Multi-stage cascaded refinement with wavelet downsampling for retinal vessel segmentation

Abstract

The morphological changes of retinal vessels play a crucial role in assisting doctors with the diagnosis of ocular and cardiovascular diseases. Retinal vessels exhibit complex and variable shapes, and current vessel segmentation methods are ineffective at capturing the features of small vessels of varying sizes. This results in difficulties in segmenting small vessels and causes vessel segmentation to suffer from discontinuities. To address this challenge, we propose a multi-stage cascaded refinement with wavelet downsampling for retinal vessel segmentation. In our network, we introduce a multi-stage cascaded structure, which first employs multi-scale feature fusion in the early stages to extract vessel shape representations of different sizes, thereby enhancing the model’s ability to capture small vessel features. To further refine the feature representation, we embed a feature refinement module at the bottom of the network, utilizing a self-attention mechanism to capture the long-range distribution continuity of the vessels. This mechanism also helps to reduce information redundancy in densely distributed vessels. Additionally, we employ wavelet downsampling as the downsampling layer in the encoder, which effectively minimizes the loss of vessel detail information during the downsampling process.Experimental results on the public datasets DRIVE, CHASE_DB1, and STARE show that the proposed method achieves AUC scores of 0.9891, 0.9910, and 0.9928, and accuracy scores of 0.9707, 0.9764, and 0.9783, respectively. These results demonstrate the superiority of our method in retinal vessel segmentation, which can significantly assist in the early diagnosis and monitoring of ocular and cardiovascular diseases.