Task-Adapted Brain MRI Foundation Models for Focal Cortical Dysplasia Detection and Localization
2Department of Neurosurgery, Sheba Medical Center
Background and Purpose: Detection of focal cortical dysplasia (FCD) on MRI remains challenging, and existing automated approaches such as MELD-Graph rely on FreeSurfer-based cortical surface reconstruction. We evaluated whether task-specific fine-tuning of pretrained brain MRI foundation models (BrainIAC and BrainDINO) improves FCD detection and localization and generalizes to a completely independent external cohort, using MELD-Graph as a reference comparator.
Methods: Two foundation models were evaluated on public datasets (ds004199: 85 FCD/85 controls; IDEAS/ds005602: 30 FCD/82 controls; total n=282) using T1-weighted and FLAIR sequences, first in a frozen linear-probe configuration and then after task-specific fine-tuning. The Sheba Medical Center cohort was completely withheld from all model development, including training, fine-tuning, model selection, calibration, and threshold selection, and was used only for final external testing (19 FCD, 17 controls; n=36). MELD-Graph was evaluated on the same external cohort for direct comparison. Sensitivity, specificity, F1 score, and lesion localization (minimum distance between the predicted region and the manual lesion mask) were assessed.
Results: Frozen foundation models showed strong pooled internal performance (AUROC, 0.833-0.903) but substantially weaker performance in the independent Sheba cohort (AUROC, 0.455-0.570). Fine-tuning improved external performance: BrainIAC sensitivity, specificity, and F1 rose from 68.4%/41.2%/0.619 to 78.9%/64.7%/0.750, and BrainDINO from 47.4%/41.2%/0.474 to 63.2%/64.7%/0.649. By comparison, MELD-Graph achieved 73.7% sensitivity, 41.2% specificity, and F1 0.651. Fine-tuned BrainIAC showed the shortest localization distance (6.2 ± 0.7 mm), compared with BrainDINO (8.9 ± 2.2 mm) and MELD-Graph (7.3 ± 1.0 mm among 14 cases with a positive MELD cluster).
Conclusion: Task-specific fine-tuning substantially improved the external performance of pretrained brain MRI foundation models for FCD detection. Fine-tuned BrainIAC achieved the best overall performance and the shortest lesion localization distance, using a volumetric T1/FLAIR pipeline without cortical surface reconstruction. Larger multicenter studies are warranted to confirm our preliminary findings.