An AI-based radiomics tool was significantly more effective than genomic sequencing for risk stratification of incidental pulmonary nodules, based on data presented at the Annual Conference of the American Association for Bronchology and Interventional Pulmonology.

An estimated 1.6 million incidental pulmonary nodules are detected annually in the US, according to a recent study. Data from incidental nodule programs suggest lung cancer detection rates range from 3% to 5%; therefore, accurate classification and risk stratification is important, said lead author Christina Bellinger, MD, professor and director of bronchoscopy and interventional pulmonology and director of the lung screening and incidental nodule program at Wake Forest University School of Medicine, Winston-Salem, North Carolina.

“Determining which nodules warrant invasive biopsy vs surveillance remains a major clinical challenge, as many nodules measuring 8-20 mm fall into an intermediate-risk category when assessed with traditional risk calculators such as the Mayo Clinic model, PET imaging, or interval surveillance,” Bellinger said.

“Artificial intelligence-based radiomic models have emerged as a promising approach to improve risk estimation and potentially reduce both unnecessary invasive procedures and delayed cancer diagnoses,” Bellinger added. However, data on the effectiveness of an AI-based radiomics tool using a Lung Cancer Prediction Score (LCP Score) and a Genomic Sequence Classifier (GSC) are lacking, she said.

Bellinger and colleagues reviewed data from 59 patients having incidental lung nodules with pathologic confirmation and GSC results who were enrolled at a single pulmonary clinic. Radiomic LCP scores were determined based on computed tomography imaging. Both the LCP Score and GSC were assessed using area under the curve (AUC). Net Reclassification Improvement (NRI) was used to assess the diagnostic value of AI compared to GSC at a risk score cutoff of 8.5 on a scale of 1 to 10.

Overall, LCP showed greater accuracy than GSC (AUC, 0.82 vs 0.59), and the AI radiomics tool showed an NRI of 45.41% over GSC. At the NRI cutoff, LCP showed significantly greater sensitivity than GSC (100% vs 26.1%) but a lower specificity (38.9% vs 86.1%). The positive predictive values were 51.1% and 54.5% for LCP and GSC, respectively, and negative predictive values were 100% and 64.6%, respectively. These data equated to identification of all 23 cancers using the AI tool compared to identification of 6 of 23 cancers using GSC, the researchers noted in their abstract.

“Based on prior studies evaluating AI-based lung cancer prediction models, along with our own clinical experience using genomic classifiers and AI-assisted risk stratification, I anticipated that the AI platform would outperform the genomic sequencing classifier in its ability to reclassify indeterminate nodules,” said Bellinger.

In practice, the results suggest that AI-based lung cancer prediction may become an additional decision-support tool for clinicians managing indeterminate pulmonary nodules, said Bellinger. “These tools are not intended to replace clinical judgment, but rather to complement existing clinical, radiographic, and patient-specific information,” she emphasized.

The study findings were limited by the relatively small study population and retrospective design, Bellinger said. “While the findings are encouraging, larger prospective studies are needed to determine how AI-based risk prediction performs when incorporated into real-world clinical workflows,” she noted.

Prior studies, such as a 2023 research letter in Respirology, have demonstrated the influence of AI risk assessment on clinical management decisions for indeterminate pulmonary nodules, Bellinger noted. “However, prospective outcome-based research is still needed to determine whether routine use of these tools translates into improved patient outcomes, reduced invasive testing, and earlier cancer detection,” she said.

In the Clinic

The current study addresses a common and difficult clinical problem of managing indeterminate pulmonary nodules, particularly those in the intermediate-risk range where the next step is often uncertain, said Jamie Garfield, MD, professor of thoracic medicine and surgery at the Lewis Katz School of Medicine at Temple University, Philadelphia.

“As both genomic classifiers and AI-based imaging tools become more available, direct comparison of these approaches is increasingly relevant for real-world decision-making,” said Garfield, who was not involved in the study.

The findings were interesting but should be interpreted with caution, Garfield emphasized. Although the AI radiomics tool performed better than the genomic classifier in this study, “these tests were not developed for the same patient population or the same clinical question,” she said. “The AI tool worked well as a rule-out test, with 100% sensitivity and negative predictive value but low specificity, while the genomic classifier was more specific but missed many cancers,” she said.

Consequently, the researchers’ statement that the AI tool identified 23/23 cancers “sounds stronger than it really is because it reflects a cutoff designed to avoid missing cancer rather than true overall diagnostic superiority,” said Garfield. AI-based CT tools may help identify low-risk nodules and potentially reduce unnecessary procedures, but these tests should not be considered interchangeable, she said. “Genomic classifiers and imaging-based tools are designed for different clinical contexts, so their value depends on where they are used and how results are incorporated into the broader clinical picture,” she noted.

Garfield concurred that prospective validation is needed in populations where each test is intended to be used. “Future studies should examine whether combining imaging-based and molecular approaches improves risk stratification, reduces unnecessary procedures, and leads to better patient outcomes, and decision-impact studies will be especially important before these tools are broadly adopted,” she added.

The study received no outside funding. Bellinger had no financial disclosures directly related to the current study but has participated in separate research studies supported by Optellum, the company that developed the AI lung cancer prediction platform evaluated in this study. Garfield had no financial conflicts to disclose.