Coformer: Collaborative Transformer for Medical Image Segmentation

Transformer has shown significant power for medical image analysis. However, the inherent design of the Transformer limits the ability to extract local features, thereby potentially affecting the performance. To address the above limitations, a Collaborative Transformer (Coformer) is proposed for medical image segmentation. In detail, the Multiscale Representation Fusion (MRF) module is designed to extract the semantic information of the fused features. During the encoding phase, local and global multi-scale feature representations are extracted by incorporating with Swin Transformer. Then, the semantic features are deeply extracted by the MRF module based on the cross-attention mechanism in the decoding phase. Comparative experiments on the well-known public Synapse Multi-Organ Segmentation dataset have demonstrated that Coformer achieves 82.39% dice score with 1.7%-12.62% improvements over the state-of-the-art methods.