JOURNAL CLUB:Deoni et al. One Component? Two Components? Three? The Effect ofIncluding a Nonexchanging ‘‘Free’’ Water Component in mcDESPOT.

Jan 14, 2013Jason Su

Motivation

• mcDESPOT provides fast, whole-brain estimation of myelin water fraction– However, its accuracy near CSF may be questionable due

to partial voluming– A third non-exchanging component in the tissue model

may allow this to be accounted for

• Lankford has recently criticized the precision of mcDESPOT with an unbiased fitting algorithm– Is this observed in practice with constrained SRC?

Theory• The mcDESPOT signal equations are expanded to include a 3rd

component or free water pool– This is a straightforward extension of the block diagonal matrices

with some key assumptions, 10 total parameters– Adds 3 new parameters: T1, T2, and volume fraction of the free

pool• 1500 < T1free <7500ms; 150ms < T2free < 1000ms; 0 < Ffree < 0.75

– The free pool is nonexchanging, avoids adding 2 possible new exchange rate parameters• Assumes that it models CSF separated by the blood-brain barrier

• The IE pool is at least 5% of the volume of the voxel

Concerns

• Is degeneracy possible? Can myelin and IE pools be confused with each other?– In 2-pool model, this can occur if MWF is allowed to be ≥0.5– What if here free pool is 0.2, the remaining 0.8 could be split

between MWF and IE interchangeably, is this taken care of?

• SRC algorithm– Ignoring some fundamental issues I’ve encountered, i.e. mean

normalization and off-resonance as a cyclic dimension– Not stated whether it usually converges or ends due to hitting

the iteration limit• In my simulations, it’s usually the latter

Methods• Simulation

– Simulated a 3-component model 50,000x each over a range of tissue parameters and fitted with 2- or 3-component models

• In-vivo at 3T – protocol is the same as with 2-component, 8 SPGR and 8 SSFP, and low-res IR-SPGR for DESPOT-HIFI B1 maps– Intra-subject repeatability

• 24yo male, 5 times over 5 weeks– Inter-subject variability

• 10 1yo infants• Why infants? To accentuate variability?

Results – Simulation

Notes

• Seems to be a fair amount of bias in Ffree estimate, almost 8%

• Histograms in general are fairly large with a width of about 0.08-0.11– My own simulations show similar or worse

performance in two-pool models with SNR of 20-30– Not sure how much noise was added in theirs

Product mcDESPOT

Modified mcDESPOT

Notes

• The 2-pool model tends to underestimate the MWF in regions near CSF compared to 3-pool– Makes sense, uses more of longer IE component as

a surrogate

• Would’ve liked to see this in MS brain with variety of lesions– Especially because there is the question of whether

edema would allow exchange with 3rd pool

Results – Curves

Notes

• BIC is a criterion that’s often used to compare fitting models– Penalizes for number of parameters, promotes a

simpler model to avoid overfitting– Closely related to Mallow’s Cp for linear models– Is a heuristic, cross validation is better for evaluating

predictive value of a model but may not work here

• SSFP fits are kind of iffy, particularly phase 0

Results – Reproducibility and Variability

Notes• Shows the characteristic behavior of low MWF CoV in

white matter, high CoV in gray• What registration was used for the inter-subject data?• Hard to interpret the intra-subject vs inter-subject results

– Are 1yo infants supposed to have about the same amount of myelination, how variable is it in development?

– Intra-subject variation is half of inter-subject, either MWF doesn’t vary much between infants or high variability in subject? Is 5 samples enough for CoV?

• Algorithm reproducibility is bad near edges/CSF?– Wonder what this looks like for 2-pool

MWF Variation in Infants

Deoni et al. Investigating white matter development in infancy and early childhood using myelin water faction and relaxation time mapping. Neuroimage. 2012 Nov 15;63(3):1038-53.

Discussion

• 3% deviation in homogeneous WM and GM compared to 2-pool– Is it worth changing model depending on location in brain?

• Is it possible in histology to examine if such a free pool in tissue is real? – The need for a 3rd pool arises out of need to account for

partial voluming, so phantoms should study that not necessarily 3-component mixture

– Examine the effects on 3rd pool as introduce more partial voluming, we could be able to precisely change the Ffree depending on voxel size?

Discussion

• Concerned about even more reduced precision compared to 2-pool– Variability of MWF near edges may indicate 3-pool

model is hard to estimate where it counts.– CoV of Ffree map?– Surprising that there is nothing added to the

acquisition• Can we really get a 3-pool fit for free from what we have

already?• CRLB would say otherwise.