Coarse-Level Perception and Fine-Level Refinement Guided Fully Convolutional Network for Retinal Vessel Segmentation

Abstract

The observation of blood vessels in human eyes is crucial for diagnosing the ophthalmological diseases. Due to the uneven illumination and low contrast of retinal images, existing retinal vessel segmentation techniques tend to miss fine vessels and the edges of faint vessels, which is very unfavourable for screening and diagnosing various diseases. To mitigate this problem, this paper proposes a coarse-level perception and fine-level refinement guided fully convolutional network for robust retinal vessel segmentation, which progressively integrate the complementary fine-level and coarse-level information of multilevel pyramid features. We develop a global semantic awareness (GSA) module and a local detail elaboration (LDE) module to help further improve the localization of retinal vessels and the refinement of small vessel branches, respectively. Extensive experiments on DRIVE dataset verify the competitive performance of the proposed model in comparison with seven state-of-the-art methods from the perspective of eight evaluation metrics.