Novel MRI sequence-agnostic deep learning model to localize seizure onset zone in pediatric drug-resistance epilepsy

Motivation: To solve an incomplete sequence problem in localizing seizure onset zone (SOZ) using clinical MRI of children with drug-resistant epilepsy, Goal(s): We develop a sequence-agnostic deep learning approach that can localize SOZ even using subsets of clinical MRI sequence data. Approach: This consisted of 1) a sequence-agnostic model with cross-sequence distillation to train sequence specific and shared feature extractors for each sequence data and 2) a shared classification head to localize SOZ sites using a specific subset of multiple sequence data. Results: Our approach provided a high accuracy of 96%/89%/86% in classifying SOZ sites using five/four/three MRI sequence data of a validation cohort. Impact: Sequence-agnostic learning approach could accurately classify seizure onset zone (SOZ) using clinical MRI sequence data of pediatric drug-resistant epilepsy. It provided a new way to better generalize the SOZ classification in multi-center studies that suffer from missing sequence data problem