Expectation-maximization with image-weighted Markov random fields to handle severe pathology

This paper describes an automatic tissue segmentation algorithm for brain MRI of children with cerebral palsy (CP) who exhibit severe cortical malformations. Many of the currently popular brain segmentation techniques rely on registered atlas priors and so generalize poorly to severely injured data sets, because of large discrepancies between the target brain and healthy (or injured) atlases. We propose a prior-less approach combined with a modification of the Expectation Maximization (EM)/Markov Random Field (MRF) segmentation by imposing a continuous weighting scheme to penalize intensity discrepancies between pairs of neighbors within each clique neighborhood, to provide robustness to the unique clinical problem of severe anatomical distortion. This approach was applied to gray matter segmentations in 20 3D T1-weighted MRIs, of which 17 were of CP patients exhibiting severe malformation. We compare our adaptive algorithm to the popular 'FreeSurfer', 'NiftySeg', 'FAST' and 'Atropos' segmentations, which collectively are state-of-The-Art surface deformation and EM approaches. The algorithm driven approach yielded improved segmentations (DSC 0.66 v 0.44 (FreeSurfer) v 0.60 (NiftySeg with 100% atlas prior relaxation) v 0.59 (FAST) v 0.64 (Atropos)) of the cerebral cortex relative to several ground-Truth manual segmentations, when compared to the existing approaches