Localization of Epileptogenicity Using Incomplete MRI Sequence Data in Children with Seizure Onset Zone Involving Temporal Lobe

Soumyanil Banerjee, Qisheng He, Min-Hee Lee, Ming Dong, Eishi Asano, Jeong-Won Jeong

Abstract: Localization of seizure onset zone (SOZ) is challenging in children with drug-resistant epilepsy. This clinical challenge becomes even more difficult when there are one or more missing MRI sequence data during the preoperative evaluation. To address this problem, we propose a novel mechanism where distillation is extensively used inside a deep neural network for prediction of SOZ in children with SOZ involving temporal lobe. We use sequence specific feature extractors to extract MRI sequence specific features, and shared feature extractors to handle the missing MRI sequences during the testing time. We utilize distillation across the different MRI sequences, both at the feature map level and the class-level, so that knowledge from the fused features is passed on to the individual MRI sequences. We addressed the non-SOZ vs SOZ class imbalance and applied our approach to perform SOZ classification of temporal lobe nodes of children with epilepsy. Our algorithm significantly outperforms all the baselines with a higher balanced accuracy, when all MRI sequences are present. Additionally, the same trained model could also handle the missing sequences during testing.

External IDs:dblp:conf/isbi/BanerjeeHLDAJ24