Nasopharyngeal Organ Segmentation Algorithm Based on Dilated Convolution Feature Pyramid

In recent years, nasopharyngeal disease has been a common disease in clinical diagnosis whose incidence is increasing. As the most direct and effective means to observe the visceral mucosa of the cavity, Electronic laryngoscope, playing a role in diagnosis and minimally invasive diagnosis and treatment in the clinic, has become an important tool in otolaryngology head and neck surgeons. It is very important for clinical medicine to accurately segment the organs in the image. The following factors make organ segmentation more difficult, such as the complex structure and background of organs in nasopharynx and larynx; the unclear edge of organs and the little color differences between organs and background in laryngoscope image. In order to accurately segment organs and distinguish different instances of the same category, this paper proposes a nasopharyngeal organ segmentation model named Dilated Pyramid-Mask (DP-Mask). The model is based on the dilated convolution feature pyramid network. In my paper, Mask R-CNN is introduced into the organ instance segmentation of electronic laryngoscope image. What’s more, in order to improve the segmentation accuracy, dilated convolution is designed in each layer of FPN to get the association of context feature information and get the segmentation of multiple organ instances in laryngoscope image. The experimental results show that the detection accuracy of DP-Mask model can reach 86.3%, Dice coefficient and mIOU can reach 0.81 and 0.88 respectively, which has high accuracy and robustness. Compared with popular U-Net and deep lab V3 algorithms, the proposed DP-Mask model improves mIOU by 5.4 and 4% respectively.