Abstract Rationale Artificial intelligence has shown promise in supporting clinicians’ performance in lung nodule stratification. Previous work demonstrated Bronchosolve’s capability as a fully automated closed-loop platform for end-to-end risk assessment of chest computed tomography (CT) scans with suspicious pulmonary nodules. However, the comparative and incremental diagnostic value of Bronchosolve when used alongside clinicians has not been examined. This study compared the discriminatory performance and accuracy of clinicians (radiologists and pulmonologists), Bronchosolve alone, and a combined (clinicians + Bronchosolve) approach across a multi-site dataset of chest CT scans with suspicious nodules. Methods We retrospectively analyzed 296 chest CT scans from 13 U.S. institutions (mean age 62.5 ± 7.1 years; 49.7% male), including both screening and incidental chest CT datasets, and representing five CT scanner manufacturers. Each case had a confirmatory biopsy or clinically documented stabilityfK1 (median 6.3 years) and a specific benign final diagnosis. Scans were independently reviewed by clinicians (5 radiologists and 5 pulmonologists) and processed by Bronchosolve. Three performance settings were evaluated: Clinicians-alone, Bronchosolve-alone, and Combined (Clinicians + Bronchosolve). Primary endpoints included sensitivity, specificity, and area under the receiver-operating-characteristic curve (AUC) with 95% confidence intervals (CIs) and pairwise p-values. Results Across all 296 cases, AUCs were 0.84 (0.83-0.86) for Clinicians, 0.87 (0.86-0.88) for Bronchosolve, and 0.88 (0.86-0.89) for the Combined approach (Figure 1). The improvement from Clinicians’ to Bronchosolve’s was statistically significant (p 0.0001), whereas the difference between Bronchosolve and Combined was not (p = 0.653). At the pre-specified high-sensitivity threshold for Bronchosolve, the Combined model increased readers’ sensitivity from 64% to 82%, which was paired with a corresponding decrease in specificity from 87% to 72%. Conclusions Bronchosolve demonstrated superior performance in standalone and combined reading formats. These findings support Bronchosolve’s potential as a scalable adjunct to detect malignant nodules early in daily clinical practice. This abstract is funded by: Imvaria Inc.
Taha et al. (Fri,) studied this question.