Abstracts

Rationale: Focal cortical dysplasia (FCD) is one of the most common causes of pharmacoresistant focal epilepsy. FCD lesions on MRI are often subtle and may require MRI post-processing to be identified and for their extent to be determined. The overarching goal of this work is to construct a multimodal deep learning framework for identifying and delineating FCD lesions, using MRI data (T1w, T2w and FLAIR), features generated from these data, and other complementary electroclinical information such as EEG/MEG localization. In the current study, we build an initial deep-learning framework using 3D T1w data from a large retrospective cohort from our center, with histopathology confirmation of type II FCD.


Methods: A total of 96 patients who had pathologically confirmed FCD II (24 with type IIA and 72 with type IIB), a 3D whole-brain preoperative T1w MRI, and at least 1 year of post-surgical follow-up were included in this study. MRI scans were acquired using 1.5T or 3T Siemens scanners. Most T1w acquisitions had 0.8×0.8×1 mm voxel resolution and 256 slices. All volumes were warped to MNI 152 space and skull stripped. Lesion labels were drawn for each patient based on expert radiology review of the available preoperative MRIs with MRIcron. 77 patients were set aside for training while the remaining 19 were used for testing. Version 2 of the nnU-Net (Isensee et al., Nat. Methods 2021) supervised deep learning model was used due to the fundamental similarities between segmentation and lesion identification tasks. All convolutional layers were 3D, implying that convolution was performed on multiple slices during the application of each convolution kernel. 5-fold cross-validation was employed during training; the training task was allowed to proceed for 1000 epochs. Optimal hyperparameters for testing were selected based on the optimum pseudo-Dice score for each epoch. Figure 1 shows the network structure.

Results: Cross-validation results yielded a mean AUC of 0.796 ± 0.042 (Figure 2A). The optimal sensitivity for the mean cross-validation curve was 0.664, and the optimal specificity was 0.838. Evaluating the trained model on the test set surpassed the cross-validation results, producing an AUC of 0.931, an optimal sensitivity of 0.845, and an optimal specificity of 0.962 (Figure 2B). Visual verification of these results (Figure 2C) showed that, when the lesion was identified, the predicted lesions conformed well to the provided lesion labels.

Conclusions: Our model showed strong efficacy in identifying and delineating the extent of FCD II lesions even though it has only been trained on standard clinical T1w images alone. The high specificity of the model was worth noting. Work is currently ongoing to improve sensitivity by adding FLAIR and T2w images, including features maps, and incorporating EEG/MEG localization data, to formulate a multimodal deep-learning framework.

Funding: None.