Fully-convolutional deep-learning based system for coronary calcium score prediction from non-contrast chest CT

The amount of calcium deposits in the coronary arteries is an important biomarker of cardiovascular disease. Coronary calcium has traditionally been quantified as an Agatston score using ECG-synchronized cardiac CT. Coronary calcium is rarely quantified from general chest CT scans, of which nearly 10 Million are performed in the US annually. We present an automatic method based on fully-convolutional deep neural network to segment coronary calcium and predict Agatston score from any non-contrast chest CT. We experimented with an internal dataset acquired through partnership with a large health organization in Israel. The dataset is composed of 1054 Chest CTs and reflects a variety of originating institutions, acquisition devices and manufacturers. In comparison to expert manual annotations, our algorithm achieved a Pearson correlation coefficient of 0.98. Bland-Altman analysis demonstrated a bias of 0.4 with 95% limits of agreement of [-189.9-190.7]). Our linearly weighted Kappa results are 0.89 for Agatston risk category assignment. We also applied our method on a very large (14,365 subjects) cohort from the National Lung Screening Trial (NLST). We demonstrate correlation of the algorithm predictions with cardiovascular-related clinical outcomes.