A45-10 Artificial Intelligence-Assisted Real-Time Monitoring System for Electronic Bronchoscopy: A Randomized Controlled Trial

Abstract Rationale Bronchoscopy examination is an important method for diagnosing respiratory diseases, but differences in operator skills may lead to inspection blind spots and incomplete coverage of anatomical sites. The current processes lack real-time assistance functions and cannot effectively monitor blind spots, record duration, or retain standard images. Therefore, Artificial intelligence (AI) assisted systems are expected to improve inspection quality and standardization. This study aims to develop and validate an AI-assisted real-time assistance system for bronchoscopy to monitor inspection blind spots, record inspection duration, and retain standard inspection images, which improves the coverage rate of bronchial anatomical sites. Methods This system includes two deep convolutional neural network (DCNN) models for filtering abnormal images and classifying qualified images into 31 standard bronchial branches. We retrospectively collected 2980 electronic bronchoscopy records from three hospitals in China, all performed using the Olympus CV-290 endoscope processor. To validate the system’s generalization ability, we used 158 cases as an external video test set for external validation. Subsequently, a prospective, multicenter, randomized controlled trial was conducted, 204 subjects from three hospital in different geographic regions of China (Shanghai, Guangzhou, and Xi’an) were randomly allocated to the experimental group and the control group according to inclusion and exclusion criteria (102 cases in the experimental group, 102 cases in the control group). The primary outcome was the comparison of bronchial anatomical site coverage rates with or without the assistance system; the secondary outcomes included average inspection duration and the sensitivity and specificity of recognition for anatomical sites. Results The average bronchial anatomical site coverage rate in the assisted group was significantly higher than in the control group (92.09% vs 82.70%, P 0.0001). The average inspection time in the assisted group was significantly longer than in the control group (221.1s vs 176.8s, P 0.0001). In 204 inspection videos, the system’s average recognition accuracy for bronchial anatomical sites was 95.79%; the sensitivity range for 31 sites was 86.83%∼100.00% (average 95.797%). Conclusions Our system can significantly improve the anatomical coverage rate of bronchoscopy examinations, with satisfactory recognition performance and clinical application potential. This abstract is funded by: None