Related Experiment Video
Updated: Jan 10, 2026
Author Spotlight: Advancing 3D Modeling for Enhanced Diagnosis and Treatment of Pulmonary Nodules in Early-Stage Lung Cancer
Published on: October 13, 2023
Potential for AI as first reader in lung cancer screening.
Roberta Eufrasia Ledda1, Camilla Valsecchi2, Federica Sabia2
1Thoracic Surgery Unit, Fondazione IRCCS Istituto Nazionale dei Tumori, Milan, Italy; Department of Medicine and Surgery (DiMeC), University of Parma, Parma, Italy.
This study compared artificial intelligence (AI) and human readings for lung cancer screening using low-dose computed tomography (LDCT). AI showed similar sensitivity but lower specificity than humans, though its high negative predictive value may support its use as a first reader.
Area of Science:
- Radiology
- Artificial Intelligence
- Oncology
Background:
- Lung cancer screening (LCS) with low-dose computed tomography (LDCT) is crucial for early detection.
- Standardizing interpretation of LDCT scans according to Lung-Imaging Reporting and Data System (LungRADS) is essential for accurate lung cancer diagnosis.
- Evaluating automated reading systems can potentially improve efficiency and consistency in LCS.
Purpose of the Study:
- To assess the agreement between human and artificial intelligence (AI)-based readings of LDCT scans in lung cancer screening (LCS) using LungRADS v1.1.
- To evaluate the diagnostic performance (sensitivity, specificity, PPV, NPV) of both human and AI readings.
- To determine the potential impact of AI on reducing the workload for LDCT interpretation.
Main Methods:
- Retrospective analysis of 4104 baseline LDCT scans from the BioMILD trial.
- Classification of LDCT scans as "negative" or "positive" based on LungRADS v1.1 categories by both a radiologist and AI software.
- Assessment of diagnostic performance using lung cancer diagnosis at 2 years as the reference standard; agreement measured by k-Cohen Index with Fleiss-Cohen weights.
Main Results:
- AI reading demonstrated comparable sensitivity (91.2% vs 89.7%) but lower specificity (75.7% vs 90.0%) than human reading.
- Agreement between human and AI readings was 83.5% (Kw 0.47).
- AI reading is expected to reduce LDCT reading workload by 74.7% and showed a high negative predictive value (99.8%).
Conclusions:
- AI reading in LCS shows comparable sensitivity to human reading but with lower specificity.
- The high negative predictive value of AI suggests its potential utility as a first reader in lung cancer screening.
- AI has the potential to significantly decrease the workload associated with LDCT interpretation in lung cancer screening programs.
