Learning feature-rich integrated comprehensive context networks for automated fundus retinal vessel analysis

In the computer-aided diagnosis of ophthalmic diseases, the automatic segmentation of retinal vessels is the most basic and critical step. The semantic segmentation networks proposed in recent years have been improving in performance, but suffer from the drawbacks of oversized models and excessive number of parameters. In this work, we focus on designing a lightweight and efficient comprehensive contextual network (CC-Net) to segment retinal vessels more accurately. The CC-Net is a U-shaped encoder-decoder structure that integrates the multi-scale pooling module and the channel fusion module. The multi-scale pooling module is specifically designed to compensate for the loss of spatial information caused by continuous pooling operations. The channel fusion module was constructed in order to allow adaptive recalibration of channel feature responses to highlight the most discriminative feature channels. To objectively validate the proposed method, we performed extensive qualitative and quantitative analyses based on three publicly available fundus datasets, CHASE_DB1, STARE, and DRIVE. The contrast-limited adaptive histogram equalization (CLAHE) method was used to enhance the contrast of the original images in the experiments without using any data augmentation strategy. The results show that the proposed CC-Net has excellent segmentation performance with a model parameter count of only 0.26 M. In addition, we use cross-training to evaluate the robustness of the method and demonstrate its strong generalization ability in the cell membrane segmentation task.