用于咽喉器官分割的空洞残差金字塔算法

Dilated Residual Pyramid Algorithm for Throat Organ Segmentation

对咽喉器官分割是喉镜图像分析以及计算机辅助诊疗的先决条件.为准确地分割器官部位,提出一种用于咽喉器官分割的空洞残差金字塔算法.首先提出空洞残差(dilated residual,DR)模块,使用多种空洞卷积提取图像不同感受野下的特征,结合残差策略提升特征多样性并加快网络训练速度;然后将DR模块与特征金字塔结合,融合多尺度特征并补充器官浅层特征,使得网络适应器官的多种形态;最后设计咽喉器官分割网络——DRP-Mask.在8000幅喉镜图像数据集上的实验结果表明,与其他5种语义分割网络相比,DRP-Mask的平均交并比提升2%~4%,比基准网络平均精度提升1.6%,实现对器官准确定位的同时也对其进行完整的分割,分割结果更贴合医生标注结果.

The segmentation of laryngeal organs is a prerequisite for laryngoscope image analysis and com-puter-aided diagnosis and treatment.To accurately segment organ parts,firstly,a dilated residual(DR)con-volution module is proposed,which uses a variety of dilated convolutions to extract the features under dif-ferent 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 8000 laryngoscope image dataset show that,compared with the other 5 se-mantic 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 per-formed,and the segmentation results are more in line with the doctor’s labeling results.