DEC-UNet: A dual-encoder feature fusion and CARAFE upsampling UNet for medical image segmentation

Aiming at the problem of low segmentation accuracy caused by complex lesion morphology and irregular lesion boundaries in medical images, a segmentation model for medical images based on DEC-UNet is proposed in the research. A dual-encoder architecture based on C-FEM and RCW-Transformer is introduced to simultaneously extract local and global features, where C-FEM is used to capture local features such as textures and boundaries, RCW-Transformer is used to capture global features such as overall morphology and spatial location. To sufficiently fuse the features extracted by the dual encoders, DCA is introduced to compute cross-attention weights between local and global features, learning their similarities and differences and preserving discriminative information from each encoder. The fused features are transmitted to the corresponding decoder levels through the skip connections. In the decoder, CARAFE is adopted for adaptive upsampling to reduce the loss of edge features, thereby improving the segmentation accuracy. Experiments are conducted on the Kvasir-SEG, ACDC, and Synapse datasets respectively. Compared with CNN-based, Transformer-based, and CNN-Transformer-based models, DEC-UNet achieves the highest DSC of 89.38%, 91.24%, and 93.83% respectively, and the lowest HD of 9.43 mm, 8.13 mm, and 6.64 mm respectively. The results demonstrate that DEC-UNet improves the accuracy of medical image segmentation.