| Objective: In the context of the great development of the intelligent TCM diagnosis.This research aims to build an automatic and high-precision tongue segmentation model,and conduct tongue segmentation research by using a variety of image recognition methods,including deep learning technology.To further verify the application value of the tongue segmentation model,this research build an analysis network based on the segmented tongue image,thereby exploring the clinical research methods and application value of tongue image based on deep learning technology by mean blood pressure prediction experimentMethods: 1.Research on tongue segmentation: Tongue segmentation experiment: The data were collected from the physical examination center of the Affiliated Hospital of Chengdu University of TCM.TFDA-1 tongue-face diagnosis apparatus was used to collect the tongue image of the research objects who met the inclusion criteria.The original data collected were labeled by Labelme for the training of deep learning model,and the data were randomly allocated according to the ratios of 80%(training set)and 20%(test set).Two kinds of deep learning models,the UNet model and the Deeplab V3 model,and two traditional image recognition methods,the active contour model(Snake)and color decomposition and threshold model(CDT),were applied to segment the tongue body.For the UNet model and Deeplab V3 model,the models were trained by the data of the training set,and then the data of the test set was used to carry out segmentation test;the Snake and CDT were directly used combined with the data of the test set to conduct the segmentation test.Finally,the segmented results of the four models were evaluated by MIo U and PA.2.Application research on intelligence of tongue image: Blood pressure prediction experiment: The data were collected from the department of endocrinology of the Affiliated Hospital of Chengdu University of TCM.The tongue images of the samples which met the inclusion criteria were collected,and the corresponding blood pressure data were extracted from the inpatient records system.After the collection,the data were randomly allocated according to 80%(training set)and 20%(test set).For the tongue images obtained,the Unet model was used to segment the tongue region,and the Res Net was used to extract the vector with images feature of 1*2048 extracted.Then the blood pressure data was augmented to the data vector of 1*2048,and fused with the above vector by using 1*1 convolution to get the fused data.The data of blood pressure(systolic/diastolic pressure)and tongue image data on the day of image shooting and the fused data of the two were used as three independent variables.Four machine learning algorithms,including GBDT,SVR,Adaboost,and RF,were used to carry out regression modeling on the mean value of blood pressures(systolic/diastolic blood pressure)in the subsequent six days.Finally,the data of the test set were used to test each model,and the results were evaluated by the Mean Absolute Error(MAE).Results:1.Research on tongue segmentation:Up to October 2019,2,048 qualified samples were collected and divided into 1,638 in the training set and 410 in the test set.The experiment results of the four tongue segmentation models were evaluated by MIo U and PA,and the details were as follows:(1)The MIo U and PA of the UNet model experiment results were respectively 91.05% and 93.31%.(2)The MIo U and PA of the Deep Lab V3 model experiment results were respectively 88.35% and 84.16%.(3)The MIo U and PA of the Snake model experiment results were respectively 76.64% and 76.26%.(4)The MIo U and PA of the CDT model experiment results were respectively 69.21% and 65.42%.2.Application research on intelligence of tongue image:Up to October 2019,325 qualified samples were collected and divided into 260 in the training set and 65 in the test set.The Mean Absolute Error(MAE)was used to uniformly evaluate the prediction results.The details were as follows:(1)The systolic blood pressure on the day of shooting was used to predict the mean systolic blood pressure of the subsequent six days: SVR was 7.31 mm Hg,RF was 6.54 mm Hg,Adaboost was 5.91 mm Hg and GBRT was 6.25 mm Hg;the mean systolic blood pressure of the subsequent six days was predicted in the same way: SVR was 7.10 mm Hg,RF was 8.61 mm Hg,Adaboost was 4.43 mm Hg,GBRT was 6.17 mm Hg.(2)The tongue image features were used to predict the mean systolic blood pressure of the subsequent six days: SVR was 15.41 mm Hg,RF was 10.39 mm Hg,Adaboost was 10.22 mm Hg and GBRT was 9.61 mm Hg;the mean systolic blood pressure of the subsequent six days was predicted in the same way: SVR was 18.03 mm Hg,RF was 13.91 mm Hg,Adaboost was 11.87 mm Hg,GBRT was 15.73 mm Hg.(3)The fused data was used to predict the mean systolic blood pressure of the subsequent six days: SVR was 7.06 mm Hg,RF was 5.49 mm Hg,Adaboost was 5.54 mm Hg and GBRT was 9.61 mm Hg;the mean systolic blood pressure of the subsequent six days was predicted in the same way: SVR was 6.32 mm Hg,RF was 8.54 mm Hg,Adaboost was 4.02 mm Hg,GBRT was 5.91 mm Hg.Conclusion:(1)Compared with traditional image recognition methods,deep learning technology can more accurately complete tongue segmentation and be conducive to the automatic recognition of TCM tongue images.(2)The UNet convolution neural network model trained in this research can effectively improve the precision of tongue segmentation,and with this model,full-automatic and high-precision segmentation of the tongue targeted area can be realized.(3)The UNet tongue segmentation model combined with the Res Net network can realize the automatic extraction of the image features.The extracted features combined with machine learning modeling can be used to explore the complex hierarchical mathematical implicit correlation between the tongue images and the clinical data.(4)Compared with the use of the blood pressure data on the day of shooting,the combined use of the blood pressure on the day of shooting can improve the prediction accuracy,indicating that the tongue image made a significant contribution to the improvement of prediction accuracy and the data fusion is an important way to explore the clinical value of TCM tongue images. |
PDF Full Text Request