Upper Airway Segmentation Based on the Attention Mechanism of Weak Feature Regions

Abstract

The segmentation and modeling of the upper airway play an important role in the analysis and diagnosis of obstructive sleep apnea. In this paper, a method based on deep learning is proposed to automatically segment the upper airway on CBCT data. The main process is to obtain a coarse segmentation result map through 3D U-Net. Then low-confidence voxels in the coarse segmentation results as weak feature regions are extracted. Secondly, the boundary attention map is constructed through the Gaussian kernel function based on weak feature regions. Then, the boundary attention mechanism is used to strengthen the learning ability of weak feature regions. Finally, a high-precision 3D model is constructed on the predicted segmentation results through the marching cube algorithm. A large number of experiments and comparisons verify that our method achieves better performance on the manually annotated dataset.