Electronic Thesis/Dissertation


Feature Selection and Classification methods for Lateralization and Localization of Epileptic Seizure Foci Open Access

Downloadable Content

Download PDF

Nearly one-third of patients with epilepsy still remain refractory to seizures even if treated with the pharmacotherapy that brings remission to the majority [1]. The standard of care to potentially cure the drug-resistant epilepsy is often surgical resection. However, despite the growing work demonstrating the effect of surgical resection, this method is still underutilized [2] since the presurgical analysis couldn't always be available due to the limits of conventional brain imaging technique to visualize the foci. Considering the limitations of conventional braining imaging techniques, a new method like Positron Emission Tomography (PET) was involved in contributing complementary diagnostic information to the presurgical evaluation. Although this method could be helpful regarding the lateralization of the foci, the precise localization is still challenging due to the low resolution and lack of anatomic landmarks. Thus, it is worthwhile to examine whether registering the PET to the anatomic parcellation provided by Magnetic Resonance Imaging (MRI) could lead to a better way to provide a more reliable prediction for the lateralization and localization of the seizure foci.In this paper, our goal is to build up a general, automated diagnostic tool that provides the prediction of seizure foci lateralization and localization (as temporal or extra-temporal). To achieve that, we use Freesurfer software to generate the virtual cortical surface where we applied the region-based study of PET based on the anatomic segmentation on it. Then, regarding the lateralization and localization, we trained two sets of Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and Naïve Bayes classification models using the relative standard uptake value (rSUV) from the region-based study as features. The novelty of our work is that it utilizes a machine learning method to model the medical features provided by PET scans to predict the presence of a seizure foci, and applies feature selection strategies to the region-based PET analysis.

Author Language Keyword Date created Type of Work Rights statement GW Unit Degree Advisor Committee Member(s) Persistent URL