Fig. 3: Confusion matrices to compare automatic and manual CAC quantification and to assess test–retest repeatability.
From: Deep convolutional neural networks to predict cardiovascular risk from computed tomography
a Comparison of CAC classes calculated by the deep-learning framework and expert readers, combining data from FHS-CT2, NLST, PROMISE, and ROMICAT-II (n = 5521). The robustness of b the deep learning framework and c expert readers to quantify CAC was assessed in 252 FHS-CT1 subjects who underwent two subsequent CT scans within 1 h (Scan 1 and Scan 2). CAC coronary artery calcium, FHS-CT117, FHS-CT217 Framingham Heart Study, (CT1) participants from the seventh examination cycle of the offspring cohort or first examination cycle of the Third Generation Cohort (2002–05) and (CT2) participants from the second examination cycle of the Third Generation Cohort (2008–11), NLST18 National Lung Screening Trial, PROMISE19 Prospective Multicenter Imaging Study for Evaluation of Chest Pain, ROMICAT-II20 Rule Out Myocardial Infarction using Computer Assisted Tomography II, CAC risk groups: very low: 0; low: 1–100; moderate: 101–300; high: >30021.
