HRM-UNet: Image Segmentation of Retinal Blood Vessels on the Basis of Improved VM-UNet

The abundance of pathologic changes provided by fundus retinal images is significant in guiding treatment of ophthalmic diseases in clinical medicine. The complexity of the structures and pathologic features of retinal blood vessels leads to many limitations in the accuracy of segmentation, so efficient and correct segmentation of retinal vascular images has become an urgent need in clinical practice. We propose a segmentation model called HRM-UNet, which uses the VM-UNet model as the basis, replaces the original down-sampling module in the encoder with a Harr wavelet-based down-sampling module, and adds residual paths in the skip connections. The model provides better performance in deeper neural networks by adding wavelet down-sampling and residual path modules that attribute to reduce the spatial resolution of feature map while retaining as much information as possible. We conducted experiments on DRIVE dataset and the results show that HRM-UNet model outperforms other mainstream network models in the retinal vessel segmentation task and improves the accuracy of retinal vessel segmentation compared to other models.