Comparing AI with Lung-RADS (Version 1.0) for Lung Cancer Detection in Low-Dose CT Surveillance

This study compared AI-based lung cancer likelihood stratification with retrospectively reconstructed Lung-RADS version 1.0 in two baseline low-dose CT screening cohorts: the U.S. National Lung Screening Trial (US-NLST) and the Pamplona International Early Lung Cancer Action Program (P-IELCAP). The primary outcome was histologically confirmed lung cancer within one year.

Study population and approach

The analysis included 2,342 US-NLST examinations with 40 one-year cancers and 3,709 P-IELCAP examinations with 27 one-year cancers. The AI system assigned scans to low-, intermediate-, high-, and highest-likelihood categories. Lung-RADS categories were reconstructed from radiologist-documented nodule information. The supplied manuscript is marked Draft.

Main findings

In US-NLST, the AI high/highest categories included 33 of 40 cancers (82.5%) among 246 scans, with a positive predictive value (PPV) of 13.4%. Lung-RADS category 3 or higher included 30 of 40 cancers among 376 scans (PPV 8.0%). In P-IELCAP, AI high/highest categories included 20 of 27 cancers (74.1%) among 207 scans (PPV 9.7%), compared with 20 of 27 among 390 scans for Lung-RADS category 3 or higher (PPV 5.1%). Both PPV comparisons were reported as statistically significant (p<0.001).

The low-likelihood AI category contained one cancer in each cohort, with observed cancer rates of 0.09% in US-NLST and 0.04% in P-IELCAP. For comparison, Lung-RADS categories 1–2 contained 10 of 40 and 7 of 27 cancers, respectively. AI high/highest categories also identified a portion of cancers retrospectively assigned Lung-RADS 1–2.

Interpretation

The findings suggest that AI stratification may concentrate more one-year cancers in a smaller high-likelihood group and classify fewer cancers as low likelihood than Lung-RADS in these cohorts. The authors note that additional studies are needed to establish clinical utility; the results do not by themselves show that AI-guided screening improves patient outcomes.

Article details

Comparing AI with Retrospectively Reconstructed Lung-RADS for Lung Cancer Detection in Low-Dose CT. Original Research manuscript submitted to Radiology: Cardiothoracic Imaging. The provided PDF is marked Draft; author names and publication details are not included on the available cover page.