CN1547149A - Method for implementing brain glioma computer aided diagnosis system based on data mining - Google Patents

Abstract

The invention is a method for brain glioma computer aided diagnosing system based on data excavating. The invention carries on attributes digitized process to data in the case bank at first by using the information in the case bank of the alioma sufferer, and creates the fuzzy membership relation to the digitized relation. Then the invention uses the fuzzy principle extraction method of the fuzzy maximal and minimal nerve network. The principle of diagnosis for excavating and discovering the alioma malignancy degree, the system is created according to the principle. The system can acquire the malignancy degree of any new case, acquires a high accuracy.

Description

Implementation method based on the brain glioma computer-aided diagnosis system of data mining

Technical field:

The present invention relates to a kind of implementation method of the brain glioma computer-aided diagnosis system based on data mining, relate to fields such as pattern-recognition, data mining and radiologic medicine, can directly apply to the computer-aided diagnosis of brain malignant grade of gliomas height.

Background technology:

The brain glioma is a kind of comparatively rare disease, and treatment depends primarily on the grade malignancy of tumour.Whether the correct decision of its grade malignancy is related to needs to carry out the very high cerebral operations of hazard level for patient.As if judging the height of grade malignancy according to patient symptom more reliably and hanging down and just may avoid unnecessary operation risk and spending.

The main foundation of the diagnosis of current glioma grade malignancy is the analysis of brain nuclear magnetic resonance image.But,, correctly judge it is the work of a difficulty owing to lack the chance of the enough cases of accumulation for most of radiologists.This strong (Yang Benqiang that waits of poplar, Wu Zhenhua, the MRI of the big cerebral gliomatosis of Zhou Lijuan diagnoses Chinese clinical medicine image magazine 2000.11 (4): 229-231) inquired into the clinical value of nuclear magnetic resonance image aspect the diagnosis of glioma grade malignancy, and in conjunction with 9 routine analysiss of cases the part experience of glioma grade malignancy diagnosis.But, can still be difficult to sum up for the correct experience of judging because case is less.Obviously, if a glioma patient grade malignancy, will cause the delay or the great operation risk of treatment by error diagnosis.For addressing this problem, can collect a large amount of cases from a big medical centre (as: one of the Huashan, Shanghai City hospital-foremost hospital of domestic neurosurgery), therefrom find the rule between sheet result and the glioma grade malignancy read of nuclear magnetic resonance image.

Computer-aided medical diagnosis is a research direction of artificial intelligence, and it is used for knowledge or the experience that the differentiation of brain malignant grade of gliomas can make more doctor utilize this aspect to accumulate.The problem that realizing this goal at first needs to solve is how to obtain diagnostic knowledge (rule between symptom and disease grade malignancy).Traditional method is to be consulted to the human expert who is good at this disease of diagnosis by special personnel, and the rule that the expert is summed up is put in order and with the foundation of their formalization as computer diagnosis then.Yet, exist in this process " bottleneck " of knowledge acquisition.Data mining (Data mining) as the new focus in machine learning field for addressing this problem the means that provide new.But must consider following ask for something:

1) correctness: reach domain expert's level, accuracy of diagnosis should surpass 80%.

2) fault-tolerance: the radiologists different for identical brain nuclear magnetic resonance image may draw different descriptions.Profile with glioma is an example, and possible description comprises: circular, oval, irregular shape, and under actual conditions, do not have accurate circle or ellipse.For this uncertainty, should keep the sane of diagnostic result as far as possible.

3) missing values: because expense or necessity aspect, the value of some project is empty in some diagnosed SARS cases.Still can effectively judge this moment.

4) intelligibility: the diagnosis rule should be able to be understood preferably succinct language description by the radiologist.

Apt é and and Weiss (Apt é Chidanand and Weiss Sholom (1997): " Data miningwith decision trees and decision rules " " Future Generation ComputerSystems ", 1997,13, pp.197-210) propose to find to diagnose rule with traditional decision-tree, but the decision tree nodes that generates is a lot, be difficult for understanding, and occur over-fitting problem (factor produces unnecessary decision tree nodes according to noise) easily.Agrawal (Agrawal R, Mannila H, Srikant R, Toivonen H., andVerkamo A.I. (1998): " Fast Discovery of Association Rules " " Advances inKnowledge Discovery and Data Mining " (Morgan Kaufmann, San Mateo, CA) UsamaM.fayyad pp.307-328) etc. the proposition rough set theory excavate correlation rule.But rough set theory is difficult for handling real number type data and uncertain description, and the regular quantity of generation is more, wordy and difficult understanding.Multilayer perceptron network based on BP training study algorithm can obtain high accuracy of diagnosis but the non-constant of intelligibility.

Summary of the invention:

The objective of the invention is at the deficiencies in the prior art, a kind of brain glioma computer-aided diagnosis system implementation method based on data mining is proposed, case storehouse according to the glioma patient who collects, excavate and find the diagnostic rule of glioma grade malignancy, and set up glioma grade malignancy computer-aided diagnosis expert system according to this rule, be used for the automatic diagnosis of glioma grade malignancy, to reduce misdiagnosis rate and patient's misery.

For realizing such purpose, the present invention adopts based on improved fuzzy minimax neural network (FuzzyMinimum-Maximum Neural Network, FMMNN) fuzzy rule extracting method, from the glioma patient's that collects case storehouse (comprising: nuclear magnetic resonance image (MRI) read sheet result and post operative diagnosis result), excavate and the diagnostic rule of discovery glioma grade malignancy, set up glioma grade malignancy computer-aided diagnosis expert system according to the grade malignancy diagnostic rule that excavates and find, utilize the diagnostic system of being set up again, the new case forecast of any input is drawn the grade malignancy of this case, thereby offer help for doctor's successive treatment and diagnosis.

The case database that the present invention utilized stores from the glioma patient's of hospital's collection information, comprise: patient's personal information (sex, age), patient's brain nuclear magnetic resonance image and expert thereof read sheet result (shape, profile, coating, oedema, the occupy-place effect, strengthening the back strengthens, blood supplies, necrosis/capsule becomes, calcification, hemorrhage, the T1 weighting, the T2 weighting), each feature description is called attribute, the malignant grade of gliomas diagnostic result that patient's postoperative obtains is divided into two classifications of height/low potential malignancy, and the record that has comprised all above-mentioned property values of certain patient and diagnostic result is called sample, and all these information are stored in the special database.

Brain glioma computer-aided diagnosis system implementation method based on data mining of the present invention is carried out as follows:

1, the digitized processing of attribute.At each attribute in the case database, add up different description forms, and sort according to size (as: age), weight description forms such as (as oedema), replace with the integer of respective sequence respectively then.Be about to property value and be mapped to an integer sequence.To indicate for non-existent property value.

2, each attribute after the logarithm value is set up fuzzy membership: the process that quantizes defines a susceptibility factor r after finishing, and sets up membership between the different values of same attribute.That is, for some attribute A, its k value v _kWith j value v _jBetween the degree of membership value be: $\mathrm{\μ}=\max (\mathrm{0,1}-r\frac{|v_j-v_k|}{\max A-\min A}).$ Here maxA is the maximum occurrences of attribute A, and minA is the minimum value of attribute A.Experiment finds that the size and the r of test errors rate do not have clear regularity, so be difficult to optimize r with high efficiency method, general r gets the value between the 0-10.

3, super box generates and expansion: to each training sample, calculate its degree of membership to the identical super box of all classifications according to its classification, and select to have the super box A of maximum membership degree _DkIf there is no the super box of identical category just generates a super box that has comprised this training sample; If there is super box A _Dk, just check super box dilation procedure that this training sample is included whether cause ambiguity error to reduce or under the prerequisite that ambiguity error does not increase the classification error rate reduce, if reduce just to carry out super box dilation procedure; Otherwise generate a new super box.

When algorithm is a sample that contains some disappearance attributes when generating super box, super box is being lacked on the minimum and maximum value that coboundary on the attribute and lower boundary be placed on this attribute.And when algorithm was the super box of sample expansion that contains some disappearance attributes, the upper and lower boundary position of super box on corresponding attribute remained unchanged.This improvement makes FMMNN can handle the disappearance attribute.

In the present invention, for strengthening the readability and the intelligibility of net result, certain sample is to the degree of membership m of the identical super box of classification _Ajk(Sample) calculate by following formula:

$m_{A_{\mathrm{jk}}}\left(\mathrm{Sample}\right)={\min }_{i=\mathrm{1,2},...p}\left({\max }_{{e\∈V}_i^{\′}}{\mathrm{\μ}}_{i,e}\left(S_i\right)\right)={\min }_{i=\mathrm{1,2},...p}\left({\max }_{e\∈V_i^{\′}}{\mathrm{\μ}}_{i,\mathrm{numeric}\_\mathrm{presention}\_\mathrm{of}\left(e\right)}\left(\mathrm{xi}\right)\right)$

Here μ is the degree of membership value between two the different values of certain attribute A that define in the step 2.

Here the ambiguity error of mentioning and be to introduce for the robustness that improves algorithm.Ambiguity error and $E_{\mathrm{fuzzy}}=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{c=1}{\overset{l}{\mathrm{\Σ}}}{(d_{\mathrm{ic}}-m_{\mathrm{ic}})}^2,$ Wherein n is the training sample sum, and l is the sample class sum, d _IcBe the true degree of membership value of i sample class c, m _IcIt is i sample obtaining according to current sorter normalization degree of membership value to classification c.

4, overlapping detection: the super box of identical category allows overlapping, so as long as detect in the super box of certain class whether contain different classes of sample, or whether the different classes of sample that contains surpass a predetermined value, if just illustrate and exist overlappingly between different classes of super box, otherwise that explanation does not have is overlapping.

5, super box shrinks: to there being overlapping super box, the check shrinkage operation whether cause ambiguity error to reduce or under the prerequisite that ambiguity error does not increase the classification error rate reduce, if just carry out shrinkage operation, otherwise do not carry out.

6, super box is carried out extra expansion: after above-mentioned steps is carried out and finished, if the expression that quantizes of certain training sample has identical degree of membership value to different classes of super box, will obtain the small expansion of extra θ/10 (or littler) so with the generic super box of this sample.These small expansions only occur in and can make on the border that this sample correctly classified.In this course, if the small expansion of super box causes ambiguity error and E _FuzzyReduce or the classification error rate reduces, just carry out super box dilation procedure; Otherwise do not carry out.

If be modified in 7 any one super box step in front, then repeat since the operation in the 3rd step.If be not modified, illustrate and found all super boxes.Training process promptly comes to an end.The super box that generates has constituted a fuzzy classification device.

8, fuzzy rule extracts: will surpass that box changes into " if ... then ... " the fuzzy rule form, and arrangement realizes the expert system of glioma grade malignancy computer-aided diagnosis in view of the above.

Traditional FMMNN includes only the generation and the expansion of super box, and overlapping detection and super box shrink three basic operations.The present invention has done 3 improvement on this basis: 1. for improving robustness, introduce ambiguity error in the super box expansion in the 3rd and the 6th step: $E_{\mathrm{fuzzy}}=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{c=1}{\overset{l}{\mathrm{\Σ}}}{(d_{\mathrm{ic}}-m_{\mathrm{ic}})}^2$ Wherein n is the training sample sum, and l is the sample class sum, d _IcBe the true degree of membership value of i sample class c, m _IcIt is i sample obtaining according to current sorter normalization degree of membership value to classification c.2. in the super box in the 3rd step generates, when algorithm is a sample that contains some disappearance attributes when generating super box, super box is being lacked on the minimum and maximum value that coboundary on the attribute and lower boundary be placed on this attribute.And when algorithm was the super box of sample expansion that contains some disappearance attributes, the upper and lower boundary position of super box on corresponding attribute remained unchanged.This improvement makes FMMNN can handle the disappearance attribute.3. in the super box in the 3rd step generated, for strengthening the readability and the intelligibility of net result, sample was to super box A _JkThe degree of membership value change by following formula and calculate to obtain:

$m_{A_{\mathrm{jk}}}\left(\mathrm{Sample}\right)={\min }_{i=\mathrm{1,2},...p}\left({\max }_{{e\∈V}_i^{\′}}{\mathrm{\μ}}_{i,e}\left(S_i\right)\right)={\min }_{i=\mathrm{1,2},...p}\left({\max }_{e\∈V_i^{\′}}{\mathrm{\μ}}_{i,\mathrm{numeric}\_\mathrm{presention}\_\mathrm{of}\left(e\right)}\left(\mathrm{xi}\right)\right)$

After the fuzzy rule about the diagnosis of glioma grade malignancy is extracted, use the fuzzy rule of these glioma grade malignancy diagnosis to set up the expert system of glioma grade malignancy computer-aided diagnosis on computers.In actual applications, when the information (sex, age, tumor locus feature description) of waiting to diagnose patient by this expert system input, this expert system is according to the information of input and the Fuzzy Rule Sets of glioma grade malignancy diagnosis, the forecast that draws this patient's glioma grade malignancy by fuzzy reasoning, thus offer help for doctor's successive treatment and diagnosis.

Method of the present invention can obtain higher test accuracy rate.Because the degree of membership relation of setting up between the different values of same attribute makes this method can utilize more in fact similar samples, therefrom find more representative decision rule, and traditional decision Tree algorithms and correlation rule extraction can not utilize this degree of membership relation.Few at some training samples, contain uncertain attribute, require to excavate in the succinct intelligible special applications of result, method of the present invention has more practical value.

The glioma grade malignancy computer-aided diagnosis expert system that the present invention sets up, can be used for the automatic diagnosis of the good grade malignancy of glioma and backcountry doctor and young doctor's diagnosis training, can be more reliably judge the height of grade malignancy and low according to patient symptom, avoid unnecessary operation risk and spending, reduce misdiagnosis rate and patient's misery.

Description of drawings:

Fig. 1 is expert system operation interface signal of the present invention.

Embodiment:

Below in conjunction with specific embodiment technical scheme of the present invention is described in further detail.

The big cerebral gliomatosis example database that embodiment adopts has 280 parts, is provided by Huashan Hospital Affiliated To Fudan Univ, does not comprise personal informations such as name.Each case (sample) has all been chosen the attribute record that 14 MRI read the true grade malignancy of sheet result and the big glioma of patient.These records mode word records that adopt are designated as more $T=\left\{({\stackrel{\→}{x}}_l,y_l)\right\}.$ (l＝1，2，…，280)， ${\stackrel{\→}{x}}_l=\{x_{l1},x_{l2},\·\·\·,x_{\mathrm{lp}},\·\·\·,x_{l14}\}$ Contain 14 attributes, the classification of sample adds up to 2 (optimum/pernicious), with y _l{ 1,2} represents ∈.

The total system implementation procedure is as follows:

1. the digitizing of attribute: text description is mapped to integer sequence, to each attribute through above-mentioned conversion, can be with＜s ₁, s ₂..., s _pNumerical value turns to＜x ₁, x ₂..., x _p.s _iBe the literal form of expression, x _iBe digital representation.

As: sex: " women " is digitized into 0, and " male sex " is digitized into 1;

Age: directly use;

Hemorrhage: " nothing " is digitized into 1, and " acute " is digitized into 2, and " chronic " is digitized into 3, or the like.

2. fuzzy membership is definite:

For an attribute A, the degree of membership value between its k value and j the value is: $\mathrm{\μ}=\max (\mathrm{0,1}-r\frac{|v_j-v_k|}{\max A-\min A}).$ Here v _jAnd v _kBe meant two different values of attribute, maxA is the maximum occurrences of attribute A, and minA is the minimum value of attribute A.R=1 in this example.

With " age " attribute is example, if r=1, so " always " (numerical value turns to 1) and " in " (numerical value turns to 2) be respectively 0 and 0.5 to the degree of membership value of " green grass or young crops " (numerical value turns to 3).

3. super box generates and expansion:

To each training sample

Calculating it is y to all categories _lThe degree of membership of super box, and select to have the super box A of maximum membership degree _Jk(j=y _l).If do not exist so super box, algorithm will generate a new super box A _Jr(j=y _l), and:

w _jri←x _li，v _jri←x _li (1-1)

If A _JrExist, need judge then whether following formula is set up:

$\mathrm{p\θ}\≥\underset{i=1}{\overset{p}{\mathrm{\Σ}}}(\max (w_{\mathrm{jki}},x_{\mathrm{li}})-\min (v_{\mathrm{jki}},x_{\mathrm{li}}))$

If following formula is false, algorithm also will generate a new super box according to (1-1) formula, otherwise to A _JkExpand as follows:

$v_{\mathrm{jki}}^{\mathrm{new}}\←\min (v_{\mathrm{jki}}^{\mathrm{old}},x_{\mathrm{li}}),w_{\mathrm{jki}}^{\mathrm{new}}\←\max (w_{\mathrm{jki}}^{\mathrm{old}},x_{\mathrm{li}}),\∀i=\mathrm{1,2},\·\·\·,p$

The super box A in expansion back _JkTo comprise sample

θ is a preset value, gets θ=1 here.

4. overlapping detection: detect whether exist between different classes of super box overlapping.

The span of each attribute of sample all normalizes to [0,1], judges any two different classes of super box A in the example by the following method _JkAnd A _HtBetween whether exist overlappingly, and finding that the attribute of finding out overlapping scope minimum when overlapping is kept among the s.Method of discrimination is as follows: introduce two constant δ ^OldAnd i, and initialization δ ^Old=1, i=1, carry out 4 judgements to i attribute:

(a) work as v _Jki＜v _Hti＜w _Jki＜w _HtiThe time, δ ^New=w _Jki-v _Hti

(b) work as v _Hti＜v _Jki＜w _Hti＜w _JkiThe time, δ ^New=w _Hti-v _Jki

(c) work as v _Jki＜v _Hti＜w _Hti＜w _JkiThe time, δ ^New=min (w _Hti-v _Jki, w _Jki-v _Hti);

(d) work as v _Hti＜v _Jki＜w _Jki＜w _HtiThe time, δ ^New=min (w _Hti-v _Jki, w _Jki-v _Hti);

(j and h are meant classification here, and k and t are meant super box number, v _JkiAnd w _JkiBe the lower boundary and the coboundary of super box i dimension)

If do not have one can be satisfied in above-mentioned, so A _JkAnd A _HtBetween do not exist overlappingly, calculate to finish.Otherwise (having one in above-mentioned at least is satisfied) illustrated the overlay region, judges δ this moment ^NewWhether less than δ ^OldIf, then δ ^NewValue is composed and is given δ ^Old, and note s=i, if exist i+1 attribute then to make i=i+l continue above-mentioned judgement; Finish otherwise calculate.

5. super box shrinks: shrink successively there being overlapping super box.If to the shrinkage operation of certain super box cause ambiguity error to reduce or under the prerequisite that ambiguity error does not increase the classification error rate reduce, this operation will be performed so; Otherwise this operation will be prevented from.According to following processing of the corresponding work of above 4 kinds of situations

(a) work as v _Jks＜v _Hts＜w _Jks＜w _HtsThe time, $w_{\mathrm{jks}}^{\mathrm{new}}=v_{\mathrm{hts}}^{\mathrm{new}}\←\frac{w_{\mathrm{jks}}^{\mathrm{old}}+v_{\mathrm{hts}}^{\mathrm{old}}}{2};$

(b) work as v _Hts＜v _Jks＜w _Hts＜w _JksThe time, $w_{\mathrm{hts}}^{\mathrm{new}}=v_{\mathrm{jks}}^{\mathrm{new}}\←\frac{w_{\mathrm{hts}}^{\mathrm{old}}+v_{\mathrm{jks}}^{\mathrm{old}}}{2};$

(c) work as v _Jks＜v _Hts＜w _Hts＜w _JksAnd (w _Hts-v _Jks)≤(w _Jks-v _Hts) time, v _Jks ^New← w _Hts ^Old

Work as v _Jks＜v _Hts＜w _Hts＜w _JksAnd (w _Hts-v _Jks＞(w _Jks-v _Hts) time, w _Jks ^New← v _Hts ^Old

(d) work as v _Hts＜v _Jks＜v _Jks＜w _HtsAnd (w _Hts-v _Jks)≤(w _Jks-v _Hts) time, w _Hts ^New← v _Jks ^Old

Work as v _Hts＜v _Jks＜w _Jks＜w _HtsAnd (w _Hts-v _Jks＞(w _Jks-v _Hts) time, v _Hts ^New← w _Jks ^Old

6. super box is carried out extra expansion: after above-mentioned steps is carried out and finished,, support the super box of the true classification of this sample will obtain extra small expansion so if the expression that quantizes of certain training sample has identical degree of membership value to different classes of super box.These small expansions only occur in and can make on the border that this sample correctly classified.In this course, if the small expansion of super box causes E _FuzzyReduce or the classification error rate reduces, this operation will be performed so; Otherwise this operation will be prevented from.

7., then repeat since the operation in the 4th step if be modified in any one super box step in front.If be not modified, illustrate and found all super boxes that training process promptly comes to an end.The super box that generates has constituted a fuzzy classification device.

8. fuzzy rule extracts:

280 parts of cases are formed the training sample set, adopt fuzzy rule extraction algorithm to obtain following fuzzy rule based on improved FMMNN

Rule _ A1: age in (1～53) AND oedema in (not having, slight) AND blood is in (general, general+slightly many) THEN low potential malignancy

Rule _ A2: age in (34～59) AND occupy-place effect in (moderate, serious) AND strengthens the back and strengthens in (inhomogeneous) AND blood for the hemorrhage in of in (enriching) AND (not having, acute) THEN high malignancy

Last arrangement in view of the above realizes assisting the expert to do the expert system of glioma diagnosis.Their rate of accuracy reached on training set to 84.64% (to low potential malignancy: 89.94%, to high malignancy: 76.58%).

The use of rule can illustrate with following example: certain female patients 22 years old, MRI are read the sheet result and are shown: the shape of big glioma: irregular; Profile: part is clear; Coating: imperfect; Oedema: light; Occupy-place effect: moderate; Strengthening the back strengthens: inhomogeneous; Blood supplies: abundant; Necrosis/capsule becomes: have; Calcification: do not have; Hemorrhage: as not have; The T1 weighting: wait signal or etc. signal follow low signal; T2 weighting: high signal.This case is respectively 0.18 and 0.91 to the degree of membership of above-mentioned two rules, should be high malignancy.The judged result of fuzzy rule is observed consistent with operation.

The expert system operation interface that the present invention realizes as shown in Figure 1.In actual applications, the brain glioma computer-aided diagnosis system that utilizes the inventive method to set up as long as the relevant patient information of input just can forecast the grade malignancy that draws this case, thereby offers help for doctor's successive treatment and diagnosis.Patient's data can add the case database simultaneously.

Claims (1)

1, a kind of brain glioma computer-aided diagnosis system implementation method based on data mining is characterized in that comprising following concrete steps:

1) digitized processing of attribute: at each attribute in the case database, sort according to different description forms, the integer with respective sequence replaces respectively, is about to property value and is mapped to an integer sequence, will indicate for non-existent property value;

2) each attribute after the logarithm value is set up fuzzy membership: the process that quantizes defines a susceptibility factor r after finishing, and sets up membership between the different values of same attribute, that is, for some attribute A, its k value v _kWith j value v _jBetween the degree of membership value be: $\mathrm{\μ}=\max (\mathrm{0,1}-r\frac{|v_j-v_k|}{\max A-\min A}),$ Here maxA is the maximum occurrences of attribute A, and minA is the minimum value of attribute A, and r generally gets the value between the 0-10;

3) super box generates and expansion: to each training sample, calculate its degree of membership to the identical super box of all classifications according to its classification, and select to have the super box A of maximum membership degree _Dk, if there is no the super box of identical category just generates a super box that has comprised this training sample; If there is super box A _Dk, just check super box dilation procedure that this training sample is included whether cause ambiguity error to reduce or under the prerequisite that ambiguity error does not increase the classification error rate reduce, if reduce just to carry out super box dilation procedure; Otherwise generate a new super box; Ambiguity error and $E_{\mathrm{fuzzy}}=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{c=1}{\overset{l}{\mathrm{\Σ}}}{(d_{\mathrm{ic}}-m_{\mathrm{ic}})}^2,$ Wherein n is the training sample sum, the 1st, and sample class sum, d _IcBe the true degree of membership value of i sample class c, m _IcIt is i sample obtaining according to current sorter normalization degree of membership value to classification c;

Whether 4) overlapping detection: the super box of identical category allows overlapping, so as long as exist overlapping between the different classes of super box of detection;

5) super box shrinks: to there being overlapping super box, the check shrinkage operation whether cause ambiguity error to reduce or under the prerequisite that ambiguity error does not increase the classification error rate reduce, if just carry out shrinkage operation, otherwise do not carry out;

6) super box is carried out extra expansion: after above-mentioned steps is carried out and finished, if the expression that quantizes of certain training sample has identical degree of membership value to different classes of super box, to obtain extra small expansion with the generic super box of this sample so, these small expansions only occur in and can make on the border that this sample correctly classified, in this course, if the small expansion of super box causes ambiguity error and E _FuzzyReduce or the classification error rate reduces, just carry out super box dilation procedure; Otherwise do not carry out;

7), repeat then if be not modified, illustrate and found all super boxes that since the operation in the 3rd step training process promptly comes to an end that the super box of generation has constituted a fuzzy classification device if be modified in any one super box step in front;

8) fuzzy rule extracts: will surpass that box changes into " if ... then ... " the fuzzy rule form, and arrangement realizes the expert system of glioma grade malignancy computer-aided diagnosis in view of the above.