MES-Net: a new network for retinal image segmentation

Abstract

Glaucoma, diabetic retinopathy, and other eye diseases have seriously threatened people’s visual health. Whether it is clinical diagnosis or computer-aided diagnosis, the accurate segmentation of the retinal tissues and lesion areas (optic disc (OD), optic cup (OC), retinal blood vessels) are indispensable. In this paper, we propose a novel network MES-Net for the retinal image segmentation tasks. MES-Net adopts U-Net as the overall architecture, including three proposed modules: multi-scale feature pre-extraction (MFP) block, encoder spatial cascading encoding (ESCE) path, and decoder input SE block. We use MFP block to extract multi-scale semantic features from input images with different resolutions, and take ESCE path to extract deep features and improve the feature reuse rate. SE block can weaken the semantic gaps between the encoder paths and the decoder paths. For retinal vessel segmentation, the accuracy values of MES-Net tested on the DRIVE, STARE and CHASE datasets are 0.9667, 0.9724, and 0.9697, and AUC is 0.9853, 0.9897, and 0.9869, respectively. For OC and OD simultaneous segmentation, the overlap coefficients of OC and OD tested on the ORIGA dataset are 0.818 and 0.948, and the accuracy are 0.935 and 0.977, respectively. The experimental results show that the proposed method significantly improves the performance of the original U-Net and is superior to other state-of-the-art methods. Thus, it can be applied to retinal image segmentation tasks, and many other medical image segmentation problems can also benefit from the proposed method.