1. Whole Brain Myelin Imaging with mcDESPOT in Multiple Sclerosis
July 18, 2012
Jason Su

2. Outline Introduction to parametric mapping Myelin imaging and MWF
mcDESPOT measurement of 2-pool exchange
mcDESPOT in multiple sclerosis
Current and future challenges

3. What is Parametric Mapping?
Start with a signal model for your data
Collect a series of scans, typically with only 1 or 2 sequence variables changing
Fit model to data
Motivation
Reveals quantifiable physical properties of tissue unlike conventional imaging
Maps are ideally scanner independent
Insert thumbnails of such maps?
For example, DTI only changes the direction of the diffusion gradients
T1 mapping gives an image in which the T1 longitudinal relaxation time is determined for every voxel in the volume, meaning each voxel has a unit associated with it (seconds).
Relaxivity can tell you concentration of contrast via T1 and if measured over time, how fast agent is absorbed or removed, relevant in cancer tumors
Maps should be scanner independent. But are they really? That is an interesting ongoing question.

4. Parametric Mapping Some examples
FA/MD mapping with DTI – most widely known mapping sequence
T1 mapping – relevant in study of contrast agent relaxivity and diseases
B1 mapping – important for high field applications
You might call it FA or MD mapping from DTI since DTI is the acquisition strategy and sometimes all that people take from it are fiber tracking results, which I would not call parametric mapping.
You might add T2 mapping

5. T1 mapping in multiple sclerosis
T1 Mapping Motivation
T1 mapping in multiple sclerosis
DCE-MRI in tumors:
[Gd] related to T1
grade II
Here are some applications that use T1 mapping. In multiple sclerosis, the Gd enhancement of lesions can be quantitatively tracked with T1 over time, also seemingly normal appearing white matter shows elevated T1 values compared to normals. In DCE-MRI, the uptake of contrast agent in tumors can be tracked by use of T1 maps and knowledge of the relaxivity of the agent.
The DCE-MRI example doesn’t absolutely require T1 mapping – a time course of T1-weighted images can be used and is probably more commonly used. But ideally you use dynamic quantitative T1 maps to feed into the tracer kinetic models.
grade IV
grade III
Levesque et al. 2010
Tofts et al. 1999, 2003, Patankar et al. 2005

6. Relaxation Mapping T1 mapping T2 mapping
IR SE – gold standard, vary TI
Look-Locker – use multiple readout pulses to collect many TIs
DESPOT1 – vary flip angle
T2 mapping
Dual SE – vary TE
CPMG – use multiple spin echoes to collect many TEs
DESPOT2 – vary flip angle
There are many different mapping methods, but each follows this paradigm of having a model and varying select parameters.

7. T1 Mapping: Inversion Recovery
Acquire a series of images at different TIs, at each voxel fit curve across the data.
Inversion or saturation recovery: 2DFT variant is slow, can be combined with FSE, single-shot imaging to accelerate
Simple exponential recovery
Gowland &Stevenson, in Tofts ed., QMRI of the Brain, 2003
Brix et al. MRI 1990; Ropele et al. MRM 1999; Wang et al. MRM 1987

8. DESPOT1 T1 mapping Christensen 1974, Homer 1984, Wang 1987, Deoni 2003
The method that’s of primary interest stems from the DESPOT1 method shown here.
Peak is called the Ernst angle = acos(E1).
Hope that gives a little bit of motivation for why parameter mapping is useful as well as how it is accomplished. Now I’ll delve a little further into the nitty gritty of steady state, DESPOT mapping methods.
Christensen 1974, Homer 1984, Wang 1987, Deoni 2003

9. DESPOT Methods
Vary flip angle in steady state sequences like SPGR and SSFP
Fast, whole brain, higher resolution 1-2mm isotropic
Requires accurate knowledge of flip angle
B1+ transmit field inhomogeneity – problem for >1.5T
Excitation slab profile – typically known and accounted for
DESPOT1 – T1 mapping
DESPOT-HIFI – add an inversion to allow T1 and B1+ mapping
DESPOT2 – T1 and T2 mapping
DESPOT-FM – collect multiple SSFP phase cycles to map B0
mcDESPOT – multi-component T1 and T2 mapping

10. Relaxation Based Myelin Imaging
DTI is not an ideal measure of myelin (low resolution, crossing fibers problem)
T2 (or R2) has been used in the past as a crude correlate of myelin
Myelination reduces water content in brain, lower T2
T2w FLAIR is used in MS to highlight lesions
T2 mapping gives a more sensitive indicator
Examine the integrity of myelin in the brain.

11. Myelin Water Fraction
Recent methods have focused on a more specific measure: myelin water fraction (MWF)
Multiecho qT2 – vary TE, decomposes the signal into a spectrum of T2 times (UBC, MacKay)
Well validated way to produce MWF maps that represent myelin
Few slices, long acquisition time
mcDESPOT – vary flip angle, models SPGR and SSFP steady state signal
Also based on modeling relaxation and two pool exchange
Validation in progress
High resolution, whole brain, but long processing time (24 hours)
- qT2 is a well validated way to produce myelin water fraction maps which are representative of myelin in brain but limited by coverage and acq time
- mcDESPOT is a newer technique that based around modeling relaxation as well and can produce maps that are conceptually similar to qT2's MWF, ie a long and short T2 pool in exchange, validation in progress
Intra- and extra-cellular water, T2 ≈ 80ms
Myelin water, T2 ≈ 20ms

12. Fractional Anisotropy map (3T), MWF (qT2, 3T)
FA vs MWF
FA and MWF-qT2 are of same normal subject
MWF-mcD different
Fractional Anisotropy map (3T), MWF (qT2, 3T)
MWF (1.5T, mcDESPOT)

13. mcDESPOT Models tissue as two water pools in exchange
Fast relaxing water pool
Slow relaxing water pool
𝑓 𝐹 + 𝑓 𝑆 =1
Assume chemical equilibrium:
𝑓 𝐹 𝑘 𝐹𝑆 = 𝑓 𝑆 𝑘 𝑆𝐹
The SPGR and SSFP signal equations must be adapted to take into account this model
T1,F T2,F fF
kFS
T1,S T2,S fS
kSF
Chemical exchange means that no pool is actively growing, which is reasonable within a scanning session
Eliminates 2 parameter to estimate, leaving 6 + B0/M0

14. mcDESPOT Model: SPGR SPGR equation Single Component
𝑆 𝑆𝑃𝐺𝑅 = 𝑀 0 1− 𝐸 1 sin⁡(𝛼) 1− 𝐸 1 cos⁡(𝛼) 𝐸 1 = e − 𝑇𝑅 𝑇 1
Here I’ve assumed TE is minimum close to 0 to eliminate that term.
Importantly SPGR eliminates transverse magnetization which allows us to only need to treat longitudinal magn. and T1. In other words, SPGR allows us to separate T1 from T2 which is critical for getting DESPOT2 to work.

15. mcDESPOT Model: SPGR SPGR Equation Multi-Component
𝑆 𝑆𝑃𝐺𝑅 = 𝑀 0,𝑆𝑃𝐺𝑅 𝐼− 𝑒 𝐴 𝑆𝑃𝐺𝑅 𝑇𝑅 sin 𝛼 𝐼− 𝑒 𝐴 𝑆𝑃𝐺𝑅 𝑇𝑅 cos 𝛼 −1 𝑀 0,𝑆𝑃𝐺𝑅 = 𝑀 0 𝑓 𝐹 𝑓 𝑆 𝐴 𝑆𝑃𝐺𝑅 = − 1 𝑇 1,𝐹 − 𝑘 𝐹𝑆 𝑘 𝑆𝐹 𝑘 𝐹𝑆 − 1 𝑇 1,𝑆 − 𝑘 𝑆𝐹
M0 represents the longitudinal magn. of fast and slow pools

16. mcDESPOT Model: SPGR Single component fit of multi-component data
f_F = .21;
T1_F = 584;
T1_S = 1400;
T_FS = 67;
Similarly there’s a matrix model for the SSFP equations, which is quite large because need to also treat transverse magnetization and I won’t go into it here.
Deoni et al. 2008

17. The Technique
Font is small, not very black, sometime you should re-create this slide in a much clearer simpler form

18. mcDESPOT Model Fitting
Expensive non-linear curve fitting problem
24 hour per 2mm isotropic brain with 12-core CPU
Previous implementations used genetic algorithms
Currently using stochastic region of contraction
SRC is a perhaps lesser known optimization method that works by taking random samples from a range in the parameter space and gradually reducing the range until the solution is honed in on.

19. mcDESPOT Maps in Normal
T1single
T1slow
T1fast
MWF
0 – 2345ms
0 – 1172ms
0 – 555ms
0 – 0.234
0 – 328ms
0 – 123ms
0 – 9.26ms
0 – 137ms
T2single
T2slow
T2fast
Residence Time

20. ISMRM 2011 E-Poster #4643
mcDESPOT-Derived MWF Improves EDSS Prediction in MS Patients Compared to Atrophy Measures Alone
J.Su1, H.H.Kitzler2, M.Zeineh1, S.C.Deoni3, C.Harper-Little2, A.Leung2, M.Kremenchutzky2, and B.K.Rutt1
1Stanford U, CA, USA, 2TU Dresden, SN, Germany, 2U of Western Ontario, ON, Canada, 3Brown U, RI, USA
Title: I prefer the wording “Compared to Atrophy Measures Alone” over “Compared to Only Atrophy Measures”
Are we allowed to change the title so that it’s different from the original abstract?
Declaration of Conflict of Interest or Relationship
I have no conflicts of interest to disclose with regard
to the subject matter of this presentation.

21. mcDESPOT-Derived MWF Improves EDSS Prediction in MS Patients Compared to Atrophy Measures Alone
ISMRM 2011 #4643
Background
Conventional MRI measures such as lesion load have been criticized with adding little new information on top of clinical scores for multiple sclerosis (MS) patients
Measures that quantify the hidden burden of disease in white matter are urgently needed

22. mcDESPOT-Derived MWF Improves EDSS Prediction in MS Patients Compared to Atrophy Measures Alone
ISMRM 2011 #4643
Purpose
To apply mcDESPOT, a whole-brain, myelin-selective, multi-component relaxometric imaging method, in a pilot MS study
Assess if the method can explain differences in disease course and severity by uncovering the burden of disease in normal-appearing white matter (NAWM)

23. Study Demographic Data Healthy Controls All Patients CIS RRMS SPMS
mcDESPOT-Derived MWF Improves EDSS Prediction in MS Patients Compared to Atrophy Measures Alone
ISMRM 2011 #4643
Study
Demographic Data
Healthy Controls
All Patients
CIS
RRMS
SPMS
PPMS N 26 10 5 6
Mean age, yr
(SD) 42
(13) 49
(12) 41 48 58
(7) 55
Male/Female ratio
10/16
7/19
3/7
0/5
0/6
4/1
Mean disease duration, yr — 14 2
(2) 15
(10) 28
(8) 20
Mean EDSS score
3.6
(2.4)
1.7
(0.9)
2.0
(1.7)
6.4
(1.1)
5.6

24. Scanning Methods 1.5T GE Signa HDx, 8-channel head RF coil
mcDESPOT-Derived MWF Improves EDSS Prediction in MS Patients Compared to Atrophy Measures Alone
ISMRM 2011 #4643
Scanning Methods
1.5T GE Signa HDx, 8-channel head RF coil
mcDESPOT: 2mm3 isotropic covering whole brain, about 15 min.
SPGR: TE/TR = 2.1/6.7ms, α = {3,4,5,6,7,8,11,13,18}°
bSSFP: TE/TR = 1.8/3.6ms, α = {11,14,20,24,28,34,41,51,67}°
2D T2 FLAIR: 0.86 mm2 in-plane and 3mm slice resolution
3D T1 IR-SPGR: 1mm3 resolution with pre/post Gd contrast

25. Processing Methods: MWF
mcDESPOT-Derived MWF Improves EDSS Prediction in MS Patients Compared to Atrophy Measures Alone
ISMRM 2011 #4643
Processing Methods: MWF
Linearly coregister and brain extract mcDESPOT SPGR and SSFP images with FSL1
Find myelin water fraction maps using the established mcDESPOT fitting algorithm2
Myelin Water Fraction
1FMRIB Software Library. 2Deoni et al., Magn Reson Med Dec;60(6):

26. Processing Methods: Deficient MWF
mcDESPOT-Derived MWF Improves EDSS Prediction in MS Patients Compared to Atrophy Measures Alone
ISMRM 2011 #4643
Processing Methods: Deficient MWF
Non-linearly register mcDESPOT MWF maps to MNI152 standard space
Combine normals together to form mean and standard deviation MWF volumes
For each subject, calculate a z-score ([x – μ]/σ) at every voxel to determine if it is significantly deficient, i.e. MWF < -4σ below the mean
Deficient MWF Voxels

27. Processing Methods: WM
mcDESPOT-Derived MWF Improves EDSS Prediction in MS Patients Compared to Atrophy Measures Alone
ISMRM 2011 #4643
Processing Methods: WM
Brain extract MPRAGE images
Segment white and gray matter with SPM83
Filter tissue masks to reduce noise then manually edit by a trained neuroradiologist
Calculate parenchymal volume fraction (PVF) as WM+GM divided by the brain mask volume
FLAIR WM
3Statistical Parametric Mapping software package.

28. Processing Methods: Lesions & DAWM
mcDESPOT-Derived MWF Improves EDSS Prediction in MS Patients Compared to Atrophy Measures Alone
ISMRM 2011 #4643
Processing Methods: Lesions & DAWM
Non-linearly register T2-FLAIR images to MNI152 standard space
Combine normals together to form mean and standard deviation volumes
Segment lesions as those voxels with z-score > +4 and diffusely abnormal white matter > +2
Edit masks by a trained neurologist
DAWM
Lesions

29. Processing Methods: NAWM & DVF
mcDESPOT-Derived MWF Improves EDSS Prediction in MS Patients Compared to Atrophy Measures Alone
ISMRM 2011 #4643
Processing Methods: NAWM & DVF
Segment normal-appearing white matter (NAWM) as WM – DAWM – lesions
Find deficient MWF volume fraction (DVF)
Sum the volume of deficient voxels in each tissue compartment and normalize by the compartment’s volume
# deficient voxels in compartment * voxel volume / compartment volume
Normal-Appearing
White Matter

30. Segmentations and DV FLAIR WM NAWM DAWM Lesions MWF Deficient MWF
mcDESPOT-Derived MWF Improves EDSS Prediction in MS Patients Compared to Atrophy Measures Alone
ISMRM 2011 #4643
Segmentations and DV
FLAIR WM
NAWM
DAWM
Lesions
After all that processing, we can visualize the end result in a very intuitive way.
MWF
Deficient MWF
Voxels
DV in NAWM
DV in DAWM
DV in Lesions

31. mcDESPOT-Derived MWF Improves EDSS Prediction in MS Patients Compared to Atrophy Measures Alone
ISMRM 2011 #4643
Statistical Methods
Use rank sum tests to compare patient groups to normals along different measures
Perform an exhaustive search to find the best multiple linear regression model for EDSS using Mallows’ Cp4 criterion among 21 possible image-derived predictors:
PVF
log-DVF in whole brain, log-DVF in WM, log-DVF in NAWM, log-DVF in lesions
log-DV in those four compartments
mean MWF in those four compartments
volumes of those four compartments (lesion volume = T2 lesion load)
volume fractions of those four compartments with respect to the whole brain mask volume
Parameter mapping allows for an explosion of variables which we can use to predict subject disability
4Mallows C. Some comments on Cp. Technometrics. 1973;15(4):

32. Results: Mean MWF in Compartments
mcDESPOT-Derived MWF Improves EDSS Prediction in MS Patients Compared to Atrophy Measures Alone
ISMRM 2011 #4643
Results: Mean MWF in Compartments
Dotted line shows mean MWF in WM for normals. Rank sum testing was done for each bar against this
Testing was also done for RRMS vs. SPMS and CIS vs. RRMS, any significant differences are shown with a connecting bracket
Significance levels:
* p < 0.05
** p < 0.01
*** p <

33. Results: DVF in Compartments
mcDESPOT-Derived MWF Improves EDSS Prediction in MS Patients Compared to Atrophy Measures Alone
ISMRM 2011 #4643
Results: DVF in Compartments
Dotted line shows deficient MWF volume fraction in WM for healthy controls
With DVF, all patient subclasses were significantly different from healthy controls
PVF, however, fails to distinguish CIS and RR patients from normals
Key idea is that we need whole brain mapping methods to allow computations like DVF against normal populations, MWF alone does not separate groups. This is one value of mcD over qT2.

34. Results: Correlations with EDSS
mcDESPOT-Derived MWF Improves EDSS Prediction in MS Patients Compared to Atrophy Measures Alone
ISMRM 2011 #4643
Results: Correlations with EDSS
Lesion load correlates poorly with EDSS
PVF and DVF are stronger indicators of decline

35. Results: Multiple Linear Regression
mcDESPOT-Derived MWF Improves EDSS Prediction in MS Patients Compared to Atrophy Measures Alone
ISMRM 2011 #4643
Results: Multiple Linear Regression
The best linear model for EDSS contains PVF (p < 0.001), mean MWF in whole brain (p < 0.001), and WM volume fraction (p < 0.01)
Whole-brain MWF and WM volume fraction significantly improve the prediction of EDSS over that produced by PVF alone
Explains 76% of the variance in EDSS (R2 = 0.76, adjusted R2 = 0.73) compared to 56% with only PVF

36. Discussion & Conclusions
mcDESPOT-Derived MWF Improves EDSS Prediction in MS Patients Compared to Atrophy Measures Alone
ISMRM 2011 #4643
Discussion & Conclusions
DVF is able to differentiate CIS and RRMS patients from normals, whereas other measures such as PVF and mean MWF cannot
The invisible burden of disease may be more important than lesions in determining disability, since we observe a higher correlation of EDSS with DVF in NAWM than lesion load
A combination of established atrophy measures with new mcDESPOT-derived MWF are more capable in accurately estimating disability than either quantity alone

37. ISMRM 2011 E-Poster #7224
Sensitive Detection of Myelination Change in Multiple Sclerosis by mcDESPOT
J.Su1, H.H.Kitzler2, M.Zeineh1, S.C.Deoni3, C.Harper-Little2, A.Leung2, M.Kremenchutzky2, and B.K.Rutt1
1Stanford U, CA, USA, 2TU Dresden, SN, Germany, 2U of Western Ontario, ON, Canada, 3Brown U, RI, USA
Declaration of Conflict of Interest or Relationship
I have no conflicts of interest to disclose with regard
to the subject matter of this presentation.

38. Results: Mean MWF in Whole Brain
Sensitive Detection of Myelination Change in Multiple Sclerosis by mcDESPOT
ISMRM 2011 #7224
Results: Mean MWF in Whole Brain
Dotted line shows mean MWF for normals. Rank sum testing was done for each bar against this value
Testing was also done for RRMS vs. SPMS and CIS vs. RRMS, any significant differences are shown with a connecting bracket
Significance levels:
* p < 0.05
** p < 0.01
*** p <

39. Results: DVF Change Colors denote subject type
Sensitive Detection of Myelination Change in Multiple Sclerosis by mcDESPOT
ISMRM 2011 #7224
Results: DVF Change
Colors denote subject type
Arrowheads indicate the direction of change and the DVF at 1-year
Dashed lines show subjects who also had a change in EDSS
PPMS
SPMS
RRMS
CIS
Normals

40. Results: DVF in Whole Brain
Sensitive Detection of Myelination Change in Multiple Sclerosis by mcDESPOT
ISMRM 2011 #7224
Results: DVF in Whole Brain
Dotted line shows mean deficient MWF volume fraction change for normals
Definite MS patients are losing significantly more myelin than normals
Progressive patients have a greater rate of DVF increase
In previous bar charts, you used connecting brackets to show significant differences, but here you aren’t. Any reason for that?
There was no significance difference between CIS vs RR and RR vs SP despite the apparent visual differential in the bars.

41. Discussion & Conclusions
Sensitive Detection of Myelination Change in Multiple Sclerosis by mcDESPOT
ISMRM 2011 #7224
Discussion & Conclusions
DVF shows statistically significant changes in brain myelination over the study period
Progressive patients show greater disease decline that are not reflected in their EDSS disability score
EDSS and DVF appear to measure different aspects of the disease.
Patients with changes in EDSS did not actually have the largest DVF changes

42. Current and Future Work
High-Field mcDESPOT
3T: 6 min 2mm isotropic, post-correction with a B1+ map is sufficient
7T: k-T points pulse design is showing promise in flattening the transmitted field
Accelerated mcDESPOT
DISCO-based view-sharing working with DESPOT1
SSFP (DESPOT2) more challenging
Possible new applications
Alzheimer’s Disease: the myelin hypothesis
Traumatic brain injury
Novel segmentation
At 7T, nonuniformity is so great that there’s complete signal drop out, need a better solution that post-correction