CN102509273A - Tumor segmentation method based on homogeneous pieces and fuzzy measure of breast ultrasound image - Google Patents

Abstract

The invention discloses a tumor segmentation method based on homogeneous pieces and fuzzy measure a breast ultrasound image, which comprises the following steps: (1), inputting the breast ultrasound image to be segmented; (2), margin detection for the image to be segmented; (3), defining the homogeneous piece; (4), defining a weighting function, considering the homogeneous pieces of each image pixel point as a fuzzy set, defining the fuzzy measure for measuring the diversity of the fuzzy set based on the margin information, the texture information and the space information, and defining the weighting function; and (5), achieving tumor segmentation. The invention has the advantages that (1), through defining the homogeneous pieces, the problem of selecting among the neighborhood windows in a traditional method is solved, the accuracy of the area statistics information is ensured, and the segmentation precision is improved; and (2), through the fuzzy measure and the homogeneous pieces, the weighting function is defined, the partial image misclassification is basically eliminated, the precision of the breast tumor segmentation is improved, and the favorable robustness is demonstrated at the same time.

Description

Lesion segmentation approach based on the breast ultrasound image of homogeneity sheet and fuzzy mearue

Technical field

The present invention relates to the Medical Ultrasound Image Processing technical field, particularly relate to a kind of lesion segmentation approach of the breast ultrasound image based on homogeneity sheet and fuzzy mearue.

Background technology

Breast cancer is one of modal malignant tumour of women, and its M & M accounts for the first place in women's diseases.Early stage diagnosis and treatment are the keys that improves the survival rates of breast cancer.Ultrasonic examination with its do not have wound, do not have special contraindication, repeatable strong, characteristics such as cost are widely used in the auxiliary detection of tumor of breast.Yet its accuracy often depends on doctor's clinical experience, has very strong subjectivity.In order to improve the objectivity of diagnosis, press for the development computer-aided diagnosis system in the clinical practice.

It is the important component part of computer-aided diagnosis system that ultrasonoscopy is cut apart.Yet, remain a pendent global problem cutting apart automatically of ultrasonoscopy so far owing to the existence of speckle noise, the pseudo-shadow of decay in the ultrasonoscopy makes.Particularly have under the characteristic situation such as similar gray scale or texture, very easily produce serious mistake branch phenomenon, caused bigger obstacle for cutting apart automatically of ultrasonoscopy at tumour and normal surrounding tissue.

Be a good selection semi-automatic the cutting apart of ultrasonoscopy.Movable contour model (Snake) is to use semi-automatic partition method the most widely.Along with the development of Level Set Method, the researcher has proposed the movable contour model based on level set, reaches the purpose of cutting apart through curve evolvement.Wherein geodesic line active contour (GAC) model is typical representative, and the GAC model has accurate edge positioning, the algorithm stability advantage of higher.But this model only utilizes the gradient information of image, receives noise effect easily, is difficult to be partitioned into the homogeneous region in the ill-defined image.And when the object in the image had the border of recessed, the GAC model possibly make the evolution curve stop at a certain local minimum state of value, and was not consistent with the border of object.

Belaid is with substituting gradient information with gray scale and the irrelevant phase information of grey scale change in theory; Level set partitioning algorithm (Belaid A., Boukerroui D., et al based on phase place were proposed in 2011; Phase-Based Level Set Segmentation of Ultrasound Images.TransactionInformation Technology in Biomedicine; 2011,15 (1): 138-147), solved the pseudo-shadow problem in the ultrasonoscopy to a certain extent.Yet, cut apart for the tangible galactophore image of textural characteristics and can not obtain desirable segmentation result because its CONSIDERING EDGE information is not only utilized texture information and spatial information simultaneously.

Normalized Cut (NCut) method is one of important for image segmentation, and the advantage of NCut method is to have embodied well the part of image and the marriage relation of the overall situation, has avoided not having and has cut apart partially.Utilize NCut method research ultrasonoscopy segmentation problem; Through setting up suitable weighted graph; Can in the definition of weighting function, introduce various image feature informations (like area information, marginal information and prior imformation etc.); Adopt the various thoughts of cutting apart, can obtain segmentation result accurately, so in the cutting apart of natural image and MRI medical image, be used widely.Yu and Shi have proposed improved based on NCut method (Yu S.; Shi J..Segmentationgiven partial grouping constraints.IEEE Transactions on Pattern Analysisand Machine Intelligence; 2004,26 (2): 173-183).This algorithm makes destination object under a spot of interactive operation through introducing part cluster priori (such as significant characteristics of image and target location), can from background, separate effectively, has also utilized sampling policy to reduce computation burden simultaneously.This method has obtained better segmentation effect at natural image, has caused widely to pay close attention to.But the NCut method is cut apart for the breast ultrasound image, and there be limited evidence currently of is seen relevant report.Simultaneously; Said method is directly applied to second-rate ultrasonoscopy mainly have following problem in cutting apart: how (1) selects suitable neighborhood; Even if it is suitable that the neighborhood window size is selected; Fixedly neighborhood possibly crossed over different tissue regions, thereby is difficult to guarantee the homogeney of neighborhood of pixels, finally causes pixel to be divided by mistake.(2) how to define the better weight of ultrasonoscopy segmentation performance still unresolved.(3) how to guarantee lesion segmentation result's robustness.

Summary of the invention

To above-mentioned prior art; The technical matters that the present invention will solve is a definition homogeneity sheet; And define weighting function on this basis based on homogeneity sheet and fuzzy mearue, and a kind of tumour automatic division method of breast ultrasound image is proposed, it has improved the wrong phenomenon of dividing of breast ultrasound image local; Avoid the decaying influence of pseudo-shadow improves segmentation precision and robustness simultaneously.

In order to solve the problems of the technologies described above, the present invention adopts following technical scheme: a kind of lesion segmentation approach of the breast ultrasound image based on homogeneity sheet and fuzzy mearue comprises the steps:

(1) input breast ultrasound image to be split;

(2) treat split image and carry out rim detection, comprising:

(2.1) adopt the breast ultrasound image texture features of importing in filter set and the clustering algorithm extraction step (1), generate the texture primitive characteristic image;

(2.2) through fuzzy mearue the multiple information of each image slices vegetarian refreshments is combined definition rim detection function; Calculate the edge energy of each pixel, generate initial edge figure, and carry out edge optimization; Finally obtain the outline map of original image, thereby accomplish rim detection;

(3) definition homogeneity sheet; Utilize the outline map that generates in the step (2) to be each image pixel definition homogeneity sheet, said homogeneity sheet is a kind of neighborhood that reflects the neighbor similarity relation; The structure of homogeneity sheet comprises the steps:

(3.1) be that search window is confirmed at the center with the current pixel point;

(3.2) the alligatoring stage: figure goes up the straight line of crossing all directions of current point in this search window of scanning on the edge of; If the edge energy value of pixel is less than given threshold value on certain direction straight line; Think that then this pixel is candidate's homologous pints, otherwise, with its eliminating and stop the scanning of this direction straight line;

(3.3) elaboration phase: for candidate's homologous pints; Further take micronization processes to judge whether it is real homologous pints; Then the remaining pixel that is not excluded is considered to the member in this homogeneity sheet in the current search window, and exists with the form of probability;

(4) definition weighting function; The homogeneity sheet of each image slices vegetarian refreshments in the step (3) is regarded as a fuzzy set, on this basis, defines a fuzzy mearue that is used for weighing the fuzzy set diversity, and define weighting function with it according to marginal information, texture information and spatial information;

(5) lesion segmentation; Adopt the weighting function of step (4) definition, the similarity between calculating pixel is to make up the weight matrix of entire image; With the coarse position information of tumour as prior imformation; Simultaneously ultrasonoscopy is carried out random sampling and reduce computation complexity, upgrade weight matrix, the partition problem that has most of scheming is become restricted minimization problem according to prior imformation of introducing and Sampling Strategies; At last; Carry out interior inserting through the weighting of the sample point adjacent and estimate, obtain estimation, thereby realize cutting apart tumour to the breast ultrasound image segmentation with non-sample point.

Further, in the said step (2.1), filter set comprises one group of odd even wave filter and one group of DOG wave filter, and the odd even wave filter is to 1 yardstick and 0 °; 22.5 °, 45 °, 67.5 °, 90 °; 112.5 °, 135 °, the input picture of 157.5 ° of 8 directions carries out convolutional filtering; The DOG wave filter carries out convolutional filtering to the input picture of 1 yardstick, and the filter response of each pixel is one group of high dimension vector, through clustering algorithm the high dimension vector of all pixels in the image is carried out cluster; Each cluster centre is one type of texture primitive, and each pixel of input picture is endowed the label of an affiliated texture primitive type, then generates the texture primitive characteristic image.

Further, the rim detection function described in the step (2.2) is: $E(p,{\mathrm{\θ}}_t)={\mathrm{\Σ}}_{k=1}^K|E_k^L(p,{\mathrm{\θ}}_t)-E_k^R(p,{\mathrm{\θ}}_t)|+\frac{{(E_k^L(p,{\mathrm{\θ}}_t)-E_k^R(p,{\mathrm{\θ}}_t))}^2}{E_k^L(p,{\mathrm{\θ}}_t)+E_k^R(p,{\mathrm{\θ}}_t)}$

K representes texture primitive classification number, is being the center with pixel p, and radius is in the circular neighborhood of R, along θ _tDiameter on the direction is divided into left and right neighborhood with this neighborhood,

Be respectively this and put in the left neighborhood and the energy in the right neighborhood, wherein: $E_k^L(p,{\mathrm{\θ}}_t)={\mathrm{\μ}}_{{\mathrm{\η}}_p^L}\left(k\right)\×d_k^L(p,{\mathrm{\θ}}_t),$ $E_k^R(p,{\mathrm{\θ}}_t)={\mathrm{\μ}}_{{\mathrm{\η}}_p^R}\left(k\right)\×d_k^R(p,{\mathrm{\θ}}_t),$ ${\mathrm{\μ}}_{{\mathrm{\η}}_p^L}\left(k\right),$ ${\mathrm{\μ}}_{{\mathrm{\η}}_p^R}\left(k\right)$ Belong to the degree of membership of k class texture primitive with the interior pixel of right neighborhood in the left neighborhood of remarked pixel point p respectively, and satisfy constraint condition ${\mathrm{\Σ}}_{k=1}^K{\mathrm{\μ}}_{{\mathrm{\η}}_p^L}\left(k\right)=1,$ ${\mathrm{\Σ}}_{k=1}^K{\mathrm{\μ}}_{{\mathrm{\η}}_p^R}\left(k\right)=1;$ $d_k^L(p,{\mathrm{\θ}}_t),$ $d_k^R(p,{\mathrm{\θ}}_t)$ Be illustrated respectively in θ _tOn the direction, in the pixel p neighborhood left with it with right neighborhood in the similarity of the pixel that belongs to k class texture primitive,

Membership?

function expression is:

${\mathrm{\μ}}_{{\mathrm{\η}}_p^L}\left(k\right)=\frac{{\mathrm{\Σ}}_{i\∈L_p}\mathrm{\δ}(T\left(p_i\right)-k)}{\left|L_p\right|},$ ${\mathrm{\μ}}_{{\mathrm{\η}}_p^R}\left(k\right)=\frac{{\mathrm{\Σ}}_{i\∈R_p}\mathrm{\δ}(T\left(p_i\right)-k)}{\left|L_p\right|}$

L wherein _p, R _pRepresent the set of all pixels in left neighborhood and the right neighborhood respectively, | ● | the gesture of expression set, p _iBe any pixel in pixel p left side neighborhood or the right neighborhood, T (p _i) be p _iThe texture primitive category label, δ (g) is a Kronecker ' s delta function, if i.e. T (p _i)=k is δ (T (p then _i)-k)=1, otherwise δ (T (p _i)-k)=0;

Similarity?

function expression is:

$d_k^L(p,{\mathrm{\θ}}_t)=|\mathrm{OE}(p,{\mathrm{\θ}}_t)-m_k^L(p,{\mathrm{\θ}}_t)|,$ $d_k^R(p,{\mathrm{\θ}}_t)=|\mathrm{OE}(p,{\mathrm{\θ}}_t)-m_k^R(p,{\mathrm{\θ}}_t)|$

OE (p, θ wherein _t) be that pixel p is at θ _tOriented energy intensity on the direction, expression formula does

I is an input picture,

With

Be respectively that direction is θ _tEven symmetry and odd symmetry wave filter;

Be respectively at θ _tOn the direction, mean direction energy intensity in pixel p left side neighborhood and the right neighborhood,

The average energy intensity direction?

and?

The expression is:

$m_k^L(p,{\mathrm{\θ}}_t)=\frac{{\mathrm{\Σ}}_{i\∈L_p}\mathrm{OE}(p_i,{\mathrm{\θ}}_t)\mathrm{\δ}(T\left(p_i\right)-k)}{{\mathrm{\Σ}}_{i\∈L_p}\mathrm{\δ}(T\left(p_i\right)-k)},$

$m_k^R(p,{\mathrm{\θ}}_t)=\frac{{\mathrm{\Σ}}_{i\∈R_p}\mathrm{OE}(p_i,{\mathrm{\θ}}_t)\mathrm{\δ}(T\left(p_i\right)-k)}{{\mathrm{\Σ}}_{i\∈R_p}\mathrm{\δ}(T\left(p_i\right)-k)}.$

Further, said clustering algorithm is the K-Means algorithm.

Further, the search window described in the said step (3.1) confirm comprise the steps: that (x y) is the center, defines the search window that size is (2R+1) * (2R+1) a: N with the position of the pixel p on the outline map in image _p(i, j)=wherein, R is the height and the width of search window, and the borderline pixel of search window is divided into two parts for x-R≤i≤x+R, y-R≤j≤y+R}, wherein, $a_1^1(x-R,y+1),a_2^1(x-R,y),...,a_{4R+2}^1(x+R,y)$ Represent the pixel of window edge the first half, $a_1^2(x-R,y),a_2^2(x-R,y+1),...,a_{4R+2}^2(x+R,y-1)$ Represent the pixel of window edge the latter half, the pixel of the first half traveled through in the direction of the clock, to the latter half pixel by traversal counterclockwise;

The alligatoring stage described in the said step (3.2) comprises the steps; The pixel of the first half and the latter half is chosen three continuous consecutive point respectively

Obtain corresponding three lines with central point respectively

The center line of certain direction in the point by point scanning search window then

When running into net point p _j, calculate E _Max(p, p _j) value:

$E_{\max }(p,p_j)={\max }_{t\∈L(p,p_j)}{E}^{*}\left(t\right)$

Wherein, L (p, p _j) be to a p from a p _jThe set of all net points, E ^*(t) edge energy that obtains for the rim detection function calculation; If E _Max(p, p _j)≤τ sets up, and then thinks p _jBe candidate's homologous pints, and continue scanning center's line; Otherwise, think p _jNot candidate's homologous pints, and finish the scanning of center line;

Elaboration phase described in the said step (3.3) comprises the steps: for candidate's homologous pints p _jCarry out micronization processes, judge that according to the angle difference judgment criterion whether it is homologous pints, at first representes corresponding straight line, then pp with vector form _jAnd pa _iBetween angle θ _{J, i}Be defined as:

${\mathrm{\θ}}_{j,i}=\∠({\stackrel{\→}{\mathrm{pp}}}_j,{\stackrel{\→}{\mathrm{pa}}}_i)=\arccos \frac{{\left({\stackrel{\→}{\mathrm{pp}}}_j\right)}^T{\stackrel{\→}{\mathrm{pa}}}_i}{\left|\right|{\stackrel{\→}{\mathrm{pp}}}_j\left|\right|\·\left|\right|{\stackrel{\→}{\mathrm{pa}}}_i\left|\right|}$

Similarly, can calculate pp _jAnd pa _k(k=i-1, the angle theta between i+1) _{J, i-1}And θ _{I, i+1}, in order to get rid of the false homologous pints in candidate's homologous pints, define a kind of angle difference judgment criterion:

The then remaining some p that is not excluded _jBe exactly homogeneity sheet Ω _pIn the point.

Further, the weighting function in the step (4) is:

π wherein _p, π _qBe respectively homogeneity sheet Ω _pAnd Ω _qFuzzy set, r, σ are constant, d (π _p, π _q) be to be defined in two fuzzy mearues on the fuzzy set, expression formula is following:

$d({\mathrm{\π}}_p,{\mathrm{\π}}_q)={\mathrm{\Σ}}_{k=}^K{\mathrm{\λ}}_k{\mathrm{\μ}}_{{\mathrm{\π}}_p}\left(k\right)+(1-{\mathrm{\λ}}_k){\mathrm{\μ}}_{{\mathrm{\π}}_q}\left(k\right)-\min ({\mathrm{\μ}}_{{\mathrm{\π}}_p}\left(k\right),{\mathrm{\μ}}_{{\mathrm{\π}}_q}\left(k\right))$

Wherein, K representes texture primitive classification number; Fuzzy set π _pAnd π _qBe defined as respectively ${\mathrm{\π}}_p=\{(k,{\mathrm{\μ}}_{{\mathrm{\π}}_p}\left(k\right)),k=\mathrm{1,2},...,K\},$ ${\mathrm{\π}}_q=\{(k,{\mathrm{\μ}}_{{\mathrm{\π}}_q}\left(k\right)),k=\mathrm{1,2},...,K\};$ ${\mathrm{\μ}}_{{\mathrm{\π}}_p}\left(k\right),$ ${\mathrm{\μ}}_{{\mathrm{\π}}_q}\left(k\right)$ Represent homogeneity sheet Ω respectively _pAnd Ω _qIn pixel belong to k (degree of membership of 1≤k≤K) type of texture primitive be defined as

${\mathrm{\μ}}_{{\mathrm{\π}}_p}\left(k\right)=\frac{{\mathrm{\Σ}}_{j\∈{\mathrm{\Ω}}_p}m(p,p_j)\mathrm{\δ}(T\left(p_j\right)-k)}{{\mathrm{\Σ}}_{j\∈{\mathrm{\Ω}}_p}m(p,p_j)},$ ${\mathrm{\μ}}_{{\mathrm{\π}}_q}\left(k\right)=\frac{{\mathrm{\Σ}}_{j\∈{\mathrm{\Ω}}_q}m(q,q_j)\mathrm{\δ}(T\left(q_j\right)-k)}{{\mathrm{\Σ}}_{j\∈{\mathrm{\Ω}}_q}m(q,q_j)}$

δ in the formula (g) is a Kronecker ' s delta function, m (p, p _j) be pixel p in the step (3) _jRefuse degree of receiving, λ _kBe homogeneity sheet Ω _pAnd Ω _qBetween the regulation and control coefficient, its formula is:

${\mathrm{\λ}}_k=\frac{H_k\left(p\right)}{H_k\left(p\right)+H_k\left(q\right)}$

H in the formula _k(p) and H _k(q) represent Ω respectively _pAnd Ω _qIn belong to k (1≤k≤K) type of total number of pixels of texture primitive.

Compared with prior art, the present invention has following beneficial effect:

(1) the present invention adopts fixedly neighborhood, but utilizes outline map to define a kind of dynamic neighborhood, i.e. homogeneity sheet; This homogeneity sheet can change according to different regional area situation adaptively; Has the uneven ability of the gray scale of processing; Because the homogeneity sheet has neighborhood homogeney and neighborhood adaptivity, like this, both solved the difficult problem that the neighborhood window is selected in the classic method; Guarantee the accuracy of regional statistical information again, helped improving the precision of cutting apart simultaneously.

(2) the present invention adopts that fixedly neighborhood and Euclidean distance are estimated the definition weighting function; But adopt fuzzy mearue and homogeneity sheet; Defined a new weighting function; The modification of weighting function makes the present invention eliminate the wrong phenomenon of dividing of image local basically, has improved the precision that tumor of breast is cut apart, and also makes it show good robustness simultaneously.

Description of drawings

Fig. 1 is the breast ultrasound image lesion segmentation approach structured flowchart based on homogeneity sheet and fuzzy mearue;

Fig. 2 is the edge detection results synoptic diagram;

Fig. 3 is homogeneity sheet structure and synoptic diagram as a result;

Fig. 4 is that the present invention is to a width of cloth breast ultrasound malignant tumour image segmentation result and existing two kinds of method segmentation result comparison diagrams;

Fig. 5 is that the present invention is to a width of cloth breast ultrasound malignant tumour image segmentation result and existing two kinds of method segmentation result comparison diagrams;

Fig. 6 is that the present invention is to a width of cloth breast ultrasound benign tumour image segmentation result and existing two kinds of method segmentation result comparison diagrams;

Fig. 7 repeats to cut apart 50 times robust performance curve map to a width of cloth breast ultrasound image for the present invention and existing method;

Fig. 8 repeats to cut apart 10 times figure as a result for existing method to a width of cloth breast ultrasound image;

Fig. 9 repeats to cut apart 10 times figure as a result for the present invention to a width of cloth breast ultrasound image.

Embodiment

To combine accompanying drawing and embodiment that the present invention is done further description below.

Referring to Fig. 1, the lesion segmentation approach of the breast ultrasound image based on homogeneity sheet and fuzzy mearue of the present invention may further comprise the steps:

Step 1 is imported breast ultrasound image to be split, shown in Fig. 2 (a).

Step 2 is treated split image and is carried out rim detection, obtains the edge energy value, generates outline map.The detailed process that realizes this step is following:

(2a) at first use one group of strange, even wave filter (to 1 yardstick, 0 °, 22.5 °; 45 °, 67.5 °, 90 °; 112.5 °, 135 ° and 157.5 ° of 8 directions) and one group of DOG wave filter (to 1 yardstick) respectively input picture is carried out convolutional filtering, the filter response of each pixel is one group of high dimension vector; High dimension vector to all pixels in the image carries out the K-means cluster, and each cluster centre is represented one type of texture primitive, and each pixel of input picture is endowed the label of an affiliated texture primitive type; Then generate the texture primitive characteristic image, shown in Fig. 2 (b), K representes texture primitive classification number (following all same); Decide as the case may be, get K=16 in this experiment;

(2b) obtain the edge energy of each pixel through the rim detection function calculation, the equationof structure of rim detection function is:

$E(p,{\mathrm{\θ}}_t)={\mathrm{\Σ}}_{k=1}^K|E_k^L(p,{\mathrm{\θ}}_t)-E_k^R(p,{\mathrm{\θ}}_t)|+\frac{{(E_k^L(p,{\mathrm{\θ}}_t)-E_k^R(p,{\mathrm{\θ}}_t))}^2}{E_k^L(p,{\mathrm{\θ}}_t)+E_k^R(p,{\mathrm{\θ}}_t)}$

Be the center with pixel p, radius is in the circular neighborhood of R (getting R=5 in this experiment), along θ _tDiameter on the direction is divided into left and right neighborhood with this neighborhood,

Be respectively this and put in the left neighborhood and the energy in the right neighborhood, function expression does $E_k^L(p,{\mathrm{\θ}}_t)={\mathrm{\μ}}_{{\mathrm{\η}}_p^L}\left(k\right)\×d_k^L(p,{\mathrm{\θ}}_t),$ $E_k^R(p,{\mathrm{\θ}}_t)={\mathrm{\μ}}_{{\mathrm{\η}}_p^R}\left(k\right)\×d_k^R(p,{\mathrm{\θ}}_t).$ ${\mathrm{\μ}}_{{\mathrm{\η}}_p^L}\left(k\right),$ ${\mathrm{\μ}}_{{\mathrm{\η}}_p^R}\left(k\right)$ Belong to the degree of membership of k class texture primitive with the interior pixel of right neighborhood in the left neighborhood of remarked pixel point p respectively, and satisfy constraint condition

The degree of membership computing formula does ${\mathrm{\μ}}_{{\mathrm{\η}}_p^L}\left(k\right)=\left({\mathrm{\Σ}}_{i\∈L_p}\mathrm{\δ}(T\left(p_i\right)-k)\right)/\left|L_p\right|,$ ${\mathrm{\μ}}_{{\mathrm{\η}}_p^R}\left(k\right)=\left({\mathrm{\Σ}}_{i\∈R_p}\mathrm{\δ}(T\left(p_i\right)-k)\right)/\left|R_p\right|,$ L _p, R _pRepresent the set of all pixels in left neighborhood and the right neighborhood respectively, | ● | the gesture of expression set, p _iBe any pixel in pixel p left side neighborhood or the right neighborhood, T (p _i) be pixel p among Fig. 2 (b) _iThe texture primitive category label, δ (g) is a Kronecker ' s delta function, if i.e. T (p _i)=k is δ (T (p then _i)-k)=1, otherwise δ (T (p _i)-k)=0;

Be illustrated respectively in θ _tOn the direction, belong to the similarity of the pixel of k class texture primitive in the pixel p neighborhood left with it and in the right neighborhood, computing formula does $d_k^L(p,{\mathrm{\θ}}_t)=|\mathrm{OE}(p,{\mathrm{\θ}}_t)-m_k^L(p,{\mathrm{\θ}}_t)|,$ $d_k^R(p,{\mathrm{\θ}}_t)=|\mathrm{OE}(p,{\mathrm{\θ}}_t)-m_k^R(p,{\mathrm{\θ}}_t)|,$ OE (p, θ _t) be that pixel p is at θ _tOriented energy intensity on the direction, expression formula does

I is an input picture,

With

Be respectively that direction is θ _tEven symmetry and odd symmetry wave filter, the even symmetry wave filter is the second order local derviation of Gaussian function, the odd symmetry wave filter is its corresponding Hilbert transform.

Be respectively at θ _tOn the direction, belong to the mean direction energy intensity of the pixel of k class texture primitive in pixel p left side neighborhood and the right neighborhood, calculation expression does $m_k^L(p,{\mathrm{\θ}}_t)=\frac{{\mathrm{\Σ}}_{i\∈L_p}\mathrm{OE}(p_i,{\mathrm{\θ}}_t)\mathrm{\δ}(T\left(p_i\right)-k)}{{\mathrm{\Σ}}_{i\∈L_p}\mathrm{\δ}(T\left(p_i\right)-k)},$

$m_k^R(p,{\mathrm{\θ}}_t)=\frac{{\mathrm{\Σ}}_{i\∈R_p}\mathrm{OE}(p_i,{\mathrm{\θ}}_t)\mathrm{\δ}(T\left(p_i\right)-k)}{{\mathrm{\Σ}}_{i\∈R_p}\mathrm{\δ}(T\left(p_i\right)-k)}.$

After (2c) the rim detection function E of application construction obtains the edge energy of 8 different directions; Use optimization method (Martin D. again; Fowlkes C.; Malik J.Learning to detect natural imageboundaries using local brightness.color and texture cues.IEEETransactions on Pattern Analysis and Machine Intelligence, 2004,26 (5): 530-549) carry out edge optimization; The rule of taking the victor to obtain entirely at last obtains the optimal edge energy; I.e. t=1; 2; L, 8.The optimal edge energy of each pixel is called outline map as the resulting image of gray scale, and shown in Fig. 2 (c), gray-scale value bigger (promptly white more) might be marginal point more among the figure.

Step 3, adopting the outline map in the step 2 is each pixel definition homogeneity sheet.

(x, y), Fig. 3 (a) is depicted as its homogeneity sheet organigram to any pixel p.Rectangle is represented search window among the figure, and the curve that does not exceed rectangle frame is represented arbitrary curve.The borderline pixel of search window is divided into two parts: the line that has clockwise arrow is represented the pixel of window edge the first half, is designated as $a_1^1(x-R,y+1),a_2^1(x-R,y),...,a_{4R+2}^1(x+R,y),$ The line that has counterclockwise arrow is represented the pixel of window edge the latter half, is designated as $a_1^2(x-R,y),a_2^2(x-R,y+1),...,a_{4R+2}^2(x+R,y-1).$ Pixel to the first half travels through in the direction of the clock, and the latter half pixel is by counterclockwise traveling through.

The basic process of homogeneity sheet structure is:

(3a) confirming of search window: (x y) is the center, confirms that size is the search window N of (2R+1) * (2R+1) with the pixel p on the outline map _p, R=10 in the experiment.

(3b) the alligatoring stage: to search window N _pThe first half pixel, at first choose three continuous adjacent accounting for

Obtain corresponding three lines with central point

Point by point scanning center line then

When running into net point p _j, calculate E _Max(p, p _j) value:

L (p, p _j) be to a p from a p _jThe set of all net points, E ^*(t) edge energy for calculating among the step 2c; Then to E _Max(p, p _j) value is carried out threshold decision, if E _Max(p, p _j)≤τ sets up, and thinks that then this net point is candidate's homologous pints, and continues scanning center's line; Otherwise, think p _jNot candidate's homologous pints, and finish the scanning of center line.τ decides as the case may be, gets τ=0.24 in this instance.

(3c) elaboration phase: for candidate's homologous pints p _jCarry out micronization processes, judge according to the angle difference judgment criterion whether it is homologous pints, to reach the purpose of discarding the dross and selecting the essential.At first represent corresponding straight line, then straight line pp with vector form _jWith Between angle θ _{J, i}Be defined as

Similarly, can obtain pp respectively _jWith

Angle theta _{J, i-1}And pp _jWith

Angle theta _{J, i+1}Utilize three angles that obtain, judge p _jThe condition that whether is homologous pints is:

After passing through micronization processes like this, the remaining some p that is not excluded _jBe exactly homogeneity sheet Ω _pIn point, and with it to refuse degree of receiving 1-E _Max(p, p _j) charge to homogeneity sheet Ω _pIn.

(3d) to search window N _pThe latter half pixel takes the method identical with the first half pixel to seek homologous pints, finally accomplishes homogeneity sheet Ω _pStructure.Shown in Fig. 3 (b), square frame is represented search window among the figure, and white portion is represented the homogeneity sheet of smooth region pixel, and the black region in the square frame is represented the homogeneity sheet of non-flat area pixel point.

Step 4, definition is based on the weighting function of homogeneity sheet and fuzzy mearue.

Traditional based on the NCut method adopt Euclidean distance and fixedly neighborhood define weighting function, but for noise image, as the tumor of breast image cut apart and improper.In the present invention, adopt the homogeneity sheet to substitute fixedly neighborhood and represent with fuzzy set, the fuzzy mearue with two fuzzy set diversities of measurement of redetermination replaces Euclidean distance on this basis, to define weighting function.If π _p, π _qBe respectively the homogeneity sheet Ω of pixel p and q _pAnd Ω _qFuzzy set, we just can define pixel to (p, the weighting function between q):

$W(p,q)=\left\{\begin{array}{cc}\exp (-\left(d({\mathrm{\&pi;}}_p,{\mathrm{\&pi;}}_q)\right)/\mathrm{\&sigma;})& \begin{array}{cc}\mathrm{if}& {\left|\right|X\left(p\right)-X\left(q\right)\left|\right|}_2<\mathrm{<figure-callout\; id="0"\; label="r"\; filenames="111121114155.png"\; state="\{\{state\}\}">r</figure-callout>}\end{array}\\ 0& \mathrm{otherwise}\end{array}\right.$

Wherein r gets constant 8 for connecting radius in the experiment.σ decides according to actual conditions, gets constant σ=0.06 in this instance.D (π _p, π _q) be to be defined in two fuzzy set π _pAnd π _qOn fuzzy mearue, expression formula is following:

$d({\mathrm{\&pi;}}_p,{\mathrm{\&pi;}}_q)={\mathrm{\&Sigma;}}_{k=}^K{\mathrm{\&lambda;}}_k{\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_p}\left(k\right)+(1-{\mathrm{\&lambda;}}_k){\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_q}\left(k\right)-\min ({\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_p}\left(k\right),{\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_q}\left(k\right))$

Fuzzy set π wherein _pAnd π _qBe defined as respectively ${\mathrm{\&pi;}}_p=\{(k,{\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_p}\left(k\right)),k=\mathrm{1,2},...,K\},$ ${\mathrm{\&pi;}}_q=\{(k,{\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_q}\left(k\right)),k=\mathrm{1,2},...,K\};$ ${\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_p}\left(k\right),$ ${\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_q}\left(k\right)$ Represent that respectively homogeneity sheet Ω is little _pAnd Ω _qIn pixel belong to k (degree of membership of 1≤k≤K) type of texture primitive, computing formula be respectively ${\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_p}\left(k\right)=\frac{{\mathrm{\&Sigma;}}_{j\&Element;{\mathrm{\&Omega;}}_p}m(p,p_j)\mathrm{\&delta;}(T\left(p_j\right)-k)}{{\mathrm{\&Sigma;}}_{j\&Element;{\mathrm{\&Omega;}}_p}m(p,p_j)}$ With ${\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_q}\left(k\right)=\frac{{\mathrm{\&Sigma;}}_{j\&Element;{\mathrm{\&Omega;}}_q}m(q,q_j)\mathrm{\&delta;}(T\left(q_j\right)-k)}{{\mathrm{\&Sigma;}}_{j\&Element;{\mathrm{\&Omega;}}_q}m(q,q_j)},$ M (p, p _j) and m (q, q _j) be the degree of receiving of refusing that calculates among the step 3c.λ _kBe Ω _pAnd Ω _qBetween the regulation and control coefficient, its formulate does

H _k(p) and H _k(q) represent Ω respectively _pAnd Ω _qIn belong to k (1≤k≤K) type of total number of pixels of texture primitive.

Step 5, lesion segmentation.

Represent the image as a cum rights non-directed graph G=(W), V is the set on summit for V, E, and E is the set that connects the limit on summit, and it is the weight matrix that element constitutes by weight that W is one, promptly W={W (p, q) }.The result that NCut is cut apart A, B} satisfies following formula:

$\mathrm{NCut}(A,B)=\frac{\mathrm{cut}(A,B)}{\mathrm{asso}(A,V)}+\frac{\mathrm{cut}(A,B)}{\mathrm{asso}(B,V)}$

$\mathrm{cut}(A,B)=\underset{p\&Element;A,q\&Element;B}{\mathrm{\&Sigma;}}W(p,q),$ $\mathrm{asso}(A,V)=\underset{p\&Element;A,q\&Element;V}{\mathrm{\&Sigma;}}W(p,q)$

(A B) solves the optimal dividing problem, in fact can be according to L=D through in global scope, minimizing NCut ^-1/2WD ^1/2Proper vector confirm segmentation result.

In order to be partitioned into tumor region exactly, can reduce computation burden again, further utilized prior imformation (being the rough position of tumour) and sampling policy here.Lesion segmentation simply is described below:

(5a) doctor through the selected area-of-interest that comprises tumour of mouse (region of interest, ROI), the background area of image is appointed as in the zone that RO I is outer, and with this as prior imformation, the weight between pixel is set to 1 in the background area again;

(5b) image is carried out random sampling reducing computation complexity, sampling rate lower usually (all number of pixels 1% or lower), we are taken as 1.5% here.If all sample points in the order set C presentation video background area then produce | C|-1 constraint condition.Suppose M=[M ₁, M _t... M _{| C|-1}], each constraint condition M wherein _tBe the vector of N * 1, it has two nonzero elements, M _t(i)=1, M _t(j)=-1, i, j ∈ C, thus the partition problem that has most that will scheme becomes restricted minimization problem,

Wherein W is the weight matrix that calculates according to the weighting function in the step 4, and to weigh the similarity between sample point, D is a diagonal matrix, and D (p, p)=∑ _qW (p, q), Q=D ^-1W, X are the division matrixes of N * 2, and N is all sample point numbers, X=[X ₁, X ₂], X wherein _lFor scale-of-two indication vector, if i belongs to A, X [i, 1]=1 then, otherwise, X [i, 1]=0.If i belongs to B, X [i, 2]=1 then, otherwise, X [i, 2]=0.The concrete method for solving of restricted minimization problem can be with reference to the improvement NCut algorithm of Yu.et al proposition.

(5c) after finding the solution restricted minimization problem and obtaining the pairing proper vector of sample point; Carrying out interior inserting through the weighting of the sample point adjacent with non-sample point estimates; Obtain non-sample and put pairing proper vector; This two-part proper vector is combined to carry out the K-means cluster, and cluster class number is 2.So far, just all accomplished lesion segmentation, the visible Fig. 4 (e) of final segmentation result to the breast ultrasound image.

Effect of the present invention can further specify through following experiment:

1. experiment condition

Hardware platform is: Intel (R) Pentium (R) 1CPU 2.80GHz, Windows XP Professional.

Software platform is: Matlab7.0 and Visual C++6.0.

2. experiment content and result

The present invention comes from clinical breast ultrasound evaluate image (malignant tumour 19 examples to 50; Benign tumour 31 examples) study; With proposed by the invention based on homogeneity sheet (Homogeneous Patch; HP) and fuzzy mearue (Fuzzy metric, (Phase-Based Level Set PBLS) compares with the improvement NCut method (being designated hereinafter simply as interactive NCut) that Yu.et al proposes for algorithm FM) (HP-FM) and existing level set dividing method based on phase place.

2.1 qualitative analysis

Fig. 4, Fig. 5 and Fig. 6 are one group of instance.

Fig. 4 (a) has shown that one has irregularly shaped; Local feature similar malignant breast tumor, Fig. 4 (b) is the segmentation result of PBLS algorithm, Fig. 4 (c) is the segmentation result of interactive NCut method; Fig. 4 (d) is the segmentation result of method of the present invention, and Fig. 4 (e) is manual segmentation result.Can be known that by Fig. 4 the PBLS algorithm has only utilized marginal information, not consider texture information and spatial information, the external force catching range is limited and can not converge to the depression boundary.Interactive NCut method is not considered the statistical information of regional area, so the local error in recessed zone is cut apart seriously.The present invention has considered the statistical information of regional area through utilizing homogeneity sheet and fuzzy mearue, therefore can obtain desirable segmentation result, has avoided local error to cut apart basically.

Fig. 5 (a) shown one irregularly shaped, have the malignant breast tumor of serious rear echo attenuation,

Fig. 5 (b) is the segmentation result of PBLS algorithm, and Fig. 5 (c) is the segmentation result of interactive NCut method, and Fig. 5 (d) is the segmentation result of method of the present invention, and Fig. 5 (e) is manual segmentation result.Can know that by Fig. 5 the rim detection in the PBLS algorithm is the basis with the phase information, because in fact phase information is not irrelevant with grey scale change, thus cause the failure of rim detection, and then cause the PBLS algorithm to produce a not satisfied segmentation result.Interactive NCut method is only utilized the gradient information of image, because the edge in the breast ultrasound image is not desirable step edges, is difficult to cut apart the serious breast ultrasound image of the pseudo-shadow of decay.And these two kinds of methods all do not consider the statistical information of local neighborhood, and statistical information is very favorable for the uneven ultrasonoscopy of gray scale.The present invention has fully utilized edge and area information through structure homogeneity sheet, the influence of the pseudo-shadow of having avoided effectively decaying, and the segmentation result that obtains has subjective vision effect preferably.

Fig. 6 (a) has shown a mammary gland benign tumor with regular shape, and Fig. 6 (b) is the segmentation result of PBLS algorithm, and Fig. 6 (c) is the segmentation result of interactive NCut method, and Fig. 6 (d) is the segmentation result of method of the present invention, and Fig. 6 (e) is manual segmentation result.Can know that by Fig. 6 PBLS algorithm and interactive NCut method are not considered textural characteristics, so the tangible breast ultrasound image of textural characteristics is difficult to obtain correct segmentation effect.HP-FM algorithm of the present invention has taken into full account textural characteristics, therefore Fig. 6 (a) is had desirable segmentation effect.

These several groups of instances have fully proved the validity of method of the present invention.

2.2 quantitative test

Except that the subjective vision effect, the similarity measurement of the tumor region that three kinds of different dividing methods are cut apart and the tumor region of expert's Freehandhand-drawing system is estimated the profile extraction effect.Suppose A _aBe the segmentation result tag set of algorithm, A _mBe the tag set that manual standard is cut apart, then similarity measurement SI definition is as follows:

$\mathrm{SI}=\frac{\left|A_{m}I{A}_a\right|}{\left|A_{m}U{A}_a\right|}$

The SI value is big more, and it is good more to cut apart overall performance.SI=1 is that mate fully in the zone that manual zone of cutting apart and doctor are cut apart.

The PBLS algorithm is respectively aforesaid three width of cloth tumor image similarity measurements: 83.39%, 81.75%, 81.55%.Interactive NCut dividing method is respectively the similarity measurement of this three width of cloth image: 83.46%, 75.50%, 77.18%.HP-FM algorithm of the present invention is respectively the similarity measurement result of this three width of cloth image: 88.88%, 89.96%, 88.22%.

The PBLS algorithm, the average similarity tolerance that interactive NCut algorithm and HP-FM algorithm of the present invention are cut apart 50 clinical breast ultrasound images is respectively 85.89%, 78.99%, 88.2%.Can know that from above analysis the HP-FM algorithm that the present invention proposes can obtain better segmentation performance.

2.3 robust analysis

Under the identical situation of sampling rate; Through changing random sample; And with the segmentation result of HP-FM algorithm of the present invention and interactive algorithm and manual segmentation result be goldstandard relatively; We can obtain about average minimum Eustachian distance (average minimum Euclide distance, curve map AD), as shown in Figure 7.In addition,, selected the segmentation result of 10 interactive NCut algorithms and HP-FM algorithm of the present invention, be shown in Fig. 8 (a)-(j) and Fig. 9 (a)-(j) respectively for ease of visual observation.From Fig. 7,8,9 is visible, and HP-FM algorithm of the present invention is insensitive to random sample, cut apart to have provided identical segmentation result, and interactive NCut method is responsive that segmentation result changes along with the variation of sample point at every turn.This mainly is because method of the present invention has adopted homogeneity sheet and fuzzy mearue definition weight; At first the feature difference between the homogeneity sheet than the feature difference between pixel more remarkable with have more robustness; Secondly fuzzy mearue is than Euclidean distance robust more, so overcome sensitive question.Under the situation that the identical but random sample of random sample number changes, the present invention has guaranteed good stability and repeatable.We may safely draw the conclusion by experimental result, and the present invention is more suitable for cutting apart in ultrasonoscopy.

Can explain by above emulation experiment; To cutting apart of clinical breast ultrasound image; There is certain advantage in the present invention; Overcome existing PBLS and interactive NCut method is applied in the deficiency on the breast ultrasound image, and though be qualitative analysis also, quantitative test or robust analysis, the present invention all is superior to existing P BLS and interactive NCut method.

In sum, the segmentation effect that the present invention is directed to clinical breast ultrasound image obviously is superior to existing P BLS and the interactive NCut method segmentation effect to clinical breast ultrasound image.

Claims (7)

1. the lesion segmentation approach based on the breast ultrasound image of homogeneity sheet and fuzzy mearue is characterized in that this method comprises the steps:

(1) input breast ultrasound image to be split;

(2) treat split image and carry out rim detection, comprising:

(2.1) adopt the breast ultrasound image texture features of importing in filter set and the clustering algorithm extraction step (1), generate the texture primitive characteristic image;

(2.2) through fuzzy mearue the multiple information of each image slices vegetarian refreshments is combined definition rim detection function; Calculate the edge energy of each pixel, generate initial edge figure, and carry out edge optimization; Finally obtain the outline map of original image, thereby accomplish rim detection;

(3) definition homogeneity sheet; Utilize the outline map that generates in the step (2) to be each image pixel definition homogeneity sheet, said homogeneity sheet is a kind of neighborhood that reflects the neighbor similarity relation; The structure of homogeneity sheet comprises the steps:

(3.1) be that search window is confirmed at the center with the current pixel point;

(3.2) the alligatoring stage: figure goes up the straight line of crossing all directions of current point in this search window of scanning on the edge of; If the edge energy value of pixel is less than given threshold value on certain direction straight line; Think that then this pixel is candidate's homologous pints, otherwise, with its eliminating and stop the scanning of this direction straight line;

(3.3) elaboration phase: for candidate's homologous pints; Further take micronization processes to judge whether it is real homologous pints; Then the remaining pixel that is not excluded is considered to the member in this homogeneity sheet in the current search window, and exists with the form of probability;

(4) definition weighting function; The homogeneity sheet of each image slices vegetarian refreshments in the step (3) is regarded as a fuzzy set, on this basis, defines a fuzzy mearue that is used for weighing the fuzzy set diversity, and define weighting function with it according to marginal information, texture information and spatial information;

(5) lesion segmentation; Adopt the weighting function of step (4) definition, the similarity between calculating pixel is to make up the weight matrix of entire image; With the coarse position information of tumour as prior imformation; Simultaneously ultrasonoscopy is carried out random sampling and reduce computation complexity, upgrade weight matrix, the partition problem that has most of scheming is become restricted minimization problem according to prior imformation of introducing and Sampling Strategies; At last; Carry out interior inserting through the weighting of the sample point adjacent and estimate, obtain estimation, thereby realize cutting apart tumour to the breast ultrasound image segmentation with non-sample point.

2. the lesion segmentation approach of the breast ultrasound image based on homogeneity sheet and fuzzy mearue according to claim 1; It is characterized in that: in the said step (2.1); Filter set comprises one group of odd even wave filter and one group of DOG wave filter, and the odd even wave filter is to 1 yardstick and 0 °, 22.5 °, 45 °, 67.5 °, 90 °, 112.5 °, 135 °, 157.5 °, and the input picture of 8 directions carries out convolutional filtering; The DOG wave filter carries out convolutional filtering to the input picture of 1 yardstick; The filter response of each pixel is one group of high dimension vector, through clustering algorithm the high dimension vector of all pixels in the image is carried out cluster, and each cluster centre is one type of texture primitive; Each pixel of input picture is endowed the label of an affiliated texture primitive type, then generates the texture primitive characteristic image.

3. the lesion segmentation approach of the breast ultrasound image based on homogeneity sheet and fuzzy mearue according to claim 1 and 2, it is characterized in that: the rim detection function described in the step (2.2) is: $E(p,{\mathrm{\&theta;}}_t)={\mathrm{\&Sigma;}}_{k=1}^K|E_k^L(p,{\mathrm{\&theta;}}_t)-E_k^R(p,{\mathrm{\&theta;}}_t)|+\frac{{(E_k^L(p,{\mathrm{\&theta;}}_t)-E_k^R(p,{\mathrm{\&theta;}}_t))}^2}{E_k^L(p,{\mathrm{\&theta;}}_t)+E_k^R(p,{\mathrm{\&theta;}}_t)}$

K representes texture primitive classification number, is being the center with pixel p, and radius is in the circular neighborhood of R, along θ _tDiameter on the direction is divided into left and right neighborhood with this neighborhood,

Be respectively this and put in the left neighborhood and the energy in the right neighborhood, wherein: $E_k^L(p,{\mathrm{\&theta;}}_t)={\mathrm{\&mu;}}_{{\mathrm{\&eta;}}_p^L}\left(k\right)\&times;d_k^L(p,{\mathrm{\&theta;}}_t),$ $E_k^R(p,{\mathrm{\&theta;}}_t)={\mathrm{\&mu;}}_{{\mathrm{\&eta;}}_p^R}\left(k\right)\&times;d_k^R(p,{\mathrm{\&theta;}}_t),$ ${\mathrm{\&mu;}}_{{\mathrm{\&eta;}}_p^L}\left(k\right),$ ${\mathrm{\&mu;}}_{{\mathrm{\&eta;}}_p^R}\left(k\right)$ Belong to the degree of membership of k class texture primitive with the interior pixel of right neighborhood in the left neighborhood of remarked pixel point p respectively, and satisfy constraint condition ${\mathrm{\&Sigma;}}_{k=1}^K{\mathrm{\&mu;}}_{{\mathrm{\&eta;}}_p^L}\left(k\right)=1,$ ${\mathrm{\&Sigma;}}_{k=1}^K{\mathrm{\&mu;}}_{{\mathrm{\&eta;}}_p^R}\left(k\right)=1;$ $d_k^L(p,{\mathrm{\&theta;}}_t),$ $d_k^R(p,{\mathrm{\&theta;}}_t)$ Be illustrated respectively in θ _tOn the direction, in the pixel p neighborhood left with it with right neighborhood in the similarity of the pixel that belongs to k class texture primitive,

Membership

function expression is:

${\mathrm{\&mu;}}_{{\mathrm{\&eta;}}_p^L}\left(k\right)=\frac{{\mathrm{\&Sigma;}}_{i\&Element;L_p}\mathrm{\&delta;}(T\left(p_i\right)-k)}{\left|L_p\right|},$ ${\mathrm{\&mu;}}_{{\mathrm{\&eta;}}_p^R}\left(k\right)=\frac{{\mathrm{\&Sigma;}}_{i\&Element;R_p}\mathrm{\&delta;}(T\left(p_i\right)-k)}{\left|L_p\right|}$

L wherein _p, R _pRepresent the set of all pixels in left neighborhood and the right neighborhood respectively, | ● | the gesture of expression set, p _iBe any pixel in pixel p left side neighborhood or the right neighborhood, T (p _i) be p _iThe texture primitive category label, δ (g) is a Kronecker ' s delta function, if i.e. T (p _i)=k is δ (T (p then _i)-k)=1, otherwise δ (T (p _i)-k)=0;

Similarity

function expression is:

$d_k^L(p,{\mathrm{\&theta;}}_t)=|\mathrm{OE}(p,{\mathrm{\&theta;}}_t)-m_k^L(p,{\mathrm{\&theta;}}_t)|,$ $d_k^R(p,{\mathrm{\&theta;}}_t)=|\mathrm{OE}(p,{\mathrm{\&theta;}}_t)-m_k^R(p,{\mathrm{\&theta;}}_t)|$

OE (p, θ wherein _t) be that pixel p is at θ _tOriented energy intensity on the direction, expression formula does

I is an input picture, With

Be respectively that direction is θ _tEven symmetry and odd symmetry wave filter;

Be respectively at θ _tOn the direction, mean direction energy intensity in pixel p left side neighborhood and the right neighborhood,

The average energy intensity direction

and

The expression is:

$m_k^L(p,{\mathrm{\&theta;}}_t)=\frac{{\mathrm{\&Sigma;}}_{i\&Element;L_p}\mathrm{OE}(p_i,{\mathrm{\&theta;}}_t)\mathrm{\&delta;}(T\left(p_i\right)-k)}{{\mathrm{\&Sigma;}}_{i\&Element;L_p}\mathrm{\&delta;}(T\left(p_i\right)-k)},$

$m_k^R(p,{\mathrm{\&theta;}}_t)=\frac{{\mathrm{\&Sigma;}}_{i\&Element;R_p}\mathrm{OE}(p_i,{\mathrm{\&theta;}}_t)\mathrm{\&delta;}(T\left(p_i\right)-k)}{{\mathrm{\&Sigma;}}_{i\&Element;R_p}\mathrm{\&delta;}(T\left(p_i\right)-k)}.$

4. the lesion segmentation approach of the breast ultrasound image based on homogeneity sheet and fuzzy mearue according to claim 3, it is characterized in that: said clustering algorithm is the K-Means algorithm.

5. the lesion segmentation approach of the breast ultrasound image based on homogeneity sheet and fuzzy mearue according to claim 3 is characterized in that:

The confirming of search window described in the step (3.1) comprises the steps: that (x y) is the center, defines the search window that size is (2R+1) * (2R+1) a: N with the position of the pixel p on the outline map in image _p(i, j)=wherein, R is the height and the width of search window, and the borderline pixel of search window is divided into two parts for x-R≤i≤x+R, y-R≤j≤y+R}, wherein, $a_1^1(x-R,y+1),a_2^1(x-R,y),...,a_{4R+2}^1(x+R,y)$ Represent the pixel of window edge the first half, $a_1^2(x-R,y),a_2^2(x-R,y+1),...,a_{4R+2}^2(x+R,y-1)$ Represent the pixel of window edge the latter half, the pixel of the first half traveled through in the direction of the clock, to the latter half pixel by traversal counterclockwise;

The alligatoring stage described in the step (3.2) comprises the steps: the pixel of the first half and the latter half is chosen three continuous consecutive point respectively

Obtain corresponding three lines with central point respectively

The center line of certain direction in the point by point scanning search window then

When running into net point p _j, calculate E _Max(p, p _j) value:

$E_{\max }(p,p_j)={\max }_{t\&Element;L(p,p_j)}{E}^{*}\left(t\right)$

Wherein, L (p, p _j) be to a p from a p _jThe set of all net points, E ^*(t) edge energy that obtains for the rim detection function calculation; If E _Max(p, p _j)≤τ sets up, and then thinks p _jBe candidate's homologous pints, and continue scanning center's line; Otherwise, think p _jNot candidate's homologous pints, and finish the scanning of center line;

Elaboration phase described in the step (3.3) comprises the steps: for candidate's homologous pints p _jCarry out micronization processes, judge that according to the angle difference judgment criterion whether it is homologous pints, at first representes corresponding straight line, then pp with vector form _jAnd pa _iBetween angle θ _{J, i}Be defined as:

${\mathrm{\&theta;}}_{j,i}=\&angle;({\stackrel{\&RightArrow;}{\mathrm{pp}}}_j,{\stackrel{\&RightArrow;}{\mathrm{pa}}}_i)=\arccos \frac{{\left({\stackrel{\&RightArrow;}{\mathrm{pp}}}_j\right)}^T{\stackrel{\&RightArrow;}{\mathrm{pa}}}_i}{\left|\right|{\stackrel{\&RightArrow;}{\mathrm{pp}}}_j\left|\right|\&CenterDot;\left|\right|{\stackrel{\&RightArrow;}{\mathrm{pa}}}_i\left|\right|}$

Similarly, can calculate pp _jAnd pa _k(k=i-1, the angle theta between i+1) _{J, i-1}And θ _{J, i+1}, in order to get rid of the false homologous pints in candidate's homologous pints, define a kind of angle difference judgment criterion:

The then remaining some p that is not excluded _jBe exactly homogeneity sheet Ω _pIn the point.

6. the lesion segmentation approach of the breast ultrasound image based on homogeneity sheet and fuzzy mearue according to claim 5, it is characterized in that: the weighting function in the step (4) is:

<math> <mrow> <mi>W</mi> <mrow> <mo>(</mo> <mi>p</mi> <mo>,</mo> <mi>q</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfenced open='{' close=''> <mtable> <mtr> <mtd> <mi>exp</mi> <mrow> <mo>(</mo> <mo>-</mo> <mrow> <mo>(</mo> <mi>d</mi> <mrow> <mo>(</mo> <msub> <mi>&amp;pi;</mi> <mi>p</mi> </msub> <mo>,</mo> <msub> <mi>&amp;pi;</mi> <mi>q</mi> </msub> <mo>)</mo> </mrow> <mo>)</mo> </mrow> <mo>/</mo> <mi>&amp;sigma;</mi> <mo>)</mo> </mrow> </mtd> <mtd> <mfenced open='' close=''> <mtable> <mtr> <mtd> <mi>if</mi> </mtd> <mtd> <msub> <mrow> <mo>|</mo> <mo>|</mo> <mi>X</mi> <mrow> <mo>(</mo> <mi>p</mi> <mo>)</mo> </mrow> <mo>-</mo> <mi>X</mi> <mrow> <mo>(</mo> <mi>q</mi> <mo>)</mo> </mrow> <mo>|</mo> <mo>|</mo> </mrow> <mn>2</mn> </msub> <mo>&lt;;</mo> <mi>r</mi> </mtd> </mtr> </mtable> </mfenced> </mtd> </mtr> <mtr> <mtd> <mn>0</mn> </mtd> <mtd> <mi>otherwise</mi> </mtd> </mtr> </mtable> </mfenced> </mrow></math>

π wherein _p, π _qBe respectively homogeneity sheet Ω _pAnd Ω _qFuzzy set, r, σ are constant, d (π _p, π _q) be to be defined in two fuzzy mearues on the fuzzy set, expression formula is following:

$d({\mathrm{\&pi;}}_p,{\mathrm{\&pi;}}_q)={\mathrm{\&Sigma;}}_{k=}^K{\mathrm{\&lambda;}}_k{\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_p}\left(k\right)+(1-{\mathrm{\&lambda;}}_k){\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_q}\left(k\right)-\min ({\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_p}\left(k\right),{\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_q}\left(k\right))$

Wherein, K representes texture primitive classification number; Fuzzy set π _pAnd π _qBe defined as respectively ${\mathrm{\&pi;}}_p=\{(k,{\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_p}\left(k\right)),k=\mathrm{1,2},...,K\},$ ${\mathrm{\&pi;}}_q=\{(k,{\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_q}\left(k\right)),k=\mathrm{1,2},...,K\};$ ${\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_p}\left(k\right),$ ${\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_q}\left(k\right)$ Represent homogeneity sheet Ω respectively _pAnd Ω _qIn pixel belong to k (degree of membership of 1≤k≤K) type of texture primitive be defined as

${\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_p}\left(k\right)=\frac{{\mathrm{\&Sigma;}}_{j\&Element;{\mathrm{\&Omega;}}_p}m(p,p_j)\mathrm{\&delta;}(T\left(p_j\right)-k)}{{\mathrm{\&Sigma;}}_{j\&Element;{\mathrm{\&Omega;}}_p}m(p,p_j)},$ ${\mathrm{\&mu;}}_{{\mathrm{\&pi;}}_q}\left(k\right)=\frac{{\mathrm{\&Sigma;}}_{j\&Element;{\mathrm{\&Omega;}}_q}m(q,q_j)\mathrm{\&delta;}(T\left(q_j\right)-k)}{{\mathrm{\&Sigma;}}_{j\&Element;{\mathrm{\&Omega;}}_q}m(q,q_j)}$

δ in the formula (g) is a Kronecker ' s delta function, m (p, p _j) be pixel p in the step (3) _jRefuse degree of receiving, λ _kBe homogeneity sheet Ω _pAnd Ω _qBetween the regulation and control coefficient, its formula is:

${\mathrm{\&lambda;}}_k=\frac{H_k\left(p\right)}{H_k\left(p\right)+H_k\left(q\right)}$

H in the formula _k(p) and H _k(q) represent Ω respectively _pAnd Ω _qIn belong to k (1≤k≤K) type of total number of pixels of texture primitive.

7. according to the lesion segmentation approach of claim 4 or 5 or 6 described breast ultrasound images based on homogeneity sheet and fuzzy mearue, it is characterized in that: the second order local derviation that said even symmetry wave filter is a Gaussian function, odd symmetry wave filter are corresponding Hilbert transform.

Images