NAUNet: lightweight retinal vessel segmentation network with nested connections and efficient attention

The state of retinal vessels in fundus images is a reliable biomarker for many diseases, and the accurate segmentation of retinal vessels is important for the diagnosis of related diseases. To address the problem of many layers and high complexity of deep learningbased vascular segmentation network, this paper proposes a lightweight encoderdecoder network NAUNet by reasonably reducing the number of network layers. By introducing the DropBlock regularization strategy, the local semantic information can be discarded more effectively to motivate the network to learn more robust and effective features. Efficient attention module uses appropriate crosschannel interaction to capture richer global information. In the skip connection part, the nested connection strategy is adopted to effectively fuse the feature maps gathered from the intermediate decoder and the original feature maps from the encoder, which makes up for the semantic gap caused by direct simple connection. In addition, data augmentation is performed on the original image to improve the robustness and prevent the overfitting problem caused by insufficient data. A mixed loss function is proposed to solve the problem of class imbalance in vascular images. Finally, NAUNet was tested and achieved F1 scores of 80.92%/81.25%/74.86% and AUC values of 0.9831/0.9849/0.9841 on the DRIVE, STARE and CHASE_DB1 datasets, respectively.The number of parameters for the proposed method was only 2.66 M.