CN105913086A - Computer-aided mammary gland diagnosing method by means of characteristic weight adaptive selection - Google Patents

Abstract

The invention relates to a computer-aided mammy gland diagnosing method by means of characteristic weight adaptive selection. The method comprises the steps of firstly extracting a mammary gland X-ray molybdenum target and a type-B ultrasonic image data characteristic, performing benign-malignant and clinic periodic marking on case data after characteristic extraction according to known clinical diagnosis results; performing multi-characteristic fusion on the X-ray molybdenum target image and the type-B ultrasonic image of the mammary gland of a same patient according to a cascaded manner, and obtaining the characteristic vector of a mammary gland sample; afterwards using a characteristic weight adaptive selection method on the training process of a binary balanced decision tree SVM multi-class classification algorithm based on a Gaussian kernel; and finally utilizing the characteristic weight adaptive selection method on the identification process of the binary balanced decision tree SVM multi-class classification algorithm based on the Gaussian kernel. The computer-aided mammy gland diagnosing method can improve accuracy and efficiency in breast cancer diagnosis.

Description

A kind of method applying the adaptively selected computer-aided diagnosis mammary gland of feature weight

Technical field

The present invention relates to Feature Engineering technical field, particularly a kind of application feature weight self adaptation The method of the computer-aided diagnosis mammary gland selected.

Background technology

Breast carcinoma is to occur at one of most common malignant tumor in women colony.China in recent years Investigation display, the occurrence rate of breast carcinoma is in cumulative year after year.Therefore the morning of breast carcinoma is improved Phase diagnostic accuracy becomes more and more meaningful.

At present, the main method that breast cancer diagnosis uses is by mammary gland x-ray, B ultrasonic image Deng image check, the state of an illness is analyzed by diagnosis person by the image feature such as calcification or lump. But density and focal zone close due to soft tissues such as the body of gland in mammary gland tissue, blood vessel, fat Degree all very close to, add the factors such as diagnosis person's visual fatigue so that the mistaken diagnosis of breast carcinoma of early stage and Fail to pinpoint a disease in diagnosis and the most often occur.Along with Medical Imaging Technology and the development of computer technology, utilize Computer carries out assisting diagnosis to be possibly realized；Such as: utilize digital image processing techniques, respectively Extract pathology in breast sonography, x-ray molybdenum target image relevant feature, use the machines such as SVM Device learning method carries out Classification and Identification etc. according to these features to Diagnosis of Breast tumors.

After major part research worker has extracted galactophore image feature, directly apply different SVM Sorting algorithm is classified more.But, the mammary X-ray molybdenum target of same patient and B ultrasonic image number According to deficiency is individually present, between the two or have data collision, diagnose more comprehensively and effectively for providing As a result, with ultrasonic image data, the mammary X-ray molybdenum target of same patient can be combined data to divide Analysis processes, by the mammary X-ray molybdenum target data fusion mutual with ultrasonic image data with mutual Mend, strengthening evidence, find the minimal disease that naked eyes cannot distinguish, and improve breast cancer diagnosis Accuracy rate, reduces misdiagnosis rate, rate of missed diagnosis.

Additionally, by the analysis to y-bend balancing decision tree SVM multi-classification algorithm, illustrate two The SVM multi-classification algorithm of fork balancing decision tree is based on the superiority in SVM multi-classification algorithm.Enter One step analysis understands, and the N sorting algorithm of the SVM of y-bend balancing decision tree comprises N-1 certainly altogether Plan face, for each decision surface, this is determined by each characteristic dimension of mammary gland sampling feature vectors The significance level of plan face classification is different, some characteristic dimension and classification strong correlation, some feature dimensions Spending weak relevant to classification, also some characteristic dimension is uncorrelated with classification.For each decision surface Speech, an effective feature selection approach directly affects svm classifier result, and then affects mammary gland The degree of accuracy of cancer diagnosis.At present, the problem solving feature selection has " firmly selecting " method, as Sequence forward selection procedures, sequence backward selection method, decision tree.The SVM using said method divides Class device performance can obtain a certain degree of raising, if but " firmly selecting " to be applied to y-bend and put down Problem of both the SVM multi-classification algorithm existence of weighing apparatus decision tree: the first, y-bend is balanced For each decision surface of decision tree SVM multi-classification algorithm, its important characteristic dimension is different, The characteristic dimension using " firmly selecting " method to extract is not appropriate for each decision surface；The second, screening Characteristic dimension out participates in calculating with identical significance level, it is impossible to reflect difference more accurately Significance level between decision surface kind characteristic dimension, to such an extent as to y-bend balancing decision tree cannot be guaranteed The performance of SVM multi-classification algorithm, therefore " selects " method to be not suitable for y-bend balancing decision tree firmly SVM multi-classification algorithm.Therefore, best settlement mechanism is to provide a kind of feature weight self adaptation Characteristic dimension is extracted in system of selection, improves the accuracy rate of svm classifier.

Therefore, the application is based on said method, to the mammary X-ray molybdenum target extracted and B ultrasonic image After data characteristics merges, build the many disaggregated models of SVM based on binary balance tree, by spy Levy weight adaptive selection method for many points of y-bend balancing decision tree SVM based on gaussian kernel The training process of class algorithm and identification proof procedure, thus provide valuable for clinical diagnosis " advisory opinion ", improves accuracy rate and the efficiency of breast cancer diagnosis.

Summary of the invention

In view of this, the purpose of the present invention is to propose to a kind of apply feature weight adaptively selected The method of computer-aided diagnosis mammary gland, improves accuracy rate and the efficiency of breast cancer diagnosis.

The present invention uses below scheme to realize: a kind of calculating applying feature weight adaptively selected The method of machine auxiliary Diagnosis of Breast, specifically includes following steps:

Step S1: extract mammary X-ray molybdenum target and B ultrasonic image data feature from known cases； By complete each case data of feature extraction according to known clinical diagnosis result carry out good pernicious with Clinical stages, marks；Described mark is divided into five classes: I level optimum, pernicious, pernicious II level, Pernicious III level, pernicious IV level；

Step S2: the x-ray molybdenum target image of same patient's mammary gland is used cascade with B ultrasonic image Mode carries out multiple features fusion, obtains the characteristic vector of mammary gland sample: each feature that will extract Cascade, then remove discordance data, obtain final characteristic vector；

Step S3: choose y-bend balancing decision tree SVM based on gaussian kernel as breast disease diagnosis Many disaggregated models, are used for y-bend based on gaussian kernel balance by feature weight adaptive selection method The training process of decision tree SVM multi-classification algorithm；

Step S4: feature weight adaptive selection method is used for y-bend based on gaussian kernel and puts down The identification process of weighing apparatus decision tree SVM multi-classification algorithm.

Further, described step S1 specifically includes following steps；

Step S11: given medium-scale above breast sonography focal area, a mammary gland X Line molybdenum target focal area image set；Described medium-scale this image set indicated above at least contains Breast sonography diagnostic images more than 250 width and mammary X-ray molybdenum target diagnostic image；

Step S12: by described breast sonography focal area, mammary X-ray molybdenum target focal area figure Image set is divided into training set and checking collection, and respectively from breast sonography focal area image and mammary gland X Line molybdenum target focal area image extracts corresponding feature；From mammary X-ray molybdenum target focal area Image zooming-out goes out edge, density, the form of calcification, the distribution of calcification；From B ultrasonic focal area The shape of image zooming-out tumor, plumpness, border, edge, echo mode, rear echo, Blood flow resistance index；Each case data of training set are carried out good according to known clinical diagnosis result Pernicious and clinical stages, marks.

Further, described step S2 specifically includes following steps:

Step S21: set up an effective y-bend balancing decision tree SVM；

Step S22: each for y-bend balancing decision tree SVM effective described in step S21 Individual decision surface, arranges suitable gaussian kernel parameter；

Step S23: each for y-bend balancing decision tree SVM effective described in step S21 Individual decision surface, uses feature weight adaptive selection method to solve the feature weight of this decision surface Vector s；

Step S24: use feature weight vector s that the training sample of this decision surface is weighted, logical Cross two points of SVM to be trained.

Further, described step S21 is: builds mammary gland data y-bend based on gaussian kernel and puts down Weighing apparatus decision tree SVM, from the beginning of root node, is first divided into two parts by classification, and each part is Intermediate node or be leafy node, then recurrence decomposes intermediate node until being leafy node；

Specifically include following steps:

Step S211: the 5 Ge Leilei centers marked in calculation procedure S1 respectively；

Step S212: find out two classes of 5 apoplexy due to endogenous wind: pernicious IV level is designated as c1, optimum Being designated as c2, the two Lei Lei center has maximum Euclidean distance；

Step S213: the two class in step S212 is respectively labeled as C1 class bunch and C2 Class bunch；

Step S214: in the middle of remaining 3 classes, selects there is minimum Euclidean with C1 class bunch The class of distance and described C1 class bunch are merged into a class, and are labeled as C1 class bunch, recalculate C1 Lei Culei center；

Step S215: in the middle of remaining 2 classes, selects there is minimum Euclidean with C2 class bunch The class of distance and described C2 class bunch are merged into a class, and are labeled as C2 class bunch, again count Calculate C2 Lei Culei center；

Step S216: cycle calculations S214 to S215, until 5 class distribution terminate.

Further, in described step S22, described suitable gaussian kernel parameter is set particularly as follows: Each decision surface arranges identical gaussian kernel parameter, sets described gaussian kernel parameter lambda=1.

Further, in described step S23, to the feature weight vector s's of each decision surface Solve tool to comprise the following steps:

Step S231: initialize: s_m=1/l, setting value θ, set maximum iteration time H=100, arranges gaussian kernel parameter lambda=1；Wherein, constant l represents the feature dimensions of mammary gland sample Number, θ is user-defined parameter and 0≤θ≤1, may be set to 0.1；

Step S232: calculating target function:

$J\left(s\right)=-\frac{1}{2}\underset{i,j}{\mathrm{\Σ}}{a}_i^{*}{a}_j^{*}{y}_i{y}_{j}Exp(-\mathrm{\λ}\underset{k=1}{\overset{l}{\mathrm{\Σ}}}{s}_k^2{\left(x_{ik}-x_{jk}\right)}^2)+\underset{i}{\mathrm{\Σ}}{a}_i^{*};$

Wherein y_iFor the classification of i-th sample, y_i∈{-1,+1}；x_ikFor i-th sample kth Characteristic component；a_i ^*For Lagrange multiplier, 0≤a_i ^*≤ C, constant C are penalty factor, a_i ^*Profit Minimizing method SMO by sequence to solve, feature weight vector s is as nuclear parameter；

Step S233: if not reaching the average of front ten times including this of object function J (s) The end condition that no longer reduces or do not reach maximum iteration time H, continues step S234, no Then jump out circulation；

Step S234: calculateAnd then calculate renewal direction vector D_m, converting equation is:

$D_m=\left\{\begin{array}{cc}0& s_m\≤\mathrm{\υ}\∩(\∂J/\∂s_m-\∂J/\∂s_{\mathrm{\μ}})>0\\ -\∂J/\∂s_m+\∂J/\∂s_{\mathrm{\μ}}& s_m>\mathrm{\υ}\∩m\≠\mathrm{\μ}\\ \underset{v\≠\mathrm{\μ},s_v>0}{\mathrm{\Σ}}(\∂J/\∂s_v-\∂J/\∂s_{\mathrm{\μ}})& m=\mathrm{\μ}\end{array}\right.;$

Wherein, s_mRepresent the m-th component of feature weight vector s,υ's Size needs user's sets itself, rule of thumb sets υ=0.0001；

Step S235: calculateγ_max=-s_v/D_v；Wherein, γ_maxFor Maximum step-length

S236: make s=s+ θ γ_maxD；

S237: return step S233.

Further, described step S4 specifically includes following steps:

Step S41: to identifying that sample is weighted according to the feature weight s of each decision surface；

Step S42: use the SVM recognizer of standard that the characteristic vector after weighting is carried out Identify.

Compared with prior art, the present invention utilizes the adaptively selected algorithm of feature weight to based on height The mammary gland feature of the y-bend balancing decision each decision surface of tree SVM of this core is weighted, and is to compare The new y-bend balancing decision tree svm classifier method for breast disease diagnosis.The method realizes spirit Live, there is stronger practicality, improve accuracy rate and the efficiency of breast cancer diagnosis.

Accompanying drawing explanation

Fig. 1 is that the x-ray molybdenum target image of same patient's mammary gland is owned by the embodiment of the present invention with B ultrasonic image Feature carries out the principle schematic of Feature Fusion.

Fig. 2 is that the embodiment of the present invention builds mammary gland y-bend balancing decision tree SVM principle based on gaussian kernel Schematic diagram.

Detailed description of the invention

Below in conjunction with the accompanying drawings and embodiment the present invention will be further described.

Present embodiments provide a kind of area of computer aided applying feature weight adaptively selected to examine The method of ablactation gland, specific as follows:

A () extracts mammary X-ray molybdenum target and B ultrasonic image data feature, ginseng from known cases According to BI-RADS standard, x-ray molybdenum target principal character includes lump, calcification and structural distortion. Lump describes again in terms of three: edge (clear, fuzzy, little leaflet, infiltration, asterism shape), shape State (circle, oval, leaflet shape, irregular shape), density (high, etc., low-density and containing fat Fat density)；Calcification be divided into from form typical optimum calcification, betweenness calcification (suspicious calcification), The calcification 3 kinds that high malignancy is possible；The distribution of calcification includes following 5 kinds of modes: (1) fill the air or It is dispersed in distribution (3) tufted distribution (4) line sample distribution (5) the section sample distribution of distribution (2) area-shaped. B ultrasonic image is mainly from the shape (regular, irregular) of tumor, plumpness (full, insatiable hunger Completely), border (clear, unintelligible), edge (finishing, not finishing), echo mode are (all Even, uneven), the aspect such as rear echo (strengthen, constant, decay and mix) levies As describing.It addition, it is generally acknowledged blood flow resistance index (RI) ＞ 0.70 of malignant change.Will Complete each case data of feature extraction according to known clinical diagnosis result carry out good pernicious with face Bed mark by stages (optimum, pernicious I level, pernicious II level, pernicious III level, pernicious IV level)；

B () as it is shown in figure 1, concentrate the x-ray molybdenum of same patient's mammary gland to training set and checking Target image and all features of B ultrasonic image carry out Feature Fusion, and concrete grammar is as follows: will extract Each feature cascade, then remove discordance data, obtain final characteristic vector.

C () randomly selects labeled data 80% adaptively selected for feature based weight The training process of y-bend balancing decision tree SVM multi-classification algorithm:

I) as in figure 2 it is shown, build mammary gland y-bend balancing decision tree SVM based on gaussian kernel.

Ii) for each decision surface of tree, the adaptively selected Algorithm for Solving of feature weight is used Go out the feature weight vector s of this decision surface.

Iii) feature weight vector s is used to weight, the training sample of this decision surface by two SVM is divided to be trained.

D the 20% of () residue labeled data is used for the y-bend that feature based weight is adaptively selected The identification proof procedure of balancing decision tree SVM multi-classification algorithm:

I) to identifying that sample is weighted according to the feature weight s of each decision surface.

Characteristic vector after weighting is known by the SVM recognizer ii) using standard Not.

Concrete, the present embodiment offer following steps:

S01: given medium-scale above breast sonography focal area, a mammary X-ray molybdenum Target focal area image set, this image set at least contains the breast sonography diagnostic graph of more than 250 width The mammary X-ray molybdenum target diagnostic image of picture sum.

S02: breast lesion administrative division map image set is divided into training set and checking collection, and respectively from breast Gland B ultrasonic focal area image and mammary X-ray molybdenum target focal area image zooming-out go out corresponding spy Levy, from mammary X-ray molybdenum target focal area image zooming-out go out edge, density, the form of calcification, The distribution of calcification；From the shape of B ultrasonic image focal area image zooming-out tumor, plumpness, limit Boundary, edge, echo mode, rear echo, blood flow resistance index (RI).By training set Each case data carry out good pernicious (good with clinical stages mark according to known clinical diagnosis result Property, pernicious I level, pernicious II level, pernicious III level, pernicious IV level).

S03: as it is shown in figure 1, training set and checking to be concentrated the x-ray molybdenum of same patient's mammary gland Target image and all features of B ultrasonic image carry out Feature Fusion, and concrete grammar is as follows: will extract Each feature cascade, then remove discordance data, obtain final characteristic vector.

S04: build mammary gland data y-bend balancing decision tree SVM based on gaussian kernel, saves from root Point starts, and classification is first divided into two parts, and each part is intermediate node or is leafy node, Recurrence decomposes intermediate node until being leafy node again.Leafy node be classification number be the knot of 1 Point, as in figure 2 it is shown, achievement process herein uses BDT_SVM method.

S05: for the characteristic vector of each decision surface, selects penalty factor=100, step-length Multiplier θ=0.1, maximum iteration time H=100, set gaussian kernel parameter lambda=1, pass through feature Weight adaptive selection method RF_SA is used for solving feature weight vector s.

S06: to each decision surface, application solves the feature weight vector s obtained to this decision-making The training sample weighting in face, is trained by two points of SVM.

S07: identify that process is as follows: first to identifying that sample is weighed according to the feature of each decision surface Weight s is weighted, and finally uses the SVM recognizer of standard to the characteristic vector after weighting It is identified.Final recognition result is exactly this classification belonging to identification sample.

In the present embodiment, described an effective mammary gland data y-bend balancing decision tree is set up Method specific as follows:

S011: calculate respectively marked 5 class (optimum, pernicious I level, pernicious II level, Pernicious III level, pernicious IV level) class center；

S012: find out two classes of 5 apoplexy due to endogenous wind: pernicious IV level is designated as c1, optimum is designated as c2, The two Lei Lei center has maximum Euclidean distance；

S013: the two class is respectively labeled as C1 class bunch and C2 class bunch；

S014: in the middle of remaining 3 classes, selects a class: pernicious III level (such With the Euclidean distance that C1 class bunch has minimum), it is labeled as C1 class bunch, is merged into a class, and And recalculate such class center；

S015: in the middle of remaining 2 classes, selects a class: pernicious I level (such with C2 class bunch has the Euclidean distance of minimum), it is labeled as C2 class bunch, is merged into a class, and Recalculate such class center；

S016: cycle calculations S014, S015, when 5 classes distribute end.

In the present embodiment, described feature weight adaptive selection method RF_SA is concrete such as Under:

Set all of mammary gland feature weight component s_i>=0, for the characteristic dimension of strong correlation of classifying s_iShould be relatively big, for weak relevant characteristic dimension s_iShould be less, and for the incoherent spy that classifies Levy dimension, for the characteristic vector at place " not essential ", now corresponding weight s_iShould It is 0.As a example by the decision surface of root node, other decision surfaces by that analogy: specifically include following Step:

Step S231: initialize: s_m=1/l, setting value θ, set maximum iteration time H=100, arranges gaussian kernel parameter lambda=1；Wherein, constant l represents the feature dimensions of mammary gland sample Number, θ is user-defined parameter and 0≤θ≤1, may be set to 0.1；

Step S232: calculating target function:

$J\left(s\right)=-\frac{1}{2}\underset{i,j}{\mathrm{\Σ}}{a}_i^{*}{a}_j^{*}{y}_i{y}_{j}Exp(-\mathrm{\λ}\underset{k=1}{\overset{l}{\mathrm{\Σ}}}{s}_k^2{\left(x_{ik}-x_{jk}\right)}^2)+\underset{i}{\mathrm{\Σ}}{a}_i^{*};$

Wherein y_iFor the classification of i-th sample, y_i∈{-1,+1}；x_ikFor i-th sample kth Characteristic component；a_i ^*For Lagrange multiplier, 0≤a_i ^*≤ C, constant C are penalty factor, a_i ^*Profit Minimizing method SMO by sequence to solve, feature weight vector s is as nuclear parameter；

Step S233: if not reaching the average of front ten times including this of object function J (s) The end condition that no longer reduces or do not reach maximum iteration time H, continues step S234, no Then jump out circulation；

Step S234: calculateOrderAnd then calculate renewal direction vector D_m, converting equation is:

$D_m=\left\{\begin{array}{cc}0& s_m\≤\mathrm{\υ}\∩(\∂J/\∂s_m-\∂J/\∂s_{\mathrm{\μ}})>0\\ -\∂J/\∂s_m+\∂J/\∂s_{\mathrm{\μ}}& s_m>\mathrm{\υ}\∩m\≠\mathrm{\μ}\\ \underset{v\≠\mathrm{\μ},s_v>0}{\mathrm{\Σ}}(\∂J/\∂s_v-\∂J/\∂s_{\mathrm{\μ}})& m=\mathrm{\μ}\end{array}\right.;$

Wherein, s_mRepresent the m-th component of feature weight vector s,υ's Size needs user's sets itself, rule of thumb sets υ=0.0001；

Step S235: calculateγ_max=-d_v/D_v；Wherein, γ_maxFor Maximum step-length

S236: make s=s+ θ γ_maxD；

S237: return step S233.

The foregoing is only presently preferred embodiments of the present invention, all according to scope of the present invention patent institute Impartial change and the modification done, all should belong to the covering scope of the present invention.

Claims (7)

1. the method applying the adaptively selected computer-aided diagnosis mammary gland of feature weight, its It is characterised by: comprise the following steps:

Step S1: extract mammary X-ray molybdenum target and B ultrasonic image data feature from known cases； By complete each case data of feature extraction according to known clinical diagnosis result carry out good pernicious with Clinical stages, marks；Described mark is divided into five classes: I level optimum, pernicious, pernicious II level, Pernicious III level, pernicious IV level；

Step S2: the x-ray molybdenum target image of same patient's mammary gland is used cascade with B ultrasonic image Mode carries out multiple features fusion, obtains the characteristic vector of mammary gland sample: each feature that will extract Cascade, then remove discordance data, obtain final characteristic vector；

Step S3: choose y-bend balancing decision tree SVM based on gaussian kernel as breast disease diagnosis Many disaggregated models, are used for y-bend based on gaussian kernel balance by feature weight adaptive selection method The training process of decision tree SVM multi-classification algorithm；

Step S4: feature weight adaptive selection method is used for y-bend based on gaussian kernel and puts down The identification process of weighing apparatus decision tree SVM multi-classification algorithm.

A kind of computer aided manufacturing applying feature weight adaptively selected the most according to claim 1 The method helping Diagnosis of Breast, it is characterised in that: described step S1 specifically includes following steps；

Step S11: given medium-scale above breast sonography focal area, a mammary gland X Line molybdenum target focal area image set；Described medium-scale this image set indicated above at least contains Breast sonography diagnostic images more than 250 width and mammary X-ray molybdenum target diagnostic image；

Step S12: by described breast sonography focal area, mammary X-ray molybdenum target focal area figure Image set is divided into training set and checking collection, and respectively from breast sonography focal area image and mammary gland X Line molybdenum target focal area image extracts corresponding feature；From mammary X-ray molybdenum target focal area Image zooming-out goes out edge, density, the form of calcification, the distribution of calcification；From B ultrasonic focal area The shape of image zooming-out tumor, plumpness, border, edge, echo mode, rear echo, Blood flow resistance index；Each case data of training set are carried out good according to known clinical diagnosis result Pernicious and clinical stages, marks.

A kind of computer aided manufacturing applying feature weight adaptively selected the most according to claim 1 The method helping Diagnosis of Breast, it is characterised in that: described step S2 specifically includes following steps:

Step S21: set up an effective y-bend balancing decision tree SVM；

Step S22: each for y-bend balancing decision tree SVM effective described in step S21 Individual decision surface, arranges suitable gaussian kernel parameter；

Step S23: each for y-bend balancing decision tree SVM effective described in step S21 Individual decision surface, uses feature weight adaptive selection method to solve the feature weight of this decision surface Vector s；

Step S24: use feature weight vector s that the training sample of this decision surface is weighted, logical Cross two points of SVM to be trained.

A kind of computer aided manufacturing applying feature weight adaptively selected the most according to claim 3 The method helping Diagnosis of Breast, it is characterised in that: described step S21 is: build based on gaussian kernel Mammary gland data y-bend balancing decision tree SVM, from the beginning of root node, first classification is divided into two Part, each part is intermediate node or for leafy node, then recurrence decompose intermediate node until For leafy node；

Specifically include following steps:

Step S211: the 5 Ge Leilei centers marked in calculation procedure S1 respectively；

Step S212: find out two classes of 5 apoplexy due to endogenous wind: pernicious IV level is designated as c1, optimum Being designated as c2, the two Lei Lei center has maximum Euclidean distance；

Step S213: the two class in step S212 is respectively labeled as C1 class bunch and C2 Class bunch；

Step S214: in the middle of remaining 3 classes, selects there is minimum Euclidean with C1 class bunch The class of distance and described C1 class bunch are merged into a class, and are labeled as C1 class bunch, recalculate C1 Lei Culei center；

Step S215: in the middle of remaining 2 classes, selects there is minimum Euclidean with C2 class bunch The class of distance and described C2 class bunch are merged into a class, and are labeled as C2 class bunch, again count Calculate C2 Lei Culei center；

Step S216: cycle calculations S214 to S215, until 5 class distribution terminate.

A kind of computer aided manufacturing applying feature weight adaptively selected the most according to claim 3 The method helping Diagnosis of Breast, it is characterised in that: in described step S22, described arrange suitably Gaussian kernel parameter, particularly as follows: each decision surface arranges identical gaussian kernel parameter, sets described Gauss Nuclear parameter λ=1.

A kind of computer aided manufacturing applying feature weight adaptively selected the most according to claim 3 The method helping Diagnosis of Breast, it is characterised in that: in described step S23, to each decision surface The tool that solves of feature weight vector s comprises the following steps:

Step S231: initialize: s_m=1/l, setting value θ, set maximum iteration time H=100, arranges gaussian kernel parameter lambda=1；Wherein, constant l represents the feature dimensions of mammary gland sample Number, θ is user-defined parameter and 0≤θ≤1, may be set to 0.1；

Step S232: calculating target function:

$J\left(s\right)=-\frac{1}{2}\underset{i,j}{\Σ}{a}_i^{*}{a}_j^{*}{y}_i{y}_{j}Exp(-\mathrm{\λ}\underset{k=1}{\overset{l}{\Σ}}{s}_k^2{\left(x_{ik}-x_{jk}\right)}^2)+\underset{i}{\Σ}{a}_i^{*};$

Wherein y_iFor the classification of i-th sample, y_i∈{-1,+1}；x_ikFor i-th sample kth Characteristic component；a_i ^*For Lagrange multiplier, 0≤a_i ^*≤ C, constant C are penalty factor, a_i ^*Profit Minimizing method SMO by sequence to solve, feature weight vector s is as nuclear parameter；

Step S233: if not reaching the average of front ten times including this of object function J (s) The end condition that no longer reduces or do not reach maximum iteration time H, continues step S234, no Then jump out circulation；

Step S234: calculateAnd then calculate renewal direction vector D_m, converting equation is:

$D_m=\left\{\begin{array}{cc}0& s_m\≤\mathrm{\υ}\∩(\∂J/\∂s_m-\∂J/\∂s_{\mathrm{\μ}})>0\\ -\∂J/\∂s_m+\∂J/\∂s_{\mathrm{\μ}}& s_m>\mathrm{\υ}\∩m\≠\mathrm{\μ}\\ \underset{v\≠\mathrm{\μ},s_v>0}{\Σ}(\∂J/\∂s_v-\∂J/\∂s_{\mathrm{\μ}})& m=\mathrm{\μ}\end{array}\right.;$

Wherein, s_mRepresent the m-th component of feature weight vector s,υ's Size needs user's sets itself, rule of thumb sets υ=0.0001；

Step S235: calculateWherein, γ_maxFor Maximum step-length

S236: make s=s+ θ γ_maxD；

S237: return step S233.

A kind of computer aided manufacturing applying feature weight adaptively selected the most according to claim 1 The method helping Diagnosis of Breast, it is characterised in that: described step S4 specifically includes following steps:

Step S41: to identifying that sample is weighted according to the feature weight s of each decision surface:；

Step S42: use the SVM recognizer of standard that the characteristic vector after weighting is carried out Identify.