LG-DFNet: A Medical Image Segmentation Network Paying More Attention to Boundary Contour Details

Though vision transformer (ViT) succeeds in medical image segmentation, it is not effective in boundary segmentation for lesions or organs with complex contours. To solve this problem, we propose Local-global Dynamic Fusion Network (LG-DFNet). Specifically, we construct Local Static Attention (LSA), which can determine the calculation range of attention weight according to the position of elements in the feature space. LSA enables the network to extract the local boundary texture structure of lesions or organs. In addition, we propose Global Dynamic Attention (GDA), which dynamically calculates the attention weight of a position according to the similarity of input data. GDA enhances the network's generalization ability. Based on LSA and GDA, we design Dynamic Cross Fusion Block (DCFB) for skip connection. By introducing Transpose Attention, we design Dual Attention Fusion Block (DAFB) as the core part of LG-DFNet. Extensive experiments on three public datasets show that LG-DFNet achieves better segmentation performance.