Fundus image segmentation method based on improved DeepLabv3+ network

Accurate optic disc (OD) and optic cup (OC) segmentation in fundus images is crucial for the automatic diagnosis and treatment of certain diseases. However, due to blurry boundaries and vessel occlusion in retinal fundus images, segmenting OD and OC remains a challenging task. This paper proposes an improved DeepLabv3+ network, which contains a novel sparse atrous spatial pyramid pooling (SparseASPP) module and an improved decoder architecture. Specifically, SparseASPP can enhance the ability of DeepLabv3+ to integrate multi-scale feature information, while the improved decoder architecture can effectively fuse low-resolution deep feature information to enhance the network's segmentation performance. The proposed method achieves significant improvements in segmentation accuracy on the RIGA+ dataset (including two training subsets and three test subsets) and the REFUGE dataset. These results demonstrate that our method outperforms DeepLabVv+ and other comparison methods.