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

Area of Science:

• Radiology and Oncology
• Artificial Intelligence in Medicine
• Medical Imaging Analysis

Background:

• Low-dose CT (LDCT) lung cancer screening typically focuses on nodules, potentially overlooking global lung features and underestimating long-term risk.
• Existing screening methods may not fully capture the complexity of lung cancer development.

Purpose of the Study:

• To develop and validate a novel lung cancer risk prediction model that incorporates both nodule and global lung features.
• To assess the added value of global lung features in improving 3-year lung cancer risk prediction compared to nodule-only models.

Main Methods:

• Development of ScreenLungNet, a 3-year risk prediction model using LDCT data from Chinese cohorts and clinical CT data.
• Extraction of features from multiple nodules and global lung characteristics using a vision transformer.
• Comparison of ScreenLungNet against single-nodule, multiple-nodule, and global-feature-only models using AUC, accuracy, sensitivity, specificity, PPV, and NPV.

Main Results:

• ScreenLungNet demonstrated superior 3-year lung cancer risk prediction performance (AUC 0.93-0.94) compared to nodule-only models.
• In the National Lung Screening Trial (NLST) cohort, ScreenLungNet achieved high performance (AUC 0.93, accuracy 94.8%, specificity 95.2%).
• The model maintained strong predictive capability in the NLST baseline-negative subset (AUC 0.87).

Conclusions:

• Integrating multiple nodule features and global lung features significantly enhances long-term lung cancer risk prediction.
• ScreenLungNet offers improved risk stratification and can predict risk in screening-negative participants.
• This model represents a advancement in leveraging comprehensive CT data for lung cancer screening.