Convolutional Neural Networks for Diagnostic Radiology: Seizure Lateralization in Temporal Lobe Epilepsy
Abstract number :
3.085
Submission category :
2. Translational Research / 2A. Human Studies
Year :
2022
Submission ID :
2204184
Source :
www.aesnet.org
Presentation date :
12/5/2022 12:00:00 PM
Published date :
Nov 22, 2022, 05:23 AM
Authors :
Erik Kaestner, PhD – University of California San Diego; Jun Rao, M.S. – University of California San Diego; Allen Chang, M.S. – Medical University of South Carolina; Irene Wang, Ph.D. – Cleveland Clinic; Robyn Busch, Ph.D. – Cleveland Clinic; Simon Keller, Ph.D. – University of Liverpool; Theodor Rüber, M.D. – University of Bonn; Daniel Drane, Ph.D. – Emory University; Travis Stoub, Ph.D. – Rush University; Ezequiel Gleichgerrcht, M.D. – Medical University of South Carolina; Leo Bonilha, M.D. – Emory University; Kyle Hasenstab, Ph.D. – San Diego State University; Carrie McDonald, Ph.D. – University of California San Diego
Rationale: A new frontier in diagnostic radiology is the inclusion of machine-assisted support tools that facilitate the identification of subtle lesions often not visible to the human eye. Structural neuroimaging plays an essential role in the identification of lesions in patients with epilepsy, which typically coincide with the seizure focus.
Methods: Here we explore the potential for a machine learning algorithm to determine lateralization of seizure onset using T1-weighted structural MRI scans as input. Using a dataset of 359 patients with temporal lobe epilepsy (TLE) from 7 surgical centers, we tested the ability of a convolutional neural network (CNN) algorithm based solely on T1 images to classify seizure laterality concordant with clinical team consensus at each center. This CNN was compared to a randomized model to index improvement from chance and to a logistic regression that included hippocampal volume to assay whether the CNN added to performance beyond current clinically-available measures. Furthermore, we leveraged a feature visualization technique to identify regions the CNN used to classify patients.
Results: Across 100 runs, the CNN model was concordant with clinician lateralization on average 78% (SD= 5.1%) of runs with the best performing model achieving 89% concordance (Figure 1). The CNN outperformed the randomized model (average concordance of 51.7%) on 100% of runs with an average improvement of 26.2% and outperformed the hippocampal volume model (average concordance of 71.7%) on 85% of runs with an average improvement of 6.25%. Feature visualization maps revealed that in addition to the medial temporal lobe, regions in the lateral temporal lobe, cingulate, and precentral gyrus all aided in the classification of right versus left TLE (Figure 2A) as compared to statistically significant differences which were limited to the hippocampus (Figure 2B).
Conclusions: This proof of concept study demonstrates that a CNN applied to structural MRI data could aid in the clinician-led localization of epileptogenic zone and identified extra-hippocampal regions to which radiological attention should be directed.
Funding: T32 MH018399 (E.K), K01NS12483 (E.K.), R01NS124585 (C.M.), R01NS122827 (C.M.)
Translational Research