Groupe de Travail IRMf/MEG 02/2014
Anatomic parcellation based on DTI data with FSL taking the example of SMA/preSMA Magdalena Wutte, Lucile Brun & Boris Burle SMA/preSMA
Parcellation - what for?
- connectivity “fingerprints” - allows us to separate brain areas based on their WM connections
Behrens, Johansen-Berg et al, Nature Neuroscience, 2003
Different types of parcellation - with targets: parcel an area based on its differential connection to predefined regions (e.g. thalamus Behrens2003, cingulate cortex Beckmann2009)
- blind: parcel an area based on its differential connection to the whole brain SMA/preSMA Johansen-Berg2004 Operculum (Broca area) Klein2007
Principle of blind parcellation Seed mask : dorso-medial frontal cortex Target mask : whole brain
Estimate diffusion tensors
Preprocessing (movement correction etc)
Split seed mask according to those fingerprints
Tractography Search for differential connectivity fingerprints in seed mask (cross correlation/clustering)
Data from fMRI center (copy-pasted from Olivier:)
sujet/
anat01/
sujet_anat01.nii (1x1x1mm3)
sujet_diffusion/
sujet_dwi.nii (1.89x1.89x2.33mm3)
anatomie, T1
diffusion, 36 directions + 8 images T2 (b=0) (4D image)
for all subjects the same: 0 0 0 0 0 0 0 0 1000 1000 1000 1000 1000 1000 1000 1000 1000 …
bvals
weights for each volume (0 = T2, 1000 = diffusion)
… 1000 1000 1000 1000 1000 1000
bvecs
0 0 0 0 0 0 0 0 -0.834348 0.9726 -0.169126 -0.10601 -0.830245 …
… -0.764557 0.397774 -0.723579
gradient directions
nr directions : restricts tractography, e.g. crossing fibers can not be modelled with 36 (bedpost)
The software/toolbox
FSL - FDT: FMRI Diffusion Toolbox - Linux platform - bash-based - user interface/command line (copy-pasted from Olivier:) http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/ IRM fonctionnelle IRM anat, segmentation, VBM IRM diffusion : FMRIB’s diffusion Toolbox: FDT Tract-Based Spatial Statistics: TBSS Deux tutoriaux: http://fsl.fmrib.ox.ac.uk/fslcourse/lectures/practicals/fdt http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FDT/UserGuide
Overview of analysis steps Create masks
Preprocessing
fslview Brainvisa FSL preprocessing tools (fslreorient, bet, flirt) FSL FDT toolbox (eddy_correct)
Estimate diffusion tensors FSL FDT toolbox Tractography Correlation/Clustering
[the complete pipeline/commands in wiki]
Matlab
Create mask Seed mask :
Target mask :
- one slice at x: -4 - y: -22 to y 30 - z: superior of cingulate sulcus - in MNI space - then transformed to individual spaces - goal was to determine border SMA/preSMA (around y 0, vertical line From AC)
- subsample to 4mm voxels (fslmaths) (due to processing time)
Preprocessing 1) conversion to nifti: Brainvisa 2) get all images in the same orientation (fslreorient2std)
T1
DTI
3) extraction of B0 images from *_dwi.nii ('nodif.nii'; as registration target in flirt and as reference volume in eddy_correct) 4) brain extraction/skull stripping of T1 image (bet) 5) transformation matrices from one space to the other (individual > MNI, indivT1 > indivDTI; flirt)
Preprocessing 6) Apply those matrices to the seed mask (MNI>indivT1) 7) Realignment (motion correction; FDT diffusion > eddy_correct) uses co-registration of each dti volume to nodif image
Data for next steps: - skull stripped anatomy (*_brain.nii) - preprocessed dti data (data.nii) - dti brain mask (nodif.nii) - subsampled target brain mask - seed mask
Estimate diffusion tensors: bedpostX
Pdd : principle diffusion direction
Estimate diffusion tensors : bedpostX Input : data.nii.gz, nodif_brain_mask.nii.gz, bvals, bvec Output : dti01.bedpostX/
Here : only 1 fiber estimated because we have only 36 directions Processing time : ~2-3 h
Does the same as dtifit used for TBSS; Difference : 1) can estimate crossing fibers ; 2) quantifies uncertainty of pdd (bayesian modelling) needed for probabilistic tractography (probtrackX needs its output) Different nomenclature : e.g. dyads1.nii.gz == dti_FA.nii.gz
Estimate diffusion tensors : bedpostX
Tractography
Tractography Streamline (deterministic tractography)
Principle diffusion direction
If we would have no noise and perfect data we could use deterministic, but we don't...
Tractography : probtrackX How confident are we? Measure of uncertainty. Probabilistic rather than deterministic tractography > from bedpostX : distribution of directions per voxel > run n streamlines from seed randomly choosing orientation from voxel > spatially overlap results : probabilistic map
(number of lines)
Tractography : probtrackX the dorso-medial mask
Target is the (subsampled) whole-brain seed in anat space, not diffusion space
Tractography : probtrackX
Runs for 30 min – 2h
output we need : fdt_matrix2.dot
Principle of blind parcellation with Matlab:
Cross-correlation of connectivity profiles number of fibers
fdt_matrix2.dot
native cross correlation matrix c
dorso-medial
dorso-medial
orr e
whole brain
lati on
[0.3]
dorso-medial
Clustering of the cross correlation matrix Split mask according to those clusters
Results : Group
Parcels
n=30
Tracts y = -8
y = 20
SMA preSMA threshold nr subjects
Individuals - differences
crosshair at [-4 0 0]
Limitations/Troubleshooting Clusters are sometimes not clear cut
Limitations/Troubleshooting Different algorithms different results
kmeans 2 cluster
kmeans 2 cluster
kmeans 3 cluster
kmeans 3 cluster
spectral
spectral
Useful links/sources Many images from: http://fsl.fmrib.ox.ac.uk/fslcourse/
(lecture/practical)
Our scripts + a (Marseille specific) howto available at: http://wiki.irmfmrs.free.fr/doku (questions to
[email protected] or
[email protected]) Useful links: http://imaging.mrc-cbu.cam.ac.uk/imaging/DTItutorial http://dbic.dartmouth.edu/wiki/index.php/Diffusion_Tensor_Imaging_Analysis FSL course 31.03 - 04.04 : http://fsl.fmrib.ox.ac.uk/fslcourse/current.html