Anatomical-connectivity based parcellation of hippocampus and medial temporal lobe epilepsy classification

Unilateral medial temporal lobe epilepsy (mTLE) have similar symptom in clinic. Previous studies indicate that it's difficult to find significant differences when comparing the structure of left and right mTLE directly. As is known to all, hippocampus sclerosis often causes mTLE. Therefore, it may be essential to detect the internal structure of hippocampus (HC) to confirm the lateralization of mTLE focus. We tracked fibers between each voxel of HC to each voxel of the whole brain by Diffusion Tensor Imaging and divide each side of the HC into 2 to 4 subregions based on anatomical connectivity (AC). Then, AC comparisons (HC and HC subregions to the target regions) were made between left mTLE, right mTLE and normal controls pairwisely for classification. Feature fusion was made to elevate the classification accuracy. With use of the leave-one-out crossvalidation strategy, the Support Vector Machine classifier could achieve accuracy of 81.6% for left mTLE versus controls, 79.6% for right mTLE versus controls, and 83.7% for left mTLE versus right mTLE. Therefore, AC based parcellation of hippocampus can be applied into clinic to find more specific focus of mTLE.