Overview

• The deep learning algorithm, trained on National Lung Screening Trial data, was externally tested on baseline CTs from Danish, Italian, and Dutch–Belgian screening studies.
• Across the pooled cohort, AUCs were 0.98 for one-year cancers, 0.96 for two years, and 0.94 over the full screening period versus PanCan at 0.98, 0.94, and 0.93.
• For indeterminate nodules sized 5–15 mm, the tool outperformed PanCan with AUCs of 0.95, 0.94, and 0.90 compared with 0.91, 0.88, and 0.86.
• At 100% sensitivity for cancers diagnosed within one year, the model labeled 68.1% of benign nodules as low risk versus 47.4% with PanCan, a 39.4% relative reduction in false positives.
• The external cohort included 4,146 participants with 7,614 benign and 180 malignant nodules, and the study disclosed funding from the Dutch Cancer Society and Siemens Healthineers.