Feasibility of AI as first reader in the 4-IN-THE-LUNG-RUN lung cancer screening trial: impact on negative-misclassifications and clinical referral rate.

Summary

In 3,678 baseline LDCTs, AI as a first reader produced far fewer negative misclassifications than radiologists (0.8% vs 11.1%) and reduced missed clinical referrals (2.9% vs 11.8%). These data support AI-first-read filters to safely rule out negatives and decrease radiologist workload in lung cancer screening.

Key Findings

• Among 3,678 baseline LDCTs, AI had 31 negative misclassifications (0.8%) versus 407 (11.1%) by radiologists.
• Missed clinical referrals would have been lower with AI as first reader (3/102, 2.9%) than with radiologists (12/102, 11.8%).
• AI independently ruled out negative cases without substantially increasing risk of missed referrals, supporting feasibility as a first-reader filter.

Clinical Implications

Programs can consider AI first-read filtering at baseline LDCT to rule out negatives, focusing radiologist time on indeterminate/positive studies and potentially standardizing referral decisions.

Limitations

• Single AI vendor and baseline-only assessment; external generalizability and longitudinal outcomes not addressed.
• Operational thresholds and definitions (e.g., NM criteria) may vary across programs.

Future Directions

Prospective implementation studies comparing AI-first-read pathways versus standard practice on detection rates, workload, cost-effectiveness, and patient outcomes across diverse settings.