SEFormer for medical image segmentation with integrated global and local features

Abstract This paper proposes a novel medical image segmentation method, SEFormer, which effectively leverages both local and global features representations to enhance segmentation accuracy and efficiency. To address the limitations of existing image segmentation methods based on transformers and CNNs cannot in simultaneously capturing both local and global features, we propose a novel hybrid network architecture that combines SENet, CNNs( Specifically ResNet), and Transformer. In this design, SENet is employed to enhance the global feature representation capabilities of CNNs, while the Transformer component compensates for the limitations of CNNs in capturing local and channel-specific features. In addition, to prevent feature loss during feature extraction, we draw inspiration from the concept of image pyramid models, obtain a larger receptive field, and use SE fusion to fuse local and global features in each layer of the feature pyramid, ensuring a more comprehensive and complete representation of image information. In the image segmentation task on the CHASEDB dataset, our method achieves great performance, improving the segmentation accuracy by 3.25% compared to existing methods.