Enhanced detection of subtle cortical abnormalities in focal epilepsy using 7 T MRI surface-based models and graph neural networks

Purpose: MRI detection of subtle focal cortical dysplasia (FCD)-like abnormalities remains challenging in focal epilepsy. Higher signal-to-noise ratio and spatial resolution offered by ultra-high-field 7T MRI and surface-based graph-neural-network (GNN) analysis may improve detection of subtle cortical abnormalities. We evaluated whether combining 7T MRI with a surface-based GNN classifier improves lesion detection in focal epilepsy of suspected structural origin. Methods: We analyzed paired 7T and 3T MRI datasets from 87 patients with focal epilepsy (78.1% pediatric) and 10 internal healthy control individuals. We processed T1-weighted and Fluid-Attenuated-Inversion-Recovery MRI using a surface-based framework and a pre-trained GNN classifier developed within the Multi-centre-Epilepsy-Lesion-Detection project. We compared classifier outputs with expert visual MRI assessment, clinical and surface electroencephalography (EEG) localization (all patients), stereo-EEG (ten patients) and histopathological (17 patients) findings. We evaluated diagnostic yield and lesion conspicuity, and performed within-subject comparisons between 7T and 3T. Results: Following quality controls, we included 70 patients. The 7T MRI-based classifier identified lesion clusters concordant with visual 3T MRI and electroclinical localization in 25/37 (67.6%) MRI-positive patients, electroclinical-concordant clusters in 15/33 (45.4%) 3T MRI-negatives, stereo-EEG-concordant clusters in 7/10 (70.0%) patients and surgically-concordant clusters in 11/17 (64.7%). Among classifier-positive patients (40/70, 57.1%), 7T allowed detection of previously hidden lesions in 15/40 (37.5%) patients, and improved detection of known lesions in 11/40 (27.5%). Conclusion: Combining 7T MRI with surface-based GNN analysis improves detection and characterization of FCD-like abnormalities in focal epilepsy, particularly in patients with unrevealing 3T MRI, supporting the adoption of advanced neuroimaging in presurgical epilepsy assessment.