Prediction of lung cancer risk at follow-up screening with low-dose CT: a training and validation study of a deep learning method. Issue 7 (November 2019)

• Record Type: Journal Article
• Authors: Huang, Peng; Lin, Cheng T; Li, Yuliang; Tammemagi, Martin C; Brock, Malcolm V; Atkar-Khattra, Sukhinder; Xu, Yanxun; Hu, Ping; Mayo, John R; Schmidt, Heidi; Gingras, Michel; Pasian, Sergio; Stewart, Lori; Tsai, Scott; Seely, Jean M; Manos, Daria; Burrowes, Paul; Bhatia, Rick; Tsao, Ming-Sound; Lam, Stephen
• Is Part Of: Lancet. Volume 1:Issue 7(2019)
• Page Start: e353
• Page End: e362
• Publication Date: 2019-11
• Subjects: Medical care -- Data processing -- Periodicals; Medical care -- Information technology -- Periodicals; Medical informatics -- Periodicals; 610.285
• DOI: 10.1016/S2589-7500(19)30159-1
• Languages: English
• ISSNs: 2589-7500

Abstract

Summary: Background: Current lung cancer screening guidelines use either mean diameter, volume, or density of the largest lung nodule on the previous CT scan or appearance of a new nodule to ascertain the timing of the next CT scan. We aimed to develop an accurate screening protocol by estimating the 3-year lung cancer risk after two screening CT scans using deep learning of radiologists' CT readings and other universally available clinical information. Methods: A deep learning algorithm (referred to as DeepLR) was developed using data from participants who had received at least two CT screening scans up to 2 years apart in the National Lung Screening Trial (NLST; training cohort). Double-blinded validation was done using data from participants in the Pan-Canadian Early Detection of Lung Cancer (PanCan) study (validation cohort). The primary analysis was to compare accuracy of DeepLR scores to predict lung cancer incidence at 1 year, 2 years, and 3 years with the Lung CT Screening Reporting & Data System (Lung-RADS) and volume doubling time, using time-dependent area under the receiver operating characteristic curve (AUC) analysis. Findings: The training cohort consisted of 25 097 participants from NLST and the validation cohort comprised 2294 individuals from PanCan. In the validation cohort, DeepLR showed good discrimination, with 1-year, 2-year, and 3-year time-dependent AUC values for cancer diagnosis of 0·968 (SD 0·013), 0·946 (0·013), and 0·899 (0·017), respectively. Among individuals deemed high risk by DeepLR, 94%, 85%, and 71% of incident and interval lung cancers diagnosed within 1 year, 2 years, and 3 years, respectively, after the second screening CT scan were identified. Furthermore, individuals with high DeepLR scores had a significantly higher risk of mortality (hazard ratio 16·07, 95% CI 10·15–25·44; p<0·0001) among people with high scores on Lung-RADS. Interpretation: DeepLR recognises patterns in both temporal and spatial changes and synergy among changes in nodule and non-nodule features. DeepLR scores could be used to accurately guide clinical management after the next scheduled repeat screening CT scan. Funding: Allegheny Health Network, Johns Hopkins University, Terry Fox Research Institute, and British Columbia Cancer Foundation.