Abstract:
The health of retinal vessels in fundus images is of great significance for the early diagnosis of various ophthalmic diseases and diabetic cardiovascular diseases, etc. However, the retinal blood vessels are delicate, distributed irregularly and the boundary is ambiguous. Therefore, it is difficult to accurately segment them. Based on the characteristics of retinal blood vessels, we propose a U-shaped network—rough channel attention residual U-Net (RCAR-UNet), which combines rough neurons and channel attention mechanism. Firstly, the network introduces the concept of upper and lower approximation in rough set theory to design rough neurons. Secondly, the rough channel attention module is constructed based on rough neurons, and the module uses global max pooling and global average pooling in U-Net skip connections to construct upper and lower approximation neurons, and performs weighted summation between neurons to reasonably rough the established channel dependencies, which not only contain global information but also have local characteristics, and can effectively achieve accurate rescaling of the extracted retinal vessels features. Then adding residual connections to transfer features directly from the lower to the higher layers, to help solve the network performance degradation problem and effectively extract richer retinal vascular features. Finally, in order to verify the effectiveness of the proposed RCAR-UNet model, comparison experiments are performed on three public fundus image datasets with traditional network models such as U-Net, Attention U-Net, etc. The results show that the RCAR-UNet model has high superiority in the accuracy, sensitivity and similarity of blood vessel segmentation.