In tongue diagnosis, colour information of tongue body has kept valuable information regarding the state of disease and its correlation with the internal organs. Qualitatively, practitioners may have difficulty in their judgement due to the instable lighting condition and naked eye’s ability to capture the exact colour distribution on the tongue especially the tongue with multicolour substance. To overcome this ambiguity, this paper presents a two-stage tongue’s multicolour classification based on a support vector machine (SVM) whose support vectors are reduced by our proposed
Tongue diagnosis is said to be the most active research in the advancement of complementary medicine compared to other diagnosis fields such as pulse and abdominal palpation [
Tongue colour samples: (a) light red, (b) red, and (c) deep red.
In 2010 and 2013, a tongue colour gamut descriptor has been proposed by several researchers using one class SVM [ Impossible features or values have been inputted as training examples. Several values have not been inputted during training (missing values). Redundant or irrelevant data have been included as training examples.
Similar researches in [
Therefore, creating a classifier by taking into account the most informative or meaningful features to improve the response time and generalization ability is crucial. This paper presents a two-stage classification system for tongue colour diagnosis aided with the devised clustering identifiers and proposed colour range that can improve the classification accuracy and classifier’s response time. Data selection method in our case (
Feature extraction or data transformation is a process of transforming a raw feature data into quantitative data structure or patterns for training accessibility [
To establish the standard of tongue diagnosis, the standardization fundamental in defining the most suitable informative features for any significant diseases is also essential. There are several reported works using hyperspectral images to attain meaningful tongue features for further analysis such as coating colour and sublingual veins [
Because there is abundance of tongue features or pathological details that have been accumulated via thousands of images progressively, the innovation in decision support system and intelligent image analysis has evolved for accurate and fast classification and diagnosis. In general, there are three learning algorithms which have been implemented using machine learning such as supervised, unsupervised, and semisupervised algorithms. By utilizing these algorithms, an accurate classification system to classify the most informative features with high generalization ability is desired. There are several works reported in feature disease classifications aiming to predict the mapping relationships between tongue features and diseases [
In [
There are a total of 300 tongue images after coating eliminations have been accumulated during clinical practice in the Oriental Medicine Research Centre, Kitasato University in Japan. All tongue images in this proposed research were taken by tongue image analyzing system (TIAS) that was invented by the Chiba University, Japan. TIAS is a closed box acquisition system that is used to capture the tongue image under stable condition in terms of illumination condition and tongue’s position. There are several components implemented in TIAS such as
halogen lamps as illuminators with high colour temperature to acquire adequate tongue colour information, integrating sphere which is a hollow cavity shaped with coated interior to produce equal distribution of light rays on a tongue, 1280 × 1024 pixels high-speed charged couple digital (CCD) camera to capture high-resolution 24 bit RGB (redness, greenness, and blueness) tongue images, 24-colour chart for colour correction purposes.
All the images implemented underwent colour correction procedure to maintain high colour reproducibility outcomes. Around 300 tongue images after coating elimination in [
This section describes our full procedures on our proposal of two-stage classification system by implementing an SVM classifier aided with clustering identifiers and red colour range. The preclassification starts with the most contributing colour and area analysis using
(a) Raw image, (b) image after segmentation, and (c) image after coating removal.
The outcome of the clustering procedure is the cluster image that is divided into background, red/light red, deep red, and region with transitional pixel clusters. The two most informative clusters or clustering identifiers are red/light red and deep red clusters. These two clustering identifiers will be used as training examples in SVM to preclassify the deep red tongue or red/light tongue. In the second-stage classification or final classification, the red/light red tongue images are further classified into red tongue and light red tongue based on the red colour range derived from our databases. The detail on how we choose the most informative clusters will be discussed in the next section.
The most important step in our proposed classification technique is a sampling strategy of bag-of-features reduction (or also known as feature selection) by applying clustering algorithm to define the most contributing colour and area on tongue images. Feature selection is a procedure of selecting an informative subset of nonredundant features among the original or transformed ones usually for efficiency purposes [
In east-Asian medicine perspective, the tongue accumulates several valuable information regarding the properties, location, and development and prognosis of a disease [
The
We have tested several possible number of clusters and observed the outcomes with the practitioner’s recommendation. By using
Based on our observation during the process of determining the number of cluster,
Maximum colour distances from black pixel or chromatic intensity are determined based on (
Moreover, maximum pixels’ coverage area formula can be deduced as in
(a) Deep red cluster identified by maximum pixels’ coverage area identifier and (b) and (c) red/light red clusters identified by maximum colour distance identifier.
During the accumulation of average colour of every clusters in
This section describes the theory and fundamentals of our proposed SVM model aided with clustering identifiers to reduce the number of support vectors during training and testing procedures that will lead to a fast classification system with high accuracy. A support vector machine (SVM) is a supervised machine learning method that is defined by a separating hyper plane which can be used for classification of images. Given a set of labelled training data, the algorithm outputs an optimal hyper plane which predicts the new example to fall on which side of the gap. In the image processing concept, this training algorithm of SVM builds a model of mapping pixels and assigns them into one category or the other divided by a discriminative hyper plane. A good separation is achieved by the hyper plane that has the largest margin which describes the distance to the nearest training data point (support vectors) of any class. The larger the margin, the lower the generalization error of the classifier will be. The reason why SVM insists on finding the maximum margin hyper planes is that this optimization offers the best generalization ability. In other words, it compromises better classification performance (e.g., accuracy) on the training data of the future data. In addition, SVM can also perform a nonlinear classification using the kernel method by mapping their inputs into high-dimensional feature spaces implicitly. Besides, SVM is said to have high generalization ability for classification problem compared to other machine learning algorithms even though the input space is very high [
Currently, there are two types of approaches for multiclass SVM. The first method is by considering all the data using one optimization formula and the other one is by combining several binary classifiers together and finally opt for the one with the best optimization. According to this paper [
Theoretically, the general binary (two classes) classification using SVM can be visualized in Figure
Methodically, if we apply this (
SVM concept of classification by constructed hyper plane.
The down sampling procedures or sample selection method for training purposes before classification is essential to reduce classifier’s burden and complexity. Since the tongue colour is very narrow [
Flowchart of our proposed first-stage classification.
After series of training procedures, we have discovered that maximum colour distance identifier has the best ability to classify deep red tongue and maximum pixels’ coverage area identifier has the best ability to classify red/light red tongue based on the loss function formula calculated using labelled tongue colour image databases. By using these proposed clustering identifiers, the numbers of overlapping pixels and misclassified points or outliers between the boundaries have been greatly reduced; hence, the number of support vectors is also reduced. This reduction of overlapping pixels promotes distance maximization (margin maximization) of separating hyper plane for better generalization ability. The implementation of our proposed tongue identifier as training examples is significant as it also lead to minimization of outliers that can prevent over fitting during the classification process. Moreover, SVM treats all training points equally; hence, both the noisy points and outliers will have negative impacts on the accurate classification [
This section describes the second stage of classification where red colour range is used as a final classifier between red/light red groups of tongue after first classification is done. If the new example of tongue is classified as deep red tongue after the first stage of classification, then it will not be classified further using this red colour range. Nevertheless, if the new example is classified as red/light red tongue in the first stage of classification, then, it has to be classified further using red colour range for final verification. The measurement of red colour range is done during the clustering procedures where we have accumulated hundreds of average red and light red tongue colour clusters which are labelled clinically beforehand by the practitioners’ naked eye as red and light red tongues. By naked eye, red and light red tongues look very similar because the colour range of chromatic value (
Red colour range for red and light red tongues.
Tongue colour |
| ||
---|---|---|---|
|
|
| |
Red |
|
|
|
Light red |
|
|
|
After the first-stage classification, red/light red cluster will be tested using red colour range defined in Table
The detailed procedures of our proposed tongue colour diagnosis system aided with the proposed identifiers and red colour range are illustrated in Figure
The outline of proposed computerized tongue colour diagnosis system.
This section discusses the experimental results and the performances of our proposed two-stage classification system using SVM model with clustering identifiers and red colour range. We have involved over 300 images and 600 features of our clustering identifiers as training examples in SVM to diagnose the red, light red, and deep red tongue colours automatically. These tongue images were taken by tongue image analyzing system (TIAS) on hundreds of outpatients in the Oriental Medicine Research Centre, Kitasato University in Japan, and each of the tongue colour was validated and labelled beforehand by several medical practitioners. The SVM algorithm is run in MATLAB environment with Intel® Core™ i7-3820CPU @3.60GHz. As a comparison, we have implemented raw images after segmentation and coating removal as training examples without clustering identifiers. Several types of kernels such as linear, Gaussian radial basis function (RBF), quadratic, and polynomial kernels have been tested. However, without clustering identifier, the accuracy rate measured was not convincing. The rows labelled as “Conventional SVM” in Table
Comparison of average classification accuracy and execution time of several algorithms using same database specifications.
Method | Technique/kernel | Accuracy (%) | Execution time (s) |
---|---|---|---|
Conventional SVM (only SVM) [ |
RBF | 50 | 219 |
Polynomial | 50 | 187 | |
Linear | 57 | 249 | |
Quadratic | 74 | 187 | |
|
|||
Proposed SVM with clustering identifiers (SVM + |
RBF | 63 | 166 |
Polynomial | 50 | 172 | |
Linear | 89 | 149 | |
Quadratic | 50 | 151 | |
|
|||
Neural network [ |
Conventional | 70.6 | 652 |
Methodically, we have divided the experiment into two stages, SVM classifier and classification based on
Based on the measurement result by SVM on 300 images, we have deduced that linear kernel is the best kernel that can successfully separate the light red/red and deep red tongues using combination of maximum colour distance and pixels’ coverage area identifiers. By implementing the result of
However, a low classification accuracy have been observed in light red and red tongue classification by using maximum colour distance identifier in SVM because these two colours look very similar to the naked eye. Nevertheless, during the colour analysis measurement, the difference in light red and red tongues was observed via each of the
Comparison of red colour range’s performance in classification.
|
Accuracy |
---|---|
Only chromatic attribute range ( |
63% |
Both chromatic and luminance attribute range
|
95% |
As can be seen in Table
In this work, we have proposed a two-stage classification method to diagnose three tongue colours: red, light red, and deep red. The proposed automatic colour diagnosis system is very useful for early detection of imbalances condition inside the body. According to traditional medicine perspectives, light red tongue is considered normal; red tongue is always associated to excess of heat, dehydration, hemoconcentration, or irritability; and deep red tongue is associated with blood stagnation, coldness, and so forth. The first-stage classifier is mainly based on SVM aided by
The authors declare that there is no conflict of interest regarding the publication of this paper.
This study was supported by the Japan Society for the Promotion of Science KAKENHI Grant no. 26860420 from the Ministry of Education, Culture, Sports, Science and Technology, Japan; Kitasato University Research Grant for Young Researchers; and FRGS research grant vote project (4F504) from the Ministry of Higher Education, Malaysia.