Fusion network based on the dual attention mechanism and atrous spatial pyramid pooling for automatic segmentation in retinal vessel images.

Abstract

Accurate segmentation of retinal blood vessels from retinal images is crucial to aid in the detection and diagnosis of many eye diseases. In this paper, a fusion network based on the dual attention mechanism and atrous spatial pyramid pooling (DAANet) is proposed for vessel segmentation. First, we propose a dual attention module consisting of a position attention module and a channel attention module, which aims to adaptively recalibrate features to extract effective features. And full-scale skip connections are used in the encoder to provide multi-scale feature maps for the dual attention modules. Then, atrous spatial pyramid pooling (ASPP) allows the network to capture features at multiple scales and combine high-level semantic information with low-level features through the encoder-decoder architecture. We qualitatively evaluate the model using five metrics: sensitivity, specificity, accuracy, AUC, and F1 score on DRIVE, CHASED_B1, and STARE datasets. The DAANet outperforms the work of 10 state-of-the-art predecessors in these three datasets. Furthermore, we apply the trained model to clinical retinal images. The model obtains gratifying accurate and detailed segmentation results, which demonstrates a promising application prospect in medical practices.