Coordinate Attention Residual Deformable U-Net for Vessel Segmentation

The location information of features is essential for pixel-level segmentation tasks such as retinal vessel segmentation. In this study, we proposed the CARDU-Net (Coordinate Attention Gate Residual Deformable U-Net) model based on coordinate attention mechanism for the segmentation task, which can extract effective features by accurately locating feature location information and enhance the accuracy of segmentation. The deformable convolution and residual structure with Dropblock are also introduced to refine the encoder structure of U-Net. The model is applied to DRIVE, CHASE DB1, and LUNA (2017) datasets, and the experimental results on the three public datasets demonstrate the superior segmentation capability of CARDU-Net, and the modified part is reflected by ablation experiments in this work. The results show that the CARDU-Net model performs better compared to other network models and can segment medical images accurately.