一种基于实例分割的舌体分割方法  

A Tongue Segmentation Method Based on Instance Segmentation

作　　者：谭建聪 肖晓霞 邹北骥[2] TAN Jiancong;XIAO Xiaoxia;ZOU Beiji(School of Informatics,Hunan University of Chinese Medicine,Changsha 410208,Hunan,China)

机构地区：[1]湖南中医药大学信息科学与工程学院,湖南省长沙市410208 [2]中南大学计算机学院,湖南省长沙市410083

出　　处：《中国卫生信息管理杂志》2023年第3期459-464,共6页Chinese Journal of Health Informatics and Management

摘　　要：目的使用实例分割算法提高舌体分割的准确性,并支持智能舌诊。方法将采集到的962张分辨率为5568×3712的舌面图像,预裁剪为1400×1400,图像中保留舌体、唇部和少部分皮肤;然后采用数据标注工具Labelme对舌体进行标注,并使用BlendMask算法分割舌体。结果BlendMask算法的定位精度为99.77%,而分水岭算法、GrabCut算法和Mask R-CNN算法的定位精度分别为45.20%、70.52%和93.37%。BlendMask算法的定位精度与上述3种算法相比分别提高了54.57%、29.25%和6.40%。结论BlendMask算法可以准确分割舌体,支持智能舌诊,为中医智能化舌诊提供参考。Objective To improve the accuracy in segmenting tongues and support intelligent tongue diagnosis by using instance segmentation algorithm.Methods The collected 962 tongue images with a resolution of 5568×3712 were pre-cropped to 1400×1400.The tongues,lips and a small part of the skin were retained in the images.The images were then analyzed by the data annotation tool Labelme.BlendMask algorithm were used to segment the tongues.Result Localization accuracy of BlendMask is 99.77%,while the accuracy of Watershed,GrabCut and Mask R-CNN algorithm are 45.20%,70.52%and 93.37%,respectively.The accuracy of BlendMask algorithm improved by 54.57%,29.25%and 6.40%compared with that of the above three ones,separately.Conclusion BlendMask algorithm can accurately segment the tongue,support intelligent tongue diagnosis,and provide a reference for intelligent tongue diagnosis in Traditional Chinese Medicine.

关 键 词：舌体分割 GRABCUT MASK R-CNN BlendMask 

分 类 号：R-058[医药卫生] R319