Recalibration of a Deep Learning Model for Low-Dose Computed Tomographic Images to Inform Lung Cancer Screening Intervals

Abstract

ImportanceAnnual low-dose computed tomographic (LDCT) screening reduces lung cancer mortality, but harms could be reduced and cost-effectiveness improved by reusing the LDCT image in conjunction with deep learning or statistical models to identify low-risk individuals for biennial screening.ObjectiveTo identify low-risk individuals in the National Lung Screening Trial (NLST) and estimate, had they been assigned a biennial screening, how many lung cancers would have been delayed 1 year in diagnosis.Design, Setting, and ParticipantsThis diagnostic study included participants with a presumed nonmalignant lung nodule in the NLST between January 1, 2002, and December 31, 2004, with follow-up completed on December 31, 2009. Data were analyzed for this study from September 11, 2019, to March 15, 2022.ExposuresAn externally validated deep learning algorithm that predicts malignancy in current lung nodules using LDCT images (Lung Cancer Prediction Convolutional Neural Network [LCP-CNN]; Optellum Ltd) was recalibrated to predict 1-year lung cancer detection by LDCT for presumed nonmalignant nodules. Individuals with presumed nonmalignant lung nodules were hypothetically assigned annual vs biennial screening based on the recalibrated LCP-CNN model, Lung Cancer Risk Assessment Tool (LCRAT + CT [a statistical model combining individual risk factors and LDCT image features]), and the American College of Radiology recommendations for lung nodules, version 1.1 (Lung-RADS).Main Outcomes and MeasuresPrimary outcomes included model prediction performance, the absolute risk of a 1-year delay in cancer diagnosis, and the proportion of people without lung cancer assigned a biennial screening interval vs the proportion of cancer diagnoses delayed.ResultsThe study included 10 831 LDCT images from patients with presumed nonmalignant lung nodules (58.7% men; mean [SD] age, 61.9 [5.0] years), of whom 195 were diagnosed with lung cancer from the subsequent screen. The recalibrated LCP-CNN had substantially higher area under the curve (0.87) than LCRAT + CT (0.79) or Lung-RADS (0.69) to predict 1-year lung cancer risk (P < .001). If 66% of screens with nodules were assigned to biennial screening, the absolute risk of a 1-year delay in cancer diagnosis would have been lower for recalibrated LCP-CNN (0.28%) than LCRAT + CT (0.60%; P = .001) or Lung-RADS (0.97%; P < .001). To delay only 10% of cancer diagnoses at 1 year, more people would have been safely assigned biennial screening under LCP-CNN than LCRAT + CT (66.4% vs 40.3%; P < .001).Conclusions and RelevanceIn this diagnostic study evaluating models of lung cancer risk, a recalibrated deep learning algorithm was most predictive of 1-year lung cancer risk and had least risk of 1-year delay in cancer diagnosis among people assigned biennial screening. Deep learning algorithms could prioritize people for workup of suspicious nodules and decrease screening intensity for people with low-risk nodules, which may be vital for implementation in health care systems.