7T Thalamus and MS Studies Jason Su Sep 16, 2013 Outline • 7T Thalamus and MS studies – Atrophy measures – PVF, thalamus, hippocampus – Study-specific mean brain template construction with ANTS – Manual -> Automatic segmentation with label fusion 7T Studies • Thomas, Manoj, and Ives recently used WMnMPRAGE to manually segment thalamic nuclei in normal control • Now looking to use this methodology in MS patients at 7T with atrophy and disease markers: – – – – – Brain atrophy, parenchymal volume fraction T2-FLAIR lesions Thalamic atrophy (whole or nuclei) Thalamic lesions Hippocampal atrophy Brain Atrophy • Want to measure the ratio of the brain parenchyma to the intracranial cavity volume • Previous MSmcDESPOT methodology – FSL BET the “target” image (T1w SPGR, 18deg) • The intracranial volume – Segment WM+GM MPRAGE with SPM8 • The brain parenchyma volume – Measures enlargement of ventricles mainly T1w BET T1w-MPRAGE T2-CUBE Brain Atrophy: New Method • FSL BET T2-CUBE image – Then edit to produce intracranial volume (~30min) • Remove CSF by segmenting T2-CUBE w/ SPM8 – Note that T1w input is more typical for SPM8 but CSF is so distinct on T2w it works – Then edit to produce brain parenchyma • Advantages: – More true to the definition of intracranial cavity – Better measure of the space between brain and skull T2w BET Before Editing After Editing CSF Removal Before Editing After Editing ? Difficulties • SPM8 is the weakest part of the pipeline – Often needs manual tweaking between subjects to get to a reasonable segmentation – Registration to its atlas often goes haywire, likely affected by 7T nonuniformity • Would be easy to manually threshold bright voxels as CSF but reviewers might prefer a proven toolbox – Is it better to be consistent or accurate? – Does it matter in the end since always edited as a final step? T2-FLAIR Lesions • Previous MSmcDESPOT methodology – Nonlinearly register all subjects to MNI brain – Normalize signal intensities to “robust” max in brain – Group normal controls and find their voxel-wise mean and std. dev. – Compare new subject to the control population • Voxels > 4 standard deviations above mean = lesion • Voxels > 2 standard deviations above mean = DAWM T2-FLAIR Lesions • New methodology – Create a study-specific mean brain template to use instead of MNI – More accurate registration should improve ability to detect lesions, maybe cortical? Mean Brain Template • ANTs (Advanced Normalization Tools) is emerging as the standard – Available in pre-compiled OSX binary – Provides useful parallelized script to compute a mean brain (buildtemplateparallel.sh) – But sometimes crashes whole computer! • Which image contrast? Want something that: – Provides useful image contrast throughout the brain – Is less affected by B1+ inhomogeneity, as this variation between subjects is not taken into account in these registration algorithms • Do typically correct for receive nonuniformity – Shows few lesions so they do not misguide the registration MPRAGE • Good gray/white contrast throughout brain • Uniformity isn’t ideal • Shows few lesions WMnMPRAGE • Extreme gray/white contrast throughout brain • Uniformity seems better than MPRAGE, could be visual effect due to extreme contrast • Shows some lesions T2-CUBE • Poor gray/white contrast • Decent uniformity • Shows few lesions FLAIR • Decent contrast • Uniformity is poor, signal loss in center of brain, which we care about the most for thalamus • Shows many lesions Decision • My first attempt was with WMnMPRAGE – Good contrast albeit unusual, but registration algorithms should be indifferent to that – Most importantly, this is the contrast we’re using for thalamic segmentation so it is a natural choice ANTs • Sum all subject brains without registration creating a crude template • Rigidly register subjects to this crude base • Sum these to produce the initial template • Iterate until convergence: – Nonlinearly register to the template – Sum the new registrations to create a new template ANTs Iteration 1 ANTs Iteration 2 ANTs Iteration 3 ANTs Iteration 7 ANTs Iteration 8 MNI T1w Notes • Cortical registration seems good • Still not quite converged after 8 iterations – 4 is the default number of iterations, way low – Each iteration takes 12 hours, barring any crashes • Once this is settled, can do segmentation of the lesions – Comparing FLAIR intensities to controls is fairly straightforward – May want to also do it on WMnMPRAGE for thalamic lesions Structural Atrophy • After our review of label fusion and ASHS papers, learned we can turn previous manual segmentations into automatic – Thomas has already manually segmented 2x6 controls, though in a non-accelerated WMnMPRAGE • Label fusion: – Register the previous segmentations to the new subject – Use local information about the registration accuracies to guide the decision What We Want Prior Subject 1 Prior Subject 2 New Subject Prior Subject 3 Problem • Nonlinear registration of N prior subjects is expensive to every new subject • Instead use the mean brain template as an intermediate space – Allows us to only need one registration for a new subject Proposed Approach Prior Subject 1 Prior Subject 2 Mean Brain Template New Subject Prior Subject 3 Drawbacks • Not straightforward how to adapt this to an ROIfocused registration – Accuracy of the mean template may be imperfect due to the expense of registering whole brain • As in ASHS, would like to first start from the whole brain registration – Then do a targeted registration with small FOV on thalamus or hippocampus – Do we take brains all the way to the subject space and register there or stay in the template space? – ASHS chooses to use template space, less intuitive • Argument for it has to do with producing a label probability map which plays nicely with linear interpolation unlike a binary mask • But STEPS does not expose that intermediate map Other Concerns • STEPS suggested a library in the range of 15+ subjects – May not be able to achieve reliable automatic segmentation until after the study – Before that we can at least try to get something reasonable to reduce Thomas’s tracing work Nuclei Separability by T1 • Another possibility is to incorporate T1 map information to help improve label fusion – Possibly as an included variable in AdaBoost like ASHS Thalamic Nuclei Notes • For 3/5 nuclei, the T1 dist. is distinct from some or all of its neighbors Summary Brain atrophy, parenchymal volume fraction • Pipeline complete, but needs CSF segmentation needs constant tweaking between subjects • Intracranial cavity editing done, parenchyma editing in progress • Label fusion can reduce the intracranial cavity editing work T2-FLAIR lesions • Need to iterate even more for mean brain template to converge Thalamic atrophy (whole or nuclei) • Once template is finalized, can begin playing with STEPS for label fusion Thalamic lesions • Try lesion segmentation pipeline with WMnMPRAGE Hippocampal atrophy • Thomas is working on measuring CA1 thickness • Segmentation is open question, try ASHS?