AI May Bolster Assessment of Pulmonary Nodules on Chest CT

Emerging research suggests that artificial intelligence (AI) can improve the assessment of indeterminate pulmonary nodules on chest computed tomography (CT) exams.

In a recently published retrospective study in Radiology, researchers examined the use of an AI-enabled, computer-aided diagnosis (CAD) software (Virtual Nodule Clinic, Optellum) for helping to determine the malignancy risk of indeterminate pulmonary nodules on chest CT. Six radiologists and six pulmonologists reviewed chest CT exams in 300 patients (mean age of 65) with pulmonary nodules ranging between 5 to 30 mm in size.

The researchers found that use of the CAD software improved the average sensitivity (from 52.6 percent to 63.1 percent) and specificity (87.3 percent to 89.9 percent) for high-risk nodules (malignancy risk threshold of 65 percent). For low-risk nodules (malignancy risk threshold of 5 percent), the study authors also noted improved sensitivity (94.1 percent to 97.9 percent) and specificity (37.4 percent to 42.3 percent) with the AI-powered software.

“The readers judge malignant or benign with a reasonable level of accuracy based on imaging itself, but when you combine their clinical interpretation with the AI algorithm, the accuracy level improves significantly,” noted study co-author Anil Vachani, M.D. MSCE, the director of the Lung Nodule Program and co-director of the Lung Cancer Screening Program at the Perelman School of Medicine at the University of Pennsylvania. “The level of improvement suggests that this tool has the potential to change how we judge cancer versus benign and hopefully improve how we manage patients.”

Use of the CAD software also improved interobserver agreement for the diagnosis of low-risk nodules by 21 percent and high-risk nodules by 17 percent, according to the study. Researchers also noted an overall increase of eight percent for interobserver agreement on management recommendations for indeterminate pulmonary nodules.

In an accompanying editorial, Masahiro Yanagawa, M.D., Ph.D. said the improved assessment of indeterminate pulmonary nodules and improved agreement among clinicians on low- and high-risk nodules are promising.

“These results are crucial for brining AI-based tools closer to clinical implementation for risk stratification and patient management,” wrote Dr. Yanagawa, an associate professor at the Graduate School of Medicine/Faculty of Medicine at Osaka University in Japan.

Dr. Yanagawa also noted that future prospective research is necessary to validate the results.

“The evaluation of subsolid nodules as well as ground-glass nodules (which change little over time) will be an important area for future validation of the AI system,” added Dr. Yanagawa.

One factor limiting the application of results to a clinical setting was the small number of partially solid nodules in the study, according to Vachani and colleagues. While the intent of the study was to determine the impact of the AI-enabled CAD in assessing risk based on nodule characteristics, the authors noted that not supplying patient clinical information to those reading the CT exams limits general application of the study results to a clinical setting.

Prior to interpreting the chest CT images, the study authors noted the readers were informed that the prevalence of malignancy in the pulmonary nodules being assessed was higher than what one would normally see in clinical practice, and this may have led to inflated risk estimates in the study.