Abstracts

Rationale: Identification of epileptogenic lesions plays an important role in presurgical evaluation of medically refractory epilepsy patients. In some patients structural MRI appears normal, while others have extensive lesions. In each case, quantitative evaluation may add additional information, but often requires significant manual intervention and registration to an atlas, which can be difficult in subjects with abnormal neuroanatomy. Saad et al. (1) described a framework for analysis of multi-contrast structural MR images allowing for tissue segmentation without registration to a template. We applied this segmentation to MRIs of patients with medically refractory epilepsy to assist in the visualization of a variety of epileptogenic lesions.Methods: MRIs were obtained on 5 adult healthy volunteers (1 female, ages 21-40, median age of 31), and 20 patients with medically refractory focal epilepsy. Study data were acquired on NIH Clinical Center 3T Philips Achieva MRI Scanners. Imaging included a series of 3D volumetric images (MPRAGE, T2, FLAIR, and Gradient Echo with 2 flip angles). All scans were registered to the MPRAGE using AFNI. Field bias adjustment and feature scaling in each sequence was carried out by the procedure described in Saad et al. (1). Healthy volunteer MPRAGE images were processed with FreeSurfer; white matter and cortical surfaces were visibly inspected for accuracy. FreeSurfer segmentations were used to create masks for gray matter (GM), white matter (WM), and cerebral spinal fluid for each subject, along with an additional “void” mask. Training sets for each contrast and class combination were created using the 5 healthy volunteers’ images. For the remaining subjects, voxel class likelihood was calculated using the naïve Bayes method described in Saad et al (1). The resulting class probability maps were then visually inspected by a trained epileptologist in patients with medically refractory epilepsy.Results: Two patients’ focal cortical dysplasias were identified by displaying voxels with a classified GM probability of greater than 0.25. The same contrast was used to visualize another patient’s periventricular nodular heterotopias based on their location. This procedure was also able to provide peri-lesional GM-WM segmentation in patients with more complex cortical abnormalities. This could provide useful presurgical planning information by identifying areas of GM adjacent to patients’ lesions.Conclusions: This framework allows for the creation of images that assist in the visual identification of GM and WM, as well as abnormal tissue in epilepsy patients with minimal training datasets and no manual intervention. Future work will evaluate additional contrasts and textures, as well as further quantitative analyses to optimize visualization of different pathologies. References: Saad, Z et al. (2015), Framework for Generating Class Priors From Multi-Contrast Images Without Group Volume Templates, OHBM poster, Honolulu, HI.