A novel deep-learning approach for coronary artery calcium scoring in contrast-enhanced spectral coronary CT angiography: A phantom study

Abstract

Background To evaluate whether deep-learning virtual non-contrast (VNC) images derived from dual-layer spectral coronary CT angiography (CCTA) can replace a dedicated coronary artery calcium (CAC) scoring CT scan. Methods A thoracic phantom containing five cylindrical calcifications (75 to 800 mg/cm3) was scanned on a spectral dual-layer CT with a dedicated non-contrast CAC CT protocol and non-contrast coronary CT (CCT) and contrast-enhanced scans of the CCTA protocol. For CAC CT and non-contrast CCT, the insert was water-filled; for contrast-enhanced CCTA, it was iodine-filled at 400 Hounsfield units at 120 kVp. CCTA images were reconstructed with thick (3.0 mm) and thin (0.67 mm) slices. A deep learning model was trained to detect calcium in spectral CCTA data and used to reconstruct VNC images. Agatston scores were computed per calcification; non-overlapping 95 %-confidence intervals with CAC CT defined statistical significance. Results Using thick slices, Agatston scores from VNC demonstrated no significant differences compared to CAC CT and CCT, except for significant underestimation of VNC at 100 mg/cm3 (4; 95 % CI: −4 to 12) compared to CAC CT (20; 95 % CI: 16 to 25). Using thick slices, Agatston scores from VNC did not significantly differ from CAC CT for any calcification. For both slice thicknesses, non-calcified regions were correctly quantified with an Agatston score of 0. Conclusion This phantom study demonstrates the feasibility of CAC quantification in spectral CCTA using deep learning. After clinical optimization and validation, VNC could potentially eliminate the need for separate CAC CT scans, reducing radiation exposure and clinical workload.