medwireNews: The Multicenter Epilepsy Lesion Detection (MELD) Project has trained an artificial intelligence (AI) model to identify focal cortical dysplasia (FCD) on magnetic resonance imaging (MRI), significantly improving performance over an existing algorithm.
In a test dataset of 260 patients with epilepsy (52% men; median age 18 years) due to radiologic- or histopathologic-confirmed FCD and 193 individuals with no neurologic conditions and normal MRI results (54% women; median age 29 years), the context-aware graph neural network (MELD Graph) detected lesions on MRI with a high positive predictive value (PPV) of 67% (70% sensitivity, 60% specificity).
By comparison, using the existing baseline, patch-based MELD Original Multilayer Perceptron (MLP) model, lesions were detected with a significantly lower PPV of 39% (67% sensitivity, 54% specificity).
In the 160 patients with epilepsy whose lesions were detected by both MELD Graph and MELD MLP, the MELD Graph had significantly higher precision, note the researchers.
MELD Graph takes input MRI data across whole cortical hemispheres and performs extraction of 34 surface-based features and 163,842 vertices and runs probability calculations to generate a report that depicts lesion location, model confidence, and lesional features, the researchers explain.
Mathilde Ripart (University College London, UK) and colleagues write in JAMA Neurology that, compared with previous models for detecting FCD, “the MELD Graph model incorporates whole-brain context, leading to multiple improvements: increased specificity, heightened sensitivity, especially to subtle FCD type I lesions, better accuracy, and well calibrated confidence estimates.”
The main test participants were from 20 international epilepsy surgery centers, and the findings in this dataset were corroborated in an independent cohort of 217 individuals from three additional epilepsy centers, of whom 116 had FCD (53% women; median age 22.5 years) and 101 were controls (50% women; median age 27.5 years).
MELD Graph detected lesions in this independent group with a PPV of 76% (72% sensitivity, 56% specificity) compared with a PPV of 46% (77% sensitivity, 47% specificity) with MELD MLP.
The investigators note that when they looked at the performance of MELD Graph in the test dataset according to various demographic factors, they found it detected 63.7% of lesions that had previously been missed by radiologists and 81.6% of lesions in histopathologically confirmed patients who were seizure free 1 year after surgery.
MELD Graph was also able to detect 84.6% of 13 “particularly subtle FCD type I lesions,” the researchers say, in addition to 75.4% and 76.3% of type IIA and IIB lesions, respectively.
“The development of tools, like MELD Graph, that can successfully localize FCDs across a range of scanners, ages, and subtypes, including detecting 64% of MRI–negative FCDs, therefore, may improve both seizure freedom and developmental outcomes through earlier diagnosis and surgery,” say Ripart et al.
They highlight the importance of the model's interpretability “for incorporating AI into radiological review,” noting that its “[h]igh-confidence predictions are rarely incorrect, while lower confidence outputs may highlight subtle abnormalities that warrant careful review.”
The MELD Project is releasing Graph as an open-source tool, “to enable independent validation and reuse, including prospective studies to assess whether it accelerates diagnosis, alters treatment, and ultimately improves outcomes.”
The researchers conclude that “[f]uture work, applying deep learning to multimodal data [...] will enable complex features to be learned that may not be visible or known to neuroradiologists.”
medwireNews is an independent medical news service provided by Springer Healthcare Ltd. © 2025 Springer Healthcare Ltd, part of the Springer Nature Group
