Authors :
Presenting Author: Erik Kaestner, Ph.D. – University of California San Diego
Jay Sawant, M.S. – University of California San Diego
Taha Gholipour, M.D. – University of California San Diego
Ezequiel Gleichgerrcht, M.D., Ph.D. – Emory University
Anees Abrol, Ph.D. – Georgia State University
Vince Calhoun, PhD – Georgia State University
Kyle Hasenstab, Ph.D. – San Diego State University
Leonardo Bonilha, M.D., Ph.D. – University of South Carolina
Carrie McDonald, PhD – University of California, San Diego
Rationale:
Epilepsy diagnosis and management relies heavily on human interpretation of brain magnetic resonance imaging (MRI), which would greatly benefit from artificial-intelligence (AI) support tools. Radiological interpretation requires multiple tasks, including diagnosis, classification, and localization, but the best approach to training AI on such multifaceted processes remains unclear. Here, we utilize three large epilepsy MRI datasets (n=3169) to perform two tasks; first to detect whether a temporal lobe epilepsy (TLE) signature is present , and second, to determine whether the TLE signature is associated with left or right seizure onset.
Methods:
We compared sequential models that first classify TLE versus healthy controls (HC), then lateralize L-TLE or R-LTE, against a simultaneous model trained to distinguish all three classes in a single simultaneous step. T1-weighted images were fed into an EfficientNetV2 model for 10 runs of 5-fold split (50 datapoints). Class prediction, model confidence, and saliency maps were derived.Results:
Sequential models outperformed the simultaneous model on both tasks (both ps< .001), with an average ~2.8% accuracy increase for discriminating HC from TLE and an average 12.7% accuracy increase for distinguishing L-TLE from R-TLE. The sequential model had a higher confidence in correct HC (p< .001; d = 0.63) and TLE (p< .001; d = 0.27). We found that sequential and simultaneous models had nearly mirrored model salience in HC versus TLE [HC salience: rho=0.97, p< .001; TLE salience: rho=0.95, p< .001], with strong foci on the known TLE limbic pattern involving the hippocampus, para-hippocampal cortical regions, cingulate cortex, and lateral temporal regions. However, the L-TLE versus R-TLE task had significantly lower correlation between sequential and simultaneous models (all Fisher’s Z >10.5, ps< .001); the sequential model focused less on subcortical regions like the thalamus and hippocampus and focused more on distributed cortical pathology. Across the sequential cascade, 95.4% of TLE patients had accurate classifications in either HC versus TLE and/or L-TLE versus R-TLE tasks (69.6% in both, 13.9% in HC versus TLE only, and 11.6% lateralization only).