medInria-NT 2: Kick-Off Hackfest

Multiple Sclerosis Lesion
Segmentation using an Automated
Multimodal Graph Cut
J. Beaumont, O. Commowick, C. Barillot
October 21st, 2016
Introduction
• Automatic MS lesions segmentation
• Classification based
• Lesions seen as outliers
• With respect to normal 3-classes segmentation
• Overall approach
• Robust three-class EM segmentation
• Graph-cut segmentation of lesions
2
Robust EM segmentation
• Objective: segmentation of the brain in three classes
• Not influenced by the presence of lesions
• Trimmed EM segmentation
• Reject h% of voxels with largest residuals
• Compute regular EM segmentation on remaining ones
• Output
• Mean and covariances of each class
• Mahalanobis distances to each class of each voxel
3
EM initialized graph cut segmentation
• Graph structure
• One node per voxel
• Terminal source and sink
• Works from initial landmarks (not automatic)
• Compute characteristics of each class
• Compute optimal cut between lesions and background
• Making it automatic: initialize from EM classes and
Mahalanobis distances
4
Further challenge automatization
• Execution sensitive to expected lesion load
• Our solution
• Two parameters sets
•
•
Mild lesion load
Large lesion load
• Rough lesion load computed from FLAIR and T2
• Used to select parameter set
5