An open-source artificial intelligence (AI) tool developed by Sophie Adler, a medical student at University College London (UCL), and her colleagues, is improving the detection of epileptic lesions often missed by traditional MRI analysis. The tool, presented in Sophie’s talk at the American Epilepsy Society (AES) Annual Meeting 2024, has shown potential in identifying focal cortical dysplasia (FCD) and hippocampal sclerosis, two key contributors to epilepsy.
FCD, a malformation of cortical development, is found in 42%-55% of epilepsy surgery patients with negative MRI findings, but it is often undetected by human reviewers. Previous machine learning tools for FCD detection have faced challenges, particularly with high rates of false positives.
The AI solution
Adler and her colleagues developed a neural network that analyses 33 surface-based features of the brain. Using data from over 600 patients with epilepsy and nearly 400 controls from 22 global epilepsy centres, the model achieved a sensitivity of 59% and specificity of 54%, rising to 67% with adjustments for lesion border uncertainty.
To improve accuracy, the team employed a graph convolutional neural network. This method considers the context of surrounding brain regions, reducing false positives and increasing the positive predictive value to 67%. Notably, the algorithm detected lesions in 63% of MRI-negative cases that had been overlooked by human reviewers.
Adler and her team are committed to ensuring accessibility through their open-source approach and have organised workshops to train clinicians and researchers. They also acknowledge the need to address underrepresentation of data from low-income countries to ensure the tool’s global relevance.
Clinical applications and future directions
The algorithm is being trialled with radiologists and surgeons for pre-surgical planning, including guiding electrode placement. A complementary tool to detect hippocampal sclerosis is also under development, with plans to integrate the systems for detecting multiple pathologies.
Read ACNR Epilepsy articles
