The authors propose a novel convolution neural network architecture to distinguish between different degeneration of retinal layers and their underlying causes. The proposed architecture includes a new residual unit subsuming atrous separable convolution, a novel building block to prevent gradient degradation in the residual unit where the middle (3×3) convolution operations are replaced with two different operations running in parallel: a (2×2) atrous convolution in the left branch and a (2×2) atrous separable convolution in the right branch. The proposed novel architecture is trained from scratch and outperforms other classification models while addressing the issue of gradient explosion, approaching near perfect accuracy of 99.8% and 100% for the Kermany and Srinivasan datasets, respectively.