Depth-Awareness Shared Self-Supervised Bronchial Orifice Segmentation for Center Detection in Vision-Based Robotic Bronchoscopy

Abstract

Limited by the complex porous structure of the bronchi, supervised segmentation of the bronchial orifice is difficult to perform. Since there is no gold standard for definition of the edges of bronchial orifices, manual labeling can lead to inconsistencies in calibration standards. This paper presents a self-supervised bronchial orifice segmentation method based on depth information and synthetic data. The method uses synthetic data and synthetic depth maps to train CNN and can be directly applied to real image data. We also propose an Attention Guidance Module that uses depth information to guide the neural network attention distribution. We manually segmented the bronchial orifice region as ground truth and counted the offset of the center point between the segmentation result and ground truth. The results show that the mIoU of the segmentation result is 79.52, the dice coefficient of the result is 0.86 which is better than 0.77 of the previous method, and the pixel deviation of the center point in a 128 × 128 image is 3.92 which is better than 4.35 of the previous method.