AI-Enabled Prediction of Future Lung Cancer Risk from Chest X-Ray Imaging for Personalized Lung Cancer Screening Management

Principal Investigator

Name
Jiantao Pu

Degrees
Ph.D

Institution
University of Pittsburgh

Position Title
Professor

Email
jip13@pitt.edu

About this CDAS Project

Study
PLCO (Learn more about this study)

Project ID
PLCOI-2000

Initial CDAS Request Approval
Dec 2, 2025

Title
AI-Enabled Prediction of Future Lung Cancer Risk from Chest X-Ray Imaging for Personalized Lung Cancer Screening Management

Summary
Lung cancer remains the leading cause of cancer mortality in the United States, and its prognosis strongly depends on early detection. Current screening guidelines rely primarily on age and smoking history, which do not capture the substantial inter-individual variability in lung cancer susceptibility. While low-dose CT (LDCT) is highly sensitive, it is resource-intensive, costly, and not broadly accessible, making optimized selection of screening intervals essential. Chest X-ray (CXR) imaging is widely available, inexpensive, and routinely collected in clinical care and in large national cohorts such as PLCO and NLST. Recent advances in artificial intelligence have demonstrated that CXR contains latent information about biological aging, body composition, vascular patterns, and early disease signatures that are not visible to human readers but may be predictive of future lung cancer risk.

This project aims to develop and validate a deep learning–based framework that predicts long-term lung cancer risk directly from CXR images and integrates this risk with demographic and clinical factors to support personalized lung cancer screening recommendations. We will leverage the rich imaging and outcome data from PLCO and NLST to (1) identify image-based biomarkers associated with incident lung cancer, (2) quantify how CXR-derived risk improves upon conventional risk models (e.g., PLCOm2012), and (3) explore strategies for adaptive screening intervals based on individualized CXR-based risk stratification. The results will inform precision approaches to lung cancer screening that prioritize high-risk individuals for LDCT while reducing unnecessary scans in low-risk groups.

This work has the potential to significantly advance population-level screening effectiveness, reduce overuse of CT imaging, and identify previously unrecognized imaging phenotypes associated with early lung carcinogenesis.

Aims

Aim 1. Develop a deep learning model to predict future lung cancer from baseline CXR.
We will train a convolutional neural network (CNN) or vision transformer (ViT) using PLCO CXR images linked to long-term cancer outcomes. The model will learn imaging biomarkers (e.g., chest morphology, vascularity, body composition patterns) associated with future lung cancer risk. We will evaluate performance using AUC, calibration, and time-dependent risk metrics.

Aim 2. Validate model generalizability using NLST CXR images and compare with established risk models.
We will externally validate the CXR-based model in NLST and assess its incremental value over baseline demographic/smoking-based predictors such as PLCOm2012. Model improvement will be quantified using net reclassification improvement (NRI), integrated discrimination improvement (IDI), and decision-curve analysis.

Collaborators

Jiantao Pu University of Pittsburgh