WCLC 2025 - Posters & ePosters-P1.04.06 AI Performance for Lung Nodule Growth in UK Lung Cancer Screening: Expert Consensus and Histological Validation

Pdf Summary

This study evaluates the performance of artificial intelligence (AI) in assessing lung nodule growth via Volume Doubling Time (VDT) during follow-up low-dose CT (LDCT) scans in the UK Lung Cancer Screening (UKLS) trial. While LDCT screening reduces lung cancer mortality, it creates a substantial workload due to numerous indeterminate nodules and the resource-intensive, variable nature of VDT assessment by human readers. AI has been validated for baseline screening but lacked validation for follow-up VDT assessment, particularly against histological outcomes.<br /><br />The retrospective analysis included 710 participants with 939 follow-up LDCT scans at 3 and 12 months. VDT assessments were independently conducted by a fully automated AI system and three human readers. Positive cases had nodules with solid components ≥100 mm³ and VDT ≤400 days. Validation involved comparison to a European expert panel (reference standard) and confirmed lung cancer diagnoses (gold standard).<br /><br />Results showed AI achieved the lowest negative misclassification rate at 3 months (9.1%) compared to human readers (18.2–63.6%). AI’s positive misclassification rate improved from 7.8% at 3 months to 0.9% at 12 months. In nine histologically confirmed lung cancers, AI detected rapid growth (VDT ≤400 days) in all four expert-identified positive cancers at earliest follow-up, outperforming humans who missed or delayed referrals in up to three cases. AI also identified rapid growth in three of five cancers that expert panel classified as negative due to protocol criteria, indicating AI's ability to detect risk beyond volume thresholds.<br /><br />The study concludes that AI enables earlier cancer detection by consistently identifying rapid growth ahead of clinical referral, acts as a safety net for small cancers beyond volume-based rules, and improves consistency and reliability of VDT assessments by reducing inter-reader variability. This suggests AI can standardize and enhance follow-up lung nodule evaluation in screening programs, potentially improving outcomes and workload management.

Asset Subtitle

Michael Davies

Meta Tag

Speaker Michael Davies

Topic Screening and Early Detection

Keywords

Artificial Intelligence

Lung Nodule Growth

Volume Doubling Time

Low-Dose CT

UK Lung Cancer Screening

VDT Assessment

Histological Validation

Cancer Detection

Inter-reader Variability

Screening Workload Management