UC San Diego

Temporal lobe epilepsy (TLE) is the most common form of focal epilepsy in adults. Many of those affected do not respond well to anti-seizure medications and require neurosurgical intervention. For these patients, locating the seizure focus to the left or right hemisphere is a critical first step in surgical planning. Through the use of Convolutional Neural Networks (CNN) that can survey medical diagnostic images in finer detail and extract more convoluted information than humans, we proposed that it will aid in the localization of epileptogenic regions (i.e., seizure generating regions) in the brain. We explored different pre-trained and novel CNN architectures trained on 364 patients with TLE from seven different epilepsy centers to ensure generalizability. Our findings show that a CNN model outperforms a more standard linear model like a Logistic Regression (LR) in performing the task of side-of-onset classification, with the best CNN model outperforming the average LR model by 17.39% in classification accuracy. The model also shows that information important for determining side-of-onset is not limited to the mesial temporal lobe (MTL) but is also located in extra-temporal regions like the parietal lobe, precentral gyrus, and cingulate gyrus. This study shows that the classification problem of Left versus Right TLE patients benefits from a more complex, nonlinear model and whole-brain information and therefore medical examination of TLE would benefit from incorporating machine learning to aid in clinician-led localization of the epileptogenic zone.