An Enhanced Double-Unet Segmentation Framework for the Early Detection of Diabetic Retinopathy

A segmentation method for diabetic retinopathy images utilizing an enhanced Double-Unet can assist clinicians in analyzing the length, curvature, and branching angles of blood vessels, among other parameters, to facilitate the early detection of diabetic retinopathy. Thin blood veins have little contrast, which makes it simple to lose spatial information. By adding the attention gates of fusion channel attention and spatial attention in the decoder section, the enhanced Double-Unet network highlights the vascular features and recovers the features lost during the coding stage. Additionally, multi-scale context information may be efficiently extracted, and blood vessel feature information can be enhanced by substituting the Dense Atrous Spatial Pyramid Pooling (Dense ASPP) module for the ASPP module. The proposed method was assessed using retinal vascular datasets (DRIVE, CHASEDB1, STARE, HRF) and fundus images from 40 patients at a leading hospital in Fujian Province of China. The results show that the present method is fast and has high accuracy and can achieve high accuracy, recall, and F1 scores on most of the above datasets with fewer parameters. The segmentation results offer a solid foundation for more vascular reconstruction research while also successfully overcoming the interference of thin vessels and low-contrast textures.