Diagnostic Method of Diabetes Based on Support Vector Machine and Tongue Images

Abstract

Objective. The purpose of this research is to develop a diagnostic method of diabetes based on standardized tongue image using support vector machine (SVM). Methods. Tongue images of 296 diabetic subjects and 531 nondiabetic subjects were collected by the TDA-1 digital tongue instrument. Tongue body and tongue coating were separated by the division-merging method and chrominance-threshold method. With extracted color and texture features of the tongue image as input variables, the diagnostic model of diabetes with SVM was trained. After optimizing the combination of SVM kernel parameters and input variables, the influences of the combinations on the model were analyzed. Results. After normalizing parameters of tongue images, the accuracy rate of diabetes predication was increased from 77.83% to 78.77%. The accuracy rate and area under curve (AUC) were not reduced after reducing the dimensions of tongue features with principal component analysis (PCA), while substantially saving the training time. During the training for selecting SVM parameters by genetic algorithm (GA), the accuracy rate of cross-validation was grown from 72% or so to 83.06%. Finally, we compare with several state-of-the-art algorithms, and experimental results show that our algorithm has the best predictive accuracy. Conclusions. The diagnostic method of diabetes on the basis of tongue images in Traditional Chinese Medicine (TCM) is of great value, indicating the feasibility of digitalized tongue diagnosis.

Figure 1

TDA-1 tongue instrument.

Figure 2

Tongue diagnosis analysis system.

Figure 3

The segmentation of tongue body and coating.

Figure 4

Map of classification model.

Figure 5

Flow chart of GA-SVM.

Figure 6

Result of PCA.

Figure 7

Fitness curve of SVM parameters optimized by GA.

Figure 8

The ROC curve (raw data).

Figure 9

The ROC curve (normalized data).

Figure 10

The ROC curve (normalized data with PCA).