Hybrid RA2-Net: Residual Atrous Attention Network for Vessel Classification using Fundus Images

Retinopathy is a visual complication that affects the retinal surface causing vision loss due to high blood pressure, diabetes, etc. Regular analysis of retinal vessels can detect the early signs of these retinal complications and alerts the experts for immediate treatment. As a result, with the advantage of deep learning technique, a robust Hybrid RA
²
-Net (Residual Atrous Attention Network) is implemented for segmenting the two classes of blood vessels from fundus images. The proposed hybrid RA
²
-Net is developed by improving the conventional U-Net by integrating with an RHAC (Residual Hybrid Atrous convolution), Attention channel and RCM (Recurrent Convolution Module). In the encoding path, the RHAC blocks extract the most pertinent features that enhance the visibility of pixels of vessels in the feature map. Further, the extracted features are improved by employing the channel attention and RCL in the decoding path. The efficiency of the proposed Hybrid RA
²
-Net is computed using AV dataset, dualmodal2019 and DRIVE standard dataset. On comparison, the proposed Hybrid RA
²
-Net achieved better results in terms of accuracy, dice coefficient, precision, specificity, and sensitivity than other networks like U-Net, Res-Net, and Attention-Net