CN1931086A - Computerized health index analysis system based on tongue picture - Google Patents

Abstract

The present invention provides one kind of computerized health index analysis system based on tongue picture and with high accuracy and wide application. The color and grain characteristics of tongue picture are first extracted by means of digital image processing technology, and the association model between the quantized characteristics and diseases is then established by means of Bayes network. The present invention proposes one kind of quantizing characteristic extraction and diagnosis model designing method for computerized tongue picture diagnosis, and tries to establish the internal relation between the quantized tongue picture characteristics and the diseases. Experiments results show that the said method can diagnose 13 kinds of common diseases effectively.

Description

A kind of Computerized health index analysis system based on picture of the tongue

(1) technical field

The present invention relates to a kind of analytical method, a kind of specifically Computerized health index analysis system based on picture of the tongue.

(2) background technology

Inspection of the tongue is the important method of tcm diagnosis, however since inspection of the tongue do not quantize and lack unified standard, thereby limited its application in clinical diagnosis.Simultaneously, the core of traditional Evolution of Tongue Inspection of TCM is " dialectical ", does not relate to as for getting in touch but between tongue presentation and the disease.But, " card " be the traditional Chinese medical science exclusive abstract conception, it and disease do not have clear and definite corresponding relation, are difficult for being understood by western medicine, thereby have limited " internationalization " of Evolution of Tongue Inspection of TCM greatly.

(3) summary of the invention

The object of the present invention is to provide a kind of scientific and precise, accuracy rate height, the Computerized health index analysis system that is with a wide range of applications based on picture of the tongue.

The object of the present invention is achieved like this: at first utilize digital image processing techniques to extract the quantization characteristic of tongue image, utilize Bayesian network that getting in touch between quantization characteristic and the disease carried out modeling then.

The present invention also has some architectural features like this:

1, the described tongue image quantization characteristic that extracts with digital image processing techniques comprises color and textural characteristics;

2, described color quantizing feature comprises the color average of whole tongue body and the standard deviation of tongue body zone picture element color;

3, described texture quantization characteristic is to use the texture quantization characteristic that extracts tongue image based on the texture operator of gray level co-occurrence matrixes;

4, described utilize Bayesian network between quantization characteristic and the disease get in touch the modeling of carrying out comprise Bayes classifier based on textural characteristics, based on the Bayes classifier of color characteristic with based on the Bayes classifier of associating feature.

The present invention proposes a kind of computerized inspection of the tongue method, this method at first utilizes popular digital image processing techniques to extract the color and the textural characteristics of tongue image; Then, utilize Bayesian network that getting in touch between above-mentioned quantization characteristic and the disease carried out modeling.Experimental result shows that this method can 13 kinds of common diseases of effective diagnosis.

At present, (BayesianNetworks BN) is considered to the important tool of the uncertain representation of knowledge and reasoning among the artificial intelligence study to Bayesian network, is widely applied to the fields such as modeling of complication system, becomes one of hot issue of artificial intelligence study.And Belief Network is based on the application of Bayesian probability success, also be a kind of in artificial intelligence and knowledge engineering widely used representation of knowledge form.Growing field begins to use the Belief Network algorithm at present, and has obtained good effect.

The present invention proposes a kind of method that quantization characteristic extracts and diagnostic cast designs that is used for the diagnosis of computer picture of the tongue, Evolution of Tongue Inspection of TCM does not quantize and the undue personal experience's of dependence problem with solving, and attempts to set up the internal relation between tongue image quantization characteristic and the disease.We with Bayes classifier as the diagnosis decision model, its method " excavation " quantization characteristic by machine learning and the relation between the disease and with the knowledge " curing " that obtains in the network structure and parameter of oneself.In order to verify its effectiveness, the method that we use this chapter is carried out classification diagnosis, wherein healthy people's 70 examples, 13 kinds of commonly encountered diseases 455 examples to 525 patients' tongue image.Experimental result shows, reaches 75.8% based on the overall diagnosis accuracy of the Bayes classifier of associating feature.Wherein, to the diagnosis of healthy people, pancreatitis, hypertension and cerebral infarction not only accurate but also reliable (TPR and PPV all are higher than 75%).The proof that these are dry straight be feasible based on the picture of the tongue diagnostic method of quantization characteristic classification, finally move towards clinical practice for the computer inspection of the tongue simultaneously and open up a new way.

(4) description of drawings

Fig. 1 is simple Belief Network structure;

Fig. 2 is the expansion Belief Network;

Fig. 3 is simple Belief Network for Bayes expands;

Fig. 4 is the many nets of simple Bayes;

Fig. 5 is general Belief Network;

Fig. 6 is a tongue body subregion sketch map;

Fig. 7 is an intestinal obstruction disease typical case tongue image;

Fig. 8 is a cholecystitis disease typical case tongue image;

Fig. 9 is an appendicitis disease typical case tongue image;

Figure 10 is pancreatitis disease typical case tongue image

Figure 11 is the Bayes classifier network structure of T-BNC;

Figure 12 is the Bayes classifier network structure of C-BNC;

Figure 13 is the Bayes classifier network structure of J-BNC.

(5) specific embodiment

Below in conjunction with accompanying drawing detailed process of the present invention and principle are further described:

1, Belief Network

1.1, the definition of Belief Network

Uncertain (uncertainty) is the inherent feature of nearly all medical problem, and uncertainty how to handle various ways just becomes the major issue that the computer-aided medical science system is faced.Present predominant methods is to utilize probabilistic framework with its formulism.Wherein, Bayes (Bayesian) probabilistic model makes this " reliability " can clearly express under its probabilistic framework for human subjective " reliability " formulism (belief) provides a kind of theoretical basis.Bayesian network (perhaps being called " Belief Network ") is used for creating and management probabilistic knowledge storehouse for this theory provides a kind of utility.

We are with one group of variable, { x ₁, K, x _n, represent certain Problem Areas, be a causal probability net of expression about the Bayesian network of this Problem Areas, can with succinct mode express set of variables joint probability distribution (jointprobabilitydistribution, JPD).This expression comprises one group of local condition's probability distribution (localconditionalprobabilitydistribution) and a set condition independent (conditionalindependence, CI) hypothesis.These independently suppose to make us that the Joint Distribution of the overall situation is decomposed into a series of local distribution, and this decomposition is based on the chain rule of probability:

$p(x_1,K,x_n)=\underset{i=1}{\overset{n}{\mathrm{\Π}}}p\left(x_i\right|x_1,K,x_{i-1})---\left(1\right)$

For each variable x _i, we define its " superset closes " is such one group of variable (parentset), ∏ _i { x ₁, K, x _I-1, make x _i{ x ₁, K, x _I-1Condition independence, promptly

p(x _i|x ₁，K，x _i-1)＝p(x _i|∏ _i) (2)

According to above-mentioned definition, Bayesian network can be described as a directed acyclic graph, and (directedacyclicgraph DAG), makes { X ₁, K, x _nIn each variable corresponding diagram in a node, node x _i" father " (father node) corresponding his father gather ∏ _iIn node.With each node x _iThat be associated is conditional probability distribution p (x _i| ∏ _i), and each example (instance) that corresponding superset closes all has a such distribution.Merge with formula (5-1) with (5-2), we are as can be seen for certain Problem Areas { x ₁, K, x _nIn any one Bayesian network (can define a plurality of Bayesian networks in the Problem Areas), all unique decision a joint probability distribution of all variablees, that is:

$p(x_1,K,x_n)=\underset{i=1}{\overset{n}{\mathrm{\Π}}}p\left(x_i\right|{\mathrm{\Π}}_i)---\left(3\right)$

Bayesian network has following advantage for data analysis: the first, because the Bayesian network model is containing the mutual relation between all variablees, so it can be relatively easy to handle the situation that some data item is lost.This seems particularly important for medical diagnostic system, because we always can not obtain patient's complete information.The second, Bayesian network can be used for learning cause effect relation, and this meets human basic thinking habit, helps the understanding to problem.The 3rd, an outstanding feature of Bayesian network is to have cause and effect and two kinds of semantic meaning representations of probability simultaneously, for the application (for example medical diagnosis) that priori (being usually expressed as cause effect relation) and observed data (showing as the form of probability) need be organically combined, Bayesian network is undoubtedly a kind of ideal model.The 4th, " crossing training " that being combined into of Bayesian statistic method and Bayesian network avoided model (over-fitting) problem provide a kind of effectively and important method.At last, experimental results show that the diagnostic accuracy of Bayesian network is insensitive for the numerical precision of its conditional probability, these characteristics have been widened the application prospect of Bayesian network greatly.Because a lot of scholars once criticized the use of Bayesian network, reason is need provide a large amount of numerical value for the conditional probability distribution of node when setting up Bayesian network.But if the rough estimate of numerical value just can be satisfied the demand, these criticize just so not sharp-pointed so.Owing to have above-mentioned these characteristics, Bayesian network is widely used in the various application of machine learning, wherein also comprises medical diagnosis.

Belief Network (BeliefNetwork) is Bayesian network (BayesianNetwork) again, probability net (ProbabilisticNetwork), cause effect relation network (CausalNetwork), or knowledge graph (KnowledgeMap).

Belief Network is a directed graph that meets the following conditions:

◆ the node of network is made up of a stochastic variable set, and each node has been represented a stochastic variable;

◆ it is right that the arc that directed arc is concentrated connects each junction associated.If a directed arc is arranged, represent that then node X is the immediate cause of node Y from node X to node Y;

◆ the node of each father node all has a conditional probability table, represents the influence of its father node to this node.For example, to node X, his father's nodal set is Parents (X), and its conditional probability table is P (X|Parents (X)) so;

◆ the node of each no father node all has a prior probability;

◆ this directed graph does not contain directed loop, thus Belief Network often be considered to a directed acyclic graph (DirectedAcrylicGraph, DAG)

Belief Network provides a simple and powerful method that advises knowledge at the specialist system invading the exterior by network structure.Wherein node is a stochastic variable, has represented a proposition, and arc has been represented the dependence of continuous two propositions.Represent different credibilitys of assigning a topic with probability, the dependence between the proposition is represented with conditional probability.We can be with more formal language description Belief Network: Belief Network is a variables set Y={Y1, Y2 ..., Yn} and a directed graph that defines the model M of condition dependence between each variable of this variables set.We can think that Yi is a discrete variable, and ci state arranged, and Yik represents some states of Yi.If the father node collection of Yi is π i, during by given each combination of father node collection, determine by the conditional probability distribution of Yi to the directed arc of Yi for uncle's nodal set.To variable Y i, when given his father's nodal set, it and every other variable Y1, Y2 ..., Yi-1} be condition independently ^[22]The Belief Network structure is the efficient coding to this basic assumption.According to the rule of joint probability distribution, the value collection yk={y1k of variables set Y ..., the joint probability distribution of ylk} can be expressed as:

$p\left(y_k\right)=\underset{i=1}{\overset{l}{\mathrm{\Π}}}p\left(y_{\mathrm{ik}}\right|{\mathrm{\π}}_{\mathrm{ij}})---\left(4\right)$

Wherein, π ij represents the combinations of states of the π i when the Y value is yk.

1.2, Belief Network commonly used

5 kinds of Bayesian networks are as a rule arranged: simple tree, the simple tree of expansion, Bayes expands simple tree, and Bayes sets and general Bayesian network more.5 kinds of Belief Network are also arranged on this basis, respectively they are introduced below.

1.2.1, simple Belief Network

In conjunction with Fig. 1, simple Belief Network simple in structure only has a father node, do not have other connection again.Though its is simple in structure, its classifying quality is not poor.Compared following advantage with other grader.At first its structure is very simple, as long as it is just passable to provide a father node, and does not obtain structure by study.Secondly categorizing process is very efficiently.But these two characteristics are to be that condition independently draws under the situation between all features of hypothesis.Though like this, if between the feature be not simple Belief Network Classifier effective than the grader of other many complexity under the very relevant condition.

1.2.2, expand simple Belief Network

In conjunction with Fig. 2, expand simple Belief Network Classifier and allow to expand the complete independence between each feature in the simple Belief Network by forming tree construction between the property value.The Belief Network of this structure can be passed through the Chow-liu algorithm.

1.2.3, Bayes expands simple Belief Network

In conjunction with Fig. 3, Bayes expands simple Belief Network Classifier by allowing can to form any graphic structure between the property value, is tree construction no longer only.But this structure is learnt very time-consuming, can adopt minimum to estimate method and CL algorithm.

1.2.4, the many nets of Bayes

In conjunction with Fig. 4, respectively corresponding network when many nets of Bayes grader is got different value with category attribute.The many nets of Bayes can be regarded as a general Bayes and expand simple Belief Network.No matter what value is classification get, Bayes expands Belief Network and keeps the relation between the attribute constant, and the many nets of Bayes can change the relation between the attribute when classification is got different value.Since each localized network links to each other with a classification value, category node can be regarded as the father node of all features in some sense.With simple Belief Network, expand simple Belief Network, Bayes expands simple Belief Network and compares, the many nets of Bayes are more general forms, it gives bigger free space between the attribute.Therefore it is a kind of unconfined Bayes classifier.And it might not expand simple Belief Network complexity than Bayes, because when decomposing by the class value, localized network is simple when not decomposing than class value.When the class value was not decomposed, local area network will be expressed all relations.

1.2.5, general Belief Network

In conjunction with Fig. 5, general Belief Network is another kind of nonrestrictive Belief Network classification tool.Simple Belief Network, expanding simple Belief Network, Bayes, to expand the common feature of simple Belief Network, the many nets of Bayes are father nodes that the class node is counted as all properties node.And general Belief Network is regarded the class node as a common node.General Belief Network is to think under data set to exist a joint probability density to distribute, and the many nets of Bayes are thought have a plurality of probability density distributions under a data set.Adopt any grader to depend on actual data cases.

1.3, several problem

The present invention utilizes the diagnosis decision model of Bayesian network as computerization picture of the tongue diagnostic system, should be noted that following two problems here:

The first, the effectively expressing priori although Bayesian network can make things convenient for (for diagnosis is exactly expertise), the present invention does not utilize this knowledge when the structure diagnostic cast.Therefore, network structure and the conditional probability table as the Bayesian network of picture of the tongue diagnostic cast all is to utilize statistical method to come from pathological data (for picture of the tongue diagnosis of the present invention is exactly quantized feature) middle school acquistion.Why we take this strategy, and its reason has two: one, incompatible each other or inconsistent phenomenon often appears in human assessment to probability, and can show the bias of various ways in their judgement.Same, although brainstrust can be confirmed the important dependence between the variable in certain field, they also can't specify probability distribution compatible with each other for a large amount of variablees.Therefore, for this application that comprises a large amount of variablees of computerization inspection of the tongue, the method probability distribution that " excavation " goes out from sample database of utilizing statistics is usually than specified more reliable by the expert; Its two, the expertise in traditional inspection of the tongue is all about by picture of the tongue " dialectical ", so our priori about getting in touch between picture of the tongue and the disease not also.

Second problem that should be noted that is that the present invention utilizes Bayesian network grader (Bayesiannetworkclassifier, BNC) structure diagnostic cast.Each node in the model except root node is all represented a kind of quantized color or textural characteristics, but these features directly corresponding pathological characters qualitatively in traditional inspection of the tongue.The method of this processing feature is consistent with original intention of the present invention, promptly realizes the quantification and the standardization of traditional Evolution of Tongue Inspection of TCM.

2, quantization characteristic extracts

2.1, be used for the quantization characteristic of computer picture of the tongue diagnosis

Quantification and standardization to traditional Evolution of Tongue Inspection of TCM are central topics of the present invention, and wherein " quantification " is at feature extraction.Yet traditional traditional Chinese medical science all is qualitatively with descriptive about the pathological characters of picture of the tongue, for example " pink tongue ", " thin and whitish fur " etc.Obviously, these descriptive pathological characters are the dialectical service of traditional Chinese medical science picture of the tongue, but they and be not suitable for the computer inspection of the tongue.How to extract that the computer picture of the tongue is diagnosed significant quantization characteristic is the major issue that the present invention at first will solve.Up to the present, extensively being adopted by most similar institutes, also is that more direct method is to manage to find the corresponding quantitative index for each descriptive pathological characters simultaneously, by Traditional Chinese Medicine experts quantized result is assessed then.It is feasible that this method is diagnosed for area of computer aided (Traditional Chinese Medicine experts) picture of the tongue, however it and the object of study that the is not suitable for this paper autonomous picture of the tongue diagnostic system of computer one by one.Because autonomous diagnosis of computer and computer-aided diagnosis have difference in essence: it is different diagnosing main body.And for any one medical diagnostic system, the final service object of feature extraction must be the diagnosis main body---descriptive pathological characters is served human Traditional Chinese Medicine experts, quantization characteristic is served " computerized traditional Chinese medicine expert ".Therefore, quantization characteristic extracts (computer inspection of the tongue) and does not need also should do not assessed by Traditional Chinese Medicine experts (Evolution of Tongue Inspection of TCM), and this also runs counter to the original intention of computer inspection of the tongue simultaneously.

Here it is emphasized that also that serving certain class diagnoses the feature of main body may understand for other class diagnosis main body.For example, descriptive characteristics can't be understood fully for computer; Same, the quantization characteristic of serving the computer inspection of the tongue also can be Traditional Chinese Medicine experts can't explain.This technology path of the present invention just and one of existing methods essential distinction at present.

The static pathological characters of Evolution of Tongue Inspection of TCM is regular governed, they can roughly be divided into two classes: color characteristic (for example, " red ", " purple ", " falling ", " indigo plant " etc.) and textural characteristics (for example, " approach ", " thick ", " neatly ", " stripping off " etc.).In order to keep consistent with technology path of the present invention, we adopt general color and statistics textural characteristics, whether there is corresponding relation and be indifferent to them with the foregoing description feature, and some quantization characteristic is " invisible " or even can't understand for Traditional Chinese Medicine experts, because they are computer inspection of the tongue service fully.These quantization characteristics are to determine by statistical in the training process of Bayesian network for the effectiveness of diagnostic classification.

2.1.1, the quantized color feature

The expression of colouring information is always with respect to specific color space, so the extraction of color characteristic also should be carried out in different color spaces.According to the principle of three primary colours, any color is made up of three components (perhaps being called " parameter ") usually, these parametric descriptions the position (coordinate) of this color in current color space.Color space commonly used in Computer Image Processing has RGB, HSV, CIEYxy, CIELUV and CIELAB.

Rgb color space is based on the additivity color system of the principle of three primary colours.The image of rgb format can directly show on CRT and need not any conversion, so it often is applied to utilizing in the system of CRT display image.Simultaneously have owing to it that linearity can add and be widely used in the every field of Computer Image Processing with characteristics such as light intensity are directly proportional.Yet for human visual perception, RGB is not a linear color space.

CIEXYZ is that it is made up of vision light intensity Y and two chrominance component X, Z by International Commission on Illumination (CIE) definition and device-independent tristimulus values color system.The value of X, Y, Z all is directly proportional with the physics light intensity, and in other words, their eternal right and wrong are minus.Because CIEXYZ and physical spectrum and human vision are all closely related, make it through being often used as the benchmark color space.Simultaneously, in order to satisfy requirement of actual application, International Commission on Illumination has defined two chromaticity coordinate x and y:x=X/ (X+Y+Z), y=Y/ (X+Y+Z) on the basis of CIEXYZ.Like this, any color relation can be expressed as the tlv triple of light intensity and two colourities: xyY (also making CIEYxy).

With the RGB system class seemingly, CIEXYZ since with the linear relationship of physics light intensity, thereby show relatively poor visually-perceptible concordance.In order to address this problem, International Commission on Illumination has proposed two new Munsell system CIELUV and CIELAB on the basis of CIEXYZ, make in these two systems any Euclidean distance (being color difference) and human eye have higher dependency to the perception aberration of these two kinds of colors at 2, promptly they all have good visually-perceptible concordance.Simultaneously, these two color spaces all are with device-independent, and wherein component L represents brightness, and remaining color component does not then have implication directly perceived.Therefore, CIELUV and CIELAB are commonly used to the computation vision aberration, and are not suitable for color calibration.

Different with above-mentioned color space, HSV (perhaps HSL) is called as " intuition system " (intuitivesystem), because each color all is represented as 3 components directly perceived in this system: colourity (H), saturation (S) and brightness (V or L) therefore are particularly suitable for color calibration.Though HSV can be obtained through linear transformation by RGB, it also is not suitable for the picture of the tongue diagnosis.Because the H component is discontinuous in redness and purple intersection (H=0), and the color of tongue body is in this color region mostly, is easy to cause result's deviation like this.Therefore, this paper will extract in other 4 spaces and quantize color characteristic.

No matter be healthy people or disease patient, its tongue body color presents single dominant hue usually, and we use the color average of whole tongue body as a kind of quantization characteristic for this reason.Simultaneously, owing to may be all or part of or be and strip off shape and be coated with tongue fur in tongue body surface, and the color of tongue fur often with the color different (their colour gamut has only part to intersect) of tongue body, and the standard deviation of tongue body zone picture element color value can embody this distribution on the whole, so we with it also as a kind of color characteristic (strictly speaking, the color value standard deviation should classify as the color and vein feature, and here we are not distinguished this).

Color average above-mentioned and standard deviation can be in 4 color spaces on any one color plane (perhaps being called passage), therefore we can obtain 22 quantized color features and (notice that " L " among CIELUV and the CIELAB is identical, all refer to brightness), as shown in table 1.

Table a kind of quantized color feature

Plane (color space)	Color characteristic (average)	Color characteristic (standard deviation)
			R(RGB) G(RGB) B(RGB) Y(CIEYxy) x(CIEYxy) y(CIEYxy) L(CIELUV/AB) U(CIELUV) V(CIELUV) A(CIELAB) B(CIELAB)	CR 1 CR 2 CR 3 CR 4 CR 5 CR 6 CR 7 CR 8 CR 9 CR 10 CR 11	CR 12 CR 13 CR 14 CR 15 CR 16 CR 17 CR 18 CR 19 CR 20 CR 21 CR 22

2.1.2 quantification textural characteristics

In all statistics textural characteristics, gray level co-occurrence matrixes (graylevelco-occurrencematrices) is the most frequently used one, and it is the second-order statistic to the gray value spatial distribution.The gray level co-occurrence matrixes of one sub-picture is a square formation, and any the element correspondence in the square formation the right relative frequency of picture element that meets certain spatial relation.In form, meet certain spatial relationship d=(dx, gray level co-occurrence matrixes dy) is the square formation of a G * G, the element definition in the matrix is:

$P_d(g_1,g_2)=\left|\left\{\begin{array}{c}(a,b)\∈N\×N:I(a,b)=g_1,I(a+\mathrm{dx},b+\mathrm{dy})=g_2,\\ (a+\mathrm{dx},b+\mathrm{dy})\∈N\×N\end{array}\right\}\right|---\left(5\right)$

In the formula, I () expression has the image of the N * N of G gray level, g ₁And g ₂The expression gray level, || the gesture of representative set.

Among the present invention, we use two textural characteristics that extract tongue image based on the texture operator of gray level co-occurrence matrixes.2-rank square and contrast that these two operators are respectively gray level co-occurrence matrixes are defined as follows:

$W_M=\underset{g_1}{\mathrm{\Σ}}\underset{g_2}{\mathrm{\Σ}}{p}^2(g_1,g_2)$

$W_C=\underset{g_1}{\mathrm{\Σ}}\underset{g_2}{\mathrm{\Σ}}|g_1-g_2|p(g_1,g_2)---\left(6\right)$

In the formula, p (g ₁, g ₂) be normalized co-occurrence matrix, i.e. p (g ₁, g ₂)=P _d(g ₁, g ₂)/S, wherein S represent might the right total number of gray level, formulate is as follows:

$S=\underset{g_1=0}{\overset{G-1}{\mathrm{\Σ}}}\underset{g_2=0}{\overset{G-1}{\mathrm{\Σ}}}{P}_d(g_1,g_2)---\left(7\right)$

W _MBe Image Smoothness or conforming a kind of tolerance, as all p (g ₁, g ₂) W when all equating _MReach minima.Contrast W _CIt is 1-rank square about co-occurrence matrix gray level difference.It is related to it should be noted that these two kinds of quantized textural characteristics and human visual perception almost do not have, i.e. their implications significantly not directly perceived equally.In this article, according to the characteristics and the experimental data of tongue image, we get G=64 and d=(6,6).

Different with the distribution of above-mentioned tongue body color, we find that the texture of tongue surface zones of different can have bigger difference each other, and this is not only confirmed by a large amount of clinical observations but also matches with the Evolution of Tongue Inspection of TCM holographic theory.For example, Evolution of Tongue Inspection of TCM is thought the tip of the tongue marquis cardiopulmonary, marquis's taste or the like in the tongue.In order to embody this characteristics that the tongue superficial makings distributes, we at first are divided into tongue body 5 zones: in the tip of the tongue, left tongue limit, the tongue, right tongue limit and the root of the tongue, as shown in Figure 6.Then, we take out a texture block and calculate above-mentioned 2 kinds of textural characteristics in each zone.Like this, we will obtain 10 textural characteristics for each width of cloth tongue image, and will be as shown in table 2.

Table 2 kind of quantification textural characteristics

Subregion (label)	W M	W C
			(3) right tongue limit (4) root of the tongue (5) in the tongue of the tip of the tongue (1) left tongue limit (2)	TR 1 TR 2 TR 3 TR 4 TR 5	TR 6 TR 7 TR 8 TR 9 TR 10

2.2, experiment and interpretation of result

We use the BayesianNetworkPowerPredictor by people such as Cheng exploitation to come training and testing inspection of the tongue model.PowerPredictor reads data (tongue image characteristic and classified information) from Access database file (* .mdb), through after the training training result (structure of Bayes classifier and relevant parameter) being exported with figure and text dual mode.Its training process has used the correlation analysis in the theory of information (dependenceanalysis) method, promptly utilizes condition independence (CI) test to come dependency (perhaps dependence) between the node metric.Thereby this system is divided into 3 subprocess with whole training process has shortened the training time, utilizes the wrapper algorithm to avoid training problem simultaneously.What deserves to be mentioned is that in fact the test of dependency in the training process is exactly a kind of feature selection, make to have only the most contributive feature of diagnostic classification could be got off in " survival ", and this we are desirable just.

In the experiment below, the present invention has used 525 width of cloth tongue images (to gather the patient from 525 healthy people and 13 kinds of commonly encountered diseases respectively altogether, experiment sample distributes and sees Table 3), wherein the healthy people's tongue image of 70 width of cloth comes from university student volunteer, 455 width of cloth disease tongue images come from patient in hospital, and Fig. 7-10 has provided 4 width of cloth typical case tongue image sample.

Table 3 kind of commonly encountered diseases and healthy people's tongue image sample list

The disease label	Disease name	Sample size
			D00 D01 D02 D03 D04 D05 D06 D07 D08 D09 D10 D11 D12 D13	The sense of the Healthy People intestinal obstruction cholecystitis appendicitis pancreatitis ephritis diabetic hypertension heart failure pulmonary heart disease coronary heart disease cirrhosis cerebral infarction upper respiratory tract	70 11 21 43 41 17 49 65 17 21 71 25 30 44

2.3, some problems

There is its particularity in computer inspection of the tongue of the present invention system as the application of a medical diagnosis, at first is the problem about evaluation methodology.In fact, the picture of the tongue diagnosis is typical pattern recognition problem, and its essence is classified to tongue image according to quantization characteristic exactly.So, how should we determine the classification (classified information) of tongue image? the problem of Here it is so-called interpretational criteria.Method commonly used is at present judged tongue presentation (perhaps tongue image) according to the personal experience by Traditional Chinese Medicine experts, provides diagnosis then.The problem of this method is that it does not solve the problem that Evolution of Tongue Inspection of TCM too relies on personal experience and visually-perceptible, and while traditional Chinese medical doctor's diagnosis is pattern of syndrome rather than disease normally.Yet this paper attempts to set up the contact between tongue image quantization characteristic and the disease, so we are with patient's " the gold criterion " that disease is classified as tongue image of making a definite diagnosis.Because all samples among our the inspection of the tongue data base are all from the inpatient, so we think that this classified information is reliable.At test phase, we make comparisons the diagnosis output of Bayes classifier and the category label (disease) of tongue image, and provide the assessment of grader performance in view of the above.The method has been established the basis that objectifies of our interpretational criteria.

Next is the problem about small sample set.In fact, the trial of nearly all medical statistics aspect all faces same difficulty, promptly how to obtain the sample of sufficient amount and how to guarantee that the data that obtain are reliable.At the problem of small sample set, we adopt k-heavily layering cross validation technology (stratifiedk-foldcross-validation, we get k=10) assess the performance of grader.K-retransposing checking is that data set S is divided into k sample size subclass about equally.Appoint get wherein 1 then as test set, test a classification results practising the grader that comes with all the other k-1 son training.Repeat this process k time, with the average of k subseries precision as last assessment result.And k-heavily layering cross validation technology and said method are basic identical, unique difference is to require each subclass all to have similar sample class to S to distribute, the benefit of doing like this is to reduce the variance of assessment result.

Have, we use discrete Bayesian network in all below experiments again.Here " dispersing " is meant the probability distribution of node, and promptly the conditional probability table of all nodes all is the probability distribution that disperses.Here the discretization method to all territories (perhaps being called " attribute ") is equally spaced (equalwidth), and interval number is set at 5.Simultaneously, for the advantage of Bayes classifier in inspection of the tongue is used is described, we are with nearest neighbor classifier (nearest-neighborclassifier, NNC) experiment of comparing.Same, we also with k-heavily layering cross validation technology the classification results of NNC is done assessment.Like this, each sample in the test set always is noted as the nearest with it identical classification of training set sample, and the average that obtains after then k time being repeated is as the overall accuracy of grader.

Be the problem that is provided with at last about misclassification cost table (misclassificationcosttable).In the practical application of Bayes classifier, we can adjust the distribution of nicety of grading according to demand by the weight in the change misclassification cost table.This paper, we use identical weight coefficient to all misclassifications, and this perhaps is not inconsistent with practical situation, but can't have influence in the literary composition demonstration to the effectiveness of Bayes classifier in inspection of the tongue is used.

3, Bayes classifier

3.1, based on the Bayes classifier of textural characteristics

In first experiment, we utilize the Bayes classifier of above-mentioned PowerPredictor system structure based on textural characteristics, are referred to as " T-BNC ".The network structure of this grader as shown in figure 11.Notice that the structure correspondence shown in the figure the highest output of precision among the heavy CV of 10-.Through the feature selection process in the training algorithm, in 10 initial texture features 5 appear in the network structure of Figure 11.Wherein comprise 2 W _MFeature: the TR on left tongue limit ₂TR with the root of the tongue ₅3 W _CFeature: TR ₆(the tip of the tongue), TR ₇(left tongue limit) and TR ₁₀(root of the tongue).Obviously, the textural characteristics of the corresponding the tip of the tongue, tongue limit and the root of the tongue is maximum for the contribution of picture of the tongue sample classification.Though have only the textural characteristics on left tongue limit to be selected, we find that the textural characteristics numerical value on left and right tongue limit is very approaching in fact.

The diagnostic result of T-BNC provides the accuracy of classification (truepositiverate, TPR at the 2nd row of table 4; Perhaps sensitivity) be 26.1%.Obviously, our textural characteristics that extracts and the diagnosis that is not suitable for listed disease in the his-and-hers watches 3.But, concerning the diagnosis of some disease, these textural characteristics still have certain meaning, for example appendicitis (D03), pancreatitis (D04) and coronary heart disease (D10).Because lower overall accuracy, we do not provide the PPV (positivepredictivevalues) of T-BNC.

3.2, based on the Bayes classifier of color characteristic

Test color characteristic in this part is for the suitability of tongue image classification diagnosis, is called " C-BNC " based on the Bayes classifier of color characteristic, and its network structure that gets as shown in figure 12.Same, this structure correspondence the highest CV output.Through feature selection, 12 color characteristics last " winning ".As can be seen from the figure, 6 features wherein directly link to each other with root node (classification node), and we claim such node to be " contribution node ".It should be noted that in 4 color spaces each has all comprised 1 " contribution node " at least, and this two category feature of average and standard deviation has similar importance for diagnosis.

The TPR of C-BNC diagnostic result and PPV are provided by the 3rd row of table 4 and the secondary series of table 5 respectively.Obviously, the diagnostic classification performance of C-BNC is far superior to T-BNC:TPR and has reached 62.3%.What deserves to be mentioned is, C-BNC not only is (TPR=90.2%) but also be reliable (PPV=68.5%) very accurately to the diagnosis of pancreatitis (D04), its main cause is that pancreatitis patient's tongue body color presents tangible blueness (as shown in figure 10) usually, is typical " blue tongue ".

3.3, based on the associating feature Bayes classifier

At last, we use all 32 colors and joint classification device of textural characteristics structure, are referred to as " J-BNC ".Its network structure as shown in Figure 5, the highest same corresponding CV output.Have 14 features and be selected, wherein having 9 is " contribution node " (5 color characteristics, 4 textural characteristics).Notice that the textural characteristics of last " winning " is respectively the tip of the tongue (TR ₁) and right tongue limit (TR ₄) W _MAnd the tip of the tongue (TR ₆) and the root of the tongue (TR ₁₀) W _CIn other words, in J-BNC, the maximum textural characteristics of diagnosis contribution is all come from the tip of the tongue, tongue limit and the root of the tongue, this with experiment one in the conclusion that draws of T-BNC match.

Table 4 the 4th row and table 5 the 3rd row have provided the TPR and the PPV of J-BNC diagnostic result respectively.The overall classification accuracy rate of J-BNC reaches 75.8%, be higher than T-BNC 26.1% and C-BNC 62.3%.Simultaneously, by table 4 and table 5 as can be seen, J-BNC is more satisfactory for the diagnosis of healthy people (D00), pancreatitis (D04), hypertension (D07) and cerebral infarction (D12), and their TPR and PPV all are higher than 75%.In order better to show classification results, table 6 has provided the confusion matrix (confusionmatrix) of J-BNC.

3.4, the contrast experiment

For the advantage of Bayes classifier in the tongue image diagnosis is described, we use nearest neighbor classifier to be a contrast experiment.In order to strengthen comparability, we use the identical process of training and testing with J-BNC: to unite the feature samples collection as data set; With k-heavily layering cross validation technology the classification results of NNC is done assessment.Contrast experiment's result (TPR) is as shown in table 7.Obviously, for the tongue image diagnostic classification, the performance of Bayes classifier is far superior to nearest neighbor classifier.

The diagnostic result (accuracy) of table 43 kind of Bayes classifier (%)

The disease label	T-BNC	C-BNC	J-BNC
				D00 D01 D02 D03 D04 D05 D06 D07 D08 D09 D10 D11 D12 D13	20.0 9.1 4.8 53.5 70.7 5.9 4.1 3.1 5.9 4.8 64.8 12 13.3 20.5	50.0 45.5 42.9 86.0 90.2 17.6 53.1 61.5 35.3 47.6 90.1 48.0 60.0 56.8	77.1 63.6 61.9 93.0 100 23.5 65.3 75.4 35.3 71.4 93.0 64.0 80.0 70.5
Meansigma methods	26.1	62.3	75.8

The diagnostic result (PPV) of table 5C-BNC and J-BNC (%)

The disease label	C-BNC	J-BNC
			D00 D01 D02 D03 D04 D05 D06 D07 D08	79.5 100 100 60.7 68.5 100 81.3 66.7 100	85.7 87.5 92.9 72.7 75.9 80.0 84.2 80.3 100

D09 D10 D11 D12 D13

83.3 37.9 100 66.7 80.6

88.2 54.5 88.9 75.0 93.9

The confusion matrix of table 6J-BNC

Table 7J-BNC and NNC contrast and experiment (accuracy) are (%)

The disease label	J-BNC	NNC
			D00 D01 D02 D03 D04 D05 D06 D07	77.1 63.6 61.9 93.0 100 23.5 65.3 75.4	28.6 54.5 33.3 44.2 56.1 35.3 18.4 18.5

D08 D09 D10 D11 D12 D13	35.3 71.4 93.0 64.0 80.0 70.5	52.9 52.4 46.5 40.0 20.0 43.2
			Meansigma methods	75.8	36.2

Claims (5)

1, a kind of Computerized health index analysis system based on picture of the tongue is characterized in that at first utilizing digital image processing techniques to extract the color and the textural characteristics of tongue image, utilizes Bayesian network that getting in touch between above-mentioned quantization characteristic and the disease carried out modeling then.

2, a kind of Computerized health index analysis system based on picture of the tongue according to claim 1 is characterized in that the described tongue image quantization characteristic that extracts with digital image processing techniques comprises color and textural characteristics.

3, a kind of Computerized health index analysis system based on picture of the tongue according to claim 1 is characterized in that described color quantizing feature comprises the color average of whole tongue body and the standard deviation of tongue body zone picture element color.

4, a kind of Computerized health index analysis system based on picture of the tongue according to claim 1 is characterized in that described texture quantization characteristic is to use the texture quantization characteristic that extracts tongue image based on the texture operator of gray level co-occurrence matrixes.

5, a kind of Computerized health index analysis system according to claim 1 based on picture of the tongue, it is characterized in that described utilize Bayesian network between quantization characteristic and the disease get in touch the modeling of carrying out comprise Bayes classifier based on textural characteristics, based on the Bayes classifier of color characteristic with based on the Bayes classifier of associating feature.

Images