Radiomics and artificial intelligence for coronary calcium scoring: insights from the CAC CV-PREVITAL study

Background Coronary artery calcium (CAC) score from ECG-gated computed tomography (CT) is a predictor of cardiovascular events in primary prevention. Advances in artificial intelligence (AI) have enabled automated CAC scoring. Radiomics enables high-throughput extraction of quantitative image features capturing intricate patterns. Purpose To assess the accuracy of an automated AI-based model for CAC scoring and the value of radiomics for patient stratification. Methods This is a substudy of randomized CAC-CV PREVITAL study. Participants who underwent ECG-gated CT scan with a CAC score > zero were included. Total CAC score was calculated using the Agatston method. Automated calcium segmentation was performed using a validated deep learning model and AI-derived total CAC score as well as mean CAC score were calculated. Radiomic features were extracted per calcification. Feature selection techniques were applied to remove less informative features, and unsupervised clustering of significant radiomics features was performed to identify patient subgroups. The agreement between the AI-derived total CAC score and manual total CAC score was assessed using Pearson correlation coefficient (r), intraclass correlation coefficient (ICC), Bland-Altman plot, and bootstrapped quadratic weighted Cohen’s kappa. Results A total of 259 individuals, mean age 59 (8.3) years, 36% women, were included. Total CAC score was 230 (447) (mean, SD). A CAC score ≥100 was present in 45% of participants, and a CAC score ≥300 was found in 17% of subjects. A total of 259 CT scans, one per individual, were analysed, and 851 radiomic features were extracted per calcification. After feature selection, 335 radiomic features were retained. AI-derived total CAC score and manual total CAC score were highly correlated (r=0.986, p<0.001; ICC = 0.98, 95% CI 0.96-0.99), they showed a strong agreement (kappa = 0.87, 95% CI 0.82–0.91), although Bland-Altman plot indicated the existence of moderate systematic bias (mean bias 40.7, 95% limits of agreement, -122 to 203). Subjects were grouped in two clusters based on radiomics features. The distribution of manual CAC score and AI-derived CAC scores was significantly different between the two clusters (p<0.001) (Figure 1). In particular, a higher prevalence of lower CAC scores was found in cluster 1 and a higher prevalence of higher CAC scores was detected in cluster 2. Nevertheless a total CAC score ≥ 300 was found in a small proportion of subjects in cluster 1 and some subjects in cluster 2 had a CAC score = 1-99 (Figure 1). Conclusions AI-based total CAC score shows a reliable accuracy for risk categorization compared with manual CAC score despite the presence of some systematic bias. Radiomics-based stratification may provide additional value beyond CAC score that deserves further investigation.