Bronchoscopic Localization Method Guided by Prior Knowledge of Branch Structure

The complexity of the bronchial structure creates difficulty in delivering the bronchoscope to the focus during bronchoscopy. Therefore, designing a bronchoscope navigation system is necessary to help doctors eliminate the difficulty in localization. In this paper, a new bronchoscope localization method is proposed. First, the network to extract the descriptor of the image is trained by contrastive learning, and the sub-region localization of the bronchoscope is realized by descriptor retrieval. Then, the depth information of the image is estimated using a network, and the location of the bronchoscope is obtained on the basis of the depth information. The public datasets verify that the sub-region localization accuracy is 95.3% on average, and the average localization accuracy of poses is 1.79 mm (position) and 3.6° (rotation), and the speed of localization is over 24 fps. The proposed method is only based on images to locate and does not need to manually initialize the pose. The localization performance is better than the advanced methods in recent years.