ME-GraphSAGE: Minority Class Feature Enhanced GraphSAGE for Automatic Labeling of Coronary Arteries

Automatic labeling of coronary artery segments improves efficiency in the diagnosis and treatment of coronary artery disease, but faces challenges due to the class imbalance between main and side branches. State-of-the-art methods primarily focus on position-direction and pixel features, which leads to suboptimal performance when dealing with bifurcated segments. In this paper, we propose a minority class feature enhanced GraphSAGE (ME-GraphSAGE), which alleviates class imbalance by generating minority class nodes. We extract bifurcation features from Digital Subtraction Angiography (DSA) images taken from four commonly observed views of coronary arteries. These features, along with other relevant ones, are fed into ME-GraphSAGE to enhance the accuracy of segment labeling in bifurcated regions. By combining the results from the four views, a higher-level sixteen-segment-based coronary labeling is obtained. Our method is evaluated on a dataset of 205 coronary DSA sequences. The experimental results show that ME-GraphSAGE significantly outperforms state-of-the-art methods in labeling coronary artery branches.