基于Mask R-CNN的舌图像分割研究＊

目的 舌图像分割是舌诊客观化的关键之一，易受舌体附近嘴唇和皮肤等带来的影响，而增加分割的难度。针对该问题，为确保舌图像分割的准确性，本研究提出一种基于卷积神经网络Mask R-CNN的舌图像分割方法。方法 首先用标注工具labelme对舌图像进行标注，然后进行Mask R-CNN舌图像分割模型的训练和舌图像分割测试。结果 采用该方法进行舌图像分割，获得的舌体边缘比较准确，并且四个定量评价指标均像素准确度、平均准确度、均交并比、频权交并比均高于84.6%。结论 本研究取得了较好的舌体分割效果，能够改善舌体周围的嘴唇和皮肤颜色与舌体颜色接近导致舌体分割轮廓不准确的问题，为舌图像分割提供了一种新的思路与方法，对舌诊客观化具有一定实用价值和借鉴意义。

Research on Tongue Image Segmentation Based on Mask R-CNN

Tongue image segmentation is one of the keys to the objectification of tongue diagnosis. When the color of the tongue is close to the lips or the skin, it will have a greater influence on the segmentation of tongue. This increases the difficulty of segmentation. Therefore, this paper proposed a tongue image segmentation method based on Mask R-CNN of convolutional neural network. In order to improve the accuracy of tongue image segmentation, firstly, the tongue image is labeled with the labeling tool labelme, and then the tongue image segmentation model of Mask R-CNN is trained and the test of tongue image segmentation is carried out. The tongue segmentation edge obtained by the tongue image segmentation method is more accurate. And the four quantitative evaluation indexes of the method are higher than 84.6%, which are Mean Pixel Accuracy, Mean Accuracy, Mean Intersection over Union and Mean Frequency Weighted Intersection over Union. This method achieves better tongue segmentation effect. It can solve the problem of inaccurate segmentation of the tongue image caused by the color of the lips and skin being closer to that of the tongue. This research provides a new idea and method for tongue image segmentation, which has certain practical value and reference significance for the objectification of tongue diagnosis.