ARDC-UNet retinal vessel segmentation with adaptive residual deformable convolutional based U-Net

To extract maximum features ResAttNet (RAN) network structure is chosen as an alternative to the convolutional layer and it enhances image feature extraction. Additionally, a Deformable Convolution (DC) network was included to provide a feature extraction module, improving the model's capacity to simulate vessel deformation. Apart from the two additional networks because of inadequate quality in retinal data, before model building pre-processing is done. The data is processed by CLAHE, normalization, grayscale transformation, and gamma transformation. Second, the fundamental network structure model U-net is constructed, and the ResAttNet (RAN) structure and DC network are combined to form the ARDC-UNet network. Experimental data, both quantitative and qualitative, demonstrate the efficiency and accuracy with which our ARDC-UNet can segment retinal vessels.