Abstracts

Rationale: Hippocampal sclerosis is the hallmark of drug-resistant temporal lobe epilepsy (TLE). Developments in MRI hardware generate images of brain anatomy in unprecedented detail, allowing the non-invasive identification of hippocampal subfields. We propose a novel automated subfield segmentation algorithm, SurfPatch, which combines multi-template feature matching with deformable parametric surfaces and vertex-wise patch sampling. We evaluated the accuracy with respect to manual subfield labels and assessed its clinical utility by automatically lateralizing the seizure focus in individual TLE patients. Methods: SurfPatch generates labels of the hippocampal subfields CA1-3, CA4-DG, and subiculum (SUB) based on either high-resolution (0.6 mm isotropic MPRAGE) or conventional (1.0 mm isotropic MPRAGE) T1-weighted MRI data. It relies primarily on patches, which are small image neighborhoods centered around a voxel or a vertex of interest. During the training step, SurfPatch builds a mean patch surface and a standard deviation (SD) patch surface across the entire template library (Fig. 1A). For segmentation, it nonlinearly warps each template surface to the test MRI image, re-computes patch features across warped surfaces, and normalizes features using surface-based z-scoring. It then selects a subset of templates, builds an average surface, and performs a deformation for final segmentation (Fig. 1B). We evaluated algorithm performance based on a publically available dataset from 25 healthy controls, for which hippocampal subfields had previously been manually delineated (http://www.nitrc.org/projects/mni-hisub25/). Performance of the algorithm was evaluated using Dice index in a leave-one-out manner. We applied SurfPatch to a cohort of 17 TLE patients, testing the ability of a classifier relying on segmented subfields volumes to lateralize the seizure focus. Results: Leave-one-out cross-validation using high-resolution T1-weighted at 0.6 mm resolution yielded excellent accuracies across all subfields (CA1-3: 872; CA4-DG: 833; SUB: 852). Segmenting subfields using conventional T1-weighted 1 mm resolution images provided similar performance (CA1-3: 862; CA4-DG: 814; SUB: 824; Fig. 2). Lateralization of the seizure focus in TLE patients was highly accurate (>93 %) when using SurfPatch, both based on submillimetric and millimetric T1-weighted images. Conclusions: SurfPatch automatically segments hippocampal subfields with high accuracy at various image resolutions. High performance on conventional MPRAGE MRI, the most commonly used anatomical sequence, promotes its use in the pre-surgical work-up of TLE and also for big data MRI initiatives. Funding: This research was funded by the Canadian Institutes of Health Research (CIHR MOP-57840 and CIHR MOP-123520). BCB received a CIHR postdoctoral fellowship.