Fig. 4. Results of in vitro experiments.

a In vitro experimental scenario under breathing simulation, for which the detailed design and motion analysis are illustrated in Supplementary Fig. 10. b The airway models corresponding to the two bronchial phantoms. c Learning process of the novice doctors. Path 23 in Phantom 1 was selected for bronchoscopy. Two novice doctors (a medical intern and an attending doctor) were invited to learn teleoperated bronchoscopy with and without the AI co-pilot from demonstrations. The medical intern performed three bronchoscopy trials without the AI co-pilot. Then, the expert (without the AI co-pilot) and the attending doctor (with the AI co-pilot) separately performed bronchoscopy. Image errors were recorded during procedures. The numbers of recorded frames are n = 3928, 5162, 3693, 3503 and 3544 for five independent experiments from Trial 1 to AI, respectively. Bar plots show mean ± 95% CI of image errors. d Comparison of results in Phantom 1. Paths 19 and 55 were selected for bronchoscopy, covering both sides of Phantom 1. The expert without the AI co-pilot and the attending doctor with the AI co-pilot performed bronchoscopy separately. The numbers of recorded frames are n = 6014 for Expert and n = 6094 for AI. e Comparison of results in Phantom 2. Paths 25 and 66 were selected for bronchoscopy. The numbers of recorded frames are n = 6180 for Expert and n = 5811 for AI. The results reveal that the novice doctor with the AI co-pilot could achieve and maintain a smaller image error than the expert during the bronchoscopy procedure. Line plots with error bands depict time-varying mean image errors ± 95% CI within a time window of 20 frames during each procedure. Bar plots depict the mean image error and the 95% confidence interval for all frames during the two bronchoscopy procedures in each phantom, i.e., the mean ± 95% CI image error in Phantom 1, containing Path 19 and Path 55, and in Phantom 2, containing Path 25 and Path 66.