Three-branch feature enhancement and fusion network for focal cortical dysplasia lesions segmentation using multimodal imaging

Objective

Conventional multimodal imaging, including MRI and fluorodeoxyglucose positron emission tomography (FDG-PET), has difficulty in accurately detecting subtle or blurred focal cortical dysplasia (FCD) lesions. Morphometric maps assist localization by highlighting abnormal regions, whereas wavelet-filtered images emphasize texture and edge details. Therefore, we propose a three-branch feature enhancement and fusion network (TBFEF-Net) that integrates conventional multimodal imaging, morphometric maps, and wavelet-filtered images to enhance the accuracy of FCD localization.

Methods

The proposed TBFEF-Net comprises a semantic segmentation backbone, a cross-branch feature enhancement (CFE) module, and a multi-feature fusion (MFF) module. In the semantic segmentation backbone, three UNet-based branches separately extract semantic features from conventional multimodal imaging, morphometric maps, and wavelet-filtered images. In the encoding stage, the CFE incorporates a residual-based convolutional block attention module (CBAM) to aggregate features from all branches, enhancing the feature representation of FCD lesions. While in the decoding stage, the MFF integrates edge detail features from the wavelet-filtered imaging branch into the conventional multimodal imaging branch, enhancing the ability to capture lesion edges. As a result, this approach enables more precise segmentation.

Results

Experimental results show that TBFEF-Net surpasses several state-of-the-art methods in FCD segmentation. In the primary cohort, the Dice and sensitivity reached 59.73 % and 67.13 %, respectively, while in the open cohort, the Dice and sensitivity were 54.67 % and 54.81 %, respectively.

Significance

We introduced wavelet-filtered images for the first time in FCD segmentation, offering a novel approach and perspective for FCD lesions localization.