GGAC: Multi-relational image gated GCN with attention convolutional binary neural tree for identifying disease with chest X-rays

Abstract

Using medical images for disease identification is an important application in the medical field. Graph Convolutional Network (GCN) is proposed to model multi-relational image and generate more informative image representations. Recently, the relations between medical images are utilized to identify diseases. This paper proposes a Gated GCN with Attention Convolutional Binary Neural Tree (GGAC) for Multi-Relational Image Identifying Disease. GGAC extracts the discriminative features of the image, strengthen the ability to model medical images, understands images representation deeply and then well captures the multi-modal relation between images. Firstly, an Attention Convolutional Binary Neural Tree based on the attention mechanism is designed to extract the fine-grained features of the images, and use the attention conversion operation on the edge of the tree structure to enhance the network’s acquisition of key image features. Secondly, a Gated GCN is proposed to improve GCN performance by solving the problem of the weight distribution of different neighbors in the same-order neighborhood. Thirdly, a GCN propagation rule is used to transfer messages in multi-relational Graph and then solves the message passing problem of high-dimensional feature data in GCN. Finally, we verify GGAC on a multi-relational graph constructed on the Chest X-rays14. It can be seen from the experiment that overfitting and underfitting can be solved to a certain extent through the extraction and inference of the features of the multi-relational graph, and then GGAC has better performance than the state-of-the-art methods, and keeps good in model complexity.