Automated Whole-brain Focal Cortical Dysplasia Detection using MR Fingerprinting with Deep Learning

Motivation: MRI-based detection of focal cortical dysplasia lesions aids the surgical treatment of epilepsy. This process can be enhanced by deep learning models. Incorporating quantitative tissue property information from MR Fingerprinting (MRF) may further improve model performance. Goal(s): To establish an automated pipeline for whole-brain FCD detection using MRF-based inputs. Approach: We used nnU-Net to train a model that took the following MRF-based inputs from 40 patients and 67 healthy controls: synthetic T1w image, T1 and T2 z-score maps, morphometric z-score maps, and tissue fraction maps. Model performance was evaluated by leave-one-out validation. Results: The model achieved 80% sensitivity and 1.7 false positive clusters/patient. Impact: Multiparametric MRF features from a single scan provide promising inputs for developing an AI tool to detect subtle epileptic lesions, enhancing noninvasive presurgical evaluation for patients with pharmacoresistant epilepsy.