Coronary calcium is detectable opportunistically on thoracic CT: pooled analyses show 52% prevalence, and AI models can automate scoring and predict cardiovascular death risk

Opportunistic identification of CAC on non-ECG-gated thoracic CT shows good diagnostic agreement and can detect previously unknown coronary artery disease in asymptomatic patients. Prior CAC scoring relied mainly on dedicated ECG-gated non-contrast cardiac CT; this review systematically assembles evidence for identifying CAC on routine thoracic CT, low-dose lung cancer screening CT, and PET/SPECT attenuation correction CT, extending applicability from dedicated cardiac CT to nearly any CT covering the heart. A meta-analysis of 15 studies showed a pooled correlation coefficient of 0.96 (95% CI 0.92–0.98, P<.001) between CAC assessed on non-gated CT and ECG-gated CT; a meta-analysis of 94 studies including 89,006 patients reported a pooled CAC prevalence of 52% (95% CI 46%–58%) with moderate heterogeneity between studies.

Incidental CAC on thoracic CT is an independent predictor of adverse cardiovascular outcomes, and reporting CAC can change clinical management behavior. The review elevates CAC from a purely imaging finding to an opportunistic biomarker that can trigger preventive treatment decisions, and assembles randomized and observational evidence linking CAC reporting to subsequent statin prescriptions and lifestyle change. A meta-analysis of 51,582 patients followed for 51.6 ± 27.4 months showed a pooled relative risk of 2.13 (95% CI 1.57–2.90, P=.004) for all-cause mortality and 2.61 (95% CI 2.17–3.74, P<.001) for MACE in patients with CAC; a meta-analysis of 5 studies in lung cancer screening populations showed a pooled relative risk of 3.27 (95% CI 1.88–5.68, P=.004) for MACE and all-cause mortality; the NOTIFY-1 randomized study showed 51.2% versus 6.9% (P<.001) receiving statin therapy at 6 months in the notification arm.

AI/machine learning models can automatically identify and quantify CAC across a range of thoracic CT protocols with prognostic discrimination, and can reveal gaps in preventive care. Prior CAC quantification depended on dedicated software and manual reading; this review assembles deep learning models trained on gated and non-gated CT, notes that some are commercially available and included on the FDA AI-enabled Medical Devices List, and proposes an end-to-end pipeline from identification to clinical action. A deep learning model trained by Zeleznik et al on Framingham Heart Study, NLST, PROMISE, and ROMICAT-II data showed that patients with a calcium score >300 Agatston units were nearly 4 times more likely to experience cardiovascular death over 6.7 years compared with those with no CAC (HR 3.87, 95% CI 2.45–6.11, P<.001); a model by Hagopian et al using US Veterans Affairs data showed 10-year all-cause mortality HR 3.49 (P<.0005) for CAC >400 Agatston units, identified CAC in 79% of 8,052 lung cancer screening patients, and found 30.6% of high-score patients were not taking lipid-lowering therapies.

Photon-counting CT, virtual non-contrast reconstructions, radiomics, and CAC as a multi-system biomarker offer new technical directions for CAC detection and risk stratification. The review discusses these still-early-stage technologies within the same framework as the clinical evidence for CAC, noting they may improve quantification accuracy and extend the informational value of CAC beyond cardiovascular disease. Early studies show that at 3 mm slice thickness photon-counting CT yields lower CAC scores than standard energy-integrating detector CT, and that CAC scores from virtual non-contrast images are slightly lower than those from true non-contrast images; in MESA, radiomic features of the whole heart predicted severe coronary stenosis better than cardiovascular risk factors alone; MESA and ARIC cohorts show associations between CAC and cancer, chronic kidney disease, COPD, osteoporosis, and dementia, though the authors explicitly state these associations do not demonstrate causation.

This article is aimed at clinicians, radiologists, and researchers interested in cardiovascular prevention, thoracic imaging, and opportunistic screening. Its conclusions apply to identifying and reporting CAC using visual ordinal scores or AI/machine learning tools in asymptomatic or oncology patients undergoing thoracic CT for non-cardiac indications (pulmonary disease assessment, oncological staging, lung cancer screening, PET/SPECT attenuation correction), and using this to assess modifiable cardiovascular risk factors. The authors note that routine CAC reporting can be done today using simple, rapid visual ordinal scores, while AI/machine learning quantification, radiomic analysis, and photon-counting CT are future directions that may further optimize risk stratification.

The authors explicitly state that no randomized trials assessing the clinical impact of opportunistic CAC screening on hard cardiovascular outcomes have yet been published, with cardiovascular endpoint results from NOTIFY-2 (PICTURE, NCT05588895) and ROBINSCA still awaited. Differences in scanner models, vendors, reconstruction algorithms, slice thickness, and acquisition parameters on non-gated thoracic CT can lead to under- or overestimation of CAC, and iodinated contrast may mask calcification, so the absence of CAC on thoracic CT does not mean the absence of coronary artery disease. Associations between CAC and cancer, chronic kidney disease, dementia, and other conditions may be due to shared risk factors or comorbidities rather than causation, and how this information should be used in clinical practice remains uncertain. AI/machine learning in clinical practice also faces issues including bias, generalizability, clinical validation, interoperability, data privacy, cost-effectiveness, post-market surveillance, and regulation; radiomics and virtual non-contrast reconstructions on photon-counting CT are not currently recommended for clinical use. In addition, reporting rates for CAC on thoracic CT vary widely between centers (18%–93%), and a survey of British radiologists found that only 11% routinely reported calcification, suggesting that standardized practice still needs to be advanced.