Nils Hampe was nominated for the Robert F. Wagner All Conference Best Student Paper Award at SPIE Medical Imaging 2021 for his work “Graph Attention Networks for Segment Labeling in Coronary Artery Trees”.

Coronary artery disease (CAD) is the most common cause of death worldwide. Clinical reporting of CAD includes characterization of plaque burden and grade of stenosis, as well as plaque location. The location is typically reported in form of the anatomical label of the segment the lesion resides in. To automate the reporting, we have developed a method for automatic labelling of the coronary artery segments.

Anatomical labeling of segments in the coronary artery tree is challenging due to substantial variability in coronary anatomies. In this work, we have used graph convolutional neural networks (GCNs), that fall under the umbrella term geometric deep learning, which is currently a very vibrant field of research. Using GCNs enabled end-to-end learning of representations of local segment neighborhoods directly from the coronary artery tree graphs.

Results demonstrated that our GCN-based approach achieved performance comparable to previous approaches using graph matching. Unlike earlier works our method does not require any expert knowledge in the form of hand-crafted rules. Moreover, our approach compares favorably to other works using machine learning, despite the comparably small dataset used in this work. Our approach may reduce manual workload and may be used in large-scale studies to improve risk stratification for diagnosis of CAD.