IARC 60th Anniversary - 19-21 May 2026
Session : Integrating AI and data science across the cancer screening continuum
Integrating deep learning of low-dose computed tomography with clinical data for lung cancer risk prediction
PHELLAN ARO R. 1, LAM S. 2,3, WARKENTIN M. 1,4, LIU G. 5,6, DIERGAARDE B. 7,8, WILSON D. 8,9, YUAN J. 8,10, AL-SAWAIHEY H. 11, MURISON K. 1,5, KHODAYARI MOEZ E. 1, BRHANE Y. 1, MEZA R. 3, MYERS R. 2, HUNG R. 1,5
1 Prosserman Centre for Population Health Research, Lunenfeld-Tanenbaum Research Institute, Sinai Health System , Toronto, Canada; 2 Department of Respiratory Medicine, University of British Columbia, Vancouver, Canada; 3 Department of Integrative Oncology, British Columbia Cancer Research Institute, Vancouver, Canada; 4 Department of Oncology, Cumming School of Medicine, University of Calgary, Calgary, Canada; 5 Dalla Lana School of Public Health, University of Toronto, Toronto, Canada; 6 Department of Medical Oncology and Hematology, Princess Margaret Cancer Centre, Toronto, Canada; 7 Department of Human Genetics, University of Pittsburgh School of Public Health, Pittsburgh, United States; 8 Cancer Epidemiology and Prevention Program, UPMC Hillman Cancer Center, Pittsburgh, United States; 9 Department of Medicine, University of Pittsburgh, Pittsburgh, United States; 10 Department of Epidemiology, University of Pittsburgh School of Public Health, Pittsburgh, United States; 11 Royal Cornwall Hospitals NHS Trust, Truro, United Kingdom
Background: Lung cancer is the leading cause of cancer deaths globally. Low-dose computed tomography (LDCT) screening reduces lung cancer mortality. Within screening programs, efficient patient management requires accurate lung cancer risk assessment. Segmentation-free deep learning (DL) models such as Sybil can improve screening efficiency but require extensive validation and possible improvement. Objectives: Based on 4 LDCT screening cohorts, we investigated whether integration of deep learning based on LDCT scans and clinical data can improve lung cancer risk prediction. Methods: A total of 52,482 LDCT scans of 22,469 participants from 4 LDCT screening programs with a median follow-up period of 7 years were included in the analysis: National Lung Screening Trial (NLST) for model development, and external validation using the Pan Canadian Early Detection of Lung Cancer Study (PanCan), Toronto site of the International Early Lung Cancer Action Plan (IELCAP-Toronto), and the Pittsburgh Lung Screening Study (PLuSS). Sybil is a deep learning model that accepts a single, whole LDCT scan as input and extracts features from it using a pretrained 3D Resnet-18 encoder. Key clinical and epidemiological factors were evaluated for their added predictive value beyond deep learning model. The area under the receiver operating characteristic curve (AUC) was estimated for lung cancer risk within 1 to 6 years, stratified by pulmonary nodule presence and size. Results: Sybil’s AUC ranged from 0.93 in year 1 and reduced to 0.79 in year 6 in the independent validation cohorts. The predictive performance was suboptimal in the absence of documented nodules (AUC=0.64), and for scans with small nodules (AUC=0.61) in year 6. Our new model, Sybil-Epi, trained with baseline scans, achieved higher predictive performance (AUC=0.83, 95% CI 0.81 to 0.85) compared to Sybil (AUC=0.80, 95% CI 0.78 to 0.82) in year 6. The difference is most notable when nodules are absent, with Sybil-Epi AUC of 0.76 (95% CI 0.70 to 0.82) and Sybil AUC of 0.64 (95% CI 0.57 to 0.70). Model interpretability was assessed based on attention maps and SHapley Additive exPlanations framework (SHAP). Conclusions and Implications: Sybil performs well for short-term lung cancer risk, but the predictive accuracy was suboptimal when nodules were absent on LDCT scans. Our integrated Sybil-Epi model with deep learning and clinical-epidemiological factors significantly improved model predictive performance. It offers the possibility to optimize screening intervals and to maximize the screening efficiency.