An Artificial Intelligence-Assisted Bronchial Automatic Identification Using Deep Learning Technique

Bronchial intubation, facilitated by fiberoptic bronchoscopy, is a crucial aspect of lung isolation technology. However, intelligent examination of bronchoscope still needs to be explored to improve uneven quality of examination and reduce high training cost. Consequently, the development of an artificial intelligence (AI)-powered computer-assisted intubation system could significantly enhance the anesthesiologists’ accuracy in identifying the left and right main bronchus, as well as the bronchial lumen (Left, Right). A total of 1800 bronchoscopy images were collected from Shanghai Changzheng Hospital to develop an AI-assisted system based on convolutional neural network (CNN) model. A clinical evaluation was designed to assess the accuracy of bronchoscopy doctors’ identification of lumen with and without AI-assisted system. For the independent test dataset, the optimal accuracy of models ranged between 81.5% and 96.1%. In the YOLO series models, the accuracy of YOLOv3, YOLOv5 and YOLOv7 were 93.4%, 95.1% and 96.1% respectively. In the Faster RCNN model, the accuracy of VGG-16 and ResNet-50 were 81.5% and 82.6% respectively, the Frames Per Second (FPS) were 55, 64, 96, 143, 67 respectively. Furthermore, the clinical evaluation was conducted by 9 anesthetists who were divided into three groups according to their work experience. The identification results indicated that the YOLOv5 model showed better performance (accuracy of 100%) than almost all human experts (average accuracy of 81.25%, 92.5%, and 90.0%). Consequently, the YOLOv5 model was selected as the final AI-assisted system. In conclusion, our AI-assisted system has significantly improved the identification accuracy of the anatomical structure of bronchial lumen, providing a valuable reference for bronchoscope-guided endobronchial intubation.