Advanced Deep Learning Models Enhance Piriform Cortex Segmentation for Epilepsy Surgery

Motivation: The piriform cortex (PC) is emerging as an important structure for surgery for drug-resistant mesial temporal lobe epilepsy, where resection improves postsurgical outcomes. Current segmentation methods remain suboptimal, necessitating advanced solutions. Goal(s): To compare state-of-the-art deep learning (DL) models and label fusion for automated PC segmentation. Approach: MRI scans from EPISURG (n=49, epilepsy side balanced) and Hammersmith (n=8 controls, gender balanced) were preprocessed (bias correction, skull-stripping, registration) and evaluated through five-fold cross-validation (70/30 training/validation) using segmentation metrics. Results: Transformer-inspired/based models achieved the highest accuracy, with MedNeXtL reaching median Dice scores of 0.90 (right) and 0.91 (left), outperforming nnU-Net and label fusion methods. Impact: Advanced DL models with attention mechanisms, such as MedNeXtL and SwinUNETR, significantly enhance piriform cortex segmentation accuracy. This improvement holds promise for refining surgical margins for patients with drug-resistant temporal lobe epilepsy, potentially leading to better postsurgical outcomes.