Comparing AI-Enabled Opportunistic Detection of Lung Cancer on CAC Scans versus Chest CT: A Pooled MESA and Framingham Analysis

Submitted manuscript | Original research review

Lung cancer is often diagnosed in people who do not meet current screening criteria. This pooled cohort study examines whether an AI system can use coronary artery calcium (CAC) scans, obtained for cardiovascular assessment, to identify lung cancer opportunistically, and compares performance with paired chest CT.

Study design

The analysis included 3,816 participants from MESA Examination 5 and the Framingham Heart Study who had paired CAC and chest CT scans and at least two years of follow-up. The multi-model AI combined whole-chest analysis, nodule-focused assessment, and safety-net checks to classify scans as low, intermediate, or high likelihood. The outcome was lung cancer diagnosed within two years.

Key findings

Fifteen lung cancers were diagnosed during follow-up. The AI high-likelihood group captured 10 of 15 cancers on CAC scans (sensitivity 66.7%; positive predictive value [PPV] 7.5%) and 12 of 15 on chest CT (sensitivity 80.0%; PPV 5.9%). An automated nodule-size threshold of 8 mm yielded sensitivity/PPV of 73.3%/2.8% on CAC and 93.3%/2.3% on chest CT. For both scan types, PPV was higher for the AI high-likelihood group than for the size threshold. Two-year cancer rates in the AI low-likelihood groups were 0.14% on CAC and 0.0% on chest CT.

Interpretation

The results suggest that AI could identify lung cancer risk on CAC scans acquired for another clinical purpose, with higher PPV than the evaluated nodule-size threshold. This retrospective pooled-cohort analysis supports prospective evaluation; it does not establish that opportunistic AI screening improves patient outcomes.

Abstract

Background: Many lung cancers arise outside current screening eligibility criteria. Objective: To assess AI-enabled opportunistic lung cancer detection on CAC and paired chest CT scans versus a Fleischner-anchored nodule-size threshold. Methods: Pooled analysis of MESA and Framingham participants with paired scans; a multi-model AI assigned low-, intermediate-, or high-likelihood categories. The endpoint was lung cancer within two years. Results: Among 3,816 participants, 15 cancers were diagnosed. High-likelihood AI detected 10/15 on CAC (sensitivity 66.7%; PPV 7.5%) and 12/15 on chest CT (sensitivity 80.0%; PPV 5.9%). The automated 8-mm threshold had PPVs of 2.8% and 2.3%, respectively. Conclusions: AI showed higher PPV than the nodule-size threshold on both scan types, warranting prospective evaluation.

Status: Submitted to Radiology: Cardiothoracic Imaging. The full submitted manuscript is not attached.

Download the abstract

Download Abstract (PDF)