In East Asia, nearly 50% of lung cancer patients are never-smokers, underscoring a disease pattern often driven by other factors such as genetic or familial predisposition rather than tobacco exposure.¹ However, efforts to implement effective screening strategies for never-smokers have been challenged by limitations in imaging resources, radiologist availability, and the high cost of surveillance.¹ At the 2025 ASCO Annual Meeting, Dr. Li, Siu-Ching Molly from The Chinese University of Hong Kong presented interim findings from the LC-SHIELD study (Lung Cancer Screening in HIgh-Risk Non-SmokErs with Artificial InteLligence Device), which evaluates the feasibility and utility of artificial intelligence (AI)-driven lung nodule detection in the high-risk non-smoker population.¹

LC-SHIELD is a prospective, observational study assessing LungSIGHT, an AI-based detection system developed for identifying pulmonary nodules ≥5mm.¹ The AI platform employs a multi-task learning framework trained on publicly available datasets, including the public benchmark LIDC-IDRI dataset, to detect and classify nodules in low-dose computed tomography (LDCT) images.¹ Eligible participants included adults aged 50-75 years who had never smoked (defined as lifetime exposure to <100 cigarettes) and met the following criteria: having at least one first-degree relative with lung cancer, no personal history of malignancy, and no prior tuberculosis or chronic lung disease.¹ Subjects with chest computed tomography (CT) imaging or pulmonary interventions within the past two years were excluded.¹ Between July and December 2024, 405 participants were enrolled and underwent LDCT screening.¹ Those with AI-positive findings were scheduled for follow-up at 3, 6, 12, and 24 months depending on risk, while AI-negative cases were planned for a repeat assessment at 24 months.¹ All scans were retrospectively reviewed by radiologists blinded to the AI results to validate the AI findings and assess concordance in nodule detection.¹

The AI algorithm (LungSIGHT) classified 49% (n=200) as positive (presence of lung nodule ≥5mm) and 51% (n=205) as negative.¹ Among the 200 AI-positive cases, 61% (n=122) were also confirmed positive by radiologists (true positives), while 39% (n=78) were deemed false positives.¹ Lung cancer was histologically confirmed in 3 participants (0.7%), including two stage 1 and one stage 3 tumors.¹ All three cases were EGFR mutation positive.¹ A key aspect of the study was iterative fine-tuning of the AI algorithm using local imaging data from the first 181 subjects enrolled in the study.¹ After fine-tuning, sensitivity was relatively stable (from 88% to 81%), while specificity improved markedly (from 52% to 85%) along with concordance (from 67% to 83%).¹ In the validation cohort (n=224), the final model achieved a balanced performance: sensitivity of 73%, specificity of 77%, and concordance of 76%.¹

In conclusion, interim findings from the LC-SHIELD study suggest that AI-assisted first-reader of screening LDCT in high-risk never smokers is feasible.¹ After fine-tuning with local calibration data, LungSIGHT demonstrated high sensitivity and specificity.¹ The use of AI may help reduce costs and workload while improving efficiency, particularly in areas with high screening demand.¹