ScreenLungNet: Personalized Long-Term Prediction of Lung Cancer Risk from a Single Low-Dose CT Screening.

Abstract

Background Lung cancer screening with low-dose CT (LDCT) primarily relies on identifying nodules, disregarding global lung features, and may underestimate long-term cancer risk. Purpose To determine whether considering global lung features alongside nodules could improve long-term lung cancer risk prediction. Materials and Methods In this retrospective study, ScreenLungNet, a 3-year lung cancer risk prediction model, was developed using LDCT data from the Wenling 2019-2020 screening cohort (Zhejiang Province, China) and clinical CT data from Zhejiang Cancer Hospital (2017-2023). Features from multiple nodules derived from malignancy scores and global lung features were extracted via a vision transformer and integrated into the model. To assess the added value of global lung features, single-nodule (based on the highest-risk nodule), multiple-nodule (based on multiple-nodule features), and global (based on global lung features) lung models were constructed for comparison. Model performance was assessed with the area under the receiver operating characteristic curve (AUC), accuracy, specificity, sensitivity, positive predictive value (PPV), and negative predictive value (NPV). Results Model training included 19 869 participants from the Wenling 2019-2020 screening cohort (mean age, 60.93 years ± 6.30 [SD]; 13 123 male; 343 patients with lung cancer) and 1802 patients from Zhejiang Cancer Hospital. Finally, independent internal and external tests were conducted on the Wenling 2021 (<i>n</i> = 9932) and National Lung Screening Trial (NLST) (<i>n</i> = 14 966) cohorts, respectively. ScreenLungNet outperformed other models for 3-year lung cancer risk prediction (AUC range, 0.93-0.94; <i>P</i> < .001). In the NLST cohort, ScreenLungNet performed well (AUC, 0.93 [95% CI: 0.92, 0.94]; accuracy, 94.8%; sensitivity, 84.9%; specificity, 95.2%; PPV, 44.0%; NPV, 99.3%; all <i>P</i> < .02). Its performance remained strong in the NLST baseline-negative subset (AUC, 0.87; accuracy, 97.0%; sensitivity, 72.1%; specificity, 97.5%). Conclusion A model integrating multiple nodules and global lung features outperformed nodule-only models for long-term lung cancer risk prediction and enabled risk prediction in screening-negative participants. © The Author(s) 2025. Published by the Radiological Society of North America under a CC BY 4.0 license. <i>Supplemental material is available for this article.</i> See also the editorial by Staziaki in this issue.