Identification of MRI Features of Temporal Lobe Epilepsy with Deep Learning
Abstract number :
3.135
Submission category :
2. Translational Research / 2D. Models
Year :
2022
Submission ID :
2203908
Source :
www.aesnet.org
Presentation date :
12/5/2022 12:00:00 PM
Published date :
Nov 22, 2022, 05:22 AM
Authors :
Allen Chang, MS – Medical university Of South Carolina; Rebecca Roth, B.A. – Emory University; Eleni Bougioukl, B.S. – Brandeis University; Dan Drane, Ph.D. – Emory; Robert E. Gross, M.D. Ph.D. – Emory; Simon Keller, Ph.D. – University of Liverpool; Theodor Ruber, M.D. – University Hospital Bonn; James Welch, B.S. – Medical University of South Carolina; Leonardo Bonilha, M.D. Ph.D – Emory University; Ezequiel Gleichgerrcht, M.D. Ph.D. – Medical University of South Carolina
Rationale: Surgery success for medication refractory temporal lobe epilepsy (TLE) is associated with the correct identification of neuroimaging signs of pathology. Thus, techniques like computer assisted technologies that can aid in locating these histological abnormalities on neuroimaging would be beneficial in surgical treatment success of medication refractory TLE.
Methods: In this experiment, we tested the ability of a convolutional neural network (CNN) algorithm to identify TLE disease versus healthy controls and individuals with Alzheimer’s disease using T1-weighted magnetic resonance imaging (MRI) scans. Furthermore, we leveraged feature visualization techniques to identify potential regions the CNN used to differentiate disease types.
Results: We found that our properly trained CNN models had a mean accuracy of 86.84% (SD = 1.33%) for disease prediction compared to our shuffled “random chance” control CNN models with a mean accuracy of 67.16% (SD = 1.04%). Our feature weight analysis found that the CNN activation pattern was disease dependent and included regions that are frequently associated with TLE (i.e., temporal regions) as well as extra-temporal regions, demonstrating the importance of a whole-brain approach.
Conclusions: We found that automated techniques, like deep learning, are quite successful in the identification of TLE, and therefore, have the potential to increase diagnostic accuracy and in turn surgical success if employed during the clinical evaluation of TLE.
Funding: CAPES - R01NS110347 (LB, EG)
Translational Research