CSAUNet: A cascade self-attention u-shaped network for precise fundus vessel segmentation

Abstract

Vessel segmentation contributes to precise diagnosis of fundus diseases. However, due to the existence of micro vessels and class imbalance, it remains a challenging task to accurately segment vessels from fundus images. To address this issue, a coarse-to-fine framework, namely cascade self-attention U-shaped network (CSAUNet), is proposed in this paper. Firstly, a self-attention U-shaped network (SAUNet) is pre-trained to roughly locate the fundus vessels. Afterwards, the pretrained SAUNet is cascaded with a residual self-attention UNet (Res-SAUNet), and the rough segmentation results of the pretrained SAUNet are regarded as the spatial attention maps to force the Res-SAUNet to focus on significant regions. In addition, self-attention modules (SAMs) are introduced in both the SAUNet and Res-SAUNet to effectively model the long-range dependencies across fundus vessels. The experimental results show that the accuracy, sensitivity, specificity and dice similarity coefficient (DSC) of CSAUNet on three public databases, including DRIVE, STARE and synthetic images (SI), reach 96.76%/83.4%/98.1%/82.07%, 97.28%/83.04%/98.62%/84% and 0.9915/0.9539/0.9962/0.9612, respectively. In addition, the coarse-to-fine framework can improve the DSC and sensitivity on DRIVE, STARE and SI by 1.24%/2.37%, 1.69%/1.04% and 0.8%/0.72%, respectively, and the incorporation of SAMs can further improve the DSC on DRIVE, STARE and SI by 0.37%, 0.7% and 0.49%, respectively, which indicates that CSAUNet can improve the segmentation performance for fundus vessels.