perfusion and the interactions between the tracer and the imagingtarget. Quantification of the dynamic behaviours of the tracer distribution requiresmeasurements on the uptake and washout of the tracer in the tissue as well asthe delivery of the tracer from arterial blood into the tissue. In order to quantifythe kinetics of the radiotracer distribution, a biomathematical model (known ascompartmental model) is often used. The quantification of biological parameters of12interest is done by representing different biochemical states of the radiotracer andits metabolites as different compartments.Quantification of tracer kinetics require an input function describing the concen-tration of non-metabolized compounds in the plasma of arterial blood as a functionof time. However, using arterial plasma as input function requires invasive bloodsampling during the scan. The plasma input model is sometimes replaced by refer-ence tissue model (RTM).EBGBE fyfyrynwy hissuy aodylsFigure 1.4: Systematic diagram of the reference tissue model [3]. Using a referenceregion, parameter Xp can be eliminated.Fig.1.4 shows an example of compartmental model known as the RTM [3] thatis commonly used in PET studies of the brain. In this particular model, there arefour different compartments, representing different concentrations of radiotracersin different states. The exchange of molecules between different compartments aredescribed by rate constants K.13One compartment (Xc ) represents the radiotracers in the plasma of the arterialblood, which delivers the injected radiotracers to the site of interest. Xc is theradiotracer concentration in the plasma (i.e. arterial input function) and representsthe availability of the input radiotracers. Once the radiotracers reach tissues, theycan either specifically bind to the site of interest or stay non-specifically bound. Thespecifically bound compartment (Xf) and non-displaceable compartment (XaW) to-gether are called the target tissue compartment. In the non-displaceable compart-ment (XaW), the radiotracers may bind to non-target sites or remain unbound (free).The reference tissue compartment (Xe) is often used as an indirect measure for theinput radiotracer concentration in the plasma without measuring Xc directly. Thechoice of the reference region depends on the radiotracer of interest. The referenceregion is chosen so that there is no specific target binding sites but have a similardistribution volume (ratio of radiotracer concentration in the tissue compartmentand the plasma compartment) as in target compartment (K ′1Rk′2 = K1Rk2) [29][3][1].The goal for dynamic PET studies is to estimate the rate constants (K) de-scribing the exchange of molecules between compartments. For example, change inradiotracer concentrations in the three tissue compartments can be described by thefollowing differential equations:dXaWdt= K1Xc − k2XaW − k3XaW + k4XfdXfdt= k3XaW − k4XfdXedt= K ′1Xc (t)− k′2Xe(t)(1.9)Winying eotzntivlHowever, estimation of individual rate constants (micro-parameters) can be difficult.Estimation of macro-parameters (as combinations of micro-parameters) is more ro-bust. One of such biologically meaningful macro-parameter most commonly used inPET studies of the brain is the binding potential (BP). BP is a combined measureof the availability/density and affinity of target binding sites [30] at equilibrium:We = WmtxRKW (1.10)where Wmtx is the receptor availability (the number of binding sites) and 1RKW isthe binding affinity.14There are several definitions of BP, but in our studies we used non-displaceablebinding potential (BPaW) which is defined as [29]:WeaW =k3k4(1.11)where k3 and k4 are the rate constants for the radiotracer exchanging between thenon-displaceable and the specifically bound compartment. So for a given amountof radiotracer in the target tissue, BPaW is the proportion of radiotracers that arespecifically bound to target binding sites in the tissue at equilibrium.In RTM, there are four independent parameters: 1) g1 = K1RK′1, the ratio oftracer delivery rates between the target and reference region; 2) k2, rate constantfor transporting tracers from the target tissue to plasma; 3) k3, rate constant fortransporting tracers from the non-displaceable to specifically bound tissue compart-ment; 4) BPaW, ratio of tracer transporting rates between the non-displaceable andspecifically bound tissue compartments. However, estimating all four parameterscan be complex and non-robust when the noise level is high.EBGBF gimpliyd fyfyrynwy hissuy aodylMany PET studies use simplified reference tissue model (SRTM) instead to reducethe number of parameters. In SRTM, the non-displaceable (XaW) and specificallybound (Xf) compartments are combined into a single tissue compartment (Xt) [3][31]. This simplification eliminates k3 and reduces the number of parameters tobe three (g1 = K1RK′1, k2 and BPaW) instead of four. The change of radiotracerconcentration in the tissue compartment is described as:dXt(t)dt= K1Xp(t)− k2tXt(t) (1.12)where k2t is the overall rate constant between the tissue and plasma compartments.The ratio of the exchange rates between the plasma and tissue compartments isthen:K1Rk2t = (K1Rk2)(1 +WeaW) (1.13)Taking all together, the following expression can be derived:Xt(t) = g1Xe(t) + [k2 − g1k21 +WeaW]Xe(t)z−k2tP1+BcND (1.14)15The three parameters of interest (g1 = K1RK′1, k2 and BPaW) can then be estimatedby minimizing the sum of squared errors between the model fit and TAC for eachpre-defined ROI or voxel using non-linear regression.EBGBG Logun Gruphiwul UnulysisLogan graphical analysis is another way to estimate the parameters of interest.However, instead of solving a non-linear regression problem as in the case of SRTM,Logan graphical analysis estimates the parameters of interest in a linearized format.Estimating parameters in a linear model is much faster and is not affected by localminima. Logan graphical analysis transforms multiple temporal measurements ofplasma and tissue uptake into a linear plot [32]. For a two tissue compartment modelshown in Fig.1.4, the radioactivity concentration in a given ROI can be describedas:gdI(t) = XaW(t) + Xf(t) + kc (t)Xc (t) (1.15)where kc is the regional blood volume.Then, combining Eq.1.15 with the differential equations describing the changein radiotracer concentration between different compartments (Eq.1.9), we get:∫ t0 gdI(t′)dt′gdI(t)= [K1k2(1 +k3k4) + kc ]∫ t0 Xc (t′)dt′gdI(t)+ int (1.16)Based on this equation, a plot of∫ P0 ebI(t′)wt′ebI(t) versus∫ P0 VP (t′)wt′ebI(t) is linear after sometime t∗ (depends on the tracer).For two compartment model, assuming kp is negligible, the ratio of the slopes ofthe linear plot obtained in target and reference region becomes:K1k2(1 + k3k4 )Krxf1 Rkrxf2= 1 +k3k4= 1 +WeaW (1.17)where the dependence on plasma volume is removed because K1Rk2 ≈ 1Rfp (fp isthe free/unbound tracer fraction to plasma proteins).Compared to SRTM, the Logan graphical analysis method is independent of anyspecific compartmental models. However, there is a noise-induced negative bias inthe BPaW values obtained with Logan graphical analysis which can be particularlyproblematic at voxel-level.16EBGB4 gummuryIn summary, the coincidence events measured by the PET detectors are recon-structed into a 3D volumetric radioactivity image while correcting for scatter andrandom coincidence events, attenuation and detector variabilities. The 3D volu-metric radioactivity image at each time frame reflects the dynamic behaviours (orkinetics) of the PET tracer distribution. Careful quantification of the tracer kineticswith mathematical models allows us to obtain accurate estimates of specific aspectof brain function. In the work presented in this thesis, we used both SRTM andLogan graphical analysis to investigate disease-induced changes in various neuro-transmitter systems for PD.EB4 durkinson's DisyusyPD is the second most common progressive neurodegenerative disorder and affectsapproximately 0.3% of the entire population. In particular, PD affects approxi-mately 100,000 Canadians with a cost of about $2.5-5 billion annually. Some riskfactors include genetic mutation, age, smoking, gender, diet and alcohol consump-tion [33]. Despite a large body of literature on PD, the disease origin is still unclear.EB4BE gymptomsThe most common symptoms of PD are motor deficits, manifesting in the form ofresting tremor, rigidity and bradykinesia. These motor symptoms are traditionallyassociated with dopaminergic denervation in the nigrostriatal neurons [6]. Motorsymptoms start to appear when 50% of dopaminergic neurons have died whichresults in approximately 80% reduction in the striatal dopaminergic content [34].This suggests a relatively long prodromal stage where the patients show no motorsymptoms but several disease-related neurochemical changes take place.In addition to the motor symptoms, there is increased recognition of the im-portance of disease-related non-motor symptoms, including depression, cognitivedecline, sleep disturbance and autonomic dysfunction. Some of these non-motorsymptoms can precede the occurrence of motor symptoms by years or even decades[35] [11]. Exact causes for the non-motor symptoms are largely unknown, but it wassuggested that they may be associated with alterations in the non-dopaminergicneurotransmitter systems. In particular, several inAvivo and post-mortem studies17showed that progressive alterations in the serotonergic system are related to severalnon-motor symptoms in PD [36][11].Therefore, studying progressive disease-related alterations in the dopaminergicand non-dopaminergic systems in early disease and prodromal stage may providenew insights into the mechanisms related to disease origin and the occurrence ofnon-motor symptoms. However, studying subjects in the prodromal stage wherethere are no motor symptoms can be challenging. A more effecient way to studythe prodromal stage is to look at subjects at a higher risk of developing PD.EB4BF LyuwinyAfiwh fypyut Kinusy F (LffKF) aututionMutations in leucine-rich repeat kinase 2 (LRRK2) genes are the most commongenetic risk factors of PD and the most common mutation of the LRRK2 gene isG2019S. The prevalence of LRRK2 G2019S mutation is approximately 4% in fa-milial cases and 1% in sporadic cases [37][38][39]. The penetrance of the mutationdepends on various factors such as age, ethnic group and environmental modifiers[40]. In the manifest stage, PD patients with and without LRRK2 mutation showsimilar patterns of dopaminergic denervation [41] [42]; however, LRRK2 mutation istypically associated with a lower non-motor symptoms burden before disease onsetand a slower progression of motor symptoms after disease onset [43][37]. In addition,asymptomatic LRRK2 mutation carriers also show increased dopamine turnover, in-creased serotonergic function and increased cholinergic function compared to healthycontrols [44–46].Compensatory mechanisms in PD are not well understood, however they areexpected to play a major role in determining disease progression in both prodro-mal and manifest stages. Upregulation in neurotransmitter functions observed inasymptomatic LRRK2 mutation carriers may reflect mutation-specific effects or riskfactors for PD. Taken together, studying changes in the LRRK2 mutation carriersin the asymptomatic and manifest stages can provide new insights into the contribu-tion of this genetic mutation to PD. By generalizing the observations obtained withasymptomatic LRRK2 mutation carriers to the prodromal stage of sporadic PD, wecan also gain new understanding on the disease origin and potential compensatoryeffects against disease-inducing mechanisms.18EB4BG hryutmyntThere is currently no cure for PD and medications are used to primarily reducemotor symptoms. The most common medication is dopamine replacement ther-apy, most commonly in the form of levodopa or dopamine agonist. Levodopa is adopamine precursor and can be converted into dopamine. Since the occurrence ofmotor symptoms is due to dopamine depletion, administration of levodopa helps totemporarily diminish the motor symptoms. Dopamine agonists bind to dopaminereceptors and directly activate the dopamine receptors to produce a pharmacologicalresponse [16]. Patients generally respond well to treatment in early stage; however,these treatments can lead to severe treatment-induced complications in 5-10 years.The most common levodopa-related complication are dyskinesia (involuntary mus-cle movement), motor fluctuations (wearing on and off of medication effect) andimpulse control disorders which can greatly impact the patients’ quality of life.EB4B4 gummuryStudying neurochemical changes in both prodromal and manifest stages of PD pro-vides the following important information. First, it helps to identify sequences ofpathological changes in different neurotransmitter systems which can provide in-sights into the pathogenesis and mechanisms of PD. Second, studying the relation-ships between multiple neurotransmitter systems allows better understanding of thenon-motor deficits and treatment-induced complications. Third, continued longitu-dinal follow-up of at-risk subjects and manifest subjects improves the estimationof the duration of the prodromal period as well as better understanding of diseaseprogression. All these implications can potentially lead to finding a cure for PD.In particular, we focused on two neurotransmitter pathways closely related to themotor and non-motor symptoms in PD: the dopaminergic and serotonergic systems.EBI byurotrunsmittyr gystyms und dEh fudiotruwyrsPET with dopaminergic tracers provides the most direct way to identify dopaminedeficiency and track the dopaminergic aspect of disease progression. Non-dopaminergicPET tracers look at neurochemical changes that occur before or alongside abnor-malities in the dopaminergic system, and could provide additional insights into theneuropathology of the disease, development of non-motor deficits and assessment of19compensatory strategies. In order to fully understand the neurochemical changesand relationships between multiple neurotransmitter systems in both the prodro-mal and manifest stage, we aim to examine 1) presynaptic dopaminergic functionwhich is a more direct measure of disease effect and disease progression, 2) seroton-ergic function which is associated with non-motor symptoms of PD and providesinsights about potential compensatory mechanisms in the serotonergic system, and3) dopamine release in response to pharmacological stimuli which can be used toexamine medication response.EBIBE Dopuminyrgiw gystymFigure 1.5: VMAT2 is labeled by [11X]-DTBZ, and the membrane DAT can belabeled by [11X]-MP. Dopamine D2 receptor availability can be labeled by [11X]-RAC, which is sensitive to synaptic levels of dopamine [4]irvxzrs for zstimvting przs–nvptix yopvminz funxtion[11X]-dihydrotetrabenazine (DTBZ) and [11X]-d-threo-methylphenidate (MP)are commonly used to estimate presynaptic dopaminergic denervation in PD. di-hydrotetrabenazine (DTBZ) labels the vesicular mono-amine transporter type 2(VMAT2) binding sites located on the presynaptic vesicles and MP labels the mem-brane dopamine transporter (DAT) (Fig.1.5). In keeping with the distribution ofnigral cell loss and with postmortem neurochemical studies of PD, both markers ofpresynaptic dopamine function show a characteristic pattern of asymmetric and pro-gressive reduction of tracer uptake going from posterior to anterior putamen, with20Figure 1.6: Neuronal projections of four dopamine systems in human brain [5]the caudate nucleus remaining relatively preserved (Fig.1.7 C). For both tracers,there is an exponential decline of tracer binding as a function of disease durationwith the decline being most rapid in early disease stage.DTBZ tracer predominantly binds to VMAT2 in the dopaminergic terminals.Major disease or treatment induced regulatory changes do not affect VMAT2 density,so VMAT2 density is deemed the most direct measure of disease progression. Onthe other hand, DAT was shown to contribute to maintaining relatively constantsynaptic dopamine levels by removing extracellular dopamine [47, 48]. MP bindingtherefore may reflect certain degree of regulatory functional changes in populationsat higher risk of PD [41] [49] (Fig.1.7 B) and in early PD [50] [51].irvxzrs for zstimvting yopvminz rzlzvsz[11X]-raclopride (RAC) is a dopamine D2 receptor antagonist (Fig.1.5) and iscommonly used to evaluate dopamine release (changes in synaptic dopamine con-centrations) in response to a range of pharmacological or physiological stimuli. Theantagonist binds to the D2 receptor and competes with dopamine released in thesynapse in response to a stimulus, resulting in lower RAC binding during scans21Figure 1.7: DAT binding to healthy controls (A), LRRK2-NMC (B) and LRRK2-PD(C). LRRK2-PD shows most reduced DAT binding throughout the striatum (C).Striatal DAT binding is also reduced in LRRK2-NMC compared to healthy controls(B) [6]performed after a stimulus has been administrated [52].Different disease-induced changes in the synaptic dopamine levels may be relatedto different patient responses to therapy [52]. It was shown that in advanced PDsubjects (S10 years disease duration), improvement in bradykinesia and rigidityscores (but not tremor) following dopamine medication administration significantlycorrelated with reduction in RAC binding, suggesting an increased dopamine releaseinto the synapse [53, 54]. The same study also showed increased dopamine release inpatients with dyskinesia compared to patients without dyskinesia. In a PD cohortwith milder disease (average disease duration 7.9 years), patients with levodopa-induced dyskinesia were also shown to have greater dopamine release compared topatients with a stable medication response 1 hour after levodopa administrationbut negligible level of dopamine release at 4 hours [55]. However, the relationshipbetween dopamine release and response to levodopa and risk of dyskinesia in earlyPD (Q5 years of disease duration) where the dopaminergic function is still relativelyintact is largely unknown.EBIBF gyrotonyrgiw gystymSeveral PET studies using the radioligand [11X]-3-amino-4-(2-dimethylaminomethyl-phenylsulfanyl)-benzonitrile (DASB), which binds to the serotonin transporter (SERT),have provided evidence for the serotonergic system involvement in several non-motordeficits associated with PD [36]. Serotonergic neurons originate from the raphe nu-clei of the brainstem and project to the entire brain (Fig.1.8); the serotonergicsystem is hypothesized to be affected prior to the dopaminergic system and asso-ciated with some prodromal deficits of PD [56] [57]. On the other hand, LRRK2-NMC showed increased DASB binding compared to healthy controls [44], suggesting22Figure 1.8: Serotonin projections in the brain. Serotonin is produced in nucleusraphe and projects onto other brain regions [7]possible compensatory mechanism in the serotonergic system before disease onset.Further investigation of the serotonergic system thus can provide useful insights intothe role of the serotonergic system in disease origin and progression as related tomotor and non-motor deficits, and possible regulatory mechanisms prior to onset ofmotor symptoms.In addition, serotonergic neurons can release dopamine in denervated striatumin an unregulated manner in PD through the ’false neurotransmitter’ hypothesis [58]and can also release levodopa-derived dopamine. When there is a normal amount ofdopaminergic neurons, the dopamine released from the serotonergic neurons is neg-ligible; however when there is significant loss of dopaminergic neurons, dopaminereleased from the serotonergic neurons has a larger impact. In the presence ofdopaminergic denervation, this unregulated release of dopamine from the seroton-ergic neurons causes large swings in synaptic dopamine level and may play a rolein levodopa-induced dyskinesia. In advanced PD, it was shown that serotonergicneurons contribute to abnormal dopamine release following Levodopa administra-tion, which then leads to higher risk of levodopa-induced motor complications [59].23However, the contribution of the serotonergic system to dopamine release in earlyPD, when there is still relative preservation of the dopaminergic terminals, is notknown.EBIBG gummuryIn this thesis project, we examined the applicability of several spatial network anal-ysis methods to explore the distributed changes in the serotonergic system in theearly and prodromal stages of PD using DASB tracer in the work described inChapter2 and Chapter3. We then used a spatio-temporal network analysis methodon DTBZ tracer to explore the disease-induced progressive dopaminergic deficit tobetter track disease progression in the work described in Chapter4. The progres-sive changes in the dopaminergic system may identify sequence of neurochemicalchanges related to different underlying disease-related mechanisms throughout thecourse of the disease. We then introduced a joint pattern analysis to explore thefunctional differences between the images obtained with two presynaptic dopamin-ergic tracers (DTBZ and MP) to investigate the DAT-related functional changes inearly PD in the work done in Chapter5. Finally, we investigated the relationshipsbetween dopaminergic function (measured by DTBZ and MP), serotonergic function(measured by DASB) and Levodopa-induced dopamine release (estimated by RAC)and their contributions to motor response to levodopa and future risk of treatment-induced motor complications in early PD in the work described in Chapter6. Sincestudying subtle or distributed changes with traditional analyses (e.g. T-test) canbe difficult, especially in early and prodromal stages of the disease; in the next sec-tion, the motivation for using network pattern analysis for analyzing disease-relatedchanges in PET data will be discussed.EBJ bytwork duttyrn UnulysysThe nervous systems are comprised of complex networks with rapidly changingneural activation patterns distributed across numerous cortical and sub-cortical re-gions. When applied to study brain functions, patterns represent the dynamic rela-tionships of neurons resulting in statistical dependencies (functional connectivity).These functional connectivities measure the correlation or covariance between brainregions without explicit reference to causal effects. The functional connectivitiesoften change rapidly due to the changing participation of different subsets of brain24regions under cognitive and behavioral tasks, attentional states and changes due tovarious neurodegenerative diseases. Functional connectivities are also closely linkedto structural connectivities: structural connections serve as fundamentals for func-tional properties, while functional connections help to shape the underlying struc-tural topology via synaptic modification and/or influencing cognitive or behavioralcapabilities in long terms [9] [60].Network-based analyses that quantify the spatio-temporal distribution patternsof tracer binding have the capacity to provide additional and complementary infor-mation to that achieved by standard univariate analyses alone; multivariate patternanalysis evaluates the covariance of tracer binding between multiple brain regions,and thus can provide insight into the relationships between multiple regions in addi-tion to mean differences between groups [61]. Pattern analysis also affords strongerstatistical power by reducing the need for stringent and sometimes overly conserva-tive multiple comparison corrections.Several multivariate pattern analysis methods have been used to explore func-tional brain networks using PET data; mostly focusing on disease-related metabolicpatterns assessed with [18F ]-fluorodeoxyglucose (FDG) PET. However, the appli-cation of multivariate pattern analysis methods for other PET tracers has beenlimited.Traditional analysis in the PET field has been focused heavily on the absolutetracer binding in a single brain region, which gives an univariate measure of thedisease-related effect on a specific target. For example, if we want to examine thePD-induced effect on the serotonergic system, we can compare the DASB tracerbinding values in the striatum in healthy controls and PD subjects; lower DASBtracer binding in the striatum indicates that there is a reduced number of seroton-ergic nerve terminals in the PD subjects. With multivariate pattern analysis, onthe other hand, we analyze multiple brain regions at the same time and obtain apattern reflecting the connectivity (covarying behaviour) of tracer binding in all theregions. Using the same DASB data as an example, with a multivariate approach,the end result might be a spatial pattern that is expressed more in the PD subjectscompared to healthy controls, where there is relatively reduced tracer binding in thestriatum but relatively increased tracer binding in another brain region (e.g. thehypothalamus). The idea of covarying relationships between multiple brain regions(pattern) is harder to conceptualize and interpret and has not been well establishedin the PET community. In multivariate analysis, we cannot simply interpret the25changes in a single brain region alone. Another challenge for the application ofmany multivariate analysis methods is the lack of temporal dimension in PET data.Unlike functional magnetic resonance imaging (fMRI) data (data dimension can besubject by ROI by time), the input binding potential data is often in the dimensionof subject by ROI/voxels and the temporal information of tracer kinetics is embed-ded in the binding potential values. The lack of a third dimension in the input datarestricts the connectivity matrix to the group-level instead of subject-level (moredetails in Chapter3).The application and development of multivariate pattern analysis methods haveincreased demand in the PET field. First, multivariate approach significantly im-proves the sensitivity for detecting small changes in tracer binding values comparedto traditional univariate approach. Second, it highlights the network rather thanlocalized changes, which may be more suitable to study deterministic changes inthe brain due to disease. Third, with the emergence of multi-tracer PET studiesand multi-modal imaging studies, multivariate approach provides a more effectiveway to combine and extract information from all the imaging datasets. In the fol-lowing sections, the proposed network analyses used in this thesis project will besummarized.EBJBE Linyur aodyls of gputiul duttyrnsN drinwipul ComponyntUnulysisbotivvtionMost PET studies have the large p and small n problem (i.e., when there are highernumber of features/ROIs than the number of observations/subjects) which can makediscrimination analysis between different subject groups and correlation analysiswith clinical outcomes challenging. For example, for DASB tracer, we have ap-proximately 40 different pre-defined ROIs to capture the entire projections of theserotonergic system and fewer than 20 subjects per group. In the traditional anal-ysis, we often apply T-test or correlation analysis to the tracer binding values ineach ROI to compare the differences or examine localized changes in a single ROI.However, this approach may lack the sensitivity to detect subtle disease-relatedchanges. Using dimension-reduction techniques like PCA, we can decompose high-dimensional data into a spatial pattern reflecting how a specific neurotransmittersystem changes in multiple ROI at once, which can capture the more accurate nature26of disease-induced changes in the brain in a deterministic way.Figure 1.9: PCA on example data points in two-dimensional coordinate.erinxipvl Componznt Vnvl–sis =eCV)PCA is a common way to reduce data dimensions by decomposing the data into a setof eigenvalues and eigenvectors and transforms the data into orthogonal/uncorrelatedvariables called the principal component (PC). Eigenvectors represent the directionsof the PCs and the eigenvalues represent the variance of the data points in that direc-tion. The first PC accounts for the most variance and each succeeding PC accountfor the highest variance possible while still orthogonal to the preceding PC. Usingthe two-dimensional square data example (Fig.1.9), PC1 and PC2 transform thedata into a new coordinate system and each data point is obtained as a linear com-bination of uncorrelated orthogonal basis set. When we have m-dimensional dataset, PC accounting for relatively large variance reveal the underlying structure ofthe data and PC accounting for relatively low variance are assumed to be noise [62].Mathematically, when we have a subject (N) by region (R) input data matrixY where each matrix element contains the imaging measure, we compute the PCfrom singular value decomposition (SVD) after centering (i.e., demeaning) the data.SVD decomposes Y into 3 matrices:n = USk T (1.18)where the NxN matrix U is the eigenvector matrix in subject space (referred toas subject scores), the NxN matrix S is a diagonal matrix containing singular values(square root of the eigenvalues), and V is a RxN matrix. The PCs in the region27space are then given by:e = Sk T (1.19)where the NxR matrix P is referred to as PC loadings (regional weights).Using a PCA-based scaled subprofile model (SSM), a specific spatial pattern ofFDG PET was found to accurately discriminate PD subjects from healthy controls.This Parkinson’s disease-related pattern (PDRP) was characterized by increasedpallidothalamic and pontine activity associated with relatively reduced activity inprefrontal and parietal cortex. It was also found to correlate consistently with unifiedParkinson’s disease rating Scale (UPDRS) motor scores [63], clinical response totherapy [64], and bradykinesia and executive dysfunction [65]. SSM has also beenapplied to identify the Parkinson’s disease-related cognitive pattern (PDCP) usingFDG PET, characterized by relatively increased activity in the cerebellar vermis anddentate nuclei with associated reduced activity in frontal and parietal associationareas [66].A detailed review of the mathematical principles and basic assumptions for SSMwas previously published [63]. Regional PET data were first centered by subtract-ing subject and region means to obtain the residual profiles; this ensures that theanalysis is minimally sensitive to global scaling effects. PCA then decomposes theresidual profile into orthogonal spatial covariance patterns along each PC and yieldstwo outputs of interest for further analysis: 1) Regional weights, which are loadingsfor each ROI that contributes to the spatial covariance pattern along each PC and 2)PC scores, here referred to as subject scores, which quantify the expressions of thecovariance patterns for each subject. Details on the implemented analysis pipelinecan be found in Chapter2.EBJBF bytwork hopologybotivvtionIn addition to identifying disease or mutation related pattern changes in severalneurotransmitter systems (using PCA as discussed in previous section), we canalso examine more detailed functional properties and organizations of networks andhow they change in the course of network growth and rewiring. Examining thedisease-induced or mutation-related changes in network organization can providenew insights into the neuronal architecture of the entire brain network and functional28behaviours of individual brain regions [60].Grvph ihzor– Vnvl–sisGraph is a mathematical structure used to model pairwise relations between objects.Graph theory analysis has been widely used in many disciplines to model processesin physics, biology and social science. It provides interesting insights into the localconnections between specific nodes or overall organization of the entire network. Inneuroscience, graphs represent functional connections between spatially distributedbrain regions, where nodes are different brain areas and edges are the connectionsbetween these areas derived from sparse pairwise correlation matrix [9].Figure 1.10: Illustration of some measures of network topology [8]Upon definitions of such graphs, several key graph properties can be calculatedto estimate network topology [8]. These measures help to answer questions suchas how efficiently information is integrated between sub-networks and how brainorganization is affected by various diseases at both local and global level.In Fig.1.10, nodes (brain regions) are denoted as circular dots and edges con-necting pair of nodes are the correlation between brain regions. Integration of anetwork is measured based on the shortest path lengths (green). Measures of net-work segregation are based on triangle counts (blue), which are also organized intodifferent modules (oval). Centrality (how important a node is to facilitating networkintegration) can be measured with the node degree (red) or the number of pathsconnecting a node. Hub nodes (black) participate in many routes with shortest29paths; therefore, they have a high centrality [8].In addition, many networks in nature have been characterized to have a small-world property, with high level of local clustering (C) and short characteristic pathlength (L) (Fig.1.11). Small-world networks have proven to be robust with highlocal and global efficiency (high level of segregation and high level of global integra-tion) in Internet, social networks, and more recently functional and structural brainnetworks. This small-world network is found to be disrupted by neurodegenerativediseases, indicating less efficient brain functions [67].Figure 1.11: a) random network; b) small-world network with most edges connectingneighboring nodes and a few long edges to create short-cuts for distant sub-networks[9]; c) scale-free network with most nodes have few connections to other nodes, whileseveral nodes (black connections) are connected to many other nodes.Graph theory analysis has been widely used to study functional network archi-tectures in the brain using fMRI and FDG-PET data in various diseases [68] [67][69].A recent FDG-PET study showed that PD is associated with dense metabolic ac-tive cores around the putamen, globus pallidus and thalamus and an exaggerationof small-world structures that is only partially corrected by dopaminergic treatment[69]. These findings suggest that disease-induced alterations in the metabolic net-work may be related to faulty information transmission associated with PD. Weapplied graph theory analysis to study disease-related changes in network topologyin Chapter3. Detailed mathematical descriptions of the graph theory metrics canbe found in Chapter3.30EBJBG gputioAtymporul duttyrnsbotivvtionAs described previously, disease-induced progressive changes often follow a char-acteristic spatio-temporal pattern and capturing these changes can be challengingwith traditional analysis methods, especially in prodromal and early disease stages.Modeling of the disease-related progressive changes can help us to better track dis-ease progression and understand the underlying mechanisms. Traditional analysismethods to capture these progressive changes include fitting a pre-defined modelwith spatial and temporal parameters or apply analysis in the spatial and temporaldomains separately. The previous two methods (PCA and graph theory analysis)are both temporally static methods that do not model the temporal progression ofthe spatial patterns. One needs to assume that these spatial patterns or networktopology remain static/constant over time and only the expression of these spa-tial patterns increases or decreases as disease progresses; this assumption may notnecessarily be valid or in keeping with the fundamental nature of the disease. Inthis work, we introduced and validated the use of DMD, a data-driven multivariateapproach, to extract coupled spatio-temporal patterns simultaneously.Y–nvmix boyz YzxompositionDMD is a relatively new decomposition method first developed in the field of fluiddynamics [70] and recently used to model temporal oscillations of electroencephalog-raphy (EEG) [71] and fMRI data [72] in the brain. DMD finds coherent spatio-temporal patterns in high-dimensional non-linear systems. There are several uniqueadvantages of DMD compared to other model fitting and multivariate approaches(details about DMD method is in Chapter4): 1) it is a data-driven method thatdoes not require a fixed set of governing equations or prior assumptions of the un-derlying dynamics; 2) it combines the advantages from two frequently used analysismethods: PCA for the reduction of high-dimensional data and spectral time-seriesanalysis for identifying the oscillation frequency of time-varying signals; 3) it canmodel non-linear systems effectively, unlike PCA which assumes that the relation-ships between variables is linear; 4) it can isolate/decompose the overall temporalcourse into specific dynamics [73, 74]. In Chapter4, we applied DMD to paramtericPET images to model the temporal changes of spatial patterns of tracer bindingvalues as disease progresses. Detailed mathematical descriptions of DMD can be31found in Chapter4.EBJB4 aultiAtruwyr Joint duttyrn UnulysisbotivvtionThe previous analysis methods focus on one dataset at a time, however there isgrowing recognition for joint analysis of neuroimaging data that allows us to extractinformation from complementary imaging modalities, especially when a single brainmapping cannot provide the complete picture of brain functions. In the study ofPD for example, several neurotransmitters are affected before and after symptomonset. Studying the relationships between multiple neurotransmitter systems maytherefore provide a more complete picture of the disease-related changes in the brain.Figure 1.12: Conceptual diagram of MCCA [10]Joint evttzrn Vnvl–sisA dimension reduction method, such as PCA, decomposes each data set into uncor-related components ordered by variance; components accounting for low varianceare assumed to reflect noise and are consequently removed. multi-set canonicalcorrelation analysis (MCCA) takes one step further to find maximally correlatedcomponents across multiple data sets which still carry large variation within eachdata set. MCCA utilizes second-order statistics (i.e. correlation and covariance) andassumes the data follow multivariate Gaussian distributions. For medical imaging32data, the population distribution approaches Gaussian as the number of subjectsincreases based on Central Limit Theorem [10].As illustrated by Fig.1.12, given two data sets X1 and X2, we decompose theminto two sets of associated components C1 and C2 and their corresponding mixingprofiles (inter-subject variations), A1 and A2. Since each canonical variant in A isuncorrelated within each data set, each component in C can be associated with onlyone component across modalities [75]. The generative model is then given by:mk = VkXk (1.20)where k is the number of data sets, mk ⊆ gaxik P Vk ⊆ gaxWP Xk ⊆ Yaxik , kkis the number of variables in mk, N is the number of observations in mk and D ismin(rank(mk)) (the minimal rank of the matrix mk).The MCCA algorithm transforms variates to maximize correlation across datasets while minimizing correlation within each data set by solving the following max-imization problem:maxxorr(emT1 P fmT2 ) =egk1Nk2fT√(egk1eT )(fgk2fT )(1.21)where e ⊆ gWxi1 and f ⊆ gWxi2 , and the canonical variates are given by:VT1 = emT1 andVT2 = fmT2 (1.22)P and Q are the eigenvectors of the matrices:(g−1k2gk2Nk1g−1k1gk1Nk2 − rI)e = 0 (1.23)and(g−1k1gk1Nk2g−1k2gk2Nk1 − rI)f = 0 (1.24)where r is the eigenvalue vector, gk1Nk2 is the cross-correlation of X1 and X2, andgk1 gk2 are the autocorrelation of X1 and X2. The associated components can bethen calculated using least-squares approximations:Xˆk = (VTkVk)−1VTkmk (1.25)MCCA has been successfully applied to fMRI data collected from subjects per-33forming a visuomotor task to obtain meaningful subject group level consensus [10]; ithas also been used for multimodal data fusion with fMRI, structural MRI and EEGdata collected from subjects with schizophrenia and healthy controls performing anauditory oddball task; the correlated profiles detected more specific associations be-tween brain structures and functions linked to the task. These specific associationsbetween brain structures and functions were found to be altered in subjects withschizophrenia [76].After extracting the common information from multiple data sets with MCCA,the residuals contain unique information provided by each individual data set andthe true noise residuals.m = mvommon +mrxsisutlm = mvommon +muniqux +mnoisx(1.26)orthogonal signal correction (OSC) can be used to separate MCCA residualsinto unique information and noise. OSC removes one or more directions in mrxsiwutlthat is orthogonal to mvommon to give muniqux:mvommon = mrW + Zmuniqux = mr − n (n Tn )−1n Tmr(1.27)where mr is mrxsiwutl, B is the regression weights and E is mnoisx.In Chapter5 and Chapter6, we applied the joint pattern analysis pipeline toextract common and unique information from multi-tracer PET data. Details onthe analysis pipeline can be found in Chapter5.EBK fysyurwh cvjywtivysThe first objective of this thesis work is to introduce and validate the use of differentnetwork pattern analysis methods for analyzing multi-tracer PET data. The secondobjective is to apply the network pattern analyses to investigate PD-induced net-work changes in various neurotransmitter systems in a deterministic fashion. Themain technical novelty of this thesis work is the development and validation of var-ious network pattern analysis methods, which can introduce more sensitive ways inaddition to the traditional univariate analyses for the analysis of PET data. We alsointroduce different network pattern analysis methods for more specific applications34of interest, including spatial models for extracting deterministic spatial patterns forgroup discrimination, spatio-temporal model for extracting spatio-temporal patternsrelated to disease progression and joint pattern analysis approach for combining in-formation from multi-tracer data. From a clinical point of view, these novel meth-ods, reflecting a conceptual shift on our understanding of how disease affects thebrain functions, can lead to new understanding of PD-induced changes in differentneurotransmitter systems.Following the introduction to necessary background materials for the work inthis thesis, we applied spatial models to study changes in spatial network connectiv-ity related to PD and LRRK2 mutation in a deterministic fashion in Chapter2 andChapter3. Then, in the work described in Chapter4, we introduced and validateda novel spatio-temporal model to extract spatio-temporal patterns of neurotrans-mitter changes that are more sensitive for tracking disease progression compared totraditional univariate analyses. We then introduced a novel joint pattern analysisapproach for extracting complementary information from multi-tracer PET datain Chapter5. The work done in Chapter6 then extends the method introduced inChapter5 to investigate the relationship between multi-tracer PET data and clinicaloutcomes.35Chuptyr FLinyur aodyls of gputiulduttyrnsFBE gummuryThe work described in this chapter contains published work: JC [C [uA IC Kl–uzhinAhC aiuA ZC hhvhinfvryA cC kvfviA JC bxKznzizA cC czilsonA gC bvwroukAbC VC hvxhzliA YC lilzA bC JC bxKzofinA VC JC htozsslA vny kC hossiA InBvzstigvtion of szrotonzrgix evrkinsons yiszvszBrzlvtzy xovvrivnxz pvttzrnusing pFFCrBYVhWDeZiA czuroImvgz ClinCA volC FNA ppC KJGKKEA JvnCGEFM [11]. I was responsible for development of the analysis methodology, imagepreprocessing, statistical analysis, clinical interpretation and manuscript composi-tion. Scanning procedures were performed by the staff members of the UBC PETimaging group. I. Klyuzhin, R. Mabrouk and M.J. McKeown provided feedbackson the development of the analysis methodology. S. Liu, M.A. Sacheli, A.J. Stoessland V. Sossi contributed to clinical interpretation. E. Shahinfard and N. Vafaicontributed to image pre-processing. J. McKenzie and N. Neilson contributed torecruitment of study participants. A.J. Stoessl and V. Sossi designed the study. V.Sossi was the the supervisory author involved throughout the project in the conceptformation and manuscript preparation.In the work described in this chapter, we first applied PCA to investigate ifthere exists a disease and/or LRRK2 mutation related spatial covariance pattern inthe serotonergic system using DASB PET data in a deterministic fashion. We thencompared the results obtained with PCA to other dimension reduction methods.We found that there existed a PD-specific spatial pattern in the serotonergic sys-tem and the expression of this disease pattern correlated significantly with diseaseduration and the degree of dopaminergic denervation as measured by DTBZ PET.This was the first time a significant correlation was found between changes in theserotonergic systems and disease duration and dopaminergic denervation. We also36found a LRRK2 mutation-related spatial pattern in the serotonergic system thatwas different from the disease pattern, which may reflect compensatory mechanismbefore disease onset or a characteristic of the particular mutation.From a methodological point of view, we showed that network pattern analysiscan provide complementary and/or additional information compared to traditionalunivariate analysis. It provides a more sensitive measure, reflecting covarying rela-tionships in multiple brain regions, and is thus able to detect subtle and more dis-tributed changes that may not be captured by traditional analysis. From a clinicalpoint of view, we showed that disease affects the serotonergic system in a particularpattern rather than affecting any region in isolation. The effect of such patternalteration still need to be fully elucidated.FBF IntroduwtionPD is the second most frequent progressive neurodegenerative disorder [33]. Thecardinal features are motor deficits, including tremor, rigidity and bradykinesia,traditionally associated with dopaminergic denervation in the substantia nigra[6].There is however increased recognition of the importance of the disease-inducednon-motor symptoms. Some of these non-motor abnormalities, including autonomicdysfunction, sleep disturbance, cognitive impairment and depression, can precedethe motor deficits by years or even decades[35]. The bases for these non-motordeficits are still unclear, but it is suggested that many of them may be associatedwith alterations of non-dopaminergic neurotransmitter systems[77]. In particular,several in-vivo and post-mortem studies support the hypothesis that progressive al-terations of the serotonergic system may contribute to a number of such non-motoralterations [36]. A large body of evidence has indeed been gathered from in-vivoPET studies using the radioligand [11C]-DASB which binds to the SERT. ReducedDASB binding in the caudate, thalamus, hypothalamus and anterior cingulate cor-tex has been observed in early disease (Q5 years of disease duration), and additionalreductions in the putamen, insula, posterior cingulate cortex and prefrontal cortexhave been detected in mild disease (5-10 years of disease duration). Ventral stria-tum, raphe nuclei and amygdala were only affected in advanced disease (S10 yearsof disease duration)[78]. Reduction in DASB binding in the caudate, putamen andraphe nuclei also significantly correlated with tremor severity on posture and action[79]. Studies have suggested PD subjects experiencing non-motor deficits showed up-37regulated DASB binding in some brain regions compared to subjects without suchdeficit. In particular, PD subjects with abnormal body mass index (BMI) alter-ations had significantly increased DASB binding in the raphe nuclei, hypothalamus,caudate and ventral striatum[80]. PD subjects with depressive symptoms had sig-nificantly increased DASB binding in the amygdala, hypothalamus, caudal raphenuclei, and posterior cingulate cortex compared to matched-PD patients withoutdepressive symptomatology but not compared to healthy controls[36].In addition to its relevance to manifest PD, alterations of the serotonergic sys-tem are deemed to play a role in the presymptomatic or non-motor stage of thedisease. According to Braak staging of PD pathology, Lewy body and neurite de-position occur within the raphe nuclei at stage two, whereas midbrain, substantianigra, amygdala and hypothalamus, more closely linked to the dopaminergic system,are affected at stage three before the onset of motor deficits[81, 82]. The serotoner-gic system is thus hypothesized to be affected by disease prior to the dopaminergicsystem and play a significant role in premotor deficits. Despite the likelihood ofnon-motor deficits preceding motor impairment, the investigation of prodromal ab-normalities is inherently difficult; non-motor deficits are not highly specific for PDand the disease diagnosis is currently based solely on the characteristic motor fea-tures. However, investigating subjects at a higher risk of PD, such as carriers ofspecific genetic mutations, might provide some insights into the prodromal stageof the disease. In this context, it is important to investigate mutation carriers inboth, the asymptomatic and the symptomatic stages, to enable inference about therelevance of the alterations observed in the mutation carriers to those observed insporadic Parkinson’s disease (sPD) and the progression of such alterations. Whileseveral mutations are known to increase the risk of PD, in this work we concentrateon the mutation in the LRRK2.Mutations in the LRRK2 gene are the most common genetic risk factors for au-tosomal dominantly-inherited PD, and the most common mutation of the LRRK2gene is G2019S. The prevalence of LRRK2 G2019S mutation is approximately fourpercent in familial cases and one percent in sporadic cases[37, 38]. The risk forLRRK2-NMC to develop overt disease is 28% at age 59 years, 51% at 69 years and74% at 79 years [37]; however, the estimated disease penetrance varies widely de-pending on age, ethnic group and genetic and environmental modifiers[40, 83]. Itis now commonly accepted that sPD and Parkinson’s patients with LRRK2 mu-tation (LRRK2-PD) show similar patterns of presynaptic dopaminergic degenera-38tion [41–43, 49]. Further, LRRK2-NMC have been found to have lower dopaminetransporter density [49] and increased dopamine turnover [45]. From the clinicalpoint of view sPD and LRRK2-PD are for the most part similar, albeit some dif-ferences are observed: LRRK2-PD is typically associated with a lower non-motorsymptom burden prior to the onset of overt disease and slower progression of mo-tor deficits once disease becomes manifest [37, 43, 84, 85]. Likewise, LRRK2-PDis generally associated with a lower frequency of cognitive impairment[85]. Takentogether, these findings may point towards the existence of some mutation-relatedcompensatory mechanisms that either protect from disease occurrence or positivelyinfluence disease progression. In this context we have recently found that LRRK2-NMC subjects present increased SERT density compared to healthy controls in thehypothalamus[44] and significantly increased cholinergic activity in the cortical re-gions [46]. In light of the demonstrated relevance of the serotonergic system to PDevolution and progression, further analysis of the imaging data is warranted. Inparticular, it may be informative to explore if alteration of the serotonergic systemfollows distinct spatial patterns or networks, relating to its widespread projectionsfrom the raphe nuclei. In addition to providing novel information, pattern analysismay be more sensitive to disease-related effects for SERT imaging data comparedto localized changes which are generally explored with univariate analyses.Network-based analyses have the capacity to provide additional and comple-mentary information to that achieved by standard univariate analyses alone; multi-variate pattern analysis evaluates the covariance of tracer binding between multiplebrain regions, and thus can provide insight into the interactions between multipleregions with a common projection source in addition to mean differences betweengroups[61]. Pattern analysis also affords stronger statistical power by reducing theneed for stringent and sometimes overly conservative multiple comparison correc-tions. Several multivariate pattern analysis methods have been used to explore func-tional networks using PET imaging, mostly focusing on metabolic patterns assessedwith [18F]-FDG[63]. Disease-specific spatial covariance patterns were investigatedusing PCA. Using a PCA-based SSM, a PDRP derived from FDG was found toaccurately discriminate PD subjects from healthy controls. This PDRP was charac-terized by increased pallidothalamic and pontine activity associated with relativelyreduced activity in prefrontal and parietal cortex. It was also found to correlateconsistently with UPDRS motor scores [63], clinical response to therapy[64], andbradykinesia and executive dysfunction[65]. SSM has also been applied to identify39the PDCP, also with FDG PET, characterized by relatively increased activity in thecerebellar vermis and dentate nuclei with associated reduced activity in frontal andparietal association areas[66]. Even though the SSM method is a well-documentedmethod, it has not been previously applied to neurotransmitter systems in PET.In this work we use the SSM method to investigate if (i) there is a PD specificalteration of the functional SERT network across selected brain regions, where rel-evance to PD is additionally tested by relating pattern expression to clinical scoresand dopaminergic denervation as measured by striatal [11C]-DTBZ (a marker forthe VMAT2) binding; (ii) LRRK2-PD is characterized by the same pattern; (iii)LRRK2-NMC subjects express a similar or different pattern. SSM with cross vali-dation was first applied to [11C]-DASB PET data of sPD subjects; upon definitionof the serotonergic Parkinsons disease-related pattern (SPDRP), the strength of itsexpression was evaluated in the LRRK2-PD and LRRK2-NMC groups. While inLRRK2-PD, expression of this pattern would indicate similar overall disease-relatedalterations, the presence of such a pattern in the LRRK2-NMC could be taken toreflect similar dysfunction of the serotonergic system during the prodromal phaseof disease, or an increased risk of or propensity to disease. The presence of a dif-ferent LRRK2-NMC spatial covariance pattern would suggest either compensatoryor genotype specific alterations. The analysis included imaging data used for theunivariate analysis describe in previous publication [44]. Finally, in order to checkfor consistency between this more extended analysis and our first results presentedin Wile et al., we applied complementary univariate analyses to regions which werefound to be significantly involved in the SPDRP and LRRK2-NMC patterns. Thesame DTBZ binding data as reported in Wile et al. [44] were used.FBG autyriuls und aythodsFBGBE gtudy durtiwipuntsThe study included 15 sPD, eight LRRK2-PD, nine LRRK2-NMC, and nine healthycontrols age-matched to both sPD and LRRK2-PD groups (Table2.1); two addi-tional sPD and one additional LRRK2-PD subjects compared to Wile et al.. Exclu-sion criteria included clinical history of depression, active anti-depressant therapyor medication with known serotonergic action and a BMI greater than 35. Thehealthy controls had no history of neurological or psychiatric disorders and were not40taking any medication. Disease duration was estimated as time from onset of motorsymptoms as reported by the subjects. The PD subjects were clinically evaluatedwith the movement disorder society unified Parkinsons disease rating scale part III(MDS-UPDRS) and Hoehn and Yahr scale (H&Y) to assess motor dysfunctions,Montreal cognitive assessment (MoCA) scores to assess cognitive performance andBeck depression inventory (BDI). All assessments were performed off medication.The study was approved by the Clinical Research Ethics Board of the University ofBritish Columbia and all subjects provided written informed consents.41]zvlth– xontrol seY =nRFJ) aggKGBeY =nRM) aggKGBcbC =nRN) Glowvl pBvvluz* pBvvluz for seY vs aggKGBeY pBvvluz for seY vs aggKGBcbC pBvvluz for aggKGBcbC vs ]CGznyzr rvtio =mvlzOfzmvlz) 6:3 9:6 3:5 4:5 N/A 0.4 0.675 0.637Vgz =–zvrs) 56±14 59±8 66±15 50±11 0.176 0.785 0.39 0.591Yiszvsz yurvtion =–zvrs) 3.4±2.7 7.4±5.0 N/A 0.058 N/A N/AbYhBjeYgh pvrt III sxorz 16±7 22±10 N/A 0.248 N/A N/A]ozhn vny nvhr stvgz 1.6±0.5 2±1.1 N/A 0.145 N/A N/Abontrzvl Cognitivz Vsszssmznt sxorz =boCV) 28.0±1.5 26.3±2.2 27.7±2.3 0.2 0.313 0.743 N/AWzxk Yzprzssion Invzntor– =WYI) 4.4±3.5 9±6.8 4.7±2.7 0.117 0.104 0.975 N/Acumwzr of pvrtixipvnts fiith spzxix mutvtion 6 G2019S/ 2 R1441C 9 G2019S N/A 0.206 N/A N/ATable 2.1: Characteristics of study participants. Data are meanstandard deviation, unless otherwise indicated.sPD=sporadic Parkinsons disease subjects; LRRK2-PD = manifest LRRK2 mutation carriers; LRRK2-NMC = non-manifesting LRRK2 mutation carriers; MDS-UPDRS = Movement Disorder Society Unified Parkinsons Disease RatingScale; N/A = not applicable. Disease duration was estimated as time from onset of motor symptoms as reported by thepatients. *p values for age, MoCA and BDI were calculated by ANOVA followed by post-hoc analyses. p values for genterratio were calculated by Fishers exact test. p values for disease duration and MDS-UPDRS III scores were calculated byindependent two-tailed T-test. p values for Hoehn and Yahr stage were calculated by Mann-Whitney test. All p valueswere false discovery rate-corrected for multiple comparisons. Data from 8 of 9 healthy controls, 15 of 15 sPD, 5 of 8LRRK2-PD, and 4 of 9 LRRK2-NMC (missing information was due to language barriers). †Data from 15 of 15 sPD, 5 of8 LRRK2-PD, and 6 of 9 LRRK2-NMC.42FBGBF gtutistiwul Unulysis und Cliniwul DutuWe compared age, MoCA and BDI scores between subject groups using analysis ofvariance (ANOVA) followed by post-hoc analyses using Tukey’s Honestly SignificantDifferences Procedure. We assessed gender as a binomial variable with Fishers exacttest. We ascertained differences in disease duration and MDS-UPDRS between sPDand LRRK2-PD subjects by the independent two-tailed T-test. H&Y scores werecompared between sPD and LRRK2-PD subjects using Mann-Whitney test. AllP-values were false discovery rate-corrected for multiple comparisons.FBGBG gwunning drotowolAll study participants underwent a [11C]-DASB PET scan and a T1-weighted brainMRI scan. Prior to PET imaging, patients withdrew from all anti-parkinsonianmedications for at least 12 hours. 558±2MBq of [11C]-DASB were administeredby intravenous injection over 60 seconds using an infusion pump (Harvard Instru-ments). Participants were positioned using external lasers aligning the gantry withthe inferior orbital-external meatal line, and custom fitted thermoplastic masks wereapplied to minimize head movement. All patients were scanned on a Siemens HRRT(Knoxville, TN) with a spatial resolution of 2O5mm3 [86]. Acquired data werebinned into 18 time frames (frame duration: 4×60 s, 3×120 s, 8×300 s, 3×600 s; im-age dimension=256×256×207; voxel size=1O22mm3) for a total duration of 80min.Transmission scans required for attenuation correction were performed over ten min-utes with a rotating 137Cs source. PET images were reconstructed using the 3Dlist-mode OP-OSEM algorithm[87] with 16 subsets and six iterations, with cor-rections for decay, dead-time, normalization, attenuation, scattered and randomcoincidences. After reconstruction, the images were smoothed with a 3mm FWHMGaussian filter to reduce noise. The frames were spatially realigned for each subjectwith rigid-body transformation to minimize the impact of motion artifacts duringscans. The structural MRI scans were performed on a Philips Achieva 3T MRIscanner (Phillips Healthcare, Best, NL) using the T1 turbo field echo (TFE) se-quence (TR/TE = 7.7/3O6ms; TFE shots=218; flip angle = 8◦; image dimension:256×256×170; voxel size 1mm3). sPD subjects and LRRK2 mutation carriers alsounderwent an additional (+)DTBZ PET scan to obtain complementary informationon the dopaminergic integrity. Scanning protocol, imaging processing steps andresults of the analysis of DTBZ data were previously reported [44].43FBGB4 Imugy drowyssing und UnulysisImage preprocessing was done using Statistical Parametric Mapping (SPM12) soft-ware (Wellcome Department of Cognitive Neurology, University College London,UK) running on Matlab 9.0 (Mathworks Inc., Natick, MA) and MEDx (Sensor Sys-tems, Sterling, VA). The MRI images were resampled with trilinear interpolationto match the voxel size of the PET images. The resized MRI images were thencoregistered to the corresponding mean PET images (mean of 18 frames) using nor-malized mutual information to estimate the affine transformation for each subject.These resized MRI images were then warped into the Montreal Neurological In-stitute (MNI) space and their inverse deformation fields were saved. The inversedeformation field vectors were applied to ROI templates predefined in the MNI spaceto bring them into the individual subject’s PET space in a single step. The qualityof each processing step was visually checked for all scans.The ROI templates were defined by experienced neurologists in the MNI spaceusing MRI and PET data from healthy subjects for a total of 43 non-overlappingROIs which are known to be involved in the serotonergic system. The regions in-cluded medulla, midbrain, pons, rostral raphe nucleus, ventral tegmental area and19 ROIs placed bilaterally (resulting in 38 ROIs over both hemispheres): anteriorand posterior cingulate, amygdala, caudate, cerebellum, dorsolateral prefrontal cor-tex, hypothalamus, insula, orbital frontal cortex, pedunculopontine nucleus (PPN),putamen (anterior, middle and posterior), substantia nigra, thalamus, ventral stria-tum, hippocampus, dentate nucleus, and globus pallidus. Decay-corrected TACswere calculated for each ROI using the Marsbar toolbox in SPM. BPaW [29] valuesin 42 ROIs were obtained using the SRTM with the cerebellum as reference region[31]. ROI-based BPaW values derived from SRTM were also compared with meanparametric BPaW values in each ROI derived using two-step simplified referencetissue model (SRTM2) [88]. As the BPaW values derived with the two methodswere highly comparable, we only used the SRTM method for later analyses.FBGBI gga aultivuriuty duttyrn UnulysisWe applied SSM with repeated five-fold cross validation to DASB BPaW data in42 ROIs to identify disease and mutation-related spatial covariance patterns usingin-house Matlab scripts. A detailed review of the mathematical principles and basicassumptions for SSM was previously published [63]. Regional DASB BPaW data44were first centered by subtracting subject and region means to obtain the residualprofiles; this ensures that the analysis is minimally sensitive to global scaling ef-fects. PCA then decomposes the residual profile into orthogonal spatial covariancepatterns along each PC and yields two outputs of interest for further analysis: 1)Regional weights, which are loadings for each ROI that contributes to the spatialcovariance pattern along each PC and 2) PC scores, here referred to as subjectscores, which quantify the expressions of the covariance patterns for each subject.To identify robust disease or mutation-related covariance patterns, the followingsteps were implemented:1. Subject scores along each PC were entered as independent variables into aforward stepwise logistic regression model, using group assignments as depen-dent variables. To limit the number of variables, only PCs accounting for atleast five percent of the total subject by region variance were entered into theregression model. The resulting covariance pattern was the linear combinationof regional weights which gave the lowest Akaike information criterion (AIC)score [89].2. 1000 iterations of five-fold cross validation were performed. In each iteration,subjects were randomly divided into five subsets. Four subsets (80% of thedata) were used to identify the covariance pattern and examine the effect ofsubject variation.3. The final disease or mutation-specific covariance pattern was the average ofthe covariance patterns obtained from all iterations. Regional weights forthis averaged covariance pattern were Z-transformed to determine significantregional contributions with weights greater than one (significant weights).4. Subject scores in all subject groups were then computed by projecting theresidual profiles onto the average covariance pattern using the topographicprofile rating (TPR) method [63]. TPR is used to compute the subject scores(subject by one) on the prospective covariance pattern by calculating the innerproduct of the residual profiles (subject by ROI) and the covariance pattern(ROI by one). Projected subject scores were also Z-transformed and analysis ofcovariance (ANCOVA) was performed to test the group discrimination powerof the average covariance pattern using age as a covariate (see results).455. Steps one to four were performed separately for sPD versus healthy controlsand for LRRK2-NMC versus healthy controls.6. Forward stepwise multiple regression analyses were performed between theprojected subject scores in participants with PD and age of symptom onset,disease duration, UPDRS motor score and DTBZ binding to examine thecorrelations of the subject pattern expressions with clinical measurements anddopaminergic denervation.FBGBJ inivuriuty UnulysisAfter applying network analysis, we performed ANOVA on DASB BPaW data be-tween the left and right hemispheres and between the subject groups only for thoseregions that showed significant weights in the SSM multivariate pattern analysis, inorder to limit the number of comparisons and to interpret the results of the SSManalysis in the context of absolute binding values. As no difference between theleft and right hemispheres was identified, values from the corresponding regions inthe two hemispheres were averaged to reduce the number of comparisons. Becausethe LRRK2-NMC subjects were slightly younger than the sPD subjects (2.1), wealso performed ANCOVA with age as a covariate when comparing these two groups.False positive rates were controlled at P = 0.05 using Bonferroni-Holms step-downprocedure[90].FB4 fysultsThere was no significant difference in any clinical variables between subject groups.FB4BE sdD und LffKFAdD gputiul Covuriunwy duttyrnA spatial covariance pattern (the SPDRP) was identified with significantly higherscores for sPD subjects than healthy controls (P Q0.001). The SPDRP was charac-terized by relatively decreased DASB binding in caudate, putamen and substantianigra, and relatively increased DASB binding in hypothalamus and hippocampus(Fig.2.1). This pattern had the largest contributions from PC2, PC3 and PC4, whichcollectively accounted for 28% of the total variance. When projecting data fromthe LRRK2 mutation carriers onto this pattern, LRRK2-PD showed a significantly46higher expression of SPDRP compared to healthy controls (P Q0.05) (Fig.2.2). Al-though only the sPD subject data were used to determine SPDRP, there was nosignificant difference between the sPD and LRRK2-PD subject scores. There wasone outlier in the LRRK2-NMC group with very high SPDRP expression (H814);this subject was the youngest participant in the study (more than two standard devi-ations away from the mean age of this subject group). Subject H814 showed normalexpression of LRRK2-NMC pattern compared to other LRRK2-NMC subjects andstriatal DTBZ binding in the control range. When excluding this subject from thegroup comparison, there was no significant difference between LRRK2-NMC andthe healthy controls, but a significant difference between LRRK2-NMC and sPDgroups (P Q0.05).Figure 2.1: Serotonergic Parkinsons disease-related pattern (SPDRP) identified bycomparing sporadic Parkinsons disease subjects and healthy controls. Regions withsignificant weights on the averaged SPDRP were overlaid onto a T1 MRI image.Blue (red) indicates regions with relatively decreased (increased) binding in sporadicParkinsons disease subjects compared to healthy controls.47Figure 2.2: Subject scores projected onto the serotonergic Parkinsons disease-relatedpattern (SPDPR) for all four subject groups. The three outliers were labeled asH1013 and H1079 in the LRRK2-PD group, and H814 in the LRRK2-NMC group.sPD = sporadic Parkinsons disease subjects; LRRK2-PD = manifested LRRK2mutation carriers; LRRK2-NMC = non-manifesting LRRK2 mutation carriers; * =significant at P Q0.05 level; **=significant at P Q0.01 level.Figure 2.3: Scatter plot of projected subject scores, denoting the strength of theserotonergic Parkinsons disease-related pattern (SPDRP) expression as a functionof disease durations (left) and as a function of DTBZ binding expressed as fractionsto age-matched normal controls (right) for sporadic PD and LRRK2-PD. The twooutliers are labeled in the same way as in Fig.2.2. The best line fit was done withoutthese two subjects. LRRK2-PD = manifest LRRK2 mutation carriers.SPDRP expression in sPD subjects correlated significantly with disease duration48(P Q0.01) (Fig.2.3 left); this correlation remained significant after correcting for theage of disease onset (P Q0.05). Combined LRRK2-PD and sPD subject scores alsocorrelated significantly with disease duration (P Q0.01) after removing two outliers(H1079 and H1013), both LRRK2-PD. Without exclusion of the two outliers, thecorrelation between disease duration and LRRK2-PD and sPD subject scores wasnot significant (P = 0.63). One outlier (H1079), while having a disease duration often years and significant reduced DTBZ binding compared to age-matched controls,only scored 13 on the UPDRS motor scale, indicative of unusually slow disease pro-gression. The second outlier (H1013) was initially found to be a subject withoutevidence of dopaminergic deficit (SWEDD) based on 6-[18F]-fluoro-L-dopa (18F-dopa) performed at the time of diagnosis and later developed unequivocal PD withbilateral striatal dopaminergic denervation and relatively preserved SERT bindingin the putamen and several cortical regions[91]; this subject had much longer diseaseduration compared to all other manifest subjects (17 years of disease duration com-pared to a mean of 4.2 years in all other subjects). Interestingly these two subjectsshowed the lowest SPDRP expression in the LRRK2-PD group (Fig.2.2), while theirexpressions of LRRK2-NMC pattern were in line with that of the other LRRK2-PDsubjects. SPDRP expression did not correlate with any other clinical variables.SPDRP expression in sPD and LRRK2-PD also correlated significantly withDTBZ binding in the more affected putamen (P Q0.01) (Fig.2.3 right). H1079 andH1013 did not appear as outliers in this correlation.FB4BF LffKFAbaC gputiul Covuriunwy duttyrnA pattern along which the LRRK2-NMC subject scores were significantly highercompared to those of healthy controls (P Q0.001) was also identified (Fig.2.4); Thepattern had the largest contributions from PC2, which accounted for 10% of thetotal variance. The LRRK2-NMC pattern was comprised of a relatively decreasedDASB binding in the pons, PPN, thalamus and raphe nucleus, and relatively in-creased binding in the hypothalamus, amygdala, hippocampus and substantia nigra(Fig.2.5). Neither sPD nor the LRRK2-PD showed a significant expression of thisLRRK2-NMC pattern. This pattern expression did not show correlation with anyclinical measurements or DTBZ binding.49Figure 2.4: Subject scores projected onto LRRK2-NMC pattern in all four sub-ject groups. sPD=sporadic Parkinsons disease subjects; LRRK2-PD = manifestedLRRK2 mutation carriers; LRRK2-NMC = non-manifesting LRRK2 mutation car-riers; ** = significant at P Q0.01 level.Figure 2.5: Serotonergic LRRK2-NMC pattern identified by comparing LRRK2-NMC and healthy controls. Regions with significant weights on the averagedLRRK2-NMC pattern were overlaid onto a T1 MRI image. Blue (red) indicatesregions with relatively decreased (increased) binding in LRRK2-NMC compared tohealthy controls. LRRK2-NMC = non-manifesting LRRK2 mutation carriers; PPN= pedunculopontine nucleus.FB4BG inivuriuty UnulysisOnly DASB BPaW values from regions significantly contributing to the spatial pat-tern obtained in the DASB multivariate analysis were tested in the univariate ROIanalysis; they included the caudate, putamen, hypothalamus, hippocampus, amyg-dala, substantia nigra, thalamus, pons, PPN and raphe nucleus. There was nosignificant difference in DASB binding between the left and right hemispheric ROIs50in any subject group, so average BPaW values from both hemispheres were usedto reduce the number of comparisons. There was no significant difference betweenLRRK2-PD and sPD in any brain region. Both sPD and LRRK2-PD showed sig-nificantly reduced DASB binding in the caudate compared to healthy controls (PQ0.01 for both groups), however this reduction was not significant after correctingfor multiple comparisons. LRRK2-NMC showed significantly higher DASB bind-ing compared to healthy controls in hypothalamus (P Q0.01) and hippocampus (PQ0.05). There was a significant age effect in hypothalamus (P Q0.01), pons (PQ0.05), substantia nigra (P Q0.05) and raphe nucleus (P Q0.05) in the healthycontrol group. With age as a covariate, LRRK2-NMC only showed significantlyhigher DASB binding compared to healthy controls in hypothalamus (P Q0.05).However, again, the significance of the results did not survive multiple comparisoncorrection for any region. As previously reported, all LRRK2-NMC subjects showednormal DTBZ binding compared to age-matched controls except for one subject andno correlation was obtained between DASB and DTBZ binding in any individualregion.FBI DiswussionUsing a multivariate pattern analysis, we found a significant SPDRP in the sPD sub-jects, and LRRK2-PD subjects also showed a significant elevation of this pattern.A significant positive correlation between the strength of the SPDRP expressionand disease duration was found for both manifest disease groups, with the excep-tion of two subjects who had the lowest SPDRP expression; one presented a withan unusual (milder) motor symptom progression, while the other outlier was ini-tially classified as a SWEDD and when imaged 17 years later exhibited bilateralstriatal dopaminergic denervation with relatively preserved SERT in the putamenand several cortical regions[91]. A very significant correlation between the strengthof the SPDRP expression and dopaminergic deficit in the more affected putamenwas also observed. In this case the two subjects described above no longer appearedoutliers, which is likely suggestive of a mechanistic link between dopaminergic den-ervation and serotonergic connectivity. Interestingly, no correlation between DTBZand DASB BPaW values of individual regions was found suggesting PD affects theserotonergic system on a more global network level rather than any particular regionin isolation. This observation also highlights the complementarity of the information51provided by a network approach.FBIBE dossivly Funwtionul Vusis for gdDfd und LffKFAbaCduttyrn hopogruphyIn SPDRP, the caudate showed the most significant relative reduction in DASBbinding followed by the putamen. Once disease manifests, the observed relativereductions in the striatum agree with previous findings that PD is associated withan absolute decrease in DASB binding in caudate and putamen compared to healthycontrols[36, 92]. Unlike the dopaminergic system where the posterior putamen ismore affected by the disease, there is a preferential loss of SERT in the caudate.Univariate analysis in this study failed to detect significant group separation in bothcaudate and putamen after correcting for multiple comparisons in PD subjects andLRRK2-NMC compared to healthy controls. This might be due to the fact thatonly early PD subjects were included in this study, whereas subjects involved inother studies had a wider range of disease durations. However, the pattern analysisstill accurately captured this disease effect even in an early disease stage.Relative upregulation of DASB binding in the substantia nigra was only evidentat the non-manifesting stage, as it was observed only in the LRRK2-NMC pattern;DASB binding in the substantia nigra was relatively reduced in subjects with man-ifest disease. This is consistent with the fact that disease causes more alterationsin the substantia nigra compared to other regions such as the hypothalamus andhippocampus, where binding remained relatively upregulated even in the presenceof disease. While dopaminergic neurons originate in the substantia nigra, 5HT neu-rons show highly variable densities and patterns of innervation in the basal ganglia.The substantia nigra receives the densest 5HT innervation, whereas the caudate ismore heterogeneously innervated than the putamen[93]. The relative upregulationof DASB binding in LRRK2-NMC may indicate a compensatory or protective roleof serotonergic innervation in the substantia nigra, which is no longer present oncedisease becomes manifest. Such upregulation could conceivably contribute to theincomplete penetrance of LRRK2 mutations.The relative upregulation of DASB binding in the hypothalamus and hippocam-pus observed in the LRRK2-NMC pattern is present even after disease manifests.Upregulation of DASB binding in the hypothalamus has also been found to correlatewith several aspects in PD such as depression and abnormal weight changes[78, 80].In task-related fMRI studies, the hippocampus has been shown to be preferentially52activated in mild PD subjects when performing a cognitive planning task, at theexpense of a reduction in the activation of the right caudate. This may indicate thathippocampal hyperactivity compensates for striatal deficiencies and thus possiblyinduces a shift towards the declarative memory system in response to a defect inprocedural memory, which is related to the frontostriatal system[94]. While this isa reasonable interpretation for the subjects with active disease, the reason for thisupregulation in the LRRK2-NMC, where the dopaminergic system is still mostlyintact, is still unclear. It has been observed however, that the cholinergic system,also related to cognition, is upregulated in this patient population [46].LRRK2-NMC pattern was also characterized by relatively reduced DASB bind-ing in the raphe nucleus, pons, PPN and thalamus. Because there was an overallincrease in absolute binding magnitudes in LRRK2-NMC compared to healthy con-trols, these regions were actually less upregulated compared to other regions inLRRK2-NMC, rather than reduced compared to healthy controls. The LRRK2-NMC pattern also showed relatively increased binding in the amygdala, which wasnot upregulated in the SPDRP. Abnormal activation of the amygdala in responseto processing of fearful stimuli has been observed in early disease stage even in theabsence of noticeable behavioral differences; the abnormality was partially reversedwith dopamine replacement[95]. Given that dopaminergic function in the amygdalamay be reduced in PD [96], the relative upregulation of the serotonergic systemin the asymptomatic stage may be of compensatory nature. In addition, amyg-dala hyperactivity in the serotonergic system in PD has also been associated withdepression[36]. Relative upregulation in amygdala in LRRK2-NMC pattern maypartially explain the higher propensity to depression in LRRK2-PD compared tosPD [42].FBIBF LffKFAbaC duttyrn drotywtion or CompynsutionLRRK2-NMC subjects did not show a significant expression of SPDRP, except forone subject with a clear expression of the disease pattern (H814 as shown in Fig.2.2),whose dopaminergic function appeared intact. LRRK2-NMC subjects however ex-pressed a unique pattern compared to healthy controls, which may be related tomutation specific alterations, such as the established decrease in dopamine trans-porter binding [41, 44, 45] and increase in dopamine turnover in the putamen [45].Interestingly the two patterns showed several common regions of relative upregula-tion; whether such upregulation may be protective against other disease-inducing53mechanisms or a risk factor for disease is still unclear at this stage. While clearalterations in other regions are observed in the SPDRP, notably a decreased DASBbinding in the substantia nigra, it may be reasonable to surmise that relatively in-creased expression of SERT (and presumably of 5HT nerve terminals) offers somedegree of protection of neurons or compensation to delay symptom manifestationsince the upregulation persists after disease onset. Alternatively, an increase inDASB binding could be related to reduced SERT occupancy secondary to lowerlevels of 5HT neurons in the synapse [97]. Evidence for this hypothesis is howeverquite marginal. The spread in LRRK2-NMC subject scores onto SPDRP may re-flect inherent increased risk of disease (or pre-manifest disease) in subjects with highpattern expression and the incomplete penetrance of LRRK2 mutation; most of ourLRRK2-NMC subjects were younger than 59 years, the expected age of disease on-set for the G2019S mutation-associated disease [37]. Two LRRK2-PD subjects hadR1441C mutations; however, they did not appear as outliers in either SPDRP orLRRK2-NMC pattern. All LRRK2-NMC subjects showed normal striatal dopamin-ergic innervation except for one outlier subject, whose serotonergic patterns strengthwas in line with those observed in the LRRK2-NMC group. This indicates that thechanges we observed in LRRK2-NMC pattern are generally not accompanied bydopaminergic deficit and may indeed be related to protective mechanisms. Longi-tudinal observations which would capture subject phenoconversion would help toexplain the high SPDRP expression in some of the LRRK2-NMC subjects, and theprotection or compensatory role of the serotonergic system in PD.FBIBG Compurison with inivuriuty UnulysisRegions with most significant changes in DASB binding in univariate analysis alsoshowed the most significant contributions to the patterns derived in the multivari-ate analysis (notably the relatively decreased binding in caudate in SPDRP andrelatively increased binding in hypothalamus and hippocampus in LRRK2-NMCpattern). While overall consistent, the results from the univariate analysis in thiswork differ slightly in terms of significance from those reported in our previouspublication [44] due to the involvement of a larger number of ROIs: notably theseparation of the striatum region into caudate and putamen, and the addition ofthe substantia nigra, amygdala and hippocampus. Here the main focus was to iden-tify subtle disease or genotype specific changes at a global network level, aiming tocapture all widespread serotonergic projections instead of restricting the analysis to54specific ROIs.FBIB4 Compurison with cthyr bytwork Unulysis duttyrnsWe compared SPDRP to other network patterns reported in the literature (FDGPCA-based analysis and resting-state fMRI functional connectivity analysis) to ex-plore possible relationships between the serotonergic system, metabolic activity andresting-state functional connectivity. SPDRP did not resemble FDG PDRP, FDGPDCP or fMRI-derived patterns; this was to some degree expected given the dif-ferent topology of the serotonergic network compared to the more extended braincircuitry involved in metabolic and hemodynamic activities.FBIBI LimitutionsThere are several limitations to this study. First, the multivariate analysis used toderive the patterns was based a relatively small sample size; however, we used five-fold cross validation and AIC to ensure the obtained pattern was generalizable tonew datasets. In addition, since we found no difference in either multivariate or uni-variate analyses between sPD and LRRK2-PD subjects and LRRK2-PD group wasnot involved in deriving the pattern, the significantly higher expression of SPDRP inLRRK2-PD subjects further confirms the robustness of the pattern. Second, therewas some overlaps in subject scores between groups for both patterns. This over-lap is consistent with previous findings obtained with univariate analysis, where noclear separation in DASB binding between healthy controls and early PD subjectswas seen [98]. The fact that we did see a significant group separation in early dis-ease confirmed the higher sensitivity of the multivariate approach. Importantly, thetwo subjects with markedly different disease progression exhibited the lowest SP-DRP expression and were classified as outliers in the correlation between SPDRPexpression and disease duration, which would further support the robustness of thismethod and possibly suggest that the serotonergic systems may play a significantrole in determining the course of the disease progression.FBJ dCU jyrsus cthyr gputiul aodylsThere are several limitations with the PCA plus logistic regression approach. Inthis section, I will address some of the major limitations by comparing the results55obtained with the PCA plus logistic regression approach to results obtained withother similar dimension reduction techniques and discuss the advantages and disad-vantages for each of the methods. We applied several dimension reduction methodsto investigate the PD-related spatial patterns in the serotonergic system using theDASB data. In particular, we compared the group discrimination power betweenthe healthy controls and sPD subjects, correlation between pattern expression anddisease duration, the stability of the spatial pattern and the computation complexity.FBJBE Dimynsion fyduwtion aythodseCV ICV eahBYV geCVpBvvluz =group yisxriminvtion pofizr) p=0.001 p=0.0003 p=0.0002 p=0.003kvrivnxz vxxountzy for =kV[) 21% 6.7% 13% 16%Corrzlvtion fiith yiszvsz yurvtion p=0.007 p=0.04 p=0.90 p=0.12Table 2.2: Comparison between PCA and other similar dimension reduction meth-ods in terms of group discrimination power between healthy controls and sPD sub-jects, variance accounted for, and correlation between the subject scores and diseaseduration using the DASB PET data. Overall, PCA seems to be most suitablemethod for this particular study as discussed in the text below.erinxipvl Componznt Vnvl–sisWe applied PCA plus logistic regression to the DASB data in healthy controls andsPD. After decomposing the input data with PCA, the logistic regression includedPC2 and PC4 (accounting for 21% of the total variance) and the p-value for groupdiscrimination between healthy controls and sPD is 0.001 as shown in Table2.2.There was also significant correlation between PCA subject scores and disease du-ration (p=0.007)(Fig2.7). The spatial pattern showed the most significant negativeweights in the caudate and putamen, and positive weights in the hypothalamus andhippocampus (Fig.2.6A).56Figure 2.6: PD-related spatial patterns in the serotonergic system obtained withA) PCA, B) ICA, C) PLS-DA, and D) RPCA methods. Error bars were calculatedwith leave-one-out cross validation.57Figure 2.7: Expression of the spatial patterns obtained with the PCA approach.Left: Subject scores (z-transformed) for healthy controls (label 0) and sPD subjects(label 1). Right: Scatter plot for disease duration (years) versus subject scores forsPD subjects.Inyzpznyznt Componznt Vnvl–sisFigure 2.8: Illustration for PCA versus ICA decomposition.58Figure 2.9: Expression of the spatial patterns obtained with the ICA approach.Left: Subject scores (z-transformed) for healthy controls (label 0) and sPD subjects(label 1). Right: Scatter plot for disease duration (years) versus subject scores forsPD subjects.The first limitation is that PCA decomposes the input data into orthogonal (un-correlated) components by maximizing the variance; however the constrain of or-thogonality may not make biological sense. independent component analysis (ICA)is an extension of PCA, which decomposes the data into statistically independentcomponents which may reflect independent disease-related mechanisms (Fig.2.8).Component 7 from the ICA approach gave better group discrimination power(p=0.0003) than PCA (Fig.2.9, but the single component only captured 6.7% ofthe total variance. The disadvantage of ICA is that the method is very sensitiveto initialization of the parameters, i.e. the results can be dramatically different ineach iteration. In addition, unlike with PCA, one needs to pre-define the numberof extracted components when doing ICA (in this case, we pre-defined 10 compo-nents). Defining different number of components (too few or too many than the trueunderlying independent components) can result in low stability of the results. Therelatively low stability can be seen from the relatively large variabilities (error bars)in the ICA spatial pattern (Fig.2.6B). Ideally, one might need to treat the numberof components as a hyperparameter when training the model and add additionalconstrains to minimize the effect of initialization. There was also significant correla-tion between ICA subject scores and disease duration, but the correlation strengthwas weaker than PCA (Fig.2.9).59evrtivl azvst hquvrz Yisxriminvtivz Vnvl–sisFigure 2.10: Illustration for PCA versus PLS decomposition. The red and purpledots indicate two different groups of subjects (predictors).PCA is a unsupervised learning method, i.e., it is blind to the group labels (pre-dictors) when doing group discrimination analysis. PCA first decomposes the datain a pure data-driven fashion, then group discrimination is achieved with logisticregression. This approach may not be ideal if the primary investigation goal is todo group discrimination (e.g. training a classifier to distinguish between healthycontrols and sPD). partial least square discriminative analysis (PLS-DA) is a su-pervised dimension reduction method that use the information in the predictors toguide the decomposition of the input variables (Fig.2.10), which can uncover thelatent relationship between the predictors and variables.60Figure 2.11: Expression of the spatial patterns obtained with the PLS-DA approach.Left: Subject scores (z-transformed) for healthy controls (label 0) and sPD subjects(label 1). Right: Scatter plot for disease duration (years) versus subject scores forsPD subjects.Comparing to PCA results, the first PLS-DA component accounted for 13%of the total variance and gave better group discrimination power as expected (p= 0.0002) (Fig.2.11). PLS-DA approach is ideal when the primary goal is to dogroup discrimination; however, in the case with DASB data, our goal was to ex-plore whether there existed a PD-related spatial pattern in the serotonergic systemwithout imposing the separation between groups, so PCA (unsupervised) was moresuitable for our particular exploratory study. In addition, there was no significantcorrelation between PLS-DA subject scores and disease duration; this may be due tothe PLS-DA focuses on capturing information for separating the two subject groupswhile ignoring information related to disease progression. The spatial pattern ob-tained with PLS-DA was similar to the one obtained with PCA (Fig.2.6C).61gowust erinxiplz Componznt Vnvl–sisFigure 2.12: Illustration for PCA versus RPCA decomposition.PCA can be very sensitive to outliers as shown in Fig2.12 since outliers can leadto large variance in the data. robust principal component analysis (RPCA) is avariation of the PCA algorithm. PCA decomposes the data by minimizing theobjective function (L2 norm) in the form:f(lPo) =n∑i=1w∑j=1(Q wj P zi S −mij)2 (2.1)where w is the principle components, z is the PCA scores and X is the input data.This objective function is in the form of a L2 norm (squared errors) which canamplify the effect of outliers.The RPCA method uses the following objective function instead:f(lPo) =n∑i=1w∑j=1Q wj P zi S −mij (2.2)This objective function is the form of L1 norm (absolute errors) is less sensitive tooutliers, however minimizing this objective function often requires more complexalgorithms.62Figure 2.13: Expression of the spatial patterns obtained with the RPCA approach.Left: Subject scores (z-transformed) for healthy controls (label 0) and sPD subjects(label 1). Right: Scatter plot for disease duration (years) versus subject scores forsPD subjects.As shown in Fig.2.13, RPCA captured 16% of the total variance and had asimilar group discrimination power compared to PCA (p=0.003), however therewas no significant correlation between RPCA subject scores and disease duration.The spatial pattern obtained with RPCA was similar to the one obtained with PCA.hummvr–Overall, the spatial patterns obtained with the four methods were similar whichdemonstrated the robustness of the PD-related spatial pattern in the serotonergicsystem. For our specific application (i.e. to explore if there exists a disease-relatedspatial covariance pattern in the serotonergic system), PCA seems to be the mostoptimal approach for the following reasons: 1) PCA is the simplest method withminimal need for parameter tuning, 2) PCA is a unsupervised learning method thatis suitable for exploratory analysis, 3) results obtained with PCA were relativelystable. However, in principle, all these methods are applicable for analyzing PETdata and the choice of a single method depends on the research question of interestand the structure of the data (i.e. whether there are significant outliers).63FBK ConwlusionUsing DASB PET this study provides a first evidence for the existence of a seroton-ergic spatial covariance pattern characteristic of PD (SPDRP) and a distinct patternobserved in LRRK2-NMC, who are at increased risk of PD. Even though the SPDRPwas derived using sPD subjects only, a significant expression of the same covariancepattern was also observed for LRRK2-PD. The pattern was found to be stronglycorrelated with disease duration except for two subjects, both of whom exhibited anunusual disease course, one of them being initially characterized as a subject with-out evidence of dopaminergic deficit. The pattern also correlated strongly with theputaminal dopaminergic denervation as estimated from DTBZ imaging. Comparedto previously used univariate analysis approaches, the spatial covariance methodwas found to be more sensitive in identifying disease-related abnormalities. Thisfinding suggests that disease-induced alterations of the serotonergic system, ratherthan being purely local, also affect interactions between separate regions in a diseasespecific fashion and is closely linked to abnormalities in the dopaminergic system.Likewise, the characteristic pattern observed in the LRRK2-NMC could provide newinsights into disease mechanisms and identify either early compensatory changes orrisk factors while the dopaminergic system is intact. Longitudinal observations anda larger subject cohort are required to disentangle these aspects.64Chuptyr Gbytwork hopologyGBE gummuryThe work described in this chapter contains original unpublished work. The analysispipeline developed for the work in this chapter resulted in a second author papercurrently under review.GBF IntroduwtionWe have previously shown in Chapter 2 that there exists a serotonergic spatialcovariance pattern characteristic of PD; the expression of this spatial pattern corre-lated significantly with both disease duration and the level of dopaminergic dener-vation, which indicates the pattern-based analysis may be more sensitive to disease-induced changes compared to univariate measures. We also showed there exists adistinct pattern in LRRK2-NMC subjects, who are at higher risk of PD comparedto healthy control population [11]. These findings suggest that disease-induced al-terations of the serotonergic system, rather than being purely local, also affect theinteraction between multiple brain regions. In the work described in this chapter,we extended the analysis to examine the effects of disease and LRRK2 mutation onthe network topology in the serotonergic system using graph theory analysis.Graph theory analysis has been used to study topological properties of structuraland functional network organization in healthy and diseased brains using MRI andEEG data. In functional imaging, a graph is composed of nodes (ROI) and edgesrepresenting functional connection between spatially distributed brain regions. InfMRI, the edges represent the correlated temporal oscillations between each pair ofbrain regions.In PET, however, the application of graph theory analysis has been limited.This is mainly due to the limited temporal information available in the acquiredPET data; the temporal information in the PET signal is often embedded in the65quantitative estimates of tracer kinetics (e.g. BPaW for DASB tracer). The lack oftemporal domain limits the calculation of the functional connections to a group levelinstead of at a subject level. The functional edges, instead of reflecting correlationsacross all temporal points between two brain regions in each subject, representcorrelations across subjects between two brain regions in each subject group. Thisassumes there is minimal within-subject variability in each subject group and makescomparing differences between groups and examining individual-specific changeschallenging.Graph theory analysis has been used to study metabolic topology using FDGtracer in PD [69] and has been recently extended to study network changes inneurotransmitter systems in Alzheimer’s disease [99]. In the work described inthis chapter, we applied graph theory analysis to DASB data to investigate possibledisease and LRRK2 mutation related topological changes in the serotonergic system.GBG aythods und autyriulsGBGBE gtudy durtiwipuntsFor the work described in this chapter, we used the same dataset as described inChapter 2 with the addition of three more sPD subjects. This study included 18sPD, eight LRRK2-PD, nine LRRK2-NMC, and nine healthy controls age-matchedto both sPD and LRRK2-PD groups. Exclusion criteria included clinical history ofdepression, active anti-depressant therapy or medication with known serotonergicaction and a BMI greater than 35. The healthy controls had no history of neurolog-ical or psychiatric disorders and were not taking any medication. Disease durationwas estimated as time from onset of motor symptoms as reported by the subjects.The PD subjects were clinically evaluated with the MDS-UPDRS and H&Y to as-sess motor dysfunctions, MoCA scores to assess cognitive performance and BDI. Allassessments were performed off medication. The study was approved by the Clin-ical Research Ethics Board of the University of British Columbia and all subjectsprovided written informed consents.GBGBF gwunning drotowolThe same scanning protocols were used as in the work described in Chapter 2 [11].All study participants underwent a [11C]-DASB PET scan and a T1-weighted66brain MRI scan. Prior to PET imaging, patients withdrew from all anti-parkinsonianmedications for at least 12 hours. 558±2MBq of [11C]-DASB were administered byintravenous injection over 60 seconds using an infusion pump (Harvard Instruments).Participants were positioned using external lasers aligning the gantry with the in-ferior orbital-external meatal line, and custom fitted thermoplastic masks were ap-plied to minimize head movement. All patients were scanned on a Siemens HRRT(Knoxville, TN) with a spatial resolution of 2O5mm3 [86]. Acquired data werebinned into 18 time frames (frame duration: 4×60 s, 3×120 s, 8×300 s, 3×600 s; im-age dimension=256×256×207; voxel size=1O22mm3) for a total duration of 80min.Transmission scans required for attenuation correction were performed over ten min-utes with a rotating 137Cs source. PET images were reconstructed using the 3Dlist-mode OP-OSEM algorithm[87] with 16 subsets and six iterations, with cor-rections for decay, dead-time, normalization, attenuation, scattered and randomcoincidences. After reconstruction, the images were smoothed with a 3mm FWHMGaussian filter to reduce noise. The frames were spatially realigned for each subjectwith rigid-body transformation to minimize the impact of motion artifacts duringscans. The structural MRI scans were performed on a Philips Achieva 3T MRIscanner (Phillips Healthcare, Best, NL) using the T1 TFE sequence (TR/TE =7.7/3O6ms; TFE shots=218; flip angle = 8◦; image dimension: 256×256×170; voxelsize 1mm3.GBGBG Imugy drowyssing und UnulysisA similar image processing and analysis pipeline were used as in the work describedin Chapter 2 [11].Image preprocessing was done using Statistical Parametric Mapping (SPM12)software (Wellcome Department of Cognitive Neurology, University College London,UK) running on Matlab 9.0 (Mathworks Inc., Natick, MA) and MEDx (Sensor Sys-tems, Sterling, VA). The MRI images were resampled with trilinear interpolation tomatch the voxel size of the PET images. The resized MRI images were then coregis-tered to the corresponding mean PET images (mean of 18 frames) using normalizedmutual information to estimate the affine transformation for each subject. Theseresized MRI images were then warped into the MNI space and their inverse defor-mation fields were saved. The inverse deformation field vectors were applied to ROItemplates predefined in the MNI space to bring them into the individual subject’sPET space in a single step. The quality of each processing step was checked visually67for all scans.The ROI templates were defined by experienced neurologists in the MNI spaceusing MRI and PET data from healthy subjects for a total of 43 non-overlappingROIs which are known to be involved in the serotonergic system. The regions in-cluded medulla, midbrain, pons, rostral raphe nucleus, ventral tegmental area and19 ROIs placed bilaterally (resulting in 38 ROIs over both hemispheres): anteriorand posterior cingulate, amygdala, caudate, cerebellum, dorsolateral prefrontal cor-tex, hypothalamus, insula, orbital frontal cortex, PPN, putamen (anterior, middleand posterior), substantia nigra, thalamus, ventral striatum, hippocampus, dentatenucleus, and globus pallidus. Decay-corrected TAC were calculated for each ROIusing the Marsbar toolbox in SPM. BPaW [29] values in 42 ROIs were obtainedusing the SRTM with the cerebellum as reference region [31]. Because we did notobserve any significant asymmetry in the left and right hemisphere, we used theaverage binding values in both hemispheres for bilaterally placed ROIs to reducethe number of variables. In total, we had 24 ROIs for further analyses.GBGB4 Gruph hhyory UnulysisIn the context of applying graph theory analysis to DASB data, a node is a brainregion with significant level of DASB binding (reflecting a high number of seroton-ergic nerve terminals) and an edge is the functional connection between a pair ofbrain regions across all subjects in each group. DASB BPaW values in each ROI(matrix size N×M, where N is the number of subjects and M is the number of vari-ables/ROIs, NQM) were first demeaned and were used to generate the ROI by ROIadjacency matrices (MxM, in this case, 24×24) for each subject group.Vyjvxznx– bvtrixzsWe can compute two types of adjacency matrices as input to graph theory analysis:1) full correlation matrix (normalized covariance matrix Σ) representing marginalcorrelation between variables; 2) partial correlation matrix or precision matrix (in-verse of the normalized covariance matrix Σ−1) representing conditional correlationsbetween pairs of variables given the remaining variables (conditional independence).Full correlation matrix is the most common type of input for graph theory anal-ysis in the neuroimaging field. However, when using full correlation matrix, twobrain regions might show high functional correlation when there is actually no di-68rect functional connection between them but because they are both connected to acommon source from a third region.Using partial correlation matrix as input allows us to remove these indirect con-nections assuming the functional connectivity structure in the brain is sparse [100].For neurotransmitter systems, partial correlations may reflect physical projectionsof nerve terminals from the nerve bodies. For example, the serotonergic neurons arepredominantly located in the raphe nucleus of the brain stem, and the neurons thenproject into the cortical and sub-cortical regions. Using partial correlations mayremove some of the ’indirect’ connections between the projections (e.g. between thecortical and sub-cortical regions) and only keep the ’direct’ connections originatedfrom the raphe nucleus. However, this approach might also limit the connectionsalong the physical projections of the serotonergic system.However, as with many neuroimaging studies, there are often more variables/ROI(M) than the number of observations/subjects (N), thus making the input matrixrank-deficient and the covariance matrix of the rank-deficient input matrix non-invertible. We attempted to use graphical least absolute shrinkage and selectionoperator (GLASSO) to estimate the partial correlation matrix of a rank-deficientmatrix. Details about the GLASSO algorithms were previously published [101].Briefly, in order to estimate the inverse covariance matrix (precision matrix)(Σ−1), we need to maximize the following log-likelihood:log(dzt(Σ−1))− tr(SΣ−1) (3.1)where tr denotes the trace (sum of diagonal elements) and S is an empirical covari-ance matrix.GLASSO uses a L1 regularization term (/‖Σ−1‖1, where / is the regularizationterm for the L1 norm) to increase the sparsity in the resulting precision matrix. Theestimation of the partial correlation depends on finding the optimal value for theregularization term / by minimizing AIC scores.log(dzt(Σ−1))− tr(SΣ−1)− /‖Σ−1‖1 (3.2)With either full or partial correlation matrices, we have a group-level adjacencymatrix. After the estimation of precision matrices using our DASB data, we com-pared the density/sparsity of the precision matrices (estimated by GLASSO) toestimate the amount of direct functional connections for the four subject groups.69Grvph ihzor– bztrixsThe adjacency matrices were then used to construct binary undirectional graphs at30 linearly spaced sparsity thresholds ranging from 0.2 to 0.5. A sparsity threshold of0.2 means we keep the top 20% strongest connections in the adjacency matrices. Thesame threshold was applied to each group-level graph to force graphs from all fourgroups to have the same number of edges for further analyses. The following graphtheory metrics were computed using the Brain Connectivity Toolbox in Matlab2019a. Detailed description for each of the graph theory metric was described indetails in previous publications [9]. Briefly:1. Clustering coefficient (C) is the probability that the neighbours of a node arealso neighbours of each other and is a measure of network density:X =1n∑i∈nXi =1n∑i∈n2tiki(ki − 1) (3.3)where n is the total number of nodes, Xi is the clustering coefficients of node,ti is the number of distinct triangles for node i, and ki is the degree (numberof edges connecting to a node) of node i. Normalized clustering coefficient() is the ratio between the clustering coefficient C and the average clusteringcoefficient of 10 random graphs with the same number of nodes and edges asthe tested graph.In the context of the DASB data, clustering coefficient measures the level ofintegration in the serotonergic network at a nodal level. A high clusteringcoefficient indicates there is a high level of dense local connections (formingdense clusters) in the serotonergic system.2. Characteristic pathlength (L) is the average shortest distance (pathlength)from one node to all the other nodes and is a measure of network efficiency:a =1n∑i∈nai =1n∑i∈n∑j∈nNj ̸=j dijn− 1 (3.4)where ai is the pathlength between each node i and all other nodes and dijis the pathlength from node i to node j. Normalized characteristic pathlength() is the ratio between the characteristic pathlength and the average charac-teristic pathlength of 10 random graphs.70In the context of the DASB data, characteristic pathlength measures the levelof integration in the serotonergic network at a global level. A high charac-teristic pathlength indicates there is a reduced global efficiency in connectionbetween more functionally distinct clusters.3. Small-worldness () is ratio between the normalized clustering coefficient andnormalized characteristic pathlength ( = ). A high small-worldness indi-cates the network organization in the serotonergic system is close to a small-world structure.4. Modularity is a measure of the amount of hubs/modules in a network and iscalculated as:f =∑u∈M[zuu − (∑v∈Mzuv)2] (3.5)where M is the total number of non-overlapping modules and zuv is the propor-tion of all edges that connect the nodes in module v [102]. A hub/module con-tains several densely interconnected nodes and there are relatively few edgesbetween nodes in different modules [103]. A modular network therefore tendsto show high small-worldness; however, the reverse is not always true. Mod-ularity is a measure of network segregation and a high modularity indicatesthere is well-defined modular structure in the serotonergic system.Leave-one-out cross-validation was performed in each subject group to exam-ine the stability of the graph theory metrics and visualize the group differences.Statistical significance of the group differences was estimated against the null dis-tributions of group differences using 100 iterations of random permutation of theinput matrices [99].GB4 fysultsGB4BE Udjuwynwy autriwysPartial correlation matrices obtained with GLASSO had the same number of edges(density) for all four subject groups (density = 0.4348); the regularization terms /were found to be 0.041 for healthy control group, 0.057 for sPD group, 0.024 forLRRK2-PD group and 0.087 for LRRK2-NMC group.71Figure 3.1: A) Full correlation matrices (ROI by ROI) for the four subject groupsat threshold = 0.3 (when taking the top 30% of the strongest correlations in eachsubject group) which was used as input for further analyses; ROI names are listedin Table3.1 B) Histograms of the absolute values of the full correlation coefficientsfor the four subject groups without thresholding; red vertical lines represent meanvalues for all correlation coefficients in each subject group.Full correlation matrices at 30% threshold are shown in Fig.3.1A for the foursubject groups. All ROIs in the correlation matrices were ordered the same way asthe order of ROIs in the correlation matrix for the healthy control group. Visually,all four subject groups showed distinct network organizations in the serotonergicsystem. sPD and LRRK2-PD subjects showed the most visually different networkorganizations compared to the healthy control brain network, while LRRK2-NMCsubjects showed the closest resemblance to the healthy control brain network. BothPD groups showed the most obvious disruption in the cluster involving brain stemregions, amygdala, hypothalamus and substantia nigra compared to healthy con-trols. LRRK2-NMC subjects showed a slight increase in connectivity between thestriatum and cortical regions compared to healthy controls.The mean values for the absolute correlation coefficients were 0.43 for the healthycontrol group, 0.48 for the sPD group, 0.42 for LRRK2-PD group and 0.62 forLRRK2-NMC group as shown in Fig.3.1B.72gdI cvmzF AmygdalaG MedullaH Substantia NigraI HypothalamusJ Dorsal MidbrainK Ventral Tegmental AreaL Pedunculopontine NucleusM PonsN Raphe NucleiFE Ventral StriatumFF HippocampusFG InsulaFH CaudateFI Posterior CingulateFJ Anterior Cingulate 1FK Dorsolateral Prefrontal CortexFL Orbital Frontal CortexFM Anterior Cingulate 2FN Middle PutamenGE Anterior PutamenGF ThalamusGG Dentate NucleusGH Posterior PutamenGI Globus PallidusTable 3.1: ROI names and orders for the correlation matrices shown in Fig.3.1A.GB4BF Gruph hhyory aytriwsGraph theory metrics calculated based on partial correlation matrices were ex-tremely unstable due to the high variability in the determination of indirect edgesduring cross-validation. We therefore focused on the results obtained with full cor-relation matrices.Graph theory metrics were computed and compared between subject groupsusing a range of proportional thresholds ranging from 20% to 50% as shown inFig.3.2. Graph theory metrics were not stable at lower thresholds (Q20%). Wechose to use 30% threshold for illustrating the group differences from the randompermutation test; the random permutation results were consistent across the rangeof thresholds (S20%).73Figure 3.2: Graph theory metrics for the four subject groups measured at differentsparsity thresholds (20%-50%). Error bars were obtained with leave-one-out cross-validation.At 30% threshold as shown in Fig.3.3, both PD groups showed significantly re-duced modularity (P Q0.01) and significantly lower mean clustering coefficients com-pared to healthy controls (sPD vs healthy controls: P Q0.05; LRRK2-PD vs healthycontrols: P Q0.01). However, there were no significant group differences betweenPD groups and healthy controls in small-worldness and characteristic pathlength.At 30% threshold (Fig.3.3), LRRK2-NMC showed significantly lower characteristicpathlength (P Q0.01) and significantly lower modularity (P Q0.01) compared tohealthy controls. No significant group differences were found in mean clustering74Figure 3.3: Graph theory metrics for the four subject groups measured at 30%sparsity thresholds. Error bars were obtained with leave-one-out cross-validation. *= p-value Q0.05. ** = p-value Q0.01. Significance was estimated with 100 iterationsof random permutation test.coefficients and small-worldness between LRRK2-NMC and healthy controls.GBI DiswussionGBIBE Udjuwynwy autriwysThe sparsities of the precision matrices were identical among all four subject groups,indicating the number of ’indirect’ connections identified by GLASSO was similar.We attempted to use GLASSO to estimate partial correlation coefficients for PETdata, however, finding the optimal value for the regularization term / depends onthe initial guess and the step size in the iterative process. If the initial guess for / is75too small, the estimation results may not converge; if the initial guess is too large,we might miss the the true solution completely. In our case, we used 0.02 for theinitial guess and step size of 0.001 for all four subject groups, which resulted in theprecision matrices with the same sparsity in all four groups.In addition, estimation of the precision matrices were sensitive to outliers andwithin-subject variabilities in the input data, which resulted in large variabilities insome graph theory metrics in the leave-one-out cross-validation process. This hadan especially large impact on the characteristic pathlength, since this metric is verysensitive to the presence of long (and may be ’indirect’) edges. For these reasons,we decided to use full correlation matrices as input for further analyses.Both PD groups showed clear disruption of the healthy control network, withmost obvious changes in the cluster involving brain stem regions, amygdala, hy-pothalamus and substantia nigra. LRRK2-PD group showed a more severe disrup-tion compared to sPD, which is likely due to the longer disease duration in thisgroup. However, the average correlation strengths were similar in the PD groupsand healthy controls, suggesting there may be a shift in the organization of the func-tional network rather than an global disruption of the overall functional connectivityin the serotonergic system.The average correlation strength was the strongest in LRRK2-NMC group com-pared to the healthy controls and PD groups, indicating there was an overall increasein functional connectivity in the serotonergic system; this is consistent with previousfindings that LRRK2-NMC subjects showed increased SERT binding compared tohealthy controls [44]. In addition to the overall increase in functional connectivity,there were also changes in the network organization; consistent with previous obser-vations that LRRK2-NMC showed a distinct spatial covariance pattern comparedto both healthy controls and PD subjects [11].GBIBF Gruph hhyory aytriwsGraph theory metrics were not stable at lower thresholds (Q20%) especially forcharacteristic pathlength due to large variability of longer connections in the inputmatrices at lower thresholds. Both PD groups showed lower clustering coefficientsand lower modularity compared to healthy controls. This suggests that there is notonly a disease-induced spatial pattern as shown in Chapter 2, there also seems to bedisease-induced changes in network organizations in the serotonergic system. Morespecifically, there was a reduction in local efficiency as measured by both clustering76coefficients and modularity, suggesting that disease might reduce the functionalconnectivity in normally condense local clusters. There was no significant disease-induced effect on long connection between functionally distant clusters as measuredby characteristic pathlength.LRRK2-NMC group, on the other hand, showed increased connectivity betweenfunctional segregated clusters and reduced modularity compared to healthy con-trols. The increased efficiency to functionally connect distant clusters may serve aregulatory role for the reduced local functional connectivity within dense clusters.GBIBG Limitutions und ConsidyrutionsThe relatively small sample size is a major limitation for this study. However,Veronese et.al. showed that some graph theory metrics (e.g. clustering coefficients)were stable even with small sample size (N = 10) [99], suggesting the relative ro-bustness of the graph theory metrics. The relatively small error bars obtained withleave-one-out cross-validation especially at higher thresholds as shown in Fig.3.2 andFig.3.3 also indicate the robustness of the graph theory metrics when using full cor-relation matrices as input. Another observation from this study is that some graphtheory metrics were more robust than others. In particular, we found estimation ofcharacteristic pathlength depended the most on the threshold level while metrics likeclustering coefficients and modularity were relatively stable. One should have extracaution when interpreting the less stable metrics especially at lower thresholds.Another major limitation is the lack of subject-specific adjacency matrices, mak-ing it impossible to correlate the graph theory metrics to clinical measures or uni-variate tracer binding values. To address this issue, raw dynamic PET data (withoutthe use of kinetic modelling) can be used as input to incorporate the temporal in-formation in the construction of subject-specific adjacency matrices. This approachof using tracer activity data at multiple time point during the scan has been im-plemented using FDG data [104, 105]. However, BPaW values give informationon specific tracer-target interactions while raw PET activity signal also containsadditional information on tracer binding to non-specific targets. The addition ofnon-specific binding may have a larger impact on tracer measuring neurotransmit-ter system activities compared to tracer measuring metabolic activities (FDG). Theapplicability and limitations of extending graph theory analysis to the temporal do-main of PET neurotransmitter data should be further investigated in future studies.77GBJ ConwlusionIn Chapter 2, we showed that there was an overall pattern of disease and mutationrelated alterations. In this study, we used graph theory analysis to examine thedisease and LRRK2 mutation related topological organization changes in the sero-tonergic system. Examining these topological organization changes provide moredetails about how disease or mutation induces changes in the serotonergic networkin addition to the overall pattern of alterations. In particular, we showed thatthere was reduced local efficiency in PD subjects and an increase in global efficiencyin LRRK2-NMC in the serotonergic system. These findings suggest that the PD-related spatial pattern observed in Chapter 2 was likely a result of the breakdown ofnormally densely connected brain regions. Likewise, the LRRK2-mutation-relatedpattern likely reflected a compensatory mechanism resulting in increasing functionalconnections between normally segregated clusters of brain regions. These more de-tailed examination of the serotonergic network could not otherwise be observed withmultivariate models introduced in Chapter 2. However, interpretation for the bi-ological meaning of the change in topological organizations should be confirmedwith additional analyses, especially at nodal level. From a technical point of view,we showed that graph theory analysis may be applicable to study changes in theneurotransmitter systems. The methods presented in this chapter may be useful forcomparing topological changes for multi-modality imaging data in a range of clinicalapplications.78Chuptyr 4gputioAhymporul duttyrns4BE gummuryThis chapter contains published work [uA JC[CA Kl–uzhinA IChCA bxKzofinAbCJCA htozsslA VCJCA hossiA kCA GEGEC covzl yvtvByrivznA zquvtionBfrzzmzthoy xvpturzs spvtioBtzmporvl pvttzrns of nzuroyzgznzrvtion in evrkinBsons yiszvszO Vpplixvtion of y–nvmix moyz yzxomposition to eZiC czuBroImvgz ClinC GJA FEGFJE [13]. I was responsible for development of the analysismethodology, statistical analysis, clinical interpretation and manuscript composi-tion. Scanning procedures were performed by the staff members of the UBC PETimaging group. I. Klyuzhin contributed to image preprocessing. M.J. McKeownprovided feedbacks on the development of the analysis methodology. A.J. Stoessland V. Sossi contributed to clinical interpretation. A.J. Stoessl and V. Sossi de-signed the study. V. Sossi was the the supervisory author involved throughout theproject in the concept formation and manuscript preparation.In the work described in the first two chapters, we focued on the methods thatexamine changes in spatial distribution of tracer binding in disease stage. However,the previous methods may not be optimal for extracting temporal changes in tracerbinding related to disease progression. In this chapter, we introduced and validatedthe use of a novel data-driven and equation-free approach, dynamic mode decompo-sition, to capture and quantify spatio-temporal patterns of neurodegeneration usingPET data. While this approach will be applicable to a wide range of imaging data,we tested the applicability and robustness of the approach using DTBZ PET tomodel progressive dopaminergic denervation in PD.From a methodological point of view, the proposed method has several advan-tages over traditional methods in terms of biologically-relevant information that canbe extracted from the data. 1) It considers tracer distributions in all the selectedregions at once, thus providing information not only on localized alterations, butalso on spatial patterns of such alterations, emphasizing the network behaviour of79the targets under investigation. 2) This approach incorporates both spatial andtemporal information simultaneously in a data-driven and equation-free fashion tomodel disease progression. Traditional approaches either use a pre-defined model tofit the imaging data across disease duration or use static multivariate methods (i.e.PCA or ICA) which only models spatial information. 3) While traditional methodscan only model the overall disease progression, the proposed approach allows thedecomposition of overall disease-induced changes into orthogonal temporal curves,possibly relating to independent underlying disease-related mechanisms.From a clinical point of view, we were able to, for the first time, decompose thedopaminergic denervation in the striatum associated with Parkinsons disease intotwo spatio-temporal patterns: (i) the anterior-posterior gradient in the putamen andhead-tail gradient in the caudate, which may be related to non-specific mechanismresponsible for disease progression and (ii) the dopaminergic denervation along thedorsal-ventral gradient in the putamen may reflect independent mechanisms respon-sible for disease initiation in very early stage of the disease.While the data considered in this study allowed us to validate the approachand provided new insights into Parkinsons disease, the approach appears very wellsuited for quantifying and decomposing disease-induced progressive changes in otherPET tracer datasets and in other disease datasets, and is possibly able to separatedifferent disease mechanisms and assess their topology at different disease stages.4BF IntroduwtionMany neurodegenerative diseases are characterized by progressive loss of neuronsand/or nerve terminals throughout the course of disease in the form of specific spatio-temporal patterns, in which neurodegeneration in different brain regions or networkof brain regions may follow distinctive temporal disease progression. Non-invasiveneuroimaging techniques such as PET and single-photon emission computed tomog-raphy (SPECT) are often used to track disease progression and to better understanddisease mechanisms. Many attempts have been made to explore spatio-temporalpatterns of disease progression using PET or SPECT. One common approach is tofit an exponential or another appropriate model with a fixed number of spatial andtemporal parameters to tracer binding values. Previous studies have explored thespatio-temporal patterns of dopaminergic denervation at both individual ROI level[51, 106] and at the voxel level [15, 107] using several dopaminergic PET tracers80and [123I]-ioflupane SPECT. However, with this approach, the progressive spatialpatterns are either limited to a set of pre-defined ROIs, or to being along a specificspatial axis within a brain structure. Data-driven multivariate approaches have alsobeen recently used to extract and visualize spatial patterns of tracer binding values,mainly using dimension reduction techniques such as PCA [14] or ICA. One ma-jor constraint associated with these methods is that PCA and ICA are temporallystatic techniques and do not model the temporal progression of the spatial patterns.One needs to assume that the spatial patterns obtained with PCA or ICA remainstatic/constant over time and only the expression of these spatial patterns increasesor decreases as disease progresses; this assumption may not necessarily be valid orin keeping with the fundamental nature of the disease.We introduce and validate the use of dynamic mode decomposition (DMD) toextract spatial patterns of tracer binding values associated with distinctive andorthogonal temporal disease progression curves. DMD is a relatively new decompo-sition method first developed in the field of fluid dynamics [70] and recently used tomodel temporal oscillations of EEG [108] and fMRI) data [72] in the brain. DMDfinds coherent spatio-temporal patterns in high-dimensional non-linear systems (de-tails about DMD algorithms are included in the Method section). There are severalunique advantages of DMD compared to other model fitting and multivariate ap-proaches: 1) it is a data-driven method that does not require a fixed set of governingequations or prior assumptions of the underlying dynamics; 2) it combines the ad-vantages from two frequently used analysis methods: PCA for the reduction ofhigh-dimensional data and spectral time-series analysis for identifying the oscilla-tion frequency of time-varying signals; 3) it can model non-linear systems effectively,unlike PCA which assumes that the relationships between variables are linear; 4) itcan isolate/decompose the overall temporal course into specific dynamics [73, 74].Instead of applying DMD to sequential time-series data in the frequency domain asis the case for EEG and fMRI data, we propose to extend DMD to model temporalchanges of spatial patterns of tracer binding values as disease progresses.To test the applicability and robustness of DMD to extract spatio-temporalpatterns of tracer binding values, we first applied DMD to study dopaminergic den-ervation in PD using [11C](+)DTBZ (a VMAT2 marker) PET. The motor deficit ofPD is traditionally associated with dysfunction of the nigrostriatal pathway, charac-terized by progressive loss of dopaminergic neurons in the substantia nigra and lossof their projection fibres to the striatum [6]. Neurodegeneration of the dopamin-81ergic system tends to follow a fairly well defined spatio-temporal pattern in whichthe dorsal posterior putamen contralateral to the more affected body side is affectedfirst, followed by degeneration in the ventral and anterior putamen and the caudate,as shown in Fig.4.1 [6]. PET studies have shown an exponential decline of dopamin-ergic terminals as disease progresses, where the rate of loss is highest in early disease[51, 109]. This is in broad agreement with post-mortem studies of nigral cell counts[110] and striatal tyrosine hydroxylase immunoreactivity [111]. Using longitudinaland cross-sectional PET data, our group previously showed that while initial levelsof dopaminergic loss were different, the rates of dopaminergic loss were similar indifferent striatal regions. It was therefore suggested that mechanisms responsiblefor disease progression (rate of loss) and disease initiation (initial severity of loss)may be different [51, 109]. It is thus expected that these distinct mechanisms maybe reflected by distinctive spatio-temporal patterns in the striatum, which could becaptured by DMD.Figure 4.1: [11C](+)dihydrotetrabenazine (DTBZ) PET image for a healthy control(left) and a Parkinson’s disease (PD) subject (right). PD subject showed charac-teristic asymmetric tracer uptake in the less and more affected sides and a spatio-temporal pattern of dopaminergic loss with the posterior putamen affected beforethe anterior putamen and caudate. PET = Positron Emission Tomography.824BG aythods und autyriuls4BGBE gtudy durtiwipuntsThis study included 41 PD subjects with disease duration ranging from 0 to 16 years.Disease duration was estimated as time from onset of motor symptoms as reportedby the subjects. PD subjects were clinically evaluated using the MDS-UPDRS andH&Y to assess motor dysfunction. All PD subjects were cognitively normal asassessed by MoCA (MoCA scores greater than 26). Detailed clinical characteristicsare listed in Table4.1. All assessments were performed off medication. The studywas approved by the Clinical Research Ethics Board of the University of BritishColumbia and all subjects provided informed written consent.cumwzr hzx Vgz =–zvrs) Yiszvsz yurvtion =–zvrs) ]&n iotvl bYhBjeYgh III boCVeY 41 25M/16F 61.8 ± 8.7 5.32± 4.31 1.9 ± 0.5 19.3 ± 10.5 28.0 ± 1.3Table 4.1: Clinical characteristics of all subjects. All numbers are reported asmean± standard deviation. Disease duration estimated as the time from onset ofmotor symptoms as reported by the patients. PD = Parkinson’s disease subjects;MDS-UPDRS = Movement Disorder Society Unified Parkinson’s Disease RatingScale; MoCA = Montreal Cognitive Assessment; H&Y=Hoehn and Yahr scale.4BGBF gwunning drotowolsAll study subjects underwent DTBZ PET scans and a T1-weighted MRI scan ofthe brain. The PET scans were performed on a Siemens HRRT (Knoxville, TN)with a spatial resolution of 2O5mm3 [86]. Subjects were positioned using externallasers aligning the gantry with the inferior orbital-external meatal line, and customfitted thermoplastic masks were applied to minimize head movement. Prior to PETscans, subjects were withdrawn from all anti-parkinsonian medications for at least12 hours. 320±34MBq of (+)DTBZ were administered by intravenous injectionover 60 seconds using an infusion pump (Harvard Instruments). Acquired data werebinned into 16 time frames (frame durations: 4×60 s, 3×120 s, 8×300 s, 1×600 s;image dimension=256×256×207; voxel size = 1O22mm3) with a total duration of60 minutes. Transmission scans required for attenuation correction were performedover ten minutes with a rotating 137Cs source. PET images were reconstructed usingthe 3D sinogram OP-OSEM algorithm [87] with 16 subsets and six iterations, withcorrections for decay, dead-time, normalization, attenuation, scattered and randomcoincidences. After reconstruction, images were smoothed with a 2mm FWHM83Gaussian filter to reduce noise. The frames were spatially realigned with rigid-bodytransformation to minimize the impact of motion during scans. The structural MRIscans were performed on a Philips Achieva 3T MRI scanner (Phillips Healthcare,Best, NL) using the T1 turbo field echo (TFE) sequence (TR/TE = 7.7/3O6ms; TFEshots=218; flip angle = 8◦; image dimension: 256×256×170; voxel size 1mm3).4BGBG Imuging drowyssingParametric DTBZ binding images were generated with a previously published anal-ysis pipeline [14]. To optimize the co-registration and warping quality, we used atwo-step registration pipeline. In the first step, the MRI images were first resampledto match the PET voxel size. DTBZ PET images averaged over 30-60min post-injection were then rigidly co-registered to the corresponding subject’s MRI imagesusing SPM12 software (Wellcome Department of Cognitive Neurology, UniversityCollege London, UK). The quality of the co-registration was visually inspected.The MRI images were segmented to create masks for the striatal and the referenceregion (occipital cortex) using Freesurfer [112]. Parametric activity ratio (activityratio (AR)) images were generated by dividing the voxel values in the respectiveactivity images by the mean activity in the reference region. Co-registered MRIand AR images were separated into more and less affected brain sides, contralateralto the more and less clinically affected body sides (based on MDS-UPDRS). In thesecond step, for each side separately, MRI-defined segmentation (Freesurfer) of theputamen and caudate were combined to generate a single labeled volume mask thatwas warped to a common striatal template using 3D diffeomorphic mapping [112].The resulting transformation matrix was saved and applied to the respective ARimages.4B4 Introduwing DaDWe first briefly summarize the DMD algorithm in section4.4.1 [71, 113]. Applicationof DMD in the context of neuroimaging to model disease progression is illustratedin section4.4.2 with a brief explanation of meaningful outcome measures. Tests forrobustness and reproducibility of DMD results are shown in section4.4.3. Compar-ison between the results obtained with DMD, univariate analysis and PCA can befound in section4.4.4 and section 4.4.5.844B4BE DaD UlgorithmConsider measurements taken from n features at times k∆t, where k is the indexof the temporal snapshots and ∆t is the time difference between each snapshot(temporal resolution/sampling rate). Measurements from each temporal snapshotare arranged in a column vector xk (size n by 1, k=1...t, t is the index of last timepoint). Construct two matrices mt and mt−1 (both are size n by t-1) as time-shiftedversions of each other (shifted by ∆t):mt−1 = [x1OOOxt−1]mt = [x2OOOxt](4.1)The progression from mt−1 to mt (temporally progressed by ∆t) is governed byan unknown linear operator A:mt = Vmt−1 + "t (4.2)where "t is the model residual or noise. DMD models a high-dimensional linear re-gression of the non-linear dynamics relating mt and mt−1 by the eigendecompositionof the operator A. To estimate A (size n by n), singular value decomposition is firstapplied to mt−1 so that:mt ≈ Vmt−1 = VUΣk ∗ (4.3)where U (size n by r, where r is the number of reduced dimension), Σ (size r by r)and V (size t-1 by r) are the left-singular vectors, singular values and right-singularvectors of mt−1. Then:V ≈ mtm†t−1 = mtk Σ−1U∗ (4.4)where m†t−1 is the pseudoinverse of mt−1. However when n is large, direct eigenvalueanalysis of A can be computationally expensive. A more efficient model is to projectA into the reduced dimensional space with the operator U. Now we define thereduced order model V˜:V˜ = U∗VU = U∗mtk Σ−1 (4.5)85And the eigendecomposition of V˜ (size r by r) is:V˜l = Λl (4.6)where W (size r by r) is the eigenvectors and the diagonal of Λ contains DMD eigen-values . Note that the eigenvalues for A and V˜ are identical, and their eigenvectorsare associated by a linear transformation. DMD modes (eigenvectors of A, size nby r) are defined as:Φ = mtk Σ−1l (4.7)The temporal dynamic curve is defined as:i (t) = zΩtzΩ = log(Λ)R∆(t)x1 = Ωz(4.8)where z (DMD amplitude, unitless) is used to scale the DMD mode (Φ) to matchthe input data at the first time point x1. Since Λ can be real or complex valued,the temporal dynamic curve shows an exponential behaviour when Λ is real andshows an oscillatory behaviour when Λ is complex. This assumes that the overalltemporal progression of the input data can be described by a sum of exponentialand oscillatory functions.We can then approximate X with a dynamic model:mˆ(t) = ΦzΩtzThe main message is that DMD decomposes the data into coupled spatio-temporalpatterns: spatial modes (Φ) and their corresponding temporal dynamics (T(t))(which follow either exponential or oscillatory behaviours).4B4BF DaD for hruwking Disyusy drogryssionUnlike fMRI and EEG data, where each temporal snapshot is a single scan fromthe dynamic time-series, here we want to model disease progression; we used cross-sectional data of subjects with different disease durations to represent temporalsnapshots of the disease course (Fig.4.2A). We first stretched each 3D parametricAR image into a long vector xk (size n by 1, n is the total number of voxels in the86Figure 4.2: Schematic diagram for dynamic mode decomposition (DMD) analysispipeline. (A) In the data preparation step, the 3D parametric PET tracer bindingimage of each subject is stretched into a flattened column vector. Each column vectoris then concatenated horizontally according to the disease durations of all subjects.DMD then decomposes the reshaped PET data into DMD modes (spatial patterns),each associated with an unique temporal dynamic curves. (B) The reshaped PETdata are used to constructmt−1 andmt matrices as time shifted version of each other,which are then used as input to DMD. PET = Positron Emission Tomography.87AR image, k=1...t) then temporally concatenated all k vectors according to theirdisease durations (in years) such that (Fig.4.2B):mt−1 = [x1OOOxt−1]mt = [x2OOOxt](4.9)where x1 represents the AR image vector for a subject with a disease duration ofone year.For time points with several snapshots (i.e., AR images of subjects with thesame disease duration), we used the averaged AR images as a single snapshot. EachDMDmode represents a spatial pattern of dopaminergic denervation associated witha particular exponential growth/decay curve or an oscillatory temporal curve aroundzero (as a result of complex Λ values). Only DMD modes with real Λ values wereconsidered, and DMD modes with oscillatory behaviour were considered as noise. Itis also important to note that DMD spatial modes are not orthogonal/independent,while the orthogonality of DMD remains in the temporal domain, i.e. the temporalcourses are orthogonal (loosely independent) to one another. While the decay con-stant in the temporal curves reflects the relative changes in the expression strength ofthe DMD spatial pattern per year, the DMD amplitude (z) represents the expressionstrength of the DMD spatial pattern at disease onset (t=0).We applied DMD separately to the putamen and caudate as there might bedifferent spatio-temporal patterns in the two striatal regions and we wanted toexamine the effect of disease in each striatal substructure separately. We also appliedDMD first separately to the less and more affected sides to investigate progressionwithin a lateralized structure, then to both sides together to examine if the knowndenervation asymmetry associated with PD (especially in the early stages) wouldinfluence the spatial patterns and their temporal courses.The analysis pipeline was written in Matlab and is available upon direct requestto the corresponding author, however PET data used in this study are not madeavailable publicly for reasons of patient confidentiality.4B4BG fovustnyss und fyproduwivility of DaDFor imaging studies, where the number of subjects may be relatively limited, thepatient population may not fully sample the disease progression spectrum, resultingin missing time points on the disease time course. In our case, we did not have two88time points of subject with nine and 14 years of disease duration. We performedleave-one-out cross validation on the temporal snapshots to examine the effect ofmissing time points on the DMD outputs. In this study, the sampling rate of onewas used which corresponds to one year disease duration step.4B4B4 inivuriuty UnulysisIn addition to the multivariate analysis, we also fitted a pre-defined exponentialfunction to the average DTBZ AR values in the less and more affected putamenand in the less and more affected caudate to compare the extracted DMD temporalcurves with the models fitted to the average AR values.The following exponential equation was fitted to the putamen:n (t) = az−ut + xz−wt (4.10)While the exponential equation fitted to the caudate was of the form:n (t) = az−ut (4.11)Where a, b, c, and d are the fitting parameters, t is disease duration (in years) andY is the average AR values.We used a pre-defined function with two exponential terms for the putamenand one exponential term for the caudate to match the number of spatio-temporalpatterns extracted by DMD.We also performed correlation analyses between the temporal expressions ofDMD mode 1, DMD mode 2 and the sum of DMD mode 1 and mode 2 and theaveraged DTBZ AR values in the less and more affected putamen and in the lessand more affected caudate.4B4BI Compurison vytwyyn DaD und dCUDMD and PCA are conceptually similar in the sense that both involve decompo-sition of the input data. However, DMD models coupled temporal-spatial patternswhile PCA only models variance in the data without accounting for any temporalinformation. In other words, if the input signal is composed of two sources withdifferent temporal courses, PCA may fail to un-mix the two sources. Mathemati-cally, the first step of the DMD algorithm is the same as PCA, while the variable A89captures the temporal dynamic of the PCA mode from one time point to the next:Ut = VUt−1 (4.12)To illustrate the impact of the differences between the two approaches on theoutcomes, we applied both DMD and PCA to DTBZ AR values in the less affectedputamen and compared the spatial patterns of DMD and PCA and their associatedsubject/temporal scores as a function of disease duration.4BI fysultsDMD decomposed DTBZ PET data into two coupled sets of distinctive spatio-temporal patterns in the putamen and a single spatio-temporal pattern in the cau-date. The temporal curves were highly robust as shown by the small variations ob-tained in the leave-one-out cross validation. We kept the first two pairs of coupledDMD modes and their temporal courses that accounted for 98% of total variancein the data for the putamen and the first pair of coupled DMD modes and theirtemporal courses that accounted for at least 96% of total variance for the caudatefor further analysis. Temporal curves for later DMD modes showed mostly oscil-latory behaviour around zero, so were considered as non-meaningful. There weresignificant correlations between the averaged DTBZ AR values in the more and lessaffected putamen and caudate and expression of DMD mode 1, mode 2 and the sumof the two modes in the putamen and caudate (P Q0.01). The correlations betweenthe averaged DTBZ AR and the expression of the sum of DMD mode 1 and 2 wasthe strongest which is expected since the temporal expression of the sum of the twoDMD modes reflect the overall progression change in DTBZ binding.90azss vffzxtzy eutvmzn borz vffzxtzy eutvmzn azss vffzxtzy Cvuyvtz borz vffzxtzy Cvuyvtzboyz F YbY vmplituyz =z) 26.77±0.69 21.38±0.96 36.64±0.17 33.14±0.46Yzxv– xonstvnt -0.027±0.004 -0.028±0.005 -0.026±0.002 -0.026±0.004boyz G YbY vmplituyz =z) 19.61±0.42 7.22±2.01 N/A N/AYzxv– xonstvnt -0.65±0.05 -1.18±0.18 N/A N/Aiotvl vvrivnxz zxplvinzy =:) 97.7 97.7 96.2 97.0Table 4.2: DMD output parameters. All numbers are reported as meanstandarddeviation. DMD amplitudes (unitless) and decay constants determine the interceptand shape of the exponential temporal curves in each striatal region. Temporalcurves with more negative decay constants drop more quickly with increasing dis-ease duration compared to the temporal curves with decay constants closer to zero.Higher DMD amplitude represents higher expression of the DMD spatial pattern atdisease onset. The total percentage variance explained were calculated with the firsttwo modes for putamen and with the first mode for caudate. Standard deviation ofthe DMD parameters were calculated from leave-one-out cross-validation. DMD =dynamic mode decomposition.4BIBE dutumynIn both the less and more affected putamen, the first DMD modes showed charac-teristic anterior-posterior gradients that were almost identical between the two sides(Fig.4.3A). These anterior-posterior patterns were associated with temporal curveswith almost identical decay constants for both the less and more affected putamen(Fig.4.4A). The DMD amplitude (intercept of the temporal curves), however, wasmuch higher in the less affected putamen compared to the more affected putamen.This indicates the expression of the anterior-posterior gradient in the putamen washigher in the less affected putamen at disease onset.The second DMD mode showed a dorsal-ventral gradient in the less affectedputamen (Fig.4.3A) associated with a temporal curve that decreased sharply in theearly stage of the disease (Q5 years) (Fig.4.4B). For the more affected putamen, thesecond temporal curve neared zero after approximately one year of disease duration(Fig.4.4B). Detailed DMD output parameters are listed in Table 2.Combining the temporal curves for the first two DMD modes (Fig.4.4C), weobserved that the initial progression rates and intercepts in the less and more affectedputamen were quite different; the difference was mainly dominated by the secondDMD mode (dorsal-ventral gradient). The progression rates then became similarin the less and more affected putamen as the first DMD mode (anterior-posteriorgradient) started to dominate. The combined temporal curve also highly resembledthe exponential function fitted to the averaged AR values in the less and moreaffected putamen (Fig.4.4D).91Figure 4.3: DMDmodes (spatial patterns) in the less and more affected putamen (A)and caudate (B). DMD was applied to the less and more affected sides separately. Inthe putamen, DMD mode 1 showed an anterior-posterior gradient and DMD mode2 showed a dorsal-ventral gradient in both the less and more affected sides. DMDmode 1 in the caudate showed a head-tail gradient in both the less and more affectedsides. Spatial patterns are displayed as maximum intensity projection onto the entireregion of interest. DMD = dynamic mode decomposition. LM = lateral-medial. AP= anterior-posterior. DV = dorsal-ventral.92Figure 4.4: (A) The first DMD temporal curve in the less and more affected puta-men, associated with the anterior-posterior gradient. (B) The second DMD temporalcurve in the less and more affected putamen, associated with the dorsal-ventral gra-dient. (C) Combined DMD temporal curves for the first and second DMD modesin the less and more affected putamen. (D) Averaged DTBZ activity ratios in theless and more affected putamen versus disease duration and the best exponentialfit curve. Error bars were generated from leave-one-out cross validation. DMD =dynamic mode decomposition. DTBZ = dihydrotetrabenazine.4BIBF CuudutyThe first DMD modes showed a head-tail gradient in both the less and more affectedcaudate (Fig.4.3B). This gradient was associated with temporal curves with almostidentical decay constants in both the less and more affected caudate (Fig.4.5A).The intercept for the temporal curves was higher for the less affected caudate thanthat for the more affected caudate, however, the differences in gradient expression93Figure 4.5: (A) The first DMD temporal curve in the less and more affected caudate,is associated with the head-tail gradient. (B) Averaged DTBZ activity ratios in theless and more affected caudate versus disease duration and the best exponentialfit curve. Error bars were generated from leave-one-out cross validation. DMD=dynamic mode decomposition. DTBZ = dihydrotetrabenazine.became smaller as the two curves tended to converge in later disease stage. Again,this DMD temporal curve highly resembled the exponential function fitted to theaveraged AR values in the less and more affected caudate (Fig.4.5B).4BIBG Compurison vytwyyn DaD und dCUAs shown in Fig.4.6, the first spatial patterns (DMD mode 1 and PCA pattern1) obtained from DMD and PCA showed very similar anterior-posterior gradientin the less affected putamen; the relationship between the subject scores relatedto the PCA-defined spatial pattern and disease duration was however much lessrobust compared to the DMD temporal curve. The second spatial patterns showedslight dorsal-ventral gradient in the putamen for both DMD and PCA and verysimilar temporal expression of the patterns as a function of disease duration. Thedifferences in the spatial patterns and scores between DMD and PCA were due tothe additional temporal information embedded in the DMD algorithm. The similardistributions of DMD and PCA scores of pattern 2 suggest the second temporalcomponent extracted by DMD also accounts for approximately the second largestvariance (approximately 3.6% of variance accounted for by the second PCA pattern)in the data in this particular case.94Figure 4.6: Comparison between DMD temporal expression of mode 1 (left) andmode 2 (right) and PCA scores of PCA pattern 1 (left) and PCA pattern 2 (right)in the less affected putamen. Both DMD temporal expressions and PCA scoresare Z-transformed. Spatial patterns are displayed as maximum intensity projectiononto the entire region of interest. DMD = dynamic mode decomposition. PCA =principal component analysis4BJ DiswussionIn this work, we showed the first application of DMD to visualize and quantifydisease-induced progressive changes in neurotransmitter activities. We first de-scribed the implementation of DMD to extract spatio-temporal patterns relatedto disease progression using neuroimaging data and then tested the method on awell-established DTBZ PET dataset to model dopaminergic denervation in PD. Themethod was found to be robust with respect to typical uncertainties in the estima-95tion of disease duration and is expected to be easily applicable to a wide range ofimaging data.This work presents both methodological and clinically relevant advances. Interms of methodological novelties, we presented the first multivariate approach thatsimultaneously models spatial and temporal patterns related to neurodegenerationusing PET data. We compared the DMD temporal curves with a static multivariateapproach (PCA) that only models spatial information in the data. We demon-strated that even though PCA and DMD showed similarities in the spatial patternsand temporal expression of the patterns as a function of disease duration, DMD wasintrinsically able to more accurately extract spatial patterns with different progres-sive behaviours and thus provided a more powerful alternative dimension reductionmethod for modeling progressive changes.An important methodological advantage of the DMD approach is that the spatio-temporal patterns are derived in a purely data-driven and equation-free fashion at avoxel-level, which is especially important when the underlying disease mechanismsare unknown. When modeling temporal changes with traditional approaches, thechoice of the final model is usually based on estimating model residuals or otherobjective methods such as the Akaike information criterion. DMD, on the otherhand, automatically decomposes the data into an optimal number of exponentialfunctions with minimal need for parameter tuning.Another unique and important advantage of the DMD approach is its abilityto decompose temporal changes of tracer binding into orthogonal (loosely implyingindependent) temporal trajectories; while the traditional approach (i.e. pre-definedmodel fitted to the averaged AR values in individual striatal regions) can onlymodel the overall temporal changes by minimizing the least square error betweenthe data and best fit curve. The spatial patterns associated with each independenttemporal trajectory represent regions that change similarly as a function of diseaseduration: the method is thus able to identify regions that are differentially sensitiveto potentially different disease mechanisms.4BJBE DaD gputioAhymporul duttyrnsIn this work, we were able to decompose the overall disease progression curve inthe putamen into two independent temporal curves associated with distinct spa-tial patterns of dopaminergic denervation. The combined temporal expression ofthe anterior-posterior and dorsal-ventral gradients (Fig.4.4C) showed striking re-96semblance to the disease progression curves previously presented using univariatemeasures in the putamen regions obtained with a different patient population on adifferent scanner [51] and the univariate exponential fit curves obtained with aver-aged DTBZ AR values in the putamen (Fig.4.4D).While the initial expression of the anterior-posterior gradient was higher in theless affected putamen, the expression of this gradient decreased gradually at similarrates in the less and more affected putamen and still existed in later disease stage (≈15 years of disease duration) (Fig.4.4A). This indicates that the anterior-posteriorgradient is likely well maintained bilaterally over the disease course [51, 114]; thefact that the rate of decrease is the same on both sides provides further support tothe interpretation that this pattern is related to mechanisms responsible for diseaseprogression.The dorsal-ventral gradient in the putamen, however, existed only in very earlydisease, mainly in the less affected side. Previous studies showed that the rate ofdopaminergic neuron loss is highest in early disease and the level of asymmetrybetween the less and more affected sides is more prominent at this stage [51, 114].According to our analysis, this rapid change of dopaminergic function in early diseasemay be mainly due to the changes along the dorsal-ventral gradient. It is interestingto note that the curves related to the less affected side seem to be shifted in time byapproximately 5-8 years compared to the more affected side (by visual inspection),likely reflecting the asymmetric nature of disease onset followed by an asymmetriconset of clinical symptoms.The expression of the head-tail gradient in the caudate decreased gradually withdisease. The initial expression (intercept), however, was higher for the less affectedcaudate than the more affected caudate. This implies that even though there isan initial difference in the expression of this gradient in the less and more affectedcaudate at disease onset, there is no asymmetric progressive changes in the caudatealong this gradient. However, the caudate (especially the tail) is particularly proneto partial volume effect, which may suggest that the head-tail gradient could bea result of spill-out effect from the edge of the caudate to its surroundings; thereduction in temporal expression for this DMD mode may therefore only reflects anoverall magnitude change in tracer binding across the entire caudate rather than atrue head-tail gradient in the caudate. To partially address this issue, we comparedthe decay constants from the DMD temporal curve and the exponential model fittedto the averaged AR values in the entire caudate as a function of time. The decay97constants from the exponential fits were -0.018 and -0.017 for the more and lessaffected caudate respectively, which were larger than the decay constants for thehead-tail gradient extract by DMD, which implies the expression of the head-tailgradient decreases faster than the average tracer binding across the whole caudate.Even though this comparison does not fully address the potential confound due topartial volume effect, the differences in decay constants support that the decreaseof the gradient is disease-related rather than an imaging artifact.The decay constants for the anterior-posterior gradient in the putamen and thehead-tail gradient in the caudate were very similar, providing further support tothe interpretation that these two gradients of dopaminergic denervation may bedue to non-specific mechanisms that affect all striatal sub-regions at a similar rate.The presence of a dorsal-ventral gradient in the putamen restricted to early diseasemay be related to either an independent disease-initiating mechanism or differentialinvolvement of striatal sub-regions at different disease stages. Further studies areneeded to relate the different spatio-temporal patterns with disease mechanisms.While practically and conceptually consistent with earlier hypotheses on PD ini-tiation/progression, the unique and more general contribution of this approach tothe analysis of PET data is that the method appears to be able to separate indepen-dent temporal patterns of disease progression and identify the regional contributionsto each of such patterns. It thus provides more direct evidence for the existence ofdifferent disease-related mechanisms and an assessment of their topological char-acteristics and relevance at different stages of disease, thus providing much moredetailed information compared to traditional analysis methods.4BJBF LimitutionsThere are several limitations of this study. The first major limitation is the rela-tively small sample size used in this study. Second, DMD assumes there is a fixedtemporal resolution in the input data; we took the sampling rate (∆ t) to correspondto one year of disease duration in our data. Since the uncertainty in the determi-nation of disease duration in PD ranges from 0.5 to 1 year, we randomly shifteddisease durations of all subjects by ±1 year; the DMD outputs for the randomlyshifted data did not change appreciably. Third, cross-sectional data were used tomodel disease progression, for which individual subject variability could representa potential confounding factor. However, obtaining longitudinal data on the samesubject at short disease intervals is often not practical. While we acknowledge that98longitudinal data may provide a more accurate description of disease progression,previous studies showed very similar disease progression curves obtained with cross-sectional data alone[50, 109] and with a mixture of cross-sectional and longitudinaldata [51, 115]. The subject variabilities are partially accounted for by taking theaveraged images over subjects with the same disease duration. More importantly,DMD models the overall progression across all time points, where each time pointis represented by a different subject, so it is less sensitive to heterogeneity in diseaseprogression between subjects compared to univariate analyses. Fourth, paramet-ric AR images were used instead of BPaW images as a measure of dopaminergicdenervation; this choice was made out of convenience as these data were readilyavailable. We previously applied PCA to both parametric AR and BPaW imagesin a subset of the same subject cohort involved in this study; the resulting spatialpatterns were virtually identical [14]. In principle, DMD can be applied to differenttypes of parametric images, since it primarily works with covariance patterns ratherthan absolute voxel values. Fifth, even though striatal VMAT2 is predominantlyexpressed on dopaminergic nerve terminals, it is also expressed on other monoamin-ergic neurons. However, DMD may be less sensitive to confounds introduced bythe lack of tracer selectivity compared to traditional analyses. In principle, DTBZsourced from different nerve terminals (i.e. dopaminergic and non-dopaminergic)should follow different temporal progression curves, which could potentially be cap-tured by DMD. Furthermore, the different spatio-temporal patterns extracted byDMD may be related to the differential onset times of VMAT2 denervation for theputamen and caudate and for the less and more affected striatal sides. For exam-ple, the VMAT2 denervation occurs first in the putamen (especially in the posteriorputamen) before affecting the caudate. The absence of the second spatio-temporalpattern in the caudate may be due to either the late onset time of VMAT2 denerva-tion that is outside the dynamic range of disease duration included in this study orthat the caudate is less sensitive to disease-initiating mechanisms and therefore onlypresents a single spatio-temporal pattern related to disease progression. Similarly,the lack of the second spatio-temporal pattern (the dorsal-ventral gradient) in themore affected putamen is likely due to the earlier onset of VMAT2 denervation inthe more affected side compared to the less affected side. Ideally, we need to includedata obtained in the prodromal stage of the disease to examine the expression ofthe dorsal-ventral gradient in the more affected putamen. Similarly, for the caudate,we might need to explore data outside the range of disease durations included in99this study (0 to 15 years) to examine the presence of other spatio-temporal pat-terns. Indeed, the different spatio-temporal patterns partially reflect the onset timeof the VMAT2 denervation in different striatal sub-regions, but they also reflectthe dynamic/progressive changes throughout the range of disease examined in thisstudy. This is supported by the finding that the decay constants for the first DMDspatio-temporal patterns were the same for the putamen and caudate. As we knowthat the time of onset of VMAT2 denervation is earlier for the putamen comparedto the caudate, the similar decay curves for the two structures reflect the similarprogressive behaviours of VMAT2 denervation from 0 to 15 years of disease, ratherthan entirely reflecting the onset time of VMAT2 denervation in the two regions.4BK ConwlusionIn this work, we introduced the DMD approach to PET data and demonstrated thatthis approach is able to capture spatio-temporal patterns of dopaminergic denerva-tion in PD. This approach appears very well suited to model disease- or treatment-induced progressive changes in imaging data. This proposed method has severaladvantages over traditional methods in terms of biologically-relevant informationthat can be extracted from the data: first, it considers tracer distributions in all theselected regions at once, thus providing information not only on localized alterations,but also on spatial patterns of such alterations, emphasizing the network behaviourof the targets under investigation. Second, this approach incorporates both spatialand temporal information simultaneously in a data-driven and equation-free fash-ion. Thirdly, it allows the decomposition of overall disease-induced progression intoorthogonal temporal curves, possibly relating to independent mechanisms. In thisstudy, we were able to, for the first time, decompose the dopaminergic denerva-tion in the striatum associated with PD into two spatio-temporal patterns: (i) theanterior-posterior gradient in the putamen and head-tail gradient in the caudate,which may be related to non-specific mechanisms responsible for disease progressionand (ii) the dopaminergic denervation along the dorsal-ventral gradient in the puta-men which may reflect independent mechanisms responsible for disease initiation inthe very early stage of the disease. While the data considered in this study allowedus to validate this approach and provide some novel insights into PD progression,the method can be easily applied to data obtained with other tracers and related toother diseases.100Chuptyr IJoint duttyrn UnulysisIBE gummuryThe work described in this chapter contains published work: [uA JC[CA Kl–uzhinAICA bxKznzizA JCA czilsonA cCA hhvhinfvryA ZCA YinzllzA KCA bxKzofinAbCJCA htozsslA VCJCA hossiA kCA GEFNC Joint pvttzrn vnvl–sis vpplizy toeZi YVi vny kbViG imvging rzvzvls nzfi insights into evrkinsonsyiszvsz inyuxzy przs–nvptix vltzrvtionsC czuroImvgz ClinC GHA FEFMJKC[12]. I was responsible for development of the analysis methodology, image pre-processing, statistical analysis, clinical interpretation and manuscript composition.Scanning procedures were performed by the staff members of the UBC PET imag-ing group. K.Dinelle, E. Shahinfard and N. Vafai contributed to image preprocess-ing. J. McKenzie and N. Neilson contributed to recruitment of study participants.I. Klyuzhin and M.J. McKeown contributed to the development of the analysispipeline. A.J. Stoessl and V. Sossi designed the study and contributed to clinicalinterpretation. V. Sossi was the the supervisory author involved throughout theproject in the concept formation and manuscript preparation.In the work described in previous chapters, we focused on a single PET tracer ata time. However, the emerging of multi-tracer and multi-modality imaging studiesreally highlights the need for novel methods that can effectively incorporte informa-tion from different aspects of the brain. In this chapter, we propose the use of anovel joint pattern analysis approach to extract common and unique spatial distri-bution patterns of VMAT2 (measured by [11C]-DTBZ PET) and DAT (measuredby [11C]-MP PET) in early PD subjects in striatal and extrastriatal regions. To thebest of our knowledge, the proposed joint pattern analysis approach has never beenapplied to study the interactions between multiple neurotransmitter systems usingmulti-tracer PET imaging.Compared to the univariate approach, this approach has several unique advan-tages: firstly, it is able to consider tracer distribution in all regions of interest at101once, thus providing information not only on localized target alteration, but also onthe spatial patterns of such alterations and can thus detect relative disease-inducedalterations in different regions. The method successfully captures the known asym-metry and rostro-caudal gradient characteristic of dopaminergic denervation in earlyPD in a completely data driven manner. Secondly, the approach can decompose thecommon information in VMAT2 and DAT distributions into orthogonal (loosely in-dependent) spatial patterns of characteristic dopaminergic changes that are eithermore sensitive to disease discrimination or to disease progression. The commoninformation in VMAT2 and DAT distributions was found to correlate more signifi-cantly with disease duration compared to any univariate measures. The orthogonal-ity of the patterns may reflect different mechanisms underlying disease initiation orprogression. Thirdly, by identifying common and unique distribution patterns foreach tracer, it can identify unique behavior of each specific target and thus possiblydiscern the relative target response to disease. Finally, while in our study this anal-ysis approach shed some new light on the effects of PD on two dopaminergic targets,the results also demonstrated the power of this type of methodological approach. Itappears very well suited to the analysis of multiple data sets, including multi-traceror even PET and MRI-derived data.IBF IntroduwtionPD is the second most frequent progressive neurodegenerative disorder [33]. Themotor deficit of PD is traditionally associated with dysfunction of the nigrostriatalpathway, characterized by progressive loss of dopaminergic neurons in the substantianigra and loss of their projection fibres to the striatum [6]. Imaging studies showthat the motor deficits start to become clinically relevant when 30 to 50% of nigraldopaminergic cells are lost[116]. Neurodegeneration of the dopaminergic systemtends to follow a fairly well defined spatio-temporal pattern in which the dorsal pos-terior putamen contralateral to the more affected body side is affected first, followedby degeneration in the ventral and anterior putamen and the caudate, as shown inFig.5.1 [6]. The relatively long preclinical stage of PD, in which subjects remainasymptomatic despite significant dopaminergic neuronal loss, may be due to poten-tial compensatory effects taking place at different stages of dopamine (DA) process-ing including DA synthesis, release and turnover [50, 115, 117]. Such compensatorymechanisms, likely also involving other neurotransmitter systems [44, 46, 118], are102Figure 5.1: [11C]-dihydrotetrabenazine (DTBZ) PET image (left) and [11C]d-threo-methylphenidate (MP) PET image (right) for a Parkinson’s disease (PD) subject.PD subject showed characteristic asymmetric tracer uptake in the less and moreaffected hemispheres. PD subject also showed spatio-temporal pattern of dopamin-ergic loss with the posterior putamen (putamen 3) affected before the anterior puta-men (putamen 1) and caudate.deemed responsible for minimizing the effects of dopaminergic deficits on the clini-cal behaviour prior to onset of motor deficits and are thought to persist in the veryearly stages of disease [50].There is now established recognition that PD is not just a motor disorder; pa-tients often experience non-motor deficits alongside or even before the onset of motordeficits. Non-motor deficits may be more closely related to DA projections outsidethe nigrostriatal pathway, in addition to alterations in other neurotransmitter sys-tems such as the cholinergic and serotonergic systems [11, 44, 46]. Outside thenigrostriatal pathway, the mesocorticolimbic pathway transmits DA from the ven-tral tegmental area (VTA) to the ventral striatum (VS) and to the prefrontal cortex[119]; the tuberoinfundibular pathway transmits DA from the hypothalamus to thepituitary gland [120]. A previous study reported that deficient hypothalamic DAtransmission may play a role in autonomic and endocrine abnormalities in PD [121].103Studying DA processing in these pathways in addition to the nigrostriatal pathwaymay provide a more complete picture of dopaminergic denervation in PD, especiallyin the early disease stage or even before disease onset. Changes in the metabolic andfunctional connectivity have also been identified at various stages of PD and relatedto specific clinical manifestations such as cognitive deficit. It is thus reasonable topostulate that PD imprints a general disease-related pattern on several aspects ofbrain function, with alterations in selected systems reflecting either a specific systemdisease response or specific clinical manifestation of the disease. The ability to iden-tify such patterns and their evolution as a function of disease progression/specificclinical manifestation may aid in the understanding of disease mechanisms or sub-jects’ propensity towards a particular clinical trajectory.In this work, we propose the use of a novel joint pattern analysis to studyfunctional similarities and differences between multiple PET targets. Our analysisenhances the more traditional approach where the relationship between two or moresets of imaging data is examined using univariate approaches such as correlation andt-test. Multivariate techniques, such as PCA and ICA [63, 122], have been used todecompose individual datasets into functional networks. However, instead of ana-lyzing individual datasets (i.e., comparing functional networks obtained separatelyfrom each dataset), we used a data fusion approach to explore and identify com-mon and unique information given by each dataset as functional networks. Thesecommon and unique functional networks can provide additional and more direct in-sights into the interactions between processes observable with different tracers andthe differential information provided by each individual tracer. This approach is par-ticularly suitable for neurodegenerative diseases such as PD, where disease affectsdifferent stages of DA processing as well as multiple neurotransmitter systems.Many different joint multimodal analysis techniques have been developed in theneuroimaging field, mainly applied to MRI data. One such data fusion approachcommonly used is (joint independent component analysis (jICA))[123]. More re-cently, canonical correlation analysis (CCA) has gained popularity. Unlike jICA,CCA provides a relatively less constrained solution to the data fusion problem[75, 76]. While jICA assumes that different datasets have exactly the same inter-subject covariations, the CCA models the coherence in the inter-subject covariationsto identify associations between datasets [75, 76]. CCA has been successfully appliedto the analysis of fMRI, EEG, electromyography (EMG), structural MRI and be-havioural data in PD and schizophrenia to explore common inter-subject variations104in different datasets [75]. After extracting the common information among differentdatasets, the unique information still remaining in each individual dataset can beextracted by using orthogonal signal correction (OSC), which was first introducedas a spectral preprocessing method in spectroscopic calibrations [124, 125]. OSCwas later used together with CCA to draw unique information from EEG and EMGdata [126].As this is a first application of such methodology to PET data, we chose toperform the analysis on data obtained from two fairly well characterized presy-naptic PET tracers: [11C](+)DTBZ and [11C]MP. VMAT2 binding measured by[11C](+)DTBZ is proportional to the DA terminal density [6, 127] and is used toestimate vesicular uptake and storage of DA. DA reuptake, mediated by the mem-brane DA transporter (DAT), can be targeted in PET by [11C]MP. Both VMAT2and DAT are mainly located at the nerve terminals, but can also be found in cellbodies and axons [128, 129]. DAT was shown to contribute to maintaining rela-tively constant synaptic DA levels by removing extracellular DA [47, 48], and maybe related to compensatory mechanisms in populations at higher risk of PD [41, 49]and in early PD [50, 51]. On the other hand, reduction in VMAT2 binding isdeemed a more direct measure of dopaminergic degeneration and is less susceptibleto disease-related compensation [130]. Despite the functional differences of VMAT2and DAT, traditional analyses of PET data showed high correlation between striatalalterations in DTBZ and MP binding in PD [50, 115] and some reported no differ-ential regulation of the striatal uptake of VMAT2 and DAT [131, 132]. In addition,VMAT2 and DAT distributions in the striatal regions have been studied extensivelywith traditional univariate analysis, but investigations of their distributions outsidethe nigrostriatal pathway have been quite limited, especially with imaging studies.In this study, we first examine the applicability and robustness of the proposedjoint pattern analysis approach. We then examine the novel information providedby this approach compared to traditional univariate analysis; specially, we1. Compare the common information between VMAT2 and DAT distributionsto the results from the traditional univariate analysis to test the method’sabilities to capture the characteristic dopaminergic patterns in the striatum.This serves as the main validation of the method.2. Explore the decomposition of the common information between VMAT2 andDAT in the orthogonal spatial patterns as reflecting different/independent105underlying disease-related mechanisms.3. Interpret the unique information specific to VMAT2 or DAT distributions inlight of the specific target behaviour in the early stages of PD.IBG aythods und autyriulsIBGBE gtudy durtiwipuntscumwzr Vgz Yiszvsz yurvtion =s–mptoms) Yiszvsz yurvtion =yivgnosis) bYhBjeYgh III ]ozhn & nvhr sxvlz boCV WYI aZYeY suwjzxts 15 59±8 56±34 44±29 17±9 1.6±0.5 28.0±1.5 4.4±3.5 380±220Table 5.1: Clinical characteristics of all subjects. All numbers are reported as mean±standard deviation. * disease duration estimated as the time from onset of motorsymptoms as reported by the patients. † disease duration estimated as the time ofclinical diagnosis. PD=Parkinsons disease subjects; MDS-UPDRS=Movement Dis-order Society Unified Parkinsons Disease Rating Scale; MoCA=Montreal CognitiveAssessment; BDI=Beck Depression Inventory; LED=Levodopa equivalent dose.The study included 15 early sporadic PD subjects (9 males and 6 females). Ex-clusion criteria included clinical history of depression, active anti-depressant therapyor medication and a BMI S35. Disease duration was estimated as time from onset ofmotor symptoms as reported by the subjects. PD subjects were clinically evaluatedusing MDS-UPDRS and Hoehn and Yahr scale to assess motor dysfunction, MoCAto assess cognitive performance and BDI. Detailed clinical characteristics are listedin Table5.1. All assessments were performed off medication. The study was ap-proved by the Clinical Research Ethics Board of the University of British Columbiaand all subjects provided informed written consent.IBGBF gwunning drotowolsAll study subjects underwent (+)DTBZ and MP PET scans and a T1-weightedMRI scan of the brain. The PET scans were performed on a Siemens HRRT(Knoxville, TN) with a spatial resolution of 2O5mm3 [86]. Subjects were positionedusing external lasers aligning the gantry with the inferior orbital-external meatalline, and custom fitted thermoplastic masks were applied to minimize head move-ment. Prior to PET scans, subjects were withdrawn from all anti-parkinsonianmedications for at least 12 h. An average of 300MBq with average specific activity106of 10.194Ci/mmol of (+)DTBZ and MP were administered by intravenous injec-tion over 60 s using an infusion pump (Harvard Instruments). Tracer injectionswere separated by at least 2O5 h to allow radioactive decay. Acquired data werebinned into 16 time frames (frame durations: 4×60 s, 3×120 s, 8×300 s, 1×600 s;image dimension=256×256×207; voxel size=1O22mm3) with a total duration of60min. Transmission scans required for attenuation correction were performed over10minwith a rotating 137Cs source. PET images were reconstructed using the 3Dlist-mode OP-OSEM algorithm [87] with 16 subsets and six iterations, with cor-rections for decay, dead-time, normalization, attenuation, scattered and randomcoincidences. After reconstruction, images were smoothed with a 3mm FWHMGaussian filter to reduce noise. The frames were spatially realigned with rigid-bodytransformation to minimize the impact of motion during scans. The structural MRIscans were performed on a Philips Achieva 3T MRI scanner (Phillips Healthcare,Best, NL) using the T1 TFE sequence (TR/TE = 7.7/3O6ms; TFE shots=218; flipangle = 8◦; image dimension: 256×256×170; voxel size 1mm3).IBGBG Imugy drowyssing und UnulysisThe anatomical MRI image of each subject was first coregistered with the subject’smean PET image using SPM (Wellcome Trust Centre for Neuroimaging, UniversityCollege London). Striatal regions of interest were manually placed on an averagedPET image derived from nine consecutive image slices (slice thickness 1O22mm)spanning the axial extent of the striatum, using MRI image as guidance. Fiveelliptical ROIs were placed on the striatum bilaterally one on the caudate, one onthe VS and three covering the full length of the putamen (anterior putamen 1,middle putamen 2 and posterior putamen 3). The same set of image slices wasalso used to define the occipital cortex reference region for both tracers.In addition to the manually defined striatal ROIs, we also developed a ROItemplate in MNI space using MRI images of healthy controls. The PET-coregisteredMRI images were transformed to the MNI space. The ROI template was inverse-transformed to match each subject’s PET image for further analysis. This ROItemplate contained four ROIs placed bilaterally (substantia nigra, thalamus, globuspallidus and hypothalamus), and four individual ROIs (posterior midbrain, pons,raphe nucleus and VTA). Combined with the manually-placed striatal ROIs, thisyielded a total set of 22 ROIs. Regional time-activity curves were extracted fromeach ROI and the Logan graphical method [32] was used to calculate the BPaW107values using time ranges from 17.5 to 60min. BPaW values were then rearrangedinto the more and less affected hemispheres for all bilateral ROIs based on theaverage DTBZ BPaW values in the three putamen ROIs.IBGB4 inivuriuty UnulysisTo examine the distributions of VMAT2 and DAT density, we first performed one-sample t-test (one-tailed) on the BPaW values for DTBZ and MP separately for22 ROIs. For each tracer, ROIs with BPaW values significantly greater than zero(denoted as significant binding, pQ0.05 after correcting for multiple comparisons)were used for joint pattern analysis. Unpaired two-sample t-test was performedbetween the more and less affected striatal regions in DTBZ and MP separately toexamine disease-induced asymmetry. We then tested the correlation between DTBZand MP BPaW values in all ROIs to check the correlation strength between the twodatasets. We also tested the correlation between DTBZ and MP BPaW values inthe caudate and putamen with disease duration to check the correlation strength ofunivariate measures for tracking disease progression in the early stages of disease.False positive rates were controlled at p=0.05 using Bonferroni-Holm’s step-downprocedure [90].IBGBI Joint duttyrn UnulysisWe first applied the joint pattern analysis approach to DTBZ and MP BPaW valuesin all ROIs (input data) that had significant DTBZ and MP binding to extract thecommon (canonical variates from CCA) and unique (from OSC) subject scores andthe associated spatial binding patterns in the two datasets. We then compared thesepatterns with results obtained with the ten striatal ROIs only to indirectly examinethe contributions of the extrastriatal regions to the patterns and the robustness ofthe method. Correlation analysis was performed between the subject scores andclinical measures.The joint pattern analysis approach applied to the DTBZ (X) and MP (Y)datasets is performed as follow:1. Input data matrices X and Y have dimensions [N×M1] and [N×M2] respec-tively. M1 and M2 are the number of imaging features (in this case, they areeither the ten striatal ROIs (M1=M2=10) or the ROIs with significant bind-ing determined by univariate analysis for either DTBZ (M1) or MP dataset108Figure 5.2: Illustration of the decomposition and regression step. X and Y are thewhitened input matrices (feature by subject) of non-displaceable binding potential(WeaW) values obtained from step 2. The transformed data (canonical variates) Uand V are calculated using CCA in step 3, which contains the most highly correlatedsubject scores along each component (in this case 5). The CCA weights matrices(A and B) are the regression coefficients from least absolute shrinkage and selectionoperator (LASSO) in step 4. Xresidual and Yresidual are the regression residuals.CCA = canonical correlation analysis.109(M2)). N is the number of subjects.2. Each input data matrix (X and Y) is first demeaned (mwxmxtn and nwxmxtn)and whitened (mwhitxn and nwhitxn). The whitening transformation first decor-relates the features in each input data matrix, so that the new data dimensionsare linearly independent (orthogonal); it then transforms the covariance ma-trix into an identity matrix, which ensures the variance of the data along eachnew dimension is equal to one.mwxmxtn = ZΣZT (5.1)where E and Σ are the eigenvector and eigenvalues of mwxmxtn.mwhitxn = Σ−1P2Zmwxmxtn (5.2)This step serves two important purposes: (i) to reduce feature dimension of arank-deficient input matrix into fewer components, so thatc ≥ max(rank(mwhitxn)P rank(nwhitxn));(ii) to scale all variables to have the same variance so that each variable is as-signed equal importance in the subsequent analysis. In this case, M1 andM2 are reduced into top five components to minimize the noise content dom-inant in later components, while still maintaining at least 90% of the originalvariance.3. CCA [75] is then applied to mwhitxn and nwhitxn (both matrices now have di-mension [Nx5]). CCA identifies linear relationships between the two datasetsto determine the inter-subject covariance. It seeks two mixing matrices (W1and W2) such that each pair of canonical variates Ui and Vi (i=15) has max-imum correlation across the two datasets, while the canonical variates withineach dataset are orthogonal (Ui and Uj are uncorrelated). For i=15,maxj1ij2ixorr(mwhitxnl1iP nwhitxnl2i)Ui = mwhitxnl1iki = nwhitxnl2i(5.3)Thus, the transformed data (canonical variates U and V) contain common(maximally correlated between two datasets) subject profiles, which are com-posed of subject score of each subject (representing the subject weights for110the corresponding mixing matrix). Subject scores are in Z-score form with amean of zero and a standard deviation of one.4. least absolute shrinkage and selection operator (LASSO) [14, 133] is then ap-plied to regress the canonical variates (Ui and Vi) from the original datasetsX and Y to compute regression coefficients (CCA weights A and B) and re-gression residuals (mrxsiwutl and nrxsiwutl) as shown in Fig.5.2. Ten-fold crossvalidation was used to estimate the best lambda with the cross-validated min-imum square error, where lambda is the LASSO penalty coefficient.m = UV+mrxsiwutln = k W + nrxsiwutl(5.4)Since the residuals from step 4 (mrxsiwutl and nrxsiwutl) may contain informa-tion specific to each dataset besides noise, OSC [124] is then applied to ex-tract the largest orthogonal component from the LASSO residuals deemed torepresent tracer-specific unique information, including unique subject scores(Uuniqux), unique CCA weights (Vuniqux), and true noise (mnoisx) for eachdataset.m = UV+mrxsiwutl = mvommon +mrxsiwutlm = mvommon +muniqux +mnoisx = UV+ UuniquxVuniqux +mnoisx(5.5)5. CCA loadings are defined as the correlation coefficients between each canon-ical variate (Ui or Vi) and each column of X or Y (feature values for allsubjects). CCA loadings represent the feature/region contributions to eachpair of canonical variates and are used to construct the spatial patterns.6. To determine the significance levels of the correlation between each pair ofextracted canonical variates (Ui and Vi), a non-parametric permutation test isperformed on the original datasets X and Y with 1000 iterations to constructthe empirical null distributions of the correlation coefficients for each pairof canonical variates. The p-value of the original correlation can then becomputed as the probability of observing a value at least as extreme as theoriginal correlation in the null distributions. The correlation between thepairs of canonical variates is considered statistically significant if the p-valueis Q0.05.1117. To test the stability of the CCA weights and loadings, leave-one-out validationtest is performed to compute the error bounds of the feature contributions.CCA loadings are considered statistically significant if the correlation p-valueis Q0.05 after correcting for multiple comparison.All codes were written in Matlab and are available upon direct request to thecorresponding author, however PET data used in this study are not made availablepublicly due to patients confidentiality.IB4 fysultsIB4BE inivuriuty UnulysisFigure 5.3: Scatter plots for average DTBZ and MP WeaW values in the lessaffected putamen versus disease duration (estimated from the time of symptomsonset) in months. Both DTBZ (left) and MP (right) WeaW values correlatedsignificantly with disease duration. S15 fell outside the 95% confidence interval.WeaW = non-displaceable binding potential. DTBZ = dihydrotetrabenazine. MP= methylphenidate.DTBZ BPaW values were significantly greater than zero (pQ0.05 corrected) inall 22 ROIs, while MP BPaW values were not significantly greater than zero in hy-pothalamus, posterior midbrain, pons, VTA and raphe nucleus (pS0.05 corrected).Therefore, all 22 ROIs were included for DTBZ and 16 ROIs were included for MPin the joint pattern analysis. Detailed results from univariate analysis are includedin the Supplementary Materials.112One subject (S15) appeared as outlier (fell outside the 95% confidence interval)when correlating BPaW values with disease duration (Fig.5.3). This subject had adisease duration of 23months, but had the highest BPaW values in all striatal regionsfor both DTBZ and MP (BPaW values were more than two standard deviationshigher compared to average BPaW values in all subjects in most striatal regions).Without this subject, correlations between disease duration and average DTBZ andMP BPaW values in the less affected putamen were stronger (g2=0.70, pQ0.001 forDTBZ; g2=0.45, pQ0.01 for MP). In order to find the best dopaminergic patternsrelated to disease, we first excluded this subject in the joint pattern analysis, thenincluded this subject in to examine the effect of this outlier on the results.IB4BF Joint duttyrn Unulysisevirs of xvnonixvl vvrivtzs 1 2 3 4 5Corrzlvtion g2 0.98 0.90 0.85 0.47 0.28ezrmutvtion pBvvluz 0.048* 0.033* 0.001* 0.123 0.054Table 5.2: Table 2: Correlation strength g2 and significance between each pair ofcanonical variates. *=significant at p-value = 0.05113Figure 5.4: Common spatial patterns along the first three pairs of canonical variatesfor DTBZ and MP. Stars indicate the ROIs with significant CCA loadings. ROI =region of interest; CCA = canonical correlation analysis; GP = globus pallidus; VS= ventral striatum; SN=substantia nigra; VTA = ventral tegmental area. DTBZ =dihydrotetrabenazine. MP = methylphenidate.114Common information was obtained using BPaW values in 22 ROIs for DTBZ andthe 16 ROIs for MP that exhibited significant tracer binding. For both DTBZ andMP, the top five whitened components together accounted for 91% of the variancein the original datasets. Each of the five whitened components accounted for atleast 7% of the variance. The top three pairs of canonical variates were significantlycorrelated between DTBZ and MP after permutation test (pQ0.05). The fourthand fifth pairs of canonical variates did not show high correlation across datasets(R2Q0.5) and were not significant after permutation tests (Table5.2), therefore arenot discussed in later sections.The DTBZ pattern along the first pair of canonical variates (g2=0.98 betweensubject scores in the two datasets) showed significant negative loadings in the moreaffected striatal regions (caudate and putamen), and significant positive loadings inthe substantia nigra, hypothalamus and pons (pQ0.01); the MP pattern along thiscanonical variate showed significant negative loadings in the more affected striatalregions (caudate, anterior putamen (putamen 1) and middle putamen (putamen 2)),and significant positive loadings in the less affected substantia nigra and thalamus(pQ0.05) (Fig.5.4A). Along the second pair of canonical variates (g2=0.90 betweensubject scores in the two datasets), the DTBZ pattern showed significant positiveloadings in the less affected caudate, VS and VTA, and significant negative loadingsin the thalamus and globus pallidus (pQ0.05); the MP pattern showed significantpositive loadings in the caudate and VS (Fig.5.4B). Subject scores along the firstand second pairs of canonical variates did not correlate with any clinical measures.115Figure 5.5: Correlation between subject scores and disease duration as estimatedfrom the time of symptoms onset (months) for DTBZ and MP along the third pairof canonical variates. DTBZ = dihydrotetrabenazine. MP = methylphenidate.The spatial patterns for both DTBZ and MP along the third pair of canonicalvariates included significant negative loadings in the less affected caudate and puta-men (pQ0.01) (Fig.5.4C). The subject scores along the third pair of canonical vari-ates (g2=0.85 between subject scores in two datasets) correlated significantly withdisease duration for both DTBZ (g2=0.70, pQ0.001) and MP (g2=0.51, pQ0.01)(Fig.5.5). Correlations with disease duration remained significant without the sub-ject with longest disease duration (132months) for both DTBZ (pQ0.001) and MP(pQ0.05).116Figure 5.6: Unique spatial patterns along the first three pairs of canonical variatesfor DTBZ (top) and MP (bottom). Stars indicate the ROIs with significant CCAloadings. ROI = region of interest; CCA = canonical correlation analysis; GP =globus pallidus; SN = substantia nigra; VS = ventral striatum; VTA = ventraltegmental area. DTBZ = dihydrotetrabenazine. MP = methylphenidate.The unique DTBZ pattern highlighted the asymmetry between the less and more117affected striatal regions with significant negative loadings in the more affected stri-atal regions (caudate, putamen and VS), globus pallidus and VTA, and significantpositive loadings in the pons (Fig.5.6top). The unique MP pattern showed signifi-cant positive loadings in the less affected posterior putamen (putamen 3), substantianigra and thalamus, and significant negative loadings in the more affected anterior(putamen 1) and middle putamen (putamen 2) (Fig.5.6bottom). The unique subjectscores for DTBZ and MP patterns did not correlate with any clinical measures.Including S15, the correlations between disease durations and the subject scoresalong the third canonical pairs were weaker but still significant for both DTBZand MP (pQ0.01 for DTBZ and pQ0.05 for MP). Regions with significant contribu-tions to the common and unique spatial patterns remained similar. The correlationstrength between the common information in DTBZ and MP along the third canon-ical pair and disease duration was weaker when only ten striatal ROIs were includedin the analysis (with ten striatal ROIs only, g2=0.63 and pQ0.001 for DTBZ andg2=0.31 and pQ0.05 for MP). The common and unique spatial patterns for bothDTBZ and MP however remained the same for the striatal regions as they mani-fested when all ROIs were included in the analysis.IBI DiswussionWith the univariate analysis, we found all examined ROIs in this study showedDTBZ BPaW values significantly greater than zero, while 16 out of 22 ROIs showedMP BPaW values significantly greater than zero. The joint pattern analysis de-composed the characteristic gradients of dopaminergic loss in the striatum into or-thogonal components ranked by the degree of commonality shared between VMAT2and DAT distributions: 1) disease-induced asymmetry between the less and moreaffected dorsal striatum; 2) disease-induced gradient with caudate and ventral stria-tum being relatively spared compared to putamen; 3) progressive loss in the lessaffected striatum, which correlated significantly with disease duration. The uniqueinformation revealed differences between VMAT2 and DAT distributions.IBIBE inivuriuty UnulysisOur results from univariate analysis, performed primarily to serve as reference for theoutcomes of the spatial pattern approach, agree with widely reported findings on thedopaminergic deficit distribution in the striatum [51, 128, 134] and are therefore in118keeping with existing knowledge about the disease. While not of primary relevanceto this study, the results obtained from the extrastriatal regions represent somenovel findings of interest.DTBZ BPaW values were significantly greater than zero in regions involved inthe nigrostriatal and mesocorticolimbic pathways and brain stem regions in earlyPD. Previous in-vivo 18F-FP-(+)-DTBZ PET imaging study in healthy controls[135] showed highest VMAT2 level in striatal regions and substantia nigra, followedby regions involved in the mesolimbic pathway, brain stem regions and thalamus.In healthy brains, VMAT2 level in the substantia nigra, hypothalamus and raphenucleus is approximately 40% of those in the anterior putamen, and VMAT2 levelin the posterior putamen is approximately the same as in the anterior putamen[135]. In our case of early PD, DTBZ binding in the substantia nigra and raphenucleus were also approximately 40% of those in the less affected anterior putamen,but higher than that in the posterior putamen. DTBZ binding in the hypothala-mus and thalamus are approximately 45% and 6% of the binding estimated in theanterior putamen in normal brains [135], while 60% and 28% was observed in earlyPD. The hypothalamus and thalamus seem to have better preserved dopaminergicintegrity compared to the anterior putamen; however, since VMAT2 is expressedby all monoamine neurons, more preserved DTBZ binding may be also reflective ofnoradrenergic instead of dopaminergic innervation.DAT distribution was less widely spread outside the striatal regions compared toVMAT2. We observed asymmetric MP BPaW values significantly greater than zeroin the striatum and substantia nigra in early PD, which agrees with previous imag-ing finding of an asymmetric reduction of DAT level in the same regions in early PD[128]. We also observed significant MP binding in the thalamus and globus pallidusand insignificant binding in the hypothalamus and brain stem regions; these imag-ing results are consistent with previous post-mortem immunohistochemical stud-ies which showed significant DAT expression in the thalamus (healthy human andnon-human primates) and globus pallidus (healthy human) [128, 136, 137], and nodetectable DAT level in the hypothalamus [138] and brain stem regions [139].119IBIBF Joint duttyrn UnulysisCommon InformvtionDTBZ and MP showed highly correlated subject scores along the first three pairsof canonical variates, corresponding to three distinct orthogonal spatial patterns.Of these three, only the subjects scores along the third pair correlated with diseaseprogression.Spatial pattern along the first pair of canonical variates showed the familiarearly disease-induced asymmetry between the less and more affected striatum inboth VMAT2 and DAT distributions as shown in Fig.5.4A. In univariate analy-sis, differences between the less and more affected striatal regions in either DTBZor MP binding were not significant after correction for multiple comparison. Pat-tern analysis accurately captured this characteristic asymmetric tracer reductionindependently of the number of regions involved in the analysis, indicating supe-rior robustness of this approach. Higher asymmetry in the dorsal striatum wasalso associated with more preserved binding in the substantia nigra, hypothalamusand pons for DTBZ and substantia nigra and thalamus for MP, consistent withthe fact that asymmetry in the dorsal striatum appears highest at clinical diseaseonset and decreases over time [51]. This particular pattern may be characteristicof disease presence in this range of disease duration rather than progression. Inaddition, the relatively higher preservation of dopaminergic function in the substan-tia nigra in early disease appears consistent with recent imaging finding showinggreater DAT loss at the axonal terminals compared to cell bodies in early PD [128]and may provide support for the hypothesis of an early involvement of synapses andpre-terminal axons in the neurodegenerative process followed by alterations of cellbodies [140, 141].Spatial patterns along the second pair of canonical variates reflected the well-known disease-induced rostro-caudal gradient in early disease (Fig.5.1). This gradi-ent did not correlate with disease duration, suggesting that it may be predominantlya characteristic evolving in the preclinical stage where there is still relative preser-vation of the terminal density/dopaminergic function in the caudate and VS. Thisbinding pattern in striatum was further associated with relatively decreased DTBZbinding in the thalamus and globus pallidus compared to the MP pattern wherethese regions showed no significant contributions.Spatial patterns along the third pair of canonical variates reflected the progres-120sive loss of dopaminergic function in the less affected striatum for both DTBZ andMP. This is consistent with previous findings [51] which indicate that even thoughdopaminergic tracer uptake in the more affected striatum may be more sensitive fordisease discrimination, tracer uptake in the less affected striatum provides a bettermarker to track disease progression. The inclusion of extrastriatal regions into thepattern analysis increased the pattern’s correlation strengths with disease durationsfor both tracers compared to using the striatal regions alone; this indicates thatdopaminergic denervation in other regions is also affected by disease progression inspite of the fact that the regional loadings in extrastriatal regions were by themselvesnot significant. The correlation of the MP pattern with disease duration was foundto be stronger than what observed with univariate analysis, and the correlationstrength of the DTBZ pattern with disease duration was similar to the univariateanalysis applied to the averaged less affected putamen without correction for multi-ple comparison. However, the pattern analysis results did not suffer from multiplecomparison problem that may decrease the statistical robustness of the outcomesobtained with the univariate analysis.In addition, the orthogonality (which loosely implies independency) of the threespatial patterns may imply that disease-induced asymmetry, disease-induced gradi-ents and denervation progression might be underlined by different or independentmechanisms. In a previous study [51] the asymmetry between the less and moreaffected striatal sides was shown to decrease as disease progresses; the first andthird common patterns both showed striatal asymmetry, but with higher loadingson either the more affected (first pattern) or less affected (third pattern) sides. Theorthogonality between the two patterns may reflect the fact that different mecha-nisms may be of most relative relevance to the less and more affected striatal sidesrelated to the fact that degeneration in each side occurs at different stages of disease;however, whether this difference is due to saturated dopaminergic loss in the moreaffected striatum or different underlying mechanisms still need further investigation.The same previous study [51] also showed that there is a marked rostro-caudal gradi-ent of dopaminergic deficit at disease onset. We observed that the striatal gradient isindependent of the striatal asymmetry, suggesting the two aspects of dopaminergicdenervation may be induced by different mechanisms underlying disease initiationor progression. Interestingly, this decomposition may thus also provide guidance todetermine metrics that are either more sensitive to disease discrimination (first twopairs of canonical variates) or are better suited to track disease progression (third121pair of canonical variates). Inclusion of data from healthy controls in the analysismay help to further explore this hypothesis.Overall, the common information in DTBZ and MP binding confirms the ap-plicability and robustness of the proposed joint pattern analysis approach, and wasshown to be more sensitive to specific spatio-temporal changes compared to univari-ate analysis.jniquz InformvtionWhile it can be assumed that common information mainly reflects characteristicdisease-induced alterations related to the integrity of dopaminergic function, uniqueinformation reflects patterns in which VMAT2 and DAT are differently affectedin early disease. Unique DTBZ pattern showed asymmetry between the less andmore affected striatum, similar to the common DTBZ pattern along the first pair ofcanonical variates, suggesting that VMAT2 may be more sensitive to direct diseaseeffects, i.e. dopaminergic terminal degeneration, compared to DAT. Indeed VMAT2density is deemed to be least sensitive to disease-induced regulatory changes [130].The globus pallidus and VTA appeared relatively more affected and the pons weremore preserved.DAT unique pattern showed more reduced tracer binding in the more affectedanterior and middle putamen, with relatively preserved tracer binding in the lessaffected posterior putamen. The relatively more preserved DAT in the posteriorputamen might be a compensatory response to the more severe dopaminergic lossobserved with VMAT2. While it is still debatable whether lower DAT contributesto higher synaptic DA levels by reducing DA reuptake, it has also been shownthat, in PD, higher DAT is associate with lower DA turnover, i.e. the functionalrole of DAT may be to maintain relatively constant synaptic DA levels [48]. Thesubstantia nigra and thalamus also appeared relatively more preserved with MP,similar to the common MP pattern along the first pair of canonical variates. DATis found along the projections from the substantia nigra to the thalamus and thento the striatum. The globus pallidus receives input from the thalamus, and thefact that this region appeared relatively more affected for DTBZ compared to MPmay again suggest a possible compensatory role effect of DAT in the substantia-thalamus-globus pallidus/striatum pathway in early disease. The downregulation ofVMAT2 in VTA may be a reflective of the fact that the mesolimbic pathway is alsoaffected relatively early in PD [142]. However, binding in extrastriatal regions may122not be specific to dopaminergic neurons and more accurate interpretations of thefunctional roles of the extrastriatal regions in these spatial patterns require moredetailed studies involving healthy controls.IBIBG LimitutionsThere are several limitations in this study. First of all, in order to unambiguouslydetermine if patterns are related to disease or normal topology differences in thetwo tracers, the same analysis should be further extended to DTBZ and MP datafrom healthy controls. However, since all common patterns in both DTBZ andMP highly resembled known characteristic disease-related dopaminergic changes,we believe these patterns are indeed related to disease. As an indirect comparisonof spatial patterns in the disease and healthy stages, we applied PCA to DTBZ datain PD and healthy control groups then compared the PCA patterns obtained in thePD group and healthy control group with the CCA patterns presented in the paper.Secondly, both tracers are not entirely selective for dopaminergic neurons (DTBZis taken up by monoaminergic terminals and MP is not 100% specific for DAT).Although such contribution is very small in the striatum in healthy condition, it maynot be entirely negligible in PD or in other regions. While these observations mayintroduce a potential confound in the interpretation of the data, such confounds arenot specific to this analysis method but to any approach comparing MP and DTBZdata. In addition, we applied the proposed method to ROI BPaW values instead ofparametric BPaW maps to reduce the effect of noise. Another important limitationof the study is the relatively small sample size. Results from this study will befurther confirmed with larger sample size and parametric images in the future.IBJ ConwlusionUsing two extensively used tracers, we showed that the proposed joint pattern anal-ysis approach was able to capture all disease-induced characteristic spatial and tem-poral distribution patterns with better sensitivity compared to univariate analysis.This approach can be easily extended to the analysis of a larger number of data setsand thus appears very well suited to the analysis of multiple data sets, multi-traceror multi-modality. It can be further extended to include voxel-level data. Themethod has several advantages in terms of biologically-relevant information thatcan be extracted from the data: first, it considers tracer distributions in all ROIs123at once, thus providing information not only on localized alterations, but also onspatial patterns of such alterations, emphasizing an network behavior of the targetsunder investigation. Secondly, the approach decomposes the common informationbetween data sets, in our case DTBZ and MP binding, into distinct orthogonal pat-terns of characteristic dopaminergic changes that are either more sensitive to diseasediscrimination or to disease progression and potentially resulting from somewhat in-dependent underlying mechanisms. Thirdly, it allowed to identify unique behavior ofeach specific target and thus possibly discern the relative target response to disease.While the data considered in this study allowed to validate this approach, applica-tion of this method to a larger data set, including healthy controls and/or patientswith more advanced disease and/or other tracers, is expected to provide new in-sights into the effect of disease on multiple targets, their interaction and behavioras a function of disease progression in an entirely data driven manner. Extensionof the method to voxel level data and other atypical parkinsonisms might also be ofinterest in the future.124Chuptyr JJoint duttyrn Unulysis {UppliwutionJBE gummuryA version of work described in this chapter has been submitted to a peer-reviewedjournal. I was responsible for development of the analysis methodology, image pre-processing, statistical analysis, clinical interpretation and manuscript composition.Scanning procedures were performed by the staff members of the UBC PET imag-ing group. M. Matarazzo, A.J.Stoessl and V.Sossi contributed to clinical interpre-tation. N. Vafai contributed to image pre-processing. J. McKenzie and N. Neilsoncontributed to recruitment of study participants. M.J. McKeown contributed tothe development of the analysis pipeline. A.C. Felicio, A.J. Stoessl and V. Sossidesigned the study. V. Sossi was the the supervisory author involved throughoutthe project in the concept formation and manuscript preparation.In the work described in this chapter, we extended the joint pattern analysisapproach introduced in Chapter 5 to include more PET tracers where the com-mon behaviors were not known. In particular, we performed a multi-tracer PETstudy using [11C](+)DTBZ (VMAT2 marker), [11C]MP (DAT marker), [11C]DASB(SERT marker) and [11C]RAC (D2 marker) to determine the relationships betweendopaminergic and serotonergic denervation and dopamine release estimated 1 h af-ter an oral administration of standard-release 250/25mg of levodopa/carbidopa,and their relationships with motor response to treatment in 18 early (Q6yr diseaseduration) PD subjects with stable response to medication.With the novel joint pattern analysis approach, we isolated a component ofdopamine release1h after levodopa intake that is conceptually related to increasingdopamine turnover in the presence of dopaminergic and serotonergic denervation.Importantly, we showed that the serotonergic system contributed significantly tothis specific component of dopamine release already in early Parkinsons disease.125Such relationships between dopamine release and serotonergic function was not ob-served with traditional univariate analyses and represent a unique finding obtainedwith this novel analysis method. In addition, we showed, for the first time, thatsuch relatively high and rapid dopamine release (highly likely abnormal accordingto previous findings), correlated with poorer motor response 2hr after levodopa in-take; which suggests this component of dopamine release does not translate intosustainable therapeutic efficacy even in early disease. We also showed that thisturunover-related dopamine release at baseline was higher in patients who developedmotor complications compared to patients who remained stable to medication threeyears after the baseline imaging scans. We hypothesize that this turnover-relateddopamine release may selectively contribute to higher risk of motor complicationsin later disease stage.JBF IntroduwtionThe most common clinical cardinal features of PD are motor deficits such as pro-gressive tremor, rigidity and bradykinesia. The motor deficits are traditionally as-sociated with progressive degeneration of the dopaminergic neurons originating inthe substantia nigra resulting in greatly reduced synaptic dopamine levels in thestriatum and impaired neurotransmission. An increasing number of studies nowpresent evidence of impairments in non-motor function, possibly relating to the de-generation of serotonergic, noradrenergic and cholinergic pathways [11, 46, 143, 144].There is currently no cure for the disease and treatment is predominantly based onenhancing dopaminergic neurotransmission through dopamine replacement therapyeither with the administration of levodopa or dopamine agonists. While generallyconsidered to be effective initially, many patients experience mixed responses to lev-odopa treatment [145, 146]. Besides, levodopa treatment often induces aggravatingcomplications within 5-10 years, including motor fluctuations and/or dyskinesias[147]. The exact mechanisms responsible for such variability in treatment responsesand the occurrence of complications are still not fully understood.We have previously shown that abnormal levodopa-induced changes in synapticdopamine levels (referred to as dopamine release) precede motor complications [55,148]. As the dopaminergic deficit increases with disease progression, the capacity ofthe nigrostriatal dopamine system to synthesize and store dopamine from exogenouslevodopa also diminishes; however, denervation in the dopaminergic system alone126cannot fully explain the occurrence of motor complications. It was suggested thatmotor fluctuations and dyskinesias may be related to elevated dopamine turnoveror an abnormal pattern of dopamine release, which leads to increasingly greaterswings in synaptic dopamine levels after levodopa administration. This increase indopamine turnover was also hypothesized to be a possible compensatory responsein early PD by increasing synaptic dopamine concentration [117, 149].In addition to dopaminergic deficit, serotonergic innervation leads to dysreg-ulated levodopa-induced dopamine release through the false neurotransmitter hy-pothesis [58]. The serotonergic terminals are also capable of synthesizing and storingdopamine derived from exogenous levodopa but lack the appropriate regulatory au-toreceptors and transporters that control the release and reuptake of dopamine.Animal studies showed that, in the presence of dopaminergic deficit, dysregulateddopamine release from the serotonergic terminals in the striatum leads to largeswings in the synaptic dopamine levels, which in turn, trigger dyskinesia throughpost-synaptic mechanisms [58, 150, 151].The relationship between serotonergic innervation, the degree of dopaminergicdeficit, and altered patterns of dopamine release and their contributions to responseto levodopa treatment and risk of motor complications in human studies are stilllargely unknown. It was shown that in advanced disease, processing of levodopa byserotonergic neurons may contribute to exacerbation of abnormal dopamine releasefollowing levodopa administration, which contributes to a higher risk of treatment-induced motor complications [59]. The contribution of the serotonergic system todopamine release in early disease, when there is still relative preservation of thedopaminergic system, is however not known. In addition, it is not known if thedopamine released by the serotonergic system provides any therapeutic benefit.To explore these questions, we used multi-tracer PET with [11C](+)DTBZ (tomeasure the degree of dopaminergic deficit), [11C]-MP (to measure dopamine re-uptake), [11C]-DASB (to measure serotonergic function) and double [11C]-RAC (toestimate dopamine release in response to levodopa intake). To effectively examinethe relationships between multiple imaging targets and their contributions to clinicaloutcomes, we used a novel joint multimodal analysis approach [12] to explore corre-lated information between spatial patterns of serotonergic innervation, dopaminergicdeficit and levodopa-induced dopamine release, and their relationships to motor re-sponses to levodopa treatment and future risk of dyskinesia.127JBG autyriuls und aythodsJBGBE gtudy durtiwipuntsThe study included 18 early (disease duration Qsix years) sporadic Parkinsons dis-ease subjects (11 males and seven females) recruited from year 2013 to 2019. All sub-jects had a stable response to medication entering the study (no motor or non-motorcolications with a maximum of four medication administration/day). Exclusion cri-teria included active anti-depressant therapy or medication and a BMI greater than35. Disease duration was estimated as time from onset of motor symptoms as re-ported by the subjects. All subjects were clinically evaluated using MDS-UPDRSto assess motor dysfunction, MoCA to assess cognitive performance and BDI. Allassessments were performed off medication. Motor response to levodopa treatmentwas measured as (MDS-UPDRS off levodopa −MDS-UPDRS on levodopa) 2 h afteran oral dose of levodopa/carbidopa (250/25mg, standard release). All subjects havebeen followed clinically to monitor for the occurrence of motor complications: thosesubjects that developed dyskinesia (with a minimum of four levodopa doses/day)within three years were a-posteriori considered to have been at higher risk at base-line, i.e. when the imaging study was performed. The study was approved bythe Clinical Research Ethics Board of the University of British Columbia and allsubjects provided informed written consent.JBGBF gwunning drotowolsAll subjects underwent five PET scans and an anatomical T1-weighted MRI scan ofthe brain. More specifically, we used PET scans with: 1) [11C](+)DTBZ, a VMAT2marker, to measure dopaminergic integrity; 2) [11C]-MP, a dopamine transporter(DAT) marker, to measure DA reuptake capacity; 3) [11C]-DASB, a serotonergictransporter marker, to estimate serotonergic integrity; 4) double [11C]-RAC, postsy-naptic D2 receptor marker, at baseline and 1 h after oral dose of levodopa/carbidopa(250/25mg, standard release) to estimate levodopa-induced dopamine release ex-pressed as a percentage change in RAC binding between the two scans. An exampleof the PET and MRI images is shown in Fig.6.1.The PET scans were performed on a Siemens HRRT (Knoxville, TN) with aspatial resolution of 2O5mm3 [86]. Subjects were positioned using external lasersaligning the gantry with the inferior orbital-external meatal line, and custom-fitted128Figure 6.1: MRI and four different images (averaged concentration) fromPET scans done on an example Parkinsons disease subject. DTBZ =Dihydrotetrabenazine. MP = Methylphenidate. DASB = 3-amino-4-(2-dimethylaminomethylphenylsulfanyl)-benzonitrile. RAC = Raclopride.thermoplastic masks were applied to minimize head movement. Prior to PET scans,subjects were withdrawn from all anti-parkinsonian medications for at least 12 h. Anaverage of 300MBq of DTBZ, MP and RAC, and 558MBq of DASB were admin-istered by intravenous injection over 60 s using an infusion pump (Harvard Instru-ments). Tracer injections were separated by at least 2O5 h to allow for radioactivedecay.Acquired data (image dimension: 256x256x207; voxel size: (1O22mm3) werebinned into 16 time frames (frame durations: 4×60 s, 3×120 s, 8×300 s, 1×600 s)with a total duration of 60min for DTBZ, MP and RAC scans, and 18 time frames(frame durations: 4×60 s, 3×120 s, 8×300 s, 3×600 s) for a total duration of 80minfor DASB scan. Transmission scans required for attenuation correction were per-formed over ten minutes with a rotating 137Cs source. PET images were recon-structed using the 3D list-mode ordinary OP-OSEM algorithm [87] with 16 subsetsand six iterations, with corrections for decay, dead-time, normalization, attenuation,scattered and random coincidences. After reconstruction, images were smoothedwith a 2mm FWHM Gaussian filter to reduce noise. The frames were spatiallyrealigned with rigid-body transformation to minimize the impact of motion duringscans.The structural MRI scans were performed on a Philips Achieva 3T MRI scanner(Phillips Healthcare, Best, NL) using the T1 TFE sequence (TR/TE = 7.7/3O6ms;TFE shots=218; flip angle = 8◦; image dimension: 256×256×170; voxel size 1mm3).129JBGBG Imugy drowyssing und UnulysisThe anatomical MRI image of each subject was first coregistered with the subjectsPET image averaged over all frames, using statistical parametric mapping (SPM12,Wellcome Trust Centre for Neuroimaging, University College London) running onMatlab 9.0 (Mathworks Inc., Natick, MA) and MEDx (Sensor Systems, Sterling,VA). For dopaminergic tracers, striatal ROI were manually placed by an experiencedtechnician, who was blind to the analysis outcome, on the averaged PET imagederived from nine consecutive image slices (slice thickness 1O22mm) spanning theaxial extent of the striatum, using MRI image as guidance. Four elliptical ROIswere placed on the striatum bilaterally one on the caudate and three covering thefull length of the putamen (anterior, middle and posterior). The same set of imageslices was also used to define the occipital cortex reference region for DTBZ and MP.The cerebellum was used as the reference region for RAC. DASB data, as processedfor a previous study were used [11]. Briefly, the PET-coregistered MRI images werewarped into the MNI space and their inverse deformation fields were saved. Theinverse deformation field vectors were applied to ROI templates predefined in theMNI space to bring them into the individual subject’s PET space in a single step.The quality of each processing step was visually checked for all scans. The ROItemplate also included four ROIs (one on the caudate and three on the putamen)placed on the striatum bilaterally and cerebellum as reference region. We comparedthe binding values in the manually placed ROI and MNI template-based ROI forDTBZ data, and the binding values correlated significantly in the two sets of ROIs(P Q10−6).Regional time-activity curves were extracted from each ROI and the BPaWvalues were calculated with the Logan graphical method [152] for the dopaminergictracers and the SRTM [31] for DASB. BPaW values were rearranged according tothe more and less affected hemispheres for all bilateral ROIs based on the averageDTBZ BPaW values in the three putamen ROIs. We then used the percentagechange in RAC BPaW as an estimate for levodopa-induced dopamine release 1 h:Yopaminzgzlzasz1h = 100%WeeTVutsxlinx −WeeTVtftxrLWWeeTVutsxlinx(6.1)130JBGB4 gtutistiwul Unulysisjnivvrivtz Vnvl–sisWe first used one-sample T-test to examine if the amount of dopamine release1hwas significantly greater than zero in each of the eight striatal regions. To test ifthere was a significant anterior-posterior gradient in the dopamine release1h in theputamen, we applied one-way ANOVA on the dopamine release1h values in the moreand less affected putamen regions separately. We applied two-way ANOVA on theaverage dopamine release1h values in the more and less affected putamen to test ifthere was a significant asymmetry between the two sides.To test the clinical correlations of the imaging measures in each striatal region,we performed forward stepwise multiple regression analysis with motor responseto levodopa as independent variable and dopamine release1h, DTBZ, MP, DASBbinding values, age of disease onset, disease duration, levodopa equivalent dose andMDS-UPDRS off as explanatory variables in each striatal region. We also performedtwo-sample T-test to examine if there was a significant separation in dopaminerelease1h in the averaged putamen between the subjects who developed dyskinesiasand subjects who remained stable three years after baseline scans. A p-value lessthan 0.05 was considered statistically significant. P-values were not corrected formultiple comparison.Joint evttzrn Vnvl–sisExamining the interaction between more than two imaging targets can be challeng-ing with traditional univariate analyses, especially with a relatively small samplesize. We applied a novel joint pattern analysis approach, comprised MCCA, to ex-tract information from all imaging datasets (i.e. all tracers) simultaneously. Theanalysis pipeline has been described in detail elsewhere [12]. Briefly, MCCA de-composes the input PET datasets (in this case, the four input datasets comprisedDTBZ, MP, DASB binding values and dopamine release values in eight striatalregions for 18 subjects) into sets of canonical variates. Each set of canonical vari-ates contains spatial patterns (comprised positive or negative weight for each ROI),reflecting tracer binding or dopamine release changes in different striatal regions(Fig.6.2). These patterns are identified under the constraint that their expressionsare highly correlated between subjects: i.e. along each set of canonical variates, theexpression strengths of each spatial pattern (subject scores) are highly correlated131Figure 6.2: Schematic diagram for the joint pattern analysis pipeline. Four setsof input data included the binding potential values (WeaW) for DTBZ, MP andDASB and dopamine release values in eight striatal regions. The joint patternanalysis was applied onto all tracer datasets. The outputs of the joint patternanalysis are defined by highly correlated subject scores across all tracers for eachset of canonical variates and the associated spatial patterns for each tracer for eachset of canonical variates. MCCA = Multi-set canonical correlation analysis. PD= Parkinsons disease. VMAT2 = vesicular monoamine transporter 2. DAT =dopamine transporter. SERT = serotonin transporter. DA = dopamine release.across datasets, meaning the spatial patterns along each canonical variate reflectinformation which is common in all input datasets (i.e. to all tracers). The spa-tial patterns belonging to different sets of canonical variates are orthogonal (looselyindependent) to each other and may thus be associated with different underlyingsources of variation in the data.We used this joint pattern analysis to extract common spatial patterns of dopamin-ergic innervation (DTBZ), dopamine transporter binding (MP), serotonin trans-porter binding (DASB) and levodopa-induced dopamine release1h along each setof canonical variates. The spatial patterns along each canonical variate thus re-flect the relationships between variations in dopamine release, dopaminergic andserotonergic innervation; spatial patterns among different canonical variates reflect132potentially independent underlying mechanisms relating the dopaminergic and sero-tonergic function and dopamine release. To examine the relationship between thesubjects expression of the spatial patterns and treatment response, we performedforward stepwise regression with motor response to levodopa as independent variableand the subject scores for the DTBZ, MP, DASB and dopamine release1h spatialpatterns along each set of canonical variates, age of disease onset, disease duration,levodopa equivalent dose, and MDS-UPDRS off as explanatory variables.Finally, to examine if the extracted common spatial patterns could predict futurerisk of dyskinesia, we calculated the total projection score for each subject as thesum of the subject scores for each pattern along each set of canonical variates. Theprojection scores represent how much a subject expresses the overall common spatialpatterns along each set of canonical variates. We then performed two-sample T-testto examine if there was a significant group separation between the subjects whowere still stable and subjects who developed dyskinesias within three years afterbaseline scans.We then repeated the entire analysis without the inclusion of the DASB data(i.e. with and without imposing a correlation between regional variations in thedopaminergic and serotonergic systems and dopamine release) to selectively inves-tigate the contribution from the serotonergic system to dopamine release1h and therelated response to treatment.JB4 fysultsJB4BE Cliniwul FollowAup15 out of the total 18 Parkinsons disease patients completed the three years clinicalfollow-up after the baseline scans. In these 15 patients, six developed mild to severedyskinesias. For the three subjects who did not yet reach the three year mark ,clinical follow-ups are currently available at two years, one year and eight monthsafter baselines. At the time of examination all three subjects were stable. Therewas no significant difference in age of onset, disease duration, levodopa duration andlevodopa doses at the time of baseline scans between subjects who remained stableand those that experienced dyskinesias at the clinical follow-up.133Figure 6.3: Percentage dopamine release estimated 1 h after levodopa intake in theeight striatal regions. * indicates P-value Q0.05. ** indicates P-value Q0.01. P1 =anterior putamen. P2 = middle putamen. P3 = posterior putamen.JB4BF inivuriuty UnulysisAs shown in Fig.6.3, there was significant dopamine release1h in the middle andposterior putamen in both the less and more affected sides and the more affectedanterior putamen (P Q0.01). When the subject with estimated -15% dopaminerelease1h was excluded, there was significant dopamine release1h in the less affectedanterior putamen. dopamine release1h in the caudate was not significantly differentfrom zero (P = 0.09 and 0.11 for the more and less affected caudate respectively).There was significantly higher dopamine release1h in the posterior putamen com-pared to the anterior putamen in both the more (PQ0.01) and less affected (PQ0.05)sides. There was also significant asymmetry in the dopamine release1h between themore and less affected mean putamen (P Q0.01).To our initial surprise, we also found a significant negative correlation betweendopamine release1h and motor response to levodopa (measured at 2 h) in the more134Figure 6.4: Correlation between percentage dopamine release1h and motor responseto levodopa (LD) in the more affected anterior putamen (A) and more affectedmiddle putamen (B).affected anterior putamen (P = 0.001) and middle putamen (P Q0.05) as shown inFig.6.4. No other clinical or imaging measures entered the regression.JB4BG Joint duttyrn UnulysisWe extracted the first four sets of canonical variates, together accounting for 98%of the total variance in the data. The subject scores along the first three sets ofcanonical variates were significantly correlated after the permutation test (P Q0.01).A trend towards significance was observed for the fourth set of canonical variates(P = 0.07).[irst szt of xvnonixvl vvrivtzsSubject scores along the first set of canonical variates were highly correlated amongthe four datasets (P Q0.01) (Fig.6.5A). The first common spatial patterns showedrelatively decreased DASB binding in the putamen, an anterior-posterior gradientof dopaminergic denervation in the putamen for DTBZ and MP with the degree ofdopaminergic denervation in the anterior putamen contributing more than the poste-rior putamen, and relatively increased dopamine release1hin the putamen (Fig.6.5B).Forward stepwise multiple regression showed that poorer motor response was as-sociated with higher dopamine release1h (significant negative correlation, P Q10−5,135Figure 6.5: A) Correlations between the subject scores for DTBZ, MP and DASBspatial patterns and dopamine release1h spatial pattern along the first set of canon-ical variates. B) Spatial patterns along the first set of canonical variates. DA =dopamine. MCCA = Multi-set canonical correlation analysis. P1 = anterior puta-men. P2 = middle putamen. P3 = posterior putamenFig.6.6A), more preserved dopaminergic binding, i.e. lower expression of the DTBZpattern (significant positive correlation, P Q0.001, Fig.6.6B), and shorter diseaseduration (significant positive correlation, P Q0.05, Fig.6.6C). No other clinical vari-ables or subject scores entered the regression.There was a significant difference in the projection scores onto the commonspatial patterns along the first set of canonical variates between subjects who werestable and who developed dyskinesias three years after the baseline scans (P=0.003)as shown in Fig.6.7.136Figure 6.6: Correlations between motor response to levodopa (LD) and subjectscores for the dopamine release1h pattern (A), subject scores for the DTBZ pat-tern (B) and disease duration (C) along the first set of canonical variates. DA =dopamine. LD = levodopa.Figure 6.7: Projection scores on the spatial patterns along the first set of canonicalvariates for subjects who were stable or developed dyskinesias years after the baselinescans and subjects who are currently not due for the follow-up. ** indicates P-valueQ0.01.hzxony to fourth szts of xvnonixvl vvrivtzsThere was no correlation between motor response and subject scores for any spatialpatterns along the second to fourth sets of canonical variates. In brief, spatial137patterns along the second canonical variates showed lower dopamine release1h dueto increasing loss of dopaminergic and serotonergic terminals. Spatial patterns alongthe third set of canonical variates showed disease-induced asymmetry and gradientin the striatal regions mainly in the two dopaminergic tracers, consistent with aprevious study conducted in a subset of the same subject population [12]. Spatialpatterns along the fourth set of canonical variates showed the strongest contributionof serotonergic innervation to dopamine release1h in the putamen.Joint pvttzrn vnvl–sis outxomzs fiithout thz inxlusion of YVhW yvtvWithout the inclusion of DASB data, the common spatial patterns along the firstset of canonical variates also showed turnover-related dopamine release (higherdopamine release with higher dopaminergic denervation). However, there was nocorrelation between motor response to levodopa and subject scores for dopaminerelease1h patterns or any other imaging-derived patterns. There was a significantgroup separation in the projection scores onto the common spatial patterns alongthe first set of canonical variates between subjects who were stable and who devel-oped dyskinesias three years after the baseline scans (P=0.007); however, the groupseparation strength was weaker without the inclusion of DASB data in the analysis.JBI DiswussionJBIBE inivuriuty UnulysisLevodopa-induced change in RAC BPND (here referred to as dopamine release1h)was only significantly greater than zero in the putamen (an average of 6.7%), butnot in the caudate (an average of 3.1%). Our previous study showed a positivecorrelation between dopamine release1h and disease duration in both the caudateand putamen [55]. We did not observe any significant correlation between dopaminerelease1h and disease duration, likely due to the relatively narrow range of diseaseduration (an average of 3.1 years) in our subject population compared to an averageof 7.9 years disease duration reported in the previous study. The shorter diseaseduration in this study may also explain the relatively lower dopamine release1h inboth the caudate and putamen compared to this previous study (11% in the caudateand 13% in the putamen).Nevertheless, we observed an anterior-posterior gradient and asymmetry in dopamine138release1h in the putamen already in very early disease, consistent with the pattern ofdopaminergic denervation. Previous animal and human studies showed that higherlevodopa-induced dopamine release was associated with increased denervation inboth animal and human studies [153, 154].Surprisingly, poorer motor response to levodopa was associated with higherdopamine release1h in the more affected anterior and middle putamen. While thisseems counter-intuitive, it may suggest that dopamine release1h does not play amajor role in therapeutic benefit as defined here (which was measured 2 h after lev-odopa administration). This is consistent with the results obtained using the jointpattern analysis as discussed below.JBIBF Joint duttyrn UnulysisUnlike the univariate analysis, the joint pattern approach is able to separate themeasured observables (i.e. all PET tracers) into independent components, and thuslikely isolate different mechanisms contributing to dopamine release. The first com-mon spatial patterns were composed of high dopamine release1h in the presence ofdopaminergic and serotonergic denervation in the putamen; this pattern is conceptu-ally related to dopamine turnover by both dopaminergic and serotonergic neurons inthe putamen. Dopamine turnover is related to the relative strength of the dopamineuptake and release and was thought to be a compensatory mechanism in early orpreclinical disease [117]. Both animal [153, 155] and human studies [156, 157] haveshown high dopamine turnover associated with high dopaminergic denervation.Results from the regression analysis showed that subjects with high dopamineturnover respond more poorly to levodopa (as assessed by change in MDS-UPDRS2h after levodopa intake) taking into account a relatively preserved dopaminer-gic function (lower expression of the DTBZ pattern) and shorter disease duration.The high dopamine turnover in the presence of a relatively preserved dopaminer-gic function emphasizes that dopamine turnover from the serotonergic system playsan important role in response to levodopa. Importantly, the association betweenhigher expression of the dopamine release1h pattern and poorer motor response dis-appeared when we did not enforce a correlation with DASB in the identification ofthe patterns. This finding further supports a crucial contribution of the serotonergicsystem to an abnormal dopamine turnover in early disease. This observation is fur-ther supported by the fact that the inverse correlations between dopamine release1hin the anterior and middle putamen and motor response in the univariate analysis139were weaker (P = 0.001) than the correlation found in the joint pattern analysis(P = 0.0001), which isolated a very specific component of dopamine release (i.e.associated with higher turnover).It should be noted that the patients studied here had mild disease of short dura-tion thus the baseline motor severity was not high and one might expect relativelylimited clinical response to levodopa. This raises the possibility of flooring effectin the numerical calculation of motor response to levodopa; in fact, motor responseto levodopa (change in MDS-UPDRS) was 3.1 on average and there was signifi-cant correlation between MDS-UPDRS off and motor response (P = 0.02). Thepositive correlation between motor response and disease duration and dopaminergicdenervation may to some degree correct for the flooring effect (i.e. the dopamineturnover still independently inversely correlated with motor response to levodopaafter correcting for the effects of these variables in the stepwise regression analysis).It is important to note that the correlation between motor response to levodopa andhigher dopamine turnover (expression of the dopamine release1h pattern along thefirst set of canonical variates) was still significant without including disease durationor dopaminergic denervation (expression of the DTBZ pattern along the first set ofcanonical variates) as covariates. In univariate analysis, there was no significant cor-relation between dopamine release1h and MDS-UPDRS off, only between dopaminerelease1h and motor response.Additionally, our previous studies showed that those patients who were at higherrisk of developing motor fluctuations and dyskinesias had higher dopamine release1hafter levodopa intake, but lower dopamine release 4hr after levodopa intake com-pared to patients who still exhibited a stable response to treatment three years later[55, 148]. This rapid increase in dopamine release at 1 h was therefore interpreted asabnormal and in patients with high turnover, the rapid decline in synaptic dopaminemay be sufficient to result in reduced motor benefit 2 h after levodopa.In this work, we were able to extract a component of the abnormal dopaminerelease1h, related to high dopamine turnover in the presence of dopaminergic andserotonergic denervation, that was most strongly related to response to levodopa.In addition, we showed that the serotonergic system contributes to the abnormaldopamine release. A previous study showed that in more advanced disease (Stenyears disease duration), where there are fewer dopaminergic terminals, serotoner-gic neurons contribute to abnormal dopamine release in Parkinsons disease patientswith levodopa-induced dyskinesias [59]. It was then speculated that the relative140preservation of serotonergic terminals in advanced disease could be a risk factor forthe development of motor complications. We showed that in early disease, increasein dopamine turnover also occurs in the serotonergic neurons, which contributes toabnormal dopamine release. It was suggested that competition between levodopa-derived dopamine and serotonin at serotonergic synapse leads to reduced serotoninrelease and thereby to overactivation of serotonergic terminals as compensation [53].The overactivation of serotonergic terminals may result in higher dopamine turnoverand thus higher abnormal dopamine release even in early disease. A recent studyshowed reduced serotonergic function in the striatum in carriers with A53T SNCAmutation compared to healthy controls, who also exhibit higher risk of early motorfluctuations in response to levodopa [158, 159]. We also showed that this abnormaland rapid dopamine release, likely reflective of higher risk of motor complication inthe future, does not seem to confer sustained therapeutic benefits in early diseasestage. We hypothesize that this component of the dopamine release estimate, whichseems to be associated with altered dopamine turnover, may be a more selectivepredictor for the development of future motor complications than a full estimate ofdopamine release. To test the hypothesis, we are currently following all 18 Parkin-sons subjects clinically to assess occurrence of motor and non-motor complications.In principle, the analysis results should be generalizable to other Parkinsons diseasecohorts.aimitvtionsThere are several limitations for this study. By virtue of binding to VMAT2, DTBZis not entirely selective for dopaminergic neurons. Although the contribution of theserotonergic neuron to DTBZ binding is small in the striatum in healthy condition,it may not be entirely negligible in the disease stage. While these observations mayintroduce a potential confound in the interpretation of the data, such confounds arenot specific to this analysis method. In order to capture a complete picture of allpossible patterns contributing to the motor response to levodopa, subjects scoresfor patterns from all four sets of canonical variates should be used in the regres-sion analysis; however, our current limited sample size does not fully support theinclusion of too many variables. Preliminary regression analysis between motor re-sponse and subject scores of patterns of all four sets of canonical variates showed thestrongest contribution to come from the dopamine release1h pattern along the firstset of canonical variates, consistent with the regression results obtained using only141the patterns along the first set of canonical variates (see Supplementary Materialsfor detailed analysis). Another important limitation is the relatively small samplesize in this study. This is, however, a very complex and technically demanding study,which limited the number of subjects that could be included. Another thing to noteis the collinearity among the explanatory variables in the stepwise regression anal-ysis, i.e. the expressions of the spatial patterns along each set of canonical variatesare highly correlated by construct. We also performed LASSO regression with five-fold cross validation, more robust when the independent variables are correlated,and obtained consistent results as the stepwise regression analysis.JBJ ConwlusionIn this study, we examined the relationships between dopaminergic and seroton-ergic denervation and levodopa-induced dopamine release and their relationshipswith motor response to levodopa and risk of dyskinesia in early Parkinsons dis-ease. Using a novel joint pattern analysis approach, we identified a component ofdopamine release1h after levodopa intake that is conceptually related to increasingdopamine turnover in the presence of dopaminergic and serotonergic denervation.Importantly, we showed that the serotonergic system contributed significantly to thisspecific component of dopamine release already in early disease. Such relationshipsbetween dopamine release and serotonergic function was not observed with tradi-tional univariate analyses and represent a unique finding obtained with this novelanalysis method. In addition, we were able to show that this abnormal turnover-related dopamine release does not translate into sustained therapeutic efficacy evenin early disease and may selectively contribute to higher risk of motor complicationsin later disease stage.142Chuptyr KConwlusion und FuturyDirywtionsFigure 7.1: Overview for the thesis.In the work described in this thesis, we introduced and validated the use of dif-ferent network pattern analysis methods for analyzing PD-related changes in PETdata. The introduced methods presented a novel and more sensitive way to extractmeaningful information from multi-tracer PET data in a deterministic fashion. Thisinformation then allows us to gain a deeper understanding of the possible mecha-nisms underlying PD and its progression.The thesis started with an overview of the basic principles of PET imagingfrom data collection to image reconstruction, followed by the description of theapplication of multi-tracer PET imaging in studying neurotransmitter changes inPD in Chapter1.143We then examined spatial connectivity changes in the serotonergic system in thework described in Chapters2 and 3. It was one of the first few applications of net-work analyses to study changes in the neurotransmitter systems. We showed thatthere were PD and LRRK2 mutation-related network changes in the serotonergicsystem. The PD-related changes correlated significantly with both disease durationand dopaminergic denervation, unlike other alterations observed with univariateanalysis approaches. These results show that the serotonergic system is affectedprogressively in the manifest stage alongside the dopaminergic system; in addition,the serotonergic system is also affected in the prodromal stage. The alterations in theserotonergic system may be related to compensatory mechanisms against dopamin-ergic denervation and may be related to non-motor symptoms of PD. These findingsprovide new understanding of the underlying disease mechanisms and disease origin.From a technical point of view, these methods may be particularly well-suitedfor the analysis of the following types of PET studies: for the analysis of distributedtracers (e.g. serotonergic tracer DASB) at both ROI and voxel level or for the anal-ysis of localized tracer (e.g. dopaminergic tracer DTBZ) at voxel level. These PETstudies often have a larger number of ROIs/voxels (variables) than the number ofavailable subjects. Network analyses eliminate the sometimes over stringent correc-tion for multiple comparison used in univariate analyses. More importantly, spatialpatterns of tracer binding reflecting relationships between multiple brain regions aremore sensitive to deterministic changes than localized changes of tracer binding inindividual brain regions.The method for extracting spatial patterns described in the work done in Chap-ter2 is already well-developed and has been implemented to analyze different PETtracer datasets. The graph theory analysis approach introduced in the work de-scribed in Chapter3, however, still presents methodological challenges for the anal-ysis of PET data. More effort is needed in the future to determine:1. the best method to estimate partial correlation coefficients for a rank-deficientinput matrix (i.e. number of subjects Qnumber of ROIs)2. the effects of indirect edges in the adjacency matrices on the graph theorymetrics, i.e. the differences between the graph theory metrics obtained withfull correlation coefficients and partial correlation coefficients3. the stability and robustness of different graph theory metrics for examiningnetwork changes in neurotransmitter systems144It is important to address these potential issues when applying graph theory analysisto study network changes in the neurotransmitter systems, however, this is out ofthe scope for the work done in this thesis.An interesting future application of graph theory analysis to PET data is toincorporate the temporal kinetics in the PET signal in the analysis instead of em-bedding the temporal information in a single measurement (e.g. BPaW values) foreach subject. In this case, the input matrix (containing radioactivity concentrations)is in the dimension [subject by ROI by time] and the corresponding adjacency ma-trix can be in the dimension [ROI by ROI] for a set of time points (a sliding windowapproach) for each subject. The addition of the extra dimension (time) in the inputmatrix allows for subject-level analysis (i.e. group discrimination analysis), correla-tion analysis with clinical measures, and temporal analysis to examine the dynamicsof network topology metrics for the duration of the scan. The sliding window ap-proach has been used to study the dynamic functional connectivity changes usingFDG PET data [160]. Applying similar approaches to study the dynamic functionalconnectivity changes in neurotransmitter systems where the dynamic behaviour isclosely related to the kinetics of tracer uptake to a specific site of interest or more indepth investigation on the dynamics of the graph theory metrics may be of interest.In the work described in Chapter4, we introduced a novel network approachspecifically designed for tracking temporal changes related to disease progression.Our results suggested that mechanisms underlying disease progression and diseaseinitiation may follow distinct spatio-temporal patterns; these findings, again, pro-vided more in depth understanding on the mechanism and origin of PD, and provideda better way for tracking disease progression compared to traditional univariate mea-sures.From a technical point of view, the proposed DMD method has a wide rangeof applications, especially in the field of progressive neurodegenerative disorders(e.g. PD, Alzheimer’s disease), healthy aging or modelling the kinetics of radio-pharmaceuticals. For PD, an interesting application is to apply DMD to differ-ent PET tracers targeting the presynaptic dopaminergic system to extract dis-tinct spatio-temporal patterns reflecting different aspects of dopaminergic process-ing. The differences between the spatio-temporal patterns obtained with differentdopaminergic tracers may unravel the unique roles of each aspect of dopaminergicprocessing in different stages of the disease.Another interesting future application of the DMD approach is to model PET145tracer kinetics without the explicit use of compartmental models. The time dynamicof PET tracer uptakes depends on several factors, including the uptake rate from theblood, decay of the tracer radioactivity, and the transfer between different compart-ments. The traditional way to model the tracer kinetics is the use of compartmentalmodels (as introduced in the Chapter1) which requires solving a fixed number ofpre-defined exponential equations. In theory, the different physiological processescan be separately modeled by a set of exponential functions, and the overall dynamicof the measured PET signal is the sum of these exponential functions. This bringsup the potential of using DMD to extract meaningful outcome measures from theoverall PET signal with the following advantages: 1) DMD approach can extractthe number of meaningful governing functions in a data-driven fashion without apre-defined number of functions/compartments; 2) as many multivariate patternanalyses, DMD is less sensitive to noise compared to fitting functions to univariatemeasures.In the work described in the final two chapters (Chapters5 and 6), we proposeda joint pattern analysis approach for the analysis of multi-tracer PET data. Suchan approach has not been developed in the PET field, despite the need for moreeffective and sensitive ways to extract complementary information from multi-tracerdatasets. We first demonstrated the applicability and sensitivity of the method us-ing two dopaminergic tracers in Chapter5. Then in the work described in Chapter6,we showed, for the first time, the relationship between the dopaminergic and sero-tonergic system and levodopa-induced dopamine release, and their relationship withclinical response to levodopa and future risk of levodopa-induced motor complica-tions.Results presented in Chapter6 are of particular interest for the clinical advancesof PD. The results showed that the serotonergic system contributes to abnormaldopamine release that not only does not translate to therapeutic benefits in earlydisease, but also contributes to higher risk of dyskinesia in later disease stage. Thesefindings provided evidence for personalized treatment, particularly targeting theserotonergic system, for PD patients with a higher risk of developing motor compli-cations.From a technical point of view, further analysis should be performed to examinethe applicability of the method to combine other PET data and other imaging data.However, it is expected that the proposed method is highly applicable to many othermulti-tracer PET studies and other multi-modal neuroimaging studies in a variety146of clinical applications. The joint pattern analysis approach may be the optimaltool to effectively combine and extract information to maximize the utility of theavailable data. A particular interesting future application for this method is to thesimultaneous PET and fMRI data. While the PET data capture the metabolicfunction or specific aspects of the neurotransmitter systems, fMRI data capturethe overall brain activation. Extracting complementary information from the twoimaging datasets may reveal important disease-related information.In conclusion, in the work presented in this thesis, we introduced and validatedthe use of network pattern analysis for the analysis of PET data. The challenges forapplying these network analysis methods to the analysis of PET data have mainlybeen the interpretation of the changes in a network rather than in a single local-ized region. Wider applications of these methods require careful understanding andinterpretation of the analysis outcomes. Despite the potential challenges, the newanalysis approaches may convey novel information by taking advantages of detectingsubtle changes among multiple brain regions and extracting complementary infor-mation from multiple imaging datasets. It is expected that the further developmentand application of quantitative network analysis approaches can contribute to ourunderstanding of the functional changes in the brain related to both diseases andhealthy aging.147Vivliogruphy[1] M. D. Walker, euuntitutivy xynumiw GD pyt swunning oz thy voxy unx vruinusing Lgc tomogruphs. PhD thesis, University of Manchester, 2009.[2] N. E. Bolus, R. George, B. R. Newcomer, Others, J. . Washington, and B. R.Newcomer, “PET/MRI: the blended-modality choice of the future?,” Journuloz nuwlyur myxiwiny tywhnology, vol. 37, no. 2, pp. 63–71, 2009.[3] A. A. Lammertsma and S. P. Hume, “Simplified reference tissue model forPET receptor studies,” byuroimugy, vol. 4, no. 3, pp. 153–158, 1996.[4] A. J. Stoessl, W. W. Martin, M. J. McKeown, and V. Sossi, “Advances inimaging in Parkinson’s disease.,” hhy LunwytB byurology, vol. 10, pp. 987–1001, nov 2011.[5] B. Capuano, I. T. Crosby, and E. J. Lloyd, “Schizophrenia: genesis, recep-torology and current therapeutics,” Wurrynt myxiwinul whymistry, vol. 9, no. 5,pp. 521–548, 2002.[6] A. J. Stoessl, “Neuroimaging in Parkinson’s disease: from pathology to di-agnosis.,” durkinsonism : rylutyx xisorxyrs, vol. 18 Suppl 1, pp. S55–9, jan2012.[7] M. S. K. Sodhi and E. Sanders-Bush, “Serotonin and brain development,”Intyrnutionul ryviyw oz nyuroviology, vol. 59, pp. 111–174, 2004.[8] M. Rubinov and O. Sporns, “Complex network measures of brain connectivity:Uses and interpretations,” byuroImugy, vol. 52, pp. 1059–1069, 2009.[9] O. Sporns, D. R. Chialvo, M. Kaiser, and C. C. Hilgetag, “Organization,development and function of complex brain networks,” 2004.[10] J. Sui, T. Adali, Q. Yu, J. Chen, and V. D. Calhoun, “A review of multivariatemethods for multimodal fusion of brain imaging data,” feb 2012.148[11] J. F. Fu, I. Klyuzhin, S. Liu, E. Shahinfard, N. Vafai, J. McKenzie, N. Neilson,R. Mabrouk, M. A. Sacheli, D. Wile, M. J. McKeown, A. J. Stoessl, andV. Sossi, “Investigation of serotonergic Parkinson’s disease-related covariancepattern using [11C]-DASB/PET,” byuroImugyN Wliniwul, vol. 19, pp. 652–660,jan 2018.[12] J. F. Fu, I. Klyuzhin, J. McKenzie, N. Neilson, E. Shahinfard, K. Dinelle,M. J. McKeown, A. J. Stoessl, and V. Sossi, “Joint pattern analysis applied toPET DAT and VMAT2 imaging reveals new insights into Parkinson’s diseaseinduced presynaptic alterations,” byuroImugyN Wliniwul, vol. 23, p. 101856,jan 2019.[13] J. F. Fu, I. S. Klyuzhin, M. J. McKeown, A. J. Stoessl, and V. Sossi, “Noveldata-driven, equation-free method captures spatio-temporal patterns of neu-rodegeneration in Parkinson’s disease: Application of dynamic mode decom-position to PET,” byuroImugyN Wliniwul, vol. 25, p. 102150, jan 2020.[14] I. S. Klyuzhin, J. F. Fu, A. Hong, M. Sacheli, N. Shenkov, M. Matarazzo,A. Rahmim, A. Jon Stoessl, and V. Sossi, “Data-driven, voxel-based analysisof brain PET images: Application of PCA and LASSO methods to visualizeand quantify patterns of neurodegeneration,” dLog cbY, vol. 13, no. 11, 2018.[15] I. S. Klyuzhin, J. F. Fu, N. Shenkov, A. Rahmim, and V. Sossi, “Use ofGenerative Disease Models for Analysis and Selection of Radiomic Features inPET,” IYYY hrunsuwtions on fuxiution unx dlusmu ayxiwul gwiynwys, vol. PP,pp. 1–1, mar 2018.[16] J. F. Fu, Invystigution oz durkinson's xisyusy rylutyx puttyrn unx ultyryxxopuminy rylyusy puttyrn in tryutmyntAinxuwyx wompliwutions unx nonAmotorsymptoms oz durkinson's xisyusy. PhD thesis, University of British Columbia,2016.[17] S. M. Ametamey, M. Honer, and P. A. Schubiger, “Molecular imaging withpet,” Whymiwul ryviyws, vol. 108, no. 5, pp. 1501–1516, 2008.[18] C. S. Levin and E. J. Hoffman, “Calculation of positron range and its ef-fect on the fundamental limit of positron emission tomography system spatialresolution,” dhysiws in myxiwiny unx viology, vol. 44, no. 3, p. 781, 1999.149[19] M. Phelps and E. Hoffman, “N. mull ani, c. higgins, m. ter-pogossian,” designconsiderations for a positron emission transaxial tomograph (pet iii),”,” IYYYhrunsB BiomyxB YngB bgAFG, pp. 516–522, 1976.[20] J. A. Fessler and J. M. Ollinger, “Signal processing pitfalls in positron emissiontomography,” IYYY gignul drowyssing auguziny, vol. 14, no. 1, pp. 43–55,1997.[21] D. L. Bailey, euuntiwution in GD positron ymission tomogruphy. PhD thesis,University of Surrey, 1996.[22] R. D. Evans and A. Noyau, hhy utomiw nuwlyus, vol. 582. McGraw-Hill NewYork, 1955.[23] H. E. Johns, dhysiws oz ruxiology. Charles River Media, 1983.[24] C. Thompson, A. Dagher, D. Lunney, S. Strother, and A. Evans, “A techniqueto reject scattered radiation in pet transmission scans,” in dhysiws unx YngiAnyyring oz Womputyrizyx aultiximynsionul Imuging unx drowyssing, pp. 244–253, International Society for Optics and Photonics, 1986.[25] M. Xu, P. Cutler, and W. Luk, “Adaptive, segmented attenuation correctionfor whole-body pet imaging,” IYYY hrunsuwtions on buwlyur gwiynwy, vol. 43,no. 1, pp. 331–336, 1996.[26] J. Qi and R. H. Huesman, “Theoretical study of lesion detectability of map re-construction using computer observers,” IYYY trunsuwtions on myxiwul imugAing, vol. 20, no. 8, pp. 815–822, 2001.[27] A. Alessio, P. Kinahan, yt ulB, “Pet image reconstruction,” buwlyur myxiwiny,vol. 1, pp. 1–22, 2006.[28] L. A. Shepp and Y. Vardi, “Maximum likelihood reconstruction for emissiontomography,” IYYY trunsuwtions on myxiwul imuging, vol. 1, no. 2, pp. 113–122, 1982.[29] R. B. Innis, V. J. Cunningham, J. Delforge, M. Fujita, A. Gjedde, R. N.Gunn, J. Holden, S. Houle, S.-C. Huang, M. Ichise, and Others, “Consensusnomenclature for in vivo imaging of reversibly binding radioligands,” Journuloz Wyryvrul Bloox Flow : aytuvolism, vol. 27, no. 9, pp. 1533–1539, 2007.150[30] M. A. Mintun, M. E. Raichle, M. R. Kilbourn, G. F. Wooten, and M. J. Welch,“A quantitative model for the in vivo assessment of drug binding sites withpositron emission tomography,” Unnuls oz byurologyN cffiwiul Journul oz thyUmyriwun byurologiwul Ussowiution unx thy Whilx byurology gowiyty, vol. 15,no. 3, pp. 217–227, 1984.[31] R. N. Gunn, A. A. Lammertsma, S. P. Hume, and V. J. Cunningham, “Para-metric imaging of ligand-receptor binding in PET using a simplified referenceregion model,” byuroimugy, vol. 6, no. 4, pp. 279–287, 1997.[32] J. Logan, J. S. Fowler, N. D. Volkow, A. P. Wolf, S. L. Dewey, D. J. Schlyer,R. R. MacGregor, R. Hitzemann, B. Bendriem, S. J. Gatley, and Others,“Graphical Analysis of Reversible Radioligand Binding from TimeActivityMeasurements Applied to [N-11C-Methyl]-(-)-Cocaine PET Studies in Hu-man Subjects,” Journul oz Wyryvrul Bloox Flow : aytuvolism, vol. 10, no. 5,pp. 740–747, 1990.[33] L. M. L. de Lau and M. M. B. Breteler, “Epidemiology of Parkinson’s disease,”hhy Lunwyt byurology, vol. 5, no. 6, pp. 525–535, 2006.[34] T. Wichmann and M. R. DeLONG, “Pathophysiology of Parkinson’s disease:the MPTP primate model of the human disorder,” Unnuls oz thy byw morkUwuxymy oz gwiynwys, vol. 991, no. 1, pp. 199–213, 2003.[35] A. Schrag, L. Horsfall, K. Walters, A. Noyce, and I. Petersen, “Prediagnosticpresentations of Parkinson’s disease in primary care: A case-control study,”hhy Lunwyt byurology, vol. 14, no. 1, pp. 57–64, 2015.[36] M. Politis and C. Loane, “Serotonergic dysfunction in Parkinson’s diseaseand its relevance to disability,” hhy gwiyntiw korlx Journul, vol. 11, no. 11,pp. 1726–1734, 2011.[37] D. G. Healy, M. Falchi, S. S. O’Sullivan, V. Bonifati, A. Durr, S. Bressman,A. Brice, J. Aasly, C. P. Zabetian, S. Goldwurm, J. J. Ferreira, E. Tolosa,D. M. Kay, C. Klein, D. R. Williams, C. Marras, A. E. Lang, Z. K. Wszolek,J. Berciano, A. H. V. Schapira, T. Lynch, K. P. Bhatia, T. Gasser, A. J. Lees,N. W. Wood, and o. b. o. t. I. L. International LRRK2 Consortium, “Phe-notype, genotype, and worldwide genetic penetrance of LRRK2-associated151Parkinson’s disease: a case-control study.,” hhy LunwytB byurology, vol. 7,pp. 583–90, jul 2008.[38] C. J. Bardien S, Lesage S, Brice A, “Genetic characteristics of leucine-richrepeat kinase 2 (LRRK2) associated Parkinson’s disease.,” durkinsonism fylutDisorx, vol. 17, no. 7, pp. 501–8, 2011.[39] M. Farrer, J. Stone, I. F. Mata, S. Lincoln, J. Kachergus, M. Hulihan, K. J.Strain, and D. M. Maraganore, “LRRK2 mutations in Parkinson disease,”byurology, vol. 65, no. 5, pp. 738–740, 2005.[40] K. Marder, Y. Wang, R. N. Alcalay, H. Mejia-Santana, M.-X. Tang, A. Lee,D. Raymond, A. Mirelman, R. Saunders-Pullman, L. Clark, L. Ozelius, A. Orr-Urtreger, N. Giladi, S. Bressman, F. t. L. A. J. LRRK2 Ashkenazi JewishConsortium, R. N. Alcalay, M.-X. Tang, H. M. Santana, E. Roos, M. Orbe-Reilly, S. Fahn, L. Cote, C. Waters, P. Mazzoni, B. Ford, E. Louis, O. Levy,L. Rosado, D. Ruiz, T. Dorovski, P. Greene, L. Clark, K. S. Marder, T. Gure-vich, A. B. Shira, M. G. Weisz, K. Yasinovsky, M. Zalis, A. Thaler, Y. Bal-ash, S. Hertzel, Z. G. Or, H. Kobo, A. Hillel, A. Shkedy, A. Orr-Urtreger,N. Giladi, A. Deik, M. J. ames Barrett, J. Cabassa, M. Groves, A. L. Hunt,N. Lubarr, M. S. Luciano, J. Miravite, C. Palmese, R. Sachdev, H. Sarva,L. Severt, V. Shanker, M. C. Swan, J. Soto-Valencia, B. Johannes, R. Or-tega, L. Ozelius, R. Saunders-Pullman, D. Raymond, and S. Bressman, “Age-specific penetrance of LRRK2 G2019S in the Michael J. Fox Ashkenazi JewishLRRK2 Consortium.,” byurology, vol. 85, pp. 89–95, jul 2015.[41] J. R. Adams, H. Van Netten, M. Schulzer, E. Mak, J. Mckenzie, A. Strongosky,V. Sossi, T. J. Ruth, C. S. Lee, M. Farrer, T. Gasser, R. J. Uitti, D. B. Calne,Z. K. Wszolek, and A. J. Stoessl, “PET in LRRK2 mutations: Comparison tosporadic Parkinson’s disease and evidence for presymptomatic compensation,”Bruin, vol. 128, no. 12, pp. 2777–2785, 2005.[42] C. Marras, B. Schuele, R. P. Munhoz, E. Rogaeva, J. W. Langston, M. Kasten,C. Meaney, C. Klein, P. M. Wadia, S. Y. Lim, R. S. I. Chuang, C. Zadikof,T. Steeves, K. M. Prakash, R. M. A. De Bie, G. Adeli, T. Thomsen, K. K. Jo-hansen, H. A. Teive, A. Asante, W. Reginold, and A. E. Lang, “Phenotype inparkinsonian and nonparkinsonian LRRK2 G2019S mutation carriers,” byuArology, vol. 77, pp. 325–333, jul 2011.152[43] C. Marras, R. N. Alcalay, C. Caspell-Garcia, C. Coffey, P. Chan, J. E.Duda, M. F. Facheris, R. Ferna´ndez-Santiago, J. Ru´ız-Mart´ınez, T. Mestre,R. Saunders-Pullman, C. Pont-Sunyer, E. Tolosa, and B. Waro, “Motor andnonmotor heterogeneity of LRRK2-related and idiopathic Parkinson’s dis-ease,” aovymynt Disorxyrs, vol. 31, pp. 1192–1202, aug 2016.[44] D. J. Wile, P. A. Agarwal, M. Schulzer, E. Mak, K. Dinelle, E. Shahinfard,N. Vafai, K. Hasegawa, J. Zhang, J. McKenzie, N. Neilson, A. Strongosky, R. J.Uitti, M. Guttman, C. P. Zabetian, Y.-S. S. Ding, M. Adam, J. Aasly, Z. K.Wszolek, M. Farrer, V. Sossi, and A. J. Stoessl, “Serotonin and dopaminetransporter PET changes in the premotor phase of LRRK2 parkinsonism:cross-sectional studies,” hhy Lunwyt byurology, vol. 16, pp. 351–359, may 2017.[45] V. Sossi, R. de la Fuente-Ferna´ndez, R. Nandhagopal, M. Schulzer, J. McKen-zie, T. J. Ruth, J. O. Aasly, M. J. Farrer, Z. K. Wszolek, and J. A. Stoessl,“Dopamine turnover increases in asymptomatic LRRK2 mutations carriers,”aovymynt Disorxyrs, vol. 25, pp. 2717–2723, dec 2010.[46] S. Y. Liu, D. J. Wile, J. F. Fu, J. Valerio, E. Shahinfard, S. McCormick,R. Mabrouk, N. Vafai, J. McKenzie, N. Neilson, A. Perez-Soriano, J. E. Arena,M. Cherkasova, P. Chan, J. Zhang, C. P. Zabetian, J. O. Aasly, Z. K. Wszolek,M. J. McKeown, M. J. Adam, T. J. Ruth, M. Schulzer, V. Sossi, and A. J.Stoessl, “The effect of LRRK2 mutations on the cholinergic system in manifestand premanifest stages of Parkinson’s disease: a cross-sectional PET study,”hhy Lunwyt byurology, vol. 17, pp. 309–316, apr 2018.[47] E. R. Block, J. Nuttle, J. J. Balcita-Pedicino, J. Caltagarone, S. C. Watkins,S. R. Sesack, and A. Sorkin, “Brain Region-Specific Trafficking of theDopamine Transporter,” Journul oz byuroswiynwy, vol. 35, pp. 12845–12858,sep 2015.[48] V. Sossi, R. R. De La Fuente-Ferna´ndez, M. Schulzer, A. R. Troiano, T. J.Ruth, and A. J. Stoessl, “Dopamine transporter relation to dopamine turnoverin Parkinson’s disease: A positron emission tomography study,” Unnuls ozbyurology, vol. 62, pp. 468–474, nov 2007.[49] R. Nandhagopal, E. Mak, M. Schulzer, J. McKenzie, S. McCormick, V. Sossi,T. J. Ruth, A. Strongosky, M. J. Farrer, Z. K. Wszolek, A. J. Stoessl, D. B.153Calne, and D. W. Dickson, “Progression of dopaminergic dysfunction in aLRRK2 kindred: a multitracer PET study.,” byurology, vol. 71, pp. 1790–5,nov 2008.[50] C. S. Lee, A. Samii, V. Sossi, T. J. Ruth, M. Schulzer, J. E. Holden, J. Wudel,P. K. Pal, R. De La Fuente-Fernandez, D. B. Calne, and A. J. Stoessl, “In vivopositron emission tomographic evidence for compensatory changes in presy-naptic dopaminergic nerve terminals in Parkinson’s disease,” Unnuls oz byuArology, vol. 47, pp. 493–503, apr 2000.[51] R. Nandhagopal, L. Kuramoto, M. Schulzer, E. Mak, J. Cragg, C. S. Lee,J. McKenzie, S. McCormick, A. Samii, A. Troiano, T. J. Ruth, V. Sossi,R. De La Fuente-Fernandez, D. B. Calne, and A. J. Stoessl, “Longitudinalprogression of sporadic Parkinson’s disease: A multi-tracer positron emissiontomography study,” Bruin, vol. 132, pp. 2970–2979, nov 2009.[52] V. Sossi, K. Dinelle, G. J. Topping, J. E. Holden, D. Doudet, M. Schulzer,T. J. Ruth, A. J. Stoessl, D. La Fuente-Fernandez, and Others, “Dopaminetransporter relation to levodopa-derived synaptic dopamine in a rat model ofParkinson’s: an in vivo imaging study,” Journul oz nyurowhymistry, vol. 109,no. 1, pp. 85–92, 2009.[53] C. Loane, P. Su, and M. Politis, “Different dyskinesias in Parkinson’s diseaseand their relationship to levodopa,” 2011.[54] N. Pavese, A. H. Evans, Y. F. Tai, G. Hotton, D. J. Brooks, A. J. Lees,and P. Piccini, “Clinical correlates of levodopa-induced dopamine release inParkinson disease: a PET study.,” byurology, vol. 67, pp. 1612–7, nov 2006.[55] R. De La Fuente-Ferna´ndez, V. Sossi, Z. Huang, S. Furtado, J. Q. Lu, D. B.Calne, T. J. Ruth, and A. J. Stoessl, “Levodopa-induced changes in synapticdopamine levels increase with progression of Parkinson’s disease: Implicationsfor dyskinesias,” Bruin, vol. 127, pp. 2747–2754, nov 2004.[56] B. Lavoie and A. Parent, “Immunohistochemical study of the serotoninergicinnervation of the basal ganglia in the squirrel monkey,” Journul oz WompurAutivy byurology, vol. 299, pp. 1–16, sep 1990.154[57] H. Braak, E. Ghebremedhin, U. Ru¨b, H. Bratzke, and K. Del Tredici, “Stagesin the development of Parkinson’s disease-related pathology,” 2004.[58] M. Carta, T. Carlsson, A. Mun˜oz, D. Kirik, and A. Bjo¨rklund, “Involvementof the serotonin system in l-dopa-induced dyskinesias,” durkinsonism unx fyAlutyx Disorxyrs, vol. 14, jul 2008.[59] M. Politis, K. Wu, C. Loane, D. J. Brooks, L. Kiferle, F. E. Turkheimer,P. Bain, S. Molloy, and P. Piccini, “Serotonergic mechanisms responsiblefor levodopa-induced dyskinesias in Parkinson’s disease patients,” Journul ozWliniwul Invystigution, vol. 124, pp. 1341–1349, mar 2014.[60] M. P. van den Heuvel and H. E. Hulshoff Pol, “Exploring the brain network:A review on resting-state fMRI functional connectivity,” 2010.[61] C. Habeck, Y. Stern, and t. A. D. N. Alzheimer’s Disease Neuroimaging Ini-tiative, “Multivariate data analysis for neuroimaging data: overview and ap-plication to Alzheimer’s disease.,” Wyll viowhymistry unx viophysiws, vol. 58,pp. 53–67, nov 2010.[62] M. C. Campbell, J. Markham, H. Flores, J. M. Hartlein, A. M. Goate, N. J.Cairns, T. O. Videen, and J. S. Perlmutter, “Principal component analysisof PiB distribution in Parkinson and Alzheimer diseases,” byurology, vol. 81,no. 6, pp. 520–527, 2013.[63] D. Eidelberg, “Metabolic brain networks in neurodegenerative disorders: afunctional imaging approach,” hrynxs in nyuroswiynwys, vol. 32, no. 10,pp. 548–557, 2009.[64] T. Eckert, C. Tang, and D. Eidelberg, “Assessment of the progression ofParkinson’s disease: a metabolic network approach,” 2007.[65] C. Lozza, J.-C. C. Baron, D. Eidelberg, M. J. Mentis, M. Carbon, and R.-M. M. Marie´, “Executive processes in Parkinson’s disease: FDG-PET andnetwork analysis,” Humun Bruin aupping, vol. 22, no. 3, pp. 236–245, 2004.[66] C. Huang, P. Mattis, C. Tang, K. Perrine, M. Carbon, and D. Eidelberg,“Metabolic brain networks associated with cognitive function in Parkinson’sdisease,” byuroimugy, vol. 34, no. 2, pp. 714–723, 2007.155[67] Y. Liu, M. Liang, Y. Zhou, Y. He, Y. Hao, M. Song, C. Yu, H. Liu, Z. Liu, andT. Jiang, “Disrupted small-world networks in schizophrenia,” Bruin, vol. 131,no. 4, pp. 945–961, 2008.[68] N. Tuovinen, K. Seppi, F. de Pasquale, C. Mu¨ller, M. Nocker, M. Schocke,E. R. Gizewski, C. Kremser, G. K. Wenning, W. Poewe, A. Djamshidian,C. Scherfler, and M. Seki, “The reorganization of functional architecture inthe early-stages of Parkinson’s disease,” feb 2018.[69] J. H. Ko, P. G. Spetsieris, and D. Eidelberg, “Network Structure and Functionin Parkinson’s Disease,” Wyryvrul Wortyfi, vol. 28, pp. 1–15, oct 2017.[70] P. Schmid, “Dynamic mode decomposition of numerical and experimentaldata,” JB Fluix aywh, vol. 656, pp. 5–28, 2010.[71] B. W. Brunton, L. A. Johnson, J. G. Ojemann, and J. N. Kutz, “Extract-ing spatialtemporal coherent patterns in large-scale neural recordings usingdynamic mode decomposition,” Journul oz byuroswiynwy aythoxs, vol. 258,pp. 1–15, jan 2016.[72] J. Casorso, X. Kong, W. Chi, D. Van De Ville, B. T. T. Yeo, and R. Lie´geois,“Dynamic mode decomposition of resting-state and task fMRI,” 2018.[73] J. L. Proctor and P. A. Eckhoff, “Discovering dynamic patterns from infectiousdisease data using dynamic mode decomposition,” Intyrnutionul Hyulth, vol. 7,pp. 139–145, mar 2015.[74] K. Taira, S. L. Brunton, S. T. M. Dawson, C. W. Rowley, T. Colonius, B. J.McKeon, O. T. Schmidt, S. Gordeyev, V. Theofilis, and L. S. Ukeiley, “ModalAnalysis of Fluid Flows: An Overview,” UIUU Journul, vol. 55, pp. 4013–4041, dec 2017.[75] N. M. Correa, Y.-O. Li, T. Adali, and V. D. Calhoun, “Fusion of fMRI, sMRI,and EEG data using canonical correlation analysis,” in FDDM IYYY IntyrnuAtionul Wonzyrynwy on Uwoustiws, gpyywh unx gignul drowyssing, pp. 385–388,IEEE, apr 2009.[76] N. M. Correa, T. Eichele, T. Adali, Y.-O. Li, and V. D. Calhoun, “Multi-setcanonical correlation analysis for the fusion of concurrent single trial ERP andfunctional MRI.,” byuroImugy, vol. 50, pp. 1438–45, may 2010.156[77] A. J. Stoessl, “Functional imaging studies of non-motoric manifestations ofParkinson’s Disease,” durkinsonism unx fylutyx Disorxyrs, vol. 15, pp. S13–6, dec 2009.[78] M. Politis, K. Wu, C. Loane, F. E. Turkheimer, S. Molloy, D. J. Brooks, andP. Piccini, “Depressive symptoms in PD correlate with higher 5-HTT bindingin raphe and limbic structures,” byurology, vol. 75, no. 21, pp. 1920–1927,2010.[79] C. Loane, K. Wu, P. Bain, J. David, P. Piccini, and M. Politis, “Serotonergicloss in motor circuitries correlates with severity of action-postural tremor inPD,” byurology, vol. 80, pp. 1850–1855, may 2013.[80] M. Politis, C. Loane, K. Wu, D. J. Brooks, and P. Piccini, “Serotonergicmediated body mass index changes in Parkinson’s disease,” byuroviology ozxisyusy, vol. 43, no. 3, pp. 609–615, 2011.[81] H. Braak, E. Ghebremedhin, U. Ru¨b, H. Bratzke, and K. Del Tredici, “Stagesin the development of Parkinson’s disease-related pathology,” Wyll unx tissuyrysyurwh, vol. 318, no. 1, pp. 121–134, 2004.[82] H. Braak, K. Del Tredici, U. Ru¨b, R. A. I. de Vos, E. N. H. J. Steur, andE. Braak, “Staging of brain pathology related to sporadic Parkinson’s disease,”byuroviology oz uging, vol. 24, no. 2, pp. 197–211, 2003.[83] J. Trinh, E. K. Gustavsson, C. Vilarin˜o-Gu¨ell, S. Bortnick, J. Latourelle,M. B. McKenzie, C. S. Tu, E. Nosova, J. Khinda, A. Milnerwood, S. Lesage,A. Brice, M. Tazir, J. O. Aasly, L. Parkkinen, H. Haytural, T. Foroud, R. H.Myers, S. B. Sassi, E. Hentati, F. Nabli, E. Farhat, R. Amouri, F. Hentati,and M. J. Farrer, “DNM3 and genetic modifiers of age of onset in LRRK2Gly2019Ser parkinsonism: a genome-wide linkage and association study,” hhyLunwyt byurology, vol. 15, pp. 1248–1256, nov 2016.[84] M. Sierra, P. Sa´nchez-Juan, M. I. Mart´ınez-Rodr´ıguez, I. Gonza´lez-Aramburu,I. Garc´ıa-Gorostiaga, M. R. Quirce, E. Palacio, J. M. Carril, J. Berciano,O. Combarros, and J. Infante, “Olfaction and imaging biomarkers in premotorLRRK2 G2019S-associated parkinson disease,” byurology, vol. 80, pp. 621–626, feb 2013.157[85] S. Srivatsal, B. Cholerton, J. B. Leverenz, Z. K. Wszolek, R. J. Uitti, D. W.Dickson, D. Weintraub, J. Q. Trojanowski, V. M. Van Deerlin, J. F. Quinn,K. A. Chung, A. L. Peterson, S. A. Factor, C. Wood-Siverio, J. G. Goldman,G. T. Stebbins, B. Bernard, B. Ritz, R. Rausch, A. J. Espay, F. J. Revilla,J. Devoto, L. S. Rosenthal, T. M. Dawson, M. S. Albert, I. F. Mata, S. C.Hu, K. S. Montine, C. Johnson, T. J. Montine, K. L. Edwards, J. Zhang, andC. P. Zabetian, “Cognitive profile of LRRK2-related Parkinson’s disease,”aovymynt Disorxyrs, vol. 30, pp. 728–733, apr 2015.[86] H. W. A. M. de Jong, F. H. P. van Velden, R. W. Kloet, F. L. Buijs, R. Boel-laard, and A. A. Lammertsma, “Performance evaluation of the ECAT HRRT:an LSO-LYSO double layer high resolution, high sensitivity scanner,” dhysiwsin ayxiwiny unx Biology, vol. 52, pp. 1505–1526, mar 2007.[87] C. Comtat, F. Bataille, C. Michel, J. Jones, M. Sibomana, L. Janeiro, andR. Trebossen, “OSEM-3D reconstruction strategies for the ECAT HRRT,”IYYY gymposium Wonzyrynwy fyworx buwlyur gwiynwy FDDHB, vol. 6, no. 3,pp. 3492–3496, 2004.[88] Y. Wu and R. E. Carson, “Noise reduction in the simplified reference tissuemodel for neuroreceptor functional imaging,” Journul oz Wyryvrul Bloox Flow: aytuvolism, vol. 22, pp. 1440–1452, dec 2002.[89] H. Akaike, “An information criterion (AIC),”auth gwi, vol. 14, no. 153, pp. 5–9, 1976.[90] S. Holm, “A Simple Sequentially Rejective Multiple Test Procedure A Sim-ple Sequentially Rejective Multiple Test Procedure,” gourwyN gwunxinuviunJournul oz gtutistiws gwunx J gtutist, vol. 6, no. 6, pp. 65–70, 1979.[91] D. J. Wile, K. Dinelle, N. Vafai, J. Mckenzie, J. K. Tsui, P. Schaffer, Y. S. Ding,M. Farrer, V. Sossi, and A. J. Stoessl, “A scan without evidence is not evidenceof absence: Scans without evidence of dopaminergic deficit in a symptomaticleucine-rich repeat kinase 2 mutation carrier,” aovymynt Disorxyrs, vol. 31,pp. 405–409, mar 2016.[92] S. J. Kish, J. Tong, O. Hornykiewicz, A. Rajput, L.-J. J. Chang, M. Guttman,and Y. Furukawa, “Preferential loss of serotonin markers in caudate versusputamen in Parkinson’s disease,” Bruin, vol. 131, pp. 120–131, oct 2008.158[93] B. Scholtissen, F. R. J. Verhey, H. W. M. Steinbusch, and A. F. G. Leentjens,“Serotonergic mechanisms in Parkinson’s disease: Opposing results from pre-clinical and clinical data,” Journul oz byurul hrunsmission, vol. 113, pp. 59–73,jan 2006.[94] A. Dagher, A. M. Owen, H. Boecker, and D. J. Brooks, “The role of the stria-tum and hippocampus in planning: A PET activation study in Parkinson’sdisease.,” Bruin, vol. 124, pp. 1020–1032, may 2001.[95] A. Tessitore, A. R. Hariri, F. Fera, W. G. Smith, T. N. Chase, T. M. Hyde,D. R. Weinberger, and V. S. Mattay, “Dopamine modulates the response ofthe human amygdala: a study in Parkinson’s disease.,” hhy Journul oz nyuroAswiynwy N thy offiwiul journul oz thy gowiyty zor byuroswiynwy, vol. 22, pp. 9099–103, oct 2002.[96] Y. Ouchi, E. Yoshikawa, H. Okada, M. Futatsubashi, Y. Sekine, M. Iyo, andM. Sakamoto, “Alterations in binding site density of dopamine transporterin the striatum, orbitofrontal cortex, and amygdala in early Parkinson’s dis-ease: Compartment analysis for [11C]-CFT binding with positron emissiontomography,” Unnuls oz byurology, vol. 45, pp. 601–610, may 1999.[97] J. H. Meyer, A. A. Wilson, N. Ginovart, V. Goulding, D. Hussey, K. Hood,and S. Houle, “Occupancy of serotonin transporters by paroxetine and citalo-pram during treatment of depression: A [C-11]DASB PET imaging study,”Umyriwun Journul oz dsywhiutry, vol. 158, pp. 1843–1849, nov 2001.[98] M. Politis, K. Wu, C. Loane, L. Kiferle, S. Molloy, D. J. Brooks, and P. Piccini,“Staging of serotonergic dysfunction in Parkinson’s disease: an in vivo 11 C-DASB PET study,” byuroviology oz xisyusy, vol. 40, no. 1, pp. 216–221, 2010.[99] M. Veronese, L. Moro, M. Arcolin, O. Dipasquale, G. Rizzo, P. Expert,W. Khan, P. M. Fisher, C. Svarer, A. Bertoldo, O. Howes, and F. E.Turkheimer, “Covariance statistics and network analysis of brain PET imagingstudies,” gwiyntiw fyports, vol. 9, p. 2496, dec 2019.[100] A. Das, A. L. Sampson, C. Lainscsek, L. Muller, W. Lin, J. C. Doyle, S. S.Cash, E. Halgren, and T. J. Sejnowski, “Interpretation of the Precision Matrixand Its Application in Estimating Sparse Brain Connectivity during Sleep159Spindles from Human Electrocorticography Recordings,” byurul Womputution,vol. 29, pp. 603–642, mar 2017.[101] J. Friedman, T. Hastie, and R. Tibshirani, “Sparse inverse covariance estima-tion with the graphical lasso,” tech. rep., 2007.[102] M. Girvan and M. E. J. Newman, “Community structure in social and biologi-cal networks.,” drowyyxings oz thy butionul Uwuxymy oz gwiynwys oz thy inityxgtutys oz Umyriwu, vol. 99, pp. 7821–6, jun 2002.[103] E. Bullmore and O. Sporns, “Complex brain networks: graph theoretical anal-ysis of structural and functional systems,” butury duvlishing Group, vol. 10,no. 3, pp. 186–198, 2009.[104] D. G. Tomasi, E. Shokri-Kojori, C. E. Wiers, S. W. Kim, . B. Demiral, E. A.Cabrera, E. Lindgren, G. Miller, G.-J. Wang, and N. D. Volkow, “Dynamicbrain glucose metabolism identifies anti-correlated cortical-cerebellar networksat rest,” Journul oz Wyryvrul Bloox Flow : aytuvolism, vol. 37, pp. 3659–3670,dec 2017.[105] S. Passow, K. Specht, T. C. Adamsen, M. Biermann, N. Brekke, A. R. Craven,L. Ersland, R. Gru¨ner, N. Kleven-Madsen, O.-H. Kvernenes, T. Schwar-zlmu¨ller, R. A. Olesen, and K. Hugdahl, “Default-mode network functionalconnectivity is closely related to metabolic activity,” Humun Bruin aupping,vol. 36, pp. 2027–2038, jun 2015.[106] S. J. Colloby, E. D. Williams, D. J. Burn, J. J. Lloyd, I. G. McKeith, andJ. T. O’Brien, “Progression of dopaminergic degeneration in dementia withLewy bodies and Parkinson’s disease with and without dementia assessed using123I-FP-CIT SPECT,” Yuropyun Journul oz buwlyur ayxiwiny unx aolywulurImuging, vol. 32, pp. 1176–1185, oct 2005.[107] S. Badoud, N. Nicastro, V. Garibotto, P. R. Burkhard, and S. Haller, “Dis-tinct spatiotemporal patterns for disease duration and stage in Parkinson’s dis-ease,” Yuropyun Journul oz buwlyur ayxiwiny unx aolywulur Imuging, vol. 43,pp. 509–516, mar 2016.[108] B. W. Brunton, L. A. Johnson, J. G. Ojemann, and J. N. Kutz, “Extract-ing spatialtemporal coherent patterns in large-scale neural recordings using160dynamic mode decomposition,” Journul oz byuroswiynwy aythoxs, vol. 258,pp. 1–15, jan 2016.[109] C. S. Lee, M. Schulzer, R. De La Fuente-Fernandez, E. Mak, L. Kuramoto,V. Sossi, T. J. Ruth, D. B. Calne, and A. J. Stoessl, “Lack of regional selec-tivity during the progression of Parkinson disease: Implications for pathogen-esis,” Urwhivys oz byurology, vol. 61, pp. 1920–1925, dec 2004.[110] J. H. Kordower, C. W. Olanow, H. B. Dodiya, Y. Chu, T. G. Beach, C. H.Adler, G. M. Halliday, and R. T. Bartus, “Disease duration and the integrityof the nigrostriatal system in Parkinson’s disease,” Bruin, vol. 136, pp. 2419–2431, aug 2013.[111] T. N. Alerte, A. A. Akinfolarin, E. E. Friedrich, S. A. Mader, C.-S. Hong, andR. G. Perez, “-Synuclein aggregation alters tyrosine hydroxylase phospho-rylation and immunoreactivity: Lessons from viral transduction of knockoutmice,” byuroswiynwy Lyttyrs, vol. 435, pp. 24–29, apr 2008.[112] B. Fischl, D. H. Salat, E. Busa, M. Albert, M. Dieterich, C. Haselgrove,A. van der Kouwe, R. Killiany, D. Kennedy, S. Klaveness, A. Montillo,N. Makris, B. Rosen, and A. M. Dale, “Whole brain segmentation: auto-mated labeling of neuroanatomical structures in the human brain.,” byuron,vol. 33, pp. 341–55, jan 2002.[113] J. H. Tu, C. W. Rowley, D. M. Luchtenburg, S. L. Brunton, and J. N. Kutz,“On Dynamic Mode Decomposition: Theory and Applications,” 2013.[114] J. Mu¨ller, G. K. Wenning, K. Jellinger, A. McKee, W. Poewe, and I. Litvan,“Progression of Hoehn and Yahr stages in parkinsonian disorders: A clinico-pathologic study,” byurology, vol. 55, pp. 888–891, sep 2000.[115] R. Nandhagopal, L. Kuramoto, M. Schulzer, E. Mak, J. Cragg, J. McKenzie,S. McCormick, T. J. Ruth, V. Sossi, R. De La Fuente-Fernandez, and A. J.Stoessl, “Longitudinal evolution of compensatory changes in striatal dopamineprocessing in Parkinson’s disease,” Bruin, vol. 134, pp. 3290–3298, nov 2011.[116] D. J. Brooks, “Imaging dopamine transporters in Parkinson’s disease,”Biomurkyrs in ayxiwiny, vol. 4, pp. 651–660, oct 2010.161[117] V. Sossi, R. de la Fuente-Ferna´ndez, J. E. Holden, M. Schulzer, T. J. Ruth,and J. Stoessl, “Changes of Dopamine Turnover in the Progression of Parkin-son’s Disease as Measured by Positron Emission Tomography: Their Relationto Disease-Compensatory Mechanisms,” Journul oz Wyryvrul Bloox Flow :aytuvolism, vol. 24, pp. 869–876, aug 2004.[118] E. Bezard, C. E. Gross, and J. M. Brotchie, “Presymptomatic compensationin Parkinson’s disease is not dopamine-mediated,” apr 2003.[119] C. P. Weingarten, M. H. Sundman, P. Hickey, and N. kuei Chen, “Neuroimag-ing of Parkinson’s disease: Expanding views,” dec 2015.[120] E. Dailly, F. Chenu, C. E. Renard, and M. Bourin, “Dopamine, depressionand antidepressants,” dec 2004.[121] F. Javoy-Agid, M. Ruberg, H. Taquet, B. Bokobza, Y. Agid, P. Gaspar,B. Berger, J. N’Guyen-Legros, C. Alvarez, and F. Gray, “Biochemical neu-ropathology of Parkinson’s disease,” Uxvunwys in byurology, vol. 40, pp. 189–198, 1984.[122] A. Vo, W. Sako, K. Fujita, S. Peng, P. J. Mattis, F. M. Skidmore, Y. Ma,A. M. Ulu, and D. Eidelberg, “Parkinson’s disease-related network topogra-phies characterized with resting state functional MRI,” Humun Bruin aupAping, vol. 38, pp. 617–630, feb 2017.[123] V. D. Calhoun and J. Sui, “Multimodal Fusion of Brain Imaging Data: A Keyto Finding the Missing Link(s) in Complex Mental Illness,” may 2016.[124] T. Fearn, “On orthogonal signal correction,” WhymomB IntyllB LuvB gystB,vol. 50, pp. 47–52, jan 2000.[125] S. Wold, K. Esbensen, and P. Geladi, “Principal component analysis,” WhymoAmytriws unx intylligynt luvorutory systyms, vol. 2, no. 1, pp. 37–52, 1987.[126] X. Chen, X. Chen, R. K. Ward, and Z. J. Wang, “A joint multimodal groupanalysis framework for modeling corticomuscular activity,” IYYY hrunsuwtionson aultimyxiu, vol. 15, pp. 1049–1059, aug 2013.162[127] Y. Masuo, D. Pe´laprat, D. Scherman, and W. Roste`ne,“[3H]Dihydrotetrabenazine, a new marker for the visualization of dopamin-ergic denervation in the rat striatum,” byuroswiynwy Lyttyrs, vol. 114, no. 1,pp. 45–50, 1990.[128] P. Fazio, P. Svenningsson, Z. Csele´nyi, C. Halldin, L. Farde, and A. Varrone,“Nigrostriatal dopamine transporter availability in early Parkinson’s disease,”aovymynt Disorxyrs, vol. 33, pp. 592–599, apr 2018.[129] D. Sulzer, S. J. Cragg, and M. E. Rice, “Striatal dopamine neurotransmission:Regulation of release and uptake,” aug 2016.[130] T. V. Borght, M. Kilbourn, T. Desmond, D. Kuhl, and K. Frey, “The vesicularmonoamine transporter is not regulated by dopaminergic drug treatments,”Yuropyun Journul oz dhurmuwology, vol. 294, pp. 577–583, dec 1995.[131] M. Karimi, L. Tian, C. A. Brown, H. P. Flores, S. K. Loftin, T. O. Videen,S. M. Moerlein, and J. S. Perlmutter, “Validation of nigrostriatal positronemission tomography measures: Critical limits,” Unnuls oz byurology, vol. 73,pp. 390–396, mar 2013.[132] L. Tian, M. Karimi, S. K. Loftin, C. A. Brown, H. Xia, J. Xu, R. H. Mach, andJ. S. Perlmutter, “No differential regulation of dopamine transporter (DAT)and vesicular monoamine transporter 2 (VMAT2) binding in a primate modelof Parkinson disease.,” dlog ony, vol. 7, no. 2, p. e31439, 2012.[133] R. Tibshirani, “Regression shrinkage and selection via the lasso: A retrospec-tive,” Journul oz thy foyul gtutistiwul gowiytyB gyriys BN gtutistiwul aythoxolAogy, vol. 73, no. 3, pp. 273–282, 2011.[134] V. Kaasinen and T. Vahlberg, “Striatal dopamine in Parkinson disease: Ameta-analysis of imaging studies,” dec 2017.[135] K.-J. J. Lin, Y.-H. H. Weng, C.-J. J. Hsieh, W.-Y. Y. Lin, S.-P. P. Wey, M.-P. P. Kung, T.-C. C. Yen, C.-S. S. Lu, and I.-T. T. Hsiao, “Brain Imagingof Vesicular Monoamine Transporter Type 2 in Healthy Aging Subjects by18F-FP-(+)-DTBZ PET,” dLog cbY, vol. 8, p. e75952, sep 2013.163[136] L. Prensa, M. Cossette, and A. Parent, “Dopaminergic innervation of humanbasal ganglia.,” Journul oz whymiwul nyurounutomy, vol. 20, pp. 207–213, dec2000.[137] M. A. Sanchez-Gonzalez, “The Primate Thalamus Is a Key Target for BrainDopamine,” Journul oz byuroswiynwy, vol. 25, pp. 6076–6083, mar 2005.[138] K. Koblinger, T. Fu¨zesi, J. Ejdrygiewicz, A. Krajacic, J. S. Bains, and P. J.Whelan, “Characterization of A11 neurons projecting to the spinal cord ofmice,” dLog cbY, vol. 9, no. 10, p. e109636, 2014.[139] S. Sharma, L. H. Kim, K. A. Mayr, D. A. Elliott, and P. J. Whelan, “Paralleldescending dopaminergic connectivity of A13 cells to the brainstem locomotorcenters,” gwiyntiw fyports, vol. 8, p. 7972, dec 2018.[140] R. E. Burke and K. O’Malley, “Axon degeneration in Parkinson’s disease,”aug 2013.[141] Z. Kurowska, J. H. Kordower, A. J. Stoessl, R. E. Burke, P. Brundin, Z. Yue,S. T. Brady, J. Milbrandt, B. D. Trapp, T. B. Sherer, and S. Medicetty, “IsAxonal Degeneration a Key Early Event in Parkinson’s Disease?,” Journul ozdurkinson's Disyusy, vol. 6, pp. 703–707, oct 2016.[142] J. L. W. Bosboom, D. Stoffers, and E. C. Wolters, “Cognitive dysfunction anddementia in Parkinson?s disease,” Journul oz byurul hrunsmission, vol. 111,pp. 1303–1315, oct 2004.[143] C. Delaville, P. D. Deurwaerde`re, and A. Benazzouz, “Noradrenaline andParkinson’s Disease,” Frontiyrs in gystyms byuroswiynwy, vol. 5, p. 31, 2011.[144] P. Huot, S. H. Fox, and J. M. Brotchie, “The serotonergic system in Parkin-son’s disease,” drogryss in byuroviology, vol. 95, pp. 163–212, oct 2011.[145] A. J. Hughes, S. E. Daniel, S. Blankson, and A. J. Lees, “A ClinicopathologicStudy of 100 Cases of Parkinson’s Disease,” Urwhivys oz byurology, vol. 50,pp. 140–148, feb 1993.[146] R. A. Hauser, P. Auinger, and D. Oakes, “Levodopa response in early Parkin-son’s disease,” aovymynt Disorxyrs, vol. 24, pp. NA–NA, dec 2009.164[147] B. Picconi, L. F. Herna´ndez, J. A. Obeso, and P. Calabresi, “Motor com-plications in Parkinson’s disease: Striatal molecular and electrophysiologicalmechanisms of dyskinesias,” jul 2018.[148] R. De La Fuente-Fernndez, J. Q. Lu, V. Sossi, S. Jivan, M. Schulzer, J. E.Holden, C. S. Lee, T. J. Ruth, D. B. Calne, and A. Jon Stoessl, “Biochemicalvariations in the synaptic level of dopamine precede motor fluctuations inParkinson’s disease: PET evidence of increased dopamine turnover,” Unnulsoz byurology, vol. 49, no. 3, pp. 298–303, 2001.[149] V. Sossi, R. de la Fuente-Ferna´ndez, M. Schulzer, J. Adams, and J. Stoessl,“Age-related differences in levodopa dynamics in Parkinson’s: implications formotor complications,” Bruin, vol. 129, pp. 1050–1058, apr 2006.[150] K. Kannari, H. Yamato, H. Shen, M. Tomiyama, T. Suda, and M. Matsunaga,“Activation of 5-HT1A but not 5-HT1B receptors attenuates an increase inextracellular dopamine derived from exogenously administered l-DOPA in thestriatum with nigrostriatal denervation,” Journul oz byurowhymistry, vol. 76,pp. 1346–1353, dec 2001.[151] M. Lundblad, S. af Bjerke´n, M. A. Cenci, F. Pomerleau, G. A. Gerhardt,and I. Stro¨mberg, “Chronic intermittent L-DOPA treatment induces changesin dopamine release,” Journul oz byurowhymistry, vol. 108, pp. 998–1008, feb2009.[152] J. Logan, J. S. Fowler, N. D. Volkow, G.-J. Wang, Y.-S. Ding, and D. L.Alexoff, “Distribution Volume Ratios without Blood Sampling from GraphicalAnalysis of PET Data,” Journul oz Wyryvrul Bloox Flow : aytuvolism, vol. 16,pp. 834–840, sep 1996.[153] E. D. Abercrombie, A. E. Bonatz, and M. J. Zigmond, “Effects of l-DOPA onextracellular dopamine in striatum of normal and 6-hydroxydopamine-treatedrats,” Bruin fysyurwh, vol. 525, pp. 36–44, aug 1990.[154] J. Tedroff, M. Pedersen, S. M. Aquilonius, P. Hartvig, G. Jacobsson, andB. L˚angstro¨m, “Levodopa-induced changes in synaptic dopamine in patientswith Parkinson’s disease as measured by [11C]raclopride displacement andPET,” byurology, vol. 46, pp. 1430–1436, may 1996.165[155] D. B. Calne and M. J. Zigmond, “Compensatory Mechanisms in DegenerativeNeurologic Diseases,” Urwhivys oz byurology, vol. 48, p. 361, apr 1991.[156] S. J. Kish, K. Shannak, A. Rajput, J. H. N. Deck, and O. Hornykiewicz, “Ag-ing Produces a Specific Pattern of Striatal Dopamine Loss: Implications forthe Etiology of Idiopathic Parkinson’s Disease,” Journul oz byurowhymistry,vol. 58, pp. 642–648, feb 1992.[157] H. Bernheimer, W. Birkmayer, O. Hornykiewicz, K. Jellinger, and F. Seitel-berger, “Brain dopamine and the syndromes of Parkinson and Huntington.Clinical, morphological and neurochemical correlations.,” Journul oz thy nyuArologiwul swiynwys, vol. 20, pp. 415–55, dec 1973.[158] M. Kasten and C. Klein, “The many faces of alpha-synuclein mutations,”aovymynt Disorxyrs, vol. 28, pp. 697–701, jun 2013.[159] H. Wilson, G. Dervenoulas, G. Pagano, C. Koros, T. Yousaf, M. Pi-cillo, S. Polychronis, A. Simitsi, B. Giordano, Z. Chappell, B. Corcoran,M. Stamelou, R. N. Gunn, M. T. Pellecchia, E. A. Rabiner, P. Barone, L. Ste-fanis, and M. Politis, “Serotonergic pathology and disease burden in the pre-motor and motor phase of A53T -synuclein parkinsonism: a cross-sectionalstudy.,” hhy LunwytB byurology, vol. 18, pp. 748–759, aug 2019.[160] H. F. Wehrl, M. Hossain, K. Lankes, C.-C. Liu, I. Bezrukov, P. Martirosian,F. Schick, G. Reischl, and B. J. Pichler, “Simultaneous PET-MRI revealsbrain function in activated and resting state on metabolic, hemodynamic andmultiple temporal scales,” butury ayxiwiny, vol. 19, pp. 1184–1189, sep 2013.166

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            data-media="{[{embed.selectedMedia}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0391842/manifest

Comment

Related Items