SF-U-Net: Using Accurate Shape Estimation and Feature Restoration to Improve Retinal Vessel Segmentation

Abstract

The features of retinal blood vessels are very essential indicators playing an important part in the process of judging and diagnosing the eye diseases for doctors. Sometimes, these features can also be the indicators for the examination of hypertension, coronary heart disease and diabetes. However, retinal blood vessels are often very small and complex in distribution, which brings toughness to the doctors when doing the operations of the segmentation of retinal blood vessels. Although the deep learning manners represented by U-Net has performed very well in the field of the segmentation of the images of the retinal blood vessel in recent years, the above-mentioned inconvenience still cannot be effectively settled. For the purpose of improving the correct rate of the segmentation and settling the above-mentioned inconvenience we propose a network called SF-U-Net, which uses accurate shape estimation and feature restoration to achieve the improvement of the accuracy. We follow the structure of Fully Convolutional Networks (FCN) and Skip Connection of U-Net and use deformable convolution to accurately capture the shape of blood vessels when extracting features at the coding layer to overcome the problem of complex blood vessel distribution. At the decoding layer, we adopt a novel dual-stream up-sampling method to achieve accurate feature restoration. Experimental results show that our SF-U-Net is capable of improving the segmentation results of retinal blood vessels conspicuously. In the experiment, we use both fundus image datasets called DRIVE and CHASE-DB1 and the experimental results of multiple indicators on them surpass other deep-learning methods significantly. The experimental results of the SF-U-Net model on a variety of indicators on DRIVE dataset exceed the experimental performances of the currently most advanced methods. The mean accuracy is 0.9602 the area under the curve (AUC) is 0.9848 and the sensitivity is 0.8567.