Xiao, David, Forero, Yency, Kammer, Michael N., Chen, Heidi, Paez, Rafael, Heideman, Brent E., Owoseeni, Oreoluwa, Johnson, Ian, Deppen, Stephen A., Grogan, Eric L., & Maldonado, Fabien. (2025). Radiomic ‘Stress Test’: Exploration of a deep learning radiomic model in a high-risk prospective lung nodule cohort. *BMJ Open Respiratory Research, 12*(1), e002687. https://doi.org/10.1136/bmjresp-2024-002687
Lung nodules—small spots that appear on lung scans—are often biopsied to check for cancer. However, many of these nodules turn out to be harmless. The Lung Cancer Prediction (LCP) score is a deep learning tool that analyzes CT scans and has been shown to work well in identifying whether a nodule might be cancerous when it’s found by chance. But it hasn’t yet been tested in situations where doctors have already recommended a biopsy.
In this study, researchers looked at lung nodules that had already been biopsied at a large medical center. They used the Mayo Clinic’s traditional prediction model to estimate how likely each nodule was to be cancerous, dividing them into low, medium, or high risk using guidelines from the British Thoracic Society. Then, they compared how well three different models could predict cancer: the Mayo model, the LCP radiomic model, and a new integrated model that combined the LCP score with key clinical details like the patient’s age, whether the nodule had spiky edges (spiculation), and whether it was located in the upper part of the lung.
The study included 321 nodules total—196 cancerous and 125 benign (non-cancerous). The Mayo model had an accuracy score (AUC) of 0.69, the LCP model had a similar score of 0.67, but the integrated model performed best with an AUC of 0.75. It also had a better F1 score, which balances how well the model correctly identifies cancer and avoids false alarms. Importantly, the integrated model correctly reclassified 8 benign nodules from medium to low risk, meaning those patients might have avoided a biopsy—and no cancer cases were mistakenly downgraded.
In summary, combining the LCP deep learning score with a few key patient details improved the ability to predict whether lung nodules were cancerous. This approach may help reduce the number of people who undergo unnecessary, invasive lung biopsies.
Figure 1
Receiver operating characteristic (ROC) curves for all models. AUC, area under the receiver operating characteristic curve; LCP, Lung Cancer Prediction score; Mayo, Mayo model; Mayo Select, Mayo model excluding all radiographic variables; ROC, receiver operating characteristic.
