Magnetic resonance imaging-based ranking of resection-related cortical candidates for presurgical localization in epilepsy using a prior-aware graph attention transformer.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology and Nuclear Medicine, Xuanwu Hospital Capital Medical University, Beijing, China.
- Beijing Key Laboratory of Magnetic Resonance Imaging and Brain Informatics, Beijing, China.
- Key Laboratory of Neurodegenerative Diseases, Ministry of Education, Beijing, China.
- Beijing Key Laboratory of Innovation and Translation of Central Nervous System Radiopharmaceuticals, Beijing, China.
Abstract
Structural magnetic resonance imaging (MRI) is fundamental to presurgical localization in epilepsy, but subtle epilepsy-related abnormalities may not always be apparent on routine review. This study aimed to develop and externally validate an MRI-only graph attention transformer for ranking resection-related cortical candidates and to evaluate its reader-level utility. Graph Attention Transformer for Epilepsy-Related Candidate Zones (GATEZ) was developed using preoperative three-dimensional T1-weighted MRI from the publicly released IDEAS (Imaging Database for Epilepsy and Surgery) database. Participants with a 12-month International League Against Epilepsy class 1 outcome were split into training, validation, and internal test sets (n = 171/37/37). Each participant was represented as a 1000-parcel cortical graph with five regional morphometric features and an individualized Morphometric Inverse Divergence network; postoperative resection masks served as the surgical reference standard for model supervision. External validation used 183 consecutive surgical participants with epilepsy who underwent hybrid <sup>18</sup>F-fluorodeoxyglucose (<sup>18</sup>F-FDG) positron emission tomography (PET)/MRI, with MRI only used for model inference. Reader-level utility was evaluated in a blinded three-reader study comparing MRI alone, MRI plus <sup>18</sup>F-FDG PET, and MRI plus GATEZ. In the internal test cohort, GATEZ placed at least one resection-overlapping parcel within the Top-10 ranked candidates in 92% of participants, with a mean Top-10 positive predictive value of 62%. Performance remained stable in the independent external cohort, with an 87% Top-10 hit rate and 59% mean Top-10 positive predictive value. Node-level area under the precision-recall curve was .29 internally and .27 externally, indicating stable enrichment of resection-related regions among the highest ranked candidates. In the blinded reader study, MRI + GATEZ improved detection compared with MRI alone (74%-78% vs. 58%-66% across readers; adjusted p ≤ .001 for all readers) and performed similarly to MRI + FDG (78%-80% across readers; adjusted p ≥ .34 for all readers). GATEZ generates a concise Top-K shortlist of resection-related cortical candidates and may serve as a practical second-look aid for presurgical localization.