Feasibility and limitations of deep learning-based coronary calcium scoring in PET-CT: a comparison with coronary calcium score CT

Abstract

Objective This study aimed to determine the feasibility and limitations of deep learning–based coronary calcium scoring using positron emission tomography‑computed tomography (PET‑CT) in comparison with coronary calcium scoring using ECG‑gated non‑contrast‑enhanced cardiac computed tomography (CaCT).

Materials and methods A total of 215 individuals who underwent both CaCT and PET‑CT were enrolled in this retrospective study. The Agatston method was used to calculate the coronary artery calcium scores (CACS) from CaCT, PET‑CT(reader), and PET‑CT(AI) to analyse the effect of using different modalities and AI‑based software on CACS measurement. The total CACS and CACS classified according to the CAC‑DRS guidelines were compared between the three sets of CACS. The differences, correlation coefficients, intraclass coefficients (ICC), and concordance rates were analysed. Statistical significance was set at p < 0.05.

Results The correlation coefficient of the total CACS from CaCT and PET‑CT(reader) was 0.837, PET‑CT(reader) and PET‑CT(AI) was 0.894, and CaCT and PET‑CT(AI) was 0.768. The ICC of CACS from CaCT and PET‑CT(reader) was 0.911, PET‑CT(reader) and PET‑CT(AI) was 0.958, and CaCT and PET‑CT(AI) was 0.842. The concordance rate between CaCT and PET‑CT(AI) was 73.8%, with a false‑negative rate of 37.3% and a false‑positive rate of 4.4%. Age and male sex were associated with an increased misclassification rate.

Conclusions Artificial intelligence–assisted CACS measurements in PET‑CT showed comparable results to CACS in coronary calcium CT. However, the relatively high false‑negative results and tendency to underestimate should be of concern.