基于Mask Scoring R-CNN的齿痕舌象识别

目的:提出一种基于Mask Scoring R-CNN和迁移学习的舌象特征识别方法。方法:首先使用CNN提取特征，使用ResNet-101和特征金字塔网络（FPN）的主干网络，可以从低层次和高层次的网络中提取特征，根据不同比例绘制金字塔特征的级别。接着使用区域生成网络将从主干网络中提取的特征生成候选感兴趣区域（ROI）。最后为每个ROI检测并分割齿痕。结果:在232例样本的测试集上进行测试，F1分数为0.95，准确率为0.93，精确率为0.99，召回率为0.914。结论:该方法能够在小样本舌象数据集上有效识别齿痕特征、准确定位齿痕位置、标定齿痕大小、提取齿痕个数，该方法具有良好的有效性、通用性、泛化性，能够为后续齿痕严重程度分析提供依据。同时为疾病预防、移动医疗保健或从生物信息学角度跟踪疾病进展提供客观、方便的计算机辅助舌诊方法。

Tooth-marked tongue recognition using Mask Scoring R-CNN

Abstract: Objective To identify tongue features based on Mask Scoring R-CNN and transfer learning. Methods After the features were extracted by convolutional neural network, the backbone networks of ResNet-101 and feature pyramid network were used to extract features from low-level and high-level networks, and the levels of pyramid features were plotted according to different proportions. Region generation network was then used to generate candidate regions of interest from the features extracted from the backbone network. Finally, tooth marks in each region of interest were detected and segmented. Results The test on the test set with 232 samples showed that F1 score was 0.95, and that the accuracy rate, precision rate and recall rate of the proposed method were 0.93, 0.99 and 0.914, respectively. Conclusion Using the proposed method can effectively identify the features of tooth marks, accurately locate the position of tooth marks, calibrate the size of tooth marks, and extract the number of tooth marks on small-sample tongue image data set. The proposed method which has good effectiveness, generality and generality provides a basis for the severity analysis of tooth marks and serves as an objective and convenient computer-assisted tongue diagnosis method for monitoring disease progression from the perspective of disease prevention, mobile healthcare or bioinformatics.