107P - Artificial intelligence-based volumetric classification of pulmonary nodules in Chinese baseline lung cancer screening population (NELCIN-B3)

Background

Developments in artificial intelligence (AI) systems to assist radiologists in reading low-dose computed tomography (LDCT) could improve lung cancer screening efficiency by triaging negative screenings. Our aim is to evaluate the performance of an AI system as an independent reader when used to rule-out negative scans (with no nodules/nodules <30 mm³) in LDCT lung cancer baseline screening in a Chinese population.

Methods

362 individuals from the Netherlands and China Big 3 (NELCIN-B3) study with at least one non-calcified solid-component nodule were included. Volumetric nodule measurement was performed on all scans by two experienced radiologists and a fully automated AI lung cancer screening software (AVIEW LCS, v1.1.39.14, Coreline Soft) independently. The largest non-calcified solid-component nodule was determined for each scan. Discrepancies between two radiologists or the AI were reviewed by a consensus panel and stratified into two groups based on NELSON-plus/EUPS protocol threshold: PM was indeterminate/positive nodule (≥100 mm³) classified by radiologists/AI, which at consensus read was negative (<100 mm³); NM was negative result classified by radiologists/AI (<100 mm³), which at consensus read were indeterminate/positive (≥100 mm³).

Results

When looking at the discrepancies for the largest solid-component nodule per participant; 34 (9.4%; 23 PM, 11 NM) discrepancies were reported for AI, compared to 31 (8.6%; 0 PM, 31 NM), and 27 (7.5%; 8 PM, 19 NM) discrepancies for radiologist 1, and 2, respectively. 13 out of 23 PM findings by AI were due to vessel segmentation, 7 were non-nodular morphology, and 3 were incorrect nodule type classification.

Conclusions

In a Chinese baseline lung cancer screening population, the use of an AI system as an independent reader to triage negative scans resulted in the lowest negative-misclassification result compared to radiologists.

Legal entity responsible for the study

The authors.

Funding

Supported by the Royal Netherlands Academy of Arts and Sciences (grant PSA_SA_BD_01). Y.M. supported by the China Scholarship Council (CSC no. 202008440409).

Disclosure

All authors have declared no conflicts of interest.