Authors: (including presenting author): :
Ma KC (1), Yee KS (1)(2), Tseng CZ (1)(2)
Affiliation: :
(1) Department of TB & Chest, Wong Tai Sin Hospital (2) Department of Medicine and Geriatrics, Kwong Wah Hospital
Keyword 1: :
AI lung cancer detection
Keyword 3: :
Service improvement
Keyword 5: :
Deep Learning
Introduction: :
Lung cancer remains a leading cause of cancer-related mortality worldwide. Earlier detection through imaging can reduce preventable deaths[1], but implementing efficient, scalable early-detection service models is challenging because CT interpretation and downstream workup require substantial specialist capacity.Sybil is a deep learning model originally trained using low-dose CT data from the U.S[4]. National Lung Screening Trial (NLST) to predict future lung cancer risk from CT scans. While Sybil has shown promising performance in its development setting, its generalizability to Chinese populations and real-world clinical settings in Hong Kong has not been established. Local validation is essential before considering deployment to support risk stratification, clinical decision-making, or pathway redesign for earlier detection.
Objectives: :
To evaluate the real-world discriminatory performance of the Sybil deep learning model for lung cancer risk prediction in a Hong Kong Chinese population undergoing CT thorax imaging, and to describe key clinical differences between cancer and non-cancer groups in this cohort.
Methodology: :
We performed a retrospective validation study including 458 individuals aged 43–82 years who underwent CT thorax scans between 2022 and 2024 in hospitals under the Hong Kong Hospital Authority. The cohort comprised 256 participants with histologically confirmed lung cancer and 202 individuals without evidence of lung cancer (non-cancer controls). Demographic and clinical characteristics were compared between groups, including smoking status and radiologic nodule size. Sybil-generated risk predictions were obtained from CT scans and model discrimination was assessed using the area under the receiver operating characteristic curve (AUC). Group comparisons were evaluated using standard statistical testing, with significance defined as p< 0.05.
Result & Outcome: :
Results: The lung cancer group had a higher prevalence of smoking compared with non-cancer controls (86% vs 24%). The mean nodule size was larger among participants with lung cancer than among controls (2.1 cm vs 1.1 cm; p< 0.05). Sybil demonstrated strong discriminatory performance for lung cancer in this Hong Kong Chinese population, achieving an AUC of 0.87. These findings suggest good separation of cancer versus non-cancer cases based on CT-derived model predictions in routine clinical imaging data. Conclusion: In a real-world cohort of patients undergoing CT thorax imaging within the Hong Kong public hospital system, the Sybil deep learning model showed good performance in distinguishing lung cancer from non-cancer individuals (AUC 0.87). This local validation supports the potential feasibility of using Sybil to enhance early detection pathways—such as triaging higher-risk scans for expedited review, informing follow-up intensity, or supporting service models that improve efficiency while maintaining clinical safety.[5] Further work should include calibration assessment, prospective evaluation in screening-relevant populations (e.g., asymptomatic high-risk individuals undergoing low-dose CT), and impact analyses examining workflow integration, downstream investigations, and patient outcomes. 1. The National Lung Screening Trial Research Team, Reduced Lung-Cancer Mortality with Low-Dose Computed Tomographic Screening. New England Journal of Medicine, 2011
2. Black, W.C., et al., Cost-Effectiveness of CT Screening in the National Lung Screening Trial. New England Journal of Medicine, 2014
3. Pinsky, P.F., Assessing the benefits and harms of low-dose computed tomography screening for lung cancer. Lung Cancer Manag, 2014
4. Mikhael, P.G., et al., Sybil: A Validated Deep Learning Model to Predict Future Lung Cancer Risk From a Single Low-Dose Chest Computed Tomography. Journal of Clinical Oncology, 2023
5. Pinsky, P.F., et al., National Lung Screening Trial Findings by Age: Medicare-eligible versus under-65 Population. Ann Intern Med, 2014
