Automated segmentation of postsurgical resection cavities on magnetic resonance imaging in focal epilepsy: A Multicentre Epilepsy Lesion Detection study
Jieun Seo, Mathilde Ripart, Helene Kaas, Cornelius Kronlage, Ben Sinclair, Lucy Vivash, Merran R. Courtney, Terence J. O'Brien, Siby Gopinath, Harilal Parasuram, Sedat Kandemirli, Natally Alarab, Lillian Lai, Marcus Likeman, Kai Zhang, Jiajie Mo, Georgian Ciobotaru, James Galea, Philip Sequeiros‐Peggs, Khalid Hamandi, Hua Xie, Venkata Sita Priyanka Illapani, William D. Gaillard, Nathan T. Cohen, Alexander G. Weil, Florence Henrichon‐Goulet, Kenza S. Lahlou, Aristides Hadjinicolaou, Agustín Ibáñez, Gonzalo M. Rojas‐Costa, Horst Urbach, Lara Bücheler, Marcel Heers, Adrián Valls Carbó, Rafael Toledano, Giulia Nobile, Costanza Parodi, Domenico Tortora, Alessandro Consales, Antonella Riva, Mariasavina Severino, Martin Tisdall, Felice D'Arco, Kshitij Mankad, Aswin Chari, Maria H. Eriksson, Rory J. Piper, J. Helen Cross, Torsten Baldeweg, Sofia González‐Ortiz, Jose Pariente, Nuria Bargalló, Yawu Liu, Reetta Kälviäinen, Carmen Barba, Matteo Lenge, Renzo Guerrini, Masaki Iwasaki, Daichi Sone, Hiroyuki Maki, Tomoki Imokawa, Noriko Sato, Julien Jung, Francisco Sepulveda, Daniel Mansilla, Andres Goycoolea, Ingeborg Lopez, Antonio Napolitano, Alessandro De Benedictis, Luca De Palma, Maria Camilla Rossi‐Espagnet, Nikolaos Kondylidis, Kostakis Gkiatis, Kyriakos Garganis, Joshua Pepper, Stefano Seri, John S. Duncan, Clarissa L. Yasuda, Lucas Scárdua‐Silva, Marina K. M. Alvim, Fernando Cendes, Antonio G. Gennari, Ruth O'Gorman Tuura, Georgia Ramantani, Mariam Josyula, Joel Stein, Nishant Sinha, Kate Davis, R. Edward Hogan, Luigi Maccotta, Sophie Adler, Konrad WagstylAbstract
Objective
Quantitative assessment of extent of tissue resection following epilepsy surgery requires accurate delineation of the resection cavity on postoperative magnetic resonance imaging (MRI). Current methods for resection cavity masking are time‐consuming and labor‐intensive, and existing automated approaches exhibit variable segmentation accuracy, particularly on extratemporal resections. We developed MELD‐PostOp, a deep learning tool trained and evaluated on a large, heterogeneous cohort to automatically segment resection cavities.
Methods
The study included 1.5‐ and 3T postoperative three‐dimensional T1‐weighted MRI images from the Multicentre Epilepsy Lesion Detection (MELD) project ( n subjects = 969, 27 centers) and from the EPISURG dataset ( n = 133). The cohort included 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 combined with the original 285 to train the final MELD‐PostOp model ( n = 965). A Stratified Test Cohort ( n = 50) and Independent Test Cohort ( n = 87) were 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 (Epic‐CHOP, ResectVol, and RESSEG).
Results
MELD‐PostOp achieved a median DSC of .85 and HD95 of 3.61 on the combined test cohort, outperforming Epic‐CHOP (DSC .69, HD95 9.67), ResectVol (DSC .66, HD95 15.05), and RESSEG (DSC .43, HD95 32.67), with significant improvements seen in both temporal and especially extratemporal resections. The model detected 98.5% (135/137) of resection cavities. MELD‐PostOp runtime was 17 s per MRI, compared to 612 s (ResectVol), 3205 s (Epic‐CHOP), and 4 s (RESSEG). MELD‐PostOp performance remained high across clinical and imaging subgroups (median DSC > .8).
Significance
MELD‐PostOp is an open‐source research tool that provides an accurate, efficient, and generalizable solution for postoperative resection cavity segmentation using only postoperative MRI scans.