Validation of the automated segmentation for the hippocampal subfields using Large Deformation Diffeomorphic Metric Mapping on 7T MR images and its application to volumetric and shape analysis for patients with temporal lobe epilepsy
Trinh, Thi Lan Huong
MetadataShow full item record
This thesis first aims to validate the Large Deformation Diffeomorphic Metric Mapping (LDDMM) method used in the automated segmentation of three major subfields of the hippocampus known as Cornu Ammonis (CA), dentate gyrus (DG) and subiculum (SUB) on the 7T T1 MR images with the isotropic resolution of 0.6 × 0.6 × 0.6 〖mm〗^3. Our results showed that using a population template to generate automated subfield segmentation gave better results than diffeomorphically transferring the subfields directly from one subject to another. Our averaged distances are within one voxel size (Left CA: 0.73 mm, left DG: 0.70 mm, left SUB: 0.66 mm, right CA: 0.79 mm, right DG: 0.71 mm, right SUB: 0.74 mm). Our Dice overlap coefficients are reasonable enough for thin structures and more consistent among different subfield than others (Van Leemput et al., 2012) (Left CA: 0.72, left DG: 0.76, left SUB: 0.76, right CA: 0.67, right DG: 0.76, right SUB: 0.70). Based on the validation results, we further implemented our volumetric and shape analysis for the hippocampus and its subfields to the study of patients with temporal lobe epilepsy (TLE) to detect any malformation that has been observed in histopathology of hippocampal sclerosis (HS). 25 subjects in the study were divided into four groups: Group A (control group), Group B (patients with unilateral temporal lobe epilepsy with clear 3T abnormality), Group C (patients with unilateral TLE with no clear 3T abnormality) and Group D (patients with subtle malformation of cortical development seen on 3T). The histogram with Gaussian kernel estimation of the Jacobian determinants showed enormous local shrinkage on all subfields in Group B, on left CA and left DG in Group D and left SUB in Group C. In the permutation test with the two-sample Kolmogorov-Smirnov statistics and in the Wilcoxon Mann-Whitney test with False Discovery Rate adjusted p-values, we obtained significant results indicating differences in the subfields among groups that can be explained by the two major pathological processes of HS observed in patients with TLE. Thus, we can detect specific changes on subfields for patients with TLE based on in-vivo ultrahigh field medical images.