Can neural networks support identification of epileptogenic foci in children with suspected focal cortical dysplasia (FCD)?

Motivation: Success of surgical resection in focal epilepsy is highly dependent on identification of epileptogenic foci. Machine learning algorithms are proving useful in supporting detection of pediatric focal cortical dysplasia (FCD) lesions. Goal(s): Evaluating 2 algorithms predicting FCD lesions, developed by the multi-centre epilepsy lesion-detection project (MELD): the multi-layer-perceptron-based classifier [MELD_MLP] and the newer graph-neural-network based algorithm [MELD_GRAPH] aiming to increase specificity/accuracy. Is the new model better for routine clinical integration? Approach: Comparison of true and false-positive rates for MELD_Graph vs MELD_MLP on 17 children Results: MELD_MLP detects more lesions than MELD_Graph. However MELD_Graph produces less false positives, resulting in better positive predictive value Impact: Clinical validation is essential for AI-based Focal Cortical Dysplasia lesion-detection algorithms. A local trade-off between sensitivity and specificity is necessary when selecting an appropriate decision support AI/machine-learning tool.