Automated Intraoperative Visual Detection of Pediatric Epileptogenic Brain Lesions Using a Machine Learning Classifier

450,000 children with epilepsy in the United States suffer lifelong disability and are at risk of sudden death. Surgical treatment of epilepsy is limited by the ability to visually discriminate between normal and abnormal brain tissue using visual light surgical microscopes: resection of excessive tissue can lead to neurologic injury, while insufficient resection often does not lead to durable cures. We propose a machine-learning-based segmentation model to identify epileptogenic, abnormal tissue thereby improving accuracy of surgical resection. We collected 62 frames from the live stream of an operating microscope during a pediatric epilepsy surgery. We trained a random forest classifier to segment full frame images into pathological tissue or background. We achieved an average specificity of 0.99, sensitivity of 0.34, and intersection over union of 0.28, despite the constraints of a limited dataset. Machine learning classifiers can avoid misclassification of normal brain and provide an initial benchmark for future model development.