Assessing the Usefulness of AI in Lung Cancer Detection on Chest Radiographs: A Retrospective Study Using NLST Data
Principal Investigator
Name
Dennis Robert
Degrees
M.B.B.S, M.M.S.T.
Institution
Qure.ai Technologies Private Limited
Position Title
Director of Clinical Research
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-1400
Initial CDAS Request Approval
Mar 5, 2025
Title
Assessing the Usefulness of AI in Lung Cancer Detection on Chest Radiographs: A Retrospective Study Using NLST Data
Summary
Lung cancer remains the leading cause of cancer-related mortality worldwide, and early detection is crucial for improving survival. While the National Lung Screening Trial (NLST) established the superiority of low-dose CT (LDCT) over chest X-ray (CXR) for lung cancer screening, CXRs continue to be widely used in resource-limited settings.
Artificial intelligence (AI)-powered radiology tools, such as Qure.ai’s qXR, have demonstrated potential in detecting lung abnormalities on CXRs. However, their effectiveness in high-risk screening populations has not been rigorously validated. This study will evaluate qXR’s performance in lung cancer detection by comparing AI-generated findings from NLST CXRs with radiologist interpretations and confirmed lung cancer diagnoses. The study will utilize NLST ACRIN chest X-rays data to assess qXR's performance in three key tasks:
- Identifying lung cancer cases accurately to improve early detection.
- Differentiating benign abnormalities from malignant findings to reduce unnecessary follow-ups.
- Comparing qXR’s diagnostic accuracy with radiologists to assess its potential clinical impact.
A retrospective case-control design will be used, with participants classified into three groups:
- Cancer-Positive: Participants with a valid screening result (screen positive or screen negative) who were diagnosed with lung cancer during screening phase.
- Cancer-Negative: Participants not diagnosed with lung cancer throughout the study
The study will apply qXR to NLST CXRs, extract AI-predicted abnormalities, and compare them with radiologist interpretations and lung cancer confirmation data. Performance evaluation will include sensitivity, specificity, false positive/negative rates, ROC/AUC analysis, and agreement metrics.
Artificial intelligence (AI)-powered radiology tools, such as Qure.ai’s qXR, have demonstrated potential in detecting lung abnormalities on CXRs. However, their effectiveness in high-risk screening populations has not been rigorously validated. This study will evaluate qXR’s performance in lung cancer detection by comparing AI-generated findings from NLST CXRs with radiologist interpretations and confirmed lung cancer diagnoses. The study will utilize NLST ACRIN chest X-rays data to assess qXR's performance in three key tasks:
- Identifying lung cancer cases accurately to improve early detection.
- Differentiating benign abnormalities from malignant findings to reduce unnecessary follow-ups.
- Comparing qXR’s diagnostic accuracy with radiologists to assess its potential clinical impact.
A retrospective case-control design will be used, with participants classified into three groups:
- Cancer-Positive: Participants with a valid screening result (screen positive or screen negative) who were diagnosed with lung cancer during screening phase.
- Cancer-Negative: Participants not diagnosed with lung cancer throughout the study
The study will apply qXR to NLST CXRs, extract AI-predicted abnormalities, and compare them with radiologist interpretations and lung cancer confirmation data. Performance evaluation will include sensitivity, specificity, false positive/negative rates, ROC/AUC analysis, and agreement metrics.
Aims
- Evaluate the sensitivity and specificity of qXR for lung cancer detection, using confirmed cancer diagnoses as the ground truth.
- Compare the agreement between qXR and NLST radiologists in detecting lung nodules and other abnormalities.
- Estimate the proportion of missed cancers detected by qXR, focusing on cases where radiologists reported a negative screen result, but cancer was later diagnosed.
- Assess the clinical impact of AI-based screening by investigating if there are any reductions in false positives and false negatives screening results when using AI.
Collaborators
Dr. Dennis Robert, Qure.ai
Dr. Santosh S, Qure.ai
Manoj Tadepalli, Qure.ai