Deep Machine Learning Predicts Cancer Risk in Follow-Up Lung Screening

Background: Lung cancer screening protocols for the next scheduled computed tomography (CT) or early recall CT are based on growth or intensity change of lung nodules in the baseline scan or size of new incident nodules. We aimed at developing an accurate deep machine learning (ML) risk prediction tool to estimate the person-level 3-year lung cancer risk after the next scheduled repeat screening CT. Methods: Two ML predictors (ML1 and ML2) were developed from 25,097 participants who had received follow-up CT screenings in the National Lung Screening Trial (NLST). Double-blinded validation was performed using the Pan-Canadian Early Detection of Lung Cancer Study (PanCan). Performance of ML predictors was compared with Lung-RADS and volume doubling time (VDT) using time-dependent ROC analysis. Added values of ML predictors to Lung-RADS in identifying individuals with high cancer incidence risk and aggressive lung cancers were explored. Findings: In the PanCan validation cohort, ML2 consistently gave the highest time dependent AUC values of 0.968±0.013 and 0.946±0.013 respectively for cancer diagnosis within 1 and 2 years compared to 0.944±0.016 and 0.908±0.019 for Lung-RADS (p= 0.202, 0.048); and 0.830±0.300 and 0.777±0.029 for VDT(p<0.001). ML predictors performed better than LungRADS or VDT in stratifying lung cancer incidence and mortality risks. Interpretation: ML tool that recognizes patterns in both temporal and spatial changes as well as synergy among changes in nodule and non-nodule features may be used to accurately guide clinical management in a longitudinal screening program. Funding Statement: This study was funded by Allegheny Health Network Cancer Research, Johns Hopkins University Discovery Award, P30 CA006973, the Terry Fox Research Institute, and the BC Cancer Foundation.Declaration of Interests: There is no competing interest or conflict of interest from all authors. Ethics Approval Statement: We have institutional approved Material Transfer Agreement (MTA) touse data from both the National Lung Screening Trial (NLST) and the Pan-Canadian Early Detection of Lung Cancer (PanCan) among authors from Johns Hopkins University, University of British Columbia-British Columbia Cancer Agency, and National Cancer Institution to conduct work related to our submitted manuscript.