M3U-CDVAE: Lightweight retinal vessel segmentation and refinement network

Retinal vessels have high curvature and diverse morphology, making them difficult to segment, especially tiny vessels. At present, the retinal vessels are mainly annotated manually by experts, which is difficult to meet the vast clinical needs. To solve the above problems, we propose an effective network M3U-CDVAE. It adopts the architecture of a segmentation-refinement network to denoise and optimizes segmentation results. Firstly, we design a lightweight segmentation network M3U with an encoder-decoder structure. Then, the Hierarchical Feature Fusion (HFF) unit combines the intermediate features generated by the segmentation network with the pre-segmentation results and connects them to the corresponding layer in the next sub-model. Finally, Convolutional Denoising Variational Auto-Encoder (CDVAE) is used as the refinement network to remove the background noise and optimize segmentation results. We conduct exhaustive ablation experiments to demonstrate the improvement brought by our contribution. At the same time, we carry out comparison experiments on DRIVE, STARE, and HRF datasets to illustrate the effectiveness of the proposed method. Experimental results exhibit that the proposed method is superior to most state-of-art methods.