Authors :
Presenting Author: Bo Kyu Choi, MD – Yonsei University College of Medicine
Jungyon Yum, MD – Resident, Department of neurology, Yonsei University College of Medicine; Yu Rang Park, PhD – Associate Professor, Department of Biomedical Systems Informatics, Yonsei University College of Medicine; SooMi Cho, MD – Fellow, Department of Neurology, Yonsei University College of Medicine; Seungwon Song, MD – Resident, Department of neurology, Yonsei University College of Medicine; Kyung Min Kim, MD – Assistant Professor, Department of Neurology, Yonsei University College of Medicine; Youngno Yoon, MD – CEO, Bright Data LLC, Yongin, Korea
Rationale: Idiopathic generalized epilepsy is a common and important epilepsy, accounting for fifteen to twenty percent of all epilepsies. Among these, juvenile myoclonic epilepsy (JME) and epilepsy with generalized tonic-clonic seizures only (GTCA) have similar clinical presentations, posing diagnostic challenges even for specialists. While conventional MRI scans usually show no abnormalities, recent advanced techniques reveal microstructural and functional changes in JME. This study aims to develop a deep learning model for distinguishing between healthy controls (HC), JME, and GTCA patients based on brain MRI scans.
Methods: A total of 196 patients (127 JME, 37 GTCA, and 32 HC) with 3D T1-weighted images were enrolled. The patients were divided into training, test, and validation sets in an 8:1:1 ratio. All MRI images were preprocessed using FreeSurfer and then resampled using SimpleITK in Python, resulting in 3D images with size of (128,128,128). The deep learning model employed was a 3D-ResNet model. We conducted binary classification for each group and the performance of the model was evaluated in the test set by measuring the area under the receiver operating curve (AUROC), accuracy, sensitivity, and specificity.
We also applied the Gradient-weighted Class Activation Mapping (Grad-CAM) algorithm to understand which regions of the brain MRI images were the most important for the prediction using the deep learning model.Results: No significant differences were found among the three groups in terms of sex (female; JME, 46.5%; GTCA, 62.2%, HC, 50.0%, p-value, 0.126), age (JME, 23.3 years; GTCA, 26.6 years; HC, 28.9 years; p-value, 0.143), age at onset (JME, 15.2 years; GTCA, 16.6 years; p-value, 0.113), disease duration (JME, 8.1 years; GTCA, 10.1 years; p-value, 0.378), febrile seizures (JME, 11.8%; GTCA 18.9%; p-value, 0.264), or family history of epilepsy (JME, 15.7%; GTCA, 5.4%; p-value, 0.104). The deep learning model showed the AUROC value of 0.78. The AUROC predicting the classification of JME, GTCA, and HC were 0.76, 0.71, and 0.86, respectively. The accuracy, precision, recall, and F1-score of the model were 0.64, 0.41, 0.64, and 0.5, respectively.
Conclusions: In this study, we presented a deep learning-based model for distinguishing between JME, GTCA, and HC using 3D MRI data. We believe that this model could be of significant assistance in clinical settings for disease differentiation.
Funding: There is no funding source in this study.