Fcdnets: Surface-Graph Neural Networks for Automated Detection of Focal Cortical Dysplasia on Multi-Contrast Mri Surface Maps

Focal cortical dysplasia (FCD) is a leading cause of drug-resistant focal epilepsy, especially in children. However, up to 30−40% of FCD lesions are “MRI-negative,” eluding visual detection on standard scans. Even expert neuroradiologists miss about one-third of these subtle lesions. Prior automated FCD detection approaches have shown promise but face important limitations. Conventional morphometry pipelines (e.g., FreeSurfer-based thickness or junction maps) rely on hand-tuned thresholds and often struggle with small cohorts and site-specific bias. Voxel-level deep learning methods, including 3D CNN and transformer models, can improve sensitivity but tend to produce excessive false positives and lack interpretability. In this work, we propose FCDNets, a graph neural network (GNN) that learns from multi-contrast MRI cortical surface maps to detect FCD lesions. Trained using only open-access MRI datasets (85 patients and 85 controls from OpenNeuro ds004199; 101 patients and 177 controls from MELD-Public), FCDNets achieves vertex-level AUROC 0.975±0.006, lesion Dice 0.64±0.12, and 82% patient sensitivity at ≤5 cm2 false-positive cortex in cross-site five-fold CV. Compared with a 3-D CNN baseline, FCDNets yields +28 percentage points higher sensitivity (82 % vs. 54 %) and 45 % fewer false-positive clusters (0.9 vs. 1.6). On MELD-Public, it maintains AUROC 0.782 and 63% sensitivity. The surface-based approach produces interpretable lesion probability maps aligned with known FCD imaging biomarkers. This open, multi-center study demonstrates the potential of surface-GNNs to aid in localizing subtle epileptogenic lesions that were previously MRIoccult, which may accelerate diagnosis and surgical treatment for drug-resistant epilepsy.