AI performance for nodule volume doubling time in the follow-up of the UKLS lung cancer screening study compared to expert consensus and histological validation

Aim:

To validate an artificial intelligence (AI) software for automated assessment of nodule growth by volume doubling time measurement (VDT) on protocol-mandated follow-up low-dose CT (LDCT) scans from the UK lung cancer screening (UKLS) trial. 

Methods:

This validation study included 710 UKLS participants with 939 LDCT follow-up scans (361 3-month and 578 12-month). Follow-up scans were assessed independently by both AI and human readers. A positive finding warranting referral was defined as the largest nodule with a solid component ≥ 100 mm³ showing VDT ≤ 400 days at follow-up. Performance was benchmarked against the European expert panel (reference standard) and then the histological outcomes (gold standard). 

Results:

Against the expert panel, AI achieved the lowest 3-month negative misclassification (NM) rate (1/11, 9.1 %), versus human readers (range: 18.2–63.6 %). AI’s positive misclassification (PM) rate was initially 7.8 % (28/361) at 3 months but decreased to 0.9 % (5/578) at 12 months. Against histological outcomes of 9 screen-detected lung cancers, AI identified VDT ≤ 400 days in all 4 cancers also deemed positive by the expert panel at the earliest 3-month follow-up, while human readers missed or delayed referrals in 1–3 of these. AI also identified VDT ≤ 400 days in 3 of 5 cancers that the panel classified as negative, primarily due to their sub-threshold volume (<100mm³). 

Conclusions:

The automated AI system showed strong VDT assessment performance in follow-up screening, outperforming human readers in the early identification of rapid growth in histologically-confirmed cancers, thus supporting its potential to enhance risk stratification and facilitate earlier lung cancer detection.