Abstract: The shapes of arteries and veins are highly similar in RGB retinal imaging,and their inherent structures are both subtle and complex,making it difficult for most retinal image processing models to achieve ideal results.To improve the accuracy of arteriovenous segmentation and reduce training costs,a retinal segmentation model based on segment anything model is proposed—SAM-Retina.SAM-Retina adopts a feature fusion,adaptive image encoder,and mask decoder architecture,using structure-and-function dual-modal retinal images simultaneously containing RGB as well as 570 nm and 610 nm single wavelength images instead of a single mode(RGB) image as input.The features of these three images are fused through a feature fusion.While retaining the pre-trained parameters of the image encoder on a large-scale natural image dataset,the model's feature extraction capability on retinal medical images is enhanced by inserting an adapter module and updating it within the vision transformer(ViT) block.The static prompt embedding instead of prompt encoder is adopted to remove the input and encoding process of prompts in the original SAM segmentation process.During the experimental phase,the model is trained and evaluated on the DualModal2019 and HRF datasets,and compared with U-Net,CRU-Net,and TW-GAN.The experimental results show that SAM-Retina is more advanced than other models in various evaluation indicators and the employment of dual-modal image also improves segmentation perfor-mance without increasing the model size.

Key words: Dual-modal retinal image,Arteriovenous segmentation,Image encoder,Vision transformer,Static prompt embedding