GLFF: A Global-Local Feature Fusion Model to Segment Epileptic Focus of Focal Cortical Dysplasia from Multi-Channel MR Images

Focal cortical dysplasia is a primary cause of drug-resistant epilepsy that necessitates resection of the epileptic focus through neurosurgery for effective treatment. However, a significant challenge lies in the accurate localization of the epileptic focus from MR images pre-neurosurgery. Structural changes may be subtle or entirely absent, thereby complicating the process of epileptic focus localization for clinic doctors or radiologists. This process is heavily reliant on experience, and it demands substantial time and labor. Additionally, variations exist between intra- and inter-observer detection results of the epileptic focus. The convolutional neural network has been explored for segmenting the epileptic focus from MR images. However, the limited receptive field size of these convolutional models makes it challenging to discriminate the epileptic focus, indicating that the performance of existing methods needs further improvement. In this paper, we propose a Global-Local Feature Fusion (GLFF) model that jointly learns global semantic features and local detailed features, utilizing a transformer encoder and a convolutional encoder from symmetric sub-volumes. These features are consolidated in the decoder to predict the epileptic focus. We executed experiments to compare the proposed method with existing ones. The proposed method achieves a Dice score of 0.408± 0.277 and a detection sensitivity of 82.4%, superior to three other models. As such, the proposed method could potentially be a valuable tool to aid doctors in quickly and accurately detecting the epileptic focus.