Abstracts

Rationale: Focal cortical dysplasia (FCD) type II is a major pathology in patients undergoing surgical resection to treat drug-resistant epilepsy. MRI post-processing methods have been proposed to improve detection of FCD lesions. In this study, we utilized novel surface-based MRI morphometry and machine learning for automated lesion detection in a cohort of patients with histologically confirmed type II FCD. To test the robustness of the methods, we chose a mixed cohort from different epilepsy centers using different MRI scanners. We also performed ROC analysis to obtain the optimal threshold for automated lesion detection. Methods: Sixty-one patients with drug-resistant epilepsy and histologically proven FCD type II were included in the study. The patients were evaluated at three different epilepsy centers using 3 different MRI scanners. The normal control database was constructed using scans from 120 normal controls, which were also obtained from different scanners including the 3 used for patients. T1-volumetric sequence was used for the post-processing following methods delineated by Adler &Wagstyl et al. (NeuroImage Clinical, 2017). Cortical surface reconstructions were generated using FreeSurfer software. As shown in Figure 1, features such as cortical thickness, gray-white matter blurring, cortical shape abnormality, sulcal depth and curvature were calculated, compared left-to right within the same patient, and normalized by the control database. These features are then incorporated into a nonlinear neural network that can be trained to identify lesional vertices. In this study, we are optimizing the value by which the output probability map from the classifier is thresholded. After thresholding, maps are clustered into neighbor-connected clusters and the top cluster is considered the putative lesion location. For all the patients included in this study, manual lesion masks were mapped to the surfaces; vertices were classified as being either lesional or non-lesional. The neural network classifier was trained on surface-based features and classification based on manual lesion masks; then performance was evaluated using leave-one-out cross-validation. Success of detection was defined by overlap between the final cluster and the manual labeling. All 61 included patients were evaluated to quantify the sensitivity of the classifier, and we also assessed specificity in 35 additional healthy volunteers. Results: The threshold of 0.9 showed optimal sensitivity of 72% and specificity of 91% (Table 1). The area under the curve for the ROC analysis was 0.75 which suggests a highly discriminative classifier. Using this threshold, subgroup analysis revealed that scans from three different centers or different scanners were not significantly different from each other. Although a lower sensitivity in children was observed as compared to adults, statistical significance was not reached. Conclusions: Automated surface-based MRI morphometry showed robust performance across cohorts from different centers and scanners. The currently proposed method may be a valuable tool to improve FCD detection in presurgical evaluation for patients with drug-resistant epilepsy. Funding: none