Multi-relational Graph Convolutional Neural Networks for Carotid Artery Stenosis Diagnosis via Fundus Images

The accumulation of carotid plaque leads to carotid artery stenosis, which in turn increases the risk of cerebrovascular disease. Noninvasive diagnosis of carotid stenosis using fundus images offers a promising approach. However, the challenge lies in extracting relevant features from these images, as convolutional neural networks(CNNs) or Transformers, which focus solely on individual images, fail to consider the interdependencies between them, leading to limited diagnostic accuracy. To address this issue, we propose a novel and effective network by combining CNNs and multi-relational graph convolutional neural networks(MGCNs). Firstly, we feed the input images into four distinct branches, which consist of CNNs or Transformers, with each branch associated with a particular relation. This process generates unique feature vectors for each branch. Secondly, we construct a multi-graph for the four kinds of clinical data, such as gender, age, sex and pid, to obtain four adjacency matrices. Finally, the feature vectors and the corresponding four adjacency matrices are input into the graph convolutional network layer respectively to obtain the prediction features, and then the prediction results are obtained through the fully connected layer. Experiments are carried out on a private dataset and the results demonstrate that the accuracy of the proposed algorithm is 10%-20% higher than that of the comparison model. Our code is available at https://github.com/momoyrz/Carotid-stenosis.