• From the Image Sciences Institute (S.G.M.v.V., N.L., M.A.V., I.I.), Departments of Radiology (B.K.V., T.L., P.A.d.J., W.B.V.), Experimental Cardiology (I.E.M.B.), and Radiotherapy (D.H.J.G.v.d.B.), and Imaging Division (H.M.V.), University Medical Center Utrecht, Heidelberglaan 100, 3584 CX Utrecht, the Netherlands; Department of Radiology and Nuclear Medicine, Radboud University Medical Center, Nijmegen, the Netherlands (N.L.); Departments of Biomedical Engineering and Physics (S.G.M.v.V., I.I.) and Radiology and Nuclear Medicine (I.I.), and Amsterdam Cardiovascular Sciences (I.I.), Amsterdam University Medical Center, University of Amsterdam, the Netherlands; Department of Radiology and Nuclear Medicine, Radboud University Medical Center, Nijmegen, the Netherlands (N.L.); Department of Cardiology, Meander Medical Center, Amersfoort, the Netherlands (I.E.M.B.); Department of Medicine, University of Mississippi Medical Center, Jackson, Miss (A.C.); and Department of Radiology and Radiological Sciences, Vanderbilt University Medical Center, Nashville, Tenn (J.G.T., J.J.C.).

Background Although several deep learning (DL) calcium scoring methods have achieved excellent performance for specific CT protocols, their performance in a range of CT examination types is unknown. Purpose To evaluate the performance of a DL method for automatic calcium scoring across a wide range of CT examination types and to investigate whether the method can adapt to different types of CT examinations when representative images are added to the existing training data set. Materials and Methods The study included 7240 participants who underwent various types of nonenhanced CT examinations that included the heart: coronary artery calcium (CAC) scoring CT, diagnostic CT of the chest, PET attenuation correction CT, radiation therapy treatment planning CT, CAC screening CT, and low-dose CT of the chest. CAC and thoracic aorta calcification (TAC) were quantified using a convolutional neural network trained with (a) 1181 low-dose chest CT examinations (baseline), (b) a small set of examinations of the respective type supplemented to the baseline (data specific), and (c) a combination of examinations of all available types (combined). Supplemental training sets contained 199-568 CT images depending on the calcium burden of each population. The DL algorithm performance was evaluated with intraclass correlation coefficients (ICCs) between DL and manual (Agatston) CAC and (volume) TAC scoring and with linearly weighted κ values for cardiovascular risk categories (Agatston score; cardiovascular disease risk categories: 0, 1-10, 11-100, 101-400, >400). Results At baseline, the DL algorithm yielded ICCs of 0.79-0.97 for CAC and 0.66-0.98 for TAC across the range of different types of CT examinations. ICCs improved to 0.84-0.99 (CAC) and 0.92-0.99 (TAC) for CT protocol-specific training and to 0.85-0.99 (CAC) and 0.96-0.99 (TAC) for combined training. For assignment of cardiovascular disease risk category, the κ value for all test CT scans was 0.90 (95% confidence interval [CI]: 0.89, 0.91) for the baseline training. It increased to 0.92 (95% CI: 0.91, 0.93) for both data-specific and combined training. Conclusion A deep learning calcium scoring algorithm for quantification of coronary and thoracic calcium was robust, despite substantial differences in CT protocol and variations in subject population. Augmenting the algorithm training with CT protocol-specific images further improved algorithm performance. © RSNA, 2020 See also the editorial by Vannier in this issue.