Coronary Ostia Localization Using Residual U-Net with Heatmap Matching and 3D DSNT.

Localization of coronary ostia landmarks in Computed Tomography Angiography (CTA) volumes is a crucial step in developing various automatic diagnostic procedures. In this study, we propose a one-step method of coronary ostia landmark localization that utilizes a residual U-Net with heatmap matching and 3D Differentiable Spatial to Numerical Transform (DSNT). We evaluate the method using two datasets: a Coronary Computed Tomography Angiography (CCTA) dataset containing 201 scans and a publicly available ImageTBAD dataset containing 77 CTA scans annotated with coronary ostia landmarks. On the CCTA dataset we report median Euclidean distance error - 1.14mm on the left coronary ostium and 0.98mm on the right coronary ostium. On the ImageTBAD CTA dataset we report median Euclidean distance error - 3.48mm on the left coronary ostium and 2.97mm on the right coronary ostium. Our evaluation shows that the proposed method improves accuracy of coronary ostia landmark localization when compared to other known methods.