ECR 2026: Poster Presentation
Jiye Kim, Maurits Engbersen, Jeroen Mollink, Maarten Poirot, Ieva Kurilova, Brittaney Everitt, Akira Hasegawa, Bryan Haslam
Abstract
Purpose
Lung nodule detection and tracking on low-dose chest CT is time intensive and subject to substantial interobserver variability, impacting early-stage lung cancer detection. This study aimed to determine whether AI-assisted lung nodule detection improves radiologists’ diagnostic performance, consistency, and efficiency across nodule sizes.
Materials and Methods
A blinded, fully crossed multi-reader multi-case study was performed using 300 low-dose chest CT examinations (134 abnormal, 166 normal) containing 253 reference-standard lung nodules. Fourteen radiologists independently reviewed all cases with and without AI assistance, separated by a 4-week washout period. Readers were instructed to detect and localize lung nodules and assign suspicion scores. Diagnostic performance was assessed using localized receiver operating characteristic (LROC) analysis, lesion-level sensitivity, exam-level specificity, and interobserver agreement (Fleiss’ kappa). Interpretation time was compared between aided and unaided reading. Size-dependent detection performance was evaluated using pre-specified subgroup analyses stratified by nodule size.
Results
Mean LROC AUC increased by 21.2% with AI, from 0.612 (95% CI: 0.540–0.685) to 0.742 (95% CI: 0.680–0.804; p<0.001). Lesion-level sensitivity improved by 27.8% from 0.547 to 0.700 (p<0.001), while exam specificity decreased from 0.835 to 0.774 and remained non-inferior. Interobserver agreement for Lung-RADS increased by 40% (κ: 0.375 to 0.525). Size-dependent analysis demonstrated the largest gain for small nodules (4–6 mm), with LROC AUC improving by 54.9% (0.388 to 0.602). After removal of extreme outliers, mean interpretation time was reduced by 25.1 seconds per case (18%), not accounting for reporting time savings.
Conclusion
AI-assisted lung nodule detection significantly and consistently improved radiologists’ diagnostic accuracy, interobserver consistency, and interpretation efficiency on chest CT. The greatest performance gains were observed for small nodules, which are particularly relevant for early lung cancer detection.
Disclaimers:
Disclaimers: Chest Suite comprises multiple applications including Veye Lung Nodules, Veye Reporting, DeepHealth Lung AI, DeepHealth Viewer and HealthCCSng. Veye Lung Nodules and Veye Reporting are manufactured by Aidence B.V. and distributed by DeepHealth, Inc. Neither Veye Lung Nodules nor Veye Reporting are FDA cleared for distribution in the US. DeepHealth Lung AI is FDA 510(k) pending. DeepHealth Viewer is manufactured by eRAD, Inc. and distributed by DeepHealth, Inc. in the US. HealthCCSng is manufactured by Nanox AI, Ltd. and distributed by DeepHealth, Inc. DeepHealth Lung AI is 510(k) pending. Any claims made about Chest Suite may reference claims associated with its individual components. Not all products and functionalities are commercially available in all countries
Any claims made about Chest Suite may reference claims associated with its individual components.
Not all products and functionalities are commercially available in all countries. For clearance and commercial availability in your geography of functionalities listed and compatibility with other systems, please contact your account manager.