Diffusion spectrum imaging-based machine learning for temporal lobe epilepsy lateralization.
Authors
Affiliations (3)
Affiliations (3)
- Department of Radiology and Nuclear Medicine, Xuanwu Hospital, Capital Medical University, Beijing 100053, China; Department of Radiation Oncology, Xuanwu Hospital, Capital Medical University, Beijing 100053, China; Beijing Key Laboratory of Magnetic Resonance Imaging and Brain Informatics, Beijing 100053, China.
- Department of Radiology and Nuclear Medicine, Xuanwu Hospital, Capital Medical University, Beijing 100053, China; Beijing Key Laboratory of Magnetic Resonance Imaging and Brain Informatics, Beijing 100053, China.
- Department of Radiology and Nuclear Medicine, Xuanwu Hospital, Capital Medical University, Beijing 100053, China; Beijing Key Laboratory of Magnetic Resonance Imaging and Brain Informatics, Beijing 100053, China. Electronic address: [email protected].
Abstract
Accurate preoperative lateralization of temporal lobe epilepsy (TLE) remains challenging, particularly in cases with subtle or MRI-negative lesions. This study aimed to overcome limitations of conventional MRI by developing a diffusion spectrum imaging (DSI)-based machine learning approach for noninvasive TLE lateralization. We retrospectively analyzed DSI scans from 49 unilateral TLE patients (29 left, 20 right) and 25 healthy controls (HC). Local connectome fingerprints and quantitative anisotropy (QA) features were extracted. A support vector machine (SVM) was trained to classify patients from controls and to identify the epileptogenic hemisphere. Model performance was evaluated using 10-fold stratified cross-validation, with feature selection and dimensionality reduction performed within each training fold. The DSI-based SVM achieved high accuracy in distinguishing TLE from HC. With fingerprint features, accuracy was 97.3% (sensitivity 0.959, specificity 1.000); QA features yielded the same accuracy 97.3% (sensitivity 0.980, specificity 0.960). For lateralization among patients, the fingerprint model reached 100% accuracy versus 91.8% for QA. In the three-class classification task (left TLE, right TLE and HC), the models achieved accuracies of 78.4% (fingerprint) and 73.0% (QA). The fingerprint-based classifier yielded F1-scores of 0.943 for HC, 0.727 for LTLE, and 0.650 for RTLE; QA achieved F1-scores of 0.875, 0.677, and 0.632, respectively. DeLong's test found no significant AUC differences. DSI-derived metrics combined with machine learning enable accurate, noninvasive lateralization of TLE. This approach addresses clinical necessities by reliably detecting epileptogenic zones, including cases with subtle structural abnormalities, offering significant potential to enhance presurgical decision-making and patient outcomes.