Lung cancer is the leading cause of cancer-related deaths globally, resulting in over 1.8 million fatalities each year. Early detection through low-dose computed tomography (LDCT) screening significantly improves survival rates by identifying malignancies early. While screening benefits are well-recognised, large-scale implementation poses challenges, particularly the increased workload for radiologists reviewing numerous CT scans. Artificial intelligence has been suggested as a solution, providing automated detection and classification of lung nodules, often using a ‘rule-in’ approach to differentiate between benign and malignant nodules. However, these systems have not yet been fully integrated into clinical practice due to concerns about reliability.
An alternative, the ‘rule-out’ method, assumes that about half of CT scans in screening populations show no significant nodules. If AI can accurately identify these negative scans, it could significantly reduce radiologists' workload. A recent study published in the European Journal of Cancer evaluates a commercially available AI system using the rule-out approach in the UK Lung Cancer Screening (UKLS) trial, comparing AI performance with human radiologists and estimating the potential reduction in workload.
AI Performance in Lung Nodule Classification
The study assessed 1,252 baseline LDCT scans using both AI and human readers, comparing their classifications against an expert radiological reference standard and histological lung cancer outcomes. AI was evaluated on two levels: its classification accuracy compared to an expert panel and its ability to detect histologically confirmed lung cancers. AI correctly classified lung nodules with fewer misclassifications than human readers. When compared to the expert reference standard, AI achieved a negative predictive value (NPV) of 92.0%, indicating that it was highly reliable in identifying scans without significant lung nodules.
When benchmarked against histological lung cancer outcomes, AI demonstrated even stronger performance. Of the 31 baseline-round lung cancers detected in the study, AI successfully identified all cases, with only one being misclassified as negative due to its small volume, falling below the 100mm³ threshold. This resulted in an NPV of 99.8%, meaning that nearly all lung cancers requiring follow-up would have been correctly identified. This performance was comparable to or better than human readers, with AI producing fewer discrepancies and misclassifications.
Recommended Read: Advancing Global Low-Dose CT Screening for Lung Cancer
These findings suggest that AI can be an effective tool in lung cancer screening, particularly in reducing false-negative classifications. By automating the detection and classification of lung nodules, AI has the potential to improve screening efficiency without compromising diagnostic accuracy. However, considerations regarding threshold settings and validation across different screening populations remain essential for ensuring the reliability of AI-assisted screening.
Impact on Radiologist Workload
The primary benefit of integrating AI into lung cancer screening is the potential for reducing radiologist workload. The study estimated that AI, when used as a first-reader to rule out negative cases, could reduce CT-reading workload by up to 79%. This reduction is achieved by classifying scans as negative when no significant lung nodules (≥100mm³) are present, allowing radiologists to focus on reviewing only indeterminate or positive cases.
In practical terms, this means that AI could remove a substantial number of scans from the radiologist’s workflow, streamlining the screening process. AI correctly classified a majority of negative scans, ensuring that radiologists spend their time analysing only those cases requiring further evaluation. Minimum workload reduction was estimated at 67%, with a potential maximum of 79% when considering additional negative cases from the UKLS trial.
These findings highlight the efficiency gains that AI can bring to lung cancer screening, making large-scale programmes more feasible. However, while AI shows promise in reducing workload, its clinical integration must be carefully managed. The study acknowledges that AI should operate within an oversight framework, where flagged cases are still subject to human review. This ensures that potential errors are minimised and maintains confidence in AI-assisted screening.
Considerations for Clinical Integration
Despite its strong performance, several key considerations must be addressed before AI can be fully integrated into lung cancer screening. Trust in AI-driven diagnoses remains a central challenge, particularly as no universal standard exists for determining acceptable AI performance in clinical practice. Establishing clear regulatory guidelines and defining appropriate performance thresholds will be crucial for ensuring that AI can be safely and effectively deployed.
Additionally, this study focused solely on baseline scans, leaving questions about AI’s ability to assess lung nodule progression over multiple screening rounds. Longitudinal tracking of nodules, particularly in follow-up scans, is essential for assessing malignancy risk. Future research should evaluate AI’s performance in monitoring nodule growth over time, as well as its effectiveness in different healthcare settings and diverse patient populations.
Another factor to consider is the management of incidental findings. AI algorithms are primarily designed to assess lung nodules, but screening scans often reveal additional findings unrelated to lung cancer, such as cardiovascular abnormalities or other lung diseases. Standardising protocols for handling incidental findings will be important to ensure that AI-driven screening does not overlook clinically significant conditions.
The study findings indicate that AI can significantly enhance lung cancer screening by improving efficiency and reducing the workload for radiologists. By acting as a first-reader, AI can quickly classify negative CT scans, allowing radiologists to focus on complex cases. With a high accuracy and a negative predictive value (NPV) of 99.8%, AI is reliable in ruling out negative cases.
The potential workload reduction of up to 79% underscores its transformative impact on screening programmes. However, further validation across different populations and regulatory frameworks is necessary before full integration into clinical practice. Overall, AI-assisted lung cancer screening offers a promising avenue for early detection and optimised healthcare resources, with the potential to improve outcomes globally.
Source: European Journal of Cancer
Image Credit: iStock
