Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Characterization of myelin water imaging using a gradient and spin echo sequence in human brain and spinal… Ljungberg, Emil 2016

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

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

Item Metadata

Download

Media
24-ubc_2016_may_ljungberg_emil.pdf [ 13.46MB ]
Metadata
JSON: 24-1.0228626.json
JSON-LD: 24-1.0228626-ld.json
RDF/XML (Pretty): 24-1.0228626-rdf.xml
RDF/JSON: 24-1.0228626-rdf.json
Turtle: 24-1.0228626-turtle.txt
N-Triples: 24-1.0228626-rdf-ntriples.txt
Original Record: 24-1.0228626-source.json
Full Text
24-1.0228626-fulltext.txt
Citation
24-1.0228626.ris

Full Text

Characterization of Myelin Water Imaging Using a Gradient and SpinEcho Sequence in Human Brain and Spinal CordbyEmil LjungbergB.Sc., Lund University, 2013A THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFMaster of Applied ScienceinTHE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES(Engineering Physics)The University of British Columbia(Vancouver)April 2016c© Emil Ljungberg, 2016AbstractMyelin water imaging is a quantitative magnetic resonance imaging technique that can be usedas an in vivo biomarker for myelin in the central nervous system. In 2007, a paradigm shift tookplace when the standard sequence for myelin water imaging changed from a multi-echo spin echosequence to a gradient and spin echo (GRASE) sequence. The GRASE sequence has so far onlybeen applied to brain imaging, and reproducibility between different scan vendors has not beenassessed. In this study I present the first implementation of myelin water imaging using GRASEin human cervical spinal cord. The reproducibility of myelin water imaging in the spinal cord wasfound to be high (coefficient of variation = 6.1%, Cronbach’s α = 0.89). A multicenter repro-ducibility study of myelin water imaging in brain between two scan vendors (Siemens and Philips)was also performed. Results from the two scanners were found to be highly correlated but witha significant offset in myelin water fraction of 4.3 %. Together, these two studies provide strongevidence of the reproducibility of myelin water imaging. It is an important step forward in thedevelopment of bringing myelin water imaging to the mainstream.iiPrefaceThe work presented in this thesis was carried out at the MRI Research Centre at the University ofBritish Columbia between May 2015 and April 2016. In addition to the formal supervision by Dr.Alex MacKay, I also received very helpful supervision from Dr. Shannon Kolind.The data presented in chapter 2 was collected and analyzed by the author Emil Ljungbergduring the fall of 2015 and spring of 2016. I was the lead of the development of the project. Thedata presented in the main study presented in chapter 3 was collected prior to the start of this thesisproject by Dr. Shannon Kolind. I contributed with data collection of the sub-studies, data analysisof all the data as well as leading the project. Lisa Lee contributed with parts of the image analysisin this chapter.Published WorkThe results in chapter 2 were accepted for oral and poster presentation the International Society forMagnetic Resonance in Medicine (ISMRM) 2016, Singapore.• Moving Towards Clinically Feasible Myelin Water Assessment in Human Cervical SpinalCord. Emil Ljungberg, Irene Vavasour, Roger Tam, Youngjin Yoo, Alexander Rauscher,David Li, Anthony Traboulsee, Alex MacKay, Shannon Kolind; Proc. Intl. Soc. Mag. Reson.Med. Singapore, May 2016 (accepted)iiiParts of the results in chapter 3 were accepted for poster presentation at the American Academyof Neurology (AAN) 2016, Vancouver• Examining the consistency of myelin-specific magnetic resonance imaging across sites andscanner manufacturers in vivo. Lisa Lee, Emil Ljungberg, Alex MacKay, Alexander Rauscher,David Li, Anthony Traboulsee, Chase Figley, Shannon Kollind; AAN Vancouver, April 2016(accepted)Study EthicsParticipating volunteers in this study consented under the approval of the Clinical Ethics Researchboard at the University of British Columbia (New MRI for Neuroimaging, H14-01083).ivTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiAbbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviiiAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xix1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Magnetic Resonance Imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.1.1 The Physics of MRI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.1.2 T1, Spin-Lattice Interactions . . . . . . . . . . . . . . . . . . . . . . . . . 51.1.3 T2, Spin-Spin Interactions . . . . . . . . . . . . . . . . . . . . . . . . . . 61.1.4 Spatial Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81.2 The Central Nervous System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111.3 Myelin Water Imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121.3.1 T2 Relaxation in vivo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141.3.2 Acquisition - Pulse Sequences for MWI . . . . . . . . . . . . . . . . . . . 151.3.3 Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211.3.4 Applications of MWI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251.3.5 Alternatives to MWI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26v1.4 Spinal Cord MRI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341.4.1 Studying Spinal Cord Pathology Using MRI . . . . . . . . . . . . . . . . . 351.4.2 Motion and Flow Artifacts in MRI . . . . . . . . . . . . . . . . . . . . . . 351.4.3 Flow and Motion is Correlated to the Cardiac Cycle . . . . . . . . . . . . 381.4.4 Motion Correction Techniques . . . . . . . . . . . . . . . . . . . . . . . . 381.4.5 Insufficient Field of View . . . . . . . . . . . . . . . . . . . . . . . . . . 421.4.6 Magnetic Field Inhomogeneities . . . . . . . . . . . . . . . . . . . . . . . 431.5 Overview of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452 Rapid Myelin Water Imaging in Human Cervical Spinal Cord . . . . . . . . . . . . 472.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472.2 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482.3 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482.4 Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492.5 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492.5.1 Study Population . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492.5.2 MRI Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492.5.3 MWI Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 512.5.4 Image Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 522.5.5 Statistical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 532.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 552.6.1 Comparison Between GRASE and MSE MWI . . . . . . . . . . . . . . . 552.6.2 Reproducibility of GRASE MWI . . . . . . . . . . . . . . . . . . . . . . 562.6.3 Cross Sectional Area Measurements . . . . . . . . . . . . . . . . . . . . . 592.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 602.7.1 Comparison Between GRASE and MSE MWI . . . . . . . . . . . . . . . 602.7.2 Reproducibility of GRASE MWI . . . . . . . . . . . . . . . . . . . . . . 612.7.3 Cross Sectional Area Measurements . . . . . . . . . . . . . . . . . . . . . 632.7.4 Imaging Artifacts and Remedies . . . . . . . . . . . . . . . . . . . . . . . 652.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 703 Multi-vendor Reproducibility of Myelin Water Imaging . . . . . . . . . . . . . . . . 713.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713.2 Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 733.3 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73vi3.4 Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 733.5 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 743.5.1 Study Population . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 743.5.2 MRI Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 743.5.3 Image Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 783.5.4 Statistical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 783.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 793.6.1 Myelin Water Fraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 793.6.2 Geometric Mean T2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 803.6.3 Refocusing Flip Angle . . . . . . . . . . . . . . . . . . . . . . . . . . . . 823.6.4 T2 Decay Curves and Distributions . . . . . . . . . . . . . . . . . . . . . 833.6.5 Simulations of Changing the Refocusing Flip Angle . . . . . . . . . . . . 833.6.6 MR Experiment of Changing Refocusing Flip Angle . . . . . . . . . . . . 843.6.7 MWI Analysis Without Stimulated Echo Correction . . . . . . . . . . . . 863.6.8 Fat Suppression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 863.6.9 Effect of Myelin T2 Window . . . . . . . . . . . . . . . . . . . . . . . . . 873.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 883.7.1 Flip Angle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 883.7.2 Geometric Mean T2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 903.7.3 Fat Saturation Pre-pulse . . . . . . . . . . . . . . . . . . . . . . . . . . . 903.7.4 Implications for Future Multicenter Studies . . . . . . . . . . . . . . . . . 913.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 924 Conclusions and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 944.1 Rapid Myelin Water Imaging in Human Cervical Spinal Cord . . . . . . . . . . . . 944.1.1 Impact of the Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 944.1.2 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 954.1.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 964.2 Multi-vendor Reproducibility of Myelin Water Imaging . . . . . . . . . . . . . . . 974.2.1 Impact of the Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 974.2.2 Limits of the Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 974.2.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 984.3 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100viiA Supporting Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109A.1 Changing the flip angle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109A.2 Mathematical Derivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110A.2.1 Even-odd echo refocusing . . . . . . . . . . . . . . . . . . . . . . . . . . 110viiiList of TablesTable 1.1 Compilation of myelin water fraction (MWF) from various studies in humanbrain and spinal cord. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27Table 2.1 Puls sequence parameters used for the gradient- and spin-echo (GRASE) andmulti echo spin-echo (MSE) for myelin water imaging (MWI) in the spinal cord. 50Table 2.2 Linear and orthogonal regression parameters between the average MWF ob-tained from the repeated GRASE scan and the MSE scan. . . . . . . . . . . . . . 56Table 2.3 MWF in human cervical spinal cord in vivo from previous studies. A rangein vertebral levels is indicated by a - and specific disc location between twovertebrae with /. The white matter (WM) region of interest (ROI) is the averageof the dorsal column (DC) and corticospinal tract (CCST) ROI. If no ROI isspecified, the MWF is an average of the ROI includes the whole cord. MWF fromthe GRASE in the present study is the average of the two repeated scans. . . . . 58Table 2.4 Summary of reproducibility measures of measured MWF with the repeated GRASEscans. ROI: dorsal column (DC), corticospinal tract (CCST), gray matter (GM). . 59Table 2.5 Difference between scans in mm2 and p-value for two tailed paired t-test be-tween the two measurements. . . . . . . . . . . . . . . . . . . . . . . . . . . . 60ixTable 3.1 Pulse sequence parameters used for the two scanners. (∗ Fat-saturation consistedof a spectral fat saturation pulse at 3.4 ppm off the main water peak resonancefrequency.) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75xList of FiguresFigure 1.1 Schematic description of how a radio frequency (RF) pulse perturbs the magne-tization vector M away from the z-axis. After the RF pulse have been applied,the transverse magnetization Mt(t) decays with a time constant T2, green linein the rightmost graph, while the longitudinal magnetization recovers with atime constant T1, red line in the rightmost graph. . . . . . . . . . . . . . . . . 5Figure 1.2 Plot showing the dependence of T1 and T2 on correlation time τc at 3 T. Ashort correlation time is equivalent to freely moving spins with long T1 and T2.Conversely, long correlation time corresponds to solids with very short T1 andT2. Figure generated using the theoretical calculations presented in Ref. [13] . 7Figure 1.3 Pulse sequence diagram of a spin-echo sequence to the left and corresponding,simplified, k-space diagram to the right. Note the asymmetric slice selectiongradient for the 180◦ RF pulse which is used to balance the slice selective gra-dient for the 90◦ RF pulse. In the k-space diagram, the phase encoding stepsare represented by the 4 vertical arrows of varying length. . . . . . . . . . . . 9xiFigure 1.4 The central nervous system (CNS) consists of the brain and spinal cord. Thespinal cord is divided in three segments (A). Histological images of the brain(B) and the spinal cord (C) shows the typical distribution of white matter (WM)and gray matter (GM) (The brain is sample from a Macaque monkey). Onthe micro structure level of the CNS we find single neurons as shown in (D)consisting of a cell body, myelinated axon, and dendrites. If we zoom in on across section of an axon we can observe the multi layered structure of myelinaround the axon (E). Image sources: (A) Ref. [16] (B) Ref. [78] (C) Ref. [56],(D) Ref. [34] (E) Ref. [64] . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13Figure 1.5 Pulse sequence diagram of the 3D GRASE sequence as visualized by the Philipsscanner simulator software environment. For each spin echo, which occurs inbetween the 180◦ RF pulses, there are 3 gradient lobes for readout, as indicatedby the three red lines at the bottom. Positive gradients move in the positivedirection of k-space, and negative gradients in the negative direction. Whena 180◦ RF pulse is played out, the position in k-space is flipped, equivalent tomultiplying each coordinate by -1. kx translates to −kx etc. . . . . . . . . . . . 20Figure 1.6 Example of single exponential synthetic T2 decay curves generated using theextended phase graph (EPG) algorithm. In (a) α is varied between 60-180◦ andαre f con = 180◦ . In (b) T2 is varied between 20-200 ms. . . . . . . . . . . . . . 22xiiFigure 1.7 Histological verification of myelin water imaging (MWI) adapted with permis-sion from Laule et al. [42]. (a) First echo from multi echo sequence at echotime (TE)=10 ms. (b) MWI clearly showing higher MWF in white matter in thebrain which is well correlated with luxol fast blue (LFB) staining for myelinin (c). Note the clear correspondence between the bright pixels in (b) show-ing high MWF and dark blue pixels in (c) showing a high myelin content fromstaining. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26Figure 1.8 Comparison of myelin estimates obtained with GRASE and multicomponentdriven equilibrium single pulse observation of T1/T2 (MCDESPOT). It is clearthat the techniques produce significantly different estimates of myelin in theCNS. MCDESPOT estimates a more or less homogeneous distribution of myelinin WM compared to GRASE which shows a wide range of MWF values in WM.Figure adapted with permissions from the author from Zhang et al. [94] . . . . 29Figure 1.9 Variation of fraction of signal associated with myelin water ( fm) along the cordobtained with MCDESPOT by Kolind et al. [39]. Adapted from Ref. [39] withpermissions from the author. . . . . . . . . . . . . . . . . . . . . . . . . . . . 36Figure 1.10 (a) Diagram of the Fourier relations behind the ghosting motion artifact. Whenthe object is static, the object is correctly reconstructed by the inverse Fouriertransform. When the object is moving, the data in the Fourier plane can bedescribed as the data from the static object multiplied by an oscillation withthe modulation frequency. The reconstructed object will then appear in severalcopies. Image adapted with permission from: Common Artifacts in MRI, byQing-San Xiang, PHYS542, UBC, 2014. (b) Example of flow artifact fromcerebrospinal fluid (CSF) in the spinal cord. . . . . . . . . . . . . . . . . . . . 37xiiiFigure 1.11 Example of pulse sequence using a spatial saturation pulse to null the signalfrom inflowing spins. The first 90◦ pulse excites a slice covering the area thatneeds to be suppressed, this is followed by strong crusher gradients in the fre-quency and phase encoding direction to destroy any magnetization from theslice. After this, a regular pulse sequence can be performed. . . . . . . . . . . 41Figure 1.12 Gradient designs to null zeroth moment (a) and first moment (b). With twolobes only m0(t) = 0 after the gradient. With three lobes m0(t) = m1(t) = 0after the gradient. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42Figure 2.1 Screen shot from scanner interface showing imaging volume, slices, shim boxand the spatial saturation band used in the multi-echo fast field echo (MFFE).Imaging volume was the same for all scans but saturation band was only usedfor MFFE. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51Figure 2.2 Overview of the image analysis pipeline using tools from the Spinal Cord Tool-box with the ROI used in the analysis. . . . . . . . . . . . . . . . . . . . . . . 53Figure 2.3 (a) Correlation plot for the MWF between the GRASE and MSE. Fit statisticsshown in table 2.2. (b) Bland-Altman plot comparing the GRASE and MSEsequence. Mean=-0.0091, 95% confidence interval=mean±0.039. MWF fromthe GRASE in a and b is the average of the two repeated scans. . . . . . . . . . 55Figure 2.4 (a) Example of T2 decay curves obtained from all three scans in DC. Decaycurve amplitude was normalized to 1 at the first echo. (b) T2 distributionsobtained from MWI analysis. From left to right the peaks in the spectrum areattributed to: myelin water (< 15 ms), intra-extra cellular water (≈ 70 ms), andCSF > 2000 ms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57Figure 2.5 Overview of the different scans from the first 6 subjects obtained from themiddle slice registered to multi-echo fast field echo (MFFE) space. . . . . . . . 57xivFigure 2.6 Bland-Altman analysis of MWF obtained from the repeated GRASE scans. Mean=-0.0068, 95% confidence interval=mean±0.051. . . . . . . . . . . . . . . . . . 59Figure 2.7 Box plot of cross sectional area (CSA) measurement obtained with the MFFE,GRASE, and MSE scans. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60Figure 2.8 Schematic explanation of the concept behind even-odd echo refocusing for asimple MSE sequence. Spin echoes are labeled E1, E2, E3, and E4. K is definedin 2.3. At even echoes the net phase is zero, thus high signal. . . . . . . . . . . 67Figure 2.9 (a) Example of flow artifact in a MFFE scan. Circled at the top is a replica ofthe CSF inside the cord which originates from the pulsating flow of CSF. (b)Example of how fold over artifacts can affect the GRASE image. . . . . . . . . 70Figure 3.1 Statistics from http://www.pudmed.gov showing number of publications match-ing keyword multicenter trial MRI between 1990 and 2014. Data retrieved onMarch 1 2016. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72Figure 3.2 Overview of all the manually drawn ROI on a single slice from the magnetiza-tion prepared rapid acquisition gradient echo (MPRAGE) scan. . . . . . . . . . 79Figure 3.3 (a) Linear relationship between Siemens and Philips-acquired MWF data. Lin-ear regression parameters (95% confidence interval): slope=0.84 (0.67, 1.01),y-intercept: -0.023 (-0.051, 0.0048), R2 = 0.82. Orthogonal regression: slope=0.91(0.67, 1.16), y-intercept=-0.034 (-0.068, -0.00086). (b) Bland-Altman analy-sis comparing MWF results obtained from the Siemens and Philips scanners.Mean difference=-0.043, 95% confidence interval=(0.0081,-0.095) . . . . . . . 80Figure 3.4 Overview of MWF, geometric mean T2 (GMT2), and flip angle from both cen-ters in one subject. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81xvFigure 3.5 MWF histograms from subject 1 (a), 2 (b), and 3 (c). white matter (WM) datashown in blue, gray matter (GM) in red. Data from Siemens is shown as solidlines and Philips as dashed lines. . . . . . . . . . . . . . . . . . . . . . . . . . 81Figure 3.6 GMT2 histograms from subject 1 (a), 2 (b), and 3 (c). white matter (WM) datashown in blue, gray matter (GM) in red. Data from Siemens is shown as solidlines and Philips as dashed lines. . . . . . . . . . . . . . . . . . . . . . . . . . 82Figure 3.7 Histograms for the estimated refocusing flip angle from both scanners for eachof the three subjects, (a), (b), and (c). . . . . . . . . . . . . . . . . . . . . . . 82Figure 3.8 (a) T2 decay curves and (b) T2 distributions obtained from the genu ROI fromboth scanners. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83Figure 3.9 Simulations showing how (a) MWF is affected by the flip angle with the trueMWF in dashed lines and the MWF calculated with the EPG analysis in solidlines. In (b), the relationship between estimated flip angle and input flip angleis shown. It is a striking 1:1 relationship for all flip angles except those closeto 180◦ . Note that all 5 lines overlap here and only the green line is visible. . . 84Figure 3.10 Graph showing the estimated flip angle distribution for the whole brain givenprescribed flip angle: α = 140,160,180◦ . The purple line shows the flip angledistribution acquired on the Siemens scanner from the same subject. . . . . . . 85Figure 3.11 Graphs showing variations in whole brain WM for (a) MWF and (b) for GMT2for varying flip angle (α). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85Figure 3.12 Without stimulated echo correction the MWF is significantly reduced in the datacollected by the Siemens scanner. . . . . . . . . . . . . . . . . . . . . . . . . 86xviFigure 3.13 Results from magnetic resonance imaging (MRI) experiment with fat saturationpulse performed on the Philips scanner with subject 3. (a) The MWF was notsignificantly affected, and the flip angle distribution as shown in (b) showedonly a slight variation in shape. . . . . . . . . . . . . . . . . . . . . . . . . . . 87Figure 3.14 (a) Full brain white matter histograms with varying myelin T2 window rangewith upper limit set to 30, 35, and 40 ms. (b) Difference in average MWFbetween Siemens and Philips within each ROI for varying T2 window. . . . . . 88Figure 3.15 Subject 3 scanned on the Philips scanner with and without fat suppression (spectral presaturation with inversion recovery (SPIR)). Red arrows point outthe different reduced signal from subcutaneous fat when fat suppression is uti-lized. Green arrows point out image artifacts in the scan obtained without fatsuppression. Note how many of these artifacts are absent in the image obtainedwith fat suppression. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92Figure A.1 Example of frequency encoding gradient lobes used for readout for each echoin GRASE and MSE. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110xviiAbbreviationsMRI magnetic resonance imagingGRASE gradient- and spin-echoMSE multi echo spin-echoMCDESPOT multicomponent driven equilibrium single pulse observation of T1/T2MFFE multi-echo fast field echoMWI myelin water imagingMWF myelin water fractionfM fraction of signal associated with myelin waterGMT2 geometric mean T2EPG extended phase graphNNLS non-negative least squaresTE echo timeTR repetition timeCNS central nervous systemMS multiple sclerosisCSF cerebrospinal fluidWM white matterGM gray matterxviiiAcknowledgmentsFirst and foremost, I would like thank my supervisors Dr. Alex MacKay and Dr. Shannon Kolindfor their outstanding support during my thesis. Through countless discussions, mainly about theGRASE sequence, I gained valuable knowledge that I will take with me in my future career. Iwould also like to thank Dr. Anthony Traboulsee, who have provided valuable insight on myproject from a clinical point of view. This have forced me to constantly consider the practicalapplications of my work, which also is a strong driving factor for continuing doing my research. Iwould also like to thank: our head MRI technologist Laura Barlow and the whole team at the MRIscanner for outstanding support; Dr. Irene Vavasour for countless consultations during my day today work, your knowledge about everything that is going on is an invaluable resource to everyonein the lab; Dr. David Li for valuable input on spinal cord MRI; Dr. Rick White for helpful adviceon statistics; Julia Schubert for proofreading and constant support during the thesis; Lisa Lee forhelping out with the multicenter study and producing a poster for AAN; and all the volunteers thatparticipated in this project.Finally I want to thank my family who, although only available over Skype, have been the mostimportant support during this last year. Thank you, without you this would not have been possible.xixChapter 1IntroductionThe profound study of nature is the most fertile source of mathematical discoveries.Joseph Fourier (1769-1830)THE year was 2003, and Paul C. Lauterbur and Peters Mansfield have traveled to Stockholm,Sweden, to receive the Nobel Prize in Medicine for their outstanding contribution to thefield of magnetic resonance imaging (MRI). The motivation from the scientific committee read:These discoveries have led to the development of modern MRI which represents a breakthrough inmedical diagnostics and research. Although several full body imaging modalities had been avail-able since the discovery of x-rays in the early 1900’s, MRI emerged as the first full body imagingtechnique without ionizing radiation. The combination of unprecedented image contrast and notusing ionizing radiation made MRI a true, and much needed, breakthrough in medical imaging.During the last three decades since MRI was first introduced into hospitals, new imaging proto-cols have emerged from research facilities all around the world, enabling detailed study of specificdiseases. One condition that has gained great interest from researchers is multiple sclerosis (MS),partly because of the large prevalence around the world but also because of the capability of MRIto image pathology in the central nervous system (CNS).1The main focus of this thesis is a field in MRI research called myelin water imaging (MWI)where myelin, a type of fat found around nerve cells in the central nervous system, is the maincontrast in the images. To understand how MWI works I will begin with an introduction to MRI,how data is acquired in MRI, and how contrast is generated. This will lead us further into a briefoverview of the CNS, and finally how MWI attempts to quantify myelin using T2 relaxation.Chapter 2 of this thesis focuses on MWI in human spinal cord, in vivo, and because of this Ihave dedicated an extensive section in this chapter to image artifacts in MRI which are specific tospinal cord imaging.1.1 Magnetic Resonance ImagingThis section will cover a basic description of the physical phenomena that enables imaging of thehuman body using MRI. The concepts that are described and derivations that are made can befound in any basic book on MRI. Therefore, explicit references will only be made in cases whereit is needed for context. The reader is referred to standard works such as ref. [30] and ref. [8] forfurther details.1.1.1 The Physics of MRIAs the name implies, MRI produces an image through a phenomenon known as nuclear magneticresonance. The nucleus of interest in MRI is that of hydrogen, and it will later be shown that mainlyaqueous hydrogen atoms (H2O) are of interest in MRI. Like all particles, the hydrogen nucleus hasa property known as spin. The spin of the hydrogen nucleus can only take two values, 1/2 (up) or−1/2 (down). These states are normally degenerated, i.e., they have the same energy. If the spin isplaced in an external magnetic field, the degeneracy is broken and the up and down quantum statesnow have different energy, a phenomenon known as Zeeman splitting. Now, the up state possessesthe lowest energy and is thus a slightly preferred state over the down state. In a large ensemble ofspins placed in an external magnetic field, the fraction of spins in each state can be estimated using2the Boltzmann distribution as:R=NparNantipar= e−∆EkT (1.1)where ∆E is the quantum energy difference between the parallel and anti-parallel states, k is Boltz-mann’s constant and T is temperature. If the ensemble is placed in a magnetic field, the energydifference between the states depends on the strength of the applied magnetic field. As a numericalexample, an ensemble of spins at room temperature placed in a magnetic field with a strength of 1.5T will have a 0.00102% excess of spins in the parallel state. Since spins with opposite direction willcancel each other out, therefore only the excess of parallel spins will contribute to the measuredmagnetization. Thus, the signal in MRI is very small.When the hydrogen nucleus is placed in an external magnetic field it will begin to precessaround the direction of the field with a frequency known as the Larmor frequency ω:ω = γB (1.2)where γ is the gyro magnetic ratio (267 ·106 rads−1T−1 for 1H) and B is the strength of the externalmagnetic field measured in Tesla (T). Thus, in a 3 T magnetic fieldω = 267 ·106 ·3= 8.01 ·108 rad/s= 127.5 MHz. The precession of the net magnetization of a large ensemble of spins when placed ina magnetic field is similar to that of a pendulum placed in a gravitational field. If the pendulum isnot aligned with the gravitational field lines it will swing from side to side until it finally reaches itsequilibrium. The same applies for the ensemble of spins, but instead of swinging from side to sidethe net magnetization vector will approach the equilibrium position while precessing, thus tracingout a spiral in space.3In equilibrium, net magnetization from the spin ensemble will align with the direction of themagnetic field fig 1.1a. Using a radio frequency (RF) magnetic pulse (RF pulse)1 at a frequencycorresponding to the Larmor frequency, the magnetization vector can be rotated into the transverseplane. This will cause the spins to precess around the z-axis and produce a magnetic flux that canbe detected using a receiver coil. This is the signal that is measured in MRI. The shape and durationof the RF pulse determines the flip angle α , i.e., the angle between magnetization vector after theRF pulse and the principal direction of the magnetic field. After the RF pulse is applied, the spinswill gradually return to the equilibrium position, a rate characterized by two time constants T1 andT2, see figure 1.1.To put this in a mathematical context we observe the results from the the seminal paper from1964 by Felix Bloch: the Bloch equation. This is by far the most commonly used description of thetime evolution of the magnetization vector M = (Mx,My,Mz) of nuclear spins in a magnetic field[12]. The Bloch equation shown in (1.3) explains how the change of magnetization, M, dependson the magnetic field B and the time constants T1 and T2dMdt= γM×B−Mxxˆ+MyyˆT2− (Mz−M0)zˆT1(1.3)Solutions to the Bloch equation are usually given for the longitudinal direction z which is thedirection of the applied magnetic field, (1.4), and the transverse plane xy, (1.5);Mz(t) =M0+(Mz(0)−M0)e−t/T1 (1.4)MT (t) =M0 · e−t/T2 , MT =√M2x +M2y (1.5)1The term Radio frequency and RF pulse might be confusing in this context. Radio frequency refers to any type ofelectromagnetic wave in the frequency range of 3 kHz to 300 GHz. In MRI we are interested in affecting the precessingprotons at the Larmor frequency, which is around 128 Mhz, i.e., in the radio frequency range.4Figure 1.1: Schematic description of how a RF pulse perturbs the magnetization vector Maway from the z-axis. After the RF pulse have been applied, the transverse magnetiza-tion Mt(t) decays with a time constant T2, green line in the rightmost graph, while thelongitudinal magnetization recovers with a time constant T1, red line in the rightmostgraph.After excitation by a 90◦ RF pulse, the z component, Mz(t), recovers with a time constant T1 and thexy component, MT (t), decays with a time constant T2, as seen in figure 1.1. From (1.4) and (1.5), itcan be seen that the signal from samples with different T1 and T2 constants will decay at differentrates, and can thus produce contrast between different tissue using specialized sequences. T1 andT2 are two fundamental properties of tissue that are used in MRI. However, the physical systemdescribed in (1.3) is an over simplification of reality which may not hold true in tissue composedof several components with different T1 and T2.1.1.2 T1, Spin-Lattice InteractionsBy observing the solutions to the Bloch equation it can be seen that T1 characterizes the decay andregrowth of the longitudinal magnetization Mz. The theory behind this relates to what is knownas Spin-Lattice interactions. After applying a RF pulse, the whole system of spins is perturbedfrom the equilibrium state. Like any system in nature, it will strive to reduce entropy and return toequilibrium. This can be achieved by exchange of energy between the spins and the surroundingatom lattice. The exchange of energy between the spins and lattice is proportional to the difference5in longitudinal magnetization between the current state and equilibrium (Mz−M0), as shown bythe third term in (1.3). The spin-lattice interactions causing T1 decay are non-reversible.1.1.3 T2, Spin-Spin InteractionsThe lowest echo time (TE) available on the MRI scanner will limit the range T2 components that willcontribute to the contrast in the image. Human MRI scanners used in clinic and research usuallyhave a lower limit TE around 5 ms for the most common pulse sequences. If a material with T2=5ms and TE=5 ms only 37% of the initial magnetization will remain when the signal is received.Solid materials exhibit a T2 on the sub-millisecond scale and are not detectable using conventionalMRI methods2. Therefore, in the remainder of this discussion on T2 relaxation, only the signal fromaqueous protons is considered.The T2 relaxation, describing the decay of the gross transverse magnetization component (1.5),is related to spin-spin interactions which are affected by the environment that water protons arecontained within. The rate of which water molecules randomly reorient decides the T2. A completeexplanation to the T2 relaxation phenomenon requires an extensive quantum mechanical derivation.Instead, a classical description of this phenomenon is presented. Consider the following scenario:water molecules contained in a given space will tumble around, rotating, stretching, bouncing offeach other at a high frequency, characterized by a short correlation time called τc. With a shortτc, protons will experience an essentially constant magnetic field from neighboring protons as theproton is moving too fast to experience any fluctuations. As the volume is decreased, motionis restricted by attractive forces from the boundaries causing longer τc. Aqueous protons nowhave access to slow moving, non-aqueous protons of the boundary with a very short T2. Theslow moving aqueous protons will now experience a time varying magnetic field from neighboringprecessing protons. This will manifest as loss of phase coherence, i.e., the ensemble of spins will2Ultrashort TE (UTE) techniques are available for acquiring an image from non-aqueous protons. UTE utilize specialexcitation and read out sequence to achieve a TE in the range of 0.05-0.2 ms. For more information about UTE pleaserefer to Ref. [65]6Figure 1.2: Plot showing the dependence of T1 and T2 on correlation time τc at 3 T. A shortcorrelation time is equivalent to freely moving spins with long T1 and T2. Conversely,long correlation time corresponds to solids with very short T1 and T2. Figure generatedusing the theoretical calculations presented in Ref. [13]fan out in the transverse plane, reducing the net transverse magnetization, resulting in a shorter T2.The full scope of this is best understood by plotting T2 against τc, see figure 1.2. In this figure it isclearly seen that for short τc, a high tumbling frequency results in a long T2, while a long τc resultsin short T2.The transverse magnetization can also be lost due to magnetic field inhomogeneities causingdephasing of the spins. This is quantified by the parameter T2′. In contrast to pure T2 decay, theeffects of T2′ are to some extent recoverable. While the effects causing T2 decay are random, themagnetic field inhomogeneities causing T2′ are considered static and we can reverse the effect toregain the magnetization. This is achieved by creating a spin echo using a 180◦ RF refocusingpulse (this will be discussed in detail in the next section). Decay of the transverse magnetizationcan therefore be characterized by yet another parameter T2∗ as1T ∗2=1T2+1T ′2. (1.6)71.1.4 Spatial LocalizationSo far we have only dealt with how to produce transverse magnetization that can be detected thescanner, the next step is to deduce from where the signal originates. The principle of spatial lo-calization in MRI is widely disparate from that of any other medical imaging modality. Whereas aspatial detector is used in most other techniques, MRI employs electromagnetic coils that detect theflux of the precessing spins inside the body that is being imaged.There are three main features of spatial localization that are important in MRI: slice selection,frequency encoding, and phase encoding. Magnetic field gradients that are capable of producing aspatially varying magnetic field are key to achieving spatial localization. At this point it is usefulto discuss these concepts with regards to a simple MRI pulse sequence. Although very simple, thespin echo sequence, shown in figure 1.3 serves as a good starting point for our discussion on spatiallocalization.A 90◦ RF pulse is played out to generate a transverse magnetization component. In this firststep, only a single slice in a plane perpendicular to the z-axis of the sample is affected. This isachieved by applying a magnetic field gradient along the z-axis, thus creating a spatial variationof the Larmor frequency of spins along the z-axis. To excite only a single slice, a slice selectiveRF pulse can be achieved by tuning the RF frequency to that of the Larmor frequency correspond-ing to a certain position along the z-axis. As shown in figure 1.3, the RF pulse has a sinc-likeshape which, through Fourier analysis, can be shown to be a rectangular function in the frequencydomain. Because of the magnetic field gradient along the z-axis, the spatial excitation profile isapproximately equal to the Fourier transform of the RF pulse. Therefore, a sinc RF pulse will excitea single uniform rectangular slice.As soon as the slice have been excited, a T2∗ decay of the magnetization starts. Next, a 180◦RF pulse is applied at t = τ to later produce a spin echo at t = 2τ = TE. The 180◦ pulse flips themagnetization in the xy plane, which causes the previously dephasing spins to come together again8Figure 1.3: Pulse sequence diagram of a spin-echo sequence to the left and corresponding,simplified, k-space diagram to the right. Note the asymmetric slice selection gradientfor the 180◦ RF pulse which is used to balance the slice selective gradient for the 90◦ RFpulse. In the k-space diagram, the phase encoding steps are represented by the 4 verticalarrows of varying length.and the net magnetization MT increases, reaching a maximum at t = 2τ . After the 180◦ RF pulseis applied, a gradient along the y-axis is used to create phase-encoding. This is a process whichgives yet another part of the spatial localization. Slice selection provides the z coordinate and thephase encoding provides either the x or y coordinate, depending on the settings on the scanner.However, for the sake of consistency in this report, I will refer to phase encoding as being alongthe y direction. During phase encoding, a linear gradient with strength Gy is applied along they-axis to produce a spatially varying magnetic field in the sample. By applying this gradient for atime τphase, spins will acquire a phase φ that can be calculated asφ = γ · y ·G · τphase. (1.7)9The phase encoding step is repeated for a range of amplitudes of Gy in order to relate the rate ofchange of phase to position within the sample. The rate of change of phase is equal to frequencyand thus the y coordinate of the signal can be recovered using an inverse Fourier transform.The last step in pulse sequence is signal read out, or frequency encoding. Similar to phaseencoding, a gradient is applied along the x-axis to create a spatially varying field. The MRI signalis acquired while the gradient is applied causing the spins to precess with slightly different Larmorfrequency depending on their position along the x-axis in the sample.To sum up, a slice selective RF pulse is applied together with a z-gradient to isolate a singleslice, i.e., the z coordinate. A gradient in the y-direction is applied to phase encode spins in thesample and a gradient in the x direction is applied to frequency encode the spins. These conceptsare hard to visualize and when advanced combinations of these components are used to buildsophisticated pulse sequences, we need a tool to visualize the data acquisition process. For thispurpose, MRI physicists created what is known k-space, a visual representation of data in a spatialfrequency domain, where data points are measured in the unit m−1. It was developed in the early1980s by Ljunggren [48] and Twieg [77] and has since become the standard for visualization of dataacquisition in MRI. To exemplify this, we look at the pulse sequence presented in figure 1.3 and itscorresponding, very simplified, k-space diagram to the right in the same figure. Frequency encodinggradients are represented by the kx axis, and phase encoding along the ky axis. A positive gradienttranslates to movement in a positive direction in k-space and the position is determined by the timeintegral of the gradient. If we denote the size of k-space as covering the range [−kymax,kymax],[−kxmax,kxmax] we findkmax = γGT (1.8)with gradient amplitude (G) and duration (T ). From (1.8) we can see that if either gradient am-plitude, G, or gradient duration, T , is reduced kmax will also be reduced. If the size of k-space isreduced, the resolution in the image domain is reduced. Inversely, if the resolution of k-space is10reduced, by lower sampling rate during read out, field of view (FOV) of the image is reduced. Thiswill have practical implications that will become more obvious in the section where we discussdifferent pulse sequences.1.2 The Central Nervous SystemThe central nervous system (CNS) consists of the brain and the spinal cord. The two main tissuetypes in the CNS are gray matter (GM) and white matter (WM). The distribution of GM and WM isvery distinct in both brain and spinal cord, see figure 1.4. In the spinal cord GM is found in thecenter, while in brain, GM is mostly found in the outer layer, the cortex.A generic nerve cell, found in the CNS, can be divided into three main components; the cellbody, the myelinated axon, and the axon terminals, see figure 1.4. There is a clear spatial distri-bution of neurons in the CNS: GM is mainly composed of nerve cell bodies whereas WM mainlyconsists of myelinated axons. The diameter of an axon can vary between 1 and 20 µm and thelength can span from a few millimeters up to one meter. To maintain efficient communication be-tween neurons, myelin is used as an insulator around the axons to speed up transmission of nervesignals. Myelin is a fatty substance which is wrapped around the axon in a spiral. In between thelayers of myelin are compartments of water, which will be referred to as myelin water, see figure1.4. Water outside and inside the cell are named extra and intra cellular water respectively. Inaddition to water, cerebrospinal fluid (CSF) is also present in the central nervous system.The spinal cord plays a crucial part in the CNS as the connection between the brain and thebody. It directs signals from the brain to the body, and receives sensory input which it relays backto the brain. As a structure, the spinal cord can be divided into four main segments: cervical,thoracic, lumbar, and sacral, see figure 1.4. Each segment is further divided into vertebrae whichare labeled with a letter corresponding to the section (Cervical, Thoracic, Lumbar, and Sacral).11The second vertebra is thus referred to as C2 while the 8th vertebra is called T1 as there only are 7cervical vertebrae.One of the most common diseases of the CNS is multiple sclerosis (MS). Patients with MSsuffer from demyelination, a loss of myelin, in the CNS. As the myelin is stripped from the axons,transmission of nerve signals is inhibited and this manifests in varying degrees of physical disabilityand cognitive decline. Myelin also provides trophic support to the axons and demyelination cantherefore also lead to axonal loss which is irreversible. The disease is today diagnosed usingthe McDonald criteria[60]. It is outside the scope of this report to recapture all of these criteria,but a brief overview will prove useful for further discussions. Two concepts are important in thediagnosis of MS: symptoms disseminated in time and symptoms disseminated in space. UsingMRI, it is possible to identify lesions (areas of acute damage) on T1 and T2 weighted scans andsymptoms disseminated in space can therefore be assessed by detecting lesions in the brain orspinal cord. Similarly, symptoms disseminated in time can be established from a follow-up MRI.In addition to lesions in the CNS, patients with MS typically exhibit atrophy (loss of tissue) in thebrain and spinal cord[38].Lesions on MRI are sensitive to demyelination but very unspecific[54]. Quantitative MRI tech-niques can providing quantitative measures of the tissue micro structure that are more specific to aparticular pathology. Myelin water imaging (MWI) is an example of such a technique. This thesiswill center around MWI imaging and the the following section will provide a detailed descriptionof data acquisition and analysis of MWI.1.3 Myelin Water ImagingMyelin water imaging (MWI) is a quantitative MRI technique with the ability to measure myelinin the CNS. Throughout this section I will describe the basic underlying physics, data acquisi-12Figure 1.4: The CNS consists of the brain and spinal cord. The spinal cord is divided in threesegments (A). Histological images of the brain (B) and the spinal cord (C) shows thetypical distribution of white matter (WM) and gray matter (GM) (The brain is samplefrom a Macaque monkey). On the micro structure level of the CNS we find single neu-rons as shown in (D) consisting of a cell body, myelinated axon, and dendrites. If wezoom in on a cross section of an axon we can observe the multi layered structure ofmyelin around the axon (E). Image sources: (A) Ref. [16] (B) Ref. [78] (C) Ref. [56],(D) Ref. [34] (E) Ref. [64]13tion, analysis and applications of MWI. The section will be concluded with a brief overview ofalternatives to MWI with their individual strengths and weaknesses.1.3.1 T2 Relaxation in vivoThe process of quantifying the T2 in tissue can be accomplished with different degree of precisiondepending on the method used. To find the T2 time in a voxel, at least two measurement withdifferent TE are required. From this, the solution to the Bloch equation presented in (1.5) can beapplied as:MT (TE1) =M0 · e−TE1/T2 , MT (TE2) =M0 · e−TE2/T2 (1.9)→MT (TE2)eTE2/T2 =MT (TE1)eTE1/T2 → T2 = TE2−TE1log(MT (TE1)/MT (TE2)) (1.10)It is obvious that this method would be very susceptible to noise as we only fit to two data points,but it is used in some studies[4]. To improve this, multiple data points can be used to find a best fitline[1]. With a set of n data points acquired with different TE the data can be fit to the followingexpressionMT (TEi) =M0e−TEi/T2 i= 1,2...n. (1.11)With this approach, the aim is to find the T2 value that minimizes the residuals in the fitting pro-cedure. However, this assumes that each voxel only contains a single T2 component and, as afore-mentioned, in the CNS we know that there are multiple water compartments exhibiting different T2times. To resolve this, Mackay et al. [50] proposed that the T2 decay from a single voxel obtainedfrom n measurements should be described as a superposition of M T2 components according to thefollowing expressionyi =M∑j=1s j · e−ti/T2, j , i= 1,2, ...,n (1.12)where yi is the signal amplitude at time point ti, and s j is the amplitude of the j:th T2 component.Some readers might recognize the solution of this problem, finding the T2 j and s j components, as14the discrete inverse Laplace transform. This is an ill posed problem, i.e., an unambiguous solutioncannot be readily found. Others might be more likely to see the set of real exponential functions inthe sum in (1.12) as base functions in an expansion yi with s j as expansion coefficients. With thisapproach, we reach the same conclusion since the set of real exponential functions with differentexponent coefficients do not form an orthogonal basis, thus resulting in an ambiguous expansion.A solution to (1.12) yields a spectrum of T2 times for the sample volume. In the context ofMWI, typical T2 times in white matter tissue are: myelin bi-layer water between 10-55 ms, intraand extra cellular water 70-95 ms, and cerebrospinal fluid (CSF) greater than 1000 ms[50]. Asuccessful measurement of a sample volume of WM should thus result in three significant peakslocated at the aforementioned T2 times. Once the T2 distribution has been obtained, the myelinwater fraction (MWF) is calculated as the ratio of the sum of T2 times, s(T2, j), assigned to myelinwater, around 10-50 ms, to the sum over all T2 components. Typical MWF reported in literature isaround 11% in WM human brain and about 22% in spinal cord[57].1.3.2 Acquisition - Pulse Sequences for MWIT2 relaxation data for MWI consists of a 4D data set with n > 32 3D volumes, each acquired at adifferent TE. From this data set, a T2 distribution is obtained from each voxel by solving equation(1.12), using the T2 decay with n data points as input data. In the seminal paper by Mackay et al.[50] a 32 echo spin-echo single slice sequence with an echo spacing of 15 ms, a TR of 3 s, and 180◦refocusing pulses was used. As MRI scanners have developed, so have the sequences for MWI. Thesingle slice technique was followed by a 7-slice sequence allowing almost full cerebral coverage[52]. This was later followed by a 3D gradient- and spin-echo (GRASE) [21] adapted for MWI by[62], allowing full cerebral coverage in less than 15 minutes.This section will describe the properties of pulse sequences currently used for MWI, differencesbetween 2D and 3D acquisitions, variations between scan vendors and MRI safety related to MWI.152D and 3D AcquisitionMRI data can be acquired in both two and three dimensions. Both methods yield similar images butthe acquisition process is different. In 2D MRI, multiple slices are acquired with a thickness corre-sponding to the bandwidth of the RF excitation pulse. It is common to choose the slice encodingdirection along the inferior-superior direction of the patient (z). Phase and frequency encoding isthen used to encode the x and y position in the image. In 3D MRI on the other hand, a slice selectionpulse with large bandwidth is used to excite a thick slab, equivalent to the whole imaging volume.A phase encoding gradient is then applied also in the slice encoding direction, yielding spatial loca-tion of spins in the z direction. The phase encoding gradient in the slice encoding direction decideswhich plane in the now three dimensional k-space will be filled.The time required for a general 2D acquisition can be calculated by [8]T2Dacq = TR ·NEX ·Nphase ·Nacq (1.13)where NEX is the number of image averages and Nacq is the number of acquisitions required toobtain all the slices. Some sequences utilize multiple averages to improve image quality. Thisis achieved by acquiring data with the same parameters multiple times and averaging the results.While it will improve, it is at the expense of increase acquisition time. In 2D MRI it is very commonto acquire interleaved data, i.e., the scanner does not have to wait a full TR before acquiring thenext slice. Instead, a non-adjacent slice is excited and it is assumed that the two slices will notexchange magnetization. If the time to obtain all slices is less than the TR then Nacq = 1.The time required for a 3D acquisition differs by an additional phase encoding term asT3Dacq = TR ·NEX ·Nphase1 ·Nphase2. (1.14)16It is obvious here that as long as Nacq = 1, 3D acquisitions will have a longer duration than 2D bya factor of Nphase2, which is directly related to the resolution in that direction.While adding another dimension will increase acquisition time, it will generally increase thesignal to noise ratio (SNR). The SNR in a given MRI acquisition can be expressed as[8]SNR≈ ∆x∆y∆z√Tacq,total (1.15)where ∆x∆y∆z is the voxel volume and Tacq,total is the total time the receiver coil is acquiring asignal from each voxel. In a 2D acquisition, data from a certain voxel is only acquired when thecorresponding slice is selected. Thus, Tacq,total =Nphase1 ·NEX ·Tacq where Tacq is the readout time.In a 3D acquisition however, data is acquired from the whole volume at once and we instead readTacq,total = Nphase1 ·Nphase2 ·NEX ·Tacq. From this we can find a rough estimate of the SNR gainfrom a 3D sequence asSNR3DSNR2D=√Nphase2. (1.16)However, it should be noted here that this is a rough estimate and interleaved 2D sequences canemploy longer TR time to weigh up for the loss in SNR.Multi-Echo Spin EchoThe most common and oldest pulse sequence for MWI is the Carr Purcell Meiboom Gill (CPMG)sequence consisting of a 90◦ excitation pulse followed by a number of 180◦ refocusing pulses. Mostimplementations of the CPMG sequence for MWI use composite refocusing pulses where a block ofthree RF pulses (90◦ y-180◦ x-90◦ y) replace the single 180◦ . However, some implementations ofmulti-echo spin echo sequences for MWI use a refocusing flip angle lower than 180◦ . Thus, theyare not a pure CPMG sequence anymore. To avoid confusion, any implementation of the multi-echospin echo sequence will hereafter be referred to as a multi echo spin-echo (MSE) sequence and,unless stated otherwise, the flip angle is 180◦ .17The MSE sequence can be found in numerous studies in MWI[42, 43, 46, 51]. It typicallyconsists of a 90◦ excitation pulse followed by 32 spin echoes generated by 180◦ refocusing pulses.The first 180◦ pulse is referred to as the flip angle, α , while the 31 subsequent RF pulses are referredto as refocusing control flip angle αre f con. While αre f con can be adjusted and lowered down to aslow as 135◦ , α is kept at 180◦ . The echo spacing is set to 10 ms and TR on the order of 1000-1500s.In contrast to conventional MRI sequences, in MSE 32 imaging volumes are acquired, one foreach echo. This implies that 32 different k-spaces are filled during the acquisition. At each spinecho, one data point in the decay curve is sampled, corresponding to one line in the correspondingk-space.GRASE - Gradient and Spin EchoAn accelerated option to the MSE sequence is gradient- and spin-echo (GRASE). The most signifi-cant difference between MSE and GRASE is that for each spin echo, three k-space lines are acquiredby using frequency encoding gradients with shifting polarity, similar to a short echo planar imag-ing (EPI) readout. A pulse sequence diagram showing the Philips implementation of a 3D GRASEsequence shown in figure 1.5. This shows how each spin echo is accompanied by three frequencyencoding lobes for readout. A 3D version GRASE is the main sequence for MWI in this thesis anda detailed study of the pulse sequence diagram in figure 1.5 is beneficial.The current GRASE protocol for MWI was first established by Ma¨dler and MacKay [53] andlater further developed and accelerated by Prasloski et al. [62]. The GRASE sequence is initiated bya 90◦ excitation pulse followed by a train of 180◦ pulses. With each 180◦ RF pulse, a slice encoding(second phase encoding) gradient is applied, which defines which plane in the 3D k-space will beoccupied during readout. This is shown in figure 1.5 where the second phase encoding gradienthas been separated from the main gradient lobe for clarity. However, in real implementations ofthe sequence they will be merged together as it gives the same result. In between each spin echo,184 phase encoding lobes and 3 frequency encoding lobes are applied. The short gradient lobes usedfor phase encoding are commonly referred to as blips. Before the first frequency encoding gradient,a negative gradient lobe translates to a position along the negative ky axis in k-space. Readout isperformed through k-space, a positive blip moves the next readout to a kx line higher up in k-spaceand so forth. The same procedure is now performed for every subsequent spin echo. For each spinecho, 3 k-space lines are acquired for the k-space corresponding to that echo time, resulting in anacceleration by a factor of 3 compared to the MSE sequence.The second gradient echo readout is the only echo with true T2 weighting, as it coincides withthe spin echo. It is therefore necessary that the second gradient echo corresponds to a k-space lineclose to the center as this will give the main contrast in the image. The first and third gradient echowill be affected by T2∗ contrast as readout is performed before and after the spin echo. Therefore,the k-space ordering in the GRASE sequence is set up to encode the center of k-space with thesecond gradient echo, and the upper and lower part of k-space along the ky axis with the first andthird gradient echo. This way, the T2∗ weighting can be considered to be negligible since centerof k-space, which decides the main contrast in the image, is acquired by the second gradient echowhich produce true T2 contrast.Vendor Specific Pulse SequencesPulse sequence names used in this thesis are typical of Philips scanners as that is the main scannervendor used in this project. Not all scanners have standard sequences readily available for collect-ing multi-echo data as used in the aforementioned sequences. Pulse programming may requiredto produce the desired sequence in some cases. This leads to potential differences in implementa-tion between sites and scanner manufacturers. Chapter 3 of this thesis is dedicated to evaluatingmulti-center and multi-vendor reproducibility of the GRASE sequence.19Figure 1.5: Pulse sequence diagram of the 3D GRASE sequence as visualized by the Philips scanner simulator softwareenvironment. For each spin echo, which occurs in between the 180◦ RF pulses, there are 3 gradient lobes for readout,as indicated by the three red lines at the bottom. Positive gradients move in the positive direction of k-space, andnegative gradients in the negative direction. When a 180◦ RF pulse is played out, the position in k-space is flipped,equivalent to multiplying each coordinate by -1. kx translates to −kx etc.20MRI SafetyMRI is an inherently safe medical imaging modality as long as pulse sequences adhere to safetystandards. There are two factors that are considered here: specific absorption rate (SAR) andperipheral nerve stimulation (PNS). SAR is the energy deposited to the subject due to the RF pulses,which is proportional to the time integral of the squared RF pulse amplitude. Higher flip anglerequires stronger RF pulses, yielding higher SAR. It can therefore sometimes be required to reducethe flip angle to reduce SAR to comply with safety standards. In MWI this is accomplished bylowering αre f con. The second factor is PNS, which is caused by rapidly varying magnetic fields andtherefore directly related to the gradient slew-rate. The GRASE sequence utilizes EPI readout foreach spin echo which utilize maximum gradient slew-rates and may thus produce PNS. Both SARand gradient slew-rate are controlled by the scanner and there is no risk of exceeding the safetylimits. However, in special cases such as patients with MRI safe metallic implants, lower SAR andgradient slew-rate may be required. To reduce SAR, the refocusing flip angle can be reduced andrepetition time (TR) increased. To reduce the gradient slew-rate, the bandwidth is typically reducedwhich results in lower amplitude of the gradients and thus a smaller difference when the gradientsswitch polarity. Lower bandwidth will result in larger voxel size, i.e., reduced spatial resolution.31.3.3 Data AnalysisA desirable method for solving (1.12) should make as few initial assumptions about the problemas possible to enable studies of unknown systems. With this in mind, Mackay et al. [50] usedthe method of non-negative least squares (NNLS), similar to the least squares method with theadditional constraint of non-negative components, to decompose the decay data. The end result ofa NNLS analysis is a T2 distribution corresponding to the water compartments present in the samplevolume.3The gold standard in MRI safety is found at http://www.mrisafety.com.21(a) (b)Figure 1.6: Example of single exponential synthetic T2 decay curves generated using the EPGalgorithm. In (a) α is varied between 60-180◦ and αre f con = 180◦ . In (b) T2 is variedbetween 20-200 ms.The algorithm used in this work is based on the NNLS analysis described by Whittall andMacKay [86] which later was further developed by Prasloski et al. [61] to include stimulated echocorrection using the extended phase graph (EPG) algorithm[32]. The main steps of the NNLS anal-ysis are as follows:1. A synthetic set of decay curves is generated using the EPG algorithm with a predefined set ofnα = 8 flip angles ranging from 50-180◦ and a set of nT2 = 40 T2 times ranging from T2minto T2max ms. This yields a set of nα · nT2 decay curves ˜yi, j with i denoting the flip angle αand j the T2. Figure 1.6 show examples of single exponential decay curves with different T2and α .2. For each voxel in the multi-echo data set, a decay curve yˆ is extracted. A NNLS fit, X , is thenevaluated for each flip angle fromXα =argmin(Xα )||yˆ−Cα ·Xα ||22 (1.17)22where column j in Cα is the synthetic decay curve for T2 time j and flip angle α generatedin the previous step. For each Xα , χ2α is calculated asχ2α =∑(yˆ−Cα ·Xα)2. (1.18)3. The χ2α is a measure of the goodness of fit for each flip angle. To find the optimal flip angle,the smallest χ2α has to be found. With only 8 flip angles used in the analysis, it is improbablethat any of these would provide the true minimal χ2α . Therefore, a spline interpolation of allχ2α is calculated and the minimum of this curve is extracted. The flip angle that minimizesthe spline interpolation is set as the optimal flip angle, αcalc.4. A new set of synthetic decay curves are generated using the EPG algorithm with the new flipangle αcalc and the same range of T2 times as in step 1. This basis setCcalc will be used to fitto the data set and find the T2 distribution.5. The new set of decay curves are now fit to the original decay data yˆ. The optimal set of T2components s j is found by minimizing the regularized NNLS equation expressed as[86]||Ccalcs j− yi||22+µ||Hs j− f ||22 (1.19)where µ is a trade-off parameter determining how much the solution is affected by the con-straints prescribed by H. In this application of the NNLS algorithm, H is the identity matrixand f = 0. The solution is limited to a certain range commonly set to χ2 < 1.02χ2min whereχ2min is the χ2 calculated for a regularized solution. The NNLS algorithm will find the largestvalue µ that keeps χ2 within the desired limits.6. The final T2 distribution in a given voxel is now s j. The T2 distribution is now divided intotwo parts: one attributed to myelin water, typically set to T2< 40 ms, and one part attributed23to the intra-extracellular, set to 40 ms < T2 < 200 ms. The part of the distribution associatedwith each component is hereafter referred to as T2dist(sf) and T2dist(mf) (sf=short fractionfor myelin, mf=middle fraction intra-extra cellular).7. The last step calculates the output parameters for each voxel. This includes• The myelin water fraction (MWF)MWF =∑T2dist(sf)∑T2dist(1.20)• The geometric mean T2 (GMT2) of the intra-extracellular T2 peakGMT2 =exp(T2dist(mf) · log(T2times(mf)))∑T2dist(mf)(1.21)• True flip anglealpha = αcalc (1.22)The current MATLAB implementation of the algorithm utilizes multicore functionality andtakes about two hours to run on a modern workstation PC, although this is of course highly depen-dent on image resolution and size of the region of interest. For smaller structures like the spinalcord, the volume should be masked out for significantly shorter computation time.Recently, an accelerated version of the analysis algorithm designed to utilize the multicorefeatures of graphical processing units was developed by Yoo and Tam [90]. The algorithm outlinedabove regularizes the T2 distribution in each voxel through the χ2 fit as shown in (1.19). Yoo andTam took this one step further by including non-local regularization of the T2 distribution.241.3.4 Applications of MWIThere is a wide range of applications for MWI including: traumatic brain injury[87], age-relatedmicro-structural changes in the brain[10], schizophrenia[24] and multiple sclerosis (MS) [58, 81,82]. Given the high prevalence of MS4 it is of high interest to investigate the possibilities of usingMWI for more detailed studies of this neurodegenerative disease. MRI has become an integral part indiagnosing and tracking disease progression MS, but as mentioned in section 1.2, MRI is only usedfor qualitative assessment in clinic. MWI has the ability to take this to the next level by quantifyingmyelination and demyelination in the CNS, thus providing a potential in vivo biomarker for recoveryand progression in patients with MS. This has many benefits over traditional anatomical imagingbut since data acquisition and analysis is complex, to date it is exclusively used in research settings.Laule et al.[42] scanned brain samples from 13 patients with verified MS to compare the MWIresults with optical measurements of luxol fast blue (LFB) stained samples. When brain samplesare stained with LFB, myelin appears blue. The amount of myelin in a given sample volume isthen easily quantified by measuring the optical density, or absorbency, using a back-lit scanner.The results of the study showed a strong correlation between MWI and optical density, averageR2 = 0.67 with SD= 0.13. An example of their analysis is showed in figure 1.7. T2 relaxation datawas acquired with a 32 spin echo CPMG measurement using composite pulses with TE=10 ms andTR=3000 ms, similar to the experiment by Mackay et al. [50]. The results from Laule et al. [42]became a landmark in the development of MWI and proved the potential of using the MWF as an invivo biomarker for myelin in the CNS.MWI has gained ground in research in MS in the last decade. Laule et al. [40] studied 33 MSpatients and 18 healthy controls and found a -16.6% reduction in MWF in normal appearing WMin MS compared to controls. Similar observations have been made in spinal cord MWI. In a studyby Laule et al. [44], MWF was found to be lower in patients with primary progressive multiple4in 2011 an estimated 100,000 Canadians were living with MS[72]25Figure 1.7: Histological verification of myelin water imaging (MWI) adapted with permissionfrom Laule et al. [42]. (a) First echo from multi echo sequence at TE=10 ms. (b) MWIclearly showing higher MWF in white matter in the brain which is well correlated withluxol fast blue (LFB) staining for myelin in (c). Note the clear correspondence betweenthe bright pixels in (b) showing high MWF and dark blue pixels in (c) showing a highmyelin content from staining.sclerosis (PPMS) compared to healthy controls. However, the change was not significant whichmotivates additional work to improve the sensitivity of spinal cord MWI. They also measured asignificant decrease in MWF over 2 years in the 24 PPMS patients enrolled in the study (-10.5%,p = 0.01). It is here important to note that the MWF in the spinal cord is significantly higher thanin the brain, see 1.1 for a comparison between previous studies..1.3.5 Alternatives to MWIMyelin water imaging using multi-echo acquisitions is not the only method for studying myelinin the CNS. The following section will introduce some common alternatives to MWI with theirrespective strengths and weaknesses.26Location MWF (SD) [%] Protocol StudyCervical cord WM 30 (4) 3D MSE MacMillan et al. [51]Cervical cord GM 5 (3) 3D MSE MacMillan et al. [51]Cervical cord WM & GM 22 (2) Single Slice CPMG Minty et al. [57]Brain Minor Forceps 7 (1) MSE Prasloski et al. [62]Brain Minor Forceps 8 (0.5) GRASE Prasloski et al. [62]Brain Internal Capsule 18 (1) MSE Prasloski et al. [62]Brain Internal Capsule 17 (1) GRASE Prasloski et al. [62]Table 1.1: Compilation of myelin water fraction (MWF) from various studies in human brainand spinal cord.mcDESPOTThe closest alternative to MWI, with a similar goal of obtaining a measure of myelin in the CNS, ismulticomponent driven equilibrium single pulse observation of T1/T2 (MCDESPOT)[20]. MCDESPOTquantifies multiple T1 and T2 components in tissue which can be used to obtain the fraction of sig-nal associated with myelin water ( fm), the MCDESPOT approximation of the MWF. While the endresult carries great resemblance to MWI, data acquisition is fundamentally different. Instead ofacquiring multiple echoes at varying echo times, MCDESPOT utilizes a combination of spoiled gra-dient echoes and balanced steady-state free precession with varying flip angle to acquire MRI data.In addition to myelin estimates, MCDESPOT can also produce a T1 map over the volume.In a study by Kolind and Deoni [39], MCDESPOT was used to assess myelin content in thecervical cord, C1-C7. Data was acquired from seven healthy volunteers with an in-plane resolutionof 1x1x1.5 mm3 and a total acquisition time of 26 min. To assess reproducibility, five out of theseven volunteers were scanned again after 4-10 weeks. Average fm in the full cord was found to be20.5% (SD=0.6%), and average fm in GM to be 10.9% (SD=1.2%). The most apparent differenceto previous studies using MWI in the cord is the higher MWF values in GM, compared to MacMillanet al. in table 1.1. The reason for this discrepancy is currently not known but one explanation couldbe that the MSE sequence is known to underestimate the MWF in regions with low SNR, such as the27gray matter in the spinal cord[11]. It could also be due to the fact that the study by MacMillan et al.[51] used a significantly smaller GM region of interest (ROI) than Kolind and Deoni [39] whichwould reduce partial volume effects. Alternatively, there may be magnetization transfer effectsinfluencing the MCDESPOT results[20]. Another interesting result from Kolind and Deoni [39] isthe small standard deviation within each ROI compared to previous studies, see table 1.1. Thisis likely due to the much larger imaging volume. Kolind and Deoni collected data from the fullcervical cord while multi echo studies commonly use single or multi slice sequences covering twovertebral segments in the cervical cord.Another recent study by Kolind et al. [38] investigated the efficacy of using MCDESPOT as abiomarker for progression in patients with PPMS. 15 patients diagnosed with PPMS and 11 healthycontrols received brain and spinal cord MRI with the MCDESPOT protocol. A significant differencein fm in the global normal appearing WM was found between controls and PPMS patients (-11%,p=0.01). The same was found in the spinal cord where fm was reduced by 19% for PPMS patientscompared to controls. fm in brain was significantly correlated with the time to complete a 9 holepeg test, a standard test to asses cognition, while fm in spinal cord was significantly correlated tothe timed 25 foot walk[66], a standard test used to assess physical disability.In conclusion, MCDESPOT provides insight into the tissue micro structure with the same end-goal as MWI but with a different data acquisition approach. However, MCDESPOT still lacks his-tological verification, such as that done by Laule et al. [42] for MWI. Studies have also shownthat magnetization transfer effects can significantly affect the fm[93]. A practical limitation of theMCDESPOT is the requirement for a rather large field of view. While this might not be an issuein brain where whole brain images usually are required, in spinal cord imaging this can increaseacquisition time. While MSE and GRASE allow for a small field of view (FOV), only including a fewvertebral segments, MCDESPOT have currently only been implemented with a large FOV coveringthe whole cervical cord. If the whole cervical cord is of interest, MCDESPOT is currently the fastest28Figure 1.8: Comparison of myelin estimates obtained with GRASE and MCDESPOT. It isclear that the techniques produce significantly different estimates of myelin in the CNS.MCDESPOT estimates a more or less homogeneous distribution of myelin in WM com-pared to GRASE which shows a wide range of MWF values in WM. Figure adapted withpermissions from the author from Zhang et al. [94]approach, but if only a small segment is required, MSE or GRASE is the faster option. The dataanalysis in MCDESPOT is also more involved and computationally expensive than MWI with a widerange of parameters that can be adjusted[19].Recent studies have compared myelin measures between T2 relaxation MWI and MCDESPOTand found the results to be significantly different[94], see figure 1.8. Myelin estimates fromMCDESPOT are significantly higher than those obtained from GRASE (0.3 compared to 0.1). Aninteresting aspect of this is that MCDESPOT studies of the spinal cord report fm values compara-ble to those in the brain, while MWI using GRASE produces almost twice as high MWF values inthe spinal cord compared to the brain. It is still not fully understood what biophysical property iscausing this discrepancy in myelin measures between the two techniques.Diffusion MRIA different approach for studying the CNS is diffusion weighted MRI (DWI). It is outside the scopeof this thesis to provide a thorough introduction to DWI and readers are referred to [2, 3, 5] foran introduction. In essence, DWI measures diffusion of water in the CNS by applying diffusion29weighting gradients in a wide range of directions. The signal attenuation along the various gradientdirections is a measure of the diffusion along that certain direction. Diffusion tensor imaging (DTI)is the most common post-processing analysis of DWI data. In DTI, a diffusion tensor is calculatedfor each voxel which represents water diffusion inside that voxel. The tensor is characterized by3 eigenvectors (µ1,µ2,µ3) and corresponding eigenvalues (λ1,λ2,λ3). Eigenvalues are ordered byascending magnitude. A spherical tensor is characteristic of isotropic diffusion, i.e., diffusion isequally probable in every direction, λ1 = λ2 = λ3. An elongated tensor is by the same reasoningcharacteristic of anisotropic diffusion, i.e., diffusion is more probable along a certain direction. Tocharacterize the shape of the diffusion tensor, the following parameters are commonly used:FA =√32√(λ1− λˆ )2+(λ2− λˆ )2+(λ3− λˆ )2√λ 21 +λ 22 +λ 23, λˆ = (λ1+λ2+λ3)/3, (1.23)MD =λ1+λ2+λ33, (1.24)AD = λ‖ = λ1, (1.25)RD = λ⊥ =λ2+λ32. (1.26)DWI is today a standard sequence available on every scanner, and because of this DTI hasgained a lot of traction in various fields, including studies of myelin. However, with the increaseduse of DTI, the association between DTI metrics, as described in (1.23)-(1.26), and pathology isthe subject of a lot of controversy. Song et al. [70] found an increase in λ⊥ in shiverer mice5compared to controls, which was suggested to be a reflection of the loss of myelin allowing agreater degree of freedom for water diffusion. In line with this argument, λ‖, which is believed tobe a marker for axonal integrity, was unchanged between the two groups. To further support this,another study by Song et al. [71] performed DTI in a mouse model with retinal ischemia. After 35Shiverer mice has a genetic modification causing a hypomyelinated CNS resulting in a characteristic shivering gait.30days, a significant loss of axons, as verified by histology, was correlated with decreased λ‖, whileλ⊥ remained constant. After 5 days, a loss of myelin as verified by histology was correlated withan increase in λ⊥. Their studies quickly gained traction in the wider scientific community and havesince been used as a reference for correlations between diffusion metrics and pathology in the CNS.Some of the controversy regarding DTI metrics have been well summarized by Wheeler-Kingshottet al. [85]. In this article, Wheeler-Kingshott highlights the issues of using radial diffusivity (RD)and axial diffusivity (AD) in voxels with crossing fibers. To support their statement, they performedcomputer simulations where two perpendicular diffusion tensors were used to represent two cross-ing fibers. Each of these tensors was then adjusted to reflect a biophysical change. Demyelinationin one of the fibers was simulated by increasing the radial component for one of the tensors. Axonalloss in one of the fibers was simulated by decreasing the axial component of one the tensors, allin line with the results from Song et al. The artificial data set was then analyzed using a standardsingle tensor DTI procedure. When demyelination was simulated, both AD and RD increased whilefractional anisotropy (FA) decreased. The simulated axonal loss manifested as decreased AD andRD, and decreased FA. Although simplistic, these simulations highlight the issue with using RDand AD in tissue with complex microstructure.A variety of advanced diffusion models have been designed to resolve some of the issues pre-sented above. One model that has gained traction over the last years is neurite orientation dispersionand density imaging (NODDI)[92]. In contrast to DTI which does not make any assumptions aboutunderlying tissue structure, NODDI is a model-based approach assuming three diffusion compo-nents: isotropic diffusion CSF, hindered diffusion in extra cellular water, and restricted diffusionin intra cellular water. Another significant difference is that up to 100 diffusion directions areacquired, yielding longer acquisition times. The end result of the analysis consits of three compo-nents: intra-cellular volume fraction vic, orientation dispersion index (ODI), and isotropic volumefraction viso. NODDI does not make the assumption that each voxel only contains one tensor but in-31stead each voxel is attributed an orientation dispersion as an indication of the fiber coherence withinthe voxel. Studies using NODDI have shown that multiple combinations of ODI and vic yields thesame FA [28, 92]. This lends further support to the notion that regular DTI parameters do not tellthe full story.Magnetization TransferAs established in section 1.2, myelin is a layered structure that surrounds axons. Water within themyelin layers, the myelin water, exhibit a short T2 due to spin-spin interactions with the neighbor-ing protons. Magnetization transfer imaging (MT) aims to quantify the exchange of magnetizationthat is occurring between the macromolecule protons and the aqueous protons, henceforth referredto as the solid pool and water pool. The line width, the range of Larmor frequencies, for the macro-molecule bound protons is far wider than that of the aqueous protons and it is this feature that MTutilizes[31]. By applying a RF pulse with a frequency off-resonance for the aqueous protons, alarge quantity of the macro molecule protons will be excited. Momentarily after the RF irradiation,exchange of magnetization is taking place between the two pools. This exchange will influence thelongitudinal magnetization component as it is related to spin-lattice interactions, see section 1.1.2.MT data is therefore collected by acquiring two sets of data, first with an off-resonance pulse to ex-cite the solid pool, and then one without. From these two acquisitions, the magnetization transferratio (MTR) is then calculated asMTR =(1−MSM0)(1.27)where MS is the measured magnetization with the saturation pre-pulse and M0 is the measuredmagnetization without the saturation pulse.MT has been used extensively for quantifying myelin. In a study by Schmierer et al. [67] MTRwas compared with myelin staining in post mortem MS brain. MT data was acquired with a dualspin-echo sequence, with and without a 1 kHz off resonance saturation pre-pulse. Myelin was32measured in histology using LFB staining. Linear regression between MTR and transmittance fromLFB stained samples showed a correlation with r = −0.84, p = 0.001. Using LFB, areas with lowmyelin content will have a high transmittance, thus, the linear relation shown here is the oppositefor what would be expected for myelin content. This is in line with the expected results from thetheory presented above. Higher myelin content means a higher concentration of the solid poolwhich can be saturated, and thereafter exchange magnetization with the liquid pool, generating ahigh MTR.In a study by Vavasour et al. [79] MTR was compared to MWF in a cohort of 19 subjects (10healthy controls and 9 MS patients). MTR data was acquired using a spin-echo sequence with andwithout a 2 kHz off-resonance pulse. MWF data was acquired using a single slice 32 echo MSEsequence with TE=10 ms, and TR=3 s. As expected, MWF was lower in MS patients comparedto controls, most significantly in white matter. In line with the results presented by Schmiereret al. [67], MTR was found to be lower in MS patients. Linear correlation between MTR and MWFrevealed a correlation of r = 0.5, p= 0.005.However, just as DWI metrics can loose their biophysical interpretation in tissue with crossingfibers, MTR can be affected in a similar way. This was concluded in a study by Vavasour et al. [82]which showed a significant correlation between water content and MTR r = −0.65, p < 0.0001.Inflammation or edema could affect both MWF (which is a fraction, dependent on the total watercontent) and MTR (which is dependent on bulk water content). However, in a study of a pre-clinical model of MS, it was shown that MTR is more sensitive to the physiological changes tomyelin induced by inflammation, while MWF was a more specific indicator of myelin content intissue[27]. Another issue with MTR is the fact that it is not inherently specific to myelin but rathermacromolecules in general and it is therefore possible to obtain a signal from MTR in a region withno myelin but other macromolecules.33Reflections on quantitative MRIAll quantitative MRI techniques have strengths and limitations. Here, I have given a brief overviewof a few quantitative MRI techniques all with the same goal of measuring myelin in the brain andspinal cord. With widely disparate techniques they all expose some biophysical property of the tis-sue microstructure in the CNS. When results from two techniques, attempting to measure the samething, obtained different results one could either claim one technique is superior to another, but itwould rather be better to try and understand which biophysical property the two techniques observedifferently. This is for example seen when myelin estimates obtained with MWI and MCDESPOTare compared. It is not yet fully understood in which situations and why these techniques agreeand sometimes disagree. There have been many comparative studies between quantitative MRItechniques but more work is needed to fully resolve this question[79, 94].1.4 Spinal Cord MRIDespite its significance in the CNS, the spinal cord receives less attention in MRI than the brain.6Of course, many of the diseases of the CNS that are studied occur in the brain. However, there isa consensus in the scientific community that the field of spinal cord MRI is hampered by technicalissues in data acquisition and analysis[74]. The main issues in spinal cord MRI are consideredto be the inhomogeneous magnetic field created by the inhomogeneous environment in the spine,the small cross sectional area of the cord and physiological motion[74]. Due to these issues, anapparent trend in the literature is that new MRI techniques are first developed for the brain, andthen optimized for the spinal cord. In chapter 2 I will present the first implementation of a 3DGRASE sequence for MWI of the spinal cord. This section aims to provide an introduction to spinalcord MRI in terms of the anatomy and physiology of the cord as well as typical issues associatedwith spinal cord MRI.6A search on pubmed.org for spinal cord & MRI yields about 1000 results while brain & MRI yields close to 9000results. Search performed on March 6th 2016.341.4.1 Studying Spinal Cord Pathology Using MRIAt first glance the spinal cord might appear simply as a bundle of nerve fibers stretching downthe spine, but there is much more to the story. Along its long axis, the spinal cord varies in crosssectional area (CSA) and myelin content. This was shown by Kolind and Deoni [39] who measuredthe fm using MCDESPOT along the cervical cord. The results of their study, see figure 1.9a, show adifference of almost 2% between the maximum fm at the C2 level to the minimum at C6. Similarly,the CSA varies along the cord with a maximum area around 90 mm2 at C4 level and minimum of< 40 mm2 below T3[25].Neuropathological studies of the spinal cord will typically strive to measure: tissue microstruc-ture using either DWI or some form of myelin imaging, CSA as a measure of atrophy, or lesions.MWI in human spinal cord has been proven to be highly useful in studies of MS. Laule et al. [45]found a significant decrease in MWF in patients with PPMS over two years (-10.5%, p = 0.01),while no change was found in the group of matched controls. Kolind et al. [38] found a 19% de-crease in MWF in cervical cord between controls and PPMS patients using MCDESPOT. Atrophyin the spinal cord has been shown to correlate strongly with disability in patients with MS[49, 73].The CSA can be as much as 10 mm2 smaller in patients with a progressive stage of MS comparedto controls. Recent studies have shown that normalization factors such as age, sex and even spinalcord length might be necessary when CSA is used as a measure for disease progression[47, 59].Most MRI techniques that have been used to study neuropathology in the brain using MRI havealso been implemented in the spinal cord for the same purpose. This includes many techniquesused to study WM such as: MT, DTI, MCDESPOT and MWI.1.4.2 Motion and Flow Artifacts in MRIThe origin of motion artifacts in spinal cord MRI has been studied and many special techniqueshave been developed to overcome these problems [22, 23, 91]. It is known that the position and35(a)Figure 1.9: Variation of fm along the cord obtained with MCDESPOT by Kolind et al. [39].Adapted from Ref. [39] with permissions from the author.velocity of the spinal cord is strongly correlated to the cardiac cycle. The pulsating flow of CSF thatruns through the subarachnoid cavity of the spinal cord is also connected to the cardiac cycle. It isof great interest to suppress the signal from the pulsating CSF as it will introduce motion artifactsin the image. The remainder of this section will be dedicated to discussing typical imaging artifactspresent in spinal cord MRI and techniques used to reduce them.The time it takes for readout in the frequency encoding direction can be considered fast enoughto image the instantaneous position of the object and thus not influenced by motion. However,significant motion can take place between phase encoding steps. If the object moves between thephase encoding steps, artifacts will appear in the phase encoding direction, independent of thephysical direction of the movement. If the movement is periodic, such as the beating heart, ghostscan appear in the image. This can be explained by the Fourier relation shown in figure 1.10a. Whenthe object is static, the Fourier transform and inverse Fourier transform, corresponding to imagereadout and reconstruction will produce the correct image of the object. If the object is moving,the Fourier transform will be modulated with a function corresponding to the frequency of theoscillation. This will produce replicas of the object in the reconstructed image where spacing of36the replicas, or ghosts, will be inversely proportional to the frequency of the motion in the phaseencoding direction.So far, we have discussed motion artifacts in terms of body motion, but the same idea appliesfor fluid flow within the object. Ghosts appearing from flow are usually seen in large vessels, suchas the aorta, or from CSF. An example of flow artifacts can be seen in figure 1.10b where thepulsating CSF in the spinal canal produce a series of ghosts up and down in the image. There aregenerally two types of flow to consider depending on the slice orientation: in plane and throughplane flow. In spinal cord MRI, in plane flow is commonly seen in a sagittal slice where the fulllength of the spinal cord is imaged while through plane is most prevalent in axial slices with CSFflowing in and out of the slice. The two types of flow will appear differently in the final image andrequire different correction methods which will be discussed later.(a) (b)Figure 1.10: (a) Diagram of the Fourier relations behind the ghosting motion artifact. Whenthe object is static, the object is correctly reconstructed by the inverse Fourier trans-form. When the object is moving, the data in the Fourier plane can be described as thedata from the static object multiplied by an oscillation with the modulation frequency.The reconstructed object will then appear in several copies. Image adapted with per-mission from: Common Artifacts in MRI, by Qing-San Xiang, PHYS542, UBC, 2014.(b) Example of flow artifact from CSF in the spinal cord.371.4.3 Flow and Motion is Correlated to the Cardiac CycleDuring the cardiac cycle, the heart will contract (systole) and relax (diastole). During systole theblood pressure will increase as oxygenated blood is flowing out of the heart, through the arteriesand out into the body. This process will influence the position of the spinal cord. Vavasour et al.[83] found an average displacement of the spinal cord in the caudal cranial (C/C) direction of 5mm, taking place at 50% of the cardiac cycle which is the time point when the heart is maximallycontracted, between systole and diastole. Figley et al. [23] also studied this correlation betweenspinal cord motion and the cardiac cycle and found movement of the spinal cord in the anteriorposterior (A/P) direction, with a peak average displacement of about 0.25 mm in the cervical andupper thoracic region, at 50% through the cardiac cycle. Displacement in the left right directionwas also studied but was found to be much smaller than the A/P direction.It is known that the flow velocity of CSF is correlated to the cardiac cycle. It flows caudallyduring systole and cranially during diastole. The CSF that surrounds the brain is pushed down thespinal column as the blood volume in the craniospinal cavity increases[33]. The flow velocity ofCSF is thus correlated to the cardiac cycle. This was shown by Vavasour et al. [83] who reportedpeak flow velocities around 2 cm/s at 50% through the cardiac cycle. Their results also verify thetheory of positive flow velocity during the first half of the cardiac cycle and negative in the secondhalf.1.4.4 Motion Correction TechniquesMotion correction techniques can be divided into two categories: motion tracking and non-motiontracking techniques. The non-motion tracking techniques include: averaging of several scans, sat-uration bands to suppress moving spins, gradient moment nulling and special k-space trajectories.Motion tracking techniques on the other hand include data collection of physiological data such38as pulse and respiration to trigger or gate the image acquisition. This section will outline the ideabehind cardiac triggering, spatial saturation bands, and gradient moment nulling.Cardiac triggeringTo correct for spinal cord motion and make the spinal cord appear to be in a fixed position, datamust be acquired at the same phase of the periodic motion. When the motion is related to the heart,the process is called cardiac triggering. The idea is simple: acquire all lines in k-space during thesame phase of the cardiac cycle, i.e., when the position of the spinal cord is the same. This willmake the spinal cord appear static in the images, and if done properly, produces an image free ofartifacts.The two most common techniques for cardiac triggering are using a peripheral pulse unit (PPU)or echocardiography (ECG). By synchronizing the signal from the trigger unit, the timing of imageacquisition can be adjusted to the same point of the cardiac cycle. While it definitely is preferredto perform cardiac triggering in spinal cord MRI, it comes with a few drawbacks. First, since dataacquisition now is linked to the heart rate, a fixed heart rate is assumed when calculating the scantime. If the heart rate changes during acquisition this can prolong the scan. Second, there is a delayin the pulse signal detected by the PPU and ECG compared to the pulse of the heart. While it canbe assumed to be a constant delay, it has to be carefully set up during protocol development.The biggest consideration for cardiac triggering is scan time. In the example of MWI, eachecho train is typically 320 ms, and the TR is on the order of 1500 ms. If we consider a standardpatient with a pulse of 60 beats per minute (BPM), the time available between the R-wave of twoconsecutive heartbeats (the RR-interval) is 1 s. In this case, 2 RR would be required for each echotrain as 1RR< TR. This will effectively make the time between each echo train 2 s, increasing thetotal time by 25%. If the heart rate would decrease, 2 RR would still be required, but the TR wouldincreasing. In scans with shorter acquisition time, many interleaved slices can be acquired withina single RR and the additional time for cardiac triggering is not that significant.39Spatial Saturation BandsSpatial saturation bands can be used to either completely null the signal from a moving structurewithin a slice, such as the motion of swallowing in a sagittal slice of the spinal cord, or to null inand outflow of spins through a slice [8]. To null the signal from spins implies that the transversemagnetization is destroyed and the signal from these spins will not be detected. In spinal cordimaging, saturation bands are most commonly used to suppress the swallowing motion. Flowsuppression using saturation bands is only needed for high flow velocities such as in the aorta.The flow of CSF is slow enough to not cause significant ghosting effects and a more simple flowsuppression method called gradient moment nulling is used instead.To null the signal from spins at a certain location within the slice, a 90◦ spatially selective RFpulse is applied to the area that needs to be suppressed, see pulse sequence in figure 1.11. Instead ofrefocusing the excited slice, strong gradients in both the frequency and phase encoding directionsare applied. This will dephase all spins in the volume making the total transverse magnetizationzero; the spins become what we call saturated.Spatial saturation pulses are used in conjunction with many pulse sequences to minimize theeffect of motion. However, it is not recommended to use in MWI since the saturation bands poten-tially could cause a magnetization transfer effect and affect the MWF.Gradient Moment NullingGradient moment nulling (GMN) is used to reduce motion artifacts from flowing spins in fluidsuch as CSF. To do this one has to ensure that the moving spins have not acquired any additionalphase at the end of the gradient, in the case of slice and phase encoding, or at the time of readoutfor frequency encoding. If there is a wide spread in phase among the spins, the total transversemagnetization will be reduced, leading to a lower signal. To describe this phenomenon we describe40Figure 1.11: Example of pulse sequence using a spatial saturation pulse to null the signalfrom inflowing spins. The first 90◦ pulse excites a slice covering the area that needs tobe suppressed, this is followed by strong crusher gradients in the frequency and phaseencoding direction to destroy any magnetization from the slice. After this, a regularpulse sequence can be performed.the position of a given spin with the initial position x0, speed v and acceleration a to be expressedbyx(t) = x0+ v0t+12at2. (1.28)The phase accumulated by this spin under the influence of a gradient G(t) is then given byφ(t) = γ∫ t0G(τ)x(τ)dτ = γ∫ t0G(τ)(x0+ v0τ+12aτ2)dτ. (1.29)The n:th gradient moment mn(t), which the reader might recognize from statistics, is then definedbymn(t) =∫ t0G(τ)τndτ. (1.30)41(a) (b)Figure 1.12: Gradient designs to null zeroth moment (a) and first moment (b). With twolobes only m0(t) = 0 after the gradient. With three lobes m0(t) = m1(t) = 0 after thegradient.The practical implementation of GMN consists of designing the gradient in such a way that thegradient moment is zero at the end of the gradient. Nulling m1(t) will affect spins flowing witha constant velocity, and m2(t) spins with constant acceleration. The zeroth moment accounts forstationary spins. From this it can be seen that the total phase accumulated during a gradient isrelated to the gradient moment. Example of gradient waveforms with GMN for the zeroth and firstorder nulled is shown in figure 1.12. A gradient with m0 nulled consists of two gradient lobeswith opposite sign. A gradient with m1 nulled consists of three gradients lobes, as seen in figure1.12. In general, to null the n:th gradient moment, n+2 gradient lobes are required. It is commonpractice in spinal cord MRI to employ GMN of the first order to suppress CSF flow artifacts. Higherorder gradient moments could be considered but it would require additional gradient lobes whichsignificantly increases scan time.1.4.5 Insufficient Field of ViewSince MRI data is collected in the Fourier domain (k-space) the FOV is related to resolution ink-space. To avoid aliasing after reconstruction (an inverse Fourier transform) the FOV needs to42include the full imaging volume. This can difficult when imaging the cervical spinal cord. In theoptimal case, the field of view would be constrained to only the neck. However, in patients withhigh shoulders the top of the shoulders could be in the same inferior/superior plane as the imagingvolume. If this happens, a signal will be received from the shoulders as the slice selective excitationpulse will excite spins there as well. When the FOV is set to only the neck, i.e., not covering thewhole volume that is being imaged, the signal from the shoulder will fold over inside the neck andinterfere with the image.Fold over artifacts will only appear in the phase encoding direction as the frequency encodingdirection is usually either oversampled or digitally filtered. There are therefore a few options forreducing fold over artifacts: (a) increase the FOV to include the shoulders (b) swap phase and fre-quency encoding directions, and (c) apply spatial saturation bands. In spine imaging, option (b)could appear attractive as the body is smaller in the sagittal plane than the coronal plane. However,this would put the throat area in the phase encoding direction, thus increasing the risk of motionartifacts. Further more, most receiver coils for spinal cord imaging are surface coils, which pre-cludes the use parallel imaging acceleration in the anterior posterior direction. Increasing the FOVis the best option and can be easily done using a built in feature on the scanner called fold-oversuppression. This will increase the field of view by a factor n, while at the same time increasingthe number of phase encoding steps by the same factor to keep the same resolution. After imagereconstruction, only the original field of view is displayed and the user only sees the image thatwas desired from the beginning. However, this results in increased acquisition time.1.4.6 Magnetic Field InhomogeneitiesAt the boundary between two materials with different magnetic susceptibilities, a local magneticfield gradient will appear[68]. The spine is an inhomogeneous structure where boundaries betweenbone, cartilage and CSF can produce local magnetic field gradients. This can cause two types ofartifacts: image distortion, and echo-shifting. Distortion refers to the mis-registration of spins. The43spin location is determined through the phase and frequency encoding process. If an additionallocal, and unknown, gradient influences the spins as well, the reconstructed image will show thesespins to another location. The shift will be proportional to the strength of the local gradient. If thegradient is along the readout direction, the spins will be shifted along this direction. If the gradientinstead is in the phase encoding direction, we will see a slight shearing effect on the image. Toreduce the shift in the image, a stronger frequency encoding gradient has to be applied. As anexample, consider the phase acquired by a spin affected by both the frequency encoding gradient(Gx) and local gradient created by susceptibility differences (G′x)φ(t)′ =−2piγ(Gxx+G′x)t ′ =−2pikxx(1+G′xGx), kx = γGxt ′. (1.31)This shows clearly that if the frequency encoding gradient Gx is increased the shift is minimized.If the frequency encoding gradient is increased, the size of k-space will increase and the voxel sizewill decrease unless the sampling rate also is decrease. In either case the SNR will decrease. Thiscan be seen by observing the expression for SNR previously presentedSNR≈ ∆x∆y∆z√Tacq,total. (1.32)where ∆x∆y∆z is the voxel size and Tacq,total is the total time data is acquired from each voxel.Higher resolution implies smaller voxel size, thus lower SNR. The same problem will appear in thephase encoding and slice selection directions. Instead of a pure shift a shearing effect will appear.With the phase encoding in y as an example:φ(t)′ =−2piγ(Gxx+G′yy)t ′ =−2pikx(x+G′yGxy). (1.33)44From this we see that a local gradient in the phase encoding direction G′y will change the encodedx′ position decided by the phase. This position will be set by the y position, thus a shift in k-space.The analogous argument can be made for the slice selection by switching y for z. From this it canbe concluded that while increasing the readout gradient Gx will reduce SNR, it will also decreasegeometric distortion from magnetic field inhomogeneities.The other detrimental effect related to local field gradients is echo-shifting. This is only presentin gradient echo sequences, thus a potential problem in a GRASE sequence. A gradient echo isproduced when the time integral of the gradient is zero. This occurs because the applied gradienteffectively brings the spins in phase, thus producing a coherent magnetization. A pulse sequenceemploying gradient echoes will have gradients designed for readout at the points where the integralof the gradients is zero. However, if additional gradients influence the spins, i.e., the local fieldgradients, the gradient echo will not appear where it is expected. The echo is expected to appear attime TE, but instead it now appears at TE’ which depends on the strength of the local field gradients.The shift in the frequency encoding gradient will be proportional to the phase difference expressedin (1.31).1.5 Overview of ThesisMost studies in the literature on myelin water imaging (MWI) utilize either a single slice or 3D multiecho spin-echo (MSE) sequence, similar to the protocol proposed by Mackay et al. [50]. With theintroduction of gradient- and spin-echo (GRASE) as an accelerated option to the traditional MSE forMWI, new opportunities have opened up. With an acquisition time less than 15 min for full cerebralcoverage, MWI using GRASE is a feasible option as a quantitative MRI technique for neurologicalstudies.The goal of this thesis is to present further evidence of the strengths of MWI using GRASE.Spinal cord MWI has so far only used MSE. In chapter 2 I will present a version of the standard45GRASE protocol for spinal cord imaging. The new GRASE sequence is evaluated in terms of re-producibility and compared to the previous MSE sequence. In chapter 3 I investigate multi centerreproducibility of GRASE between two scan vendors. Multi center studies with MWI have onlyinvestigated the reproducibility of the MSE sequence. For MWI with GRASE to gain wider traction,multi center and multi vendor reproducibility is integral.46Chapter 2Rapid Myelin Water Imaging in HumanCervical Spinal Cord2.1 IntroductionExtending down the spine, measuring a cross sectional area no larger than 1 cm2, the spinal cordis the essential connection between the brain and the rest of the body. Despite its indisputablesignificance in the central nervous system (CNS), and known involvement in many neurologicaldiseases, imaging of the spinal cord using magnetic resonance imaging (MRI) is lagging behindbrain MRI[74]. This is largely because the spinal cord is inherently difficult to image. It is a smallstructure, thus requiring a high resolution to resolve fundamental anatomy such as white and graymatter. On a clinical 3T MRI, anatomical scans can produce a resolution as low as 0.3 mm2 in plane.However, the majority of quantitative MRI techniques are limited by signal to noise ratio (SNR) andrequire larger voxel dimensions. More so, compared to the relatively static brain, the spinal cordis a moving structure surrounded by pulsating cerebrospinal fluid (CSF). As discussed in detail in47chapter 1, the spinal cord moves about 1 mm along the inferior superior axis and the surroundingCSF reaches velocities up to about 0.9 cm/s[83].Using a gradient- and spin-echo (GRASE) sequence, Prasloski et al. [62] were able to acceleratebrain myelin water imaging (MWI) data acquisition by a factor of 3, down to 15 min, approachingwhat is referred to as clinically feasible times. As is usually the case, this was first only appliedin the brain, leaving the multi echo spin-echo (MSE) sequence as the only option for spinal cordMWI[51]. The MSE protocol for spinal cord MWI has an acquisition time of over 20 min, makingit difficult to integrate into clinical research protocols which usually are limited to about 40 minof effective scan time. In this chapter I will present an implementation of the previously describedGRASE sequence[62] for MWI in cervical spinal cord. The goal was to investigate the reproducibil-ity and fidelity of the GRASE sequence and evaluate whether it is a viable alternative to the MSEsequence for spinal MWI, thus making spinal cord MWI a feasible option for quantitative MRI inclinical applications.2.2 ObjectivesTo adapt the 3D GRASE protocol previously developed by Prasloski et al. [62] for spinal cordimaging and evaluate the reproducibility in healthy volunteers. The new sequence will also becompared to the previous MSE sequence used by MacMillan et al. [51].2.3 MotivationSince the implementation of the accelerated GRASE sequence, full cerebral MWI can be achievedin under 15 min[62]. However, spinal cord MWI is still utilizing the older MSE sequence whichtakes more than 20 min with coverage of two vertebral levels[51]. Given the known involvementof the spinal cord in neurodegenerative diseases such as multiple sclerosis (MS) and neuromyelitisoptica (NMO)[36], it is of great interest to accelerate spinal cord MWI to the same clinically feasibletimes as brain MWI.482.4 HypothesisThe 3D GRASE sequence has previously been implemented for brain MWI with high reproducibilityand a strong correlation to the previous MSE sequence[61]. However, given the issues associatedwith spinal cord MRI such as motion, CSF flow, and susceptibility artifacts (as explained in chapter1) it was our belief that including gradient echoes, as are used in GRASE, would introduce moreartifacts. Our main hypothesis was that, despite known issues with artifacts, the GRASE sequencecould be adapted for spinal cord to produce high quality and reproducible MWI data.2.5 Methods2.5.1 Study Population12 healthy volunteers with no documented injuries or disease in the spinal cord participated inthe study (7F, 5M; age: median=27, range=20-47). All volunteers provided signed consent beforeparticipating under the approval of the University of British Columbia Clinical Research and EthicsBoard.2.5.2 MRI ExperimentsThe 3D GRASE protocol for spinal cord imaging was evaluated in terms of: scan-to-scan repro-ducibility and correlation to a previously used MSE sequence[51]. MRI data was acquired at theUBC MRI Research Centre on a 3.0 T human MRI scanner (Philips Achieva 3T, Philips Medi-cal Systems, Best, The Netherlands) using a 6-channel spine coil. The 3D GRASE sequence byPrasloski et al.[62] was adapted for a smaller field of view (FOV) and higher resolution, the full setof pulse sequence parameters are presented in table 2.1. A 3D MSE sequence, identical to that usedby MacMillan et al. [51], was also acquired with pulse sequence parameters also outline in table2.1. As the main goal of this study was to reduce the acquisition time, cardiac triggering was not49performed as this would increase the scan time significantly for subjects with slower cardiac cyclesand also introduce repetition time (TR) variability.GRASE MSEEcho spacing 10 ms 10 msTR 1500 ms 1500 msFOV (AP,RL,FH) 180x150x40 mm3 180x135x40 mm3Slices per slab 16 8Acquired resolution (AP,RL,FH) 0.75x0.75x5 mm3 0.70x1.41x5 mm3Reconstructed resolution (AP,RL,FH) 0.63x0.63x2.5 mm3 0.7x0.7x5 mm3Slice oversampling 1.2 1Averages 1 1Fat suppression No NoExcitation flip angle 90◦ 90◦Refocusing flip angle 180◦ 135◦EPI factor 3 -Acquisition Time 8.5 min 23.5 minTable 2.1: Puls sequence parameters used for the GRASE and MSE for MWI in the spinal cord.For segmentation and anatomical contrast between white matter (WM) and gray matter (GM),a multi-echo fast field echo (MFFE) sequence was acquired with the following parameters: echotime (TE)1=6.6 ms, TE2−5=8.2 ms, TR=815 ms, reconstructed resolution=0.3x0.3x2.5 mm3, scantime=5 min. A spatial saturation slab was placed over the pharynx to reduce motion artifacts fromswallowing. All scans (GRASE, MSE and MFFE) were centered at the C2/C3 disc parallel to thecord, see figure 2.1.Reproducibility of the GRASE sequence was assessed by acquiring two GRASE scans, one at thebeginning and one at the end of the scan session. Since no previous studies have used GRASE forMWI, the MSE sequence was used as a reference which would allow comparison of myelin waterfraction (MWF) values available in the literature.50Figure 2.1: Screen shot from scanner interface showing imaging volume, slices, shim boxand the spatial saturation band used in the MFFE. Imaging volume was the same for allscans but saturation band was only used for MFFE.2.5.3 MWI AnalysisMyelin water fraction (MWF) maps were obtained from the T2-relaxation data, acquired by theGRASE and MSE, using a non-negative least squares algorithm with stimulated echo correction[61](see section 1.3.3) and non-local spatial regularization developed in-house[90]. The non-localregularization algorithm is particularly beneficial for MWI analysis in spinal cord as it is moreeffective for images with low SNR. It is also useful for small structures, such as GM in the spinalcord which could be over-smoothed in local regularization where all neighboring pixels contributeequally. The range of T2 times associated to myelin water was chosen to be 10 <T2< 40 ms.512.5.4 Image AnalysisThe main steps of the image analysis pipeline are outlined in figure 2.2. The GRASE and MSEscans were registered to the MFFE scan using the Spinalcord Toolbox[18]. Due to variable contrastbetween multi-echo data (GRASE and MSE) registration was based on spinal cord segmentationperformed using the PropSeg tool from the Spinalcord Toolbox[89]. This limits the accuracy of theregistration but is adequate for this study since registration only is performed between scans fromthe same patient acquired in the same scan session. The MFFE scan was registered to the AMU-MNI-POLY spinal cord template using a non-linear transform based on both anatomy and spinalcord segmentation[25]. Regions of interest (ROI) were transformed from template space to MFFEspace using the inverse transform acquired from the previous step. The spinal AMU-MNI-POLYtemplate contains 30 isolated white-matter tracts[7] but due to the low resolution in the acquireddata some of these tracts were combined to form a larger ROI. Tract 00, 01, 15, and 16 werecombined to form the dorsal column (DC); tract 02 and 17 were combined to form the corticospinaltract (CCST). A gray matter (GM) ROI from the spinal cord template was also used. All ROI arepartial volume masks, i.e., the mask values range from 0-1 to indicate the volume fraction of eachROI in a given voxel. Prior knowledge from using the spinal cord template has suggested thatmanual intervention can be required to achieve precise ROI delineation. Because of this, each ROIwas thresholded at 0.5 and binarized. This allows for easier visual inspection and adjustmentswhen required. The value at which the masks were binarized was decided from visual inspection,and kept constant throughout the analysis for all subjects.All ROI were cropped to only include the middle 10 slices of the MFFE to avoid artifacts fromwrap around effects in the slice direction, especially in the MSE where no slice oversample wasused. MWF maps were transformed to MFFE space and the average MWF was calculated for eachROI (GM, DC, CCST, and full cord). By transforming all MWF data to the same space, only oneset of ROI has to be produced. This reduces inconsistencies between scans and ensures that the52Figure 2.2: Overview of the image analysis pipeline using tools from the Spinal Cord Toolboxwith the ROI used in the analysis.reproducibility measure actually is a measure of the scan reproducibility and not ROI registration.Since all scans were transformed to MFFE space, potential bias induced by the transform will bethe same for all scans.While MWF is the main reproducibility measure of interest in this study, average cross sectionalarea (CSA) of the cord was also calculated for each scan. The purpose of this was to investigateif inconsistencies in MWF could be traced back to variations in CSA. This could in turn be dueto patient motion or repositioning of the patient that was not captured during registration. Sincethe spinal cord varies in CSA along its long axis, displacement in that direction would be pickedup by CSA measurements. The CSA was calculated using spinal cord segmentation from the 10thecho in multi-echo data (GRASE and MSE) and the accumulated MFFE scan using the SpinalcordToolbox. One of the later echoes in multi-echo sequences was used to produce clear T2 contrastfor registration.2.5.5 Statistical AnalysisThe most common method for assessing a correlation between two variables x and y is linear re-gression. This assumes that the independent variable x is known without uncertainty. However, if53regression is performed between two measurements where both are associated with some measure-ment error, the ordinary least square algorithm used to find the slope in linear regression might notbe the best choice. An alternative to linear regression, which takes these uncertainties into accountis orthogonal regression which utilizes the total least square algorithm for finding the slope. Insteadof minimizing the error between the best fit line and the measured data along the y-axis, the erroris measured orthogonally from the best fit line to the data point. Therefore, correlation in MWFbetween MSE and GRASE was assessed using both linear and orthogonal regression with averageMWF obtained from the isolated ROIS (GM, CCST, DC). Bland-Altman analysis using the same ROIwas also used to evaluate any biased trend in the difference between the two techniques.To quantify the reproducibility of the GRASE scan, several measures were used. Many studiesin the literature report reproducibility of MWI in terms of the coefficient of variation (CV) calculatedusing[69]CV =(1+14n)· STDEV(MWFGRASE1,MWFGRASE2)(MWFGRASE1+MWFGRASE2)/2. (2.1)It is clear that the CV will depend strongly on the mean value within the ROI and therefore largerCV will be reported for ROIS with inherently low MWF, such as GM. To avoid this issue, intra-class correlation (ICC) or Cronbach’s α can be used instead[14]. Both Cronbach’s α and ICC willapproach 1 for higher reproducibility where values larger than 0.7 are considered acceptable. Inthis study, reproducibility will be reported in terms of: CV, ICC(3,k), Cronbach’s α , and meandifference between the measurements. This will allow for comparison to previously reported val-ues of reproducibility in the literature. A two sample paired t-test was used to test for significantdifferences in MWF and CSA between the different sequences.Statistical analysis was carried out using MATLAB (Mathworks, Natick MA, USA) and RStudio (RStudio Inc, Boston MA, USA).54(a) (b)Figure 2.3: (a) Correlation plot for the MWF between the GRASE and MSE. Fit statisticsshown in table 2.2. (b) Bland-Altman plot comparing the GRASE and MSE sequence.Mean=-0.0091, 95% confidence interval=mean±0.039. MWF from the GRASE in a andb is the average of the two repeated scans.2.6 ResultsTwo subjects were removed from the analysis. The first had severe motion artifacts, making theanalysis impossible. The second was affected by phase wrap, causing the edge of the shoulder tofold into the center of the cord.2.6.1 Comparison Between GRASE and MSE MWIAverage whole cord segment MWF was in good agreement between the two sequences: 22.9%(SD=2.5%) for GRASE and 22.5% (SD=3.4%) for MSE. The same holds for DC, CCST and GM ROI,see table 2.3. A two sided paired t-tests showed no significant difference between MWF obtainedwith the GRASE and MSE. Linear and orthogonal regression using MWF obtained in each ROI isshown in figure 2.3a with fit parameters shown in table 2.2. Using linear regression the slopebetween MSE and GRASE was 0.9 for all ROI (95% confidence interval 0.72-1.07) with a R2 = 0.8,with orthogonal regression the slope was 1.0 (95% confidence interval 0.80, 1.21).Bland-Altman analysis of the MWF obtained from the two sequences is shown in figure 2.3bwith no apparent trend for larger differences in MWF for ROIS with higher MWF. This is contrary to55Slope (95% Conf. Interval) Y-intercept (95% Conf. Interval) R2Linear regression 0.88 (0.71, 1.07) 0.018 (-0.028, 0.064) 0.8Orthogonal regression 1.00 (0.80, 1.21) -0.012 (-0.066, 0.046) -Table 2.2: Linear and orthogonal regression parameters between the average MWF obtainedfrom the repeated GRASE scan and the MSE scan.results from Bland Altman analysis obtained by Prasloski et al. [61] who observed a trend towardsgreater difference in MWF between GRASE and MSE for WM ROI with higher MWF values in brain.This appears as a positive slope in the Bland Altman plot.T2 decay curves (figure 2.4a) and T2 distributions (figure 2.4b) from GRASE and MSE are qual-itatively similar. The oscillating pattern in the end of the decay curve in figure 2.4a is attributedto flow artifacts caused by even-odd echo refocusing. At first glance the T2 distributions in figure2.4b might appear to differ between sequences but given the small ROI, here about 10-20 voxels,this is expected. The important fact is that within all 3 ROI, there are two peaks: one around 70 msand one at the end of the x-axis around 15 ms. This represents the intra-extra cellular water andmyelin water peaks, respectively. At the other end of the T2 spectrum, around 2000 ms, a hint ofthe tail of another peak attributed to cerebrospinal fluid (CSF) appears as well.A visual comparison between the MWF obtained with GRASE and MSE from the first six subjectsis shown in figure 2.5 together with the anatomical MFFE. From this subset of scans, the GRASEsequence appears to produce the best anatomical detail with the characteristic GM structure shapedlike a butterfly in the center of the spinal cord.2.6.2 Reproducibility of GRASE MWIA summary of all reproducibility measures is presented in table 2.4. CV for the GRASE showedoverall good reproducibility with an average CV for the full cord of 6.1%. There is a larger spreadin reproducibility between the ROI when ICC(3,k) and Cronbach’s α is used, ranging from 0.3 to0.89. The repeated GRASE sequence produced qualitatively similar MWF maps, figure 2.5, with the56(a) (b)Figure 2.4: (a) Example of T2 decay curves obtained from all three scans in DC. Decay curveamplitude was normalized to 1 at the first echo. (b) T2 distributions obtained from MWIanalysis. From left to right the peaks in the spectrum are attributed to: myelin water(< 15 ms), intra-extra cellular water (≈ 70 ms), and CSF > 2000 ms.Figure 2.5: Overview of the different scans from the first 6 subjects obtained from the middleslice registered to multi-echo fast field echo (MFFE) space.57Study Sequence B0 ROI MWF (SD) [%]Minty et al. [57] Single slice MSE 1.5T C2-C7 21.8 (2.1)Laule et al. [41] Single slice MSE 1.5T C2/C3 25.7 (1.4)Wu et al. [88] Single slice MSE 1.5T Upper cervical 26 (2)MacMillan et al. [51] Multi slice MSE 3TC4/C5 DC 30.6 (1.2)C4/C5 CCST 28.4 (1.2)C4/C5 GM 4.9 (3.9)Present StudyGRASE 3TC2-C3 DC 29.3 (2.6)C2-C3 CCST 26.0 (3.1)C2-C3 GM 15.0 (2.3)C2-C3 22.9 (2.5)MSE 3TC2-C3 DC 30.3 (2.7)C2-C4 CCST 26.0 (3.2)C2-C3 GM 16.9 (3.7)C2-C3 22.5 (3.4)Table 2.3: MWF in human cervical spinal cord in vivo from previous studies. A range invertebral levels is indicated by a - and specific disc location between two vertebrae with/. The WM ROI is the average of the DC and CCST ROI. If no ROI is specified, the MWFis an average of the ROI includes the whole cord. MWF from the GRASE in the presentstudy is the average of the two repeated scans.characteristic butterfly shaped gray matter tracts, here shown in darker color, in the center of thecord.The average difference in full cord MWF between the first and second GRASE was 1.2 percent-age points (PP). A paired two sided t-test revealed a significant difference (p < 0.05,ρ = 0.85)for whole cord MWF obtained by the repeated GRASE scan. However, the Bland-Altman plot com-paring the repeated GRASE scan, in figure 2.6, shows that the difference in MWF for each ROI isscattered around a mean close to 0. There is no noticeable trend for larger difference between therepeated scans with higher mean MWF in the ROI indicating the difference between the two scansis similar over the whole range of MWF values.58Region of InterestReliability Measure Full Cord DC CCST GMAverage CV 6.1 7.75 8.08 11.50Cronbach’s α 0.89 0.3 0.67 0.53ICC(3,k) 0.89 0.31 0.68 0.53Average difference 1.2 0.98 0.6 0.52t− test no difference p= 0.04 p= 0.4 p= 0.6 p= 0.5Table 2.4: Summary of reproducibility measures of measured MWF with the repeated GRASEscans. ROI: dorsal column (DC), corticospinal tract (CCST), gray matter (GM).Figure 2.6: Bland-Altman analysis of MWF obtained from the repeated GRASE scans. Mean=-0.0068, 95% confidence interval=mean±0.051.2.6.3 Cross Sectional Area MeasurementsCSA measured on the MFFE is significantly lower than the two GRASE scans, shown in figure 2.7and table 2.5. The MFFE shows the lowest variability within the group, SD=5 mm2, while theGRASE and MSE have a SD≈ 8 mm2. This is not surprising though since the MFFE has higherresolution and thus higher sensitivity. A striking observation is that CSA from the second GRASEscan is significantly lower than the first one (p = 0.012 from paired t-test). Both GRASE scansproduce higher CSA than the MSE, although only the first GRASE scan produce significantly higherCSA (p = 0.0017). CV for CSA for the repeated GRASE scan was on average 1.5%. This suggests59that the potential magnetic susceptibility artifacts are insignificant and do not induce distortion inthe image.Figure 2.7: Box plot of cross sectional area (CSA) measurement obtained with the MFFE,GRASE, and MSE scans.Difference in CSA (t-test statistic)mFFE GRASE 1 GRASE 2mFFE - - -GRASE1 -4.06 (p= 0.012) - -GRASE2 -3.06 (p= 0.026) 0.99 (p= 0.012) -MSE -1.97 (p= 0.15) 2.01 (p= 0.0017) 1.01 (p= 0.09)Table 2.5: Difference between scans in mm2 and p-value for two tailed paired t-test betweenthe two measurements.2.7 Discussion2.7.1 Comparison Between GRASE and MSE MWIWhen the GRASE sequence was first implemented for brain MWI by Prasloski et al. [62], linearcorrelation between GRASE and MSE showed a slope of 1.14 (R2 = 0.95), compared to the presentstudy where the slope of the linear regression was 0.88. However, the MSE protocol used by60Prasloski et al. study had a flip angle of 180◦ and therefore the two protocols are not directlycomparable. Bland-Altman analysis by Prasloski et al. found a positive slope in the Bland-Altmanplot suggesting a bias for ROI with higher MWF. However, with p = 0.06 this was not significant.Bland-Altman analysis in this study covered a larger range of MWF values (0.1-0.35) and no trendin the data was found, indicating that neither of the sequences are biased at any level.From the T2 distributions in figure 2.4b it can be seen that the MSE tends to produce wider T2distributions occasionally ranging beyond the T2 window assigned to myelin as shown by the blackdashed line. The T2 distribution from the GRASE however tends to stay within the limits. If theintra-extra cellular T2 peak leaks into the myelin T2 window, it will be accounted for as myelinwater and thus contribute to the MWF. However,since the GRASE appeared to produce consistentT2 distributions with the intra-extra cellular T2 peak outside the myelin window we decided to keepthe original window limits.The myelin estimates from WM and GM obtained from this study are in good agreement withprevious studies. The most apparent difference is the higher MWF obtained from the GM ROI inthe current study (15%) compared to that reported by MacMillan et al. [51] (5%). The GM maskused by MacMillan et al. only included the anterior horns while the GM mask used in the currentstudy covered the whole GM structure. This will likely introduce partial volume effects from WMand thus increase the MWF. However, MWF values obtained from the DC and CCST are in goodagreement with lower MWF in the CCST compared to the DC.2.7.2 Reproducibility of GRASE MWIThe average difference in MWF between the two GRASE scans was 0.012 and the difference wasfound to be significant at p = 0.04. This is surprising as no difference between the scans wasexpected. The cause of this bias for higher MWF values in the second scan is unknown, howeverit could be related to the significant decrease in CSA observed in the second GRASE scan, which is61further discussed in section 2.7.3. This warrants for further studies to investigate if the MWF resultsare influence by when in the scan protocol the GRASE data is acquired.Since this is the first study to use a 3D GRASE sequence for spinal cord imaging there is noliterature to compare the obtained reproducibility to. MacMillan et al. [51] studied the scan-to-scanreproducibility of the MSE sequence, reporting reliability in terms of Cronbach’s α of 0.65 in DCand 0.82 in CCST. No values were reported for full cord reproducibility. These values are higherthan those obtained from the current study (0.3 in DC and 0.68 in CCST). However, Cronbach’s αfor full cord was estimated to 0.89 in the present study which is considered as highly reproducible.Meyers et al. [55] performed a multicenter reproducibility study of MWI with five subjects whoreceived brain MWI at 6 different sites using equivalent Philips Achieva 3.0T MRI scanners usinga 7 slice MSE protocol. Here, reproducibility was quantified in terms of the CV. The average CVfrom all sites was 3.99%, ranging from 2.47% to 8.10%. Looking at individual scans from each sitethe CV ranges from 0.15% to 25%. Lower CV is expected for brain imaging since larger ROI canbe used for the brain and there is also less motion. However, the results obtained from the presentstudy with an average CV of 6%, it is not far away from the 4% achieved by Meyers et al. for brainMWI.An alternative to GRASE and MSE for myelin imaging is multicomponent driven equilibriumsingle pulse observation of T1/T2 (MCDESPOT). In a study by Kolind and Deoni [39], MCDESPOTdata was acquired from the whole cervical spinal cord in 26 min. They reported MWF of 25.1% inthe DC and 10.9% in GM with a reproducibility within subjects of CV=2.6%. The lower CV obtainedfrom this study can in part be explained by the larger ROI used, covering the whole cervical cord.A larger volume will average out noise and better reproducibility can be achieved. No CV wasreported for WM or GM ROI.Reproducibility as described in this study is a measure of the precision of the GRASE sequencefor spinal cord imaging. This is an important parameter to include when planning a clinical study62using GRASE for MWI. High reproducibility, precision, allows for smaller sample sizes, thus amore affordable study. To determine the sample size required in a clinical trial to reach the desiredstatistical power, the CV can be used. Consider a clinical study with a cohort of MS patients wherethe the MWF is expected to decrease between two scan sessions. The number of patients requiredto reach statistical significance (α = 0.05,1−β = 0.8) can be expressed as[6]n=(4 ·CV (in %)100(1−1/ f ))2, CV < 50%,1 < f < 1.5 (2.2)where f = µ1/µ2 and µ is the MWF at the first and second scan session. To put this into context,Laule et al. [44] measured a change in MWF of -10.5% over 2 years in patients with primaryprogressive multiple sclerosis (PPMS). At baseline the mean MWF was 22.1%, thus a decrease by10.5% gives f = 1.105. With full cord CV=6.09%, as reported in this study, (2.2) would equate ton = 7. This tells us that if a 10% decrease is expected, 7 patients are required to ensure that thelongitudinal change is significant at a level α = 0.05,β = 0.2, where α indicates the probabilityfor type 1 errors and β for type 2 errors. In the study by Laule et al. the change over 2 yearswas significant (p = 0.01), however, the 2.6% decrease found at year 1 from baseline was notsignificant.2.7.3 Cross Sectional Area MeasurementsThe significantly smaller CSA in the second GRASE compared to the first GRASE is an interestingobservation that requires further investigation. From a physiological point of view, it could beexplained by the fact that the patient usually become more relaxed during the scan session and thustheir heart rate drops. Since the flow of CSF and motion of the cord is closely related to the cardiaccycle, the smaller CSA could be a reflection of less motion. However, it seems unlikely that thiswould be the explanation for all subjects as many of the volunteers are experienced MRI volunteersand their heart rates would remain fairly constant during the scan session. In a study by Wang et al.63[84], the area of the spinal cord was found to be significantly reduced after 14 hours of dehydration.Even though this is far beyond the time frame of this study, it is an indication that physiologicalfactors such as hydration will affect the spinal cord CSA. Thus, it can not be excluded that theobserved reduction in CSA is due to some, currently unknown, physiological change during thescan session. From a physical point of view, the observed difference in CSA between the repeatedGRASE scans, may be related to the gradients and that the temperature is slightly higher at the endof the scan session than in the beginning. How and why this would affect the CSA is currentlyunknown. One hypothesis is that gradient temperature increases after the first GRASE acquisitionwhich would affect the passive shim, and thus the affect the shimming of the imaging volume whichmight lead to geometric distortions. If the gradient temperature is constant after the first scan thenthat could explain why the data first GRASE acquisition is significantly different from the secondGRASE and the MSE. Further work should also be dedicated to investigate if the B1 field, i.e., theflip angle, change between the repeated GRASE scans.It was shown in this study that CSA measurements should be acquired with a high resolutionscan, such as the MFFE instead of GRASE to improve precision. This can be seen by the smallerconfidence interval for the MFFE in figure 2.7. The fact that the two sequences produce differentCSA measures is likely due to the difference in contrast (T2∗ in MFFE and T2 in GRASE) causingthe automated segmentation tool to estimate the cord boundary differently. Similar results havebeen found previously when comparing CSA obtained with T1 and T2 weighted scans[37]. Thishighlights the importance of consistency in study design. Any change in the pulse sequence thatwill change the contrast will also change CSA measurements. The difference in CSA from therepeated GRASE scan might also suggest that scans should be placed in the same order within onesession.642.7.4 Imaging Artifacts and RemediesThe greatest advantage of the GRASE sequence is the shorter acquisition time, which has manybenefits. Shorter scan time improves patient comfort and also reduces the risk of patient motionwhich will cause artifacts. In MWI, 32 echoes are collected in each TR and k-space is graduallyfilled throughout the whole acquisition, meaning that any motion artifacts are difficult to correct. Tocontrast this, in advanced diffusion protocols for instance, up to 100 diffusion volumes are acquiredsequentially using an echo planar imaging (EPI) sequence. This allows for potential motion artifactsbetween volumes to be corrected for in post processing. In a structure like the spinal cord thishas great implications as the spinal cord constantly moves due to the cardiac cycle, regardlessof patient motion. While cardiac gating would reduce these artifacts, Summers et al. [75] foundthat the quiescent phase of the cord velocity lasts about 40% of the cardiac cycle at the C2 level.To fit the whole echo train within this window would require a heart rate lower than 75 beatsper minute (BPM) which cannot be guaranteed. Additionally, the acquisition time would increasesignificantly. With a minimum TR of 1500 ms as used in the current study, the actual time betweenacquisition will depend on the heart rate if cardiac triggering is used. For example, with a heartrate of 75 BPM, each cardiac cycle lasts 0.8 s. In this case, 2 heart cycles would suffice to reacha TR of 1500 ms. However, if the heart rate drops to 60 BPM, each cycle lasts 1 s and while 2cycles would suffice the TR would increase to 2000 ms, i.e., by 30%. To optimize a sequence withcardiac triggering the TR be variable with the lower limit of the minimum prescribed TR. In theGRASE sequence used in the present study TR is short enough to introduce some T1 weighting inthe end result. This is not considered to an issue since the TR is kept constant. However, if theTR is variable, the T1 weighting will differ between k-space lines and potentially influence theresults. From a practical point of view, the scan time could potentially within a range large enoughto prolong the scan session beyond the allotted time. With this in mind, cardiac triggering was notconsidered to be a feasible option and, as it turns out, it does not seem to be required.65Another factor also related to the cardiac cycle is flow of CSF inside the spinal column. Justas the cord itself moves, the pulsating CSF can potentially cause flow artifacts. In the MFFE thiswas reduced by using flow suppression, equivalent to gradient moment nulling as explain in sec-tion 1.4.4. While this suppresses most flow artifacts, some artifacts are still present as shown infigure 2.9a. Gradient moment nulling is not possible with the GRASE or MSE sequence since thiswould require modification of the gradient waveforms and thus, no form of flow artifact reductiontechnique was used. The oscillating data points seen in the last part of the decay curve, as shownin figure 2.4a, may be due to inhomogeneities in the B1 field, resulting in a reduced flip angle.However, this is accounted for in the extended phase graph (EPG) algorithm and therefor not con-sidered to be an issue. Another explanation to the oscillating data points is the constantly flowingCSF. This could cause a phenomenon known as even-odd echo refocusing which can be explainedusing the the theory presented in 1.4.4. When a spin moving with a constant velocity v experiencea constant magnetic field gradient with strength G, the net phase shift compared to a stationary spincan be expressed asφ =∫ t2t1γGvtdt = K · (t22 − t21) K =γGv2(2.3)When a 180◦ radio frequency (RF) pulse is applied, the phase is inverted, i.e., φ1 =−φ0. Considera simple multi-echo spin echo sequence where a frequency encoding gradient is applied after the90◦ excitation pulse and in between the 180◦ refocusing pulses. The phase acquired before the first180◦ pulse is φ = KT 2. After the RF pulse the sign of the phase is changed and then it increasesquadratically with time. At the time of the first echo, t = 2T , an additional phase of 3KT 2 havebeen acquired and the net phase is now φ = 2KT 2. This process is outlined in figure 2.8 wherethe incremental phase represents the acquired phase during the applied gradient shown in green.If the argument above is extended to the second echo, it can be shown that the net phase will beφ = 0, i.e., the spins are in phase. The lower the net phase dispersion is, the higher the acquired66signal will be. Thus, even echoes will in this situation produce a higher signal than odd echoes.The reduction in signal intensity observed at odd echoes will depend on the gradient strength G,flow velocity v and gradient duration T . For more details on even-odd echo refocusing, pleaserefer to the appendix section A.2.1. The appearance of flow artifacts in the decay curve is furtherFigure 2.8: Schematic explanation of the concept behind even-odd echo refocusing for a sim-ple MSE sequence. Spin echoes are labeled E1, E2, E3, and E4. K is defined in 2.3. Ateven echoes the net phase is zero, thus high signal.supported by the presence of CSF in the T2 distribution in figure 2.4b. At T2=2 s, there is a hintof a tail of another peak that could be attributed to CSF In brain MWI this is only seen in ROI closeto the ventricles, but given the small size of the cord, it can be expected in almost any ROI. Onetechnique to reduce the signal from CSF is to apply an inversion recovery pulse prior to the echotrain acquisition[57]. This will however reduce SNR and was not considered to be a feasible optionin this study where high in-plane resolution was desired. Another explanation to the peak at T2=2s is a constant signal offset in the measured data. When magnitude data is created from the realand imaginary part after the inverse Fourier transform in the reconstruction process, the standarddeviation of the noise (σ ) from the real and imaginary part combine in a non-linear process. Ina voxel with no signal the following mean µM and standard deviation σM can be observed in themagnitude data [29]µM = σ√pi2, σ2M = σ2 ·(2− pi2). (2.4)67From this it can be seen that even though no signal is originating from a voxel, the magnitudedata will produce a small signal. This signal is present throughout the echo train and in situationswhere this offset is large, i.e., in areas with low SNR the analysis algorithm will interpret this as acomponent with very long T2.The main difference between the GRASE and MSE sequence is the use of gradient echoes forread out. This was hypothesized to pose challenges when imaging the spinal cord. As mentioned insection 1.4.6, the heterogeneous magnetic susceptibility in the spine will cause an inhomogeneousmagnetic field which in turn can cause susceptibility artifacts such as distortion and echo-shifting.Both of these artifacts will be reduced with increased gradient strength in the frequency encodingdirection, Gx. In this study, in-plane square voxels were used for the GRASE, but the sequencemight be possible to accelerate even more by acquiring rectangular voxels with a larger dimensionin the phase encoding direction. This will not influence image distortion or echo-shifting artifacts,as previously shown in (1.31).To reduce potential artifacts from swallowing, spatial saturation bands were used with the MFFEsequence. However, this is not a viable option for MWI. The spatial saturation band will apply aspatially selective RF pulse for the area that is being suppressed, but this implies that for some otherlocation in the imaging volume, this will act as an off resonance pulse, similar to a magnetizationtransfer pulse. It has been shown previously that applying a magnetization transfer pre-pulse priorto the the echo-train acquisition will reduce the amplitude of the low T2 components, thus reducingthe MWF[80]. The other solution for reducing influence of artifacts from swallowing is to choosethe phase encoding direction in the right left direction. This also the preferred choice in order togain the biggest time improvement with sensitivity encoding (SENSE). Since the receiver coil isa surface coil with receiver elements in a horizontal plane, parallel imaging can only be used ineither the foot head or right left direction.68Other issues associated with gradient echo sequences are specific absorption rate (SAR) andgradient slew-rate limits. The GRASE protocol used in this study ran with the highest allowed SARand gradient slew-rate. If patients with metal implants are scanned with the GRASE sequence, theSAR and gradient slew-rate have to be lowered. The standard way to approach this is to reduce thebandwidth which decreases the height of the read out gradient and thus lowers gradient slew-rate.To reduce SAR, the TR is usually increased. Currently unpublished work from our research groupsuggest that the MWF is affected by TR. Caution should therefore be taken before conducting crosssectional studies in a cohort where some patients are scanned using a protocol with modified SARand gradient slew-rate levels. Whether a lower bandwidth will influence the results significantlyhas yet to be determined. However, arguments previously made regarding image distortion andecho shifting suggest there might be problems associated with lower bandwidth.One of the subjects in the current study was excluded due to a fold over artifact from theshoulder. As this was the first time it occurred, it was not anticipated and proper action was nottaken prior to the scan. This caused a severe artifact in one of the GRASE scans, rendering it useless.However, for the second GRASE scan fold-over suppression was used with satisfying results, seefigure 2.9b. With fold-over suppression, the FOV is extended beyond the shoulders and the artifactdisappears.69(a) (b)Figure 2.9: (a) Example of flow artifact in a MFFE scan. Circled at the top is a replica of theCSF inside the cord which originates from the pulsating flow of CSF. (b) Example ofhow fold over artifacts can affect the GRASE image.2.8 ConclusionsThis study shows the first successful implementation of GRASE protocol for in vivo MWI in humancervical spinal cord. MWF measures obtained with the GRASE sequence are highly reproduciblewith an average CV of 6.1 and Cronbach’s α of 0.89. By using a GRASE instead of MSE theacquisition time was reduced by a factor of 3 down to 8 minutes achieving a clinically feasiblescan time. This study provides strong encouragement for future studies to use GRASE for spinalcord MWI. However, there are still unresolved issues that should be the subject of further studies.The small, but significant, difference in MWF and CSA in the repeated GRASE scans should beinvestigated. It should also be investigated if CSF contamination in the T2 decay data will influencethe results.70Chapter 3Multi-vendor Reproducibility of MyelinWater Imaging3.1 IntroductionAs research centers around the world become more connected, the utility for multi-center trials in-creases. By involving multiple sites in a clinical study, a larger number of subjects can be includedwhile keeping the work load at each individual site at a manageable level. Over the last 20 years,the number of multicenter studies utilizing magnetic resonance imaging (MRI) have increased dra-matically as can be seen by the number of references in the literature, visualized in figure 3.1. Inorder for a multicenter study to utilize a specific MRI technique, reproducibility between scannersand scanner vendors needs to be verified. Even though the same protocol is used, minor differencesin pulse sequence implementation and hardware can cause variability in the data. Because of this,numerous studies have focused on verifying reproducibility of various qualitative and quantitativeMRI techniques.71Figure 3.1: Statistics from http://www.pudmed.gov showing number of publications matchingkeyword multicenter trial MRI between 1990 and 2014. Data retrieved on March 12016.Reproducibility of myelin water imaging (MWI) has previously been investigated, with themost significant contribution recently made by Meyers et al. [55] where reproducibility of MWIwas compared between 6 different sites, all using Philips Achieva 3 T MRI scanners. T2 relaxationdata was collected using a 3D multi echo spin-echo (MSE) sequence with 7 slices, 32 echoes, 10 msecho spacing and 1000 ms repetition time (TR). The average inter-site coefficient of variation (CV)for myelin water fraction (MWF), for all subjects, was found to be 4.68%. To date there is onlya single study on the reproducibility of MWI between two scanner vendors. This was performedby Chia et al. [17] who scanned a single subject on both a 1.5 T GE Signa and a 1.5 T SiemensSonata. An equivalent single slice (5 mm slice) MSE sequence was used at both sites. Regions ofinterest (ROI) were manually drawn and MWF was evaluated within each ROI. Their results showedCV ranging from 1% in the genu of the corpus callosum to 40% in the head of caudate nucleus.In 2012, Prasloski et al. [62] introduced gradient- and spin-echo (GRASE) for MWI which wouldaccelerate the acquisition by a factor of 3. This enabled full cerebral MWI in less than 15 min,thus increasing the utility of MWI in clinical studies. The GRASE sequence has been successfully72implemented for MWI in numerous studies since[9, 10, 87, 94], but no studies have been dedicatedto validating reproducibility between multiple scanner vendors. In this chapter, I will attempt tobridge this gap by investigating the reproducibility of MWI using a 3D GRASE at two different 3 TMRI scanners: a Philips Achieva and a Siemens Verio. The goal of this study was to investigate ifMWI produces the same results when scanners from different vendors are used. This is an integralstep in making MWI using 3D GRASE a mainstream application that can be used at any centeraround the world.3.2 ObjectiveTo assess the reproducibility of MWI using GRASE between two scanner vendors and investigatewhich factors in sequence implementation could potentially cause discrepancy in the results. Ifthe results do not correlate, or an offset appears, a solution to reduce this discrepancy will beinvestigated.3.3 MotivationMulticenter MRI studies are becoming increasingly popular. Quantitative MRI techniques need tobe evaluated for multicenter use before they can be adapted for research or clinical studies takingplace across multiple institutes. MWI can be acquired in clinically feasible times using a 3D GRASEsequence, but it has not been tested on multiple scanner vendors. This study will bridge that gap,taking MWI one step closer to the mainstream.3.4 HypothesisMyelin water fraction (MWF) acquired from two different scanners should be highly correlatedgiven that the same pulse sequence is used. In the present study, an equivalent implementation ofthe GRASE sequence was used at both centers and we expect the results to be highly correlated.73Potential offset in the results should be possible to explain from differences in sequence implemen-tation and/or hardware differences.3.5 Methods3.5.1 Study PopulationThree healthy volunteers (2F, 1M; ages 27, 33, 38) were scanned at two sites: the Kleysen Institutefor Advanced Medicine at the University of Manitoba, Winnipeg, and the UBC MRI ResearchCentre at the University of British Columbia, Vancouver. All volunteers provided signed consentbefore participating under the approvals of both The University of Manitoba Biomedical ResearchEthics Board and The University of British Columbia Clinical Research and Ethics Board.3.5.2 MRI ExperimentThe MRI experiments conducted in this study can broadly be divided into two categories: the mainstudy which included 3 healthy volunteers being scanned at two different sites, and additionalsub-studies investigating specific differences observed in the results from the main study. Forconsistency, the methods and results for all experiments are collated in the Methods and Resultssections, respectively.Main StudyEach subject was scanned at two sites using two different 3 T MRI scanners: Philips Achieva,Philips Medical Systems, Best, The Neatherlands, and Siemens Verio, Siemens Healthcare, Erlan-gen, Germany. Both scanners were equipped with an equivalent implementation of a 3D gradient-and spin-echo (GRASE) sequence for acquisition of multi-echo MWI data1. T2 relaxation for MWI1GRASE is commonly referred to as Turbo Gradient Spin Echo (TGSE) on Siemens scanners. However, the sequencewill be referred to as GRASE throughout this thesis.74Siemens PhilipsEcho Spacing 10 ms 10 msTR 1030 ms 1030 msfield of view (FOV) (AP,RL,FH) 240x180x110 mm3 240x180x110 mm3Acquired Resolution 1.5x1.5x5 mm3 1.5x1.5x5 mm3Reconstructed Resolution 1.5x1.5x5 mm3 1x1x2.5 mm3Averages 1 1Fat suppression Yes NoExcitation flip angle 90◦ 90◦Refocusing flip angle 180◦ 180◦Partial Fourier Reconstruction 6/8 NoSENSE Factor - 2EPI Factor 3 3Table 3.1: Pulse sequence parameters used for the two scanners. (∗ Fat-saturation consistedof a spectral fat saturation pulse at 3.4 ppm off the main water peak resonance frequency.)was collected at both sites using a whole-brain 3D GRASE sequence, parameters are outline in table3.1.To accelerate data acquisition. the Philips sequence utilized parallel imaging with sensitiv-ity encoding (SENSE) reconstruction for accelerated imaging, with a SENSE factor of 2[63]. TheSiemens sequence was unable to utilize SENSE and 6/8 partial Fourier reconstruction in the sliceencoding direction was used instead. Simply put, when SENSE is used, the acquisition is dividedbetween a set of coils elements in the receiver coil array. The data that is acquired from each coilcan be combined to an image, by beforehand calculating the coil profile. This will reduce acqui-sition time since each coil can sample a less dense k-space. With a SENSE factor of 2, the FOVfrom each coil is half of the desired FOV, but by combining the scans from each set of coils throughclever reconstruction algorithms, the full FOV can be recovered. Partial Fourier reconstruction ex-ploits symmetry in k-space. Instead of acquiring data symmetrically in k-space, the acquisitiononly covers parts of k-space. For instance, using partial k-space in the slice encoding direction ina 3D acquisition would mean that data is collected from kzmin to kzmax where kzmin 6= −kzmax. The75missing parts of k-space are assumed to be symmetric around kz = 0. Since these acquisition set-tings only affect reconstruction, and read out is unaffected, we assume that it will have insignificanteffect on quantitative T2 data analysis.An anatomical scan was also acquired on the Siemens scanner using a 3D magnetization pre-pared rapid acquisition gradient echo (MPRAGE) sequence (echo time (TE)=2.5 ms, TR=1900 ms,TI=900 ms, reconstructed resolution=1x0.5x0.5 mm3). This scan was used for registration pur-poses as outlined in the next section.The Effect of Varying the Refocusing Flip AngleBoth scanners had the refocusing flip angle prescribed to 180◦ , but due to differences in hardwareand pulse sequence implementation, this cannot always be guaranteed. Initial analysis of the datafrom the main study suggested that the refocusing flip angle was different between the two scanners.The influence of flip angle on MWF was studied by Prasloski et al. [61] who reported a minorincrease in MWF for a refocusing flip angle around 150◦ . We set out to verify how this wouldinfluence the results using both computer simulations and MRI experiments on the Philips scanner.Computer simulations were carried out using the following method: a set of simulated decaycurves was generated using the extended phase graph (EPG) algorithm for a range of 40 refocusingflip angles between 100-180◦ . Each decay curve was generated from two T2 components withT2 of 20 ms and 80 ms and amplitude Amyelin and AIE , respectively. For each flip angle, five T2combinations were evaluated: Amyelin = 0.05,0.0875,0.125,0.1625,0.2 and AIE = 1−Amyelin. Forfurther details about the EPG algorithm, see section 1.3.3.MRI experiments were carried out by scanning subject number 3 on the Philips scanner andsetting the prescribed flip angle to 140◦ , 160◦ , and 180◦ . This was done to verify results obtainedfrom the computer simulations. Please refer to section A.1 for detailed information about how tochange the refocusing flip angle on the Philips scanner.76Effect of Fat SaturationFat saturation pre-pulses are used with some pulse sequences to reduce the signal from fat in theimaging volume. The most common fat saturation technique on Philips scanners is spectral presat-uration with inversion recovery (SPIR). An off-resonance inversion pulse is applied to only affectfat followed by a strong spoiler gradient to destroy any transverse magnetization. The excitationpulse is applied at an inversion time (TI) after the inversion pulse when the longitudinal magneti-zation from fat is zero. This will effectively attenuate the signal from fat for the imaging sequence.Another technique for fat saturation, available on Siemens scanners, is Spectral Fat Saturation. Thistechnique is similar to SPIR: an off resonance radio frequency (RF) pulse is applied to selectivelyexcite fat followed by a spoiler gradient to destroy the transverse magnetization from fat. It differsfrom SPIR in that there is no inversion time to also destroy the longitudinal magnetization. Theadvantage of this is that the fat saturation process is shorter.In practice, fat saturation is very similar to applying a magnetization transfer pulses prior to theimaging sequence. Vavasour et al. [80] showed that this will attenuate the short T2 component morethan the long T2 component. This should then result in a decreased MWF. The MRI experimentsin that study used a Carr Purcell Meiboom Gill (CPMG) sequence on a 1.5 T scanner, therefore theeffects on a GRASE sequence on a 3 T scanner might not be analogous.GRASE scans acquired on the Siemens scanner in the main study used fat-suppression by de-fault. However, fat suppression was not used on the Philips scanner. Therefore, the effect of fatsaturation on the Philips scanner was tested. Subject 3 was scanned again on the Philips scannerusing an identical GRASE protocol as the main study but with the addition of SPIR fat suppression.To match the fat suppression used on the Siemens scanner a 400 Hz off resonance pulse should beused but due to limitations on the scanner the frequency was set to 360 Hz off resonance from thewater peak.773.5.3 Image AnalysisMulti-echo data from the GRASE sequence was analyzed voxel by voxel using in-house scriptsimplementing the analysis algorithm outlined in section 1.3.3 in MATLAB (Mathworks, NatickMA, USA). The distribution of T2 times associated to myelin water was chosen to be T2< 40 ms.MWF, flip angle (alpha), and geometric mean T2 (GMT2) were saved from the analysis. T2 decaycurves and distributions were obtained from isolated ROI for qualitative comparison between thescanners as well. One scan was also analyzed without using stimulated echo correction. This resultsin skipping the step where the optimal flip angle is calculated and instead each voxel is assigneda flip angle of 180◦ . Previous studies have shown that the MWF should decrease significantly asthe flip angle is lowered when stimulated echo correction is not used[61]. The goal of this was toverify that the flip angle approximation done by the EPG algorithm indeed is caused by a low flipangle. If this is the case then the MWF should be significantly reduced when the EPG analysis is notused.If the intra-extra cellular T2 peak in the T2 distribution falls inside the window of T2 timesassociated with myelin water, this will influence the MWF. To ensure that this is not affecting theresults, scans from both sites for one subject were analyzed by setting the upper limit of the myelinwater T2 window to 25, 30, 35, and, the default, 40 ms.A set of ROI were drawn manually on a center slice of the high resolution anatomical MPRAGE,see figure 3.2. Whole brain white matter (WM) and gray matter (GM) ROI were obtained using fslFAST[95] on the MPRAGE. Output images from multi-echo analysis were transformed to MPRAGEspace using a linear transformation, as implemented in fsl FLIRT[35].3.5.4 Statistical AnalysisLinear and orthogonal regression was carried out for the MWF values obtained from the isolatedROI outlined in figure 3.2. Bland-Altman analysis was used to investigate if there was an offset78Figure 3.2: Overview of all the manually drawn ROI on a single slice from the MPRAGE scan.in measured MWF between the two centers. Statistical analysis was carried out using MATLAB(Mathworks, Natick MA, USA) and R Studio (RStudio Inc, Boston MA, USA).3.6 Results3.6.1 Myelin Water FractionMyelin estimates from both scanners were highly correlated, see linear fit in figure 3.3a, but theSiemens scanner produced significantly higher MWF values with an average difference of 0.046(p < 0.001) measured over all ROI. Visual inspection of the MWF maps reveal clear differencesbetween the two scanners, see figure 3.4. Histograms of the MWF from both scanners in WM and79(a) (b)Figure 3.3: (a) Linear relationship between Siemens and Philips-acquired MWF data. Linearregression parameters (95% confidence interval): slope=0.84 (0.67, 1.01), y-intercept:-0.023 (-0.051, 0.0048), R2 = 0.82. Orthogonal regression: slope=0.91 (0.67, 1.16),y-intercept=-0.034 (-0.068, -0.00086). (b) Bland-Altman analysis comparing MWF re-sults obtained from the Siemens and Philips scanners. Mean difference=-0.043, 95%confidence interval=(0.0081,-0.095)GM are shown in figure 3.5. The GM histograms obtained from the Siemens scanner shows thatthere are significantly fewer voxels with low MWF compared to the Philips scanner. This can alsobe seen in figure 3.4 where the MWF maps from the Siemens scanner appear to have a smallerdynamic range. The shape of the histograms for WM are more similar between the two scanners,with the Siemens data being shifted slightly towards higher values.3.6.2 Geometric Mean T2The geometric mean T2 is a measure of the mean T2 time of the intra-extra cellular peak on alogarithmic scale. Maps of the GMT2 are shown in figure 3.4. Histograms for the same data areshown in figure 3.6 where the GMT2 is summarized for whole brain WM and GM ROI. The GMT2obtained from the Siemens scanner is significantly lower with an average difference of -3.6 ms(p< 0.001), calculated over all ROI.80Figure 3.4: Overview of MWF, GMT2, and flip angle from both centers in one subject.(a) (b) (c)Figure 3.5: MWF histograms from subject 1 (a), 2 (b), and 3 (c). white matter (WM) datashown in blue, gray matter (GM) in red. Data from Siemens is shown as solid lines andPhilips as dashed lines.81(a) (b) (c)Figure 3.6: GMT2 histograms from subject 1 (a), 2 (b), and 3 (c). white matter (WM) datashown in blue, gray matter (GM) in red. Data from Siemens is shown as solid lines andPhilips as dashed lines.3.6.3 Refocusing Flip AngleFlip angle maps estimated by the analysis algorithm from the GRASE acquisitions at both sitesare shown in figure 3.4. Data from the Siemens scanner shows a lower flip angle than the dataacquired with the Philips scanner. Histograms for the flip angle over the whole brain reveal a clearpattern for each scanner, see figure 3.7. The Siemens scanner consistently showed a peak in flipangle between 120-130◦ with a similar histogram shape for all three subjects. The Philips scannerproduced consistent histograms with a peak flip angle >170◦ .(a) (b) (c)Figure 3.7: Histograms for the estimated refocusing flip angle from both scanners for each ofthe three subjects, (a), (b), and (c).82(a) (b)Figure 3.8: (a) T2 decay curves and (b) T2 distributions obtained from the genu ROI fromboth scanners.3.6.4 T2 Decay Curves and DistributionsTo gain further understanding about the observed differences between the scanners, T2 decay curvesand distributions were analyzed. One example of T2 decay curves and distributions obtained fromthe MWI analysis is presented in figure 3.8. Data in this figure is obtained from the genu ROI.The two graphs presented in figure 3.8 shows the underlying cause of the statistics presented inhistogram in the previous sections. The oscillating pattern in the decay curve obtained from theSiemens data is characteristic of a lower flip angle (see chapter 1 section 1.3.3). In the T2 distribu-tions, it can be seen that the myelin peak appears to have the same amplitude from both scannersbut the the T2 peak from intra-extra cellular water is wider in the Philips scanner compared to theSiemens. If only the intra-extra cellular peak is wider, and the myelin peak remains constant, theMWF will decrease. Thus, the higher MWF observed in data obtained by the Siemens scanner couldbe explained by the narrower intra-extra cellular T2 peak.3.6.5 Simulations of Changing the Refocusing Flip AngleFigure 3.9a shows that, based on simulations, the effect of flip angle on MWF should be minor.MWF appeared to have a slight local minimum around 160◦ , but the variations were much smaller83than those observed in our in vivo study. Simulations also showed that given an input flip angle,the analysis algorithm can estimate the flip angle with high precision, see figure 3.9b. Only forinput flip angles close to 180◦ is there a deviation from the clear 1:1 relationship between input andoutput flip angle.(a) (b)Figure 3.9: Simulations showing how (a) MWF is affected by the flip angle with the true MWFin dashed lines and the MWF calculated with the EPG analysis in solid lines. In (b), therelationship between estimated flip angle and input flip angle is shown. It is a striking1:1 relationship for all flip angles except those close to 180◦ . Note that all 5 linesoverlap here and only the green line is visible.3.6.6 MR Experiment of Changing Refocusing Flip AngleManually changing the flip angle on the Philips scanner revealed similar flip angle histograms asthose acquired with the Siemens scanner, see figure 3.10. It can here be seen that the peak of theflip angle distribution is consistently lower than the input flip angle. Neither the MWF nor the GMT2were significantly affected by a lower flip angle as shown in figure 3.11. The slight variations thatcan be observed are not comparable to results obtained from the Siemens scanner shown in purplein the histograms in figure 3.11.84Figure 3.10: Graph showing the estimated flip angle distribution for the whole brain givenprescribed flip angle: α = 140,160,180◦ . The purple line shows the flip angle distri-bution acquired on the Siemens scanner from the same subject.(a) (b)Figure 3.11: Graphs showing variations in whole brain WM for (a) MWF and (b) for GMT2 forvarying flip angle (α).853.6.7 MWI Analysis Without Stimulated Echo CorrectionWhen the analysis was run without stimulated echo correction the measured MWF was significantlyreduced in the Siemens scanner, see figure 3.12 but barely affected in the data obtained with thePhilips scanner. This is in line with previous studies showing reduced MWF with lower flip anglesif stimulated echo correction is not performed[61] which lends credence to the flip angle estimatesproduced by the EPG algorithm.Figure 3.12: Without stimulated echo correction the MWF is significantly reduced in the datacollected by the Siemens scanner.3.6.8 Fat SuppressionUsing fat suppression on the Philips scanner had minor influence on the results. As seen by thehistograms in figure 3.13, MWF did not change significantly and there was only a slight change inthe flip angle distribution.86(a) (b)Figure 3.13: Results from MRI experiment with fat saturation pulse performed on the Philipsscanner with subject 3. (a) The MWF was not significantly affected, and the flip angledistribution as shown in (b) showed only a slight variation in shape.3.6.9 Effect of Myelin T2 WindowThe consistently lower GMT2 seen on the Siemens scanner (figure 3.6) could suggest that the intra-extra cellular T2 peak is leaking into the myelin peak window, which by default is set to 0−40 ms.If this is the case, the MWF measured by the Siemens scanner would decrease if the upper limit ofthe myelin window was decreased. To find out if this is the case, the data was reanalyzed with theupper limit of the myelin T2 window set to 25, 30, and 35 ms. This should be done with certainprecautions; if the window width is reduced too much, it could potentially also cut off the myelinwater T2 peak. We would therefore look for a plateau where the MWF does not decrease anymore.It the window is further decreased it would eventually reach the myelin peak and the MWF woulddecrease.Figure 3.14 shows how MWF distribution in WM is affected by the adjusted window size, andalso how the average MWF within each ROI is affected. Even though the average MWF is clearlydecreasing, as can be seen in the histogram (figure 3.14a), the change seems to be on the sameorder for both scanners, as seen by the ROI comparison (figure 3.14b).87(a) (b)Figure 3.14: (a) Full brain white matter histograms with varying myelin T2 window rangewith upper limit set to 30, 35, and 40 ms. (b) Difference in average MWF betweenSiemens and Philips within each ROI for varying T2 window.3.7 DiscussionThe measured difference in MWF between the Philips and Siemens scanners is consistent overall three subjects, as is the difference in flip angle distribution. This is most certainly due to asystematic difference in the data acquisition between the two scanners to which there can be manydifferent explanations. Throughout this present study several factors that potentially could causethis discrepancy have been investigated.3.7.1 Flip AngleThe first and most obvious clue is the difference in flip angle as seen in figure 3.4 and 3.7. Flipangles from the Siemens scanner were much lower, around 130◦ , than the Philips scanner whichaveraged around 170◦ . This is unlikely to be the underlying cause as it was shown in the presentstudy using computer simulations that the MWF remains almost constant over a wide range of flipangles. However, it should be noted that the estimated flip angle is not necessarily a true measure ofthe flip angle that is transmitted during the scan. The fitting algorithm chooses the flip angle whichis the best fit to the decay curve and it is therefore possible that some other factor is causing the88decay curve to appear as if the flip angle is lower. Future work should therefore include acquisitionof a flip angle map (B1 map) of the sample during the same scan session to validate the flip anglemaps generated by the algorithm.The MRI experiments carried out on the Philips scanner using reduced flip angle did not showany trend of increased MWF as seen in the Siemens scanner. However, the shape of the flip angledistribution, seen in the histogram in figure 3.10, carries great resemblance to that produced by theSiemens scanner shown in figure 3.7. The histogram generated by the 140◦ flip angle from thePhilips scanner in particular seems to line up with those produced by the Siemens scanner. Thissupports the idea that the refocusing flip angle that is transmitted at the Siemens scanner actuallyis around 140◦ .A noteworthy feature of the data from the Philips scanner is the flip angle profile in the xyplane. The flip angle increases from the edges but has a local minimum in the center of the brain.This is a phenomenon similar to phase wrap. The magnetization components after a refocusingpulse (M′) with an angle θ can be expressed asM′=Mx′My′Mz′=M00sinθcosθ . (3.1)Since only the transverse magnetization can be detected, the following signal will be measuredMT =√Mx′2+My′2 =√M2x +M2y · cos2 θ . (3.2)Since cos2 θ is symmetric around θ = 180◦ it is clear that the analysis cannot separate betweenα = 170◦ and α = 190◦ . Therefore, the local minimum appearing in the center of the brain islikely due to α > 180◦ . This will also affect the shape of the histograms. As the histogram willfold over at 180◦ , it will increase in amplitude for flip angles slightly lower than 180◦ . This is89clearly seen in figure 3.10 where the histogram for flip angle 180◦ has a significantly higher peakthan the other flip angles.3.7.2 Geometric Mean T2Further clues were provided by studying the GMT2. As seen in figure 3.6, the GMT2 is consistentlylower from the Siemens scanner. One hypothesis regarding this was that the intra-extra cellularpeak might be leaking into the myelin water window. By changing the upper cutoff for the myelinwindow it was shown that a lower cutoff resulted in a lower MWF in both scanners (figure 3.14).The fact that the MWF from both scanners still decreased about the same as seen by the ROI com-parison in figure 3.14b suggests that the slightly lower GMT2 from the Siemens scanner has aninsignificant effect on the results. However, it is still an interesting observation that the Siemensscanner appeared to estimate shorter T2 times than the Philips scanner. This was not observed incomputer simulations with varying flip angle, nor in MR experiments with adjusted flip angles.3.7.3 Fat Saturation Pre-pulseFat saturation can be beneficial for certain image sequences when signal from fat negatively im-pacts the appearance of the image. The use of fat saturation for MWI however is not completelyunderstood at this time. The study by Vavasour et al. [80] provided some support for the idea thata magnetization pre-pulse could affect the results significantly. However, magnetization pre-pulsessuch as those used by Vavasour et al. differ from standard fat saturation techniques. SPIR for in-stance uses a spectrally fat suppressing pulse together with an inversion recovery to null the signalfrom fat. However, if fat suppression would influence the T2 spectra it would result in decreasingthe amplitude of the short T2 component, thus lowering the MWF. This is contrary to the resultsobserved in the present study where the Siemens scanner, which used fat saturation, reports higherMWF values.90MRI experiments carried out on the Philips scanner using SPIR revealed no significant differencein the MWF results, figure 3.13. This could be due to the inversion time between the saturation pulseand excitation. The fat saturation technique used on the Siemens scanner has no inversion time,thus the magnetization transfer effect would be expected to be larger. Since the fat suppression onthe Philips and Siemens scanners differ significantly in their implementation, it cannot be taken forgranted that the spectral fat suppression used on the Siemens scanner does not influence the MWF.An interesting result from the experiment with fat suppression on the Philips scanner was thereduced amount of artifacts in the raw GRASE data. In figure 3.15, the first echo obtained from thePhilips scanner is shown, with and without SPIR. When the default setting is used, the subcutaneousfat around the brain gives a high signal, and some image artifact in the posterior part of the brainis visible. If SPIR is used, the signal from the subcutaneous fat is attenuated (as expected) and theimaging artifact also seems to be reduced. Considering the fact that neither the MWF not GMT2 waschanged, additional work should be dedicated to further investigation about the use of SPIR for fatsaturation in MWI using GRASE.As aforementioned, the estimated flip angle from the analysis is the flip angle associated withthe decay curve that best fits the data, and if some other factor would influence the decay curve,this could lead to estimating the wrong flip angle. The characteristic of a flip angle lower than 180◦is that the amplitude of the first echo is lower than the second echo. This was shown in chapter 1figure 1.6. It may be that the fat saturation used in the Siemens scanner reduces the signal in thefirst echo but that this is recovered in stimulated echoes from subsequence RF pulses in the secondecho.3.7.4 Implications for Future Multicenter StudiesIf the offset in MWF cannot be resolved it is possible to include the difference in a statistical model.It is common for large multicenter studies to include the scan site as a covariate in the statisticalanalysis[26, 76]. The most simple and straightforward approach would be to include the site as a91Figure 3.15: Subject 3 scanned on the Philips scanner with and without fat suppression(SPIR). Red arrows point out the different reduced signal from subcutaneous fat whenfat suppression is utilized. Green arrows point out image artifacts in the scan obtainedwithout fat suppression. Note how many of these artifacts are absent in the imageobtained with fat suppression.covariate and either subtract the difference (0.043) from the MWF maps obtained from the Siemensscanner or subtract half the difference from the MWF maps from the Siemens scanner and add halfthe difference to the MWF maps obtain from the Philips scanner. However, it should be noted thatthe observed difference has a rather large 95% confidence interval (-0.106, 0.013) which impliesthat the statistical model used in the study will be less sensitive. This will only affect cross sectionalstudies where a cohort split between two centers is combined. However, longitudinal changes foreach study participant is still possible.3.8 ConclusionsThe initial hypothesis for this project was that given an equivalent implementation of the GRASEsequence, the two scanners should produce equivalent MWF results. In the main study it becameapparent that there is a systematic difference in the data acquisition resulting in higher MWF values92from the Siemens scanner. To resolve this issue, additional sub-studies were performed to test theinfluence of various sequence parameters.The large discrepancy in flip angle was the most distinct difference between the two scanners.MRI experiments with the Philips scanner was able to reproduce the lower flip angle estimatedfrom the Siemens scanner, but no difference in MWF was observed. This was supported by com-puter simulations showing that that MWF is not expected to change with flip angle. However, ifstimulated echo analysis is not used, the MWF from the Siemens scanner is significantly decreasedas expected. Another potential difference was the use of fat saturation on the Siemens scanner. Anattempt to replicate this on the Philips scanner was performed, but again, no difference in MWF wasobserved. Further analysis of the T2 distributions revealed that the GMT2 is systematically lowerfrom the Siemens scanner. To ensure that the T2 window limits used for calculating the MWF didnot influence the results, the MWF was calculated with a range of window sizes. This revealed nodifference in MWF between the two scanners.The current GRASE protocols used for MWI have been thoroughly developed for the Philipsscanner at the Vancouver site. Additional work should now be dedicated to investigate potentialdifferences in the implementation of the GRASE sequence on the Siemens scanner that could causethe observed differences in MWF, flip angle, and GMT2. Further investigation is however hamperedby the fact that the scanners are located at two different sites (Winnipeg and Vancouver) whichmakes additional scan-rescan experiments complicated. Even though there is an offset in MWFbetween the two scanners, it is a encouraging to see that the correlation between the two scannersis strong.93Chapter 4Conclusions and Discussion4.1 Rapid Myelin Water Imaging in Human Cervical Spinal CordThe successful implementation of myelin water imaging (MWI) using gradient- and spin-echo(GRASE) for spinal cord MWI is an important step forward in the development of making MWIa mainstream application. With an acquisition time of less than 10 minutes, it is on par with otherquantitative magnetic resonance imaging (MRI) techniques for spinal cord imaging. Few studieshave utilized MWI for spinal cord imaging in clinical studies, but given the lower acquisition timeand still high reproducibility presented in this thesis it will hopefully be an attractive option forfuture studies.4.1.1 Impact of the ResultsThis study has shown that spinal cord MWI can be acquired using a 3D GRASE sequence with highreproducibility and in a clinically feasible time frame. Data is acquired with high resolution andwell preserved anatomical detail and contrast between white matter (WM) and gray matter (GM)can be observed on most scans. These results present a strong foundation for future spinal cordstudies to implement MWI using GRASE.944.1.2 LimitationsThe current implementation of the GRASE sequence for MWI in spinal cord only covers two verte-brae, thus limiting the scope of the analysis. The field of view (FOV) can be extended with the tradeoff of longer acquisition time. Another limiting factor in terms of FOV is scanning of individualswith high shoulders, typically caused by short muscles in the neck resulting in a shrugged posture.This can cause phase wrap and interfere with the imaging procedure.The GRASE is a gradient and radio frequency (RF) intense sequence commonly pushing thelimits for peripheral nerve stimulation (PNS) and specific absorption rate (SAR). Individuals withmetallic implants in the spine cannot be scanned with these high gradient slew-rates and SAR levelsand thus the sequence has to be specially adjusted for these individuals. However, changing scanparameters for certain individuals in a study will cause inconsistency in the data and may rendercross sectional analysis invalid.Due to the small size of the spinal cord and lack of anatomical contrast between other regionsthan WM and GM, segmentation of specific tracts is difficult and limited by the use of an atlas asused in this study. Even though segmentation is manually inspected, it is difficult to ensure that theROI are correctly placed. Furthermore, since the region of interest (ROI) are very small, down to 10voxels in an axial slice, ROI averages are strongly influence by partial volume effects along edge ofthe ROI. In the brain, many structures can be localized on a high resolution T1 or T2 weighted scan,but this is not possible in the spinal cord. This also ties into image registration between volumes.Compared to the structure of the brain, there are few landmarks for registration between images ofthe spinal cord. Here we use spinal cord segmentation for registration which, while robust, couldbe improved by also including anatomical landmarks.954.1.3 Future WorkThe next step in the development of MWI in the spinal cord is to scan patients with spinal cordpathology to verify that MWI using GRASE also detects abnormalities in myelin content. Wepropose a small study with 12 patients including: 4 multiple sclerosis (MS), 4 neuromyelitis op-tica (NMO), and 4 spinal cord injury (SCI) patients. In MS patients we would expect to find focallesions, extending less than one vertebrae with significant demyelination. Depending on the stageof the disease and current disease modifying therapy, longitudinal changes in myelin water frac-tion (MWF) and cross sectional area (CSA) could be detected as well. In NMO patients on the otherhand, extended regions of demyelination, extending more than one vertebrae, is expected. Depend-ing on which level the patients with SCI have the injury we could expect either focal or diffusechanges in myelin content.In terms of image acquisition, the effect of non-square acquisition voxels should be investi-gated. By reducing the number of phase encoding steps, acquisition time will be decreased and itshould not affect the possibility of susceptibility artifacts. It should be further investigated if it ispossible to develop a version of the GRASE sequence which is safe for patients with spinal cordimplants and still is within a clinically feasibly time frame. This is a particularly important stepas there are great applications for MWI in studies of SCI. This would also allow for a standardizedprotocol that can be used for all studies.Although this project did not focus on the image analysis in particular, it became evident thatbetter techniques for tissue segmentation in the spinal cord are needed. Recent work by Broschet al. [15] have shown very promising results for segmentation of structures in the brain usingdeep learning. Similar techniques should be tested for spinal cord GM segmentation to see if itcould increase the accuracy of segmentation. If good GM segmentation can be achieved, this wouldallow for more robust template registration by using the GM segmentation as a landmark in theregistration process.964.2 Multi-vendor Reproducibility of Myelin Water ImagingWhat at first appeared to be a straightforward reproducibility study of MWI turned into thoroughinvestigation of the underlying physics of the GRASE sequence and how various scan parameterscan affect the MWF. Even though the main underlying cause of the difference between the scannersis unknown, this study has investigated many potential factors and will limit the scope of potentialfactors to investigate in the future.4.2.1 Impact of the ResultsThis is the first study aiming to quantify the reproducibility of MWI using GRASE between twodifferent scan manufacturers. The results are highly correlated, providing a strong incentive tocontinue exploring this topic. It is desired to find the cause of the different MWF obtained fromthe two scanners. If this is not possible, the scan site can be included as a covariate in statisticalmodels for multicenter studies. Multi center and multi vendor reproducibility is an important stepforward in bringing MWI using GRASE to the mainstream.4.2.2 Limits of the ResultsThe main limitation in this study is that only two scanners were used. It is therefore not conclusivethat differences that are measured were due to major differences between the scan vendors oractually caused by individual variations between the scanners. It has previously been shown thatMWI using multi echo spin-echo (MSE) is reproducible between multiple Philips scanners of thesame model[55]. Therefore, it would be beneficial to carry out the same experiments on an identicalSiemens Verio 3 T at another site to determine if the observed differences are inherent to onescanner or if the results are reproducible between scanners. If the results are reproducible betweenSiemens scanners, that would be strong evidence that there is a difference in implementation of theGRASE sequence between Philips and Siemens that is causing the difference in MWF.974.2.3 Future WorkAt this stage, the observed difference in MWF obtained with the two scanners is not completelyunderstood and future work is required to unravel the cause. The first step will be to obtain detailedinformation about the GRASE sequence used by the Siemens scanner. In this study, only the basicparameters such as echo time (TE), repetition time (TR), and resolution were matched, with theassumption that other factors in the implementation of the sequence would not affect the results.However, given the outcome of the study and attempts to replicate the results on the Philips scan-ner, there seems to be a fundamental difference in the pulse sequence sequence implementation. Ifit is impossible to completely match the protocols, additional work should be dedicated to adaptingthe analysis algorithm to account for differences in the sequences. For instance, the current imple-mentation assumes that the first refocusing pulse is 180◦ and that the subsequent refocusing pulsescould be lower. It could be that the first RF pulse in the Siemens implementation of the GRASEsequence is lower than 180◦ , thus causing the analysis algorithm to fit a lower flip angle.In terms of hardware differences, the receiver head coil is the only part that can be changed.In this study a 12 channel channel head coil was used with the Siemens scanner while a 8 channelhead coil was used on the Philips scanner. Future work should investigate if the number of channelsin the head coil will affect the results, or the amount of imaging artifacts.Even though fat saturation is not recommended for MWI since magnetization transfer effectsare believed to affect the results, we observed less imaging artifacts when using spectral presatu-ration with inversion recovery (SPIR) on the Philips scanner. It was also noticed that no apparentdifference could be seen in the MWF histogram suggesting that SPIR will not affect the myelin mea-sures but does reduce artifacts by suppressing subcutaneous fat around the brain. This needs to befurther investigated in order to fully understand how SPIR affects the MWF. This could be a poten-tial development for better brain MWI. Once the effects of brain MWI with SPIR are investigated,98it should also be applied to spinal cord MWI as I see great potential to reduce imaging artifacts bysuppressing the large volume of fat outside the spine.4.3 Concluding RemarksMyelin water imaging (MWI) has been around for over 20 years and as new quantitative MRI tech-niques constantly emerge, MWI continues to develop to deliver higher resolution and shorter acqui-sition times. Through this work I have shown additional capabilities of using GRASE for MWI. Ihave extended the scope of applications for GRASE to spinal cord imaging and initiated the studyof multi-vendor reproducibility. Together these two studies constitute an important step in thedevelopment of MWI using GRASE.As the field of quantitative MRI keeps growing I hope that results from multiple techniquestogether will increase our understanding of the origin of the MRI signal in the central nervoussystem (CNS). This will directly improve clinical research aiming to study pathology in brain andspinal cord. It is important that the advances made in technical research, such as that presentedin this thesis, are translated into practical clinical applications, as that is what is driving the fieldforward. During the course of my thesis research I have been fortunate to work with a strong teamof physicists and clinicians. I believe it is integral to engage in translational research for projectslike this to reach clinical applications.This thesis had an initial exploratory approach and the scope of the thesis only became clearwhen several experiments had been carried out. In essence, it is very well summarized by theDanish physicist Lene Hau:Physics is about questioning, studying, probing nature. You probe, and, if you’relucky, you get strange clues.- Dr. Lene Hau99Bibliography[1] S. Balamoody, T. G. Williams, C. Wolstenholme, J. C. Waterton, M. Bowes, R. Hodgson,S. Zhao, M. Scott, C. J. Taylor, and C. E. Hutchinson. Magnetic resonance transverserelaxation time T2 of knee cartilage in osteoarthritis at 3-T: a cross-sectional multicentre,multivendor reproducibility study. Skeletal Radiol., 42(4):511–20, 2013. → pages 14[2] R. Bammer. Basic principles of diffusion-weighted imaging. Eur. J. Radiol., 45(3):169–84,2003. → pages 29[3] R. Bammer and F. Fazekas. Diffusion imaging of the human spinal cord and the vertebralcolumn. Top. Magn. Reson. Imaging, 14(6):461–476, 2003. → pages 29[4] C. M. Bauer, H. Jara, and R. Killiany. Whole brain quantitative T2 MRI across multiplescanners with dual echo FSE: Applications to AD, MCI, and normal aging. Neuroimage, 52(2):508–514, 2010. → pages 14[5] C. Beaulieu. The basis of anisotropic water diffusion in the nervous system - A technicalreview. NMR Biomed., 15(7-8):435–455, 2002. → pages 29[6] G. V. A. N. Belle and D. C. Martin. Sample Size as a Function of Coefficient of Variationand Ratio of Means. Am. Stat. Assoc., 47(3):165–167, 1993. → pages 63[7] M. Benhamou, V. Fonov, M. Taso, A. L. Troter, M. Sdika, L. Collins, V. Callot, andJ. Cohen-Adad. Atlas of white-matter tracts in the human spinal cord. Proc Intl Soc MagReson Med, 22:0013, 2014. → pages 52[8] M. A. Bernstein, K. F. King, and X. J. Zhou. Handbook of MRI Pulse Sequences. ElsevierScience, 2004. ISBN 9780080533124. → pages 2, 16, 17, 40[9] T. Billiet, S. Deprez, B. Maedler, R. Peeters, H. Zhang, A. Leemans, T. Dhollander,D. Christiaens, and F. Amant. Investigating the long-term effects of systemic chemotherapyon brain white matter using multi-shell diffusion MRI and myelin. Proc. Intl. Soc. Mag.Reson. Med., 22, 2014. → pages 73100[10] T. Billiet, M. Vandenbulcke, B. Ma¨dler, R. Peeters, T. Dhollander, H. Zhang, S. Deprez,B. R. Van den Bergh, S. Sunaert, and L. Emsell. Age-related microstructural differencesquantified using myelin water imaging and advanced diffusion MRI. Neurobiol. Aging, 36(6):2107–2121, 2015. → pages 25, 73[11] T. a. Bjarnason, C. R. McCreary, J. F. Dunn, and J. R. Mitchell. Quantitative T2 analysis:The effects of noise, regularization, and multivoxel approaches. Magn. Reson. Med., 63(1):212–217, 2010. → pages 28[12] F. Bloch. Nuclear induction. Phys. Rev., 70(7-8):460–474, 1946. → pages 4[13] N. Bloembergen, E. Purcell, and R. Pound. Relaxation effects in nuclear magnetic resonanceabsorption. Phys. Rev., 73(7):679–715, 1948. → pages xi, 7[14] G. Bravo and L. Potvin. Estimating the reliability of continuous measures with cronbach’salpha or the intraclass correlation coefficient: Toward the integration of two traditions. J.Clin. Epidemiol., 44(4-5):381–390, 1991. → pages 54[15] T. Brosch, L. Tang, Y. Yoo, D. Li, A. Traboulsee, and R. Tam. Deep 3D ConvolutionalEncoder Networks with Shortcuts for Multiscale Feature Integration Applied to MultipleSclerosis Lesion Segmentation. IEEE Trans. Med. Imaging, 0062(c):1–1, 2016. → pages 96[16] Cancer Research UK. Diagram of the spinal cord CRUK, 2014. → pages xii, 13[17] C. Chia, T. Bjarnason, A. MacKay, and G. Pike. Cross-site reproducibility of myelin waterestimates. Proc ISMRM, 293(i):2520, 2006. → pages 72[18] J. Cohen-Adad, B. De Leener, M. Benhamou, D. Cadotte, D. Fleet, A. Cadotte, M. G.Fehlings, J. P. P. Paquette, W. Thong, M. Taso, and Others. Spinal Cord Toolbox: anopen-source framework for processing spinal cord MRI data. In Proc. 20th Annu. Meet.OHBM, Hamburg, Ger., page 3633, 2014. → pages 52[19] S. C. L. Deoni and S. H. Kolind. Investigating the stability of mcDESPOT myelin waterfraction values derived using a stochastic region contraction approach. Magn. Reson. Med.,73(1):161–169, 2015. → pages 29[20] S. C. L. Deoni, B. K. Rutt, T. Arun, C. Pierpaoli, and D. K. Jones. Gleaning multicomponentT1 and T2 information from steady-state imaging data. Magn. Reson. Med., 60(6):1372–1387, 2008. → pages 27, 28[21] D. A. Feinberg and K. Oshio. GRASE (gradient- and spin-echo) MR imaging: a new fastclinical imaging technique. Radiology, 181(2):597–602, 1991. → pages 15[22] C. R. Figley and P. W. Stroman. Investigation of human cervical and upper thoracic spinalcord motion: Implications for imaging spinal cord structure and function. Magn. Reson.Med., 58(1):185–189, 2007. → pages 35101[23] C. R. Figley, D. Yau, and P. W. Stroman. Attenuation of Lower-Thoracic, Lumbar, andSacral Spinal Cord Motion: Implications for Imaging Human Spinal Cord Structure andFunction. Am. J. Neuroradiol., 29(8):1450–1454, 2008. → pages 35, 38[24] S. W. Flynn, D. J. Lang, A. L. Mackay, V. Goghari, I. M. Vavasour, K. P. Whittall, G. N.Smith, V. Arango, J. J. Mann, A. J. Dwork, P. Falkai, and W. G. Honer. Abnormalities ofmyelination in schizophrenia detected in vivo with MRI, and post-mortem with analysis ofoligodendrocyte proteins. Mol Psychiatry, 8(9):811–820, 2003. → pages 25[25] V. Fonov, A. Le Troter, M. Taso, B. De Leener, G. Le´veˆque, M. Benhamou, M. Sdika,H. Benali, P.-F. Pradat, D. Collins, V. Callot, and J. Cohen-Adad. Framework for integratedMRI average of the spinal cord white and gray matter: The MNIPolyAMU template.Neuroimage, 102:817–827, 2014. → pages 35, 52[26] L. Friedman, G. H. Glover, and T. F. Consortium. Reducing interscanner variability ofactivation in a multicenter fMRI study: Controlling for signal-to-fluctuation-noise-ratio(SFNR) differences. Neuroimage, 33(2):471–481, 2006. → pages 91[27] P. J. Gareau, B. K. Rutt, S. J. Kariikb, and J. R. Mitchebb. Magnetization Transfer andMulti-component T2 Measurements with Histopathologic Correlation in an ExperimentalModel of Multiple Sclerosis. 8:2000, 2000. → pages 33[28] F. Grussu, T. Schneider, H. Zhang, D. C. Alexander, and C. A. M. Wheeler-kingshott. Invivo estimation of neuronal orientation dispersion and density of the human spinal cord.Neuroimage, 111:590–601, 2015. → pages 32[29] H. Gudbjartsson and S. Patz. The Rician distribution of noisy MRI data. Magn. Reson.Med., 34(6):910–914, 1995. → pages 67[30] E. M. Haacke, R. W. Brown, M. R. Thompson, and R. Venkatesan. Magnetic ResonanceImaging: Physical Principles and Sequence Design. Wiley, 1999. ISBN 9780471351283.→ pages 2[31] R. M. Henkelman, G. J. Stanisz, and S. J. Graham. Magnetization transfer in MRI: Areview. NMR Biomed., 14(2):57–64, 2001. → pages 32[32] J. Hennig. Multiecho imaging sequences with low refocusing flip angles. J. Magn. Reson.,78(3):397–407, 1988. → pages 22[33] M. C. Henry-Feugeas, I. Idy-Peretti, B. Blanchet, D. Hassine, G. Zannoli, andE. Schouman-Claeys. Temporal and spatial assessment of normal cerebrospinal fluiddynamics with MR imaging. Magn. Reson. Imaging, 11(8):1107–1118, 1993. → pages 38[34] Q. Jarosz. Diagram of Neuron. → pages xii, 13102[35] M. Jenkinson and S. Smith. A global optimisation method for robust affine registration ofbrain images. Med. Image Anal., 5(2):143–156, 2001. → pages 78[36] H. Kearney, D. H. Miller, and O. Ciccarelli. Spinal cord MRI in multiple sclerosisdiagnostic,prognostic and clinical value. Nat. Rev. Neurol., 11(6):327–338, 2015. → pages 48[37] G. Kim, F. Khalid, V. V. Oommen, S. Tauhid, R. Chu, M. A. Horsfield, B. C. Healy, andR. Bakshi. T1- vs. T2-based MRI measures of spinal cord volume in healthy subjects andpatients with multiple sclerosis. BMC Neurol., 15(1):124, 2015. → pages 64[38] S. Kolind, A. Seddigh, A. Combes, B. Russell-Schulz, R. Tam, V. Yogendrakumar, S. Deoni,N. A. Sibtain, A. Traboulsee, S. C. R. Williams, G. J. Barker, and P. A. Brex. Brain and cordmyelin water imaging: A progressive multiple sclerosis biomarker. NeuroImage Clin., 9:574–580, 2015. → pages 12, 28, 35[39] S. H. Kolind and S. C. Deoni. Rapid three-dimensional multicomponent relaxation imagingof the cervical spinal cord. Magn. Reson. Med., 65(2):551–6, 2011. → pages xiii, 27, 28, 35,36, 62[40] C. Laule, I. M. Vavasour, G. R. W. Moore, J. Oger, D. K. B. Li, D. W. Paty, and A. L.MacKay. Water content and myelin water fraction in multiple sclerosis. J. Neurol., 251(3):284–293, 2004. → pages 25[41] C. Laule, I. M. Vavasour, J. D. Vavasour, J. Oger, D. W. Paty, A. L. Mackay, and D. K. Li.Cervical Cord Abnormalities in Primary Progressive Multiple Sclerosis : Atrophy andMyelin Water Changes at Baseline. Baseline, 10(2):2004–2004, 2004. → pages 58[42] C. Laule, E. Leung, D. K. B. Li, A. L. Traboulsee, D. W. Paty, A. L. MacKay, and G. R. W.Moore. Myelin water imaging in multiple sclerosis: quantitative correlations withhistopathology. Mult. Scler., 12(6):747–753, nov 2006. → pages xiii, 18, 25, 26, 28[43] C. Laule, P. Kozlowski, E. Leung, D. K. B. Li, A. L. MacKay, and G. R. W. Moore. Myelinwater imaging of multiple sclerosis at 7T: Correlations with histopathology. Neuroimage, 40(May 2005):1575–1580, 2008. → pages 18[44] C. Laule, I. M. Vavasour, Y. Zhao, a. L. Traboulsee, J. Oger, J. D. Vavasour, a. L. Mackay,and D. K. B. Li. Two-year study of cervical cord volume and myelin water in primaryprogressive multiple sclerosis. Mult. Scler., 16(6):670–7, 2010. → pages 25, 63[45] C. Laule, I. M. Vavasour, Y. Zhao, A. L. Traboulsee, J. Oger, J. D. Vavasour, A. L. Mackay,and D. K. B. Li. Two-year study of cervical cord volume and myelin water in primaryprogressive multiple sclerosis. Mult. Scler., 16(6):670–677, 2010. → pages 35[46] I. R. Levesque, C. L. L. Chia, and G. B. Pike. Reproducibility of in vivo magnetic resonanceimaging-based measurement of myelin water. J. Magn. Reson. Imaging, 32(1):60–68, 2010.→ pages 18103[47] E. Ljungberg, S. H. Kolind, R. Tam, M. S. Freedman, D. K. Li, and A. L. Traboulsee.Correcting Cervical Spinal Cord Area for Cord Length Strengthens Correlation BetweenAtrophy and Ambulation Motivation. In Eur. Congr. Treat. Res. MS, Barcelona, 2014. →pages 35[48] S. Ljunggren. A simple graphical representation of Fourier-based imgaing methods. J.Magn. Reson., 54:338–343, 1983. → pages 10[49] C. Lukas, M. H. Sombekke, B. Bellenberg, H. K. Hahn, V. Popescu, K. Bendfeldt, E. W.Radue, A. Gass, S. J. Borgwardt, L. Kappos, Y. Naegelin, D. L. Knol, C. H. Polman, J. J. G.Geurts, F. Barkhof, and H. Vrenken. Relevance of spinal cord abnormalities to clinicaldisability in multiple sclerosis: MR imaging findings in a large cohort of patients.Radiology, 269(2):542–52, 2013. → pages 35[50] A. Mackay, K. Whittall, J. Adler, D. Li, D. Paty, and D. Graeb. In vivo visualization ofmyelin water in brain by magnetic resonance. Magn. Reson. Med., 31(6):673–677, 1994. →pages 14, 15, 21, 25, 45[51] E. L. MacMillan, B. Ma¨dler, N. Fichtner, M. F. Dvorak, D. K. B. Li, A. Curt, and A. L.MacKay. Myelin water and T2 relaxation measurements in the healthy cervical spinal cordat 3.0T: Repeatability and changes with age. Neuroimage, 54(2):1083–1090, 2011. → pages18, 27, 28, 48, 49, 58, 61, 62[52] B. Ma¨dler and A. MacKay. In-vivo 3D Multi-component T2-Relaxation Measurements forQuantitative Myelin Imaging at 3T. Proc. 14th Sci. Meet. Int. Soc. Magn. Reson. Med., 11(2004):2112, 2006. → pages 15[53] B. Ma¨dler and a. MacKay. Towards whole brain myelin imaging. Proc. Intl. Soc. Mag.Reson. Med, 759(2005):1973, 2007. → pages 18[54] S. Mallik, R. S. Samson, C. A. M. Wheeler-Kingshott, and D. H. Miller. Imaging outcomesfor trials of remyelination in multiple sclerosis. J. Neurol. Neurosurg. Psychiatry, 85(12):1396–1404, 2014. → pages 12[55] S. M. Meyers, I. M. Vavasour, B. Ma¨dler, T. Harris, E. Fu, D. K. Li, A. L. Traboulsee, A. L.MacKay, and C. Laule. Multicenter measurements of myelin water fraction and geometricmean T2 : Intra- and intersite reproducibility. J. Magn. Reson. Imaging, 38(6):1445–1453,2013. → pages 62, 72, 97[56] S. Mikula, J. M. Stone, and E. G. Jones. BrainMaps.org - Interactive High-ResolutionDigital Brain Atlases and Virtual Microscopy. Brains. Minds Media, 3:bmm1426–bmm1426, 2008. → pages xii, 13[57] E. P. Minty, T. a. Bjarnason, C. Laule, and A. L. Mackay. Myelin water measurement in thespinal cord. Magn. Reson. Med., 61(4):883–892, 2009. → pages 15, 27, 58, 67104[58] J. Oh, E. T. Han, M. C. Lee, S. J. Nelson, and D. Pelletier. Multislice brain myelin waterfractions at 3T in multiple sclerosis. J. Neuroimaging, 17(2):156–163, 2007. → pages 25[59] N. Papinutto, R. Schlaeger, V. Panara, A. H. Zhu, E. Caverzasi, W. A. Stern, S. L. Hauser,and R. G. Henry. Age, gender and normalization covariates for spinal cord gray matter andtotal cross-sectional areas at cervical and thoracic levels: A 2D phase sensitive inversionrecovery imaging study. PLoS One, 10(3):1–18, 2015. → pages 35[60] C. H. Polman, S. C. Reingold, B. Banwell, M. Clanet, J. A. Cohen, M. Filippi, K. Fujihara,E. Havrdova, M. Hutchinson, L. Kappos, F. D. Lublin, X. Montalban, P. O’Connor,M. Sandberg-Wollheim, A. J. Thompson, E. Waubant, B. Weinshenker, and J. S. Wolinsky.Diagnostic criteria for multiple sclerosis: 2010 Revisions to the McDonald criteria. Ann.Neurol., 69(2):292–302, 2011. → pages 12[61] T. Prasloski, B. Ma¨dler, Q. S. Xiang, A. MacKay, and C. Jones. Applications of stimulatedecho correction to multicomponent T2 analysis. Magn. Reson. Med., 67(6):1803–1814,2012. → pages 22, 49, 51, 56, 76, 78, 86[62] T. Prasloski, A. Rauscher, A. L. MacKay, M. Hodgson, I. M. Vavasour, C. Laule, andB. Ma¨dler. Rapid whole cerebrum myelin water imaging using a 3D GRASE sequence.Neuroimage, 63(1):533–539, oct 2012. → pages 15, 18, 27, 48, 49, 60, 61, 72[63] K. P. Pruessmann, M. Weiger, M. B. Scheidegger, and P. Boesiger. SENSE: Sensitivityencoding for fast MRI. Magn. Reson. Med., 42(5):952–962, 1999. → pages 75[64] Roadnottaken. Myelinated neuron, 2007. → pages xii, 13[65] M. D. Robson, P. D. Gatehouse, M. Bydder, and G. M. Bydder. Magnetic Resonance: AnIntroduction to Ultrashort TE (UTE) Imaging. J. Comput. Assist. Tomogr., 27(6), 2003. →pages 6[66] R. A. Rudick. A clinically meaningful measure of disability. Am. Acad. Neurol., 74(17):S8–S15, 2010. → pages 28[67] K. Schmierer, F. Scaravilli, D. R. Altmann, G. J. Barker, and D. H. Miller. Magnetizationtransfer ratio and myelin in postmortem multiple sclerosis brain. Ann. Neurol., 56(3):407–415, 2004. → pages 32, 33[68] J. Schneck. The role of magnetic susceptibility in magnetic resonance imaging: MRImagnetic compatibility of the first and second kinds. Med. Phys., 23(6):815–850, 1996. →pages 43[69] R. Sokal and F. Rohlf. Biometry: the principles and practice of statistics in biologicalresearch. Freeman and Co, New York, 4 edition, 2012. → pages 54105[70] S.-K. Song, S.-W. Sun, M. J. Ramsbottom, C. Chang, J. Russell, and A. H. Cross.Dysmyelination Revealed through MRI as Increased Radial (but Unchanged Axial)Diffusion of Water. Neuroimage, 17(3):1429–1436, 2002. → pages 30[71] S.-K. Song, S.-W. Sun, W.-K. Ju, S.-J. Lin, A. H. Cross, and A. H. Neufeld. Diffusion tensorimaging detects and differentiates axon and myelin degeneration in mouse optic nerve afterretinal ischemia. Neuroimage, 20(3):1714–1722, 2003. → pages 30[72] C. Statistics Canada. Neurological conditions, by age group and sex, household populationaged 0 and over, 2010/2011, occasional (number unless otherwise noted), 2011. → pages 25[73] V. L. Stevenson, S. M. Leary, N. A. Losseff, G. J. M. Parker, G. J. Barker, Y. Husmani, D. H.Miller, and A. J. Thompson. Spinal cord atrophy and disability in MS A longitudinal study.Neurology, 51(1):234–238, 1998. → pages 35[74] P. W. Stroman, C. Wheeler-Kingshott, M. Bacon, J. M. Schwab, R. Bosma, J. Brooks,D. Cadotte, T. Carlstedt, O. Ciccarelli, J. Cohen-Adad, a. Curt, N. Evangelou, M. G.Fehlings, M. Filippi, B. J. Kelley, S. Kollias, a. Mackay, C. a. Porro, S. Smith, S. M.Strittmatter, P. Summers, and I. Tracey. The current state-of-the-art of spinal cord imaging:Methods. Neuroimage, 84:1070–1081, 2014. → pages 34, 47[75] P. Summers, P. Staempfli, T. Jaermann, S. Kwiecinski, and S. Kollias. A preliminary studyof the effects of trigger timing on diffusion tensor imaging of the human spinal cord. Am. J.Neuroradiol., 27(9):1952–1961, 2006. → pages 65[76] S. J. Teipel, M. Wegrzyn, T. Meindl, G. Frisoni, A. L. W. Bokde, A. Fellgiebel, M. Filippi,H. Hampel, S. Klo¨ppel, K. Hauenstein, and M. Ewers. Anatomical MRI and DTI in thediagnosis of Alzheimer’s disease: A european multicenter study. J. Alzheimer’s Dis., 31(SUPPL. 3):33–47, 2012. → pages 91[77] D. B. Twieg. The k-trajectory formulation of the NMR imaging process with applications inanalysis and synthesis of imaging methods. Med. Phys., 10(5):610–621, 1983. → pages 10[78] UCLA Interactive Neuroscience Collection. The Heal National Digital Library. → pagesxii, 13[79] I. M. Vavasour, K. P. Whittall, A. L. MacKay, D. K. B. Li, G. Vorobeychik, and D. W. Paty.A comparison between magnetization transfer ratios and myelin water percentages innormals and multiple sclerosis patients. Magn. Reson. Med., 40(5):763–768, 1998. → pages33, 34[80] I. M. Vavasour, K. P. Whittall, D. K. Li, and a. L. MacKay. Different magnetization transfereffects exhibited by the short and long T2 components in human brain. Magn. Reson. Med.,44(6):860–6, 2000. → pages 68, 77, 90106[81] I. M. Vavasour, C. Laule, D. K. B. Li, J. Oger, G. R. W. Moore, A. Traboulsee, and A. L.MacKay. Longitudinal changes in myelin water fraction in two MS patients with activedisease. J. Neurol. Sci., 276(1-2):49–53, 2009. → pages 25[82] I. M. Vavasour, C. Laule, D. K. B. Li, A. L. Traboulsee, and A. L. MacKay. Is themagnetization transfer ratio a marker for myelin in multiple sclerosis? J. Magn. Reson.Imaging, 33(3):713–718, 2011. → pages 25, 33[83] I. M. Vavasour, S. M. Meyers, E. L. MacMillan, B. Ma¨dler, D. K. B. Li, A. Rauscher,T. Vertinsky, V. Venu, A. L. MacKay, and A. Curt. Increased spinal cord movements incervical spondylotic myelopathy. Spine J., 14(10):2344–2354, 2014. → pages 38, 48[84] C. Wang, R. C. Tam, E. Mackie, D. K. B. Li, and a. L. Traboulsee. Dehydration affectsspinal cord cross-sectional area measurement on MRI in healthy subjects. Spinal Cord, 52(8):616–20, 2014. → pages 63, 64[85] C. A. Wheeler-Kingshott and M. Cercignani. About axial and radial diffusivities. Magn.Reson. Med., 61(5):1255–1260, 2009. → pages 31[86] K. P. Whittall and A. L. MacKay. Quantitative interpretation of NMR relaxation data. J.Magn. Reson., 84(1):134–152, 1989. → pages 22, 23[87] A. D. Wright, M. Jarrett, I. Vavasour, E. Shahinfard, S. Kolind, P. van Donkelaar, J. Taunton,D. Li, and A. Rauscher. Myelin Water Fraction Is Transiently Reduced after a Single MildTraumatic Brain Injury A Prospective Cohort Study in Collegiate Hockey Players. PLoSOne, 11(2):e0150215, 2016. → pages 25, 73[88] Y. Wu, A. S. Field, A. L. Alexander, J. Fleming, H. Zhang, and I. D. Duncan. Quantificationof Myelin Water in Human Cervical Spinal Cord in vivo. 11:2004, 2004. → pages 58[89] M. C. Yiannakas, A. M. Mustafa, B. De Leener, H. Kearney, C. Tur, D. R. Altmann, F. DeAngelis, D. Plantone, O. Ciccarelli, D. H. Miller, J. Cohen-Adad, and C. A. GandiniWheeler-Kingshott. Fully automated segmentation of the cervical cord from T1-weightedMRI using PropSeg: Application to multiple sclerosis. NeuroImage Clin., 10:71–77, 2016.→ pages 52[90] Y. Yoo and R. Tam. Non-local spatial regularization of MRI T2 relaxation images for myelinwater quantification. In MICCAI, volume 8149, pages 614–621, 2013. → pages 24, 51[91] M. Zaitsev, J. Maclaren, and M. Herbst. Motion artifacts in MRI: A complex problem withmany partial solutions. J. Magn. Reson. Imaging, 2015. → pages 35[92] H. Zhang, T. Schneider, C. a. Wheeler-Kingshott, and D. C. Alexander. NODDI: Practical invivo neurite orientation dispersion and density imaging of the human brain. Neuroimage, 61(4):1000–1016, 2012. → pages 31, 32107[93] J. Zhang, S. H. Kolind, C. Laule, and A. L. MacKay. How does magnetization transferinfluence mcDESPOT results? Magn. Reson. Med., 74(5):1327–1335, 2015. → pages 28[94] J. Zhang, S. H. Kolind, C. Laule, and A. L. Mackay. Comparison of myelin water fractionfrom multiecho T2 decay curve and steady-state methods. Magn. Reson. Med., 73(1):223–232, 2015. → pages xiii, 29, 34, 73[95] Y. Zhang, M. Brady, and S. Smith. Segmentation of brain MR images through a hiddenMarkov random field model and the expectation-maximization algorithm. Med. Imaging,IEEE Trans., 20(1):45–57, 2001. → pages 78108Appendix ASupporting MaterialsA.1 Changing the flip angleThe current implementation of the gradient- and spin-echo (GRASE) sequence on the Philips scan-ner makes a distinction between the first and the 31 subsequent RF refocusing pulses. In thesequence development tool the first RF refocusing pulse is referred to as echo while the followingare referred to as ME. Only the ME is possible to change from the console and there referred to asthe refocusing pulse control. By default, it is set to 180◦ but for various reasons such as specificabsorption rate (SAR) issues, it can be lowered. To change all of the 32 refocusing RF pulses, bothflip angle parameters were adjusted in the sequence development tool.109A.2 Mathematical DerivationsA.2.1 Even-odd echo refocusingIn spin-echo echo-train sequences such as GRASE and multi echo spin-echo (MSE), flow insidethe imaging volume can cause an artifact known as even-odd echo refocusing. For stationaryspins, the phase dispersion will be 0 at every spin echo. However, if spins are moving at a constantvelocity, the phase dispersion will be non-zero for every other spin echo. To provide a mathematicaldescription, consider the readout gradient used in the GRASE sequence as shown in figure A.1. TheFigure A.1: Example of frequency encoding gradient lobes used for readout for each echo inGRASE and MSE.accumulated phase by a spin moving with a constant velocity under the influence of a gradient G(t)applied between time t1 and t2 can be expressed asφ =∫ t2t1G(t)γvtdt (A.1)110First, consider the read-out gradients used in the GRASE sequence shown in figure A.1. Let thephase accumulated at the spin echo, in the middle of the positive gradient lobe, be φ1. Then φ1 isφ1 =∫ (t3−t2)/2t1G(t)γvtdt =−∫ t2t1Gγvtdt+∫ (t3−t2)/2t2Gγvt =Gγv2(t21 −7t224− t2t32+t234). (A.2)Using the same method, the phase accumulated between the spin echo and the end of the read outgradient complex φ2 can be calculated asφ2 =∫ t4(t3−t2)/2G(t)γvtdt =∫ t3(t3−t2)/2Gγvtdt−∫ t4t3Gγvtdt =Gγv2(7t234+t2t32− t224− t24). (A.3)To simplify this, set the total duration of the read out gradient complex to τ . Then t2 = t1 + t3,t3 = t1+2τ/3, and t4 = t1+τ . Assuming that the three gradient lobes are of equal duration, φ1 andφ2 can then be expressed asφ1 =−Gγv2(4t1τ3+ t21 +7τ236)=−Gγv2((t1+2τ3)2− τ24)(A.4)φ2 =Gγv2(t21 +2τt13− 5τ236)=Gγv2((t1+τ3)2− τ24)(A.5)This result should then be compare to a simple rectangular read out gradient, as used in theMSE sequence, also shown in figure A.1. Using analogous reasoning, the accumulated phase at thecenter of the gradient lobe φ1, the time of the spin echo, for the MSE sequence is thenφ1 =Gγv2((t1+τ ′2)2− t21)=Gγv2(τ ′t1+τ ′24). (A.6)The phase accumulated between the spin echo and the end of the gradient φ2 is thenφ2 =Gγv2((t1+ τ ′)2−(t1+τ ′2))=Gγv2(τ ′t1+3τ ′24). (A.7)111We can now compare the total phase accumulated during the read out gradients from both theGRASE and MSE as φtot = φ1+φ2:GRASE φGRASE,tot = φ1+φ2 =−Gγv2(2t1τ3+τ23)(A.8)MSE φMSE,tot =Gγv2(2τ ′t1+ τ ′2)(A.9)In the example given in figure A.1 τ = 3τ ′. Thus, the difference in accumulated phase during theread out gradient by the two sequences is|φGRASE,tot |− |φMSE,tot |= Gγv2(2τ ′t1+9τ ′2− (2τ ′t1+ τ ′2))=Gγv2·2τ ′2 (A.10)The result of this shows us that a large phase dispersion can be expected in the GRASE sequencegiven that the conditions above hold. The phase dispersion causing an attenuation in the signalin odd echoes is proportional to the phase accumulated during the readout gradient, thus a largerdifference in detected signal between even and odd echoes can be expected for the GRASE sequence.It should be noted that the gradient waveforms are simplified in this example and that there typicallyare additional gradients applied before the first spin echo which could influence the even-odd echorefocusing pattern.112

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            data-media="{[{embed.selectedMedia}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0228626/manifest

Comment

Related Items