Deep learning to identify high-risk smokers for lung cancer screening from chest radiographs.

Abstract

1519 Background: Appearance on chest radiography may inform selection of high-risk smokers for lung cancer screening CT, beyond Centers for Medicare & Medicaid Services (CMS) eligibility criteria. Methods: A convolutional neural network (CXR-LC) predicting 12-year incident lung cancer from the chest radiograph image, age, sex, and smoking status (current/former) was developed in 41,856 persons aged 55-74 from the Prostate, Lung, Colorectal & Ovarian trial (PLCO). The final model was tested in held-out smokers from PLCO (n=5,615, 37.9% CMS eligible, 12-year follow-up), and externally in the National Lung Screening Trial (NLST, n=5,493, all CMS eligible, 6-year follow-up). Sensitivity was compared at a fixed screening population size defined by CMS eligibility. Ordinal CXR-LC risk score (low/indeterminate/high/very high) was based on development dataset 12-year probability thresholds (<2%/2-<3.3%/3.3-<8%/≥8%). Results are provided in test datasets only. Results: In the PLCO test dataset, CXR-LC was more sensitive than CMS eligibility at a fixed screening population size (74.9% vs. 63.8%, p=0.01) and missed 30.8% fewer lung cancers. CXR-LC risk groups were associated with incident lung cancer in PLCO test dataset smokers (very high vs. low CXR-LC risk: 12.4 vs 1.1 lung cancers/1,000 person-years) with external testing in NLST (all CMS eligible: 12.7 vs 2.3) (Table). This association was robust to adjustment for radiologist findings and the PLCOm2012 risk score. Conclusions: CXR-LC identified smokers at high risk of incident lung cancer, beyond CMS eligibility. [Table: see text]