Abstract:
Multimodal medical images have advantages such as comprehensive information, improving diagnostic accuracy, detecting therapeutic effects, and prognostic evaluation, which can provide doctors with more comprehensive and accurate information. This paper proposes a multimodal brain tumor image segmentation network that combines edge feature extraction to utilize the comprehensive information between multimodal medical images. The main design idea of the network is as follows: Firstly, the pre activation residual block has been improved, and a dual residual pre activation module has been proposed to replace the encoding and decoding layers in U-Net. Additionally, the network width has been increased in the decoding layer, enhancing the network's feature extraction ability; Then, an edge feature extraction module was designed to extract edge features from different modalities and levels, and integrate them with the segmentation results, enhancing the network's ability to express edge lesions; Finally, using the BraTS2018 and BraTS2022 datasets to validate network performance, the similarity coefficients Dice in the entire tumor area, tumor core area, and enhanced tumor area reached 89.80%, 90.40%, 82.60%, and 91.73%, 93.06%, and 87.00%, respectively. The proposed network has high accuracy and relatively low redundancy for lesion areas with unclear boundaries. Overall performance of the proposed network is better than most existing networks.