CN104143101A - Method for automatically identifying breast tumor area based on ultrasound image - Google Patents

Abstract

The invention discloses a method for automatically identifying a breast tumor area based on an ultrasound image. The method comprises the following steps of acquiring the ultrasound image of the breast, and preprocessing the ultrasound image; segmenting the ultrasound image subjected to preprocessing through an image segmentation method to obtain a plurality of segmented subareas; extracting a grey level histogram, texture features, gradient features and morphological features of the ultrasound image, and combining the grey level histogram, the texture features, the gradient features and the morphological features of the ultrasound image with two-dimensional position information to obtain high-dimensionality feature vectors; selecting the most effective feature subset of the high-dimensionality feature vectors through feature ordering based on biclustering and a selection method; performing learning classification on the selected most effective feature subset through a classifier, and then automatically identifying the breast tumor area. By means of the method, the breast tumor area can be identified automatically from segment results of the breast tumor ultrasound image, therefore, automation performance of computer-aided diagnosis is improved, manual operation of clinical doctors is reduced, and subjective influence of clinical doctors is reduced.

Description

A kind of tumor of breast region automatic identifying method based on ultrasonoscopy

Technical field

The present invention relates to computer-aided diagnosis field, particularly a kind of tumor of breast region automatic identifying method based on ultrasonoscopy.

Background technology

Nowadays, breast cancer has become one of principal disease threatening women's life, however the mankind to its concrete pathology also in unknown state, early find that early treatment has just become the major way being effectively treated.Along with social development and the progress of technology, ultrasonic imaging becomes one of main detection mode of tumor of breast with its non-intruding, low, the simple and reliable feature of cost.

Yet relying on the diagnosis of ultrasonoscopy is that experience is dependent, inevitably can be subject to clinician's subjective factor impact, thereby can greatly reduces accuracy and the reliability of diagnostic result.Therefore,, for fear of the impact of subjective factor and accuracy and the reliability of raising diagnostic result, a lot of achievements in research has been put forward out and obtained to the technology of computer-aided diagnosis.

In the Mammogram Analysis system based on ultrasonoscopy having proposed, after ultrasonoscopy is cut apart, conventionally need clinician's artificial selection from all subregion being partitioned into go out tumor region, just can proceed follow-up analysis operation, this has seriously reduced the robotization performance of computer-aided diagnosis, and has introduced clinician's subjectivity.

Therefore people need a kind of new tumor region recognition methods to overcome the shortcoming existing in prior art.

Summary of the invention

The shortcoming that the object of the invention is to overcome prior art, with not enough, provides a kind of tumor of breast region automatic identifying method based on ultrasonoscopy.

Object of the present invention realizes by following technical scheme:

A tumor of breast region automatic identifying method based on ultrasonoscopy, the step that comprises following order:

S1. obtain breast ultrasound image, and it is carried out to pre-service;

S2. use image segmentation algorithm to cut apart pretreated ultrasonoscopy, obtain a plurality of subregions of cutting apart;

S3. extract grey level histogram, textural characteristics, Gradient Features, the morphological feature of ultrasonoscopy, add two-dimensional position information, obtain high-dimensional proper vector;

S4. to high-dimensional proper vector, by the feature ordering based on double focusing class and system of selection, choose the most effective character subset, specifically comprise the following steps:

(1) extract double focusing class Seed Points: the element in each is listed as is separately done cluster, and the result of each cluster is regarded as the Seed Points of a double focusing class; For the data matrix of the capable C row of a given R, apply a kind of graduation clustering algorithm that condenses to each row of objective matrix;

(2) heuristic structure double focusing class: utilize a kind of didactic method that the double focusing class Seed Points extracting in the first step is done and expanded, according to predefined criterion structure double focusing class;

(3) eliminate the double focusing class of redundancy: first, the column number ascending sort that the set of double focusing class is comprised according to each double focusing class; Then, from the double focusing class that comprises minimum column number, start iteration, detect the subset whether it belongs to the double focusing class coming below, if so, eliminate this double focusing class;

(4) feature ordering and selection: after data matrix is extracted to double focusing class, need to from these double focusing classes, extract the information that can weigh each feature, calculate the corresponding double focusing class of each feature score value, score value is higher, show that this feature more can express this data set, tightr with contacting of further feature, then according to needed number of features, select the character subset that score value is the highest, be the most effective character subset;

S5. select after the most effective character subset, use sorter to carry out learning classification to it, can reach the tumor of breast region object of identification automatically.

In step S1, described pre-service adopts full Variation Model, reaches the object of filtering and noise reduction by minimizing full variation, and full variation is defined as

$\mathrm{TV}\left[u\right]=\underset{\mathrm{\Ω}}{\∫}|\▿u|\mathrm{dx}$

Wherein, Ω represents continuous signal domain, presentation video gradient, dx represents an element of Ω.

Owing to will not damaging the local edge of image in level and smooth speckle noise, selected and strengthened all highly effective full Variation Models (Total Variation, TV) to eliminating speckle noise and edge.

In step S2, a kind of in described image segmentation algorithm threshold method, clustering procedure, maximum a posteriori, Markov random field method, dynamic programming, Bayes's level set method, the partitioning algorithm based on graph theory.

In step S4 (1) step by step, described a kind of cohesion graduation clustering algorithm is specific as follows:

A, calculate maximal value and the minimum value of the distance of all raw data;

B, each element of all raw data is initialized as to an independent class;

C, set up apart from array, record the distance between all kinds of, be initialized as the distance of each raw data;

D, all data are done to following circulation: find the minimum value apart from array, if this value is less than threshold tau, these two classes are classified as to a class, recalculate apart from array, until be more than or equal to τ apart from the minimum value of array, stop circulation;

After cohesion graduation cluster, will obtain a series of double focusing class Seed Points of objective matrix:

[C _s(i,j),N _cl(j)]＝HC(j,τ),j＝1,...,C

Wherein, HC representative cohesion graduation clustering algorithm, τ is the default threshold value of clustering algorithm, C _s(i, j) represents the Seed Points of i double focusing class of j row, N _cl(j) represent the total number of j row double focusing class Seed Points.

In step S4 (2) step by step, described does and expands double focusing class Seed Points, and its method comprises following steps:

A, the line number comprising according to each Seed Points are done ascending sort;

B, from comprising minimum line number R _jseed Points start, to the corresponding row of other row expansions, form a R _jthe matrix M of row C row;

C, for matrix M, calculate its all square residue score:

$\mathrm{MSRS}=\frac{1}{\left|R\right|\left|C\right|}\underset{i\∈R,j\∈C}{\mathrm{\Σ}}{(e_{\mathrm{ij}}-e_{\mathrm{iC}}-e_{\mathrm{Rj}}+e_{\mathrm{RC}})}^2$

$e_{\mathrm{iC}}=\frac{1}{\left|C\right|}\underset{j\∈C}{\mathrm{\Σ}}{e}_{\mathrm{ij}},e_{\mathrm{Rj}}=\frac{1}{\left|R\right|}\underset{i\∈R}{\mathrm{\Σ}}{e}_{\mathrm{ij}},e_{\mathrm{RC}}=\frac{1}{\left|R\right|\left|C\right|}\underset{i\∈R,j\∈C}{\mathrm{\Σ}}{e}_{\mathrm{ij}}$

Wherein, R is the line number of matrix, and C is matrix column number, e _ijit is the element value of the capable j row of this matrix i; If the MSRS calculating is greater than the threshold value of setting, for each point in matrix, row or the row at this place are removed in calculating, calculate the MSRS of new matrix, the difference that must divide with new and old matrix M SRS is weighed this point and is expert at or is listed as the contribution margin to MSRS, deletes row or the row at the some place of MSRS contribution margin maximum; Continuous this process of iteration, until the MSRS value of matrix is less than the threshold value presetting, this matrix is a double focusing class of finally trying to achieve.

(4) step by step of described step S4, specific as follows:

A, data matrix is extracted to double focusing class after, need to from these double focusing classes, extract the information that can weigh each feature: the feature ordering scheme based on double focusing class has been considered the correlativity of feature and two factors of the separation property of sample, wherein the correlativity of feature represents by correlativity score, for weighing the tightness degree contacting between the feature of character subset, the separation property of sample represents by separation property score, for weighing the independent degree of a feature; For k feature, if it appears at double focusing class subset Z _kin arbitrary double focusing class in, the correlativity score of this feature and classification score are expressed as follows respectively:

$\mathrm{correlation}\_\mathrm{score}=\underset{i=1}{\overset{n_{b,k}}{\mathrm{\Σ}}}\frac{n_{f,k}\left(i\right)}{n_c}$

$\mathrm{separability}\_\mathrm{score}=\frac{n_{s,k}}{n_r}\sqrt{\underset{i=1}{\overset{n_{b,k}}{\mathrm{\Σ}}}{({\mathrm{\μ}}_{i,k}-{\mathrm{\μ}}_{a,k})}^2/n_{b,k}}$

Wherein, n _b,krepresent Z _kthe number of middle double focusing class, n _f,k(i) represent Z _kin the number of the feature that comprises of i double focusing class, n _s,krepresent Z _kin the number of unduplicated row, μ _i,krepresent Z _kin the average of k feature of i double focusing class, μ _a,krepresent μ _i,kaverage;

The final score of k feature divides value representation by double focusing class:

$\mathrm{bicluster}\_\mathrm{score}=a\·\stackrel{\‾}{\mathrm{correlation}\_\mathrm{score}}+\stackrel{\‾}{\mathrm{separability}\_\mathrm{score}}$

Wherein, with represent respectively correlation_score and separability_score normalization result; A is a constant, the contribution to final score for balance correlation_score and separability_score;

The bicluster_score score value of feature is higher, shows that this feature more can express this data set, tightr with contacting of further feature;

B, according to needed number of features, select the character subset that score value is the highest.

In step S5, a kind of in described sorter K nearest neighbor classifier, neural network classifier, support vector machine classifier.

Compared with prior art, tool has the following advantages and beneficial effect in the present invention:

1, groundwork principle of the present invention is: first with machine learning method, tumor region and non-tumor region are learnt, after image is cut apart, used the sorter training to carry out Classification and Identification to tumor region and non-tumor region, automatically identify tumor region; Utilize method provided by the present invention, can from the segmentation result of Ultrasound Image of Breast Tumor, automatically identify tumor region, thereby improve the robotization performance of computer-aided diagnosis, reduce clinician's manual operations, reduce clinician's subjectivity impact.

2,, due to complicated factors such as imaging circumstances and image-forming principles, ultrasonoscopy has the features such as geneogenous speckle noise, many shades and low contrast, makes the oeverall quality of ultrasonoscopy poor.The present invention carries out the pre-service such as denoising before ultrasonoscopy is cut apart to ultrasonoscopy, suppress the harmful effect that speckle noise brings.

3, grey level histogram, textural characteristics, Gradient Features, morphological feature are chosen in feature extraction of the present invention, be to show that these features have stronger ability to express for the tumor of breast in ultrasonoscopy because pain is crossed research, can know and describe tumor region and the characteristic of non-tumor region and the similarities and differences between the two.

4, the present invention carries out the reason of feature selecting: research shows, with suitable method, reduce the dimension of feature, remove the feature of redundancy poor efficiency, select to have most the Feature Combination of ability to express, not only can not reduce the accuracy rate of Classification and Identification, but also can improve to a certain extent classification performance.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of a kind of tumor of breast region automatic identifying method based on ultrasonoscopy of the present invention;

Fig. 2 is GLCM example calculation (a) data matrix (b) the GLCM schematic diagram of method described in Fig. 1;

Fig. 3 is the double focusing class Seed Points expansion schematic diagram of method described in Fig. 1;

Fig. 4 is that the automatic recognition accuracy of method described in Fig. 1 is with the variation diagram of feature set size.

Embodiment

Below in conjunction with embodiment and accompanying drawing, the present invention is described in further detail, but embodiments of the present invention are not limited to this.

Adopt all kinds of programming languages such as C language all can realize well the present invention.

As shown in Figure 1, a kind of tumor of breast region automatic identifying method based on ultrasonoscopy, comprises following five steps:

Step (1) pre-service: due to complicated factors such as imaging circumstances and image-forming principles, ultrasonoscopy has the features such as geneogenous speckle noise, many shades and low contrast, makes the oeverall quality of ultrasonoscopy poor.Therefore, before ultrasonoscopy is cut apart, need to carry out the pre-service such as denoising, suppress the harmful effect that speckle noise brings.Owing to will not damage the local edge of image in level and smooth speckle noise, the present invention has selected and has strengthened highly effective full Variation Model (Total Variation to eliminating speckle noise and edge, TV), by minimizing full variation, reach the object of filtering and noise reduction, full variation is defined as

$\mathrm{TV}\left[u\right]=\underset{\mathrm{\Ω}}{\∫}|\▿u|\mathrm{dx}$

Wherein, Ω represents continuous signal domain, presentation video gradient, dx represents an element of Ω.

Step (2) image is cut apart: use image segmentation algorithm to cut apart pretreated ultrasonoscopy, obtain numerous subregions of cutting apart.Available image segmentation algorithm has a lot, as: threshold method, clustering procedure, maximum a posteriori, Markov random field method, dynamic programming, Bayes's level set method and the partitioning algorithm based on graph theory etc.

Step (3) feature extraction: research shows, grey level histogram, textural characteristics, Gradient Features, morphological feature etc. have stronger ability to express to the tumor of breast in ultrasonoscopy, in order to be described clearly the characteristic of tumor region and non-tumor region and the similarities and differences between the two, the present invention selects above-mentioned several category features to extract, add the positional information of bidimensional, can obtain the proper vector of 73 dimensions.Specific as follows:

Grey level histogram: if the gray shade scale of Ultrasound Image of Breast Tumor is divided into L level, can explain by following probability distribution formula for arbitrary grey level histogram of cutting apart subregion:

p _i＝n _i/N,p _i≥0,i＝1,2,...,L

Wherein, the pixel number that N is this subregion, n _iit is the pixel number of i gray level.Yet, directly from grey level histogram, obtain the feature of image region directly perceived not, therefore calculate the first-order statistics amount of following several conventional Description Image grey level histograms as the quantization characteristic of subregion:

(1) gray average:

$\mathrm{\μ}=\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}i\·p\left(i\right)$

(2) gray standard deviation:

${\mathrm{\σ}}^2=\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}{(i-\mathrm{\μ})}^{2}p\left(i\right)$

(3) measure of skewness:

$\mathrm{Skew}=\frac{1}{{\mathrm{\σ}}^2}\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}{(i-\mathrm{\μ})}^{3}p\left(i\right)$

(4) kurtosis:

$\mathrm{Kurtosis}=\frac{1}{{\mathrm{\σ}}^4}\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}{(i-\mathrm{\μ})}^{4}p\left(i\right)-3$

(5) energy:

$\mathrm{Energy}=\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}p{\left(i\right)}^2$

(6) entropy:

$\mathrm{Entropy}=-\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}p\left(i\right){\log }_{2}p\left(i\right)$

Textural characteristics: gray level co-occurrence matrixes is a kind of method of comparatively conventional Description Image textural characteristics, what in fact it described is the space dependency characteristic between pixel grey scale in image.That each point in gray level co-occurrence matrixes records is θ at an angle to each other, and the point that phase mutual edge distance is d, to the probability occurring in whole subregion, can be explained with following formula:

p(i,j,d,θ)＝#{(x ₁,y ₁)(x ₂,y ₂)|f(x ₁,y ₁)＝i,f(x ₂,y ₂)＝j,

|(x ₁,y ₁)-(x ₂,y ₂)|＝d,∠((x ₁,y ₁)(x ₂,y ₂))＝θ}

Wherein, (x ₁, y ₁), (x ₂, y ₂) be respectively two points of centering positions in subregion, f (x _n, y _n) be the gray shade scale of relevant position pixel, ∠ represents two angles between point, # represents that the point that meets described condition is to the total degree occurring in whole subregion.Fig. 2 is an example that calculates gray level co-occurrence matrixes, θ=0 wherein, d=1.

Yet, in actual applications, the seldom direct textural characteristics using gray level co-occurrence matrixes as image, but as asking grey level histogram feature, calculate some statistics and carry out representation feature.Make p _ijrepresent the capable j column element of i value in gray level co-occurrence matrixes, L represents the gray shade scale of dividing, and conventional statistic expression formula is as follows:

(1) maximum probability:

Max?Pro＝maxp _ij

(2) diversity factor:

$\mathrm{Dissimilarity}=\underset{i,j=0}{\overset{L-1}{\mathrm{\Σ}}}{p}_{\mathrm{ij}}|i-j|$

(3) contrast:

$\mathrm{Contrast}=\underset{i,j=0}{\overset{L-1}{\mathrm{\Σ}}}{p}_{i,j}{(i-j)}^2$

(4) homogeneity degree:

$\mathrm{Homogeneity}=\underset{i,j=0}{\overset{L-1}{\mathrm{\Σ}}}\frac{p_{\mathrm{ij}}}{1+{(i-j)}^2}$

(5) unfavourable balance square:

$\mathrm{Inverse}=\underset{i,j=0}{\overset{L-1}{\mathrm{\Σ}}}\frac{p_{i,j}}{1+|i-j|/L}$

(6) energy:

$\mathrm{Energy}=\sqrt{\underset{i,j=0}{\overset{L-1}{\mathrm{\Σ}}}{p_{\mathrm{ij}}}^2}$

(7) entropy:

$\mathrm{Entropy}=-\underset{i,j=0}{\overset{L-1}{\mathrm{\Σ}}}{p}_{\mathrm{ij}}\ln p_{\mathrm{ij}}$

(8) average:

$\mathrm{Mean}=\underset{i,j}{\overset{L-1}{\mathrm{\Σ}}}{\mathrm{ip}}_{\mathrm{ij}}$

(9) standard deviation:

$\mathrm{sd}=\sqrt{\underset{i,j=0}{\overset{L-1}{\mathrm{\Σ}}}{p}_{\mathrm{ij}}{(i-\mathrm{Mean})}^2}$

(10) degree of correlation:

$\mathrm{Correlation}=\underset{i,j=0}{\overset{L-1}{\mathrm{\Σ}}}\frac{p_{\mathrm{ij}}(i-\mathrm{Mean})(j-\mathrm{Mean})}{{\mathrm{Sd}}^2}$

Following statistic need to be used the marginal probability distribution of gray level co-occurrence matrixes, the expression formula of four kinds of marginal probability distributions respectively:

$p_x\left(i\right)=\underset{j}{\mathrm{\Σ}}{p}_{\mathrm{ij}}$

$p_y\left(j\right)=\underset{i}{\mathrm{\Σ}}{p}_{\mathrm{ij}}$

$p_{x+y}(k=i+j)=\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{p}_{\mathrm{ij}}$

$p_{x-y}(k=|i-j\left|\right)=\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{p}_{\mathrm{ij}}$

And average (11):

$\mathrm{SumMean}=\underset{i=0}{\overset{2L-2}{\mathrm{\Σ}}}{\mathrm{ip}}_{x+y}\left(i\right)$

And standard deviation (12):

$\mathrm{SumSd}=\sqrt{\underset{i=0}{\overset{2L-2}{\mathrm{\Σ}}}{(i-\mathrm{SumMean})}^2{p}_{x+y}\left(i\right)}$

And entropy (13):

$\mathrm{SumEntropy}=-\underset{i=0}{\overset{2L-2}{\mathrm{\Σ}}}{p}_{x+y}\left(i\right){\log }_2{p}_{x+y}\left(i\right)$

(14) differ from average:

$\mathrm{DifMean}=\underset{i=0}{\overset{2L-2}{\mathrm{\Σ}}}{\mathrm{ip}}_{x-y}\left(i\right)$

(15) differ from standard deviation:

$\mathrm{DifSd}=\sqrt{\underset{i=0}{\overset{2L-2}{\mathrm{\Σ}}}{(i-\mathrm{DifMean})}^2{p}_{x-y}\left(i\right)}$

And entropy (16):

$\mathrm{SumEntropy}=-\underset{i=0}{\overset{2L-2}{\mathrm{\Σ}}}{p}_{x-y}\left(i\right){\log }_2{p}_{x-y}\left(i\right)$

(17) edge entropy:

$\mathrm{HX}=\mathrm{HY}=-\underset{i}{\mathrm{\Σ}}{p}_x\left(i\right){\log }_2{p}_x\left(i\right)$

(18) associating edge entropy:

$\mathrm{HXY}1=-\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{p}_{\mathrm{ij}}{\log }_2\left(p_x\left(i\right)p_y\left(j\right)\right)$

(19) separate edge entropy:

$\mathrm{HXY}2=-\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{p}_x\left(i\right)p_y\left(i\right){\log }_2\left(p_x\left(i\right)p_y\left(j\right)\right)$

(20) information correlation:

$\mathrm{INFH}1=\frac{\mathrm{Entropy}-\mathrm{HXY}1}{\max (\mathrm{HX},\mathrm{HY})}$

$\mathrm{INFH}2=\sqrt{1-\exp (-2(\mathrm{HXY}2-\mathrm{Entropy}))}$

Histogram of gradients:

(1) gradient amplitude and the gradient direction of each pixel in the piece of zoning.For compute gradient amplitude, need to first try to achieve VG (vertical gradient) and horizontal gradient:

VG (vertical gradient):

f _h(x,y)＝f(x+1,y)-f(x-1,y)

Horizontal gradient:

f _v(x,y)＝f(x,y+1)-f(x,y-1)

Wherein, f (x, y) is illustrated in the gray-scale value of the pixel of position (x, y).On this basis, gradient amplitude and gradient direction calculate respectively by following formula:

Gradient amplitude:

$m(x,y)=\sqrt{f_h{(x,y)}^2+f_v{(x,y)}^2}$

Gradient direction:

$\mathrm{\θ}(x,y)=\arctan \frac{f_h(x,y)}{f_v(x,y)}$

(2) histogram generates.Through the calculating of the first step, each pixel can calculate corresponding gradient amplitude and gradient direction, is similar to grey level histogram, the cumulative gradient amplitude with the pixel of identical gradient direction, thus generate the histogram based on gradient.In order to reduce the complexity of calculating and to improve arithmetic speed, the orientation average of 360 degree is quantified as K interval, and the gradient direction dropping between same zone is considered to identical.After having added up, this histogram of gradients can represent with a K dimensional vector.

(3) K dimensional vector is done to normalization.K dimensional vector after normalization is exactly last for describing the feature of histogram of gradients.

Shape facility: in current research, comparatively conventional and effectively shape facility mainly contain following four kinds: girth, area, complex-shaped degree and like circularity.

Complex-shaped degree and seemingly circularity are tried to achieve by following formula respectively:

ShapeComplexity＝Perimeter ²/Area

CircularDegree＝4π·Area/Perimenter ²

Wherein Perimeter and Area represent respectively subregion girth and the area after normalization.

Position feature: by following expression formula, calculate,

L _x＝1-2·|w/2-x|/w

L _y＝1-2·|h/2-y|/h

Wherein, w and h represent respectively width and the height of ultrasonoscopy, and (x, y) represents the center position of subregion.If subregion present position more approaches the central point of ultrasonoscopy, the value that its position feature calculates is larger.

Step (4) feature selecting: after feature extraction, can extract more high-dimensional characteristic information from Ultrasound Image of Breast Tumor.Yet research shows, reduces the dimension of feature with suitable method, removes the feature of redundancy poor efficiency, selects to have most the Feature Combination of ability to express, not only can not reduce the accuracy rate of Classification and Identification, but also can improve to a certain extent classification performance.Feature selection approach has a lot, as Variance Score, Laplacian Score and Fisher Score etc., the character subset that the present invention has selected a kind of feature ordering based on double focusing class and system of selection to choose the most applicable specific classification object, concrete operations are as follows:

According to the present invention for specific classification object, adopt to extract the algorithm of double focusing class, what this pair of clustering algorithm extracted is the double focusing quasi-mode that column element is equal.This double focusing class extraction algorithm mainly comprises three steps:

(1) extract double focusing class Seed Points: the element in each is listed as is separately done cluster, and the result of each cluster is regarded as the Seed Points of a double focusing class.For the data matrix of the capable C row of a given R, apply a kind of cohesion graduation clustering algorithm (Agglomerative Hierarchical Clustering) to each row of objective matrix.The specific descriptions of cohesion graduation clustering algorithm are as follows:

1) calculate maximal value and the minimum value of the distance of all raw data

2) each element of all raw data is initialized as to an independent class

3) set up apart from array, record the distance between all kinds of, be initialized as the distance of each raw data

4) all data are done to following circulation: find the minimum value apart from array, if this value is less than threshold tau, these two classes are classified as to a class, recalculate apart from array.Until be more than or equal to τ apart from the minimum value of array, stop circulation.

After cohesion graduation cluster, will obtain a series of double focusing class Seed Points of objective matrix:

[C _s(i,j),N _cl(j)]＝HC(j,τ),j＝1,...,C

Wherein, what HC represented is cohesion graduation clustering algorithm, and τ is the default threshold value of clustering algorithm, C _s(i, j) represents the Seed Points of i double focusing class of j row, N _cl(j) represent the total number of j row double focusing class Seed Points.

(2) heuristic structure double focusing class: utilize a kind of didactic method that the double focusing class Seed Points extracting in the first step is done and expanded, according to predefined criterion structure double focusing class.Expansion algorithm is as described below:

1) line number comprising according to each Seed Points is done ascending sort.

2) from comprising minimum line number R _jseed Points start, to the corresponding row of other row expansions, form a R _jthe matrix M of row C row.

3), for matrix M, calculate its all square residue score (Mean Square Residue Score, MSRS):

$\mathrm{MSRS}=\frac{1}{\left|R\right|\left|C\right|}\underset{i\∈R,j\∈C}{\mathrm{\Σ}}{(e_{\mathrm{ij}}-e_{\mathrm{iC}}-e_{\mathrm{Rj}}+e_{\mathrm{RC}})}^2$

$e_{\mathrm{iC}}=\frac{1}{\left|C\right|}\underset{j\∈C}{\mathrm{\Σ}}{e}_{\mathrm{ij}},e_{\mathrm{Rj}}=\frac{1}{\left|R\right|}\underset{i\∈R}{\mathrm{\Σ}}{e}_{\mathrm{ij}},e_{\mathrm{RC}}=\frac{1}{\left|R\right|\left|C\right|}\underset{i\∈R,j\∈C}{\mathrm{\Σ}}{e}_{\mathrm{ij}}$

Wherein, R is the line number of matrix, and C is matrix column number, e _ijit is the element value of the capable j row of this matrix i.If the MSRS calculating is greater than the threshold value of setting, for each point in matrix, row or the row at this place are removed in calculating, calculate the MSRS of new matrix, the difference that must divide with new and old matrix M SRS is weighed this point and is expert at or is listed as the contribution margin to MSRS, deletes row or the row at the some place of MSRS contribution margin maximum.Continuous this process of iteration, until the MSRS value of matrix is less than the threshold value presetting, this matrix is a double focusing class of finally trying to achieve.Fig. 3 has provided double focusing class Seed Points expansion schematic diagram.

(3) remove the double focusing class of redundancy: owing to having generated a large amount of double focusing classes in second step, and the situation that exists the double focusing class of some smaller redundancies to be comprised by larger double focusing class.The double focusing class of removing redundancy is very necessary, because the double focusing class repeating can cause the double counting while assessing the separation property of sample and the correlativity of feature in subsequent characteristics rank scheme.Eliminating the double focusing class of redundancy carries out in two steps: first, and the column number ascending sort that the set of double focusing class is comprised according to each double focusing class; Then, from the double focusing class that comprises minimum column number, start iteration, detect the subset whether it belongs to the double focusing class coming below, if so, eliminate this double focusing class.

After data matrix is extracted to double focusing class, need to from these double focusing classes, extract the information that can weigh each feature.Feature rank scheme based on double focusing class has been considered the correlativity of feature and two factors of the separation property of sample, wherein the correlativity of feature represents by correlativity score, for weighing the tightness degree contacting between the feature of character subset, the separation property of sample represents by separation property score, for weighing the independent degree of a feature.For k feature, if it appears at double focusing class subset Z _kin arbitrary double focusing class in, the correlativity score of this feature and classification score are expressed as follows respectively:

$\mathrm{correlation}\_\mathrm{score}=\underset{i=1}{\overset{n_{b,k}}{\mathrm{\Σ}}}\frac{n_{f,k}\left(i\right)}{n_c}$

$\mathrm{separability}\_\mathrm{score}=\frac{n_{s,k}}{n_r}\sqrt{\underset{i=1}{\overset{n_{b,k}}{\mathrm{\Σ}}}{({\mathrm{\μ}}_{i,k}-{\mathrm{\μ}}_{a,k})}^2/n_{b,k}}$

Wherein, n _b,krepresent Z _kthe number of middle double focusing class, n _f,k(i) represent Z _kin the number of the feature that comprises of i double focusing class, n _s,krepresent Z _kin the number of unduplicated row, μ _i,krepresent Z _kin the average of k feature of i double focusing class, μ _a,krepresent μ _i,kaverage.

The final score of k feature divides value representation by double focusing class:

$\mathrm{bicluster}\_\mathrm{score}=a\·\stackrel{\‾}{\mathrm{correlation}\_\mathrm{score}}+\stackrel{\‾}{\mathrm{separability}\_\mathrm{score}}$

Wherein, with represent respectively correlation_score and separability_score normalization result; A is a constant, the contribution to final score for balance correlation_score and separability_score;

The bicluster_score score value of feature is higher, shows that this feature more can express this data set, tightr with contacting of further feature.

Described step (5) Classification and Identification: select after the most effective character subset, use sorter to carry out learning classification to it, can reach the tumor of breast region object of identification automatically.Conventional sorter has: K nearest neighbor classifier (K-nearest Neighbor, KNN), neural network classifier (Neural Networks, NNs) and support vector machine classifier (Support Vector Machine, SVM) etc.

In order to verify correctness of the present invention and validity, done following experiment:

In experiment, use VC++6.0 to realize the present invention.Experiment verifies based on 46 Ultrasound Image of Breast Tumors (malignant tumour and benign tumour each 23), and the size of every image is 400 * 300.These 46 Ultrasound Image of Breast Tumor Shi You Tumor Hospital Attached to Zhongshan Univ. provide, by Philip medical ultrasonic image instrument HDI5000SonoCT (transducer: L12-550mm broadband linear array transducer; Frequency: 7.1MHz) obtain.Experiment test host parameter is: system is Windows732 bit manipulation system, and processor is AMD5000+ (2.60GHz), inside saves as Kingston (2.00GB).

Ultrasound Image of Breast Tumor is after image is cut apart, and each subregion in image is all regarded as a sample instance, and each sample instance is marked as tumour (positive sample) or non-tumour (negative sample), represents respectively with 1 and-1; Then, each subregion is extracted to the five category feature information such as grey level histogram, gray level co-occurrence matrixes, histogram of gradients, shape and position, each sample instance can extract the feature of 73 dimensions altogether, each sample instance can with one 74 dimension vector representation (1 dimension classification information and 73 dimension characteristic informations).All sample instance are extracted after features, used 5 folding cross validations to carry out algorithm evaluation: all sample sets are equally divided into 5 group data sets at random, choose wherein one group as test set, remaining sample set is as training set; Iteration five times, until each group data set all selected mistake as test set.For each proof procedure, first the feature selection approach based on double focusing class is applied to training set, double focusing class score value according to feature sorts, then the character subset that is 1～73 to size is respectively tested on test set, by its classification performance, verifies correctness of the present invention and validity.

Choose accuracy (Accuracy), susceptibility (Sensibility) and specificity (Specificity) as quantitative evaluation index.Wherein, accuracy is illustrated in the ratio that the sample of correctly being classified in all samples accounts for total sample, and susceptibility represents that the sample of correctly being classified in positive sample accounts for the ratio of positive sample, and specificity represents that the sample of correctly being classified in negative sample accounts for the ratio of negative sample.Be defined as follows respectively:

$\mathrm{Accuracy}=\frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{FP}+\mathrm{TN}+\mathrm{FN}}$

$\mathrm{Sensibility}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$

$\mathrm{Specificity}=\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}$

Wherein, TP represents the number that positive sample is correctly validated, and TN represents the number that negative sample is correctly validated, and FN represents that positive sample is erroneously identified as the number of negative sample, and FP represents that negative sample is erroneously identified as the number of positive sample.

Fig. 4 is that automatic recognition accuracy is with the variation diagram of character subset size.As can be seen from this figure, accuracy rate does not improve along with the increase of number of features, illustrates in primitive character space and has some redundancy feature information, and these features not only can not be improved the effect of classification, have reduced on the contrary the accuracy rate of automatic identification; Consider three indexs such as accuracy, susceptibility and specificity, as can be seen from Figure 4, when selecting front 25 dimensional features when (mainly comprising position feature, shape facility and gray level co-occurrence matrixes), the accuracy rate of identification is the highest automatically.

Following table has provided accuracy, susceptibility and the specificity of automatically identifying under the condition of selecting front 25 dimensional features, can find out, three indexs, all more than 97%, have absolutely proved correctness of the present invention and validity.

Accuracy	Susceptibility	Specificity
			0.983±0.013	0.985±0.019	0.974±0.035

Above-described embodiment is preferably embodiment of the present invention; but embodiments of the present invention are not restricted to the described embodiments; other any do not deviate from change, the modification done under Spirit Essence of the present invention and principle, substitutes, combination, simplify; all should be equivalent substitute mode, within being included in protection scope of the present invention.

Claims (7)

1. the tumor of breast region automatic identifying method based on ultrasonoscopy, is characterized in that the step that comprises following order:

S1. obtain breast ultrasound image, and it is carried out to pre-service;

S2. use image segmentation algorithm to cut apart pretreated ultrasonoscopy, obtain a plurality of subregions of cutting apart;

S3. extract grey level histogram, textural characteristics, Gradient Features, the morphological feature of ultrasonoscopy, add two-dimensional position information, obtain high-dimensional proper vector;

S4. to high-dimensional proper vector, by the feature ordering based on double focusing class and system of selection, choose the most effective character subset, specifically comprise the following steps:

(1) extract double focusing class Seed Points: the element in each is listed as is separately done cluster, and the result of each cluster is regarded as the Seed Points of a double focusing class; For the data matrix of the capable C row of a given R, apply a kind of graduation clustering algorithm that condenses to each row of objective matrix;

(2) heuristic structure double focusing class: utilize a kind of didactic method that the double focusing class Seed Points extracting in the first step is done and expanded, according to predefined criterion structure double focusing class;

(3) eliminate the double focusing class of redundancy: first, the column number ascending sort that the set of double focusing class is comprised according to each double focusing class; Then, from the double focusing class that comprises minimum column number, start iteration, detect the subset whether it belongs to the double focusing class coming below, if so, eliminate this double focusing class;

(4) feature ordering and selection: after data matrix is extracted to double focusing class, need to from these double focusing classes, extract the information that can weigh each feature, calculate the corresponding double focusing class of each feature score value, score value is higher, show that this feature more can express this data set, tightr with contacting of further feature, then according to needed number of features, select the character subset that score value is the highest, be the most effective character subset;

S5. select after the most effective character subset, use sorter to carry out learning classification to it, can reach the tumor of breast region object of identification automatically.

2. the tumor of breast region automatic identifying method based on ultrasonoscopy according to claim 1, is characterized in that: in step S1, described pre-service adopts full Variation Model, reaches the object of filtering and noise reduction by minimizing full variation, and full variation is defined as

$\mathrm{TV}\left[u\right]=\underset{\mathrm{\Ω}}{\∫}|\▿u|\mathrm{dx}$

Wherein, Ω represents continuous signal domain, presentation video gradient, dx represents an element of Ω.

3. the tumor of breast region automatic identifying method based on ultrasonoscopy according to claim 1, it is characterized in that: in step S2, a kind of in described image segmentation algorithm threshold method, clustering procedure, maximum a posteriori, Markov random field method, dynamic programming, Bayes's level set method, the partitioning algorithm based on graph theory.

4. the tumor of breast region automatic identifying method based on ultrasonoscopy according to claim 1, is characterized in that: in step S4 (1) step by step, described a kind of cohesion graduation clustering algorithm is specific as follows:

A, calculate maximal value and the minimum value of the distance of all raw data;

B, each element of all raw data is initialized as to an independent class;

C, set up apart from array, record the distance between all kinds of, be initialized as the distance of each raw data;

D, all data are done to following circulation: find the minimum value apart from array, if this value is less than threshold tau, these two classes are classified as to a class, recalculate apart from array, until be more than or equal to τ apart from the minimum value of array, stop circulation;

After cohesion graduation cluster, will obtain a series of double focusing class Seed Points of objective matrix:

[C _s(i,j),N _cl(j)]＝HC(j,τ),j＝1,...,C

Wherein, HC representative cohesion graduation clustering algorithm, τ is the default threshold value of clustering algorithm, C _s(i, j) represents the Seed Points of i double focusing class of j row, N _cl(j) represent the total number of j row double focusing class Seed Points.

5. the tumor of breast region automatic identifying method based on ultrasonoscopy according to claim 1, is characterized in that: in step S4 (2) step by step, described does and expand double focusing class Seed Points, and its method comprises following steps:

A, the line number comprising according to each Seed Points are done ascending sort;

B, from comprising minimum line number R _jseed Points start, to the corresponding row of other row expansions, form a R _jthe matrix M of row C row;

C, for matrix M, calculate its all square residue score:

$\mathrm{MSRS}=\frac{1}{\left|R\right|\left|C\right|}\underset{i\∈R,j\∈C}{\mathrm{\Σ}}{(e_{\mathrm{ij}}-e_{\mathrm{iC}}-e_{\mathrm{Rj}}+e_{\mathrm{RC}})}^2$

$e_{\mathrm{iC}}=\frac{1}{\left|C\right|}\underset{j\∈C}{\mathrm{\Σ}}{e}_{\mathrm{ij}},e_{\mathrm{Rj}}=\frac{1}{\left|R\right|}\underset{i\∈R}{\mathrm{\Σ}}{e}_{\mathrm{ij}},e_{\mathrm{RC}}=\frac{1}{\left|R\right|\left|C\right|}\underset{i\∈R,j\∈C}{\mathrm{\Σ}}{e}_{\mathrm{ij}}$

Wherein, R is the line number of matrix, and C is matrix column number, e _ijit is the element value of the capable j row of this matrix i; If the MSRS calculating is greater than the threshold value of setting, for each point in matrix, row or the row at this place are removed in calculating, calculate the MSRS of new matrix, the difference that must divide with new and old matrix M SRS is weighed this point and is expert at or is listed as the contribution margin to MSRS, deletes row or the row at the some place of MSRS contribution margin maximum; Continuous this process of iteration, until the MSRS value of matrix is less than the threshold value presetting, this matrix is a double focusing class of finally trying to achieve.

6. the tumor of breast region automatic identifying method based on ultrasonoscopy according to claim 1, is characterized in that: (4) step by step of described step S4, specific as follows:

A, data matrix is extracted to double focusing class after, need to from these double focusing classes, extract the information that can weigh each feature: the feature ordering scheme based on double focusing class has been considered the correlativity of feature and two factors of the separation property of sample, wherein the correlativity of feature represents by correlativity score, for weighing the tightness degree contacting between the feature of character subset, the separation property of sample represents by separation property score, for weighing the independent degree of a feature; For k feature, if it appears at double focusing class subset Z _kin arbitrary double focusing class in, the correlativity score of this feature and classification score are expressed as follows respectively:

$\mathrm{correlation}\_\mathrm{score}=\underset{i=1}{\overset{n_{b,k}}{\mathrm{\Σ}}}\frac{n_{f,k}\left(i\right)}{n_c}$

$\mathrm{separability}\_\mathrm{score}=\frac{n_{s,k}}{n_r}\sqrt{\underset{i=1}{\overset{n_{b,k}}{\mathrm{\Σ}}}{({\mathrm{\μ}}_{i,k}-{\mathrm{\μ}}_{a,k})}^2/n_{b,k}}$

Wherein, n _b,krepresent Z _kthe number of middle double focusing class, n _f,k(i) represent Z _kin the number of the feature that comprises of i double focusing class, n _s,krepresent Z _kin the number of unduplicated row, μ _i,krepresent Z _kin the average of k feature of i double focusing class, μ _a,krepresent μ _i,kaverage;

The final score of k feature divides value representation by double focusing class:

$\mathrm{bicluster}\_\mathrm{score}=a\·\stackrel{\‾}{\mathrm{correlation}\_\mathrm{score}}+\stackrel{\‾}{\mathrm{separability}\_\mathrm{score}}$

Wherein, with represent respectively correlation_score and separability_score normalization result; A is a constant, the contribution to final score for balance correlation_score and separability_score;

The bicluster_score score value of feature is higher, shows that this feature more can express this data set, tightr with contacting of further feature;

B, according to needed number of features, select the character subset that score value is the highest.

7. the tumor of breast region automatic identifying method based on ultrasonoscopy according to claim 1, is characterized in that: in step S5, and a kind of in described sorter K nearest neighbor classifier, neural network classifier, support vector machine classifier.