NIDDK (NIH)

Neural Networks for Detecting Subtle Epileptogenic Lesions and Supporting Clinical Decision-Making

The webinar introduced the Multi‑center Epilepsy Lesion Detection (MELD) project, showcasing a graph convolutional neural network designed to identify subtle focal cortical dysplasia (FCD) lesions in drug‑resistant epilepsy patients. The session also highlighted NIDDK’s role in providing data resources and its forthcoming cloud‑based analytics environment. Radiologists miss roughly 30% of FCD lesions on standard MRI, prompting the development of AI tools such as MELD’s multilayer perceptron (MLP) and DeepFCD. By harmonizing multi‑site MRI features, applying ComBat batch correction, and training on over 600 patient datasets, the new graph neural network achieved a 67% overall detection rate and correctly identified 63% of lesions previously deemed MRI‑negative. During a live demonstration, participants evaluated a patient scan, revealing how raw feature visualizations increased confidence in lesion localization. The model’s predictions overlapped well with expert‑annotated masks, yet produced an average of two false‑positive clusters per case, resulting in a low positive predictive value and requiring radiologist review. If refined, this technology could reduce reliance on invasive diagnostics, shorten time to surgery, and improve seizure‑freedom outcomes. The NIDDK cloud platform promises broader access to harmonized datasets and AI tools, accelerating collaborative research and clinical translation.