OCTA Retinal Image Segmentation Method Based on Improved SegNet

At present, the segmentation effect of OCTA retinal image is not ideal, and there exists such problems as discontinuous segmentation of larger vessels, continuous segmentation of discontinuous capillaries and so on. Therefore, this paper proposes a model DAS-Net for vessel segmentation of OCTA retinal image. DAS-Net uses the basic structure of SegNet model, adds deformable convolution and CBAM module, and optimizes the loss function of the model. Experiments on datasets ROSE, OCTA-500_3M and OCTA-500_6M have shown that DAS-Net achieves better segmentation effect. Compared with U-Net, SegNet and CS-Net, DAS-Net gets the highest score in accuracy, dice coefficient and specificity, and it can solve the above problems to some extent.