Prediction using graph neural networks of coronary artery diameters

In the diagnosis, prevention, and treatment of ischemic heart disease, it is important to predict the diameters of coronary arteries in order to detect stenosis, assess hemodynamic risk, and optimize revascularization. Arterial networks within organs are fractal systems consisting of bifurcations. Predicting the diameters of daughter vessels based on the known diameter of the parent segment is of great practical interest. Known prediction models are empirical. This paper proposes the use of graph neural networks to predict the diameters of daughter vessels of bifurcations. Purpose – to develop graph neural networks for predicting the diameters of coronary arteries of the heart. Graph neural networks were used for the first time to predict the diameters of coronary arteries. A graph convolution network, a graph attention network, a GraphSAGE network, and a graph neural network based on a transformer were implemented and studied. The quality of the implemented networks, as well as a multilayer perceptron and two well-known hydrodynamic models, was evaluated on a test data set, showing the advantage of graph neural networks over known models, especially when predicting smaller diameters of bifurcation vessels. The best results were demonstrated by a graph neural network based on a transformer with determination coefficients of 0.96 when predicting the larger diameter of the daughter vessel and 0.73 for the smaller vessel diameter. The developed neural networks can serve as the basis for prediction systems for clinical practice.