Abstract: The segmentation of laryngeal organs is a prerequisite for laryngoscope image analysis and computer-aided diagnosis and treatment. To accurately segment organ parts, firstly, a dilated residual (DR) convolution module is proposed, which uses a variety of dilated convolutions to extract the features under different receptive fields of the image, combined with the residual strategy to improve the feature diversity and speed up the network training speed. Secondly, the DR module is combined with the feature pyramid to fuse multi-scale features and supplement the low-level features of organs, so that the network can adapt to shape variations of organs. Finally, on this basis, the throat organ segmentation network DRP-Mask is designed. The experimental results on the 8 000 laryngoscope image dataset show that, compared with the other 5 semantic segmentation networks, the mIoU of DRP-Mask is improved by 2% to 4%, and the mAP of the benchmark is improved by 1.6%. At the same time, as accurate positioning, complete segmentation is performed, and the segmentation results are more in line with the doctor's labeling results.