Incidental lung nodules are commonly detected on CT scans, and while most are benign, some may represent early-stage lung cancer. Accurately stratifying nodules is challenging as over-investigating low-risk nodules can lead to complications, while triaging high-risk nodules as low-risk can delay cancer diagnosis. Guidelines incorporating nodule size as a critical risk factor have been developed, but the malignancy rate in solid nodules ≥15 mm is still variable.
Additional non-invasive biomarkers are needed to identify those at the highest risk. The British Thoracic Society guidelines currently use a Brock score threshold of ≥10% to trigger further investigation of solid nodules, which may not stratify large nodules well. Furthermore, the COVID-19 pandemic has disrupted diagnostic services, making decision support even more critical.
Radiomics models have been developed for nodule classification, but no studies have explored the utility of radionics in large nodules (≥15 mm). Through the Lung Imaging Biobank for Radiomics and AI research (LIBRA), the authors aimed to develop a pipeline for multi-center radiomics research, create a radiomics algorithm to accurately classify large lung nodules, and develop a decision-support tool to reduce delayed cancer diagnosis rates in the broad 10–70% Herder risk group.
As per The Lancet, a team of researchers has established a national pipeline for AI-based lung cancer early diagnosis research through the LIBRA study. The pipeline incorporates heterogeneous data from multiple institutions and scans vendors to develop the LN-RPV, an artificial intelligence algorithm targeted at large lung nodules.
The LN-RPV algorithm performed better than the median radiologist and can be integrated with the British Thoracic Society (BTS) guidelines to reduce the risk of delayed cancer treatment. The BTS algorithm and Herder score are widely used across the UK for nodule stratification. The LN-RPV can potentially improve early cancer diagnosis and treatment by highlighting which patients are high-risk and recommending they be fast-tracked to intervention.
The LN-RPV consists of only two features compared to the seven values input for the Herder score. This could streamline or automate the process of nodule risk calculation. For centers without routine access to PET or where PET scanning will be delayed, the LN-RPV could indicate malignancy probability earlier.
The LN-RPV’s first feature is the nodule surface-to-volume ratio, defined as the surface area divided by the total volume (SNS_s2v). The second feature is the gray-level co-occurrence matrix (GLCM) correlation (GLCM_Correl). GLCM features have been used to classify benign and malignant lesions in other disease groups, including breast cancer. In non-small cell lung cancer (NSCLC), GLCM features are associated with the degree of tumor immune infiltration, PDL1 expression, and patient survival.
Although the LN-RPV retained good performance in the external test set (accuracy 76%), this data was obtained from public imaging databases which may not closely match the setting in which the algorithm is intended. Therefore, additional external testing with large, representative datasets is required before generalizable clinical use. Prospective evaluation in a real-world nodule MDT is the next step to verify its clinical utility.
Aside from the external test set, there are some limitations to consider. Firstly, the model does not incorporate changes in radionics features over time, which is an area for future development. Secondly, though an auto-segmentation pipeline has been developed, a truly integrated solution that unifies all pre-processing, segmentation, and extraction steps into a single program has not yet been developed. Finally, blaming the PET as unfavorable when missing could underestimate Herder score performance in the clinical decision-support scenario.
In conclusion, the LN-RPV algorithm can potentially improve early cancer diagnosis and treatment by highlighting which patients are high-risk and recommending they be fast-tracked to intervention. However, additional external testing with large, representative datasets is required before generalizable clinical use. Prospective evaluation in a real-world nodule MDT is the next step to verify its clinical utility.
