ARSA-UNet: Atrous residual network based on Structure-Adaptive model for retinal vessel segmentation

Abstract

Retinal vessel segmentation remains a challenging task due to the myriad morphological characteristics of the retinal vessel and the complexity of the retinal background. To address the challenge, a novel ARSA-UNet network is proposed in this paper. Firstly, a structure adaptive layer is proposed as a substitute for the traditional convolutional layer to enhance the convolution performance. It aims to adapt different feature mappings through model structure adjustment and improve the learning efficiency of the network. Then, the proposed atrous residual path is incorporated in the hopping layer to dynamically adjust the receptive field to obtain deeper multi-scale semantic information and richer spatial information. Finally, a feature-screening fused module is added to the decoder to suppress redundant features, achieving a more effective fusion of upper- and lower-layer information. In addition, to mitigate the network degradation caused by the deepening of layers, a multi-scale deep supervised mechanism is introduced so that the deep network can also be adequately trained. Extensive experiments have been conducted on the mainstream datasets DRIVE, CHASE, and STARE, with results suggesting that the Sen value of the proposed method achieves the optimal results by the metrics set out in all three datasets. Meanwhile, F1, Acc and Spe fare well in most of the datasets. Overall, ARSA-UNet outperforms other state-of-the-art methods for the procedure. Source code is available at https://github.com/LaboratoryofCSBC/SADS-UNet.