An artificial intelligence system has outperformed the average radiologist in estimating whether difficult-to-assess lung nodules are cancerous. That is the main finding of the international LUNA25 Challenge, led by Colin Jacobs of the Department of Medical Imaging at Radboudumc.
The study focused on small lung nodules found during lung cancer screening, between 5 and 15 millimeters in size, that are often hard to classify. Many of these abnormalities are harmless, but some are early cancers. Better estimates of which nodules are suspicious could help detect lung cancer sooner while also reducing unnecessary follow-up scans and procedures.
International AI challenge
In the LUNA25 Challenge, research teams from around the world developed AI systems to estimate the cancer risk of these nodules on low-dose CT scans. The systems were tested on an independent dataset from three major European lung cancer screening trials in Denmark, Italy, the Netherlands, and Belgium. In parallel, 65 radiologists from 23 countries assessed the same nodules, allowing a direct comparison between AI and human experts.
The best-performing AI system, developed by a start-up company in Germany, performed better than the average radiologist and outscored 62 of the 65 individual radiologists. At a clinically relevant threshold, the system correctly identified 12% more cancers and produced 20% fewer false alarms than the average radiologist.
“This finding underscores the potential of AI to act as a decision-support tool that can assist radiologists in one of the most challenging nodule categories seen in lung cancer screening.” says Dré Peeters, PhD candidate at the Diagnostic Image Analysis Group, Department of Medical Imaging at Radboudumc.
Towards clinical implementation
Lung cancer remains the leading cause of cancer-related death worldwide. Screening with low-dose CT scans can help detect the disease earlier in people at high risk, but interpreting the scan results is not always straightforward. The findings of this study suggest that AI could become a useful decision-support tool in screening programs by helping radiologists assess suspicious nodules more accurately and efficiently.
The researchers emphasize that further prospective studies are needed before such systems can be used in routine clinical care. Future research should now focus on how AI can be safely integrated into screening workflows and existing nodule management guidelines.
About the publication
The study was published in Radiology: Artificial Intelligence as: Benchmarking of AI and Radiologists for Indeterminate Lung Nodule Malignancy Risk Estimation on Screening CT: The LUNA25 Challenge, by D. Peeters, B. Obreja, N. Antonissen, Z. Saghir, U. Pastorino, M. Silva, G. H. de Bock, H. Gietema, F. Gleeson, M. A. Heuvelmans, S. Lam, G. Litjens, F. Mohamed Hoesein, C. Schaefer-Prokop, E. Scholten, A. Snoeckx, E. H. F. M. van der Heijden, R. Vliegenthart, M. Prokop, and C. Jacobs, on behalf of the LUNA25 Consortium.
