Emerging AI CT Tool May Provide Viable Complement to Lung-RADS Classification

New research reveals that an artificial intelligence (AI)-enabled computer-assisted diagnosis (CAD) tool may offer significantly enhanced screening sensitivity and specificity for detecting malignant lung nodules on computed tomography (CT) in comparison to the Lung-RADS classification system.

In a retrospective review of data from 963 patients with a total of 1,331 lung nodules, researchers compared a machine learning classification tool (RevealAI™-Lung) to the Lung-RADS classification system. Trained with imaging data from the National Lung Screening Trial (NLST), the classification software reviews CT scans and nodule location to compute a “malignancy Similarity Index” (mSI) numerical value ranging from 0 (benign) to 1 (cancer), according to the study, which was recently published in the Journal of the American College of Radiology.

In an NLST cohort involving 704 nodules, the study authors said the combination of the mSI score and the Lung-RADS classification system improved sensitivity by 68 percent rate and specificity by 17 percent in comparison to Lung-RADS alone. In a validation cohort involving data from a lung screening program, the combination of the mSI score and Lung-RADS improved sensitivity by 25 percent and specificity by 33 percent in comparison to Lung-RADS alone, according to the study. In another cohort, including 187 patients who had lung nodules detected outside of a lung cancer screening program, the study authors noted a 117 percent improvement in sensitivity and an 18 percent improvement in specificity for the combination of the mSI score and Lung-RADS in comparison to Lung-RADS alone.

“When combined with existing clinical risk models, (RevealAI-Lung) provides additional diagnostic value and more accurate prediction of malignancy risk in multiple cohorts,” wrote Scott J. Adams, M.D., Ph.D., who is affiliated with the Department of Medical Imaging at the University of Saskatchewan in Saskatoon, Canada, and colleagues.

In a subsequent analysis of malignant nodules in the NLST cohort and CT scans obtained one year prior to diagnostic confirmation, Adams and colleagues found that 53 patients had high mSI values and the average time to confirmed diagnosis for these patients was 502 days after the CT scan. The study authors pointed out that mSI scoring would have reclassified 36 percent of patients, initially diagnosed with a Lung-RADS category under 4B, as high risk.

“We demonstrate that mSI identifies nodules as likely malignant over 1 year before they were definitively diagnosed in practice, demonstrating the utility of supervised machine learning applications for early detection and management of lung cancer,” noted Adams and colleagues.

The study authors also maintained that the mSI’s ability to accurately classify low-risk nodules could help prevent unnecessary procedures.

“For example, in the NLST, 24% of thoracotomies, 75% of bronchoscopies, and 67% of needle biopsies were performed on patients who did not have confirmed lung cancer,” noted Adams and colleagues.

In regard to study limitations, the authors conceded that the criteria for the retrospective study may have led to a sample bias toward higher-risk nodules that would warrant a more detailed follow-up. Even with multi-year radiology follow-up, Adams and colleagues acknowledged that some cancerous nodules may be characterized as benign. They also pointed out that the challenges of tissue sampling with small lung nodules may contribute to inaccuracies with pathological confirmation.