1. MSmcDESPOT A look at the road behind and ahead October 30, 2009

2. The Technique mcDESPOT (multi-component driven equilibrium single pulse observation of T1/T2) is a quantitative MR technique that characterizes many of the key parameters relevant to MRI A series of spoiled gradient echo (SPGR) and phase-cycled steady-state free precession (SSFP) scans are collected at different sets of flip angles The signal from a single voxel across all these scans is modeled as the combination of two different pools of water, a fast and slow pool in exchange with each other A fitting algorithm (stochastic region of contraction) computes the optimal set of parameters that characterizes the observed signal at each voxel in the brain

3. The Technique The final result is a set of 10 maps defining MR parameters throughout the entire brain: – Fast pool T1, T2, and residence time – Slow pool T1 and T2 – Single pool T1, T2, and M 0 – this is when we do not model each voxel as the sum of two pools – B0 off-resonance – Fast volume fraction – this is how much each pool contributes to a voxel’s signal or alternatively, what fraction of a voxel is occupied by each pool We attribute the fast pool to water trapped between the lipid bilayers of the myelin sheath, while the slower-relaxing species is believed to correspond to the less restricted intra- and extracellular pools – This needs further histological verification but we will continue under this premise – Thus we rename the fast volume fraction to the “myelin water fraction” (MWF), our key parameter of interest

4. The Study Given this technique which we believe can characterize myelination in the brain, we move our sights to examine a disease that is characterized by demyelination: multiple sclerosis 23 normals + 2 pending 25 MS patients, 5 in each of 5 classes (low-risk CIS, high-risk CIS, RR, SP, PP) Each scanned at 1.5T to avoid B1 inhomogeneity and assumed flip angle inaccuracy: – mcDESPOT protocol at 2mm 3 isotropic – 32-direction DTI sequence at 2.5mm 3 – T2/PD FSE at 0.43mm 2 in-plane and 6mm slice resolution – FLAIR at 0.86 mm 2 in-plane and 3mm slice resolution – MPRAGE pre and post Gd constrast for patients at 1mm 3

5. Preprocessing for mcDESPOT Prior to running the fitting algorithm, we must run the SPGR and SSFP images through a preprocessing pipeline All the following steps are achieved using the FMRIB Software Library (FSL) Throughout this presentation, we will focusing our attention on a single SPMS patient, affectionately known as P025

6. Preprocessing mcDESPOT – Step 1 1.Linear coregistration with trilinear interpolation (FSL FLIRT) – so that each voxel across all the images is the same piece of physical tissue

7. Preprocessing mcDESPOT – Step 2 2.Brain extraction from skull (FSL BET) – to reduce computation time

8. Preprocessing DTI – Step 1 Similarly, the diffusion weighted images must also be coregistered for eddy current correction and brain extracted

9. Preprocessing DTI – Step 2

10. Processing Now that the data is all prepared, it is run through the parameter fitting program The mcDESPOT volumes are processed with our own code The diffusion volumes are fitted with FSL’s dtifit

11. mcDESPOT Maps

12. mcDESPOT Maps - MWF

13. DTI Maps Fraction AnisotropyMean Diffusivity

14. Postprocessing Postprocessing involves bringing these various maps and scans into a standard space so that they can be compared with each other on a voxel per voxel basis We use the 2mm 2 MNI152 T1 standard space template and the 1mm 2 FMRIB58 FA map, an average of FA maps from 58 subjects, each nonlinearly registered to the MNI brain

15. Postprocessing – mcDESPOT The mcDESPOT coregistration target for each subject is nonlinearly registered to the MNI brain and this warp field is in turn applied to the 10 maps The warp field is found with FSL’s FNIRT using an 8mm 3 warp resolution

16. Standard Space Reg. – SPGR Target MNI MNI152 2mmmcDESPOT SPGR Registration Target

17. Postprocessing – DTI The DTI FA map for each subject is nonlinearly registered to the FMRIB58 map Alternatively, we could register it to the mcDESPOT target and use the already computed warp field

18. Standard Space Reg. – FMRIB58 FA

19. Standard Space Reg. – DTI FA

20. Postprocessing – Clinical Each clinical scan for each patient is linearly registered to the mcDESPOT target Then the target->MNI warp is applied

21. Analysis Whole brain MWF, z-score based thresholding Would like to move onto tissue-specific MWF study, particularly in these types: WM, GM, NAWM (normal-appearing white matter), NAGM, and lesions only This brings us to the tricky issue of segmentation

22. Current Issues – Segmentation Lesion segmentation is proving to be a very difficult task Ultimately we’d like to use the lesion mask to subtract from our other tissue classifications to produce “normal-appearing” tissues Here’s some pictures of our results so far using FSL’s FAST with a variety of channels (SPGR, FLAIR, T2, PD)

23. Clinical-MPRAGE

24. WM Segmentation – SPGR 3 class

25. WM Segmentation – SPGR 4 class

26. WM Seg. – SPGR-FLAIR 4 class

27. GM Segmentation – SPGR 3 class

28. GM Segmentation – SPGR 4 class

29. Lesion Segmentation – SPGR 3 class

30. Lesion Seg. – SPGR-FLAIR 3 Class

31. CSF Segmentation – SPGR 3 Class

32. CSF Seg. – SPGR-T2-PD 3 class

33. Segmentation Questions What is the best way to obtain a lesion mask? Should we do operations on the masks we have, like subtracting the out-of-brain CSF from the SPGR-FLAIR 3 class CSF mask? Are there other existing tools out there for either automatic or semi-manual lesion segmentation?

34. Statistical Study We intend to use the Wilcoxon rank sum test as our workhorse for statistical comparison Many of our variables are not intrinsically Gaussian so the t-test and ANOVA do not seem applicable

35. Open Questions DTI registration to standard space: FMRIB58 or via SPGR mcDESPOT target? Patient matching: gender or age first? How to deal with lesions across patients: Vrenken approach is to replace missing data with the mean of the group