A Self-supervised CNN-GCN hybrid network based on latent graph representation for retinal disease diagnosis

In this paper, we propose a self-supervised CNN-GCN hybrid network (SCGHNet) based on latent graph representation by combining a convolutional neural network (CNN) and a graph convolutional network (GCN), and the Self-Supervised Learning (SSL) is achieved by collaborative tasks. We propose three components for the latent graph representation, including self-attention feature extraction module (SAFEM), latent instance graph representation (LIGR) and latent adaptable softmax (LASoftmax). The proposed SAFEM is used to extract the local and global features from retinal images to satisfy the requirement of the multiscale observation of pathologies. The proposed LIGR is used to explore the relationships among retinal images and promote the concentration of similar samples. The proposed LASoftmax is used to improve the optimization solution. Benefiting from the LASoftmax with LIGR, the latent graph representation can be used to improve the collaborative tasks, enlarge the distance between positvie and negative samples, and obtain more accurate retinal disease diagnosis. In the experiments, our model achieves the AUC of 79.5% and the accuracy of 88.5% for the iChanllenge-AMD dataset. The proposed model outperforms other state-of-the-art SSL models in retinal disease diagnosis of age-related macular degeneration (AMD) and pathological myopia (PM), and shows generalization with outstanding diagnosis performance.