Whole brain image registration using multi-structure confidence-weighted anatomic constraints Ali R. Khan, Mirza Faisal Beg Simon Fraser University, Canada Introduction
Results
Inter-subject registration of whole brain magnetic resonance images Challenging: high anatomical variability, convoluted folding of the cortex Applications: morphometry, functional localization, atlas creation Our approach: Use automated Freesurfer [1] segmentations of multiple brain structures as simultaneous anatomic constraints (multi-structure registration), instead of as initialization [2] Weight these using trained segmentation confidence maps (SCMs) Large deformation diffemorphic framework for registration (LDDMM [3])
Multi-structure confidence-weighted LDDMM registration compared against: Free-form Deformation B-splines, IRTK [4] Single channel LDDMM 1.5T brain MR scans brains from the Internet Brain Segmentation Repository (IBSR) [5] 9 brains used for training SCMs, other 9 brains used for testing image registration
Cortical surface distances Surface Distance (mm)
Methods 1. Run Freesurfer
Subcortical and cortical Fully automated segmentation Segmentation errors? Train confidence maps
Subcortical overlap
6.0 5.0 4.0 3.0 2.0 1.0 0.0
2. Train segmentation confidence maps for each brain structure For each training subject: a. Find local errors between manual and automated structure b. Spatially normalize these to the atlas c. Compute SCM as probability of accuracy
0.73 0.74 0.79 0.73 0.74 0.78 0.8 0.8 0.83 0.84 0.85 0.86 0.73 0.75 0.79 0.77 0.8 0.81 0.67 0.72 0.71 0.65 0.65 0.73
WM GM Vent Thal Caud Put
Single-channel
Multi-structure
Propagated cortical surfaces
Pall Hipp
Freesurfer Manual
Segmentations
Amyg
SCM
Error map
Nuc Acc
0.54 0.57 0.48 0.51 0.53
0.66
IRTK Single channel Multi-structure
0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00
3. Perform multi-structure confidence-weighted LDDMM registration to atlas
Reference
16 subcortical, and 35 cortical structures used in the multi-structure registration Provides anatomical constraints to help guide the high-dimensional registration
Single-channel
Multi-structure
Dice Similarity Coefficient
Conclusions
ϕ SCMs
Subject subj Atlas Atlas
Anatomical constraints, in the form of automated segmentations, can improve brain registration More accurate volumetry, morphometry, or functional localization in brain mapping studies Limitations: SCMs generated for subcortical structures only; future work will include cortical SCMs High computational cost with LDDMM; requires high-performance computing machines
References [1] B. Fischl, et al., “Whole brain segmentation automated labeling of neuroanatomical structures in the human brain,” Neuron, vol. 33 (3), pp. 341–55, 2002. [2] A. R. Khan et al., “Freesurfer-initiated fully-automated subcortical brain segmentation in MRI using large deformation diffeomorphic metric mapping,” Neuroimage, vol. 41 (3), pp. 735–46, 2008. [3] M. F. Beg, et al., “Computing large deformation metric mappings via geodesic flows of diffeomorphisms,” International Journal of Computer Vision, vol. 61 (2), pp. 139–57, 2005. [4] Rueckert et al., “Nonrigid registration using free-form deformations: application to breast MR images,” IEEE Transactions on Medical Imaging, vol. 18 (8), pp. 712–21, 1999. [5] IBSR data was provided by the Center for Morphometric Analysis at Massachusetts General Hospital and is available at http://www.cma.mgh.harvard.edu/ibsr/.
Medical Image Analysis Lab, School of Engineering Science Simon Fraser University, 8888 University Dr, Burnaby, BC,V5A 1S6, Canada
Ali R. Khan (
[email protected]), Mirza Faisal Beg (
[email protected]) Website: http://www.autobrainmapping.com
