Automated Segmentation of Post-Surgical Resection Cavities on MRI in Focal Epilepsy: a MELD Study

作者：J. Seo, M. Ripart, H. Kaas, B. Sinclair, L. Vivash, M. Courtney, T. O'Brien, S. Gopinath, H. Parasuram, S. Kandemirli, N. Alarab, L. Lai, M. Likeman, K. Zhang, J. Mo, G. Ciobotaru, J. Galea, P. Sequeiros-Peggs, K. Hamandi, H. Xie, V. Illapani, W. Gaillard, N. Cohen, A. G. Weil, F. Henrichon-Goulet, K. Lahlou, A. Hadjinicolaou, A. Ibanez, G. M. Rojas-Costa, H. Urbach, L. Bucheler, M. Heers, A. Valls Carbo, R. Toledano, G. Nobile, C. Parodi, D. Tortora, A. Consales, A. Riva, M. Severino, M. Tisdall, F. D'arco, K. Mankad, A. Chari · 发表于：medRxiv · 年份：2026 · DOI：10.64898/2026.02.26.26347093 · 研究领域：Medicine

Objective Quantitative assessment of extent of tissue resection following epilepsy surgery requires accurate delineation of the resection cavity on postoperative MRI. Current methods for resection cavity masking are time-consuming and labour-intensive, while existing automated approaches exhibit variable segmentation accuracy, particularly on extra-temporal resections. We developed MELD-PostOp, a deep learning tool trained and evaluated on a large, international, heterogeneous cohort to automatically segment resection cavities. Methods The study included 1.5 and 3T postoperative 3D T1-weighted MRI images from the Multicentre Epilepsy Lesion Detection (MELD) project (nsubjects=969, 27 centres) and from the EPISURG dataset (n=133). The cohort included both children and adults, alongside a range of resection locations, pathologies, and MRI characteristics. Resection cavities were individually segmented in 285 subjects and used to train an nnU-Net prototype model. The prototype model was used to generate an additional 680 resection masks, which were subsequently quality-controlled, edited and then combined with the original 285 to train the final MELD-PostOp model (n=965). A Stratified (STC; n=50) and Independent Test Cohort (ITC; n=87) were masked and withheld for model evaluation. Performance was evaluated using Dice Similarity Coefficient (DSC), 95th percentile Hausdorff distance (HD95), number of predicted clusters and inference runtime; and compared against established tools...