DCF-Net: A Dual-Coding Fusion Network based on CNN and Transformer for Biomedical Image Segmentation

Convolutional neural networks (CNNs) have become the mainstream technology in the field of biomedical image segmentation. However, due to the inherent locality of convolutional operations, CNNs tend to ignore the relationships between remote pixels. In contrast, Transformer performs well in capturing global feature information but weakly in extracting fine-grained local feature information. To solve this, we propose a novel dual-coding fusion network that can effectively combine the respective advantages of the Transformer and CNN. DCF-Net adopts an encoder-decoder structure. Firstly, we propose two multi-scale feature representations based on the CNN encoder and the Swin Transformer encoder, respectively. Secondly, we reconstruct the self-attention mechanism in the Transformer branch by shifting window adaptive complementary attention, overcoming the shortcoming of the traditional self-attention mechanisms that only focus on spatial attention while ignoring channel, spatial, and channel cross-dimensional attention. Finally, in order to achieve a fine fusion of local and global features obtained from the multi-scale feature representations, we design a feature interaction fusion module in the skipping connection part. We conducted experiments on biomedical electron microscopy images, achieving a Dice coefficient score of 93.89% and a pixel accuracy of 99.01% . Our results demonstrate that the proposed DCF-Net outperforms state-of-the-art approaches.