Feature-based statistical analysis of structural MR data for automatic detection of focal cortical dysplastic lesions

This paper presents a method for fully automated detection and localization of Focal Cortical Dysplastic (FCD) lesions from anatomical magnetic resonance (MR) images of the human brain. Model-based tissue classification of the image under study was applied first such that a gray matter (GM) segmentation map is obtained of which we demonstrate that it also includes possible FCD lesions. Cortical thickness was estimated at each voxel using an appropriate distance transform applied to the binarized GM object and an FCD specific feature map was constructed by computing the ratio of cortical thickness over absolute image intensity gradient at each voxel. In absence of any prior anatomical hypothesis on the spatial location of the lesion, a statistical parameter map was constructed by evaluating the evidence for each gray matter voxel against the null hypothesis of no difference in the feature map of the patient versus similar maps obtained for a group of normal controls. Voxel clusters for which the null hypothesis was found to be improbable at optimally selected thresholds for cluster height and extent were reported as lesions. The method was applied to a surgical series of 17 FCD patient images that were compared against a group of 64 neurologically normal controls. The method correctly detected and located the FCD lesion in 9 out of 17 FCD cases (53%) using a threshold that minimized false positives and 12 of 17 (71%) using a threshold that allowed more false positive results. The detected lesions had a median volume of 7.2 cm3 versus 2.9 cm3 for the non-detected lesions. The detected lesions more often had an increased cortical thickness on T1 than the non-detected lesions (P = 0.015, Fisher's exact test). Due to a high variance of the feature maps in the temporal lobes and insula, detection of FCD lesions in these regions appeared more difficult than in other brain regions with lower variance.