UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Near infrared spectroscopy : novel signal processing methods and applications Molavi, Behnam 2013

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

Item Metadata


24-ubc_2013_fall_molavi_behnam.pdf [ 5.44MB ]
JSON: 24-1.0103293.json
JSON-LD: 24-1.0103293-ld.json
RDF/XML (Pretty): 24-1.0103293-rdf.xml
RDF/JSON: 24-1.0103293-rdf.json
Turtle: 24-1.0103293-turtle.txt
N-Triples: 24-1.0103293-rdf-ntriples.txt
Original Record: 24-1.0103293-source.json
Full Text

Full Text

Near Infrared Spectroscopy: Novel Signal ProcessingMethods and ApplicationsbyBehnam MolaviB.Sc. Electrical Engineering , Iran University of Science and Technology, 2004M.Sc. Electrical Engineering , Iran University of Science and Technology, 2007A THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFDoctor of PhilosophyinTHE FACULTY OF GRADUATE AND POSTDOCTORALSTUDIES(Electrical and Computer Engineering)The University Of British Columbia(Vancouver)October 2013c? Behnam Molavi, 2013AbstractOxygen is a critical component in living organisms and its concentration in tissue isan important parameter indicative of tissue metabolism, level of activity and healthcondition. As a result, measuring oxygen concentration in the tissue is essentialin many clinical and research applications. Near Infrared Spectroscopy (NIRS)is a non invasive method of measuring tissue oxygenation using diffusion of lightin the tissue. NIRS as a safe, non invasive and low cost monitoring technologyhas been used in a wide range of applications including monitoring muscle andbrain oxygenation, brain computer interface and rehabilitation. The motivation forthis thesis has been to develop new signal processing methods and to investigatepotential new applications for NIRS.One major characteristic of NIRS is its sensitivity to movement of the targettissue during the measurement. The effects of movements, known as motion ar-tifacts, have limited clinical applications of NIRS in ambulant patients as well asexperimental applications of NIRS monitoring in areas such as exercise scienceand sports medicine. In this thesis, we present a new method of reducing the effectof motion artifacts on NIRS signal using Discrete Wavelet Transform (DWT).One of the areas of application which can significantly benefit from reductionof motion artifacts is NIRS-based wearable sensors. In particular, a potential andunexplored application of NIRS is providing a monitoring method for people withbladder control problems, which occurs in a variety of conditions including spinalcord injury and stroke. We investigate the application of NIRS for detection ofbladder filling to capacity using a wearable wireless monitoring sensor which canbe used to warn the subject once the bladder content reaches a predefined percent-age of the full capacity.iiNIRS can be used as a functional neuroimaging method to identify brain acti-vations during practice of a motor/cognitive task. One important question in thisfield is how the activated brain areas are interconnected. We thus investigate theuse of phase information in NIRS channels to identify cortical connections and inparticular, show the applicability of this approach in identifying language networkin human infants.iiiPrefaceSome of the methods and results presented in this thesis have been published orhave been submitted for consideration for publication as journal or conference pro-ceedings articles. The list of the publications can be found below.The material in Chapter 3 was published with preliminary results and at dif-ferent stages of development using different data types in Canadian Conferenceon Electrical and Computer Engineering in 2010 [1], the Proceedings of the In-ternational IEEE EMBS Conference in 2010 [2] and proceedings of SPIE in 2011[3]. A more complete version was published in the IOP Journal of Physiologi-cal Measurement [4]. This article was featured in the ?Highlights of 2012? of theJournal of Physiological Measurement. The material in Chapter 4 was acceptedfor publication in the IEEE Transactions on Biomedical Circuits and Systems andis also under review by University of British Columbia University-Industry LiaisonOffice (UILO) for potential intellectual property (IP) protection and licensing. Apreliminary version of the method and results presented in Chapter 5 was publishedin the Proceedings of the International IEEE EMBS Conferences in 2011 [5] and2012 [6]. A modified and more detailed version of this work was submitted and iscurrently under review for consideration for publication.The research in this thesis was conducted with approval of the Research EthicsBoard (CREB) of the University of British Columbia (approval number H02-80575).The publications resulting from this thesis are as follows:Journal Articles? B. Molavi and G. A. Dumont. Wavelet-based motion artifact removal forfunctional near-infrared spectroscopy. Physiological measurement, 33(2):iv259-270, 2012. ([4])? B. Molavi, B. Shadgan, A. Macnab, G. A. Dumont. Non-invasive opticalmonitoring of bladder filling to capacity using a wireless NIRS device. IEEETransactions on Biomedical Circuits and Systems, 2013, In press.? B. Molavi, L. May, J. Gervain, J. F. Werker, G. A. Dumont. Analyzing rest-ing state functional connectivity in the language system using near infraredspectroscopy. Submitted, 2013Refereed Conference Papers? B. Molavi, J. Gervain, G. A. Dumont, H. A. Noubari. Functional connectiv-ity analysis of cortical networks in Functional Near Infrared Spectroscopyusing phase synchronization. 34th Annual International Conference of theIEEE Engineering in Medicine and Biology Society (EMBS), pp. 5182?5185, 2012. ([6])? B. Molavi, J. Gervain, and G. A. Dumont. Estimating cortical connectivityin functional near infrared spectroscopy using multivariate autoregressivemodeling. 33rd Annual International Conference of the IEEE Engineeringin Medicine and Biology Society (EMBS), pp. 2334?2337, 2011. ([5])? B. Molavi, G. A. Dumont, B. Shadgan, and A. J. Macnab. Attenuationof Motion Artifact in Near Infrared Spectroscopy Signals Using a WaveletBased Method. Proceedings of SPIE 7890, 2011. ([3])? B. Molavi and G. A. Dumont. Wavelet based motion artifact removal forFunctional Near Infrared Spectroscopy. 32nd Annual International Confer-ence of the IEEE Engineering in Medicine and Biology Society (EMBS), pp.5-8, 2010. ([2])? B. Molavi, G. A. Dumont, and B. Shadgan, ?Motion artifact removal frommuscle NIR Spectroscopy measurements,? in IEEE Canadian Conference onElectrical and Computer Engineering, pp. 1-4, 2010. ([1])vTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiList of Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiiiAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvi1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Introduction to Near Infrared Spectroscopy . . . . . . . . . . . . 11.2 Theory of NIRS . . . . . . . . . . . . . . . . . . . . . . . . . . . 41.2.1 Modified Beer-Lambert Law . . . . . . . . . . . . . . . . 41.2.2 Photon Diffusion in Tissue . . . . . . . . . . . . . . . . . 61.3 Applications of NIRS . . . . . . . . . . . . . . . . . . . . . . . . 71.4 NIRS Instrumentation . . . . . . . . . . . . . . . . . . . . . . . . 91.5 Safety Considerations . . . . . . . . . . . . . . . . . . . . . . . . 111.6 Limitations of NIRS . . . . . . . . . . . . . . . . . . . . . . . . 121.6.1 Penetration Depth and Spatial Resolution . . . . . . . . . 121.6.2 Light Coupling . . . . . . . . . . . . . . . . . . . . . . . 121.6.3 Interferences . . . . . . . . . . . . . . . . . . . . . . . . 13vi1.7 Motivation for this Thesis . . . . . . . . . . . . . . . . . . . . . . 151.8 Thesis Contributions . . . . . . . . . . . . . . . . . . . . . . . . 181.9 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202.1 Interference Reduction in NIRS . . . . . . . . . . . . . . . . . . 202.2 NIRS in Urology . . . . . . . . . . . . . . . . . . . . . . . . . . 232.3 Functional Brain Connectivity Using functional Near Infrared Spectroscopy(fNIRS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252.4 NIRS for Monitoring TMS . . . . . . . . . . . . . . . . . . . . . 272.5 Summary and Conclusion . . . . . . . . . . . . . . . . . . . . . . 293 Motion Artifact Removal . . . . . . . . . . . . . . . . . . . . . . . . 323.1 Artifact Removal Using Discrete Wavelet Transform . . . . . . . 333.2 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363.2.1 Performance Evaluation . . . . . . . . . . . . . . . . . . 383.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 393.3 Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403.3.1 Performance Evaluation . . . . . . . . . . . . . . . . . . 423.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 433.4 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . 464 Bladder Filling Monitoring with NIRS . . . . . . . . . . . . . . . . . 524.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524.2 Detection of Bladder Filling to Capacity Using NIRS . . . . . . . 534.2.1 Electronics . . . . . . . . . . . . . . . . . . . . . . . . . 554.2.2 Firmware . . . . . . . . . . . . . . . . . . . . . . . . . . 584.2.3 PC Interface . . . . . . . . . . . . . . . . . . . . . . . . 604.3 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . 614.4 In Vitro Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 624.4.1 In Vitro Setup . . . . . . . . . . . . . . . . . . . . . . . . 624.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 634.5 In Vivo Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 644.5.1 Materials and Method . . . . . . . . . . . . . . . . . . . 64vii4.5.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 644.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 684.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 715 Cortical Connectivity Analysis . . . . . . . . . . . . . . . . . . . . . 725.1 Functional Connectivity Using Multivariate Autoregressive Mod-eling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 735.1.1 Materials and Method . . . . . . . . . . . . . . . . . . . 735.1.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 785.2 Analyzing Resting State Functional Connectivity in The HumanLanguage System Using Near Infrared Spectroscopy . . . . . . . 805.2.1 Material and Methods . . . . . . . . . . . . . . . . . . . 825.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 875.3 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . 916 Custom fNIRS System Design . . . . . . . . . . . . . . . . . . . . . . 956.1 Background and Motivation . . . . . . . . . . . . . . . . . . . . 956.2 Instrument Design . . . . . . . . . . . . . . . . . . . . . . . . . . 966.2.1 Light Sources . . . . . . . . . . . . . . . . . . . . . . . . 976.2.2 Source Modulation . . . . . . . . . . . . . . . . . . . . . 976.2.3 Light Detection . . . . . . . . . . . . . . . . . . . . . . . 986.2.4 Digital Lock-in Amplification . . . . . . . . . . . . . . . 986.2.5 User Interface . . . . . . . . . . . . . . . . . . . . . . . . 1016.3 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . 1016.4 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1026.4.1 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 1026.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 1036.5 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . 1067 Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . . 1087.1 Motion Artifact Removal from fNIRS . . . . . . . . . . . . . . . 1087.2 Wireless NIRS for Monitoring Bladder Contents . . . . . . . . . . 1107.3 fNIRS Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . 1137.4 Custom-made fNIRS Device for TMS Monitoring . . . . . . . . . 115viiiBibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117ixList of TablesTable 3.1 Normalized Mean Squared Error between processed signal andoriginal artifact-free signal . . . . . . . . . . . . . . . . . . . 40Table 3.2 Median of Normalized Mean Squared Error (NMSE) and At-tenuation in artifacts energy (in dB) for proposed method . . . 45Table 4.1 System level parameters of the sensor . . . . . . . . . . . . . . 67Table 5.1 Pairwise comparison between selected ROIs inside and outsidelanguage network (Tukey?s test). . . . . . . . . . . . . . . . . 89xList of FiguresFigure 1.1 Absorption spectra for human tissue . . . . . . . . . . . . . . 3Figure 1.2 Typical fNIRS data power spectrum . . . . . . . . . . . . . . 14Figure 3.1 Simulated comparison of artifact removal methods . . . . . . 37Figure 3.2 Comparison of power spectra after artifact removal . . . . . . 38Figure 3.3 Coherence between original and processed signal . . . . . . . 39Figure 3.4 Experiment setup . . . . . . . . . . . . . . . . . . . . . . . . 41Figure 3.5 Optodes placement . . . . . . . . . . . . . . . . . . . . . . . 42Figure 3.6 Artifact removal example . . . . . . . . . . . . . . . . . . . . 43Figure 3.7 Artifact attenuation . . . . . . . . . . . . . . . . . . . . . . . 44Figure 3.8 Comparison with wavelet denoising . . . . . . . . . . . . . . 46Figure 3.9 Comparison with adaptive wavelet denoising . . . . . . . . . 47Figure 3.10 Artifact power attenuation versus NMSE for 3 subjects . . . . 48Figure 3.11 Correlation with artifact intensity . . . . . . . . . . . . . . . 49Figure 3.12 Performance comparison with common NIRS artifact removalmethods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50Figure 4.1 Sensor block diagram. . . . . . . . . . . . . . . . . . . . . . 55Figure 4.2 Sensor components . . . . . . . . . . . . . . . . . . . . . . . 56Figure 4.3 Timing of sampling. . . . . . . . . . . . . . . . . . . . . . . 58Figure 4.4 Snapshot of the LED driving current. . . . . . . . . . . . . . 59Figure 4.5 Detector?s signal . . . . . . . . . . . . . . . . . . . . . . . . 60Figure 4.6 PC user interface screen shot. . . . . . . . . . . . . . . . . . 61Figure 4.7 In vitro setup. . . . . . . . . . . . . . . . . . . . . . . . . . . 63xiFigure 4.8 In vitro data. . . . . . . . . . . . . . . . . . . . . . . . . . . 64Figure 4.9 Sensor placement for in vivo device test. . . . . . . . . . . . . 65Figure 4.10 Voiding result. . . . . . . . . . . . . . . . . . . . . . . . . . 65Figure 4.11 Comparison of detected light attenuation in full and emptybladders. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66Figure 4.12 Voiding (reference NIRS). . . . . . . . . . . . . . . . . . . . 67Figure 5.1 Experimental design. . . . . . . . . . . . . . . . . . . . . . . 76Figure 5.2 fNIRS optode locations . . . . . . . . . . . . . . . . . . . . . 77Figure 5.3 Connectivity matrices. . . . . . . . . . . . . . . . . . . . . . 79Figure 5.4 Temporal evolution of connections. . . . . . . . . . . . . . . 80Figure 5.5 fNIRS experiment setup. . . . . . . . . . . . . . . . . . . . . 82Figure 5.6 Optode placement on the head. . . . . . . . . . . . . . . . . . 83Figure 5.7 Signal recorded from two channels with high degree of con-nection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88Figure 5.8 Group level RSFC maps . . . . . . . . . . . . . . . . . . . . 89Figure 5.9 The connectivity map for the 2 subgroups . . . . . . . . . . . 90Figure 5.10 Correlation between connection strengths in the two subgroups. 90Figure 5.11 ROI connections comparison. . . . . . . . . . . . . . . . . . 91Figure 5.12 Lateralization index for the language network. . . . . . . . . . 91Figure 6.1 Block diagram of the overall fNIRS system. . . . . . . . . . . 96Figure 6.2 Block diagram of the laser diode driver module. . . . . . . . . 97Figure 6.3 Direct digital synthesizer block diagram. . . . . . . . . . . . 98Figure 6.4 In phase-quadrature demodulation scheme. . . . . . . . . . . 99Figure 6.5 Arterial occlusion test. . . . . . . . . . . . . . . . . . . . . . 104Figure 6.6 Isometric contraction results. . . . . . . . . . . . . . . . . . . 105Figure 6.7 Total hemoglobin changes in response to motor activity . . . . 105Figure 6.8 Block averaged hemodynamic response. . . . . . . . . . . . . 106xiiList of AbbreviationsADC Analog to Digital ConverterAMT Active Motor ThresholdAPD Avalanche PhotodiodeAR AutoregressiveAUC Area Under the CurveBOLD Blood Oxygen-Level DependentCBSI Correlation-based Signal ImprovementCW Continuous WaveDAC Digital to Analog ConverterDAQ Data AcquisitionDC Directed CoherenceDDS Direct Digital SynthesisDPF Differential Pathlength FactorDWT Discrete Wavelet TransformEEG ElectroencephalographyFD Frequency DomainxiiifMRI functional Magnetic Resonance ImagingfNIRS functional Near Infrared SpectroscopyFWT Fast Wavelet TransformGLM Generalized Linear ModelGUI Graphical User InterfaceHRF Hemodynamic Response FunctionLD Laser DiodeLED Light Emitting DiodeLUTS Lower Urinary Tract SymptomsMAD Median Absolute DeviationMBLL Modified Beer-Lambert LawMCU Microcontroller UnitMEG MagnetoencephalographyMEP Motor Evoked PotentialMRI Magnetic Resonance ImagingMS Multiple SclerosisMVAR Multivariate AutoregressiveNCDF Normal Cumulative Distribution FunctionNEP Noise Equivalent PowerNIRS Near Infrared SpectroscopyNMSE Normalized Mean Squared ErrorOD Optical DensityxivPC Personal ComputerPCA Principal Component AnalysisPCB Printed Circuit BoardPET Positron Emission TomographyPFS Pressure Flow StudyPM Premotor areaPPG PhotoplethysmogrophyRMS Root Mean SquareROI Regions Of InterestRSFC Resting State Functional ConnectivityrTMS repetitive Transcranial Magnetic StimulationSNR Signal to Noise RatioSAD Sum of Absolute DifferencesSPI Serial Peripheral InterfaceSURE Stein?s Unbiased Risk EstimatorTD Time DomainTIA Transimpedance AmplifierTIWT Translation Invariant Wavelet TransformTMS Transcranial Magnetic StimulationTOI Tissue Oxygenation IndexUTI Urinary Tract InfectionxvAcknowledgmentsThis dissertation could not have been completed without the support and encour-agements I received from many people.First and foremost, I would like to offer my sincere gratitude to my supervisor,Prof. Guy A. Dumont, for his continuous guidance, advice and patience throughoutmy PhD studies. He has not only been a caring and dedicated supervisor, but alsoa great supportive mentor allowing me to pursue my ideas and interests.I am extremely grateful to my co-supervisor, Prof. Janet F. Werker for herwonderful support and guidance. I consider myself very fortunate for having theopportunity to work with her and benefit from her valuable experience and ex-pertise. I admire her for being so dedicated, supportive and caring towards herstudents.I would also like to thank my PhD supervisory and examining committee mem-bers, Prof. Matthew Yedlin, Prof. Darlene Reid, Prof. Edmond Cretu and Prof.Jane Wang for dedicating their valuable time and effort to reviewing this thesis andproviding advice and feedback which has contributed significantly to the improve-ment of this thesis.I would like to extend my gratitude to my external examiner, Prof. MartinWolf for his invaluable feedback and comments on the thesis. His expert adviceand suggestions had a significant role in improving this work.Throughout this project, I had the opportunity to work with a number of groupsand researchers. It was my great pleasure to collaborate with Prof. Andrew Mac-nab, who generously shared his experience and expertise in the field of biomedicaloptics with me. I would like to especially thank Dr. Babak Shadgan for offering hisvaluable clinical and technical expertise, advice and mentorship in all stages of myxvistudies. I would also like to thank Prof. Hossein A. Noubari and Prof. Lara Boydfor their support and guidance. I wish to thank Prof. Lukas Chrostowski and Prof.David Jones for offering their technical advice, support and fruitful discussions ondevelopment of the optical systems used in this dissertation.I owe special thanks to Dr. Henny Yeung, Dr. Judit Gervain, Dr. Krista Byers-Heinlein, Lilian May and Alison Gruel for their help and support in collecting theexperimental data, providing feedback and their contribution in developing meth-ods and ideas presented in Chapter 3, Chapter 5 and the rest of the thesis.Many thanks to my great friends and colleagues at the University of BritishColumbia, Ali Shahidi Zandi, Pedram Ataee, Christopher Brouse, Masoud Rahjoo,Mehrnoush Javadi, Sara Khosravi, Prasad Shrawan, Parastoo Dehkordi, Reza Taf-reshi, Ping Yang, Mande Leung, Walter Karlen, Jin oh Hahn and Klaske van Heud-sen who provided a friendly environment, kept me in good company and made mygraduate studies such a wonderful and rewarding experience.I would also like to express my sincere appreciation to my brother, Dr. BehzadMolavi for his encouragements, inspirations and support in pursuing my PhD.My sincere gratitude to my parents, whose never-ending love has given meinspiration and hope in every step of my life.Finally, I wish to express my deepest appreciation to my wife and my bestfriend, Zamzam Kordi without whom none of this would have been possible. Herunconditional love and support at times of despair and frustration gave me thecourage, hope and motivation to continue this journey.xviiDedicationTo my parents, and my wife ...xviiiChapter 1Background1.1 Introduction to Near Infrared SpectroscopyOxygen is a critical component in living organisms and its concentration in tissue isan important parameter indicative of tissue metabolism, level of activity and healthcondition. As a result, measuring oxygen concentration in the tissue is essential inmany clinical and research applications.In 1977 Jobsis introduced the method of measuring local tissue oxygenationnon-invasively using near infrared light [7]. Absorption of the light by living tissueis lower in this wavelength range (600-1000 nm) which results in maximum pen-etration depth. Below this wavelength range, hemoglobin strongly absorbs lightand above this range, water is the major absorber. Figure 1.1 shows the absorptionspectra of light energy as a function of wavelength in the vicinity of NIR wave-lengths for important species in tissue. Within this range, light attenuation occursas a result of scattering and absorption by chromophores such as oxygenated anddeoxygenated hemoglobin (HbO2 and HHb) which are important biological in-dicators, lipid, water and cytochrome oxidase. This method of using NIR lightto measure chromophore concentration (or concentration changes) in the tissue isknown as Near Infrared Spectroscopy (NIRS). In practice, a light point source isplaced on the tissue surface to shine light into the tissue. A light detector, also lo-cated on the surface, but at a distance from the source detects diffusively reflectedlight from within the tissue and extracts information on the chromophore concen-1CHAPTER 1. BACKGROUNDtration from the detected light intensity and phase. The intensity attenuation is dueto absorption by blood chromophores as well as scattering and can be related to thechanges in concentration of the chromophores.The unique type of information provided by NIRS is quite different from thatobtained with pulse oximetry or Photoplethysmogrophy (PPG), even though thebasic principles are similar. NIRS and PPG are based on the same principle thatNIR light can penetrate in the tissue and is mostly absorbed by hemoglobin species.However, PPG measures arterial oxygen saturation which is the percentage ofhemoglobin in arteries that is bound to oxygen:SpO2 =aHbO2aHbO2 +aHHb(1.1)where SpO2 is the peripheral oxygen saturation, and aHbO2 and aHHb are the ar-terial HbO2 and HHb concentrations, respectively. This fraction which is normallyclose to 100%, is a centrally controlled parameter and only changes in response tocritical conditions.NIRS measures tissue oxygenation changes or in some cases, tissue oxygensaturation which is the percentage of hemoglobin bound with oxygen in the tis-sue. This parameter can change significantly in response to increase or decrease inoxygen demand in the tissues. For example, it has been shown that under an in-cremental inspiratory threshold loading study, the SpO2 stayed relatively constantthroughout the experiment (? 96%), while NIRS showed decrease in tissue oxy-genation of a non active muscle and increase in oxygenation of the active muscle[8].In general, NIRS can be employed in any application where hemodynamicsof a tissue is of importance. Because of its non-invasive nature and the uniquetype of information it provides about tissue hemodynamics, NIRS has become verypopular and has found a wide range of applications such as monitoring muscle orbrain oxygenation [9], brain computer interface [10], rehabilitation [11] and cancerdetection [12].Brain functional studies are among the application areas in which NIRS ispromising. Due to neurovascular coupling, the neural activation which is accom-panied by hemodynamic changes resulting from an increase in oxygen demand,2CHAPTER 1. BACKGROUND650 700 750 800 850 900 950 1000 105010?3Wavelength (nm)Extinction coefficient (mm?1  ?M?1 )  00.05Water absorption ? a (mm?1 )O2HbHHbWaterFigure 1.1: Absorption spectra of oxygenated and deoxygenated hemoglobin(HbO2 and HHb) and water from 600nm to 1000nm. The absorptionincreases significantly below 600nm and above 1000nm.can be detected using NIRS in superficial areas of the brain [13]. This method,known as fNIRS, is widely used to detect activations in brain cortex in responseto different stimulations. The information recorded by functional Near InfraredSpectroscopy (fNIRS) is very similar to that obtained by functional Magnetic Res-onance Imaging (fMRI) and Positron Emission Tomography (PET). However,fNIRS is less expensive, portable, non ionizing, non restraining and has highertemporal resolution. Also, the process of hemodynamic changes detection is dif-ferent between fMRI and NIRS. In fMRI, magnetic properties of hemoglobin areused to detect changes in HHb [14]. Losing oxygen causes hemoglobin to demon-strate paramagnetic properties. In fNIRS, the difference in absorption spectra ofHbO2 and HHb is the principle of detection. As a result, NIRS is sensitive to bothHbO2 and HHb while fMRI is only sensitive to HHb. This is an important differ-ence as the change in HbO2 in response to stimulation is larger and more correlatedwith Blood Oxygen-Level Dependent (BOLD) response than HHb [15].3CHAPTER 1. BACKGROUND1.2 Theory of NIRS1.2.1 Modified Beer-Lambert LawIn order to derive chromophore concentrations, raw light intensity readings needto be converted to concentrations or concentration changes. The Beer-Lambertlaw describes the attenuation of light when propagating in a non scattering, butabsorbing medium:I = I0e??al (1.2)where I0 is the intensity of the light entering the medium, I is the light intensity atlocation x= l with l in cm and ?a is the attenuation coefficient in cm?1. In presenceof multiple absorbants, this coefficient is related to the absorbants? concentrationas?a = ?i?ici (1.3)where ci and ?i are the concentration and extinction coefficient of the ith absorberin molL?1 and Lmol?1cm?1, respectively. In a highly scattering medium, such ashuman tissue, this equation is no longer valid as absorption is not the only mech-anism resulting in light intensity attenuation. The modified form of this equationknown as the Modified Beer-Lambert Law (MBLL) takes scattering into consid-eration and explains the relationship between chromophores? concentration andreflected optical density [16]:OD =?log II0= ?clB+G (1.4)where OD is the optical density, I0 is the incident light intensity, I is the detectedlight intensity, ? and c are the same as described above, l is the distance betweenwhere the light enters the tissue and where the detected light exits the tissue, Bis a pathlength factor that accounts for increases in the photon pathlength causedby tissue scattering, and G is a factor that accounts for the constant losses such asthose caused by measurement geometry [16].A change in the concentration of the chromophore will result in a change inthe reflected light?s intensity which is sensed at the detector. When concentration4CHAPTER 1. BACKGROUNDchanges, ? and distance l remain constant. Assuming B and G also remain constant,we can rewrite equation 1.4 as?OD =?log IFinalIInitial= ??CLB (1.5)where ?OD = ODFinal ?ODInitial is the change in the optical density, IFinal andIInitial are the measured intensities before and after the change in the chromophoreconcentration, and ?C is the change in concentration. L is determined by the probesgeometry, ? is the property of the chromophore and B is often referred to as theDifferential Pathlength Factor (DPF) and can be determined with very short pulsesof light and has been tabulated for various tissues.In order to consider the contribution of two or more chromophores, we need towrite equation 1.5 for 2 different chromophores and make measurements at morethan 1 wavelength. In this way, OD changes at wavelength ? would be:?OD? =(??HbO2? [HbO2]+ ??HHb? [HHb])B? L (1.6)by measuring the ?OD in 2 wavelengths, one can solve for changes of concentra-tion in HbO2 and HHb using?HHb =??2HbO2?OD?1B?1 ? ??1HbO2?OD?2B?2(??1HHb??2HbO2 ? ??2HHb??1HbO2)L(1.7)?HbO2 =??1HHb?OD?2B?2 ? ??2HHb?OD?1B?1(??1HHb??2HbO2 ? ??2HHb??1HbO2)L(1.8)The MBLL is sufficient in cases where only the measurement of changes froma baseline is desired. This baseline is dependent on different factors (source de-tector coupling with the tissue, tissue parameters etc.). As a result, removing anoptode and replacing it in the exact same position on the tissue, will result in dif-ferent baseline reading and therefore absolute reading values are not comparable.Therefore, if absolute values of chromophores or Tissue Oxygenation Index (TOI)measurements are required, this model will not be adequate.5CHAPTER 1. BACKGROUND1.2.2 Photon Diffusion in TissueA more robust model of light-tissue interaction is based on using diffusion equa-tion to describe light propagation in highly scattering living tissue. Assuming anisotropic source, the light-tissue interaction can be described by the diffusion equa-tion [17]:? ? (D(r)??(r, t))???a(r)?(r, t)???(r, t)? t =??S(r, t) (1.9)where ?(r, t) is the photon fluence rate at position r and time t in Wcm?1,S(r, t) is the source power per volume emitted radially outward in Wcm?1, D(r) =?3(? ?s(r)+?a(r)) is the photon diffusion coefficient, ??s is the reduced scattering coeffi-cient, ?a is the absorption coefficient and ? is the speed of the light in the medium.The reduced scattering coefficient is related to scattering coefficient as ? ?s =(1?g)?s and characterizes the amount of scattering in the tissue.In order to characterize the diffusion of light in the tissue, different source andboundary conditions can be assumed. Here, we briefly look at 2 simple cases whichare relevant here.For the case of a frequency modulated sourceS(r, t) = Sdc(r)+Sac(r)e?i?t (1.10)where Sdc and Sac are the constant and time varying components of the source. Inthis case, the time varying part of the solution would also have the same frequencyas the source and will have the form of?ac(r, t) =U(r)e?i?t (1.11)assuming a homogeneous medium, the diffusion equation reduces to(?2 ? k2)U(r) =? ?DSac(r) (1.12)where k2 = ??a?i?D . For a point source at position r = r in infinite spaceSac(r) = Sac? (r) (1.13)6CHAPTER 1. BACKGROUNDThe general solutions to Equation 1.12 are of the form Cekr and Ce?kr. Onecan then solve Equation 1.12 to get [17]U(r) = ?Sac4piDre?kr (1.14)where k is a complex number with real partkr =(??a2D)1/2??(1+(???a)2)1/2+1??1/2(1.15)and imaginary partki =?(??a2D)1/2??(1+(???a)2)1/2?1??1/2(1.16)Hence, at location r on tissue surface, the light intensity would be frequencymodulated at the same frequency of the source with reduced amplitude and a phaseshift which are both functions of ?a and D.A more realistic boundary condition is a semi-infinite homogeneous medium.In this case, an image source symmetric around the boundary can be used to satisfythe boundary condition. In the limit where we are far enough from the source, thesolution can be written asU(?,z = 0)? ?S04piDe?k??2(2k(ltrzb + z2b))= A0e?kr??2 ei(?ki?+?0) (1.17)which again indicates attenuation along with a phase shift at position r withrespect to the source. In case of a non modulated source, one would get the expo-nential decrease in intensity only. The application of these formulations in derivingconcentration values in different NIRS instruments will be discussed in Section Applications of NIRSThe unique advantages of NIRS has led to its use in a variety of applications. NIRSin general, is applicable in any situation where hemodynamic monitoring of a tissue7CHAPTER 1. BACKGROUNDis required.One of the areas in which NIRS is well suited for are brain studies which in-volve recording activities in different brain areas. In these studies, neurovascularcoupling results in increase in hemodynamics that lags the neural activation. Thischange can be detected by fNIRS. The data is then converted to activation usingmathematical models incorporating the hemodynamic response function. In fNIRSstudies, one would only be interested in measuring changes in chromophore?s con-centration from a baseline and therefore, continuous wave fNIRS devices that mea-sure concentration changes using MBLL are well suited for these applications.fNIRS has been used for studying various brain disorders. Hock et al. haveexamined Alzheimer?s patients during verbal fluency and other cognitive tasks,finding decreases in HbO2 and total hemoglobin (tHb) relative to the baseline inthe parietal lobe as compared to an increase in HbO2 and tHb in healthy subjects[18, 19]. Schizophrenia patients have also been studied with fNIRS in several stud-ies. In one study, for example, ?dysregulated? patterns of HbO2 and HHb changewere found using fNIRS in frontal regions of schizophrenic patients as comparedto healthy subjects during a mirror drawing task [20]. Also, it was shown that thetypical pattern of right-lateralized activation during a continuous performance testwas absent in schizophrenic patients [21]. Depression is another condition exam-ined with NIRS. In a study by Matsuo, reduced frontal activation during a verbalfluency test in patients suffering from depression was found compared to controls[22]. These are only a few representative examples of applications of fNIRS inbrain disease conditions. A more detailed review of such applications can be foundelsewhere [23].One of the attractive areas of application for NIRS is brain studies in infantswhere use of other modalities such as fMRI or PET is either impractical or verydifficult. Also, the small head size and thin skull results in high quality and betterpenetration into brain tissue. For example, unique features of fNIRS has madeit possible to assess newborn infant?s brain response to different languages andlanguage structures [24]. NIRS has also been used to study developmental changesin the cerebral hemodynamic response to different stimulation types [25, 26].One of the more recent developments in applications of NIRS has been thestudy of bladder muscle hemodynamics during voiding to diagnose bladder ob-8CHAPTER 1. BACKGROUNDstruction/dysfunction non-invasively [27]. It has been shown that patterns of HbO2and HHb are different for normal and obstructed bladder reflecting variations in de-trusor muscle hemodynamics and oxygen supply and demand. This can potentiallyreplace invasive methods currently in practice for urodynamics tests.New applications for NIRS are constantly being developed and tested. Valida-tion of such data in different fields has turned NIRS into a promising diagnostictool.1.4 NIRS InstrumentationThere is a wide variety of NIRS instruments currently in use, both commercial([28?33]) and custom-built ([34?38]). A detailed review of different NIRS devicesand measurement techniques can be found elsewhere [23, 39, 40].In terms of operation basis, the NIRS devices can be divided into 3 categories:Continuous Wave (CW), Time Domain (TD) and Frequency Domain (FD) devices[41].TD systems use very short light pulses which are scattered and absorbed bythe tissue layers. The temporal distribution of photons as they leave the tissue isdetected by the system. The shape of this distribution gives information about thetissue scattering and absorption. TD systems provide very accurate estimations oftissue parameters (scattering coefficient ?s and absorbing coefficient ?a). However,they are usually expensive, bulky and require extensive knowledge of the systemto work with and are not particularly suitable for practical clinical applications.In FD systems, the light source is amplitude modulated at intermediate fre-quencies (around 100MHz). As discussed in Section 1.2, the amplitude and phaseof the light at the detector is a function of reduced scattering coefficient ? ?s and ab-sorption coefficient ?a. Hence by measuring amplitude and phase of the modulatedlight at the detector, the chromophore concentrations can be calculated.CW systems are the most widely used NIRS devices in practice. They radiatelight continuously into the tissue and measure the amplitude decay of the reflectedsignal. In the simple basic form, these systems can not quantify baseline absorp-tion and scattering. However, they can be made with inexpensive, widely availablecomponents. Even though CW systems are often used for measuring slow hemo-9CHAPTER 1. BACKGROUNDdynamic signals [42], they can measure the fast neuronal signals under certainconditions [43?45]. These devices usually incorporate the MBLL model for es-timating the absorption coefficient in the tissue. One of the major limitations ofusing CW NIRS devices with single source-detector pair based on MBLL is thatthey can only measure ?changes? in tissue hemodynamics. In other words, the ab-solute values of chromophores can not be measured. This means a dynamic shouldbe generated in order to observe the changes. This dynamic could be in differentforms including stimulation (brain studies), muscle activity (sports medicine, blad-der study), occlusion or drug administration. Even though some applications arestill feasible using measurement of changes from baseline only, this will rule outpossibility of using NIRS in many other areas. Therefore, several methods havebeen proposed to alleviate this problem in CW systems. Most of these techniquesuse a more realistic model for light-tissue interaction to estimate the tissue opticalparameters and hence, the concentration.A sub class of CW devices use the diffusion model to estimate parameterssuch as tissue oxygen saturation index. If we assume a non modulated source inSection 1.2, the detected amplitude at position ? would be:U(?,z = 0) = A0e?kr??2 (1.18)Therefore, if signals are recorded at multiple detectors with different ? , one canestimate the slope of ln?2U(?) and derive ?a, given some simplifying assumptionson ?s (see [46] for example). Such devices are also known as spatially resolvedNIRS and are capable of measuring absolute values of tissue oxygen saturation.Laser Diodes (LDs) and Light Emitting Diodes (LEDs) are the two commonlyused discrete wavelength light sources for NIRS. LDs have narrower spectral widththan LEDs, however, LEDs are cheaper and do not have stability issues of laserdiodes.Common detectors used in NIRS instruments are silicon PIN photodiodes,Avalanche Photodiode (APD)s and PM T! (PM T!). Silicon photodiodes are in-expensive detectors with high quantum efficiency, however, they are not very sen-sitive and are best for applications where detected light levels are high. APDs aremore sensitive and they provide internal gain through internal avalanche multipli-10CHAPTER 1. BACKGROUNDcation, but they are more expensive. PMTs have very high gain-bandwidth productand are used when very high sensitivity is required.Selection of source wavelengths in a NIRS device is an issue that can affectthe signal quality and is an important topic of interest [17]. Traditionally, fordual wavelength systems, wavelengths have been chosen in such a way that oneis above and the other one is below the isosbestic point. Isosbestic point is thewavelength at which HbO2 and HHb have equal absorption in the NIR window(around 800 nm). However, it has been shown that some specific combinations ofwavelengths may result in better separation of HbO2 and HHb and less cross talkbetween chromophores as a result of solving Equation 1.3 [47]. Optimal selectionof wavelengths has been a subject of study with dual [48?50] and multi-wavelengthNIRS configurations [51, 52].1.5 Safety ConsiderationsNIRS is a relatively safe optical method as it uses non-ionizing radiation, is gen-erally non-invasive and uses low power light radiation to examine the tissue. Themajor safety concerns for the NIRS light power are damage to the eye and skintissue. In this particular wavelength range, tissue heating is the major process ofconcern that can potentially lead to tissue damage. This process is more of a con-cern for eye as the light is focused on the retina by the lens which can increase thepotential hazards. There are many factors that affect the potential of light for caus-ing damage to the tissue including the source type (e.g. coherent vs non-coherent),source power, exposure time, wavelength and the beam spot size. The safety powerlimits for laser light sources can be found in guidelines such as IEC 60825-1 andANSI Z136 standards [53]. Similar guidelines for LED based devices are definedseparately in IEC 62471 [54]. The safety of LED based NIRS devices that are indirect contact with the tissue have also been investigated in terms of tissue heatingcaused by the radiance as well as conducted heat from the semiconductor junction[55]. As a rough estimate, the average power level of the LED or laser of less than10 mW in adults is normally considered safe [56]. Most commercial devices usepower levels of 0.5-1.5 mW which is safe even for application on newborn infants[57].11CHAPTER 1. BACKGROUND1.6 Limitations of NIRSSome limitations of NIRS that need to be considered in any application are re-viewed in this section. More detailed review of the limitations of NIRS can befound elsewhere [39, 58, 59].1.6.1 Penetration Depth and Spatial ResolutionHigh scattering in the living tissue causes light to go through a banana shape vol-ume before reaching the receiving optode on the tissue surface. As a result, thesignal detected at the detector reflects a combination of changes in hemodynamicsin the entire sampled volume. The mean penetration depth depends on the source-detector separation as well as the tissue scattering and absorption properties andhas been investigated using theoretical models for light propagation in tissue [60].As a rough estimate, the mean penetration depth or the depth of maximum sensi-tivity of NIRS can be considered to be of order of half interoptode distance [17].A more accurate equation for the penetration depth can be found in [60]. However,it should be noted that the signal still contains interferences from other layers.This also limits the spatial resolution of the NIRS. In some studies, such as sportsmedicine studies, this may not be an issue as the overall change in the muscle tis-sue, for instance, is of interest. However, in fNIRS this spatial resolution may bea limitation as one can not precisely localize activation to very specific small brainareas. However, more general and superficial areas such as dorsolateral prefrontalcortex, superior parietal cortex, and language and primary sensorimotor areas arewithin detectable limits [58].1.6.2 Light CouplingOne of the challenges in NIRS is achieving good, stable optical contact between thetissue and the optodes to get sufficient light levels [58]. One of issues in achievinga good coupling is the optode placement on the tissue. In fNIRS in particular, thisintroduces a challenge and different types of holders have been proposed to addressthis issue as a properly secured optode plays an important role in achieving a goodsignal quality [61]. The specific requirements of an optode holder depend on thesubject populations and therefore, the type of the holder used in a study on adults12CHAPTER 1. BACKGROUNDfor example, would be different from one used in infants? study [57]. In the caseof fNIRS, the optode holder secures the optodes on the tissue to provide a goodcoupling past the hair and reducing the effect of subject?s motion. A layer of haircan attenuate light and block the tissue from the source light [61]. Hair folliclesalso strongly absorb near infrared wavelengths [58]. In adults, care must be taken tomove the hair aside before placement of the optodes. Also, it has been suggestedthat holding the optodes a few millimeters away from the tissue can reduce theeffect of light obstruction by layers of hair and help in achieving a better coupling[61].1.6.3 InterferencesIn NIRS, the hemodynamic changes correlated with activity of the tissue understudy are the parameters of interest. However, NIRS signal can be contaminatedwith interferences with other sources of origin.One major source of unwanted interference is motion. Movements of varioustypes result in distortion of the data stream broadly referred to as motion artifact.Motion can result in data distortion through different mechanisms. For example,movement can cause changes in the local blood circulation. Decoupling of sourceand detector from tissue due to subject?s motion is another common source of ar-tifact. Such artifact is evident as non-physiological signal changes, often of largemagnitude, that commonly result from alterations in the apposition of the NIRSlight source and photo detector to the tissue with resulting alteration in pathlength.The distortion of motion artifact obscures trends in the data that may be relevant,and can compromise meaningful analysis of NIRS monitoring data. Removal ofsuch artifact is often an essential pre-processing step for accurate analysis of thesignal. Motion artifact removal is especially of interest in fNIRS as the level of thesignal of interest is already low and care must be taken to extract as much informa-tion as possible from the recorded data and exclude as few blocks as possible dueto motion contamination. Common approaches along with the approach developedin this thesis for this purpose are presented in Chapter 2 and Chapter 3.Another source of interference in NIRS is the systemic interference. Figure 1.2shows the power spectrum of a typical fNIRS signal segment acquired from the13CHAPTER 1. BACKGROUND0.5 1 1.5 2?80?70?60?50?40Frequency (Hz)Power/frequency (dB/Hz)Welch Power Spectral Density EstimateCardiac pulsationRespirationOther low frequencyFigure 1.2: Power spectrum of fNIRS data collected from forehead at rest.The black arrows indicate systemic interferences at different frequenciesin the fNIRS signal. The peaks around 1.1 Hz and 0.35 Hz are causedby cardiac pulsation and respiration, respectively. The peak at 0.1 Hzmay be related to the heart rate variability and the Mayer waves.forehead of a 30 year old healthy subject at rest condition using the instrumentdescribed in Chapter 6. Two most common interferences are cardiac pulsation andrespiration (shown with arrows in the figure). Cardiac pulsation (the peak around1.1 Hz in Figure 1.2) results in a small diameter change in the blood vessels as aresult of expansion and contraction of the vessels. This will lead to fluctuations inabsorption of light due to changes in blood volume within the sampling volume.This can be seen as a pulsation interference in the NIRS signal and is more evidentin HbO2 compared to HHb. Respiratory interference is related to the respiratoryvariations in arterial pulse pressure and is observed in frequencies between 0.1 and0.4 Hz [62]. Heart rate variability component can also be observed at lower fre-quencies (0.01-0.1 Hz) for longer recorded signals. Mayer waves are also anothersource of interference around 0.1 Hz (see Figure 1.2). These interferences are infact a rich source of information and contain valuable physiological informationwhich can be relevant in many applications. However, in cases where the maininterest is only the changes in tissue oxygenation, it would be essential to removethese interferences to avoid misinterpretation of the NIRS data.A more detailed description of these systemic interferences can be found in[62].14CHAPTER 1. BACKGROUND1.7 Motivation for this ThesisOne factor that has limited experimental applications of NIRS is the sensitivity ofthis method to movement of the tissue of interest during the measurement. This canhappen when, for example the subject moves spontaneously, with involuntary mus-cle contraction, with repositioning of area being monitored, or when the NIRS op-todes are not attached optimally to the tissue surface [40]. The effects of movementhave limited clinical applications of NIRS in ambulant patients as well as experi-mental applications of NIRS monitoring in exercise science and sports medicine.Movements of various types result in distortion of the data stream broadly referredto as motion artifact. Such artifact is evident as non-physiological signal changes,often of large magnitude, that commonly result from alterations in the appositionof the NIRS light source and photo detector to the tissue with resulting alterationin pathlength. The distortion of motion artifact obscures trends in the data thatmay be relevant, and can compromise meaningful analysis of NIRS monitoringdata. Hence, removal of such artifact is potentially of value and is often an essen-tial pre-processing step for accurate analysis of the signal. The importance of thispre-processing step has resulted in introduction and development of several motionartifact treatment methods in recent years [63?67]. Developing an effective and ef-ficient algorithm for removing or reducing motion artifacts in order to improve thequality of NIRS signal has been one of the major motivations for this thesis.One of the more recent applications of NIRS is in urology where the detrusoris being monitored by NIRS for hemodynamic changes. One potential and un-explored application of NIRS is providing a monitoring method for people withbladder control problems. Bladder control problem occurs in a variety of condi-tions including spinal cord injury, Multiple Sclerosis (MS) and stroke. For affectedpatients the problems that result range from accidental leakage of urine by day (in-continence) or bed wetting at night (enuresis), through to an inability to empty thebladder (urinary retention). Urinary retention has potentially serious consequences,particularly in patients with abnormal bladder function secondary to spinal cord in-jury. In such cases, not being aware that the bladder is full can lead to back pressuredeveloping in the urinary tract that risks serious damage to the kidneys. In all thesecases, a wearable non-invasive device that monitors the fullness of the bladder and15CHAPTER 1. BACKGROUNDprovides and alarm once the volume of urine in the bladder has reached a pre-setthreshold, would be beneficial. Individual patients would then, depending on theirpathology, be able to empty their bladder voluntarily to avoid incontinence. In chil-dren with nocturnal enuresis, a problem that affects 20% of children over four yearsof age[68] , a device that wakes the patient with an alarm once the bladder is full,but before incontinence occurs, has major advantages over current systems thatonly detect accidental voiding. Enabling the subject to wake, sense that his/herbladder is full, and void voluntarily before leakage occurs, would lead to condi-tioning over time to waking in response to the bladder being full, and resolve theenuresis. This is an example of an application area where motion artifact reductionis essential for long term monitoring purposes. Proper motion artifact reductionmethods should be integrated into such a device to make it practical for clinicalapplications.The interaction between spatially separated cortical regions in the human brainplays an important role in performing a particular task. Functional imaging meth-ods such as fNIRS can identify different cortical regions involved in a task. How-ever, possible functional connections and interactions among activated areas arenot directly reflected in such images without further processing. These connec-tions, in addition to providing insight into brain?s architecture and function, mayenable one to predict certain brain disorders. As discussed in Chapter 2, it hasbeen shown that functional connectivity may be able to identify broken corticalnetworks in particular disease conditions. fNIRS appears to be a suitable tool formapping brain functional connectivity specially in infants. This group of subjectswere of special interest to us as our collaborators have been working on the devel-opment of language system in infants. This research has had major impact in bothtechnical community [24, 69, 70] as well as general media [71]. One question thatwas left unanswered was how this language network and its connections form anddevelop with maturity. This question is of particular value as there is a growingbody of evidence on the brain areas involved in language processing in neonates,but less on the underlying connectivity. As a result, exploring methods of mappingfunctional connectivity with fNIRS was another area of focus in this thesis. Treat-ment of motion artifacts is a required pre-processing step for reliable results fromconnectivity analysis.16CHAPTER 1. BACKGROUNDOne of the potential applications for NIRS is to monitor the effect of differenttypes of stimulation on the brain. In particular, Transcranial Magnetic Stimulation(TMS) which magnetically stimulates the brain is one of the methods that has beena subject of study using fNIRS. Due to the strong magnetic pulse created by theTMS coil, monitoring the effect of TMS on brain is a challenging task for most ofthe common neuroimaging techniques. Being dependent on optical signal for itsfunctionality, fNIRS is relatively immune to this problem and is therefore a goodfit for monitoring the brain hemodynamics and neural activations during the stim-ulation. The most common approach in this case is to study the effect of TMSon the hemodynamics of the stimulation target area or the areas closely related toit. However, one interesting question is how TMS affects and modulates the brainnetworks. This can be investigated either with concurrent TMS-NIRS to look atnetwork and connectivity changes in a shorter time scale, or by doing connectivityanalysis following TMS to study the lasting effect of TMS on a particular corti-cal network. Resting State Functional Connectivity (RSFC) analysis using fNIRSprovides a useful tool for monitoring this effect in a non expensive and easy to usemanner. Because of the special needs and requirements of using NIRS with TMS,a customizable and preferably an in house made fNIRS device is highly desirable.In particular, a custom made NIRS instrument with high enough sensitivity andsampling rate can potentially measure the fast optical signal [72] which allows oneto look at the network changes in a concurrent TMS-NIRS study and evaluate more?elastic? changes that occur in a shorter time scale [73].Development of a motion artifact reduction method as a required pre-processingstep was the initial focus of this thesis which forms a signal processing basis forthe rest of the thesis. Development of a new approach for analyzing resting statefunctional connectivity for potential use in infants as well as in TMS studies and inparticular for stroke patients to investigate changes in cortical networks in responseto TMS was the next major area of interest. In order to look at both short and longtime scale changes in connectivity, we aimed at developing and testing a custommade NIRS device appropriate for concurrent use with TMS. Our final objectivewas to use NIRS to develop a new wearable bladder monitoring sensor for patientswith bladder control problem which is an example of a device that can significantlybenefit from our artifact removal technique due to the inherent problem of motion17CHAPTER 1. BACKGROUNDwhich is a result of engagement in routine daily activities by the subject wearingthe sensor.1.8 Thesis ContributionsThe major contributions of this thesis are summarized in this section. This thesis? Introduces a novel method for removal of motion artifacts from fNIRS datausing the Discrete Wavelet Transform (DWT). This method was adopted asone of the fNIRS motion artifact removal methods in HOMER2 NIRS pro-cessing package1, an open source Matlab toolbox for analyzing fNIRS datadeveloped by MGH-Martinos Center for Biomedical Imaging and widelyused by fNIRS community. This method has been subject to independentreviews and comparison with other methods using simulated [74] and exper-imental NIRS data [75]. The article describing this method appeared in the?Highlights of 2012? collection of the IOP Journal of Physiological Mea-surement. The articles featured in this collection ?span some of the mostcutting-edge areas of biomedical physics, and collectively are a reflection ofthe most influential research published in PMEA in 2012? 2.? Introduces a method for identification of wavelet levels for optimum artifactremoval performance.? Develops a novel method and apparatus for optically detecting changes inbladder contents non-invasively with potential application for patients withbladder control problem. This method is currently under review by the Uni-versity of British Columbia University-Industry Liaison Office (UILO) forpotential intellectual property (IP) protection and technology licensing.? Introduces a method of analyzing time-varying connectivity between corti-cal regions in infants using fNIRS and Multivariate Autoregressive (MVAR)modeling.1www.nmr.mgh.harvard.edu/optics2Physiological Measurement Highlights of 2012 web page18CHAPTER 1. BACKGROUND? Develops a new method for detecting and mapping language network in in-fants using fNIRS and phase analysis on resting state hemodynamic changesin different cortical regions.? Describes the design, test and validation of a custom made TMS-compatiblefNIRS device for future use in TMS studies on stroke patients.1.9 Thesis OutlineThis chapter provided the background information on NIRS, its applications, po-tentials and limitations. In Chapter 2 the state of the art research on the four majorparts of the thesis, namely, motion artifact removal, NIRS in urology, connectivityanalysis with fNIRS and application of fNIRS with TMS is reviewed. In Chapter 3we present a novel signal processing method to identify and remove motion arti-facts from fNIRS data. We present simulation results and comparison with othercommon preprocessing methods used in fNIRS. Also, the method is applied toNIRS data collected from infants in an fNIRS study to evaluate its performance.In Chapter 4, we introduce a novel application for NIRS in the form of a proofof concept for a wireless wearable sensor to help individuals with bladder dysfunc-tion determine when their bladder is full.Chapter 5 is dedicated to investigation of the concept of cortical connectivityusing fNIRS. The first part of this chapter provides a preliminary study usingMVAR while the latter provides a more elaborate method of connectivity analysisusing phase information of the resting state hemodynamic signals.In Chapter 6 we describe a TMS compatible fNIRS experimental setup for fu-ture use in evaluating the effect of TMS in stroke subjects. A custom made fNIRSdevice has several advantages for monitoring the TMS effect. In particular, dueto special needs of NIRS devices used with TMS, being able to modify and ad-just NIRS system parameters such as power and sampling rate would be desired.Additionally, a sensitive enough device with high sampling rate may be capableof detecting fast optical signal in response to TMS which is currently not feasiblewith commercially available NIRS devices.Finally, Chapter 7 summarizes the methods, results and findings presented inthis thesis and identifies some of the possible future directions.19Chapter 2Literature ReviewIn this chapter we review the work related to the 4 main topics of this thesis, whichare interference reduction in NIRS, applications of NIRS in urology, functionalconnectivity analysis using fNIRS and applications of NIRS with TMS.2.1 Interference Reduction in NIRSEven though artifact removal can in general be beneficial in any NIRS dataset andsetup, it has received more attention in fNIRS. The fNIRS data usually contains theweak and noisy hemodynamic data and motion artifacts can have significant impacton detecting the hemodynamic response. As a result, the works reviewed here aremostly focused on applications of artifact and interference removal methods infNIRS.In general, there are 3 major approaches to identification and/or removal ofinterferences and motion artifacts from NIRS signal. One common method is touse an auxiliary input signal whose output is highly correlated with the motionartifacts. This could be from sources such as an accelerometer attached to theoptodes or an optical channel sampling from superficial layers of the tissue. Thissignal is then used to remove the artifact caused by motion from the NIRS signal.Adaptive filtering is one of the methods commonly utilized with this approachfor fNIRS motion artifact removal. Izzetoglu et al. used adaptive filtering with areference channel from an accelerometer attached to the subject?s head to cancel20CHAPTER 2. LITERATURE REVIEWmotion artifacts [76]. Accelerometer seems to be a good reference for adaptivefiltering of motion artifact as it is highly correlated with the motion. Other stud-ies have also employed accelerometer reference for baseline correction [77] andartifact removal [78].Another common form of a reference signal is NIRS data from a referencechannel. Robertson et al. applied adaptive filtering to fNIRS, using co-locatedsources and detectors in each optode to detect and remove motion artifacts [63].Essentially, the reference channel is selected such that it does not penetrate deepinto the tissue and hence, does not capture hemodynamic information from thetarget tissue. Instead, it will mostly capture changes from superficial layers whichare correlated with motion.Zhang et al. used a similar method with the scalp superficial optical measure-ment as the reference signal for removing global interferences (such as cardiacpulsation and respiration) from deeper brain functional signals [79]. Essentially,these interferences will appear stronger in the superficial layers and can act as areference for removing them from the overall signal. Multidistance optode con-figuration for acquisition of a reference signal has also been used with empiricalmode decomposition and adaptive filtering to remove physiological interferencesfrom fNIRS signal [80].This approach in general can be very effective as it incorporates a direct mea-sure of motion timing and intensity. However, it requires an additional form ofrecording such as accelerometer signal or extra optical channel to provide a refer-ence signal. Therefore, a modified hardware/experimental setup is required for thisapproach and is therefore not applicable to the large amount of NIRS data collectedin past studies by many researchers in this field.Another approach to motion artifacts removal is to identify motion contami-nated blocks or segments of NIRS data and exclude them from calculation of themean Hemodynamic Response Function (HRF). This is often used in fNIRS dataprocessing to improve quality of hemodynamic response detection when averag-ing blocks in a block design study [81]. The motion event in this case has to befirst identified with some criteria such as visual inspection of the NIRS signal orrecording motion event times while conducting the experiment. For example iden-tification of contaminated block by detecting rapid changes in total hemoglobin21CHAPTER 2. LITERATURE REVIEWconcentration in time domain has been reported [82, 83]. This is based on the gen-eral assumption that hemodynamic changes are slow and any fast change in thesignal is caused by motion. Even though this method is not uncommon in NIRSstudies, simulation studies suggest that this approach has no significant improve-ment in recovering HRF compared to no correction for artifacts [74].The last approach is using signal processing methods to identify and remove/re-duce the effect of motion artifacts from previously recorded NIRS data. In otherwords, one uses the temporal or spectral features of the artifacts to identify them.A number of methods have been suggested in the literature based on this approach.Cui et al. suggested using the negative correlation between HbO2 and HHb to de-tect motion artifacts [67]. Normally, the HbO2 and HHb changes, especially duringhemodynamic response to stimulation, are in opposite directions. In general, thechanges of the two are not highly correlated. Therefore, if a highly correlatedchange between the two species is detected, it can be attributed to the motion.Scholkmann et al. used a moving standard deviation scheme to detect motionartifacts and applied spline interpolation in the time interval of the motion to modelthe artifact and subtract it from the signal [66]. This method was reported to workwell for several types of motion artifacts, however, parameters of the method needto be properly adjusted for the type of motion artifact to be removed.This approach has also been employed for reduction of global interferencesfrom fNIRS data. For example, Zhang et al. used eigenvector based spatial filteringto remove global interferences from NIRS signal to improve HRF detection [64].In this method, global interferences are assumed to be responsible for the baselineperiod variations. Using the baseline information as a reference, the stimulationperiod data is processed such that the effect of global interference is minimized.Application of Wiener filtering has also been suggested to reduce the effect ofmotion artiacts [76]. This method requires prior knowledge of the original fNIRSsignal?s power spectrum. Kalman filtering with appropriate model on noise anddata has been applied to both fNIRS and PPG [65, 84]. Application of Kalmanfiltering requires prior assumption on distribution of noise which models the arti-facts.Wavelet decomposition is another promising approach for detection and re-moval of interferences from fNIRS signal. Continuous Wavelet transform for ex-22CHAPTER 2. LITERATURE REVIEWample, has been used to detect blocks contaminated with motion artifacts using ahard threshold on the wavelet transform amplitudes [85]. The scales at which thisthresholding occurs were identified through Monte Carlo simulation on a train-ing data set. Robertson et al. proposed a method based on the DWT along withreference channels for source and detector optodes [63]. The wavelet coefficientswhere shrunk only if they exceeded a threshold in the data channel as well as thereference channels. Wavelet transform has also been applied for removal of globalinterferences in fNIRS signal [86].A detailed review and comparison of some of these methods can be found in[74] and [75].2.2 NIRS in UrologyThe evolution of the NIRS as a means of monitoring the hemodynamics and oxy-genation of the bladder is recent [87]. Important information can be derived withNIRS that contribute to the evaluation of patients with symptoms of bladder dys-function, and understanding is growing of the distinct patterns of change in chro-mophore concentrations that occur in the context of disease. Wireless NIRS de-vices have been utilized to evaluate bladder function in health and disease [87, 88],and NIRS is proving to be uniquely applicable to the study of bladder pathophys-iology because of the anatomic and vascular characteristics of the organ, how thebladder?s microcirculation must function to maintain perfusion as it fills and con-tracts to empty, and because of the negative effect of disorders of detrusor musclehemodynamics and oxygenation on normal voiding function [89].The earliest clinical application of NIRS in urology was reported by Stothers etal. [27]. They observed that patterns of change in HbO2 and HHb are different inhealth and disease during bladder contraction and concluded that abnormalities indetrusor muscle hemodynamics may be related to symptoms of voiding dysfunc-tion.This method was evaluated by other researchers and urologists. In a studyby Yurt et al., NIRS performance in classifying male subjects with Lower Uri-nary Tract Symptoms (LUTS) as bladder outlet obstructed or unobstructed wascompared with the gold standard uroflowmetry and urodynamic Pressure Flow23CHAPTER 2. LITERATURE REVIEWStudy (PFS) [90]. The NIRS method correctly identified 25 out of 29 obstructedpatients and 21 out of 24 unobstructed cases resulting in a sensitivity of 86.2% andspecificity of 87.5%.Amelink et al. studied the application of NIRS for in vivo monitoring of blad-der wall microvascular blood oxygen saturation [91]. Deterioration of bladder mi-crovasculature has been recognized as a cause of continuing bladder function lossand NIRS can be used as a tool to measure the oxygenation of the bladder wall todifferentiate between bladders with loss of function and normal ones. In this casehowever, the NIRS probe needs to be placed in the working channel of a standardcystoscope [91]Detrusor muscle oxygenation has also been monitored using NIRS during de-trusor overactivity contractions [92]. In their study, Vijaya et al. examined 55women with a mean age of 52 years. During involuntary contractions in the 23subjects with detrusor overactivity, they observed a statistically significant increasein HHb at maximum detrusor pressure from the baseline [92] while no change wasobserved for voluntary detrusor contraction.One bladder-related symptom of concern not addressed by current NIRS mon-itoring studies is the inability to sense when the bladder is full. This symptomoccurs in a number of conditions and for affected patients the problems that re-sult range from accidental leakage of urine by day (incontinence) or bed wettingat night (enuresis), through to an inability to empty the bladder (urinary retention).While incontinence and enuresis are troublesome, can be embarrassing, and nega-tively affect a patient?s quality of life [93], urinary retention has potentially seriousconsequences, particularly in patients with abnormal bladder function secondaryto spinal cord injury. In such cases, not being aware that the bladder is full canlead to back pressure developing in the urinary tract that risks serious damage tothe kidneys. Unrecognized, this situation increases morbidity and contributes to ashortened life expectancy.Monitoring bladder size, volume, and content non-invasively can be done us-ing ultrasonic scanning, and is common in clinical practice. This technique usesultrasonic imaging to differentiate the urinary bladder from surrounding tissues andorgans, produce volume information, and estimate urine level. The method givesaccurate results, but does not lend itself to monitoring in ambulant subjects or home24CHAPTER 2. LITERATURE REVIEWuse as it requires powerful computational resources and a complex scanning con-trol system. The requirements of gel application and control measurement alsomake most commercial ultrasonic scanners inappropriate for continuous wearablemonitoring, although some portable and wearable ultrasonic sensors for bladdermonitoring have been reported [68, 94].Bioelectrical impedance analysis is another method proposed for determiningthe volume of urine in the bladder. This technique is principally used for deter-mining extracellular and total body water, and several skin surface electrodes arerequired on the abdomen at the level of the bladder for the changes in electricalimpedance used for detection of urine volume to be measured [95].2.3 Functional Brain Connectivity Using fNIRSThe interaction between different cortical areas is responsible for coordination ofcomplex tasks. Studying this cortical connectivity helps in achieving a better un-derstanding and insight of the human?s brain architecture and can potentially helpin diagnosing certain diseases and conditions affecting the brain connectivity.Cortical connectivity can be divided into two main categories. Functionalconnectivity is defined as temporal correlations between spatially remote neuro-physiological events [96]. In functional connectivity, some measure of statisticalinterdependence is used as a measure of connectivity. Effective connectivity onthe other hand, involves identifying causal influences between cortical regions. Inother words, it shows how the information flows between different cortical areas.Cortical connectivity analysis is a recent subject of interest in functional imagingwhich studies how the activated cortical networks interconnect and coordinate toperform a particular sensorimotor/cognitive task.Functional connections between cortical regions can reveal networks consist-ing of functionally connected cortical regions which are involved in task specificactivities. In some situations, task based neuroimaging may not be an appropriatediagnostic tool as the subjects may be unable to perform a task at all [97]. Connec-tivity mapping can be beneficial in such circumstances. Functional connectivityusing fNIRS is specially important and relevant in cases where subjects can notbe transferred to scanner devices. Examples are patients in intensive care units or25CHAPTER 2. LITERATURE REVIEWmental/language development studies in infants. fNIRS provides a good substi-tute for fMRI based connectivity mapping and allows researchers or clinicians toperform same studies as the ones with fMRI on subjects with limited mobility.Functional connectivity has been a subject of study in Electroencephalography(EEG) [98], fMRI [99] and more recently in fNIRS [97, 100] and different meth-ods have been proposed for analyzing it. Cross correlation and cross coherence aretwo widely used methods for detecting functional connectivity in fMRI [101, 102].In these methods a seed region is selected and the cross correlation/coherence iscalculated between the seed region and time course from all other brain areas. Todetermine the direction of influence, Directed Coherence (DC) method has beenproposed [103]. It decomposes coherence into components that represent feedfor-ward and feedback components of the interaction between two time series. Partialdirected coherence and directed transfer function are also proposed for neural struc-ture determination [104]. In practice, these methods rely on modeling the signalwith an MVAR model.Another measure for analyzing cortical connectivity is the mutual informationbetween signals from different brain areas [105]. This method has the advantagethat it is model free and is thus not limited to linear models.In functional neuroimaging methods such as fMRI and fNIRS, one could exam-ine the task-specific functional connectivity which studies how connections differin response to a task. Sun et al. for example, used coherence and partial co-herence to evaluate task-specific connectivity in subjects performing two differenttasks where one of the tasks required more bimanual coordination [102]. Theirresults showed that even though there was no significant difference in mean activ-ity between the two tasks, there was an increase in interhemispheric connectivitybetween primary motor (M1) and Premotor area (PM) for the bimanual task whichrequired higher degree of coordination.A different type of connectivity from task-specific connectivity is the RSFC.RSFC is based on the synchronization between spatially remote and different cor-tical areas at rest which is an indication of functional connection between differentbrain networks. It was first demonstrated by Biswal et al. using BOLD-fMRI bydetecting low frequency oscillations in the motor cortex at rest [101].White et al. originally used fNIRS based RSFC analysis in five subjects to26CHAPTER 2. LITERATURE REVIEWdemonstrate the feasibility of using fNIRS as an alternative to fMRI for this typeof connectivity analysis [97]. Through simultaneous imaging over motor and visualcortices, they were able to derive robust correlation maps which were in agreementwith expected functional neural architecture.fNIRS based connectivity has since been employed to study RSFC in differentbrain networks. In a study by Zhang et al., 30 young adults were studied in aresting state followed by localizer task measurement [106]. The localizer task wasused to identify the seed channel for connectivity analysis. Using the GeneralizedLinear Model (GLM) with seed channel as the independent variable, they showed asignificant RSFC between left inferior frontal cortex and superior temporal regionswhich are associated with language.Lu et al. investigated a similar approach to study sensorimotor and auditorycortices [107]. They studied 29 adult subjects and computed the RSFC using seedbased correlation analysis and showed that the resulting networks were consistentwith previous fMRI findings.Duan et al. compared the results of RSFC with fNIRS using correlation withseed Regions Of Interest (ROI) with those obtained from fMRI to evaluate validityof fNIRS based RSFC [108]. They used simultaneous fNIRS-fMRI data from 21subjects in resting state. There was high similarity in connectivity between the bi-lateral primary motor ROI using the two methods, specially for HbO2 and BOLDfor all subjects. Also, group level sensorimotor connectivity maps showed similar-ity between the two methods and this similarity in group level was higher than inindividual level.2.4 NIRS for Monitoring TMSAs described in Chapter 1, TMS is the method of magnetically stimulating thebrain. This method has found clinical and research applications such as treatmentof psychiatric diseases and stroke rehabilitation. Using neuroimaging techniques,one can directly monitor the effect of the stimulation on brain activity and the in-teraction among brain regions. The strong electromagnetic pulse produced by theTMS coil strongly interferes with other sensitive electronic devices in the vicinityof the coil. This introduces a severe problem for other monitoring devices such27CHAPTER 2. LITERATURE REVIEWas EEG or Magnetoencephalography (MEG). Such modalities rely on very smallelectric activities for their performance which are swamped by the induced cur-rents from TMS pulse. Optical methods such as NIRS, however, are based on thechanges in parameters of diffusive light and therefore are not affected by the TMSelectromagnetic pulse. Thus, NIRS seems to be an appropriate choice for studyingthe effects of TMS on human brain. The short penetration depth of the NIRS inthis case seems to match the focal depth of the TMS coil which normally does notexceed 3-4cm.In simultaneous NIRS-TMS studies, the changes in cortical hemodynamics ei-ther at the location targeted by TMS pulses or in other areas of the brain are mon-itored during or after the stimulation. Hada et al. used a two channel device be-neath the stimulation coil during a repetitive TMS recording [109]. They observeddecrements in tHb and HbO2 concentrations and increment of HHb during andafter repetitive Transcranial Magnetic Stimulation (rTMS) at different stimulationrates and intensities. They also observed that concentration changes continued forup to 10s after stimulation.Hanaoka et al. investigated the effects of low frequency rTMS over the rightfrontal lobe on the function of the left frontal lobe by NIRS [110]. They observedsignificant changes on hemodynamics during the poststimulation baseline periodwhich was interpreted as demonstrating the activation and deactivation of the leftfrontal cortex during and after rTMS of the right frontal cortex.In a similar study, Mochizuki et al. studied interhemispheric interactions be-tween bilateral motor and sensory cortices using NIRS and rTMS [111]. Theyrecorded hemoglobin concentration changes at the right prefrontal cortex, PM,primary hand motor area (M1) and primary sensory area (S1) during and afterstimulation over the left PM, M1, and S1. They also recorded Motor EvokedPotential (MEP) to TMS over the right M1 from the left first dorsal interosseousmuscle after the conditioning TMS over left S1. They reported that TMS over PMinduced a significant HbO2 decrease at the contralateral PM and stimulation overM1 elicited a significant HbO2 decrease at the contralateral S1, and TMS over S1significant HbO2 decreases at the contralateral M1 and S1. They suggested that allthese indicate a mainly inhibitory interaction between bilateral PMs and bilateralsensorimotor cortices in humans.28CHAPTER 2. LITERATURE REVIEWMochizuki et al. also used NIRS recordings in a separate study along withTMS in 4 different conditions (3 intensities and sham stimulation) and similar toother cases, detected significant changes in HbO2 and HHb in all intensities [112].They suggested that the increase of HbO2 concentration at 100% Active MotorThreshold (AMT) under the active condition reflects an add-on effect by TMS tothe active baseline and that the decrease of HHb and tHb concentrations at 120 and140% AMT under the resting condition are due to reduced baseline firings of thecorticospinal tract neurons induced by a lasting inhibition provoked by a higherintensity TMS [112].Eschweiler et al. utilized NIRS with rTMS on the left dorsolateral prefrontalcortex of patients suffering from major depression [113]. They observed that ab-sence of a task-related increase of total hemoglobin concentration at the stimula-tion site before the first active rTMS significantly predicted the clinical responseto active rTMS [113]. They reported that clinical benefits of rTMS are predictedby low local hemodynamic responses and support the idea of activation-dependenttargeting of rTMS location.Chiang et al. have also followed similar procedure to investigate the effectof TMS using NIRS [114]. However, they mostly concentrated on finding outhow long the TMS effect lasts. More specifically, the aim of their study was tomeasure the change of HbO2 and HHb levels in the left motor cortex after 20 minof 1 Hz TMS over the right motor cortex. Subjects carried out a finger to thumbtapping task sequentially with six blocks of ten cycles (30 s on and 60 s off). Oneblock was performed before TMS and five after TMS. The results showed that thelevel of HbO2 in the unstimulated cortex increased after TMS over the contralateralhemisphere and that the increase lasted 40 min after 1 Hz stimulation. HHb wasslightly decreased during the first 15 min after stimulation.2.5 Summary and ConclusionArtifact removal is a crucial pre-processing step for almost any NIRS applicationdue to the high sensitivity of this method to motion. When a NIRS based sensor isused for continuous monitoring, the issue of motion artifacts becomes a more seri-ous problem. This specifically applies to NIRS instruments for continuous moni-29CHAPTER 2. LITERATURE REVIEWtoring of the bladder function which need to be worn by the subject at all times andare hence affected by the subject?s day to day activities.The problem of motion artifacts is not unique to wearable NIRS systems. WhenNIRS is used for monitoring brain function, which is one of the most appealing ap-plications of this technology, motion artifacts can appear from a number of differ-ent sources. In monitoring the effect of TMS for example, which is a recent area ofinterest [115, 116], several mechanisms can result in motion artifacts. In particular,the TMS pulse can result in activation of superficial scalp muscles which can poten-tially cause motion artifacts in the NIRS signal [117]. Additionally, the motion andvibrations resulting from TMS coil activation can result in slight shifts in optodeslocation which is reflected in the signal [73, 117]. In addition to these sources, thehead movements, specially in longer stimulation sessions can contribute to motionartifacts. This highlights the importance and necessity of developing proper artifactremoval methods for use in combined TMS-NIRS studies.Cortical connectivity analysis, another potential field of application of NIRS,is also susceptible to motion artifacts. Many of the connectivity analysis methodsrely on similarities and relations between signals in spatially remote channels todetect connectivity [106, 107]. Spontaneous movements of the subject, which is acommon issue specially in infants and young children, results in highly correlatedchanges across all or a large number of channels that can influence the measuredconnectivity. Being able to identify motion corrupted data segments allows theremoval or treatment of these data blocks to avoid detection of false connectivity[108].The advantages of NIRS over other neuroimaging methods has lead to a greatinterest in its use for monitoring and studying brain function. One area of interestfor fNIRS in brain functional study is the influence of electrical and magnetic stim-ulation on brain hemodynamics. In particular, the intrinsic advantages of NIRSmake it a desirable choice for use with TMS. The most common approach is tostudy the effect of TMS on the hemodynamics of the target area or the areas closelyrelated to it. However, one interesting question is how TMS affects and modulatesthe brain networks. This can be investigated either with concurrent TMS-NIRSor in an ?offline? approach [118] where the lasting effect of TMS on a particularcortical network is investigated. RSFC using fNIRS provides a useful tool for mon-30CHAPTER 2. LITERATURE REVIEWitoring this effect in a non-expensive and easy to use manner. Moreover, with theavailability of a NIRS instrument capable of measuring the fast optical signal, onecan also look at the network changes in a concurrent TMS-NIRS study and evaluatemore ?elastic? changes that occur in a shorter time scale and require instrumentswith higher temporal resolution and sensitivity to measure [73].In summary, development of a motion artifact reduction methods as a requiredpre-processing step was the initial focus of this thesis which forms a signal process-ing basis for the rest of the thesis. Development of a new approach for analyzingresting state functional connectivity for potential use in infants as well as in TMSstudies and in particular for stroke patients to investigate the changes in corticalnetworks was the next major area of interest. In order to look at short time scalechanges in connectivity, a custom made NIRS device was developed and tested. Fi-nally, a new NIRS based wearable bladder monitoring sensor was developed whichis an example of a device that can significantly benefit from the artifact removalmethods due to the inherent problem of motion as a result of constant wearing bythe subject for bladder monitoring.31Chapter 3Wavelet Based Motion ArtifactRemoval for Functional NearInfrared SpectroscopyNIRS has found usage in a wide range of applications as a powerful tool for moni-toring tissue hemodynamics. In particular, fNIRS as a subset of NIRS, has been asubject of interest for brain studies due to its non invasive, non restraining nature.However, for fNIRS to work well, it is important to reduce its sensitivity to motionartifacts. In this chapter, we introduce a new wavelet-based method for remov-ing motion artifacts from fNIRS signals. Even though this method was originallydesigned and tested on brain fNIRS data, it can in general be applied to a widerange of NIRS data collected from different tissue types. The material presentedin this chapter was published in the proceedings of the international IEEE EMBSconference in 2010 [2], proceedings of SPIE in 2011 [3] and the journal of physio-logical measurement [4]. The method presented here was independently comparedwith a number of other common methods of artifact removal using simulated andexperimental data [74, 75]. This method has also been used in HOMER2 NIRSprocessing package1, an open source MATLAB toolbox for analyzing fNIRS datadeveloped by MGH-Martinos Center for Biomedical Imaging and widely used by1www.nmr.mgh.harvard.edu/optics32CHAPTER 3. MOTION ARTIFACT REMOVALthe fNIRS community.3.1 Artifact Removal Using Discrete Wavelet TransformA majority of motion artifacts appear in the form of abrupt changes in the ampli-tude of the signal. The DWT can provide good localization in time or frequencydomain. Therefore, motion artifacts appear as isolated large coefficients in the dis-crete wavelet domain. This makes identification and removal of artifacts easier inthe wavelet domain.A signal y(t) can be expanded using the DWT asy(t) = ?kv j0k? j0k(t)+??j= j0?kw jk? jk(t) (3.1)where ? jk(t) = 2 j/2?(2 jt ? k) is the scaling function and ? jk(t) = 2 j/2?(2 jt ? k)is the wavelet function [119]. j and k are the dilation and translation parametersrespectively and j0 is the coarsest scale in the decomposition. v j0k and w jk are theapproximation and detail coefficients and ?(t) and ?(t) are the mother wavelet andscaling functions, respectively.We assume that the observed signal is composed of the physiological signal ofinterest, f (t), plus an interference term, ?(t)y(t) = f (t)+ ?(t) (3.2)Using the Fast Wavelet Transform (FWT) algorithm [120], the wavelet transformof the observed signal can be written asw jk = ?lg(l?2k)v j+1(l) j = j0 . . .J?1,k = 0 . . .2 j ?1 (3.3)v jk = ?lh(l?2k)v j+1(l) (3.4)where g(n) and h(n) are the wavelet filter bank highpass and lowpass filters re-spectively with vJ(n) = y(n) and j0 is the coarsest level [119]. y(n) is the sampledversion of y(t) with n = 0 . . .N ?1 and N = 2J . This can be written in matrix form33CHAPTER 3. MOTION ARTIFACT REMOVALasW = WY (3.5)where Y =[y(0) . . . y(N ?1)]T. W =[WJ?1 WJ?2 . . . W0 V0]Tis the N ?N DWT matrix and W =[WJ?1 WJ?2 . . . W0 V0]Tis the vector of waveletcoefficients [121]. W j is the vector of wavelet coefficients at level j i.e. (W j)k =w jk and V0 is the scaling coefficient. (.)k indicates the kth element in the vector.Writing Equation 3.2 in vector form and using discrete version of y, f and ? andapplying the wavelet transform for one level we haveW jY = W jf+W j? (3.6)which gives the relationship between wavelet coefficients of the observed signal,underlying physiological signal of interest and the noise term representing artifacts.Distribution of wavelet coefficients can be described by a mixture of Gaussians[122, 123]. One Gaussian component describes coefficients centered around zeroand one describes those spread out at larger values. Here we impose a single Gaus-sian distribution on wavelet coefficients. The hemodynamic signal is a smooth andslowly varying signal compared to motion artifacts. Therefore, most of waveletcoefficients of the fNIRS signal are spread around zero with smaller variance com-pared to motion artifact coefficients. Our model is similar to the one described byAntoniadis [122] with the assumption that only coefficients from normal distribu-tion with smaller variance in the mixture model belong to original signal. Hence,the model is reduced to a single Gaussian distribution. The wavelet coefficients ofthe observed signal, y(n), can be therefore written asw jk = w? jk + ? jk (3.7)where w? jk ? N(0,?2). The mean of the distribution is zero because the waveletcoefficients w jk are the outputs of a highpass filter. The ? jk coefficients appearas a few large coefficients across the time course of each level. For any givencoefficient, w jk, if the probability of observing values larger than w jk is less than34CHAPTER 3. MOTION ARTIFACT REMOVALan arbitrary probability, ? , we can conclude that the coefficient does not belong tothe original signal and must have been due to artifacts and thus must be removed.This probability can be written asp jk = 2(1??( |w jk|??))(3.8)where ? is the Normal Cumulative Distribution Function (NCDF). We then pro-pose to use the following thresholding scheme for the removal of the artifacts:w? jk ={w jk if p jk > ?0 if p jk < ?(3.9)? is the probability threshold which can function as a tuning parameter. In thisapproach, we are basically treating artifacts as large outliers added to the desiredcoefficients w? jk ? N(0,?2). The parameter ? indirectly determines how much theartifact power should be reduced. For the limiting case of ? ? 0, no thresholdingis applied and the signal coefficients are left intact at level j.The level selection for artifact removal is based on the degree of artifact con-tamination at each level which is defined as total number of coefficients exceedingthreshold in that particular level. Define ? = {w jk : p jk < ?} and ? j = ?k I?(w jk)where I?(x) is the indicator function. Then artifact removal is conducted in levelsthat fall in the 90th percentile of{? j}.The variance of the distribution of w? jk can be estimated using Median AbsoluteDeviation (MAD) [124]. MAD is a robust estimator of scale and is not sensitiveto outliers. In our application, artifacts behave like outliers added to the originalsignal whose variance we would like to estimate. As a result, a limited numberof artifacts does not cause a problem in estimating the variance of original signalcoefficients, W jf. The estimate of standard deviation is related to MAD in eachsubband by [124]?? j =Median(|W j|)0.6745 (3.10)where ?? j is the estimated standard deviation in scale j and W j is the set of wavelettransform coefficients in the same level.35CHAPTER 3. MOTION ARTIFACT REMOVALFollowing thresholding, the signal can be reconstructed using?Y = WT ?W =[WTJ?1,WTJ?2 . . .WT0 ,VT0]??????????WJ?1?WJ?2...?W0?V0?????????(3.11)where(?W j)k = w? jk.To avoid pseudo Gibbs phenomena near singularities or abrupt changes in thesignal which is mostly attributed to the lack of shift invariance in traditional DWT,we performed artifact removal on all possible circularly shifted versions of theoriginal signal and then undid the shifts and averaged the results [125]. This isknown as Translation Invariant Wavelet Transform (TIWT) and will reduce theundesirable effects caused by shift-variance of DWT. This can be formally statedasy?(n) = 1MM?h=1S?h (T (Sh (y(n)))) (3.12)where Sh is the shift operator and T is the operator representing DWT followed byartifact removal and M is the total number of shifts [125].3.2 SimulationWe used simulated NIRS data for preliminary evaluation of this method. In sim-ulation, we have access to the original artifact-free signal and can evaluate howwell our method can reconstruct the original signal. This usually is not possiblein experimental data as the artifact-free signal is not available. We first simulatethe NIRS signal with an Autoregressive (AR) model of order 9 and then add theartifacts. It was observed that the AR model was sufficient to model the generalbehavior of a short duration fNIRS signal. The artifact was selected from an ex-perimental fNIRS signal and superimposed on the simulated fNIRS signal. Allsimulations and Wavelet processings were performed in MATLAB (Mathworks,36CHAPTER 3. MOTION ARTIFACT REMOVAL850 900 950 1000 1050 1100 1150 1200MedianSURETIWTArtifactTime (s)Concentration Change (? Mol)  NIRS Signal with ArtifactOriginal SignalProcessed SignalMotion artifactFigure 3.1: Comparison of methods: plots from the top show the simulatedoptical signal with artifact, the proposed method with TIWT, SUREthreshold and Median filtering respectively. The original artifact-freesignal is also shown for each method.MA, USA) and using the Wavelab 850 toolbox 2.We evaluated the performance of our method using DWT and TIWT alongwith Median filtering and Stein?s Unbiased Risk Estimator (SURE) based waveletdenoising in removing the simulated artifact [126]. We wanted to verify if ourmethod offers any significant improvement over regular wavelet denoising, there-fore SURE based wavelet denoising is chosen for comparison. Median filteringis the procedure frequently used for removing impulsive noise from signals. Thelength of the median filter is selected to be twice the duration of the artifact toprovide the best artifact suppression while having minimum filter length.2www-stat.stanford.edu/ wavelab37CHAPTER 3. MOTION ARTIFACT REMOVAL0 1 2 3 4 5?40?20020406080100Frequency (Hz)Amplitude (dB)  Original SignalProposed Method (DWT)Median FilterSURE ThresholdProposed Method(TIWT)Figure 3.2: Comparison of the power spectrum of the original signal and theprocessed signal using 4 methods for a typical simulated fNIRS signal.3.2.1 Performance EvaluationWe compared the methods using Normalized Mean Squared Error (NMSE). NMSEis defined asNMSEi = 10log10?Nn=1 (yi(n)? y?i(n))2?Nn=1 y2i (n)(3.13)where yi(n) is the original signal, y?i(n) is the artifact removed signal and i is thechannel index. In order to take into account the random effect of the fNIRS signaland the artifacts, we applied each artifact removal algorithm on 100 realizations ofthe AR signal with artifact and averaged the performance of each method. Param-eter ? is set to 0.05 for an average artifact removal.We also use Magnitude Squared Coherence (MSC) as an additional measurewhich is defined asMSCyy?( f ) = |Pyy?( f )|2Py?y?( f )Pyy( f ) (3.14)where Pyy?( f ) is the cross power spectrum of signals y? and y, and Py?y?( f ) and Pyy( f )are the power spectrum densities of y? and y as a function of frequency, respectively.38CHAPTER 3. MOTION ARTIFACT REMOVAL0 1 2 3 4 500.51(a)MSC0 1 2 3 4 500.51(b)MSC0 1 2 3 4 500.51(c)MSC0 1 2 3 4 500.51Frequency (Hz)(d)MSCFigure 3.3: Coherence between original signal and a) output of our methodwith DWT, b) SURE method, c) Median filtering and d) our methodwith TIWTThe coherence has a value between 0 and 1 and measures how much y? correspondsto y at each frequency. In other words, it indicates how well the signals spectraare matched in each frequency band. The frequency bands selected for analysis are0.15 Hz wide. The power spectrum densities are calculated in each window andthen the window is shifted towards the next band. There is a 50% overlap betweenadjacent windows. The results are averaged over 100 trials.3.2.2 ResultsThe results of NMSE analysis for simulated fNIRS are summarized in table 3.1.Lower NMSE in the table indicates higher similarity to the original artifact-freesignal and hence, better performance in artifact removal. We used two sample t-test to verify if the differences in the table are significant. The data is tested for39CHAPTER 3. MOTION ARTIFACT REMOVALTable 3.1: Normalized Mean Squared Error between processed signal andoriginal artifact-free signalMethod Median SURE DWT TIWTFilter ThresholdAverage-8.95 -8.71 -9.97 -10.98NMSE (dB)Standard 1.36 1.65 2.38 1.97Deviation (dB)normality prior to the test. The difference between TIWT and other methods isstatistically significant (p < 0.01). The difference between DWT and SURE is alsosignificant (p < 0.05), however, difference between DWT and median filter is notsignificant. Visual comparison of the results for median filter and DWT reveals thatDWT better preserves the shape of the signal in regions away from artifact. Fig.3.1 demonstrates the result of applying median filter, wavelet SURE denoising, andthe proposed method using TIWT to a sample of simulated fNIRS signal.Fig. 3.2 shows the power spectrum of the signals in Fig. 3.1. The figuresuggests that median filtering significantly alters the power spectrum of the sig-nal, while the 3 wavelet based methods selectively reduce the power in frequencyranges where the energy of artifact is mostly concentrated. The results of MSC areshown in Fig. 3.3. The figure shows high coherency between the original signal andthe artifact removed signal using our method with DWT and TIWT. It is evidentthat the coherency for median filtering results is very low for higher frequencieswhich is due to smoothing effect of Median filter.3.3 ExperimentfNIRS is widely used in infant studies as the small skull size leads to a largervolume being investigated by fNIRS and also because other functional imagingmethods may not be readily applicable to infants. Functional studies in infantshave many applications in areas such as brain development in infants [127], speechperception [83], premature infant studies [128] and cognitive studies [82]. How-ever, motion artifact is a serious problem in infant fNIRS studies as the subjectsmay move spontaneously.To evaluate the performance of the method, we applied it to fNIRS data col-40CHAPTER 3. MOTION ARTIFACT REMOVALFigure 3.4: Experiment setuplected from 3 infants (two 1 day old and one 2 day old infants, 1 male). The ex-periment was approved by the University of British Columbia board of ethics andconsent form was signed by the infants? parents. A 24 channel ETG-4000 fNIRSdevice (Hitachi Medical Corporation, Tokyo, Japan) with 700nm and 830nm lasersand sampling rate of 10Hz was used for data collection. Artifact removal was per-formed on raw optical density data. Motion artifacts are a form of interferencein optical signal rather than a physiological interference. We attempt to removethem in the optical attenuation signal to avoid using artifact contaminated data incalculation of HbO2 and HHb.During the study, fNIRS optodes placed on left and right temporal regions werefixed by a gauze bandage. Optode placement is shown in Figure 3.5. The total du-ration of the processed fNIRS recording was 819.2 seconds and the first 20 chan-nels from each subject were included in the analysis. The infants were videotapedduring the experiment to determine the time instant of movements. The video sig-nal was then processed using Sum of Absolute Differences (SAD) between each 2consecutive frames to acquire a reference signal from which motion intervals wereextracted for evaluation purposes [129].A total of 29 motion events were recorded for 3 subjects (16, 7 and 6 for sub-jects 1 to 3). Motion artifacts that take place at different time instants are inde-pendent and may have different shape, duration or amplitude. Therefore, overall41CHAPTER 3. MOTION ARTIFACT REMOVALFigure 3.5: Optodes placement. The red circles and blue squares indicate thesource lasers and detectors, respectively. The numbers between the dotsindicate channel numbers.artifact attenuation was calculated for the entire set of motion events. The value ofartifact attenuation for each artifact is the median of attenuation across all channels.3.3.1 Performance EvaluationWe used artifact power attenuation and NMSE as criteria to evaluate the perfor-mance of our method. Artifact power attenuation is defined as? im = 10log10?n?Am[yHPi (n)]2?n?Am[y?HPi (n)]2 (3.15)where Am is the artifact time interval for the mth artifact in channel i and yi(n) isthe original signal and is highpass filtered to yield yHPi (n). y?HPi (n) is the highpassfiltered version of the processed signal. The purpose of highpass filtering is to re-move the effect of low-frequency physiological variations to ensure the calculatedenergies only reflect the energy of the artifact and not that of the physiologicalsignal. This measure is basically the ratio of artifact energy before and after re-moval in dB. The artifact interval is identified using the video reference signal andis selected such that it begins at the time instant the reference signal deviates frombaseline and ends when it reaches the baseline following the perturbations causedby motion. The baseline is assumed to be the stable signal level before and afterartifact.42CHAPTER 3. MOTION ARTIFACT REMOVAL880 900 920 9400.911.11.21.3Time(s)??OD  Before RemovalAfter Removal580 590 600 6102.72.82.933.  Before RemovalAfter Removal880 900 920 94050100150200250Time(s)Motion Intensity (SAD)580 590 600 61060708090100110120Time(s)Motion Intensity (SAD)Figure 3.6: Two typical motion artifacts and results of applying the proposedmethod (top) with the motion reference signal extracted from videotape(bottom).NMSE is defined asNMSEi = 10log10?n/?Am [yi(n)? y?i(n)]2?n/?Am [yi(n)? y?i]2(3.16)where i is the channel index and y?i is the mean value of yi(n). The NMSE iscalculated for the artifact-free segments of each channel. NMSE shows how muchdistortion has been introduced and complements the first criterion, which indicateshow much of the artifact power has been removed.3.3.2 ResultsTo evaluate the performance of the method, we first applied it to an artifact-freefNIRS signal of length 512 samples (corresponding to 51.2 seconds) in the ab-sence of motion artifacts which resulted in an NMSE of -13.80 dB, -17.54 dB and43CHAPTER 3. MOTION ARTIFACT REMOVAL2 4 6 8 10 12 14 16 18 200102030Attenuation(dB)  700nm830nm2 4 6 8 10 12 14 16 18 200102030Attenuation(dB)  700nm830nm2 4 6 8 10 12 14 16 18 200102030ChannelAttenuation(dB)  700nm830nmFigure 3.7: Artifact attenuation in 20 channels for 2 wavelengths for subjects1 to 3 (top to bottom).-14.84 dB for 3 subjects, respectively. We used Daubechies 5 (db5) wavelet for allexperiments and the value of ? was set to 0.1. This is equivalent to treating coeffi-cients whose probability of belonging to hemodynamic signal is less than %10 asartifacts. Next, we evaluate the performance of the method in presence of motionartifacts. Figure 3.6 shows two typical head motion artifacts in the fNIRS signal,one in the form of a short abrupt impulsive noise and one in the form of a seriesof slower variations in the signal and the filtered signal along with the motion ref-erence signal. The median of ? im and NMSEi across all channels for each subjectand for the 2 wavelengths are presented in Table 3.2. We consider the error to beacceptable if it is within 5% of the signal?s energy which translates to less than-13dB in NMSE. The value of the NMSE in the table is calculated by excludingthe first and last 300 samples in each channel in calculating Equation 3.16. Thisis to ensure the error represented in NMSE does not include errors due to edge ef-fect. Different channels in our fNIRS setup are affected differently by motion andtherefore, the artifact attenuation is not the same for all channels. This is shownin Figure 3.7 where the performance of the method in terms of artifact attenua-tion across different channels is shown for 2 wavelengths. The artifact attenuation44CHAPTER 3. MOTION ARTIFACT REMOVALTable 3.2: Median of NMSE and Attenuation in artifacts energy (in dB) forproposed methodSubject 1 Subject 2 Subject 3700nm 830nm 700nm 830nm 700nm 830nmNMSE -20.84 -21.23 -16.70 -16.97 -18.05 -17.62Attenuation 15.65 15.40 18.77 15.03 18.66 22.81ranges from 7.3 dB to 37.3 dB in subject 1, 6.0 dB to 39.2 dB for subject 2 and3.48 dB to 41.28 dB for subject 3 for both wavelengths combined. The median ofoverall artifact attenuation over the total of 29 motion events is 18.29 dB and 16.42for 700nm and 830 nm channels.As a comparison with regular wavelet denoising, we applied adaptive waveletdenoising based on SURE (Stein?s Unbiased Risk Estimator) risk to the same testdata [126]. We chose wavelet denoising levels in such a way that similarity be-tween processed and original signal is the same for SURE based denoising and ourmethod. We then compare the artifact attenuation. The comparison is made acrossall artifacts for the three subjects. The intensity and duration of each artifact isdifferent and it is fair to assume that the results of attenuation for each artifact isindependent of other artifacts for the same subject. In this way, the performance of2 methods on 29 different artifacts are compared.We chose the following metric for the similarity of the original and processedsignal [130]:d( f1, f2) =?? pi?pi(log f1(?)f2(?))2 d?2pi?(? pi?pilog f1(?)f2(?)d?2pi)2(3.17)This metric indicates how far apart the power spectra of 2 signals are. We evaluatedthe similarity in every artifact-free segment of original and processed signal for thetwo methods and averaged across each subject to derive 1 value for each subject.Two typical artifacts were chosen to qualitatively demonstrate the differencebetween proposed method and regular wavelet denoising as shown in Figure 3.8.The top two right panels show a case in which both proposed method and waveletshrinkage perform equally well in removing the artifact. The top two left panels45CHAPTER 3. MOTION ARTIFACT REMOVAL0.911.11.21.3??OD  Before RemovalAfter Removal0.911.11.21.3??OD  Before RemovalWavelet Denoising620 625 630 635 640 645 650 655 66050100150200Time(s)Motion Intensity2.833.23.4??OD  Before RemovalAfter Removal2.833.23.4??OD  Before RemovalWavelet Denoising580 585 590 595 600 605 6106080100120Time(s)Motion IntensityFigure 3.8: Comparison of proposed method with wavelet denoising for 2typical motion artifacts. Top 2 panels show the artifacts and results ofapplying proposed method. Middle panels show the results of waveletdenoising and the bottom panels are the motion reference signals ex-tracted from videotape.show the case were wavelet denoising is not capable of detecting and removingthe artifact while the proposed method has been able to attenuate the artifact. Thisis shown quantitatively in Figure 3.9. The attenuation is significantly different forthe 2 methods (2 sample t-test p<0.01). The results suggest that the proposedtechnique yields higher artifact attenuation for a given level of distortion in thesignal.The effect of varying ? from 0.01 to 0.15 on NMSE and artifact power attenu-ation for 3 subjects is shown in Figure 3.10. ? can be used as a tuning parameterto achieve desired artifact attenuation in trade off with signal distortion. The per-formance of the method is not the same for all subjects with similar ? . However,changing ? has the same effect on all subjects.3.4 Discussion and ConclusionIn this chapter, we proposed a method for reducing motion artifacts in fNIRS sig-nals using the discrete wavelet transform. The method is based on the assumptionthat motion artifacts have different characteristics in terms of amplitude and du-ration from the original signal. This difference is better highlighted in waveletdomain due to the good localization property of the DWT.46CHAPTER 3. MOTION ARTIFACT REMOVAL5101520Proposed Method SUREMethodAttenuation (dB)Figure 3.9: Comparison of proposed method with adaptive wavelet denoisingmethod (SURE based)To estimate the fNIRS signal coefficient?s distribution with the proposed method,the entire time span of the signal should be available. Therefore, our method in itscurrent form is not suitable for real time processing. A possible workaround foronline processing is to estimate the variance based on available data and updatethe estimate as more data becomes available. The effectiveness of this motion arti-fact removal method on improving accuracy of fNIRS activation maps is yet to beexamined.Artifact reduction can potentially distort the signal. It is important to be ableto control the level of signal distortion in practical applications. The value of theparameter ? in the proposed method can be set by the user to control NMSE intrade off with the intensity of artifact attenuation to reach a balance between NMSEand artifact attenuation.Evaluation of a motion artifact reduction method requires knowledge of motionevent times. Use of deliberate artifacts [76] or human observers [85] has beenproposed for this purpose. Our method of extracting motion reference signal fromvideo signal provides a non-subjective measure of motion for further evaluation ofthe artifact attenuation method.The capability of the method to reduce the artifacts in a clinical data set was47CHAPTER 3. MOTION ARTIFACT REMOVAL?30 ?25 ?20 ?15 ?101012141618202224NMSE (dB)Artifact Attenuation (dB)  ?=0.01?=0.15?=0.01?=0.15?=0.01?=0.15Subject 1Subject 2Subject 3Figure 3.10: Artifact power attenuation versus NMSE for 3 subjectsdemonstrated. The method was also compared with regular wavelet denoising andit was shown that for a given level of distortion, the proposed method yields higherartifact attenuation.Our method was based on an additive model for interference caused by motion.Assumption of additive noise model for motion artifact is not uncommon in theliterature. The Kalman filtering method used by Lee and Izzetoglu models motionartifact as additive observation noise [65, 84]. Some methods based on adaptivefiltering are also based on the assumption that motion artifact noise is additiveand can therefore be removed by subtracting the estimated noise from the signal[63, 76].The attenuation in artifact power achieved by this method may change fromone subject to another due to differences in total number of motion events and theirintensity. This has also been reported in the form of variability in Signal to NoiseRatio (SNR) across subjects in earlier works [63, 76]. There is also variability inartifact attenuation in different channels. This can be explained by noting that dueto the nature of the method, stronger artifacts are better isolated by wavelet trans-form and also can be better separated from the background fNIRS signal. There-fore, this method works best for spike-like artifacts and artifacts with significantlylarger amplitudes and shorter duration compared to physiological changes in the48CHAPTER 3. MOTION ARTIFACT REMOVAL0 10 20 30010203040Original Artifact Energy (dB)Artifact Attenuation (dB)  700nm830nm700nm Reg.830nm Reg.?10 0 10 20010203040Original Artifact Energy (dB)Artifact Attenuation (dB)  700nm830nm700nm Reg.830nm Reg.?10 0 10 20 30010203040Original Artifact Energy (dB)Artifact Attenuation (dB)  700nm830nm700nm Reg.830nm Reg.Figure 3.11: Correlation between original artifact intensity and artifact atten-uation in different channels for 3 subjects.fNIRS signal. This can explain the variability of artifact attenuation in Figure 3.7.Channels 1 and 2 for example, are located on the edge of the chevron shape optodeholder and are more likely to lose contact with the tissue due to head?s motion andtherefore yield higher attenuation in all subjects. Figure 3.11 shows the plots of ar-tifact attenuation versus original artifact energy for all channels in 3 test subjects.Original artifact energy was normalized to the energy in a reference segment of thesignal. The reference segment for each subject was manually selected such thatit represents the baseline state of the signal. There is a significant correlation be-tween artifact attenuation and original artifact energy ( R2=0.9584 and R2=0.9636for 700nm and 830nm channels in subject 1 , R2=0.8337 and R2=0.8327 for 700nmand 830nm channels in subject 2 and R2=0.9741 and R2=0.9380 Subject 3). Thisconfirms that original artifact energy explains the variability in artifact attenuation.Despite the fact that this method is designed and works best for spike-like arti-facts, it has been shown that the method could work well on more subtle types ofartifacts as well and in particular that it can be effective in reducing motion artifactsthat are correlated with the evoked cerebral response [75].The differences between our method and two other wavelet based methods forNIRS motion artifact removal should be emphasized here. Sato et al [85] used acontinuous wavelet transform based approach to detect blocks contaminated withmotion artifacts using a hard threshold on the wavelet transform amplitude in asubset of scales without attempting to remove them. These scales were identifiedthrough Monte Carlo simulation on a training data set. In the method of Robertsonet al., filtering is based on fixed threshold for each level as well as availability ofoptical motion reference signal [63]. Thresholding takes place only if the signal49CHAPTER 3. MOTION ARTIFACT REMOVALFigure 3.12: ?Box plots of the AUC0?2, AUC ratio and within-subject SDcomputed for all techniques and for both HbO (upper row) and HbR(bottom row). The red line in the box plot indicates the median, whilethe two extremities of the box plot represent the first and third quar-tile. Red crosses indicate outliers. The lines above linking the differ-ent techniques represent the significant statistical difference (p <.05 ifthe line is blue, p <.01 if the line is red).? [75] (Reprinted from Neu-roImage, Brigadoi S. et al., Motion artifacts in functional near-infraredspectroscopy: A comparison of motion correction techniques appliedto real cognitive data. Page 6, Copyright 2013, with permission fromElsevier Academic Press).amplitude is larger than the threshold in its source or detector reference signals.This is different from our method in that we do not have any reference signal andwavelet coefficient shrinking is only based on the probability of having an artifact.The wavelet level selection in our method is adaptive and changes with the degreeof contamination in the signal.The performance of some common fNIRS motion artifact removal methods,50CHAPTER 3. MOTION ARTIFACT REMOVALincluding the one introduced in this chapter, were compared using simulated andexperimental data in two recent studies [74, 75]. Using experimental NIRS datacollected from 22 subjects, Brigadoi et al. compared the performance of differentmotion artifact reduction methods [75]. The methods compared in this study in-cluded Kalman filtering [65], Correlation-based Signal Improvement (CBSI) [67],Principal Component Analysis (PCA) [131], spline interpolation [66], trial rejec-tion and the wavelet based method introduced in this chapter and also describedin [4]. In this study, subjects participated in a color-naming of a non-color wordtask during which the participants were asked to say aloud the name of the colorof the word that appeared on a computer screen [75]. In this particular task, a mo-tion artifact is caused by the jaw motion as the subject performs the task. Using 5criteria, the performances of the methods in recovering the hemodynamic responsewere compared. The criteria used in this study were AUC for the mean HRF dur-ing the first 2 seconds of stimulation (AUC0?2), the ratio of the AUC between 2to 4s to that during the first 2 seconds (AUC2?4), mean of the standard deviationof the hemodynamic response in each trial, between-subject standard deviation ofthe hemodynamic response and the number of trials averaged for every subject tocompute HRF.The summary of the results for three of the metrics in this study is reproduced inFigure 3.12 [75] 3. According to this study, the proposed Wavelet method, CBSI,Kalman and PCA 97 showed lower values of AUC0?2 with less variability. Theproposed method was found to be most effective in reducing AUC0?2. As forthe AUC ratio, CBSI and Kalman techniques had the highest AUC followed byWavelet. The proposed method along with PCA 97 outperformed other methodsin reducing the within-subject standard deviation with the Wavelet technique re-ducing standard deviation in 100% of the cases (see Figure 3.12). In this study,the proposed Wavelet method was also shown to be the only method to be able torecover all trials [75].3Reprinted from NeuroImage, Brigadoi S. et al., Motion artifacts in functional near-infrared spec-troscopy: A comparison of motion correction techniques applied to real cognitive data. Page 6,Copyright 2013, with permission from Elsevier Academic Press51Chapter 4Non-Invasive Optical Monitoringof Bladder Filling to CapacityUsing a Wireless NIRS DeviceLack of bladder fullness sensation is an issue that arises in different neurogenicconditions and in addition to influencing patients? quality of life, can result in seri-ous kidney damage. We describe a wireless wearable sensor prototype and methodfor detecting bladder fullness using NIRS. The sensor has been tested in vitro andin vivo to verify its feasibility and is shown to be capable of detecting changes inbladder content non-invasively. The work in this chapter was accepted for publica-tion in the IEEE transactions in biomedical circuits and systems and is also underreview by the University of British Columbia University-Industry Liaison Office(UILO) for potential IP protection and licensing.4.1 IntroductionThe importance of a wearable wireless device capable of monitoring bladder con-tent to help individuals with bladder control problem was discussed in Chapter 1.In this chapter, we describe the development, evaluation and pilot testing of a NIRSprototype for noninvasive optical monitoring of bladder filling to capacity usinga compact wearable wireless system. We propose using this small, low-weight,52CHAPTER 4. BLADDER FILLING MONITORING WITH NIRSinexpensive, wireless and easy to use device as a noninvasive method for moni-toring the point in time when the bladder becomes full, with lower computationalrequirements and complexity compared to ultrasonic continuous measurement sys-tems. Our method employs the properties of NIR light absorption of human tissueand water to measure changes in water content in the field beneath a NIRS device.Because the bladder rises out of the pelvis below the anterior abdominal wall asurine accumulates within the organ, this device can detect when a bladder capacitypreviously defined by ultrasound is reached. When the bladder rises into the NIRlight field as it fills, the water in the urine it contains results in high light absorptionthat generates an abrupt decrease in the light intensity sensed returning to the NIRSdevice. This event can be set to activate an alarm; potentially benefiting patientswith any of the problems related to an inability to sense when their bladder is full.4.2 Detection of Bladder Filling to Capacity Using NIRSThe major absorbing chromophores of physiologic interest in NIR wavelength win-dow, as described in Chapter 1 are HbO2 and HHb, as indicated in Figure 1.1.Water which is the main compound in urine (95% [132]), also has an absorptionpeak at 975nm and this peak can be used to detect urine content in the bladder anddifferentiate between an empty bladder, one with low volume, and a full bladder.In NIRS, light in the NIR window is used to interrogate the tissue. A lightsource (emitter optode) is placed on the skin surface, with a detector (receiveroptode) placed a few centimeters away. Changes in the light attenuation due toabsorption of the transmitted light by chromophores in the tissue (HbO2, HHb andwater) are detected by the receiver optode. The resulting changes in raw opticaldata are then converted to changes in chromophore concentration. A commonmodel used for this purpose, as described in Chapter 1 is the MBLL [16]:A? =?logII0= (?i?i(? )ci)BL+G (4.1)Where A? is the light intensity attenuation at wavelength ? , I0 is the source inten-sity, I is detected light intensity, ?i(? ) is the extinction coefficient of chromophorei at wavelength ? in Lmol?1cm?1, ci is the ith chromophore concentration in53CHAPTER 4. BLADDER FILLING MONITORING WITH NIRSmolL?1, L is the interoptode distance in cm, B is the differential pathlength fac-tor and G is an additive term to take fixed scattering losses into consideration. Thismodel is usually used in differential form to measure concentration changes in thetissue. The effective depth of penetration in this method is approximately half theinter-optode distance (see Section 1.6.1 for a discussion).Self contained wireless NIRS devices have been utilized for a wide range ofstudies involving brain, muscle, and the bladder [87]. Such devices have the ad-vantages of imposing less motion restriction, which means subjects can engage inrelatively more active physical pursuits, and suitability for longer term monitor-ing in ambulant patients [87]. Wireless NIRS devices often use LED as the lightsource. Although LED based NIRS systems have a broader spectrum comparedto laser-based NIRS devices, they have the advantages of being small, low weight,inexpensive, compact and self-contained and can be applied directly on the skinsurface without need for the fiber-optic cables required for laser systems.The hypothesis for our NIRS-based method for monitoring the level of urinein the bladder and to detect bladder filling to capacity was that with an LED lightsource using a wavelength close to the absorption peak of water at 975 nm, a self-contained NIRS device placed on the abdominal skin would detect water (urine)when the bladder enlarged into the NIR field. Ultrasound data indicates that asthe bladder fills naturally the dome of the organ rises within the abdominal cavitybringing the bladder and the urine it contains into the NIR light field [133]. The wa-ter contained in the bladder then absorbs light causing a decrease in detected lightintensity. Here, we describe a prototype of such a device as a proof of principle.While this method is similar in concept to the method presented for continuousbladder monitoring using ultrasound [68], in our method it is the urine in the blad-der (rather than the anterior wall of the bladder) which triggers the alarm. The levelof bladder fullness that corresponds to the urine capacity that needs to be detectedwill depend upon the patients symptoms, and his/her underlying medical condi-tion. In later development phases, this capacity value can be defined for individualpatients, and the fullness and position of the bladder beneath the abdominal skinthat this volume corresponds to can be assessed by ultrasound. The NIRS device isthen positioned on the abdominal skin so that it alarms when the bladder reachesthe size that corresponds to the capacity required for that patient.54CHAPTER 4. BLADDER FILLING MONITORING WITH NIRSFigure 4.1: Sensor block diagram.4.2.1 ElectronicsThe hardware consists of a 60?70?20 mm wireless NIRS device, that is worn bythe subject on the abdominal skin. The sensor can either operate offline by stor-ing the data on board or in real time via a link to a base Personal Computer (PC)through a wireless USB dongle. The block diagram of the sensor is shown inFigure 4.1. The sensor is made using commercially available components on a 2-layer Printed Circuit Board (PCB) and is enclosed in a custom made 3-dimensionally(3D) printed enclosure as shown in Figure 4.2. The device weighs 55 grams. Allcomponents except the source LED and the detector are mounted using standardsurface mounting technology. The source and detector are mounted on the frontside of the enclosure using adhesive glue and are wired to the main PCB.55CHAPTER 4. BLADDER FILLING MONITORING WITH NIRSFigure 4.2: Sensor components (top) and device exterior view with extrudedsource and detector (bottom)All the signal controls, sampling and processing are performed by a 16-bitlow power microcontroller (MCU)(MSP430F2274 Texas Instruments, TX, USA)running at 16 MHz.The source LED is a 950 nm LED (OSRAM Opto Semiconductors, 55 nmspectral half width, 16 mw nominal power) driven by a constant current driver,that in turn is controlled by a hardware timer. Even though the absorption peak ofwater is at 975 nm, the 950 nm source output is still highly absorbed by water as thespectral bandwidth of the source covers 975nm wavelength. The light detector is56CHAPTER 4. BLADDER FILLING MONITORING WITH NIRSa 5.22 mm2 silicon photodiode integrated with a Transimpedance Amplifier (TIA)(OPT101, Texas Instruments). The responsivity of the detector is 0.45 A/W at 950nm and the TIA is set to provide a gain of 6?106 V/A and a bandwidth of 2.5kHz. The amplifier?s output is filtered and sampled by 10 bit Analog to DigitalConverter (ADC) integrated on the MCU. Prior to sampling by ADC, the outputof the amplifier is filtered by an active twin T notch filter with center frequency at60Hz to remove interferences from AC power line coupling and ambient lightingfollowed by a first order lowpass filter with fc=5 kHz.The sensor is powered by a 3.7 v, 850 mAh lithium-ion polymer rechargeablebattery that provides up to 20 hours of continuous monitoring. The battery voltageis regulated down to 3.3 v through a low dropout linear regulator. The battery isrecharged through a mini-USB connection.In case of offline standalone operation the sensor can log data on the 16 kBonboard flash memory storage. The data can be later downloaded into a PC forfurther analysis.Two communication interfaces are supported: wired using USB 2.0 connectionand wireless using wireless link and a wireless dongle connected to a PC.The wireless link uses 868-915 MHz band for communication and transfersdata at 250 kbps. A wireless module based on Texas Instruments CC110L radiotransceiver is used (A110L, Anaren Microwave Inc, NY, USA). The MCU com-municates with the module over the Serial Peripheral Interface (SPI) bus at 250kHz. The wireless link allows remote start and stop of data collection through PC,download of the logged data and real-time data streaming to the PC with a rangeof up to 20m.A triple axis accelerometer (ADXL345, Analog Devices Inc., MA, USA) isused to detect motion to remove motion corrupted data segments. The accelerom-eter shares the SPI bus with the wireless module.The sensor is encapsulated in a custom made 3D printed enclosure (Verowhitepolyjet resin). An extruded feature that houses source and detector provides highercoupling with the tissue and also reduces the ambient light interference (see Figure 4.2).57CHAPTER 4. BLADDER FILLING MONITORING WITH NIRSFigure 4.3: Timing of sampling.4.2.2 FirmwareThe firmware controls the source LED timing, data sampling process, logs dataand communicates with a PC for command reception or data transmission.The scattering and attenuation of light in the tissue result in 6-7 orders of mag-nitude decrease in signal power. As a result, to have better SNR at the detectoroutput, higher source optical power is desired to increase the number of photonsthat can reach the detector. However, to limit the total tissue exposure and mini-mize the possibility of tissue thermal overheating, the power has to be kept withina safe range. An average power limit of 2 mW can be considered safe and has beenused as the limit for similar NIRS devices [57, 134]. This power level also fallswithin the safe radiation levels defined in IEC 62471. To achieve this power levelwhile having high instantaneous power, a source-switching scheme is employed asshown in Figure 4.3. We chose a 30 ms delay between the LED activation times.We also empirically found 60 mW of instantaneous power to result in well de-tectable light levels as the light exits the tissue for our interoptode distance of 3cm. As a result, the source LED needs to be activated for a maximum of 800 ?swith instantaneous power of 60 mW which corresponds to a driving current of 370mA (Figure 4.4) in order to keep the average power below 2 mW. This scheme also58CHAPTER 4. BLADDER FILLING MONITORING WITH NIRS0 0.02 0.04 0.06 0.08 0.1 0.12?2000200400600Time(s)Amplitude (mA)Figure 4.4: Snapshot of the LED driving current.reduces the total power consumption.To ensure accurate timing of LED driving pulses and ADC conversion triggers,the LED is driven directly by a hardware timer which is programmed to producepulses every 30 ms. A separate timer triggers ADC conversion for sampling theLED light level. The TIA bandwidth of 2.5 kHz results in an approximate risetime of 140 ?s for the LED pulses. Therefore, sufficient delay before sampling isnecessary to allow the transients at the detector output to settle. We used a delay of600 ?s as shown in Figure 4.5, which shows the signal at the detector output alongwith the sampling trigger signal.The detector?s output signal is initially sampled at 83 kS/s and a total of 8samples are recorded. These samples are then averaged and stored in a buffer. Thissampling rate allows use of a low order antialiasing filter and collection of sufficientsamples during the LED activation time. The next sampling cycle occurs in 30 msand follows the same pattern. This is equivalent to sampling the continuous opticalsignal at 83 KS/s, low pass filtering it with a moving average filter of length 8 andthen down-sampling the result to approximately 33 Hz. The digital averaging helpsreduce the high frequency noise.To prevent potential interference from ambient lighting, background light levelis sampled as the baseline and subtracted from the detected light level. The firmwaretherefore takes a sample from background light level 800 ?s after turning thesource LED off in each sampling cycle. This delay ensures all transients have59CHAPTER 4. BLADDER FILLING MONITORING WITH NIRS?2 0 2 4 6 8 10 12x 10?400.51Time (s)Amplitude (V)Figure 4.5: Detected signal at the detector for a source-detector separation of3 cm (blue) with the sampling trigger signal (green). Eight samples arerecorded with the source active and One sample is taken with source offwhich reflects the background lighting level.settled and the background light level is being sampled properly.The baseline corrected value is placed in a data packet along with a time stampand transmitted wirelessly to the PC. In case of offline operation, it is logged ontothe onboard flash memory.The sensor operation can be controlled either by wireless commands through aPC or, for offline data collection, by user push button on the sensor.4.2.3 PC InterfaceA Graphical User Interface (GUI) based on MATLAB (Mathworks, MA, USA)is developed for remote controlling the sensor, streaming data from the sensor andsaving it to a file for long term monitoring, downloading data stored on the sensor?smemory and processing the signal in real time or offline (linear filtering, trendremoval, etc.). A snapshot of the GUI with a sample data set is shown in Figure 4.6.The top panel shows the real time trace of the signal or the loaded data. The bottompanels contain controls for wireless operation and USB wired modes.60CHAPTER 4. BLADDER FILLING MONITORING WITH NIRSFigure 4.6: PC user interface screen shot.4.3 Performance EvaluationThe sensor?s dark noise was measured by readings obtained by placing the sensorin a dark room with no light incident on the detector for one hour. The Root MeanSquare (RMS) value of the noise in this setup was calculated as:Vn =?1N ?(x[n]? x?)2 (4.2)where x[n] is the signal read by the device in analog-digital conversion units.This process was repeated for a couple of measurements to obtain an estimate of theprototype?s noise voltage. This value was calculated to be less than 470 ?V . TheNoise Equivalent Power (NEP) was then calculated from dark noise measurementsusingPi =VnR(? )G (4.3)where Pi is the incident light equivalent power in W, R(? ) is the responsivity of61CHAPTER 4. BLADDER FILLING MONITORING WITH NIRSthe detector at ? = 950 nm an A/W, G is the TIA gain in V/A and Vn is the noisevoltage. The NEP is calculated to be approximately 180 pW. This defines thedetection sensitivity for a signal to noise ratio of unity and is the minimum lightlevel detectable by the sensor.The long-term stability of the sensor was evaluated by continuous recording ofdata from a phantom using the sensor for 30 minutes after a warm up period of 1minute. The aqueous phantom was prepared using method described in [135]. Thephantom scattering and attenuation parameters are chosen to be close to those ofabdominal tissue (in particular, abdominal fat with attenuation coefficient ?a = 3cm?1 and reduced scattering coefficient ? ?s = 3.3 cm?1 [136], see Section 4.6 for adiscussion). The phantom was made with 20% intra-lipid mixed with ink to obtaindesired optical parameters. The difference between the initial and final readingnormalized to the initial signal value was recorded as the drift. The device shows1.5% drift over the period of 30 minutes.4.4 In Vitro Evaluation4.4.1 In Vitro SetupTo verify the capability of the sensor in detecting bladder level changes in vitro,a simple setup as shown in Figure 4.7 was employed. The setup was made tosimulate the bladder, urine and the abdominal tissue during bladder filling andvoiding. A latex balloon was submerged in a phantom prepared as described in theprevious section in such a way that the balloon neck is attached to the top of thecontainer. The balloon can be filled with water from the top using a syringe. Thedistance of balloon from the side-walls was 1.5 cm when full and 6 cm when empty.The sensor was placed on the side-wall of the cylindrical container and securedwith medical adhesive tape (3M, MN, USA). The data was recorded wirelessly.62CHAPTER 4. BLADDER FILLING MONITORING WITH NIRSFigure 4.7: Schematic diagram of the in vitro setup for simulating bladderfilling and voiding (left) with a picture of the setup (right).4.4.2 ResultsFigure 4.8 shows a sample recorded data when the balloon is filled and emptied.The intensity readings from the sensor were converted to attenuation asA =?log II0(4.4)An increase in the amount of water in the optical path results in decrease inlight intensity and therefore an increase in attenuation (A). As the balloon fillingbegins around t = 9s, the absorbance increases up to the point where the balloon isfilled around t = 10s. Similarly, when voiding starts, the absorption reduces untilthe balloon is emptied.The drop in the signal level between the end of the filling and the beginning ofvoiding is caused by motion of the balloon at the end of filling cycle as the resultof ending water flow. The same occurs at the beginning of the voiding.63CHAPTER 4. BLADDER FILLING MONITORING WITH NIRS6 8 10 12 14 160.30.310.320.330.340.35Time (s)A (OD)Figure 4.8: In vitro recorded data when the balloon is filled and emptied.Red, green, black and cyan lines indicate beginning of filling, end offilling, beginning of voiding and end of voiding, respectively.4.5 In Vivo Evaluation4.5.1 Materials and MethodPilot data on 1 subject has been collected in 6 independent trials with the deviceduring voiding to verify if the sensor is capable of differentiating between full andempty bladder. The sensor is placed 2 cm above the symphysis pubis across themidline during voiding as shown in 4.9 and is secured using medical adhesive tape.The absolute intensity reading from the detector is then converted to attenuation ac-cording to Equation 4.4 and used for comparison between full and empty bladder.Data was transmitted wirelessly to a PC for recording. For this proof of princi-ple test, the motion rejection feature of the sensor using the accelerometer was notused.4.5.2 ResultsFigure 4.10 shows a typical attenuation signal recorded at source-detector sepa-ration of 3 cm. The red (solid), green (dashed) and black (dotted) vertical linesindicate permission to void, beginning of voiding and end of voiding, respectively.The signal shows a fall at the start of voiding and then plateaus around 15 s afterbeginning of voiding. This is possibly due to the fact that as the voiding begins,64CHAPTER 4. BLADDER FILLING MONITORING WITH NIRSFigure 4.9: Sensor placement for in vivo device test.70 75 80 85 90 95 1000.0450.050.0550.060.065Time (s)A (OD)Figure 4.10: Attenuation detected during voiding with source-detector sep-aration of 3 cm. The red (solid), green (dashed) and black (dotted)vertical lines indicate permission to void, beginning and end of void-ing, respectively.65CHAPTER 4. BLADDER FILLING MONITORING WITH NIRS Full Bladder Empty Bladder (OD)Figure 4.11: Comparison of detected light attenuation in full and empty blad-ders.the bladder dome is in the light path between the source and the detector. As thebladder shrinks, the urine level in the light path reduces and light intensity at thedetector increases (decrease in attenuation). At a certain point (in this case around15 s after voiding begins), even though the voiding continues, the bladder is nolonger visible to the sensor?s light and therefore no further change in detected lightintensity is observed.Figure 4.11 shows the light attenuation changes between full and empty blad-der for 6 independent trials with urine volume ranging from 300 ml to 700 ml. Asignificant difference in light absorbance is observed between pre- and post voidingstates as shown in the figure (paired t-test p<0.01). The starting point or baselineis variable among trials as a result of differences in light coupling, geometry, etc.However, there is a consistent difference between pre- and post voiding in the trialsas a result of the change in bladder content (mean of the differences: 0.022 withstandard error of mean of 0.0096).66CHAPTER 4. BLADDER FILLING MONITORING WITH NIRS30 35 40 45 5060708090Time (s)uMolFigure 4.12: Water concentration in tissue changes during voiding usingOxymon MK III (Artinis BV, The Netherlands) as a general purposedesktop reference spectrophotometer with an interoptode distance of 3cm. The red (solid), green (dashed) and black (dotted) lines indicatepermission to void, beginning and end of of voiding, respectively.Table 4.1: System level parameters of the sensorPower consumption in standby 63 mW (19 mA @ 3.3 V)Active power consumption with radio transmission 182 mW (55 mA @ 3.3 V)Active power consumption without radio transmission 122 mW (37 mA @ 3.3 V)Range 20 mLight output power < 2 mWCost < 40 $For comparison Figure 4.12 shows the concentration changes detected duringa separate voiding session when using a general purpose desktop laser-poweredreference spectrophotometer (Oxymon MK III, Artinis BV, The Netherlands) with971 nm laser for detection of water. The pattern of change in absorption is sim-ilar to those obtained with our prototype, even though the 971 nm signal is moresensitive to changes in water content.Table 4.1 shows the overall system level parameters for the designed proto-type. Active power consumption with and without radio refer to the cases whenthe device is linked to a PC and when the device is operating independently.67CHAPTER 4. BLADDER FILLING MONITORING WITH NIRS4.6 DiscussionWe have developed a novel optical method for non-invasive monitoring of bladdercapacity using a compact wireless NIRS prototype incorporating an LED with awavelength of 950 nm and demonstrated the feasibility of using this device placedon the abdominal skin to detect a signal change that indicates when the bladder andthe urine it contains have left the monitoring field of the device. Our data supportour hypothesis that when the bladder fills and enlarges, the urine within the blad-der can be detected using NIRS with a light source close to the absorption peak ofwater at 975 nm. Further validation of our NIRS-based method to detect when anindividual?s bladder capacity reaches a pre-defined limit is required, along with de-velopment of appropriate decision making process for activating filling alarm andcomparison of the data obtained to results from ?gold standard? ultrasonic blad-der scanning. Definition of the limit of bladder capacity will vary for each patientdepending on their clinical condition, but once defined and the landmarks of thebladder with this capacity established by ultrasound, the device alone should suf-fice for monitoring when the patient?s desired capacity is reached. Currently, thereis no alternative method and device for continuous bladder filling detection, a veryimportant clinical issue especially in patients with different types of urinary incon-tinence and patients with spinal cord injury.Variations in fat layer thickness are a potential limitation during NIRS measure-ments [137]. However, since the thickness of the fat layer remains constant as thebladder fills and empties, no effect on light attenuation relevant to the monitoringof bladder capacity is generated. In obese subjects light absorption by a signifi-cantly thicker fat layer can be anticipated. This would result in an overall decreasein signal level and also increase the distance of the sensor from the bladder. Thisproblem can be addressed to some extent by increasing the interoptode separationwhich effectively increases penetration depth along with shorter and higher powerLED pulses. However, if the fat layer is too thick, it could prevent sufficient NIRlight from reaching the bladder for our system to function.The output of the system has a drift as described in Section 4.3 . This driftcan mostly be attributed to slight temperature changes that result in a drift in theLED?s output intensity. Even though this drift is negligible in short term compared68CHAPTER 4. BLADDER FILLING MONITORING WITH NIRSto changes during bladder voiding for example, it can add up in long term monitor-ing. To avoid this, the device can be set to restart filling monitoring after each alarmto prevent the error from accumulating over time. This process of resetting the de-vice may also be required to account for the wide range of changes in absoluteoptical signal values read by the device. As shown in Figure 4.11 and discussed inSection 4.5.2, the parameter of interest is the change in signal attenuation. How-ever, the large variations in the initial values may be a potentially serious limitationin using the device as a continuous monitoring system. The reason is that the de-vice needs to register the initial value in order to measure the changes and if theinitial value is dependent on parameters such as coupling, determination of a fixedthreshold will be very hard. One approach to address this problem is restarting themeasurement every time the device is placed on the bladder, or when the device isrepositioned. A better and more robust alternative approach is to use methods thatcan take into account the changes in coupling such as the multidistance methodsdescribed in [138, 139].Even though water has a high absorption peak at 975 nm, HbO2 and HHb stillcontribute to absorption at this wavelength. This can be seen in Figure 4.6 andFigure 4.10 where absorption of light by HHb and HbO2 result in heart beat andrespiration systemic interference patterns appearing as small oscillations on thesignal. However, the contribution of the change in these chromophores concen-trations to the total signal attenuation during natural bladder filling compared tothat of water is relatively small [27]. Also the similarity in the data obtained fromour prototype and the reference spectrophotometer with wavelengths of 971 nmand 906 nm suggests that 950 nm wavelength is sufficiently sensitive for the mon-itoring function intended. However, if required, another pair of LEDs detectingchanges in HbO2 and HHb could be added to the device and software incorporatedto remove the effect of hemoglobin from the total attenuation signal.Similarly while motion induced artifact can be partially removed by the ac-celerometer on the sensor, incorporation of a dual source monitoring scheme woulddetect the coupling change or slow drifts caused by the small variations in the posi-tion of the device which can occur during continuous monitoring. In such a schemethe first channel would be placed over the bladder, with the second channel locatedfurther from the bladder so as to differentiate motion/coupling-induced changes in69CHAPTER 4. BLADDER FILLING MONITORING WITH NIRSthe detected intensity from those caused by bladder water content changes. Thechanges caused by motion or changes in optical coupling will be highly correlatedbetween the two channels, while changes caused by alterations in bladder capacitywill only affect the channel located on the bladder.The enclosure of our prototype was produced using a stereolithography based3D printer and due to limitations we had for the material, it was made using awhite color resin. This is not a recommendable option for a NIRS device. Thewhite color of the enclosure results in a dominantly scattering with low absorptionmedium for the light exiting the tissue while ideally, the enclosure should absorb allphotons leaving the tissue to closely simulate a semi-infinite boundary condition.For this reason, it is desirable to have a dark enclosure. Therefore, even though ourresults indicate the feasibility of this method and approach, this limitation needs tobe addressed in the future.The performance of our prototype was initially evaluated using a liquid phan-tom as described in Section 4.3 and Section 4.4. The values used for ?a and ? ?swere from data measured at 1064 nm [136]. The value of ?a at this wavelengthmay be too high compared to that of the prototype?s source LED wavelength at950 nm. This implies that the sensor?s performance is likely to be better than theresults of the in vitro test. Additionally, the use of a solid phantom is preferableto a liquid phantom as it provides more stable properties and does not require acontainer which may affect the results as the light passes through it to reach thephantom [140].The PC connectivity, in addition to providing an alternative method for devicecontrol as well as data processing and storing, can potentially be beneficial in caseswhere remote monitoring of a subject?s bladder activity is of interest. In UrinaryTract Infection (UTI), for example, which is a common condition in spinal cordinjury patients, the frequency of voiding increases and access to this informationcollected in normal daily life conditions by the clinician is important in treatmentof patients. In this case, the limited range of connection might limit the usageof wireless link to indoors only. However, the same benefits could potentially beoffered by replacing the PC with a smart phone in the future.For our device to reliably monitor ambulant subjects consistently and withthe level of accuracy required for detection of bladder capacity in selected patient70CHAPTER 4. BLADDER FILLING MONITORING WITH NIRSgroups, additional trials and development are required. In particular the potentialeffect of different body postures, positions and clinical conditions needs to be eval-uated. Some MRI studies have suggested that body position in young subjects doesnot affect the shape and position of the bladder significantly [141]. This needs tobe verified on our population of interest whose physiological conditions may differfrom young subjects. Data will also need to be collected in cohorts where the agerange and diagnostic criteria match those of the patients for whom monitoring witha device such as ours is considered of potential benefit.4.7 ConclusionWe have designed and developed a compact wireless optical sensor prototype forcontinuous non-invasive monitoring of the bladder in patients who are unable tosense when their bladder is full. This is a significant clinical problem in individualswith abnormal (neurogenic) bladder function, such as patients affected by MS,stroke and/or spinal cord injury, elderly patients with incontinence, and childrenwith persistent enuresis. The device is capable of differentiating between when thebladder is empty or contains a small volume of urine and when it becomes full, byusing the absorption properties of water at a wavelength of 950 nm . With such adevice used as a sensor with an alarm, it is hence feasible to warn the subject whenthe volume of urine in his/her bladder reaches a pre-determined threshold of thebladder capacity. This would potentially enable patients at risk for urinary retentionto protect themselves from renal damage, elderly subjects prone to incontinence toretain the ability to void voluntarily, and children with problematic enuresis tobecome conditioned to when they need to wake to void. Further clinical studieswith this device are required to validate this method.71Chapter 5Cortical Connectivity AnalysisUsing fNIRSThe interaction between spatially separated cortical regions plays an important rolein performing a cognitive task. Functional imaging methods such as fNIRS are ca-pable of detecting activated areas of the brain based on hemodynamic changesassociated with increased neural activity. fNIRS as an inexpensive and portableequivalent to fMRI can help identify functional or effective connections and inter-actions among cortical areas in a particular task. In this chapter, we first presentour preliminary method and results on detecting connections between brain regionsin a speech study using fNIRS and MVAR modeling. We then describe an analy-sis method for mapping resting state cortical networks using phase synchronizationand present results of applying this analysis method to fNIRS data from neonates tomap the language network. The preliminary material in the first part of this chapterwas published in International IEEE EMBS Conference in 2011 [5]. The materialin the second part of this chapter has been submitted and is currently under reviewfor consideration for publication.72CHAPTER 5. CORTICAL CONNECTIVITY ANALYSIS5.1 Functional Connectivity Using MultivariateAutoregressive ModelingMVAR modeling is a common approach to studying the interaction between brainregions in fMRI [142] and EEG [143]. MVAR can establish a direct measure offunctional relation between brain regions.We used MVAR modeling to measure time varying connectivity between tem-poral and frontal areas of neonates brain during a neurocognitive study using fNIRS.Higher temporal resolution along with non-confining nature of fNIRS makes it anatural choice for study of functional connectivity and its temporal evolution ininfants. Study of connectivity and its changes on infant can contribute to a betterunderstanding of the early learning process.5.1.1 Materials and MethodMVAR Modeling for Time Varying ConnectivityAn AR model for multichannel fNIRS signal can be written as [144]Y (n) =p?i=1A(i)Y (n? i)+ ?(n) n = p . . .N (5.1)where Y (n) = [y1(n) y2(n) . . .yL(n)]T is the L channel fNIRS measurement at timepoint n, p is the maximum lag and N is the total number of available samples.A(i)= [a jk(i)] is an L?L matrix in which a jk(i)?s are the AR coefficients describingy j(n) in terms of yk(n? i). a jk(i) can give a measure of connection in terms ofcausality between signals in different channels and shows how much of the energyof signal in channel j can be represented by signal in channel k. ?(n) is a normalidentically and independently distributed noise with zero mean. Equation 5.1 canbe rewritten asY = XA+E (5.2)where A = [AT (1) AT (2) . . .AT (p)]T is a (p ? L)? L matrix of MVAR coeffi-cients at lags 1 to p, Y = [Y T (n) Y T (n? 1) . . .Y T (p+ 1)]T is an (N-p)?L matrix,73CHAPTER 5. CORTICAL CONNECTIVITY ANALYSISand X defined asX =??????Y (n?1) Y (n?2) ? ? ? Y (n? p)Y (n?2) Y (n?3) ? ? ? Y (n? p?1)............Y (p) Y (n? p?1) ? ? ? Y (1)??????(5.3)is an (N-p)?(L?p) matrix.The maximum likelihood estimator of A is [144]A = (XT X)?1XT Y (5.4)In order to track possible changes in a jk in the time course of the signal, onecan divide the signal into smaller segments and estimate A in each segment:Am = (XTsmXsm)?1XTsmYsm (5.5)in which Xsm and Ysm are formed by replacing Y (n) with Ysm(n):Ysm(n) =???Y (n), mW ? n < (m+1)W0, else(5.6)in which W is the sliding window width. In other words, we fit the AR modelto a small window of the signals. The window is then shifted one sample in theforward direction and the model is fitted again to the data in the new window.In order to summarize the effect of AR coefficients at different time lags be-tween 2 channels, we define a connectivity index asc jk(n) =?pi=1 a? jk(i)2?Lk=1 ?pi=1 a? jk(i)2(5.7)where a? jk(i)?s are the elements of Am. ?pi=1 a? jk(i)2 represent the contributionof signal in channel k in minimizing the prediction error of AR model in channelj. Larger value for this parameter means information in channel k can be used tobetter predict values in channel j given the past values of both channels. This pa-74CHAPTER 5. CORTICAL CONNECTIVITY ANALYSISrameter has also been referred to as Direct Causality (DC) in the literature [145].The denominator in Equation 5.7 is the sum of such effects from all other chan-nels. Normalization ensures comparable values over different subjects. c jk(n) isevaluated for every time window as defined in Equation 5.6. We now define theconnectivity matrix as C(n) = [c jk(n)]. Each element of connectivity matrix C(n)shows the causal effect of channel k on channel j at time point n.fNIRS Experiment and DataThe purpose of this experiment is to study the changes in functional connectivityin neonates brain when exposed to two different types of audio stimuli. The ex-periment was originally designed to study the ability of neonates to learn simpleunderlying structures in speech [24]. To establish the feasibility of our method,we applied it to 3 representative cases from the original study [24]. The selectedsubjects were all female with ages 2,3 and 4 days, respectively. Informed consentwas acquired from parents when the experiment was being conducted. The studydesign was approved by the ethics committee of the Azienda Ospedaliera Universi-taria di Udine, Italy where the experiments were conducted [24]. During the 22-25minute long testing session, audio stimulus was administered to subjects while thesubjects were in the state of quiet rest or sleep. The audio stimuli consisted ofconsonant-vowel syllables organized into syllable pairs and were divided into 2major ?grammar? groups named ?ABB? and ?ABC? based on their syllables rep-etition order. Each grammar was presented in blocks of 18 seconds long followedby a silence of randomly varying duration (25-35 seconds). A total of 14 blocksfor each stimulus was presented. Figure 5.1-a shows the experiment design.The hemodynamic changes associated with increased neural activity in re-sponse to the 2 types of stimuli were monitored by an fNIRS device (24 channelHitachi ETG-4000 machine with 695 and 830 nm lasers, interoptode distance of 3cm and sampling rate of 10 Hz). The optode placement and the location of chan-nels is shown in Figure 5.2. The tragus and the vertex were used as landmarks foroptode positioning to ensure data is recorded from perisylvian and anterior brainregions.Earlier study using the same dataset indicated that neonates were capable of75CHAPTER 5. CORTICAL CONNECTIVITY ANALYSIS 	  	 	  	 	 (a) Experiment?s design950 1000 1050 1100 1150 1200 1250 1300 13500.40.60.8Time(s)Connectivity index(b) Representative connectivity indexFigure 5.1: Experiment design and a representative connectivity index (be-tween channels 2 and 5 for subject 3). Red and green lines denote thebeginning of the ABB and ABC blocks, while cyan and black indicatethe end of blocks.discriminating between the grammars [24]. The discrimination was indicated bysignificant increase in HbO2 in response to one type of stimulus in temporal andfrontal regions of neonates brain. The temporal region is known to be responsi-ble for auditory processing in infants [146] while the frontal areas are responsiblefor computation of structure and higher order representations in infants and adults[146]. Since the process of learning the grammar types involves 2 spatially sep-arate areas of the brain, it is natural to assume a functional connectivity networkshould be involved. The purpose of current pilot study was to use the data collectedin the same experiment and detect possible changes in such connections as a resultof exposure to stimuli using the proposed method.Before applying functional connectivity analysis, raw optical data collected byfNIRS device was converted to changes in HbO2 and HHb concentration usingMBLL [29]. The signals were highpass filtered to remove any overall trend in thesignals. A window of length 200 samples (20 seconds) was used to estimate AR76CHAPTER 5. CORTICAL CONNECTIVITY ANALYSISFigure 5.2: Side view of fNIRS optode holder overlaid on schematic repre-sentation of neonates head. The red and blue dots indicate the sourcelasers and detectors, respectively. The numbers between the dots are thechannel numbers. The optodes are placed such that they sample datafrom perisylvian and anterior brain regions.coefficients in each step according to Equation 5.6.Channels 1 to 6 on the left hemisphere were chosen to study the functional con-nectivity. This choice is based on the fact that the temporal region (represented bychannels 3 and 6) and frontal region (represented by channels 2,5 and possibly 1)are the major areas involved in processing audio stimuli and processing structures,respectively. Earlier studies have also shown that language function is left hemi-spheric dominant [24, 147]. Therefore, we limited our study to the left hemisphereonly. Also, only HbO2 changes were analyzed for this study. It has been shownthat HbO2 is more sensitive to regional cerebral blood flow changes [24, 148].MVAR model is estimated for channels 1-6. We are interested in overall con-nectivity difference between conditions (grammars), which means a function ofC(n) must be employed to summarize the connectivity matrix in each block for theconditions. We use simple averaging asc?Bijk =1M ?n?Bi c jk(n) (5.8)to form ?CBi = [c?Bijk] where Bi is the ith block of condition B, where B is either77CHAPTER 5. CORTICAL CONNECTIVITY ANALYSIStype ?ABB? or ?ABC?. M is the total number of calculated matrices in the block.The resulting connectivity matrices ?CBi are grand averaged to yield overall con-nectivity matrix for each condition in every subject. Blocks involving motion arti-facts are excluded from this procedure. Motion artifacts are identified by changeslarger than 0.5 mMol.mm/s in the concentration changes.5.1.2 ResultsFigure 5.1-b shows a representative connectivity index between channels 2 and 5(c52(n)). The duration of each stimulus is indicated by vertical lines. The figuresuggests that the connection between the 2 channels becomes stronger when thestimulus is being presented.Figure 5.3 shows the connectivity matrix for the 3 test subjects. Self connec-tions are not shown in the figure. Connections with strength of less than 15% ofmaximum strength in each subject are not shown in the connectivity network in theright panels of Figure 5.3. In order to differentiate conditions, overall connectivitymatrix for condition ?ABC? is subtracted from that of condition ?ABB? to yieldthe difference in average connectivity between the 2 conditions. This differencematrix shows channels whose connectivity is stronger in one condition comparedto the other. This is important as there may be larger and more complicated net-works involved in accomplishing a particular task while we are only interested inconnections which are stronger for the ?ABB? grammar.All three subjects demonstrate strong connectivity between temporal and frontalareas. This is indicated by connection from channel 6 to 2 and 5 in subject 1, 6 to2 in subject 2 and 2 to 6 in subject 3.Also in subject 1, channel 6 shows strong connection with channels 3 which inturn has connection with channel 5 in temporal region. Possible explanation canbe that channel 6 is the lowest/first level of auditory processing, its output feedsinto channel 3. The next level of auditory processing, channel 3 then connects withthe frontal area, channel 5 for higher level structural processing. This can also beobserved in subject 3.The connectivity matrices provide an overall comparison of functional connec-tions between temporal and frontal areas. Another interesting analysis would be to78CHAPTER 5. CORTICAL CONNECTIVITY ANALYSISFromTo  1 2 3 4 5 612345600.050.10.15(a) Subject 1FromTo  1 2 3 4 5 612345600. Subject 2FromTo  1 2 3 4 5 612345600. Subject 3Figure 5.3: Connectivity matrices and networks for 3 test subjects. Connec-tion strength is color coded. Only connection paths which are strongerin condition ?ABB? compared to condition ?ABC? are shown. The restare set to zero. Figures on the right show a graphical representation ofthe connectivity network overlaid on a head model (lateral view).79CHAPTER 5. CORTICAL CONNECTIVITY ANALYSIS0 2 4 6 8 10 12? index(a) Subject 10 5 10 index(b) Subject 20 2 4 6 8 10?0.0500. index(c) Subject 3Figure 5.4: Temporal evolution of connection strength between temporal re-gion and frontal region. For subjects 1 and 2, plots represent connectionfrom channel 6 to 2. For subjects 3, plot represents connection fromchannel 2 to 6. r2=0.97, r2=0.87, r2=0.76 for subjects 1 to 3, respec-tively.investigate the temporal evolution of connectivity matrices ?CBi across the blocks.The hypothesis is that this evolution should be associated with learning in infantsand should therefore change as the subjects are exposed further to the stimuli. Westudied this by investigating temporal evolution of connection strength betweenrepresentative temporal and frontal channels. Channels 6 and 2 are selected as theyhave strong temporal-frontal connection in all 3 subjects . Figure 5.4 shows theplots of connection strength vs block number. Each point corresponds to averageconnectivity strength within a stimulus block. All three subjects show an increasein connection strength in the time course of the experiment.5.2 Analyzing Resting State Functional Connectivity inThe Human Language System Using Near InfraredSpectroscopyAs discussed in Chapter 1, fNIRS can measure the neuronal activity in response toa task or an environmental stimulation through neurovascular coupling in superfi-cial areas of the brain. fNIRS has been used for functional studies as a portable,less expensive and less restraining alternative to fMRI in different task based brainfunctional studies. One of the more recent areas of interest in both fNIRS andfMRI is the study of interaction between different cortical areas through their in-80CHAPTER 5. CORTICAL CONNECTIVITY ANALYSIStrinsic neuronal signaling [97, 101]. This intrinsic signaling appears in the formof slow varying spontaneous fluctuations in the BOLD signal in the absence ofstimulation. These fluctuations are correlated between brain areas that are anatom-ically and functionally connected and have been used to map brain functionalnetworks such as sensorimotor, visual, and auditory as well as higher networkssuch as language and attention [149, 150]. These networks are often mapped fromthe data collected at rest and are therefore referred to as Resting State FunctionalConnectivity (RSFC) maps.The RSFC analysis can potentially identify changes in intrinsic neural activityas a result of disease in some neurological and psychiatric conditions. Changes inconnectivity strength in different brain networks have been observed in conditionssuch as autism [151], depression [152], Alzheimer disease [153] and attention-deficit hyperactivity disorder [154].Given the advantages of fNIRS, different brain networks have been investigatedthrough RSFC using fNIRS. One of the most common methods for analyzing brainnetwork connectivities using RSFC in fNIRS is cross correlation [149]. In crosscorrelation, an fNIRS channel is selected as the seed channel and the correlationsof the signal in all other channels with the seed channel are calculated. The objec-tive is to find the cortical areas whose resting state fluctuations are similar to thatof the seed channel. Cross correlation-based functional connectivity has been in-vestigated in conjunction with fNIRS to derive connectivity maps in different brainnetworks [97, 107, 108]. One drawback of correlation based connectivity is that itcan be sensitive to detection of spurious connection as a result of presence of crosstalk between channels, systemic interference or noise [155].In this section, we have investigated the phase relation between fNIRS chan-nels and have used it as a measure of functional connectivity. Compared to methodsbased on signal amplitude such as cross correlation, phase is much less sensitiveto noise and interferences. It also does not require the assumption of stationarityfor the signals. Phase synchronization is not equivalent to coherence or frequencysynchronization and is an independent characteristic of the interrelationship be-tween two processes [155]. To evaluate the feasibility of this analysis method fordetecting functional connectivity, we applied it to a study of processing of speechvs. non-speech in newborn human infants. This type of comparison is of particular81CHAPTER 5. CORTICAL CONNECTIVITY ANALYSISFigure 5.5: fNIRS experiment setup.value for the question being asked as there is a growing body of evidence on thebrain areas involved in language processing in neonates, but less on the underlyingconnectivity.5.2.1 Material and MethodsfNIRS DataThe fNIRS data was collected from newborn infants at BC Children?s hospital,Vancouver Canada, during a separate language perception study [156]. Informedconsent was obtained from parents when the experiment was being conducted. Thestudy design was approved by the ethics committee of the University of BritishColumbia. The experiment design and setup are shown in Figure 5.5. A total of19 subjects were used in the analysis out of which two subjects were excludeddue to severe artifacts in the signals and poor data quality resulting from optodedisplacement during data collection. During the experiment, audio stimulus wasadministered to subjects while the subjects were in state of quiet rest or sleep.The audio stimuli consisted of blocks of sentences in Spanish and Silbo-Gomero.Silbo-Gomero is a whistled language that is a surrogate language of Spanish. Ituses whistles rather than speech, and was developed by shepherds in the CanaryIslands to communicate across long distances. Spanish and Silbo-Gomero wereselected as both are unfamiliar to the infants, while one is a spoken language and82CHAPTER 5. CORTICAL CONNECTIVITY ANALYSISFigure 5.6: Optode placement on the head. Blue squares and red dots indicatedetectors and sources, respectively and the numbers indicate the channelnumber.the other is not. Each block was 15 seconds long followed by 25-35 seconds ofsilence. A total of 8 blocks for each stimulus were presented in which each blockconsisted of continuous speech. The total experiment time was 22-25 minutes.The subjects? brain hemodynamic response was monitored by a 24 channelfNIRS device (Hitachi ETG-4000 machine with 695 and 830 nm lasers at a powerof 0.75 mW, interoptode distance of 3 cm and sampling rate of 10 Hz). Twochevron shaped optode holders secured nine 1 mm fibers to the head. There werea total of 4 detector and 5 source fibers on each holder resulting in 12 recordingchannels per holder. Figure 5.6 shows the placement of optodes on the subject?shead. Surface landmarks (ears or vertex) were used for the placement of the probeholder over the infant?s perisylvian area of the scalp. Channels 11 and 12 in theleft hemisphere and 23 and 24 in the right hemisphere were ideally placed abovethe infant?s ear. A stretchy cap secured the holders on the infants? head.Data AnalysisIn order to determine the phase relation between channels, we first extracted thephase of the signal in each channel using the Hilbert transform. Hilbert transformconverts a real valued signal to a complex one, known as analytic signal, whose realpart and phase correspond to the original signal and its derived phase, respectively[157]. The Hilbert transform of signal x[n] in the frequency domain is defined as83CHAPTER 5. CORTICAL CONNECTIVITY ANALYSIS[157]:Y (e j?) =? jsgn(?)(X (e j?)) (5.9)where X(e j?)is the Fourier transform of x[n] and sgn(?) is the sign functionhaving value of 1 for ? > 0 and -1 for ? < 0. The analytic signal can then bewritten asxa[n] = x[n]+ jy[n] (5.10)where y[n] is the inverse Fourier transform of Y (e j?).We use the joint probability distribution of the phases across channels to de-scribe their connectivity. A common model for probability distribution of phasewhich is the circular analogue of the Gaussian distribution is the Von Mises distri-bution. The Von Mises probability density function (pdf) is defined as [158]:f (? |? ,?) = 12piI0(?)e?cos(???) (5.11)where ? is an angle defined in the interval [?pi,pi) and I0(?) is the modified Besselfunction of order 0. The parameter ? is the equivalent of the covariance for theGaussian distribution and ? is the expected value of the angle. The probabilitydensity function of the signal phase in channel m conditioned on that of channel ncan therefore be written as:f (?m ??n|? ,?) = 12piI0(?)e?mncos(?m??n??) (5.12)?mn describes the intensity of phase correlation between signals in channels m andn. In other words, it shows how much prior information of ?n affects distribution of?m. The first moment of the distribution given in Equation 5.11 can be calculatedas [158]m1 =E[e j? ] =? pi?pie j? f (? |? ,?)d? (5.13)= I1(?)I0(?)e j?Using the first moment, one can estimate parameter ? by numerically solving theoptimization problem84CHAPTER 5. CORTICAL CONNECTIVITY ANALYSISargmin?(|m1|??????1NN?1?i=0e j?i?????)2= (5.14)argmin?(I1(?)I0(?)??????1NN?1?i=0e j?i?????)2(5.15)where N is the number of samples in the data segment and ?i is the measured phaseof the signal at time point i. Parameter ? can then be estimated using? = 6 1N ?e j?i (5.16)A close relationship exists between parameters of the distribution and the phaselocking value (PLV) which is a common measure used in EEG signal processingto detect functional connectivity through synchronization between channels. PLVis related to the distribution through the magnitude of the first circular moment ofthe phase distribution [159]PLVmn =???E[e j(?m??n)]???= I1(?)I0(?)(5.17)One advantage of using Von Mises distribution over phase locking for con-nectivity analysis is that once the parameters are estimated, one would have thedistribution function and can re-sample from the distribution to determine the sig-nificance levels. Also, the preferred phase difference is not available in PLV.To evaluate fNIRS functional connectivity, we first calculated the phase fromall fNIRS channels using the Hilbert transform as described earlier. Ideally, wewould be interested in phase relations between channels when subjects are not ex-posed to any type of stimulation to reveal intrinsic network activities. It was shownin an earlier study, however, that the infant brain produces no significant responsein the language network to Silbo-Gomero stimuli [156]. We therefore used thefNIRS data during Silbo-Gomero stimulation as an alternative to resting state. TheBOLD spontaneous fluctuations are concentrated at frequencies of less than 0.1 Hz.Therefore, the signals were first filtered with an infinite impulse response bandpass85CHAPTER 5. CORTICAL CONNECTIVITY ANALYSISfilter (IIR) between 0.02-0.08 Hz to extract spontaneous hemodynamic activitiesand reject other interferences. This frequency range is comparable to those used inother studies investigating RSFC using fMRI and fNIRS [101, 108].fNIRS data in general can be contaminated with motion artifacts as the result ofsubjects? spontaneous movements. These artifacts create interference in the formof highly correlated phase changes in fNIRS channels, especially in spatially closechannels. This interference results in very high phase correlation and can obscureunderlying phase connections between channels. Even though filtering of the mo-tion artifacts is possible, in order to minimize possibility of introducing any interdependence between channels, no artifact removal procedure was applied. Instead,the channels for all subjects were inspected visually and artifact contaminated re-gions within the Silbo-Gomero stimulation window were marked. An artifact freesegment of the data in each channel was then selected for the analysis and thephase of the selected signal segments were then derived using the Hilbert trans-form. Since the brain shows no response to this stimulation type, the stimulationonsets were ignored and the segments were selected independent of the stimula-tion onsets. The segments contained variable number of stimulation blocks andtheir length ranged from 50s to 220s.The channel with the highest activation in the grand average for the Spanishstimulation task during the original study in the left hemisphere was selected asthe seed channel for the RSFC analysis. The joint phase distribution of the seedchannel and all other channels was then estimated by calculating the phase differ-ence between the seed channel and other channels and then estimating ?mn and ?mnusing Equation 5.15 and Equation 5.16. We used a simplex derivative-free methodto solve Equation 5.15 and derive ?mn [160]. The analysis was performed in MAT-LAB (Mathworks MA, USA) and the phase coupling estimation toolbox developedby Cadieu et al. was used for parts of the analysis [161] 1.The analysis was performed on HbO2 changes only. Previous studies on theapplication of fNIRS to detect language network activity and connectivity haveshown that HbO2 is more sensitive to regional cerebral blood flow changes thanHHb with the equipment used here [24, 107, 148].1http://redwood.berkeley.edu/klab/pce.html86CHAPTER 5. CORTICAL CONNECTIVITY ANALYSISTo examine the validity and reliability of the connectivity information derivedwith this method, we divided the subjects randomly into two groups and evaluatedthe connections for each group, similar to the approach proposed in [106]. We thencompared the correlation of the connectivities between the groups.Other studies have suggested that language network is left lateralized. We veri-fied this in the network derived using our method. The lateralization was quantifiedusing [106, 162]LI = 1M/2M/2?i=1? iL ?? i?R? iL +? i?R(5.18)where M is the total number of channels, ? iL is the value of ?is in which i is thechannel number and s is the seed channel in the left hemisphere. ? i?R is the value ofthe same parameter with the channel symmetric to i in the right hemisphere. Thesignificance level of the calculated lateralization index is then evaluated. The lat-eralization index results in a number between -1 and 1 with more positive numbersindicating higher degrees of left lateralization.As the final step, we defined 4 ROI, 2 inside the language network and 2 outsidethe network and evaluated the connection strengths in these areas. In particular,channels 6 and 7 were selected inside the network, based on our prior knowledgethat the physical area they cover is in the language network, and channels 1 and12 outside the network with the optode configuration used in the current study.Channel 1 is over the frontal areas while channel 12 covers the temporal area. Thechoice of these channels as being outside the language area is justified by the factthat they showed no significant activation in response to native language or Spanish[156].5.2.2 ResultsFigure 5.7 shows the HbO2 signal from channels 7 and 9, the seed, for a typicalsubject. Qualitatively, the histogram of the phase at different time points for bothchannels does not show a clear dominant phase range as shown in Figure 5.7-b andFigure 5.7-c. However, the joint distribution histogram shown in Figure 5.7-d hasa sharp peak focused around the mean phase difference. This is also evident inthe Von Mises distribution function plot derived from estimated parameters in each87CHAPTER 5. CORTICAL CONNECTIVITY ANALYSIS900 920 940 960 980 1,000 1,020 1,040 1,060 1,080?0.4?  Ch7Ch9(a)?3 ?2 ?1 0 1 2 3050100150200250? = 0.27, ? = ?1.58Phase (rad)Count(b)?3 ?2 ?1 0 1 2 3050100150? = 0.17, ? = ?2.16Phase (rad)Count(c)?3 ?2 ?1 0 1 2 30100200300400500600700? = 2.70, ? = ?0.42CountPhase (rad)(d)Figure 5.7: Filtered HbO2 signal recorded from two channels with high de-gree of connection (a) and the distribution of the phases for each channel(b and c) along with the joint phase distribution (d). The red curve inhistograms shows the estimated probability distribution.case (shown in red). The estimated distribution parameters are also indicated in thefigure. In the case of phase histogram for individual channels (Figure 5.7- b and c),the values of estimated ?mn are much smaller than in the conditional distribution(Figure 5.7-d). This indicates a high phase relationship between the two channelsand is interpreted as connectivity.Using the method described earlier, the group level resting state functional con-nectivity map with channel 9 chosen as the seed channel was derived and is shownin Figure 5.8. The detected network includes the areas known to be associatedwith language network including the superior temporal gyrus and Broca?s area.The maps are also in agreement with those obtained for the language network inadults using correlation based fNIRS connectivity studies [106].The connectivity maps resulting from the 2 random subgroups are shown inFigure 5.9. The maps for the two subgroup cover similar areas in both hemispheres.The correlation between individual connections in the two subgroups is shown in88CHAPTER 5. CORTICAL CONNECTIVITY ANALYSIS  ?00.511.522.53318116492712105  ?00.511.522.53201513182321161419242217Figure 5.8: Group level RSFC maps with channel 9 in the language area usedas the seed channel. Left and right panels correspond to left and righthemispheres.Table 5.1: Pairwise comparison between selected ROIs inside and outsidelanguage network (Tukey?s test).Channel Pair 7-6 7-12 7-1 12-1Mean ?? -0.15 1.65 1.57 -0.09CI (95%) [-1.43 1.13] [0.37 2.94] [0.29 2.85] [-1.37 1.20]Figure 5.10 (Pearson correlation r=0.6 ).Figure 5.11 shows the results of ROI connectivity analysis where the connec-tion strength between the seed channel and 2 channels in the language area (6 and7) is compared with that with two channels outside the language system (channels1 and 12). Analysis of variance indicates significant difference between the con-nections (ANOVA p<0.01) inside and outside the language network. In particular,channels 6 and 7 connections are not different while they are both higher than thatof channels 12 and 1 in the temporal and frontal areas, respectively. Results ofmultiple comparison test are shown in Table 5.1 (Tukey?s test).The results of the lateralization analysis are shown in Figure 5.12. The aver-age lateralization index is 0.172 and is significantly different from zero (1 samplet-test, p<0.001). Here, the lateralization index is also compared for all subjectsbetween the language network and a control case. The control network is createdby choosing channel 11 as the seed channel. There is a significant difference in lat-eralization index between the language and control network (paired t-test p<0.01).89CHAPTER 5. CORTICAL CONNECTIVITY ANALYSIS  00.511.522.531 23 4 56 78 9 1011 12  00.511.522.5313 1415 16 1718 1920 21 2223 24  00.511.522.531 23 4 56 78 9 1011 12  00.511.522.5313 1415 16 1718 1920 21 2223 24Figure 5.9: The connectivity map for the 2 subgroups with channel 9 used asthe seed channel. The top and bottom panels are results of subgroups1 and 2, respectively. Left and right panels correspond to left and righthemispheres.0.5 1 1.5 2 2.5 3 3.50.511.522.53Subgroup 1Subgroup 2Figure 5.10: Correlation between connection strengths in the two subgroups.90CHAPTER 5. CORTICAL CONNECTIVITY ANALYSISCh7 Ch6 Ch12 Ch101234kFigure 5.11: ROI connections comparison between language area (repre-sented by channels 7 and 6) and outside language area (represented bychannels 1 and 12). The bars indicate the standard error of the mean.Language Control? 5.12: Lateralization index for the language network and comparisonwith control network.These results suggest left lateralization in the detected language network.5.3 Discussion and ConclusionIn the first part of this chapter, we used MVAR modeling to identify changes inconnection strength in cortical network involved in a speech perception study onneonates. The hemodynamic changes associated with increased neural activitywere detected by fNIRS device. The purpose of this pilot study was to detect thechanges in functional connectivity in response to exposure to 2 different types ofstimuli.The cortical signals were modeled as a MVAR signal in which AR coefficientsrepresented connection strength at different lags. An overall connection strengthmeasure was defined and was evaluated for every block of the 2 stimulus types.The grand average of blocks in the 3 test subjects indicated strong connectionsfrom temporal to frontal areas. Connections were also observed from lower levelaudio processing areas to higher audio processing levels which in turn mediated the91CHAPTER 5. CORTICAL CONNECTIVITY ANALYSISconnection to structural processing regions. It should be noted that this preliminarystudy has the important limitation of low number of subjects and therefore, thestatistical significance of the results could not be verified.Another observation was the temporal increase in connection strength for onetype of stimulus compared to the other across experiment blocks. This is perhapsassociated with learning in the time course of experiment. The results of this pre-liminary study are functionally and neuroanatomicaly relevant which led us to thenext part of this chapter where a different type of connectivity was analyzed on alarger dataset.We evaluated use of phase synchronization to identify resting state functionalconnectivity in the language system in infants using fNIRS. We used joint prob-ability distribution of phase between fNIRS channels with a seed channel in thelanguage area to estimate phase relations and identify the language system net-work. Our results indicate the feasibility of this method in identifying the languagesystem. The connectivity maps are consistent with anatomical cortical connectionsand are also comparable to those obtained from fMRI functional connectivity stud-ies [163, 164]. The results indicate left hemisphere lateralization of the languagenetwork.Brain networks connectivity reveals information about underlying anatomicalareas involved in a particular task. In some disease conditions, changes in corti-cal connections occurs [149]. Application of connectivity estimating methods tofNIRS enables investigation of such changes in cases where use of fMRI is not pos-sible, such as in infants and extends utilization of fNIRS in wider range of clinicalapplications.Use of fNIRS for analysis of functional connectivity offers several advantagesover more traditional fMRI based connectivity analysis. Collection of fMRI datafrom infants and young children under resting condition can be challenging. Incontrary, fNIRS is easily applicable to even newborns. Also, in cases where sub-jects to be tested are immobilized and can not be transferred to an Magnetic Reso-nance Imaging (MRI) scanner, portable fNIRS systems can replace fMRI for con-nectivity analysis. One limitation compared to fMRI is the limited penetrationdepth which means connectivity analysis will be limited to cerebral cortex.Our results are comparable to similar studies in the literature. In particular92CHAPTER 5. CORTICAL CONNECTIVITY ANALYSISZhang et al. analyzed RSFC in the language system in adults using fNIRS [106].Their results indicated significant RSFC between left inferior frontal and superiortemporal cortices which are associated with language system [165]. Using fMRI,Fransson et al. studied resting state networks in the infants brain [166]. Theyobserved similar networks in the bilateral temporal/inferior parietal cortex whichencompasses primary auditory cortex [166].The presence of motion artifacts has a significant effect on the connectivitystrength results using phase synchrony method presented here. The motion artifactresults in in-phase changes across affected channels which may result in strongerphase correlation compared to those resulting from spontaneous neuronal activity.Therefore, care must be taken to ensure segments being processed do not includemotion artifacts. Saturated channels or channels that have lost coupling to tissuedue to displacement will also have similar effect.Our fNIRS data was not collected during strict ?resting? state. We used con-tinuous blocks of data in which subjects were listening to a non speech audio stim-ulation. No significant activation compared to baseline was observed on this dataand was therefore used as the baseline [156]. Some fMRI studies have followed asimilar approach for mapping RSFC. The study by Greicius et al on default modenetwork in Alzheimer?s disease patients for example, was performed during a lowdemand cognitive task [149, 167]. An alternative approach is to regress out thetask evoked response from the data before performing RSFC analysis [168].The fNIRS signal is known to contain systemic interferences. This includesinterference from cardiac pulsation, respiration, cardiovascular autoregulation andheart rate variability. The frequency band for connectivity analysis must be chosensuch that it includes the relevant variations caused by the neurovascular couplingwhile rejecting the frequency bands containing these interferences. The cardiacinterference in our study is around 2 Hz and the very low frequency interference(heart rate variability, cardiovascular autoregulation) is around 0.01 Hz. The res-piratory fluctuation is around 0.2 Hz. The frequency band we chose for analysis(0.02-0.08 Hz) reduces the effect of these interferences and therefore, connectivitydetection as a result of these interferences is less likely.Bivariate methods in general can result in non existing spurious connectionswhen there is a propagation of information from one channel to others. A pairwise93CHAPTER 5. CORTICAL CONNECTIVITY ANALYSISmeasure of connectivity will result in detection of connection between all possibleconnection pairs, ie. direct or indirect. However, since we are only looking tofind channels which belong to the same network, use of a bivariate measure ofconnectivity can be justified. Most methods for functional connectivity mappingin the literature based on fMRI or fNIRS also use seed based methods which isrelying on the bivariate concept of finding coherence/correlation between channelswith the seed channel [149, 167].In summary, the results of this work suggest that the proposed method canbe used to reveal underlying connectivity patterns of cognitive functions in theresting state through phase relations between hemodynamic changes in differentbrain regions. The results also indicate a left lateralization in the detected networkwhich suggests the language system may be left-lateralized already in newborns.94Chapter 6Design and Validation of aCustom fNIRS Device forMonitoring TMSIn this chapter, we describe the design and development of a custom-made con-tinuous wave fNIRS instrument for monitoring the effect of TMS on the brainactivation and connectivity.6.1 Background and MotivationTMS is a method of stimulating brain using strong magnetic pulses that activatecortical neurons through electromagnetic induction. TMS has been used as an ex-perimental tool for neurophysiological and psychophysiological studies. In orderto better understand and investigate the effect of TMS on the brain, neuroimag-ing techniques have been used concurrently with TMS. This allows studying thechanges in hemodynamics and neural activity both in the target brain area as wellas the areas closely related to it. However, the strong magnetic pulse produced bythe TMS coil introduces serious challenges for common neuroimaging techniques.The optical nature of the NIRS makes it immune to this type of interference andmakes it an appropriate tool for monitoring brain hemodynamics during concurrentTMS studies. This type of combined NIRS-TMS study allows one to not only mon-95CHAPTER 6. CUSTOM FNIRS SYSTEM DESIGNFigure 6.1: Block diagram of the overall fNIRS system.itor the effect of TMS on the target area, but also to investigate cortical functionalconnectivity changes in response to stimulation. Such a change in connectivitycan occur in a short time scale that does not last long after the stimulation, or in alonger time scale that outlasts the duration of the stimulation.This chapter describes a custom-made NIRS instrument for monitoring the ef-fect of TMS on brain hemodynamics and neural activity. When combined withthe connectivity analysis method described in the previous chapter, the instrumentcan be used for analyzing the effect of TMS on RSFC with potential applicationin stroke patients to study both short term and long term effects of TMS on brainnetworks.6.2 Instrument DesignFigure 6.1 shows the overall system setup. It consists of two laser sources, deliv-ery fibers, detecting fiber, a photo detector and a Data Acquisition (DAQ) system.Laser sources are amplitude modulated and the modulation signal is driven by aMicrocontroller Unit (MCU). The output of the laser sources are launched intotwo fibers (source optodes). The light is delivered to the target tissue through thesource optode. Another fiber collects the light from the tissue and delivers it toa photo detector. The signal from the detector is amplified and sampled by a PCthrough a DAQ system. A digital lock in amplification scheme is then implementedon the PC to measure attenuation of the diffusively reflected light.96CHAPTER 6. CUSTOM FNIRS SYSTEM DESIGN6.2.1 Light SourcesThe light source consists of 2 LDs at 780 nm and 830 nm (Sanyo Electric, Japan).LD?s emitted light is focused into an angled ferrule of a single mode fiber (5.6?m core and 125 ?m cladding) using an appropriate lens. The diode, lens andfiber ending are enclosed in a housing (Thorlabs Inc., NJ, USA). A standard FCconnector is mounted on the free end of fibers for each diode. The connector isattached to a coupler located on the front panel of the device. This allows externalaccess to the laser output. The laser diodes have integrated photodiodes whichprovide feedback to ensure constant power radiation.The output power of the LD is controlled by a closed loop control mechanismas shown in Figure 6.2. The maximum LD power is initially set and is then modu-lated by the input signal. The maximum deliverable current to diodes is also limitedby the driver for the overcurrent protection of the LDs.6.2.2 Source ModulationBoth of the source diodes are amplitude modulated to allow separation of changesin amplitude due to attenuation at the two wavelengths. The sine wave modulatingfrequencies are chosen as 1 kHz and 1.25 kHz. The frequencies are selected to behigh enough to avoid 1/f noise and low enough for the MCU and also for the sam-pling rate (amount of data that needs to be stored) and also satisfy the conditionsfor digital lock-in amplification. The higher frequency must not be a multiple ofFigure 6.2: Block diagram of the laser diode driver module.97CHAPTER 6. CUSTOM FNIRS SYSTEM DESIGNFigure 6.3: Direct digital synthesizer block diagram.the lower frequency to avoid harmonic interference.A Direct Digital Synthesis (DDS) scheme is implemented for driving the laserdrivers using a 16 bit MCU (MSP430, Texas Instruments, Texas, US). The blockdiagram of the DDS is shown in figure Figure 6.3. The timer produces an interruptat 100 kHz rate. The numerically controlled oscillator produces the value of thetwo sign waves and is triggered by the timer. The values are written to a Digitalto Analog Converter (DAC). The output of DAC is filtered to reconstruct the sinewaves.6.2.3 Light DetectionA 1/8? flexible fiber optic light guide collects the light from the tissue and deliversit to the optical detector (Edmund Optics, NJ, USA). The detector is an APD mod-ule that includes the APD and the temperature compensated low noise high speedtransimpedance amplifier (C5460-01, Hamamatsu Photonics, Japan). The ampli-fier has an NEP of 0.02 pW/?(Hz) with a gain of 108 V/W which provides goodsensitivity for the low light levels from the tissue. The analog output of the APDmodule is sampled at 200 kS/s by a DAQ module (NI USB-6210) controlled by aPC. The results are read by a MATLAB script on a PC and processed in real time.6.2.4 Digital Lock-in AmplificationA digital lock-in amplification scheme is used to detect the highly attenuated lightsignal from the tissue and also separate the attenuation of the 780 nm componentfrom that of 830 nm [169]. Digital lock in eliminates the need for high cost lock inamplifiers and provide better stability and ability to measure lower frequencies. Adetailed description of the method can be found elsewhere [169]. Here we provide98CHAPTER 6. CUSTOM FNIRS SYSTEM DESIGNa brief description of the method and its implementation details.The modulated signal at the detector can be written asA(n) = Adc +Aaccos(2piffs n+?)(6.1)where Adc is the DC component of the detected signal, Aac is the amplitude of themodulated signal, f , ? , fs are the modulation frequency, signal phase and samplingfrequency respectively. To recover the amplitude Aac, a quadrature demodulationscheme as shown in Figure 6.4 can be used. The resulting in-phase (I) and quadra-ture (Q) components will then beI =A(n)cos(2piffs n)(6.2)=Adccos(2piffs n)+ Aac2cos(4piffs n+?)+ Aac2cos(?)andQ =A(n)sin(2piffs n)(6.3)=Adcsin(2piffs n)+ Aac2sin(4piffs n+?)+ Aac2sin(?)Both operations are performed on an integer number of periods. For 1 cycle,Figure 6.4: In phase-quadrature demodulation scheme.99CHAPTER 6. CUSTOM FNIRS SYSTEM DESIGNNs = fsf . The amplitude and phase can be recovered from the in-phase and quadra-ture components by lowpass filtering to eliminate the higher frequency componentsand using?Aac =?I2LP +Q2LP (6.4)?? = atanQLPILP(6.5)where ILP and QLP are the results of lowpass filtering I and Q, ?Aac and ?? are theestimated amplitude and phase, respectively.Since more than 1 modulating frequency is used in our modulation scheme,then due to finite attenuation of filter at higher frequencies, some crosstalk betweencomponents will occur. To avoid this, one can choose the modulation frequenciesfm such that they fall on the zeros of the filter [169]. For this purpose, assuming asimple moving average filter we will haveh[n] = 1Ns(6.6)where h[n] is the moving average lowpass filter and Ns is the total number of sam-ples collected. This filter has its zeros at k fsNs . So by choosingfm = k fsNs (6.7)the filter response will have zeros at the multiples of modulating frequencies.With fs = 200 kHz, f1 = 1000 Hz and f2 = 1250 Hz, we have Ns = 4000. GivenfNIRS overall sampling rate of 10Hz, Ns samples are read from the DAQ every 100ms and amplitude of the two frequency components are calculated through[I f1I f2]= 1Ns[cos2pi f10fs cos2pi f11fs ? ? ? cos2pi f1(Ns?1)fscos2pi f20fs cos2pi f21fs ? ? ? cos2pi f2(Ns?1)fs]??????A[1]A[2]...A[Ns]??????(6.8)100CHAPTER 6. CUSTOM FNIRS SYSTEM DESIGNand[Q f1Q f2]= 1Ns[sin 2pi f10fs sin2pi f11fs ? ? ? sin2pi f1(Ns?1)fssin 2pi f20fs sin2pi f21fs ? ? ? sin2pi f2(Ns?1)fs]??????A[1]A[2]...A[Ns]??????(6.9)where I f1 , Q f1 , I f2 , Q f2 are the in-phase and quadrature components for f1 andf2, respectively. The amplitude of two wavelength components is derived from?A f1ac = 2?Q2f1 + I2f1 (6.10)?A f2ac = 2?Q2f2 + I2f2The optical density and concentration changes are calculated from ?A f1ac and ?A f2ac.6.2.5 User InterfaceThe graphical user interface was prepared using MATLAB. The code collects Nssamples from the DAQ card every 100ms. The timing is controlled by a timer ob-ject in MATLAB. The samples go through I/Q demodulation according to Equation 6.3to Equation 6.5. The intensity values are then converted to Optical Density (OD)and then Beer-Lambert law converts OD to concentration changes for HbO2 andHHb. For functional NIRS studies, the GUI also delivers the stimulations to thesubject and adds stimulation markers to the data.6.3 Performance EvaluationThe system?s dark noise was measured by readings obtained by placing the deviceand receiver optode in a dark room with no light incident on the receiver optode.The RMS value of the noise in this setup was calculated and repeated for a coupleof measurements to obtain an estimate of the prototypes noise voltage. This valuewas calculated to be less than 50 ?V. The NEP was then calculated from dark noise101CHAPTER 6. CUSTOM FNIRS SYSTEM DESIGNmeasurements usingPi =VnR(? )G (6.11)where Pi is the incident light equivalent power in W, R(? ) is the responsivity ofthe detector at ? = 780 and ? = 830 nm in A/W, G is the amplifier gain and Vnis the noise voltage. With the overall gain (R(? )G) of 1.5? 108 V/W, the NEP isapproximately 0.34 pW.The drift of the measurement was evaluated by continuous recording of datafrom a phantom using the device for 30 minutes after a warm up period of 1 minute.The aqueous phantom was prepared using the method described in Chapter 4. Thephantom scattering and attenuation parameters are chosen to be close to those ofthe adult head tissue (ie. the scalp and skull with ?a = 0.4 cm?1 and ? ?s = 20 cm?1[170], see Section 6.5 for a discussion). The phantom was made with 20% intra-lipid mixed with ink to obtain desired optical parameters. The difference betweenthe initial and final reading normalized to the initial signal value was recorded asthe drift. The device shows 0.1% drift over the period of 30 minutes.6.4 ValidationWe evaluated the performance of the device through in vivo experiments. Theexperiments included arterial occlusion and isometric contraction of the forearmmuscle and the brain response to a motor task test. In all experiments, the collectedintensity data was converted to optical density and concentration changes using theMBLL.6.4.1 MethodsForearm Muscle Arterial Occlusion TestIn this test, the hemodynamic response to an arterial occlusion in the forearm ofa healthy male subject was investigated. This is a common test for validation ofcustom made NIRS instruments [171, 172]. Arterial occlusion was induced in theforearm by means of a pneumatic pressure cuff. The NIRS optodes were placedon the arm and below the pressure cuff and monitored the oxygenation of the arm102CHAPTER 6. CUSTOM FNIRS SYSTEM DESIGNtissue. The cuff was inflated up to 200mm Hg to block any blood in and out flowfrom the muscle for 1.5 minute. The cuff was then released and the NIRS recordingcontinued for another 2 minutes to monitor tissue oxygenation recovery.Isometric ContractionIn this test, the Brachioradialis muscle in the forearm was monitored with the NIRSoptodes during an isometric contraction experiment. An isometric contraction in-volves static contraction of a muscle without a change in muscle length. The sub-ject (same as in previous section) forcefully gripped an object for 30 seconds fol-lowed by one minute of recovery time.Motor Task TestMonitoring brain activation during a motor task is a common method used forevaluating performance of custom made fNIRS instruments [172]. In this test, thefNIRS optodes were placed over the hand area in the left motor cortex near locationC3 according to international 10/20 system of a healthy 30 year old right handedmale subject [173]. The source and detector fibers were secured using a custommade optode holder consisting of a 3D printed holder (Verowhite polyjet resin)tied with elastic band to the head. The source-detector separation was set to 3 cm.The subject was asked to perform a task of opening and closing his fist at a rate ofapproximately 3Hz for 30 seconds followed by 30 seconds of rest. The instructionsfor beginning and ending of the resting/task periods were provided visually througha PC and the timings of the stimulation were recorded along with the fNIRS data.6.4.2 ResultsThe result of the arterial occlusion test is shown in Figure 6.5. Black (solid) andmagenta (dotted) lines indicate time instants when cuff pressure reached maximumvalue and when the cuff was released, respectively. As the total occlusion begins,the amount of blood in the tissue remains constant. This is reflected by constanthemoglobin (tHb) detected during the occlusion. However, oxygen is being con-sumed by the tissue and the HbO2 is constantly converted to HHb. Therefore, HbO2decreases as HHb increases with almost equal changes in the two chromophores?103CHAPTER 6. CUSTOM FNIRS SYSTEM DESIGN250 300 350 400 450?101x 10?5Time (s)Concentration (M)  O2HbHHbtHbFigure 6.5: Typical HbO2 and HHb waveforms in the arterial occlusion test.Black (solid) and magenta (dotted) lines indicate time instants whencuff pressure reached maximum value and when the cuff was released,respectively.concentrations. The gradient of the chromophore changes in this case is propor-tional to tissue?s local oxygen consumption. Once the cuff is released, the changesare reversed. A hyperemic reaction can be observed where due to the auto regula-tion mechanisms, HbO2 and HHb overshoot and undershoot beyond their originalvalue once the cuff is released. The two chromophores gradually return to theiroriginal values during the recovery period.Figure 6.6 shows the hemodynamic changes in response to the isometric con-traction. Black (solid) and magenta (dotted) lines indicate starting point and theending point of the contraction, respectively. The contraction results in an increasein blood flow into the muscle along with an increase in muscle oxygen consump-tion. This can be seen in he figure as an increase in tHb and HHb along witha decrease in HbO2. Once the object is released and the muscle is relaxed, thechanges are reversed (dotted vertical line). The level shifts at the beginning and theend of contraction are the result of the grip motion.The processing of the motor task fNIRS data was performed using HOMER2toolbox for MATLAB [174]. Figure 6.7 shows the changes in total hemoglobinconcentration in response to the motor task over a period of approximately 5 min-utes. The tHb is associated with increased blood flow to the hand area in the brain.The red (solid) lines indicate the beginning of the motor task. In order to measure104CHAPTER 6. CUSTOM FNIRS SYSTEM DESIGN0 20 40 60 80 100 120 140 160 180?50510x 10?5Time (s)Concentration (M)  O2HbHHbtHbFigure 6.6: Isometric contraction results collected from Brachioradialis mus-cle in the forearm. Black (solid) line and magenta (dotted) vertical linesindicate beginning and end of contraction, respectively.100 150 200 250 300 350?2?1012x 10?6 tHbTime (s)Concentration (M)Figure 6.7: Total hemoglobin changes in response to motor activity. ThefNIRS optodes were placed over the hand area in the left motor cortexnear location C3 according to international 10/20 system of a healthy30 year old right handed male subject who was asked to perform a taskof opening and closing his fist.105CHAPTER 6. CUSTOM FNIRS SYSTEM DESIGN?15 ?10 ?5 0 5 10 15 20 25 30 35?1?0.500.511.52x 10?6Time (s)Concentration (M)  O2HbHHbtHbFigure 6.8: Block averaged hemodynamic response.the hemodynamic response, the signal was block averaged over the stimulation/restblocks. Figure 6.8 shows the resulting block averaged traces of HbO2, HHb andtHb with the stimulation starting time at zero. An increase in the HbO2 and tHbalong with a decrease in HHb is observed which matches the typical hemodynamicresponse [171, 172].6.5 Discussion and ConclusionWe described the development of a custom made fNIRS device for future use inTMS experiments to monitor brain tissue hemodynamic changes. The instrumentperformance was evaluated using in vivo tests commonly used in the literaturefor evaluation of custom NIRS instruments. The results are comparable to otherstudies [171, 172] suggesting the instrument is capable of detecting hemodynamicchanges in the tissue and in particular, in the brain and can be used for further brainstudies involving TMS.One of the tests we used to evaluate the performance of the custom device wasthe occlusion test as described in Section 6.4.1. In an occlusion test, it is expectedthat the blood flow to the tissue is fully blocked resulting in decrease in HbO2 and106CHAPTER 6. CUSTOM FNIRS SYSTEM DESIGNincrease in HHb while the tHb stays constant. In our test, however, an increasein tHb is observed as shown in Figure 6.5. This may be due to the fact that thepressure cuff is inflated to 200 mmHg which in this case may have been insufficientto fully block the blood flow. As a result, blood can still reach the tissue and hencethe increase in tHb.The performance of the custom made device was evaluated using a liquid phan-tom as described in Section 6.3. The optical parameters used in this case may betoo high and not a good representation of the overall optical characteristics of anadult head (e.g. compared to [175]). This limitation needs to be addressed in thefuture for a better characterization of the device performance. Also, as stated inChapter 4, a solid phantom is preferable and will be considered in the future.A custom made fNIRS device has multiple advantages for monitoring TMSeffect. The sensitivity and custom sampling rate of such a device allows it to po-tentially measure the fast optical signal [176] which results from small changes inscattering as a result of electrical activity of the neurons. Application of fNIRS sys-tems to monitor fast optical signal during TMS stimulation is new and promising[73]. Additionally, custom devices facilitate temporal and spatial co-registration ofNIRS data with TMS stimulations.107Chapter 7Conclusion and Future WorkSince the first introduction of NIRS by Jobsis, a lot of research has been conductedto apply this promising non-invasive optical method in different clinical applica-tions. Further research and validation in different areas is still required for NIRSto be routinely adopted by clinicians. In this thesis, we attempted to address someof the current issues in NIRS signal processing and applications.7.1 Motion Artifact Removal from fNIRSOne particular issue of interest in NIRS is the sensitivity of the data to motion ar-tifacts. In Chapter 3, we presented a novel method for removal of motion artifactsfrom fNIRS data using the discrete wavelet transform. The method relied on thedifferences between motion-induced patterns and those caused by hemodynamicchanges to identify and remove artifacts in the wavelet domain. The method wasevaluated using simulated data as well as experimental data in terms of the amountof distortion it introduced in the signal and the reduction in artifact intensity andwas shown to be effective in reduction of the motion artifacts. The balance be-tween the amount of artifact reduction and distortion introduction in our method iscontrolled by the user through a tuning parameter.Artifact reduction addresses an important issue in fNIRS signal processing.The hemodynamic response is a weak signal whose detection requires averagingover several blocks of stimulation. Often, contamination with motion artifact re-108CHAPTER 7. CONCLUSION AND FUTURE WORKsults in some of the blocks being excluded from the analysis. In some cases, suchas in infant studies for example, it is not possible to collect many data blocks andmany subjects do not make it to the end of the experiment as they may become rest-less and bored. Therefore, it is important to be able to keep as many data blocksas possible for the analysis. Our proposed method can be used as a preprocessingstep to reduce the intensity of such contaminations in order to allow keeping moredata blocks and improve the contrast to noise ratio in the detected hemodynamicresponse.The proposed method of artifact removal can be enhanced further in severalways. One major improvement would be to extend the range of artifact types thatcan be detected by the method. Our method as described in Chapter 3 targets spikeartifacts. However, artifacts resulting in change in baseline are not addressed withthis method. Such artifacts could potentially be detected in the wavelet domainusing the same principle, but they need to be processed differently.Also, in some cases, only identification of artifacts or contaminated blocks isrequired [24]. The method can be further developed to identify such data segmentsand its performance needs to be evaluated in terms of specificity and sensitivity.Another major potential future direction is to extend the method for real time pro-cessing of the data [177]. This can be of significance in wearable NIRS sensors andin particular, can be directly applied to the bladder sensor described in Chapter 4.This requires using methods for estimating the data variance in real time and as thedata is being captured.The performance of this method relies largely on two major factors. One is thecapability of the Wavelet transform to map the signal to a space where motion ar-tifacts can be better distinguished from the fNIRS signal. The other is the methodused to identify the motion artifact coefficients in the wavelet domain. In our cur-rent approach, we assumed a probability distribution for the wavelet coefficientsand simply gave the coefficients a score based on their probability of belongingto this distribution. Even though this approach was shown to work satisfactorily, itcan be further improved. The problem of identifying the motion induced coefficientcan be considered a classification problem and therefore, well known classificationmethods can be adopted and applied. In particular, one could assume a probabilitydistribution for the motion artifacts coefficients, estimate the parameters of the dis-109CHAPTER 7. CONCLUSION AND FUTURE WORKtribution and use a Bayesian classifier to classify the coefficient as motion inducedor normal. The estimation of the parameters requires use of training data as anadded step. This procedure is expected to improve the performance of the methodin terms of NMSE introduced in the signal as well as artifact attenuation.Another alternative approach which is worth comparing to in the future ischanging the parameter estimation method for the wavelet coefficients. In ourcurrent approach, the main parameter of the probability distribution of waveletcoefficients (i.e. ?? ) was estimated with MAD using the entire data in each channelwhich also included the motion artifacts. This was effective as MAD is not sensi-tive to outliers which in our case, were the motion artifacts. However, if the motionartifacts are frequent and their amplitudes are close to those of fNIRS signal, thenthey may no longer be considered outliers and the estimate of variance will thenbe affected. As a result, the performance of the method will decrease. If a longenough artifact-free segment of the data is available, then it can be used to directlyestimate the probability distribution of the fNIRS data coefficients and its param-eters empirically. This is expected to yield better results as the distribution is nolonger affected by motion artifacts. The downside is that a training process withthe artifact free data would be involved which adds an extra step to the procedure.7.2 Wireless NIRS for Monitoring Bladder ContentsWe reported design and development of a compact NIRS based wireless wearablesensor for continuous non-invasive monitoring of the bladder with potential appli-cation for bladder incontinence patients in Chapter 4. This addresses an importantclinical problem in patients with abnormal bladder function. The device was testedin vitro and in vivo as a proof of concept and was shown to be capable of differ-entiating between empty and full bladder. The results supported the feasibility ofthis device for the purpose of using it as a warning system that alerts the subjectswhen their bladder reaches a pre-determined threshold of bladder capacity. Such adevice, when fully developed as a wearable warning system, can help patients withurinary retention problem and protect them from renal damage by giving earlywarning and alarms for voiding their bladder. Currently, there is no alternativemethod and device for continuous bladder filling detection, a very important clin-110CHAPTER 7. CONCLUSION AND FUTURE WORKical issue especially in patients with different types of urinary incontinence andpatients with spinal cord injury.A major limitation of our work on this sensor presented in Chapter 4 is thelack of clinical data for validation. The purpose of our study at this stage was tointroduce and verify our method through limited multiple trials to provide evidenceof effectiveness of diffuse optics in the form of a wearable sensor for our targetapplication. This method and device need to be further validated in a clinical studywith a larger number of subjects as the next step.Some other technical challenges may limit the applicability of our proposedmethod. As discussed in Chapter 4, the thickness of the fat layer may affect theeffectiveness of our proposed device. The fat layer can cause problems by twomechanisms: 1) by inducing extra attenuation (through both absorption and scat-tering) and 2) by increasing the distance of the bladder from the tissue surface.One way to overcome the problem of increased distance is increasing the pene-tration depth by increasing the distance between source and detector. This wouldrequire an increase in the power level or using a more sensitive detector. Both ofthese solutions would also be effective in overcoming the second mechanism whichis increased attenuation. Increasing the average power may not be safe and mayresult in patients? discomfort. However, we can increase the instantaneous powerand decrease the sampling rate to keep the average power and SNR the same. Evenwith this scheme, the maximum practical limitation of the penetration depth forthis diffusive method is about 3cm.Another solution could be a tight fixation which results in mechanically push-ing the sensor deeper into the tissue. This can however, cause discomfort for thesubject. Changing the wavelength (within the optical wavelengths) does not seemto be effective in improving the penetration depth. Increasing the wavelength de-creases scattering, but attenuation due to absorption by water increases signifi-cantly. Decreasing the wavelength on the other hand, increases attenuation byabsorption by HbO2 and HHb as well as increased scattering.The power consumption of the device is another issue that requires furtherimprovement. With the current prototype, the battery life is 45 hours in standbymode, 23 hours when logging onboard without radio transmission and 15 hourswith radio transmission. Ideally, a wearable device needs to be recharged as few111CHAPTER 7. CONCLUSION AND FUTURE WORKtimes as possible and last at least a day without charging to be practical. Thepower requirements of the prototype at this stage are not low enough. In particular,since the amplifier and filter are the components consuming most of the standbypower, consumption can be significantly reduced by disabling these componentsin standby mode and only activating them when required. Improving the powerconsumption during radio transmission is another major future direction to makethis device suitable for real life patient monitoring.The connectivity of this device to a PC has two main advantages. In our proofof principle study, it was necessary for the investigator to be able to visually seethe signal as the voiding was happening in order to be able to mark different events(permission to void, start of voiding, end of voiding, potential motion artifacts,etc.). In general, it provides an alternative method for device control as well asdata processing and storing. The PC connectivity also serves a further purpose ofbeing used in future developments for clinical remote monitoring. The frequencyof bladder filling and voidings which are transferred to the PC can be uploaded toan online database to be accessed by a clinician. These information are meaningfulin some pathological conditions. For example in UTI which is a common condi-tion in spinal cord injury patients, the frequency of voiding increases and access tothese information collected in normal daily life conditions by the clinician is im-portant in treatment of patients. In such situations, the limited range of connectionmight limit the usage of wireless link to indoors only. However, the same benefitscan be offered by replacing the PC with a smart phone which can provide similarfunctionalities while being portable and mobile. Development of appropriate ap-plications and database for patient monitoring is a major future improvement forthis system.The actual process of making decisions on bladder fullness and when to triggerthe alarm is an important step and the effectiveness of the proposed overall sys-tem depends on it. However, this was beyond the scope of our work presented inChapter 4. The optimal decision making process which requires evaluation of therecordings on a case by case basis by a clinician who factors in parameters such asbladder size, anatomy, level of injury, etc. is part of the future work on this method.The drift in the output of the device can potentially cause problems in longterm patient monitoring as the accumulated drift may be mistaken for bladder fill-112CHAPTER 7. CONCLUSION AND FUTURE WORKing. The most important cause of the drift as discussed in Chapter 4 is the slighttemperature rise caused by LED operation. This rise leads to slight changes (drift)in the LED output intensity over time. This has been reported to be present inother works on NIRS systems [171, 172]. This limitation needs to be addressedfor practical application of our method. The solution suggested in Chapter 4 wasto reset the device at voiding times (alarm) to prevent the error from accumulating.A better solution could be to use LEDs with photodetector integrated to control theoutput power in a closed loop system. Such LEDs could be selected to also include2 more wavelengths for detection of HbO2 and HHb in order to estimate changesfrom these chromophores as well. This could help minimize the contribution ofthese chromophores to the detected signal and minimize the systemic interferencesas discussed in Chapter 4.Another major limitation of the method presented in Chapter 4 is the large vari-ations in the initial value of optical attenuation which is caused by the sensitivityof the device to unknown changes in attenuation caused by parameters such assensor geometry, coupling and tissue scattering. These parameters vary from oneexperiment to another or when the sensor is relocated on the tissue. Therefore, inorder to develop a fully practical and reliable system for continuous monitoringof the bladder, it is necessary to address this issue so that a threshold of bladdercapacity can be properly established and the measured attenuation can be com-pared to this threshold. Therefore, a more robust approach such as a multidistancemethod which is less sensitive to changes in coupling needs to be considered forthis method in the future [138, 139].7.3 fNIRS ConnectivityIn Chapter 5 we used fNIRS along with phase analysis to identify resting statefunctional connectivity patterns in language system. Use of phase relation betweenhemodynamic changes in fNIRS data in resting state is a novel contribution of thiswork. The agreement between results presented with those obtained from fNIRSdata in similar studies with different methods and different subject population sug-gests feasibility of the proposed approach.As stated earlier, the data used in Chapter 5 was not collected in ?strict? resting113CHAPTER 7. CONCLUSION AND FUTURE WORKstate, meaning that the subjects where not in the ideal state of rest with no particularcognitive or sensorimotor task. It is not uncommon in the literature to use lightcognitive task as a substitute for resting state ([149, 167]). Our data which wascollected during stimulation with Silbo-Gomero language to which infants showedno response, can be considered a very low demand task and therefore, the resultsare still valid. However, investigating the alternative approach of absolute restingstate data and comparing the results to those obtained in Chapter 5 would be ofinterest. This requires collection of fNIRS data from infants without any particularstimulation. The results could confirm the degree of validity of the assumption ofresting state for our low cognitive stimulation condition.An interesting question that can possibly be answered using this approach ishow brain networks develop with age. This is a relevant question in developmentalneuroscience [178] and fNIRS offers advantages that make it ideal for studies thatinvolve collecting data from young infants over other methods such as fMRI. Wehave already applied our phase-based fNIRS connectivity analysis for identifyinglanguage network in newborn infants. A future study on older subjects to identifythe same network and changes in the network during brain development couldreveal very valuable information. If fNIRS connectivity is validated as a clinicalmethod for evaluating brain network connectivity patterns, it can potentially beused for example to identify brain network developmental problems in infants longbefore the symptoms can be observed.One potential challenge in the use of the proposed method as described inChapter 5 for monitoring cortical network development is the choice of the seedchannel. The seed channel in our method was chosen as the channel inside the lan-guage network which showed high activation in language tasks. To ensure changesin the shape and location of the network with age are taken into account, one needsto precede the RSFC analysis with a localizer task to identify a proper seed channel.Another possible future direction is using NIRS based functional connectivityin a clinical study to validate its capability in discriminating between health anddisease in conditions such as stroke, depression or autism.Using the joint probability distribution function of the phase results in estima-tion of phase dependence between the two channels along with the preferred phasedifference. In the method presented in Chapter 5, only the first parameter was used.114CHAPTER 7. CONCLUSION AND FUTURE WORKHowever, the second parameter could contain significant information as well. Thephase difference between HbO2 and HHb has been shown to convey informationthat discriminates between different states as well as health and disease in partic-ular conditions [179, 180]. Applying similar approach to connectivity with ourproposed method in the future can provide further connectivity information.As discussed in Chapter 5, HbO2 is more sensitive to regional blood flowchanges and was therefore used for the connectivity analysis. However, compar-ison of the results with those obtained by HHb could be insightful as to whetherthey detect similar networks and whether this difference is only attributed to highernoise level in HHb.We discussed the issue of non existing spurious connections as a result of in-herent shortcoming of bivariate methods when there is a propagation of informa-tion from one channel to others. Even though this does not affect the detection ofnetworks as explained in Chapter 5, applying the same analysis using an equiva-lent multivariate method could reveal further information. A similar approach hasbeen applied to EEG and could be adopted to fNIRS similar to what was used inChapter 5.7.4 Custom-made fNIRS Device for TMS MonitoringIn Chapter 6, we described the development and validation of a custom madefNIRS system for monitoring the effect of the TMS on brain activation and con-nectivity. Even though commercial NIRS devices are available for different typesof studies, developing custom made devices for specific purposes is common [171,172, 181]. The functionality of the device was verified in Chapter 6 using testswhich are well documented in the literature. In order to further validate the deviceand test the hypothesis that brain activation in the motor and visual cortices canbe detected by this device, we will be using it in the practice of a novel unilateraljoystick based tracking task [182]. This task has been used by our collaborators fordifferent studies in the past research during fMRI. This past work assures us thatsignificant change in both brain function and behavior occur during the practice ofthis task [183]. In addition, the availability of fMRI data during task performancemakes comparison of the data from the device with those from fMRI easier.115CHAPTER 7. CONCLUSION AND FUTURE WORKOne major future improvement is integrating the TMS 3D stimulation localiz-ing system into fNIRS optodes so that fNIRS monitoring location and TMS stimu-lation location can both be registered on an anatomical MRI image. Currently, 3Dobject trackers are used for co-registering TMS stimulation target with an anatom-ical MRI image of the subject?s brain. A similar approach can be used to trackfNIRS probe on the head to ensure correct spatial relation between the two.rTMS has been shown to be effective in improving performance in motor skillpractice. Monitoring the accumulating effect of TMS on brain hemodynamics canbe beneficial in understanding the mechanism of effect of rTMS involved in thisprocess. Moreover, comparison of neural activation while practicing a motor skilltask before and after rTMS stimulation session can help evaluate the effect of rTMSon the brain in a qualitative manner. In particular, the fNIRS data collected inthis way can also be used in combination with the connectivity analysis methoddescribed in Chapter 5 to evaluate brain connectivity changes following the TMSstimulation.Investigating the effect of rTMS on the brain using the fast optical signal is arecent subject of interest [73]. This type of optical signal has the advantage of amuch higher temporal resolution compared to that of hemodynamic response and istherefore capable of detecting rTMS-induced changes in a much shorter time scale.Even though the capabilities of the device described in Chapter 6 for detecting fastoptical signal was not demonstrated in this thesis, this device could potentially beused for this purpose using minimum changes. Validation and application of thisdevice in an rTMS study for detecting fast optical signal is in the scope of ourfuture studies.116Bibliography[1] B. Molavi, G. A. Dumont, and B. Shadgan. Motion artifact removal frommuscle NIR Spectroscopy measurements. In CCECE 2010, pages 1?4,May 2010. ? pages iv, v[2] B. Molavi and G. A. Dumont. Wavelet based motion artifact removal forFunctional Near Infrared Spectroscopy. Annual International Conferenceof the IEEE Engineering in Medicine and Biology Society, pages 5?8,2010. ? pages iv, v, 32[3] B. Molavi, G. A. Dumont, B. Shadgan, and A. J. Macnab. Attenuation ofmotion artifact in near infrared spectroscopy signals using a wavelet basedmethod. In Proceedings of SPIE 7890, page 78900M, February 2011. ?pages iv, v, 32[4] B. Molavi and G. A. Dumont. Wavelet-based motion artifact removal forfunctional near-infrared spectroscopy. Physiological measurement,33(2):259?70, February 2012. ? pages iv, v, 32, 51[5] B. Molavi, J. Gervain, and G. A. Dumont. Estimating cortical connectivityin functional near infrared spectroscopy using multivariate autoregressivemodeling. In Annual International Conference of the IEEE Engineering inMedicine and Biology Society, pages 2334?2337, January 2011. ? pagesiv, v, 72[6] B. Molavi, J. Gervain, G. A. Dumont, and H. A. Noubari. Functionalconnectivity analysis of cortical networks in Functional Near InfraredSpectroscopy using phase synchronization. In Annual InternationalConference of the IEEE Engineering in Medicine and Biology Society,pages 5182?5185, August 2012. ? pages iv, v[7] F. F. Jo?bsis. Noninvasive, infrared monitoring of cerebral and myocardialoxygen sufficiency and circulatory parameters. Science,198(4323):1264?1267, December 1977. ? pages 1117BIBLIOGRAPHY[8] B. Shadgan, J. A. Guenette, A. W. Sheel, and W. D. Reid.Sternocleidomastoid muscle deoxygenation in response to incrementalinspiratory threshold loading measured by near infrared spectroscopy.Respiratory physiology & neurobiology, 178(2):202?209, September 2011.? pages 2[9] M. Ferrari, T. Binzoni, and V. Quaresima. Oxidative metabolism in muscle.Philosophical transactions of the Royal Society of London.Series B,Biological sciences, 352(1354):677?683, 1997. ? pages 2[10] P. M. Arenth, J. H. Ricker, and M. T. Schultheis. Applications of functionalnear-infrared spectroscopy (fnirs) to neurorehabilitation of cognitivedisabilities. The Clinical neuropsychologist, 21(1):38?57, Jan 2007. ?pages 2[11] S. Coyle, T. Ward, C. Markham, and G. McDarby. On the suitability ofnear-infrared (nir) systems for next-generation brain-computer interfaces.Physiological Measurement, 25(4):815?822, Aug 2004. ? pages 2[12] R. Choe, A. Corlu, K. Lee, T. Durduran, S. D. Konecky,M. Grosicka-Koptyra, S. R. Arridge, B. J. Czerniecki, D. L. Fraker,A. DeMichele, B. Chance, M. A. Rosen, and A. G. Yodh. Diffuse opticaltomography of breast cancer during neoadjuvant chemotherapy: a casestudy with comparison to MRI. Medical physics, 32:1128?1139, April2005. ? pages 2[13] F. Orihuela-Espina, D. R. Leff, D. R. C. James, A. W. Darzi, and G. Z.Yang. Quality control and assurance in functional near infraredspectroscopy (fNIRS) experimentation. Physics in medicine and biology,55(13):3701?3724, July 2010. ? pages 3[14] S. Huettel, A. W. Song, and G. McCarthy. Functional magnetic resonanceimaging. Sinauer Associates Publishers, Sunderland Mass., 2004. ? pages3[15] G. Strangman, J. P. Culver, J. H. Thompson, and D. A. Boas. AQuantitative Comparison of Simultaneous BOLD fMRI and NIRSRecordings during Functional Brain Activation. NeuroImage,17(2):719?731, October 2002. ? pages 3[16] D. T. Delpy, M. Cope, P. Van Der Zee, S. Arridge, S. Wray, and J. Wyatt.Estimation of optical pathlength through tissue from direct time of flight118BIBLIOGRAPHYmeasurement. Physics in medicine and biology, 33(12):1433?1442,December 1988. ? pages 4, 53[17] T. Durduran, R. Choe, W. B. Baker, and A. G. Yodh. Diffuse optics fortissue monitoring and tomography. Reports on Progress in Physics,73(7):076701, July 2010. ? pages 6, 7, 11, 12[18] C. HOCK, K. VILLRINGER, F. M ?ULLER-SPAHN, M. HOFMANN,S. SCHUH-HOFER, H. HEEKEREN, R. WENZEL, U. DIRNAGL, andA. VILLRINGER. Near Infrared Spectroscopy in the Diagnosis ofAlzheimer?s Diseasea. Annals of the New York Academy of Sciences,777(1):22?29, January 1996. ? pages 8[19] C. Hock, K. Villringer, F. Mu?ller-Spahn, R. Wenzel, H. Heekeren,S. Schuh-Hofer, M. Hofmann, S. Minoshima, M. Schwaiger, U. Dirnagl,and A. Villringer. Decrease in parietal cerebral hemoglobin oxygenationduring performance of a verbal fluency task in patients with Alzheimer?sdisease monitored by means of near-infrared spectroscopy (NIRS)correlation with simultaneous rCBF-PET measurements. Brain Research,755(2):293?303, May 1997. ? pages 8[20] F. Okada, Y. Tokumitsu, Y. Hoshi, and M. Tamura. Impairedinterhemispheric integration in brain oxygenation and hemodynamics inschizophrenia. European archives of psychiatry and clinical neuroscience,244(1):17?25, 1994. ? pages 8[21] A. J. Fallgatter and W. K. Strik. Reduced frontal functional asymmetry inschizophrenia during a cued continuous performance test assessed withnear-infrared spectroscopy. Schizophrenia bulletin, 26(4):913?919, 2000.? pages 8[22] K. Matsuo, T. Kato, M. Fukuda, and N. Kato. Alteration of hemoglobinoxygenation in the frontal region in elderly depressed patients as measuredby near-infrared spectroscopy. The Journal of neuropsychiatry and clinicalneurosciences, 12(4):465?471, 2000. ? pages 8[23] A. P. Gibson, J. C. Hebden, and S. R. Arridge. Recent advances in diffuseoptical imaging. Physics in Medicine and Biology, 50(4):R1?R43, 2005.? pages 8, 9[24] J. Gervain, F. Macagno, S. Cogoi, M. Pen?a, and J. Mehler. The neonatebrain detects speech structure. Proceedings of the National Academy of119BIBLIOGRAPHYSciences of the United States of America, 105(37):14222?14227,September 2008. ? pages 8, 16, 75, 76, 77, 86, 109[25] J. H. Meek, M. Firbank, C. E. Elwell, J. Atkinson, O. Braddick, and J. S.Wyatt. Regional hemodynamic responses to visual stimulation in awakeinfants. Pediatric research, 43(6):840?843, 1998. ? pages 8[26] P. Zaramella, F. Freato, A. Amigoni, S. Salvadori, P. Marangoni,A. Suppjei, B. Schiavo, and L. Chiandetti. Brain auditory activationmeasured by near-infrared spectroscopy (nirs) in neonates. Pediatricresearch, 49(2):213?219, Feb 2001. ? pages 8[27] L. Stothers, B. Shadgan, and A. Macnab. Urological applications of nearinfrared spectroscopy. The Canadian journal of urology, 15(6):4399?4409,December 2008. ? pages 9, 23, 69[28] M. C. van der Sluijs. New and highly sensitive continuous-wavenear-infrared spectrophotometer with multiple detectors. In Proceedings ofSPIE, volume 3194, pages 63?72. SPIE, 1998. ? pages 9[29] M. Cope and D. T. Delpy. System for long-term measurement of cerebralblood and tissue oxygenation on newborn infants by near infraredtransillumination. Medical & Biological Engineering & Computing,26(3):289?294, 1988. ? pages 76[30] M. A. Franceschini, V. Toronov, M. Filiaci, E. Gratton, and S. Fantini.On-line optical imaging of the human brain with 160-ms temporalresolution. Optics Express, 6(3):49?57, 2000. ? pages[31] C. D. Gomersall, P. L. Leung, T. Gin, G. M. Joynt, R. J. Young, W. S. Poon,and T. E. Oh. A comparison of the hamamatsu niro 500 and the invos 3100near-infrared spectrophotometers. Anaesthesia and Intensive Care,26(5):548?557, Oct 1998. ? pages[32] F. Kawaguchi, N. Ichikawa, N. Fujiwara, Y. Yamashita, and S. Kawasaki.Clinically available optical topography system. Hitachi Reveiw, 50(1),2001. ? pages[33] V. Quaresima, S. Sacco, R. Totaro, and M. Ferrari. Noninvasivemeasurement of cerebral hemoglobin oxygen saturation using two nearinfrared spectroscopy approaches. Journal of biomedical optics,5(2):201?205, April 2000. ? pages 9120BIBLIOGRAPHY[34] D. Grosenick, H. Wabnitz, H. H. Rinneberg, K. T. Moesta, and P. M.Schlag. Development of a Time-Domain Optical Mammograph and First invivo Applications. Applied Optics, 38(13):2927, May 1999. ? pages 9[35] N. L. Everdell, a. P. Gibson, I. D. C. Tullis, T. Vaithianathan, J. C. Hebden,and D. T. Delpy. A frequency multiplexed near-infrared topography systemfor imaging functional activation in the brain. Review of ScientificInstruments, 76(9):093705, 2005. ? pages[36] V. Ntziachristos, X. Ma, A. G. Yodh, and B. Chance. Multichannel photoncounting instrument for spatially resolved near infrared spectroscopy.Review of Scientific Instruments, 70(1):193?201, 1999. ? pages[37] D. Haensse, P. Szabo, D. Brown, J.-C. Fauche`re, P. Niederer, H.-u. Bucher,and M. Wolf. A new multichannel near infrared spectrophotometry systemfor functional studies of the brain in adults and neonates. Optics Express,13(12):4525, 2005. ? pages[38] A. Siegel, J. J. Marota, and D. Boas. Design and evaluation of acontinuous-wave diffuse optical tomography system. Optics Express,4(8):287?298, 1999. ? pages 9[39] M. Wolf, M. Ferrari, and V. Quaresima. Progress of near-infraredspectroscopy and topography for brain and muscle clinical applications.Journal of biomedical optics, 12(6):062104, 2007. ? pages 9, 12[40] M. Ferrari, L. Mottola, and V. Quaresima. Principles, techniques, andlimitations of near infrared spectroscopy. Canadian journal of appliedphysiology, 29(4):463?487, August 2004. ? pages 9, 15[41] R. D. Frostig. In vivo optical imaging of brain function. CRC Press, BocaRaton, 2002. ? pages 9[42] M. Fabiani, D. D. Schmorrow, and G. Gratton. Optical imaging of theintact human brain [guest editorial]. Engineering in Medicine and BiologyMagazine, IEEE, 26(4):14?16, 2007. ? pages 10[43] G. Gratton, C. R. Brumback, B. A. Gordon, M. A. Pearson, K. A. Low, andM. Fabiani. Effects of measurement method, wavelength, andsource-detector distance on the fast optical signal. NeuroImage,32(4):1576?1590, 2006. ? pages 10[44] E. L. Maclin, G. Gratton, and M. Fabiani. Optimum filtering for erosmeasurements. Psychophysiology, 40(4):542?547, Jul 2003. ? pages121BIBLIOGRAPHY[45] E. L. Maclin, K. A. Low, J. J. Sable, M. Fabiani, and G. Gratton. Theevent-related optical signal to electrical stimulation of the median nerve.NeuroImage, 21(4):1798?1804, 2004. ? pages 10[46] S. J. Matcher, P. J. Kirkpatrick, K. Nahid, M. Cope, and D. T. Delpy.Absolute quantification methods in tissue near infrared spectroscopy. InProc. SPIE 2389, pages 486?495, May 1995. ? pages 10[47] H. Sato, M. Kiguchi, F. Kawaguchi, and A. Maki. Practicality ofwavelength selection to improve signal-to-noise ratio in near-infraredspectroscopy. NeuroImage, 21(4):1554?1562, 2004. ? pages 11[48] Y. Yamashita, A. Maki, and H. Koizumi. Wavelength dependence of theprecision of noninvasive optical measurement of oxy-, deoxy-, andtotal-hemoglobin concentration. Medical Physics, 28(6):1108?1114, 2001.? pages 11[49] G. Strangman. Factors affecting the accuracy of near-infrared spectroscopyconcentration calculations for focal changes in oxygenation parameters.NeuroImage, 18(4):865?879, 2003. ? pages[50] D. A. Boas, A. M. Dale, and M. A. Franceschini. Diffuse optical imagingof brain activation: approaches to optimizing image sensitivity, resolution,and accuracy. NeuroImage, 23:S275?S288, January 2004. ? pages 11[51] A. Corlu, R. Choe, T. Durduran, K. Lee, M. Schweiger, S. R. Arridge,E. M. C. Hillman, and A. G. Yodh. Diffuse optical tomography withspectral constraints and wavelength optimization. Applied optics,44(11):2082?2093, April 2005. ? pages 11[52] A. Corlu, T. Durduran, R. Choe, M. Schweiger, E. M. C. Hillman, S. R.Arridge, and A. G. Yodh. Uniqueness and wavelength optimization incontinuous-wave multispectral diffuse optical tomography. Optics letters,28(23):2339?2341, December 2003. ? pages 11[53] Safety of laser products - Part 1: Equipment classification andrequirements. Technical report, International Electrotechnical Commission,2007. ? pages 11[54] Photobiological safety of lamps and lamp systems. Technical report,International Electrotechnical Commision, 2006. ? pages 11122BIBLIOGRAPHY[55] A. Bozkurt and B. Onaral. Safety assessment of near infrared light emittingdiodes for diffuse optical measurements. Biomedical engineering online,3(1):9, March 2004. ? pages 11[56] H. Koizumi, T. Yamamoto, A. Maki, Y. Yamashita, H. Sato, H. Kawaguchi,and N. Ichikawa. Optical topography: practical problems and newapplications. Applied optics, 42(16):3054?3062, June 2003. ? pages 11[57] J. Gervain, J. Mehler, J. F. Werker, C. a. Nelson, G. Csibra, S. Lloyd-Fox,M. Shukla, and R. N. Aslin. Near-infrared spectroscopy: a report from theMcDonnell infant methodology consortium. Developmental cognitiveneuroscience, 1(1):22?46, January 2011. ? pages 11, 13, 58[58] G. Strangman, D. A. Boas, and J. P. Sutton. Non-invasive neuroimagingusing near-infrared light. Biological psychiatry, 52(7):679?693, October2002. ? pages 12, 13[59] A. P. Gibson, J. C. Hebden, and S. R. Arridge. Recent advances in diffuseoptical imaging. Physics in Medicine and Biology, 50(4):R1?R43, 2005.? pages 12[60] M. S. Patterson, S. Andersson-Engels, B. C. Wilson, and E. K. Osei.Absorption spectroscopy in tissue-simulating materials: a theoretical andexperimental study of photon paths. Applied optics, 34(1):22?30, January1995. ? pages 12[61] S. Lloyd-Fox, A. Blasi, and C. E. Elwell. Illuminating the developingbrain: the past, present and future of functional near infrared spectroscopy.Neuroscience and biobehavioral reviews, 34(3):269?284, March 2010. ?pages 12, 13[62] V. Toronov, M. A. Franceschini, M. Filiaci, S. Fantini, M. Wolf,A. Michalos, and E. Gratton. Near-infrared study of fluctuations in cerebralhemodynamics during rest and motor stimulation: temporal analysis andspatial mapping. Medical physics, 27(4):801?815, April 2000. ? pages 14[63] F. Robertson, T. Douglas, and E. Meintjes. Motion Artefact Removal forFunctional Near Infrared Spectroscopy: a Comparison of Methods. IEEETransactions on Biomedical Engineering, 57(6):1377?1387, 2010. ?pages 15, 21, 23, 48, 49[64] Y. Zhang, D. H. Brooks, M. A. Franceschini, and D. A. Boas.Eigenvector-based spatial filtering for reduction of physiological123BIBLIOGRAPHYinterference in diffuse optical imaging. Journal of biomedical optics,10(1):11014, 2005. ? pages 22[65] M. Izzetoglu, P. Chitrapu, S. Bunce, and B. Onaral. Motion artifactcancellation in NIR spectroscopy using discrete Kalman filtering.Biomedical engineering online, 9(1):16, January 2010. ? pages 22, 48, 51[66] F. Scholkmann, S. Spichtig, T. Muehlemann, and M. Wolf. How to detectand reduce movement artifacts in near-infrared imaging using movingstandard deviation and spline interpolation. Physiological measurement,31(5):649?662, May 2010. ? pages 22, 51[67] X. Cui, S. Bray, and A. L. Reiss. Functional near infrared spectroscopy(NIRS) signal improvement based on negative correlation betweenoxygenated and deoxygenated hemoglobin dynamics. NeuroImage,49(4):3039?3046, February 2010. ? pages 15, 22, 51[68] P. Petrican and M. A. Sawan. Design of a miniaturized ultrasonic bladdervolume monitor and subsequent preliminary evaluation on 41 enureticpatients. IEEE transactions on rehabilitation engineering, 6(1):66?74,March 1998. ? pages 16, 25, 54[69] J. Gervain and J. F. Werker. Prosody cues word order in 7-month-oldbilingual infants. Nature communications, 4:1490, January 2013. ? pages16[70] L. May, K. Byers-Heinlein, J. Gervain, and J. F. Werker. Language and thenewborn brain: does prenatal language experience shape the neonate neuralresponse to speech? Frontiers in psychology, 2:222, January 2011. ?pages 16[71] Pitch perfect. The Economist, 2013. ? pages 16[72] G. Gratton and M. Fabiani. The event-related optical signal (EROS) invisual cortex: replicability, consistency, localization, and resolution.Psychophysiology, 40(4):561?71, July 2003. ? pages 17[73] N. A. Parks, E. L. Maclin, K. A. Low, D. M. Beck, M. Fabiani, andG. Gratton. Examining cortical dynamics and connectivity withsimultaneous single-pulse transcranial magnetic stimulation and fastoptical imaging. NeuroImage, 59(3):2504?2510, February 2012. ? pages17, 30, 31, 107, 116124BIBLIOGRAPHY[74] R. J. Cooper, J. Selb, L. Gagnon, D. Phillip, H. W. Schytz, H. K. Iversen,M. Ashina, and D. A. Boas. A systematic comparison of motion artifactcorrection techniques for functional near-infrared spectroscopy. Frontiersin neuroscience, 6(147), January 2012. ? pages 18, 22, 23, 32, 51[75] S. Brigadoi, L. Ceccherini, S. Cutini, F. Scarpa, P. Scatturin, J. Selb,L. Gagnon, D. A. Boas, and R. J. Cooper. Motion artifacts in functionalnear-infrared spectroscopy: A comparison of motion correction techniquesapplied to real cognitive data. NeuroImage, April 2013. ? pages 18, 23,32, 49, 50, 51[76] M. Izzetoglu, A. Devaraj, S. Bunce, and B. Onaral. Motion artifactcancellation in NIR spectroscopy using wiener filtering. IEEE Transactionson Biomedical Engineering, 52(5):934?938, 2005. ? pages 21, 22, 47, 48[77] J. Virtanen, T. Noponen, K. Kotilahti, J. Virtanen, and R. J. Ilmoniemi.Accelerometer-based method for correcting signal baseline changes causedby motion artifacts in medical near-infrared spectroscopy. Journal ofbiomedical optics, 16(8):087005, August 2011. ? pages 21[78] A. Blasi, D. Phillips, S. Lloyd-Fox, P. H. Koh, and C. E. Elwell. Automaticdetection of motion artifacts in infant functional optical topography studies.Advances in experimental medicine and biology, 662:279?284, January2010. ? pages 21[79] Q. Zhang, G. E. Strangman, and G. Ganis. Adaptive filtering to reduceglobal interference in non-invasive NIRS measures of brain activation: howwell and when does it work? NeuroImage, 45(3):788?794, April 2009. ?pages 21[80] Y. Zhang, J. Sun, and P. Rolfe. Reduction of global interference infunctional multidistance near-infrared spectroscopy using empirical modedecomposition and recursive least squares: a Monte Carlo study. Journal ofthe European Optical Society: Rapid Publications, 6:11033, June 2011. ?pages 21[81] T. Nakano, H. Watanabe, F. Homae, and G. Taga. Prefrontal corticalinvolvement in young infants? analysis of novelty. Cerebral cortex,19(2):455?463, February 2009. ? pages 21[82] M. Pen?a, A. Maki, D. Kovacic?, G. Dehaene-Lambertz, H. Koizumi,F. Bouquet, and J. Mehler. Sounds and silence: an optical topography studyof language recognition at birth. Proceedings of the National Academy of125BIBLIOGRAPHYSciences of the United States of America, 100(20):11702?11705,September 2003. ? pages 22, 40[83] J. Gervain, F. Macagno, S. Cogoi, M. Pen?a, and J. Mehler. The neonatebrain detects speech structure. Proceedings of the National Academy ofSciences of the United States of America, 105(37):14222?14227,September 2008. ? pages 22, 40[84] B. Lee, J. Han, H. J. Baek, J. H. Shin, K. S. Park, and W. J. Yi. Improvedelimination of motion artifacts from a photoplethysmographic signal usinga Kalman smoother with simultaneous accelerometry. Physiologicalmeasurement, 31(12):1585?1603, October 2010. ? pages 22, 48[85] H. Sato, N. Tanaka, M. Uchida, Y. Hirabayashi, M. Kanai, T. Ashida,I. Konishi, and A. Maki. Wavelet analysis for detecting body-movementartifacts in optical topography signals. NeuroImage, 33(2):580?587,November 2006. ? pages 23, 47, 49[86] K. E. Jang, S. Tak, J. Jung, J. Jang, Y. Jeong, and J. C. Ye. Waveletminimum description length detrending for near-infrared spectroscopy.Journal of biomedical optics, 14(3):034004, 2009. ? pages 23[87] A. Macnab and B. Shadgan. Biomedical applications of wirelesscontinuous wave near infrared spectroscopy. Biomedical Spectroscopy andImaging, 1(3):205?222, 2012. ? pages 23, 54[88] A. Macnab, B. Shadgan, K. Afshar, and L. Stothers. Near-InfraredSpectroscopy of the Bladder: New Parameters for Evaluating VoidingDysfunction. International Journal of Spectroscopy, 2011:1?8, 2011. ?pages 23[89] A. Macnab, B. Friedman, B. Shadgan, and L. Stothers. Bladder anatomyphysiology and pathophysiology: Elements that suit near infraredspectroscopic evaluation of voiding dysfunction. Biomedical Spectroscopyand Imaging, 1(3):223?235, 2012. ? pages 23[90] M. Yurt, E. Su?er, O. Gu?lpinar, O. Telli, and N. Arikan. Diagnosis of bladderoutlet obstruction in men with lower urinary tract symptoms: comparisonof near infrared spectroscopy algorithm and pressure flow study in aprospective study. Urology, 80(1):182?186, July 2012. ? pages 24[91] A. Amelink, D. J. Kok, H. J. C. M. Sterenborg, and J. R. Scheepe. In vivomeasurement of bladder wall oxygen saturation using optical spectroscopy.Journal of biophotonics, 4(10):715?720, October 2011. ? pages 24126BIBLIOGRAPHY[92] G. Vijaya, G. A. Digesu, A. Derpapas, D. C. Panayi, R. Fernando, andV. Khullar. Changes in detrusor muscle oxygenation during detrusoroveractivity contractions. European journal of obstetrics, gynecology, andreproductive biology, 163(1):104?107, July 2012. ? pages 24[93] C. H. van der Vaart, J. R. J. de Leeuw, J. P. W. R. Roovers, and A. P. M.Heintz. The effect of urinary incontinence and overactive bladdersymptoms on quality of life in young women. BJU international,90(6):544?549, October 2002. ? pages 24[94] N. K. Kristiansen, J. C. Djurhuus, and H. Nygaard. Design and evaluationof an ultrasound-based bladder volume monitor. Medical & biologicalengineering & computing, 42(6):762?769, November 2004. ? pages 25[95] S. Leonhardt, A. Cordes, H. Plewa, R. Pikkemaat, I. Soljanik, K. Moehring,H. J. Gerner, and R. Rupp. Electric impedance tomography for monitoringvolume and size of the urinary bladder. BiomedizinischeTechnik/Biomedical engineering, 56(6):301?307, December 2011. ?pages 25[96] K. J. Friston, C. D. Frith, P. F. Liddle, and R. S. Frackowiak. Functionalconnectivity: the principal-component analysis of large (PET) data sets.Journal of cerebral blood flow and metabolism, 13(1):5?14, January 1993.? pages 25[97] B. R. White, A. Z. Snyder, A. L. Cohen, S. E. Petersen, M. E. Raichle,B. L. Schlaggar, and J. P. Culver. Resting-state functional connectivity inthe human brain revealed with diffuse optical tomography. NeuroImage,47(1):148?156, August 2009. ? pages 25, 26, 27, 81[98] C. Gerloff, J. Richard, J. Hadley, A. E. Schulman, M. Honda, andM. Hallett. Functional coupling and regional activation of human corticalmotor areas during simple, internally paced and externally paced fingermovements. Brain, 121(8):1513?1531, August 1998. ? pages 26[99] L. Lee, L. M. Harrison, and A. Mechelli. A report of the functionalconnectivity workshop, Dusseldorf 2002. NeuroImage, 19(2):457?465,June 2003. ? pages 26[100] A. V. Medvedev, J. M. Kainerstorfer, S. V. Borisov, and J. VanMeter.Functional connectivity in the prefrontal cortex measured by near-infraredspectroscopy during ultrarapid object recognition. Journal of biomedicaloptics, 16(1):016008, 2011. ? pages 26127BIBLIOGRAPHY[101] B. Biswal, F. Z. Yetkin, V. M. Haughton, and J. S. Hyde. Functionalconnectivity in the motor cortex of resting human brain using echo-planarMRI. Magnetic resonance in medicine, 34(4):537?541, October 1995. ?pages 26, 81, 86[102] F. T. Sun, L. M. Miller, and M. D?Esposito. Measuring interregionalfunctional connectivity using coherence and partial coherence analyses offMRI data. NeuroImage, 21(2):647?658, February 2004. ? pages 26[103] S. Schnider, R. Kwong, F. Lenz, and H. Kwan. Detection of feedback inthe central nervous system using system identification techniques.Biological Cybernetics, 60(3):203?212, January 1989. ? pages 26[104] L. A. Baccala? and K. Sameshima. Partial directed coherence: a newconcept in neural structure determination. Biological cybernetics,84(6):463?474, June 2001. ? pages 26[105] R. Salvador, A. Mart??nez, E. Pomarol-Clotet, J. Gomar, F. Vila, S. Sarro?,A. Capdevila, and E. Bullmore. A simple view of the brain through afrequency-specific functional connectivity measure. NeuroImage,39(1):279?289, January 2008. ? pages 26[106] Y. J. Zhang, C. M. Lu, B. B. Biswal, Y. F. Zang, D. L. Peng, and C. Z. Zhu.Detecting resting-state functional connectivity in the language systemusing functional near-infrared spectroscopy. Journal of biomedical optics,15(4):047003, 2010. ? pages 27, 30, 87, 88, 93[107] C. M. Lu, Y. J. Zhang, B. B. Biswal, Y. F. Zang, D. L. Peng, and C. Z. Zhu.Use of fNIRS to assess resting state functional connectivity. Journal ofneuroscience methods, 186(2):242?249, February 2010. ? pages 27, 30,81, 86[108] L. Duan, Y. J. Zhang, and C. Z. Zhu. Quantitative comparison ofresting-state functional connectivity derived from fNIRS and fMRI: asimultaneous recording study. NeuroImage, 60(4):2008?2018, May 2012.? pages 27, 30, 81, 86[109] Y. Hada, M. Abo, T. Kaminaga, and M. Mikami. Detection of cerebralblood flow changes during repetitive transcranial magnetic stimulation byrecording hemoglobin in the brain cortex, just beneath the stimulation coil,with near-infrared spectroscopy. NeuroImage, 32(3):1226?1230, 2006. ?pages 28128BIBLIOGRAPHY[110] N. Hanaoka, Y. Aoyama, M. Kameyama, M. Fukuda, and M. Mikuni.Deactivation and activation of left frontal lobe during and afterlow-frequency repetitive transcranial magnetic stimulation over rightprefrontal cortex: A near-infrared spectroscopy study. NeuroscienceLetters, 414(2):99?104, 2007. ? pages 28[111] H. Mochizuki, T. Furubayashi, R. Hanajima, Y. Terao, Y. Mizuno,S. Okabe, and Y. Ugawa. Hemoglobin concentration changes in thecontralateral hemisphere during and after theta burst stimulation of thehuman sensorimotor cortices. Experimental Brain Research,180(4):667?675, 2007. ? pages 28[112] H. Mochizuki, Y. Ugawa, Y. Terao, and K. Sakai. Corticalhemoglobin-concentration changes under the coil induced by single-pulsetms in humans: a simultaneous recording with near-infrared spectroscopy.Experimental Brain Research, 169(3):302?310, 2006. ? pages 29[113] G. W. Eschweiler, C. Wegerer, W. Schlotter, C. Spandl, A. Stevens,M. Bartels, and G. Buchkremer. Left prefrontal activation predictstherapeutic effects of repetitive transcranial magnetic stimulation (rtms) inmajor depression. Psychiatry Research: Neuroimaging, 99(3):161?172,2000. ? pages 29[114] T.-C. Chiang, T. Vaithianathan, T. Leung, M. Lavidor, V. Walsh, andD. Delpy. Elevated haemoglobin levels in the motor cortex following 1 hztranscranial magnetic stimulation: a preliminary study. Experimental BrainResearch, 181(4):555?560, 2007. ? pages 29[115] N. Hanaoka, Y. Aoyama, M. Kameyama, M. Fukuda, and M. Mikuni.Deactivation and activation of left frontal lobe during and afterlow-frequency repetitive transcranial magnetic stimulation over rightprefrontal cortex: a near-infrared spectroscopy study. Neuroscience letters,414(2):99?104, March 2007. ? pages 30[116] H. Mochizuki, Y. Ugawa, Y. Terao, and K. L. Sakai. Corticalhemoglobin-concentration changes under the coil induced by single-pulseTMS in humans: a simultaneous recording with near-infrared spectroscopy.Experimental brain research. Experimentelle Hirnforschung.Expe?rimentation ce?re?brale, 169(3):302?10, March 2006. ? pages 30[117] R. J. Ilmoniemi and D. Kicic?. Methodology for combined TMS and EEG.Brain topography, 22(4):233?248, January 2010. ? pages 30129BIBLIOGRAPHY[118] S. Bestmann, C. C. Ruff, F. Blankenburg, N. Weiskopf, J. Driver, and J. C.Rothwell. Mapping causal interregional influences with concurrentTMS-fMRI. Experimental brain research, 191(4):383?402, December2008. ? pages 30[119] G. Strang and T. Nguyen. Wavelets and filter banks. Wellesley-CambridgePress, Wellesley, MA, 1997. ? pages 33[120] S. Mallat. A theory for multiresolution signal decomposition: the waveletrepresentation. IEEE Transactions on Pattern Analysis and MachineIntelligence, 11(7):674?693, July 1989. ? pages 33[121] D. B. Percival and A. T. Walden. Wavelet methods for time series analysis.Cambridge University Press, Cambridge ; New York, 2000. ? pages 34[122] A. Antoniadis, J. Bigot, and T. Sapatinas. Wavelet Estimators inNonparametric Regression: A Comparative Simulation Study. Statisticalsoftware, 83(6):1?83, 2001. ? pages 34[123] H. A. Chipman, E. D. Kolaczyk, and R. E. McCulloch. Adaptive BayesianWavelet Shrinkage. Journal of the American Statistical Association,92(440):1413, December 1997. ? pages 34[124] D. C. Hoaglin, F. Mosteller, and J. W. Tukey. Understanding robust andexploratory data analysis. Wiley, NY, 1983. ? pages 35[125] R. R. Coifman and D. L. Donoho. Translation-invariant de-noising. InWavelets and Statistics, pages 125?150, 1995. ? pages 36[126] X.-P. Zhang and M. D. Desai. Adaptive denoising based on sure risk. IEEESignal Processing Letters, 5(10):265?267, 1998. ? pages 37, 45[127] H. Obrig and A. Villringer. Beyond the visible?imaging the human brainwith light. Journal of cerebral blood flow and metabolism, 23(1):1?18,January 2003. ? pages 40[128] S. R. Hintz, D. A. Benaron, A. M. Siegel, A. Zourabian, D. K. Stevenson,and D. A. Boas. Bedside functional imaging of the premature infant brainduring passive motor activation. Journal of perinatal medicine,29(4):335?343, January 2001. ? pages 40[129] M. E. Al-Mualla. Video coding for mobile communications: efficiency,complexity, and resilience. Academic Press, San Diego, CA, 2002. ?pages 41130BIBLIOGRAPHY[130] T. T. Georgiou. An Intrinsic Metric for Power Spectral Density Functions.IEEE Signal Processing Letters, 14(8):561?563, 2007. ? pages 45[131] Y. Zhang, D. H. Brooks, M. A. Franceschini, and D. A. Boas.Eigenvector-based spatial filtering for reduction of physiologicalinterference in diffuse optical imaging. Journal of biomedical optics,10(1):11014, 2005. ? pages 51[132] D. F. Putnam. Composition and concentrative properties of human urine.Technical Report July, National Aeronautics and Space AdministrationReport CR-1802, Washington DC, 1971. ? pages 53[133] L. Stothers, B. Shadgan, and A. Macnab. Near-infrared spectroscopy of thedetrusor during urodynamics with simultaneous ultrasound measurementsof bladder dimensions and position. Biomedical Spectroscopy andImaging, 1(2):137?145, 2012. ? pages 54[134] J. K. Choi, M. G. Choi, J. M. Kim, and H. M. Bae. Efficient DataExtraction Method for Near-Infrared Spectroscopy (NIRS) Systems WithHigh Spatial and Temporal Resolution. IEEE Transactions on BiomedicalCircuits and Systems, 7(2):169?177, April 2013. ? pages 58[135] R. Choe. Diffuse optical tomography and spectroscopy of breast cancerand fetal brain. Doctoral dissertation, University of Pennsylvania, 2005.? pages 62[136] G. J. Mu?ller and A. Roggan, editors. Laser-induced InterstitialThermotherapy. SPIE Optical Engineering Press, Bellingham, Wash.,1995. ? pages 62, 70[137] M. C. van Beekvelt, M. S. Borghuis, B. G. van Engelen, R. A. Wevers, andW. N. Colier. Adipose tissue thickness affects in vivo quantitative near-IRspectroscopy in human skeletal muscle. Clinical science, 101(1):21?28,July 2001. ? pages 68[138] M. Wolf, P. Evans, H. U. Bucher, V. Dietz, M. Keel, R. Strebel, and K. vonSiebenthal. Measurement of absolute cerebral haemoglobin concentrationin adults and neonates. Advances in experimental medicine and biology,428:219?227, January 1997. ? pages 69, 113[139] M. Wolf, K. von Siebenthal, M. Keel, V. Dietz, O. Baenziger, and H. U.Bucher. Comparison of three methods to measure absolute cerebralhemoglobin concentration in neonates by near-infrared spectrophotometry.Journal of biomedical optics, 7(2):221?227, April 2002. ? pages 69, 113131BIBLIOGRAPHY[140] D. A. Boas. Diffuse photon probes of structural and dynamical propertiesof turbid media: Theory and biomedical applications. Phd dissertation,University of Pennsylvania, 1996. ? pages 70[141] N. K. Kristiansen, S. Ringgaard, H. Nygaard, and J. C. Djurhuus. MRIassessment of the influence of body position on the shape and position ofthe urinary bladder. Scandinavian journal of urology and nephrology,38(1):53?61, January 2004. ? pages 71[142] L. Harrison, W. Penny, and K. Friston. Multivariate autoregressivemodeling of fMRI time series. NeuroImage, 19(4):1477?1491, 2003. ?pages 73[143] B. L. P. Cheung, B. A. Riedner, G. Tononi, and B. D. Van Veen. Estimationof cortical connectivity from EEG using state-space models. IEEEtransactions on bio-medical engineering, 57(9):2122?2134, September2010. ? pages 73[144] S. Weisberg. Applied linear regression. Wiley, 2005. ? pages 73, 74[145] M. Kaminski, M. Ding, W. A. Truccolo, and S. L. Bressler. Evaluatingcausal relations in neural systems: Granger causality, directed transferfunction and statistical assessment of significance. Biological Cybernetics,85(2):145?157, August 2001. ? pages 75[146] G. Dehaene-Lambertz, S. Dehaene, and L. Hertz-Pannier. Functionalneuroimaging of speech perception in infants. Science,298(5600):2013?2015, December 2002. ? pages 76[147] J. P. Szaflarski, S. K. Holland, V. J. Schmithorst, and A. W. Byars. fMRIstudy of language lateralization in children and adults. Human brainmapping, 27(3):202?212, March 2006. ? pages 77[148] Y. Hoshi. Functional near-infrared spectroscopy: current status and futureprospects. Journal of biomedical optics, 12(6):062106, 2007. ? pages 77,86[149] D. Zhang and M. E. Raichle. Disease and the brain?s dark energy. Naturereviews. Neurology, 6(1):15?28, January 2010. ? pages 81, 92, 93, 94, 114[150] C. F. Beckmann, M. DeLuca, J. T. Devlin, and S. M. Smith. Investigationsinto resting-state connectivity using independent component analysis.Philosophical transactions of the Royal Society of London. Series B,Biological sciences, 360(1457):1001?1013, May 2005. ? pages 81132BIBLIOGRAPHY[151] V. L. Cherkassky, R. K. Kana, T. A. Keller, and M. A. Just. Functionalconnectivity in a baseline resting-state network in autism. Neuroreport,17(16):1687?1690, November 2006. ? pages 81[152] A. Anand, Y. Li, Y. Wang, J. Wu, S. Gao, L. Bukhari, V. P. Mathews,A. Kalnin, and M. J. Lowe. Activity and connectivity of brain moodregulating circuit in depression: a functional magnetic resonance study.Biological psychiatry, 57(10):1079?1088, May 2005. ? pages 81[153] S. J. Li, Z. Li, G. Wu, M. J. Zhang, M. Franczak, and P. G. Antuono.Alzheimer Disease: evaluation of a functional MR imaging index as amarker. Radiology, 225(1):253?259, October 2002. ? pages 81[154] L. Tian, T. Jiang, Y. Wang, Y. Zang, Y. He, M. Liang, M. Sui, Q. Cao,S. Hu, M. Peng, and Y. Zhuo. Altered resting-state functional connectivitypatterns of anterior cingulate cortex in adolescents with attention deficithyperactivity disorder. Neuroscience letters, 400(1-2):39?43, May 2006.? pages 81[155] P. Tass, M. G. Rosenblum, J. Weule, J. Kurths, a. Pikovsky, J. Volkmann,A. Schnitzler, and H. J. Freund. Detection of n:m Phase Locking fromNoisy Data: Application to Magnetoencephalography. Physical ReviewLetters, 81(15):3291?3294, October 1998. ? pages 81[156] L. May, J. Gervain, M. Carreiras, and J. F. Werker. The specificity of theneural response to language at birth. In fNIRS Meeting, London, UK, 2012.? pages 82, 85, 87, 93[157] A. V. Oppenheim, R. W. Schafer, and J. R. Buck. Discrete-time signalprocessing. Prentice Hall, Englewood Cliffs, N.J., 1989. ? pages 83, 84[158] K. V. Mardia and P. E. Jupp. Directional Statistics. John Wiley, Chichester,NY, 2000. ? pages 84[159] R. T. Canolty, C. F. Cadieu, K. Koepsell, K. Ganguly, R. T. Knight, andJ. M. Carmena. Detecting event-related changes of multivariate phasecoupling in dynamic brain networks. Journal of neurophysiology,107(7):2020?2031, April 2012. ? pages 85[160] J. C. Lagarias, J. A. Reeds, M. H. Wright, and P. E. Wright. ConvergenceProperties of the Nelder?Mead Simplex Method in Low Dimensions. SIAMJournal on Optimization, 9(1):112?147, January 1998. ? pages 86133BIBLIOGRAPHY[161] C. F. Cadieu and K. Koepsell. Phase Coupling Estimation fromMultivariate Phase Statistics. Neural Computation, 22(12):3107?3126,December 2010. ? pages 86[162] J. R. Binder, S. J. Swanson, T. A. Hammeke, G. L. Morris, W. M. Mueller,M. Fischer, S. Benbadis, J. Frost, S. M. Rao, and V. M. Haughton.Determination of language dominance using functional MRI: Acomparison with the Wada test. Neurology, 46(4):978?984, April 1996. ?pages 87[163] D. Perani, M. C. Saccuman, P. Scifo, A. Anwander, A. Awander, D. Spada,C. Baldoli, A. Poloniato, G. Lohmann, and A. D. Friederici. Neurallanguage networks at birth. Proceedings of the National Academy ofSciences of the United States of America, 108(38):16056?16061,September 2011. ? pages 92[164] M. Mahmoudzadeh, G. Dehaene-Lambertz, M. Fournier, G. Kongolo,S. Goudjil, J. Dubois, R. Grebe, and F. Wallois. Syllabic discrimination inpremature human infants prior to complete formation of cortical layers. InProceedings of the National Academy of Sciences of the United States ofAmerica, pages 1?6, February 2013. ? pages 92[165] Q. Zhang, E. N. Brown, and G. E. Strangman. Adaptive filtering for globalinterference cancellation and real-time recovery of evoked brain activity: aMonte Carlo simulation study. Journal of biomedical optics, 12(4):044014,2010. ? pages 93[166] P. Fransson, B. Skio?ld, S. Horsch, A. Nordell, M. Blennow, H. Lagercrantz,and U. Aden. Resting-state networks in the infant brain. In Proceedings ofthe National Academy of Sciences of the United States of America, volume104, pages 15531?15536, September 2007. ? pages 93[167] M. D. Greicius, G. Srivastava, A. L. Reiss, and V. Menon. Default-modenetwork activity distinguishes Alzheimer?s disease from healthy aging:evidence from functional MRI. Proceedings of the National Academy ofSciences of the United States of America, 101(13):4637?4642, March 2004.? pages 93, 94, 114[168] B. J. He, A. Z. Snyder, J. L. Vincent, A. Epstein, G. L. Shulman, andM. Corbetta. Breakdown of functional connectivity in frontoparietalnetworks underlies behavioral deficits in spatial neglect. Neuron,53(6):905?918, March 2007. ? pages 93134BIBLIOGRAPHY[169] J. M. Masciotti, J. M. Lasker, and A. H. Hielscher. Digital Lock-InDetection for Discriminating Multiple Modulation Frequencies With HighAccuracy and Computational Efficiency. IEEE Transactions onInstrumentation and Measurement, 57(1):182?189, January 2008. ? pages98, 100[170] V. Tuchin. Tissue optics: light scattering methods and instruments formedical diagnosis. SPIE Optical Engineering Press, Bellingham,Washington, 2000. ? pages 102[171] Z. Zhang, B. Sun, H. Gong, L. Zhang, J. Sun, B. Wang, and Q. Luo. A fastneuronal signal-sensitive continuous-wave near-infrared imaging system.The Review of scientific instruments, 83(9):094301, September 2012. ?pages 102, 106, 113, 115[172] T. Muehlemann, D. Haensse, and M. Wolf. Wireless miniaturized in-vivonear infrared imaging. Optics express, 16(14):10323?10330, July 2008. ?pages 102, 103, 106, 113, 115[173] V. Jurcak, D. Tsuzuki, and I. Dan. 10/20, 10/10, and 10/5 SystemsRevisited: Their Validity As Relative Head-Surface-Based PositioningSystems. NeuroImage, 34(4):1600?1611, February 2007. ? pages 103[174] T. J. Huppert, S. G. Diamond, M. A. Franceschini, and D. A. Boas.HomER: a review of time-series analysis methods for near-infraredspectroscopy of the brain. Applied optics, 48(10):D280?D298, April 2009.? pages 104[175] E. Gratton, S. Fantini, M. A. Franceschini, G. Gratton, and M. Fabiani.Measurements of scattering and absorption changes in muscle and brain.Philosophical transactions of the Royal Society of London. Series B,Biological sciences, 352(1354):727?735, June 1997. ? pages 107[176] G. Gratton. Fast optical imaging of human brain function. Frontiers inHuman Neuroscience, 4(June):1?9, 2010. ? pages 107[177] C. K. Kim, S. Lee, D. Koh, and B. M. Kim. Development of wireless NIRSsystem with dynamic removal of motion artifacts. Biomedical EngineeringLetters, 1(4):254?259, December 2011. ? pages 109[178] M. C. Stevens. The developmental cognitive neuroscience of functionalconnectivity. Brain and Cognition, 70(1):1?12, June 2009. ? pages 114135BIBLIOGRAPHY[179] M. L. Pierro, A. Sassaroli, P. R. Bergethon, B. L. Ehrenberg, and S. Fantini.Phase-amplitude investigation of spontaneous low-frequency oscillations ofcerebral hemodynamics with near-infrared spectroscopy: a sleep study inhuman subjects. NeuroImage, 63(3):1571?1584, November 2012. ? pages115[180] M. Reinhard, E. Wehrle-Wieland, D. Grabiak, M. Roth, B. Guschlbauer,J. Timmer, C. Weiller, and A. Hetzel. Oscillatory cerebralhemodynamics?the macro- vs. microvascular level. Journal of theneurological sciences, 250(1-2):103?9, December 2006. ? pages 115[181] M. Diop, J. T. Elliott, K. M. Tichauer, T. Y. Lee, and K. St Lawrence. Abroadband continuous-wave multichannel near-infrared system formeasuring regional cerebral blood flow and oxygen consumption innewborn piglets. The Review of scientific instruments, 80(5):054302, May2009. ? pages 115[182] L. A. Boyd and C. J. Winstein. Cerebellar stroke impairs temporal but notspatial accuracy during implicit motor learning. Neurorehabilitation andneural repair, 18(3):134?143, September 2004. ? pages 115[183] S. K. Meehan, B. Randhawa, B. Wessel, and L. A. Boyd. Implicitsequence-specific motor learning after subcortical stroke is associated withincreased prefrontal brain activations: an fMRI study. Human brainmapping, 32(2):290?303, February 2011. ? pages 115136


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



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"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items