AI can score heart-artery calcium on a lung-cancer scan. A study tests how well
Every low-dose CT for lung cancer also images the heart. A 323-scan study found an AI tool matched expert calcium scoring almost exactly — and separate data show why that number carries weight.
| Group | Value (value) |
|---|---|
| Death | 1.51 (1.28 to 1.79) |
| Death, MI or stroke | 1.57 (1.33 to 1.84) |
| Death, MI, stroke or revascularisation | 1.69 (1.45 to 1.98) |
When a person at risk gets a low-dose CT scan to screen for lung cancer, the scan also captures the heart — and the calcium lining the coronary arteries is one of the strongest available markers of cardiovascular risk, usually left unmeasured. A new study tests whether artificial intelligence can read that calcium automatically off the same scan: on 323 lung-screening CTs, a fully automated tool agreed almost perfectly with expert scoring [s1], and separate outcome data explain why the reading is worth having — calcium found incidentally on this kind of scan independently predicts death and heart attack [s2].
This is the logic of "opportunistic" screening: extract a second, unrelated piece of clinical information from an image already taken, at no extra scan, dose or appointment. The question is whether the automated number can be trusted.
What the tool measures
The coronary artery calcium (CAC) score, or Agatston score, quantifies calcified plaque in the heart's arteries. It is normally derived from a dedicated ECG-gated cardiac CT. Lung-screening scans are different — low-dose and not gated to the heartbeat, which makes the arteries harder to measure. The study set out to see whether an AI system could produce a reliable Agatston score on those non-ideal images without a human doing the manual work [s1].
What the study found
In a retrospective single-centre analysis, 323 participants undergoing low-dose CT for lung-cancer screening (55.7% male; median age 61 years) had their calcium quantified three ways: by the AI tool, by a semi-automated method requiring a technologist, and by two radiologists grading visually [s1]. Against the semi-automated reference, the AI's scores showed excellent agreement — an intraclass correlation coefficient of 0.96 and a Spearman correlation of 0.97 — with a small systematic bias of 21.5 and, the authors note, moderate limits of agreement [s1].
For the clinically important task of telling calcium apart from none, the AI reached a sensitivity of 0.97 (95% CI 0.92–0.99) and a specificity of 0.91 (95% CI 0.86–0.94) for excluding CAC [s1]. Sorted into plaque-burden categories, AI and the semi-automated method agreed almost perfectly (weighted kappa 0.92), better than the semi-automated method agreed with the two visual readers (kappa 0.84 and 0.68) [s1]. And it was faster where it counts: the semi-automated scoring took 102.0 ± 95.7 seconds per scan, against 15.5 ± 6.0 and 24.2 ± 7.4 seconds for visual grading, while the AI required no manual scoring at all (p<0.001) [s1].
The honest limits are in the design. This is a single centre, a retrospective sample, and — crucially — a test of agreement, not of truth: the AI is being checked against another software method, not against patient outcomes or a physical phantom. Excellent agreement means the tool reproduces the reference, including whatever the reference gets wrong.
Why the number matters
The reason to automate CAC on lung scans at all is that the score carries prognostic weight. In a separate study, a deep-learning algorithm quantified incidental calcium on routine, non-gated chest CTs from 5,678 adults with no known cardiovascular disease [s2]. Just over half — 52% — had a score above zero, and among those with a score of 100 or more the average 10-year cardiovascular risk was 24%, yet only 26% were taking a statin [s2].
After adjustment for age, sex, blood pressure, lipids, smoking and other factors, a score of 100 or more (against a score of zero) was associated with a higher risk of death, at a hazard ratio of 1.51 (95% CI 1.28–1.79); of the composite of death, myocardial infarction or stroke, at 1.57 (95% CI 1.33–1.84); and of that composite plus revascularisation, at 1.69 (95% CI 1.45–1.98) [s2]. Incidental calcium, in other words, is not incidental to a patient's prognosis — it flags people who are undertreated for a risk they and their doctors did not know they carried.
Why it matters
Automating CAC on scans already being done is one of the more genuinely useful ideas in imaging AI: the information is free, the prognostic evidence is solid, and the main barrier has been the manual labour of scoring. But agreement with a software reference is the start of the case, not the end of it — the recurring lesson across AI imaging, from the FDA's radiology testing gap to the MASAI mammography trial and two 2026 radiology-assist studies, is that a tool that matches a reference has not yet been shown to change what happens to patients. The open question, as with the broader outcomes gap in AI devices, is whether a flagged score actually leads to a statin prescription and a prevented event, or simply to a line in a report. For the scale of the missed risk this could surface, see our coverage of silent atherosclerosis in symptom-free adults.
What to watch
Whether automated CAC reporting is validated prospectively and against outcomes rather than against other software, whether it holds up across scanners and populations, and — the decisive step — whether health systems build the pathway that turns an automated score into treatment. A number generated for free still needs someone to act on it.
This article is informational and is not medical advice.
Sources
- [s1] Ochs K, Ensle F, Happe J, et al. "Agreement and workflow efficiency of AI-based coronary artery calcification quantification in lung cancer screening: Comparison with semi-automated and visual assessment." European Journal of Radiology Open, published online 18 July 2026. https://doi.org/10.1016/j.ejro.2026.100799
- [s2] Peng A, Dudum R, Jain S, et al. "Association of Coronary Artery Calcium Detected by Routine Ungated CT Imaging With Cardiovascular Outcomes." Journal of the American College of Cardiology, 82(12), published online 11 September 2023. https://doi.org/10.1016/j.jacc.2023.06.040
Sources
- Agreement and workflow efficiency of AI-based coronary artery calcification quantification in lung cancer screening: Comparison with semi-automated and visual assessment — European Journal of Radiology Open , July 18, 2026
- Association of Coronary Artery Calcium Detected by Routine Ungated CT Imaging With Cardiovascular Outcomes — Journal of the American College of Cardiology , September 11, 2023
More on
What is a silent heart attack, and how often does one go unnoticed?
Some heart attacks leave scar on an ECG or MRI but were never felt as one. In a long US cohort they made up more than 45% of heart attacks — and they still shortened survival.
Should healthy adults get a screening ECG?
US preventive guidance recommends against a resting or exercise ECG in symptom-free adults at low cardiovascular risk, and finds the evidence insufficient at higher risk. The tracing rarely changes treatment.
Should healthy adults get a coronary calcium scan? Guidelines disagree
It is not a routine test for everyone — the USPSTF finds the evidence insufficient. Cardiology guidelines back it for one group: intermediate-risk adults weighing whether to start a statin.
Routine CT coronary scans after chest pain did not cut later heart attacks
TARGET-CTCA randomised 3,170 emergency patients whose troponin had cleared them of a heart attack. Outpatient CT angiography-guided care matched standard care on later infarction or cardiac death.