Real-world Performance of Computer-aided Pulmonary Nodule Detection for Lung Cancer.

Purpose

To evaluate the real-world sensitivity of an artificial intelligence-based pulmonary vessel suppression and computer-aided detection system (ClearRead VIS/CADe) for lung cancer detection and to characterize nodule features associated with detection and nondetection.

Materials And Methods

This retrospective study included patients diagnosed with lung cancer within a lung cancer screening program in the authors' health care network after deployment of ClearRead VIS/CADe (August 2023 to June 2025). A fellowship-trained cardiothoracic radiologist reviewed the diagnosis, CT examination, and the most recent prior chest CT examination for each patient. Malignant nodules were characterized by size, type, location, and relationship to major vessels. For each examination, whether the malignant nodule was identified in the radiology report, annotated by CADe, and visible on vessel-suppressed images was recorded. Sensitivity was assessed overall and by nodule characteristics. Multivariable logistic regression was performed for nodules <30 mm to identify factors associated with CADe annotation.

Results

A total of 129 patients were included. Overall sensitivity of CADe annotation on diagnosis CT examinations was 72% (93/129), whereas vessel-suppressed images demonstrated higher sensitivity (92%, 119/129). CADe sensitivity varied significantly by nodule size (P<0.001), was highest for solid nodules (80%) and lower for part-solid (75%), ground-glass (38%), and cystic nodules (13%) (P<0.001). Vessel-suppressed images demonstrated high sensitivity across nodule sizes and morphology. After exclusion of masses, subpleural and paramediastinal locations remained independently associated with reduced sensitivity (P<0.05).

Conclusion

Vessel-suppressed images demonstrated high sensitivity for malignant pulmonary nodules, whereas CADe performance varied by nodule type and location.