CN104966100A - A benign and malignant image lump classification method based on texture primitives - Google Patents

Abstract

The invention provides a benign and malignant image lump classification method based on texture primitives. The method mainly comprises the following steps: dividing a lump area image obtained after normalization processing into a center area and a peripheral area; obtaining feature vectors of the lump area image according to the features of the center area and the peripheral area, and combining the feature vectors into a training sample feature matrix, and training a K nearest neighbor classifier; extracting feature vector Fq of the lump image area obtained through normalization processing of a lump area image to be identified; inputting the feature vector Fq to the trained K nearest neighbor classifier to obtain benign and malignant classification results of the lump area image to be identified. The method, with the difference of features of the center area and the peripheral area of the lump area being taken into consideration, carries out texture dictionary construction and feature extraction on the center area and the peripheral area, introduces linear discriminant analysis to the texture primitive dictionary construction and realizes effective classification of benign and malignant image lumps base on the texture primitives.

Description

The good pernicious sorting technique of image lump based on texture primitive

Technical field

The present invention relates to image lump recognition technology field, particularly relate to the good pernicious sorting technique of a kind of image lump based on texture primitive.

Background technology

Breast cancer is the modal malignant tumour of a kind of women, China oneself become one of the fastest country of breast cancer incidence growth rate, and the age of onset of China's breast cancer is rejuvenation trend.In numerous mammography technology, mammography is one of the most frequently used clinically breast cancer disease examination means.Lump is one of modal Features of breast cancer, differentiates the optimum and pernicious of breast lump by shape, edge and textural characteristics.The good pernicious classification of lump based on image procossing and pattern-recognition can provide objectively auxiliary suggestion for doctor, and reduce unnecessary live tissue puncturing, the diagnosis for clinical early-stage breast cancer has great importance.

In existing lump sorting technique, morphological feature and textural characteristics are conventional feature representation modes.Morphological feature requires higher to the segmentation precision of mass edge, but for clinical image, lump is overlapped with gland tissue, very difficult automatically and be accurately partitioned into mass edge, be thus difficult in clinical middle realization.Comparatively morphological feature, the robustness of textural characteristics is better, requires lower to the segmentation precision of mass edge.

The shortcoming of the existing lump sorting technique based on textural characteristics is: texture primitive does not consider the space distribution information of pixel, and texture primitive lost the classification information of training image when building texture dictionary.

Summary of the invention

The embodiment provides the good pernicious sorting technique of a kind of image lump based on texture primitive, to realize perniciously effectively classifying to image lump is good based on texture primitive.

To achieve these goals, this invention takes following technical scheme.

The good pernicious sorting technique of image lump based on texture primitive, comprising:

Training lump area image is normalized, lump area image after normalized is divided into central area and outer peripheral areas, described central area and outer peripheral areas are built respectively and can distinguish texture dictionary, then obtain the feature of described central area and outer peripheral areas respectively;

The feature of described central area and outer peripheral areas is merged to the proper vector of the lump area image after obtaining described normalized;

The combination of eigenvectors of the training lump area image after described normalized is become training sample eigenmatrix, training k nearest neighbor sorter;

Lump area image to be identified is normalized, the proper vector F of lump image-region after extraction normalized _q, by the proper vector F of lump area image _qbe input to the k nearest neighbor sorter trained, obtain the good pernicious classification results of described lump area image to be identified.

Preferably, described is normalized training lump area image, training image after normalized is divided into central area and outer peripheral areas, the central area of described training image and outer peripheral areas are built respectively and can distinguish texture dictionary, then obtain the central area of described training image and the feature of outer peripheral areas respectively, comprising:

Brightness normalization is carried out to the training image of image lump, image I after normalization _nrepresent, build the rectangle template with two coaxial layers, Mask _inand Mask _ex, described Mask _inand Mask _exlength and be widely respectively the length of described lump training image and wide setting multiple, use described rectangle template that described lump training image is divided into central area R _inwith outer peripheral areas R _ex, R _in=I × Mask _in, R _ex=I × Mask _ex;

Extract the central area R of described training image _inthe each neighborhood gray scale of M of each pixel, it is M that described M territory ash neighborhood gray scale is adjusted to length ²row vector, all row vector organization center region Neighborhood matrix G _in; Extract outer peripheral areas R _exm × M neighborhood gray scale of each pixel, it is M that this M × M neighborhood gray scale is adjusted to length ²row vector, all row vectors composition outer peripheral areas Neighborhood matrix G _ex;

To described central area Neighborhood matrix G _inwith the classification information matrix application linear discriminant analysis of its correspondence, obtain central area neighborhood weight row vector ω _in; To described outer peripheral areas Neighborhood matrix G _exwith the classification information matrix application linear discriminant analysis of its correspondence, obtain outer peripheral areas neighborhood weight row vector ω _ex;

By described central area neighborhood weight row vector ω _inwith described central area Neighborhood matrix G _inevery a line corresponding point be multiplied, obtain central area and can distinguish Neighborhood matrix Gd _in; By described outer peripheral areas neighborhood weight row vector ω _exwith described outer peripheral areas Neighborhood matrix G _exevery a line corresponding point be multiplied, obtain outer peripheral areas and can distinguish Neighborhood matrix Gd _ex;

To described Gd _inand Gd _exapply k mean cluster respectively, obtain central area and can distinguish texture primitive dictionary Dic _inand outer peripheral areas can distinguish texture primitive dictionary Dic _ex;

According to described training image central area Neighborhood matrix G _in, ω _inwith Dic _inobtain training image central area feature F _in; According to described training image outer peripheral areas Neighborhood matrix G _ex, ω _exwith Dic _exobtain training image outer peripheral areas feature F _ex.

Preferably, described to training image Gd _inand Gd _exapply k mean cluster respectively, obtain central area and can distinguish texture primitive dictionary Dic _inand outer peripheral areas can distinguish texture primitive dictionary Dic _ex, comprising:

The Gd corresponding to benign tumors training image _inapplication k mean cluster, obtains cluster centre Cb _in, the quantity of cluster centre is to distinguish the primitive quantity K that texture primitive dictionary comprises _dic, each center is M ²the row vector of length; The Gd corresponding to Malignant mass training image _inapplication k mean cluster, obtains cluster centre Cm _in, the quantity of cluster centre is K _dic, each center is M ²the row vector of length; Combine described Cb _inand Cm _in, form 2K _dicrow M ²the matrix of row $\left[\begin{array}{c}{\mathrm{Cb}}_{\mathrm{in}}\\ {\mathrm{Cm}}_{\mathrm{in}}\end{array}\right],$ Remove with other row vector average Euclidean apart from close K _dicindividual row vector, finally obtains K _dicindividual center, by this K _dicindividual center is as the texture dictionary Dic of central area _in;

The Gd corresponding to benign tumors training image _exapplication k mean cluster, obtains cluster centre Cb _ex, the quantity of cluster centre is to distinguish the primitive quantity K that texture primitive dictionary comprises _dic, each center is M ²the row vector of length; The Gd corresponding to Malignant mass training image _exapplication k mean cluster, obtains cluster centre Cm _ex, the quantity of cluster centre is K _dic, each center is M ²the row vector of length; Combine described Cb _exand Cm _ex, form 2K _dicrow M ²the matrix of row $\left[\begin{array}{c}{\mathrm{Cb}}_{\mathrm{ex}}\\ {\mathrm{Cm}}_{\mathrm{ex}}\end{array}\right],$ Remove with other row vector average Euclidean apart from close K _dicindividual row vector, finally obtains K _dicindividual center, by this K _dicindividual center is as the texture dictionary Dic of outer peripheral areas _ex.

Preferably, described according to described G _in, ω _inwith Dic _inobtain central area feature F _in, comprising:

By described G _inany a line G _in ⁱwith ω _incorresponding point are multiplied, and obtain central area and can distinguish neighborhood vector Gd _in ⁱ, k=1 ..., M ².; By described Gd _in ⁱquantize to Dic _inin the texture primitive nearest with its Euclidean distance quantize mark value, obtain quantize mark value L _in(Gd _in ⁱ), it is specifically calculated as:

$L_{\mathrm{in}}\left({{\mathrm{Gd}}_{\mathrm{in}}}^i\right)=\underset{j}{\arg \min }\left\{d({{\mathrm{Gd}}_{\mathrm{in}}}^i,{{\mathrm{Dic}}_{\mathrm{in}}}^j)\right\},j=1,...,K_{\mathrm{dic}}.$

Wherein d is Euclidean distance, K _dicto distinguish the texture primitive quantity that texture dictionary comprises;

After obtaining the quantification mark value of each pixel in central area, calculate the histogram of all quantification mark value, as the feature F of central area _in, it is specifically calculated as:

Wherein, N _inbe the quantity of central area pixel, δ function is:

$\mathrm{\δ}[L_{\mathrm{in}}\left({{\mathrm{Gd}}_{\mathrm{in}}}^i\right),j]=\left\{\begin{array}{cc}1,& L_{\mathrm{in}}\left({{\mathrm{Gd}}_{\mathrm{in}}}^i\right)=j;\\ 0,& L_{\mathrm{in}}\left({{\mathrm{Gd}}_{\mathrm{in}}}^i\right)\≠j.\end{array}\right..$

Preferably, described according to described G _ex, ω _exwith Dic _exobtain outer peripheral areas feature F _ex, comprising:

By described G _exany a line G _ex ⁱwith ω _excorresponding point are multiplied, and obtain outer peripheral areas and can distinguish neighborhood vector Gd _ex ⁱ, ${\mathrm{Gd}}_{\mathrm{ex}}^i\left(k\right)={\mathrm{\ω}}_{\mathrm{ex}}\left(k\right)\·G_{\mathrm{ex}}^i\left(k\right),$ k＝1,...,M ².

By described Gd _ex ⁱquantize to Dic _exin the texture primitive nearest with it quantize mark value, obtain quantizing mark value L _ex(Gd _ex ⁱ), $L_{\mathrm{ex}}\left({{\mathrm{Gd}}_{\mathrm{ex}}}^i\right)=\underset{j}{\arg \min }\left\{d({{\mathrm{Gd}}_{\mathrm{ex}}}^i,{{\mathrm{Dic}}_{\mathrm{ex}}}^j)\right\},j=1,...,K_{\mathrm{dic}}.;$

After obtaining the quantification mark value of each pixel of outer peripheral areas, calculate the feature F of histogram as outer peripheral areas of all mark value _ex, $F_{\mathrm{ex}}\left(j\right)=\underset{i=1}{\overset{N_{\mathrm{ex}}}{\mathrm{\Σ}}}\mathrm{\δ}[L_{\mathrm{ex}}\left({{\mathrm{Gd}}_{\mathrm{ex}}}^i\right),j]/N_{\mathrm{ex}};$

Wherein N _exthe quantity of outer peripheral areas pixel.

Preferably, described merges to the central area of described training image and the feature of outer peripheral areas the proper vector obtaining described training image, comprising:

By described central area feature F _inbe multiplied by factor lambda, and with described outer peripheral areas feature F _exjoin end to end, as proper vector F=[the λ F of described training image _inf _ex].

Preferably, the described proper vector training k nearest neighbor sorter utilizing described training image, comprising:

The combination of eigenvectors of all training images is become training sample eigenmatrix, combined training sample characteristics matrix and its classification information, training k nearest neighbor sorter, is input to the distance dis (F of described k nearest neighbor sorter ₁, F ₂) be specifically calculated as follows:

$\mathrm{dis}(F_1,F_2)=\frac{1}{2}\underset{j=1}{\overset{K_{\mathrm{dic}}}{\mathrm{\Σ}}}\frac{{[F_1\left(j\right)-F_2\left(j\right)]}^2}{F_1\left(j\right)+F_2\left(j\right)}+\frac{1}{2}\underset{i=K_{\mathrm{dic}}+1}{\overset{2K_{\mathrm{dic}}}{\mathrm{\Σ}}}\frac{{[F_1\left(i\right)-F_1\left(i\right)]}^2}{F_1\left(i\right)+F_2\left(i\right)}$

Above-mentioned F ₁, F ₂be respectively the proper vector of two training images.

Preferably, described is input to the proper vector of the lump area image to be identified after described normalized the k nearest neighbor sorter trained, and obtains the good pernicious classification results of described lump area image to be identified, comprising:

By described training image feature extracting method, extract the proper vector F of image to be identified _q, by the proper vector F of image to be identified _qbe input to the k nearest neighbor sorter trained, determine in the proper vector of the training image in described k nearest neighbor sorter with F _qnearest front K width lump image, belongs to the quantity L of benign tumors in calculating K width lump image ₁with the quantity L belonging to Malignant mass in K width lump image ₂if, L ₁>=L ₂, then lump image to be identified is judged as optimum, if L ₁<L ₂, then lump image to be identified is judged as pernicious.

The technical scheme provided as can be seen from the embodiment of the invention described above, the embodiment of the present invention is solve the problem that texture primitive does not consider pixel space distribution information, from tumor growth Track character, design many coaxial layers template, by region and outer peripheral areas centered by lump Region dividing.Consider the feature otherness of central area and outer peripheral areas, respectively texture dictionary built to central area and outer peripheral areas and extract feature.Lose the problem of classification information when building texture dictionary for solving texture primitive, linear discriminant analysis is incorporated into texture primitive dictionary and builds by the present invention, proposition can realize the good malignant characteristics of lump and extract by subregion texture primitive, achieves and perniciously effectively to classify to image lump is good based on texture primitive.

The aspect that the present invention adds and advantage will part provide in the following description, and these will become obvious from the following description, or be recognized by practice of the present invention.

Accompanying drawing explanation

In order to be illustrated more clearly in the technical scheme of the embodiment of the present invention, below the accompanying drawing used required in describing embodiment is briefly described, apparently, accompanying drawing in the following describes is only some embodiments of the present invention, for those of ordinary skill in the art, under the prerequisite not paying creative work, other accompanying drawing can also be obtained according to these accompanying drawings.

Fig. 1 is the processing flow chart of the good pernicious sorting technique of the image lump based on texture primitive of the embodiment of the present invention;

Fig. 2 is the lump area sample image for training in the step 1 in the treatment scheme of the good pernicious sorting technique of the image lump based on texture primitive of the embodiment of the present invention;

Fig. 3 is that training image is divided into central area R by the above-mentioned rectangle template of use that the embodiment of the present invention provides _inwith outer peripheral areas R _exschematic diagram;

Fig. 4 is that the embodiment of the present invention is based on lump area image to be identified in the step 14 in the treatment scheme of the good pernicious sorting technique of image lump of texture primitive.

Embodiment

Be described below in detail embodiments of the present invention, the example of described embodiment is shown in the drawings, and wherein same or similar label represents same or similar element or has element that is identical or similar functions from start to finish.Being exemplary below by the embodiment be described with reference to the drawings, only for explaining the present invention, and can not limitation of the present invention being interpreted as.

Those skilled in the art of the present technique are appreciated that unless expressly stated, and singulative used herein " ", " one ", " described " and " being somebody's turn to do " also can comprise plural form.Should be further understood that, the wording used in instructions of the present invention " comprises " and refers to there is described feature, integer, step, operation, element and/or assembly, but does not get rid of and exist or add other features one or more, integer, step, operation, element, assembly and/or their group.Should be appreciated that, when we claim element to be " connected " or " coupling " to another element time, it can be directly connected or coupled to other elements, or also can there is intermediary element.In addition, " connection " used herein or " coupling " can comprise wireless connections or couple.Wording "and/or" used herein comprises one or more arbitrary unit listing item be associated and all combinations.

Those skilled in the art of the present technique are appreciated that unless otherwise defined, and all terms used herein (comprising technical term and scientific terminology) have the meaning identical with the general understanding of the those of ordinary skill in field belonging to the present invention.Should also be understood that those terms defined in such as general dictionary should be understood to have the meaning consistent with the meaning in the context of prior art, unless and define as here, can not explain by idealized or too formal implication.

The invention provides the good pernicious sorting technique of a kind of image lump based on texture primitive, particular flow sheet as shown in Figure 1, comprises sorter training process and uses the sorter trained to carry out the good pernicious assorting process of lump:

One, sorter training process, specifically comprises the following steps:

Step 1: brightness normalization is carried out, to reduce the impact of image I contrast change to arbitrary width training lump area image I.Its concrete grammar is:

I _n(x,y)＝[I(x,y)-I _min]/(I _max-I _min)

Wherein I (x, y) is the gray-scale value of pixel (x, y), I _minand I _maxthe minimum of image I and maximum gradation value, I _n(x, y) is the gray-scale value after pixel (x, y) normalization.

Step 2: to arbitrary width training image, builds the two-value rectangle template with two coaxial layers: Mask _inand Mask _ex, Mask _inand Mask _exlength and be widely respectively the long and wide α of training image doubly, α is the coefficient of setting.Use above-mentioned rectangle template that training image is divided into central area R _inwith outer peripheral areas R _ex.Be specially: R _in=I _n× Mask _in, R _ex=I _n× Mask _ex.

Step 3: to arbitrary width training image, extract M × M neighborhood gray scale of each pixel in picture centre region, it is M that this M × M neighborhood gray scale is adjusted to length ²row vector, all row vector organization center region Neighborhood matrix G _in.If training image belongs to benign tumors, then the classification information of each pixel in this picture centre region is 1, if training image belongs to Malignant mass, then the classification information of each pixel in this picture centre region is 2, the classification information organization center area classification information matrix of all pixels.

Step 4: to arbitrary width training image, extract M × M neighborhood gray scale of each pixel of outer peripheral areas, it is M that this M × M neighborhood gray scale is adjusted to length ²row vector, all row vectors composition outer peripheral areas Neighborhood matrix G _ex.If training image belongs to benign tumors, then the classification information of each pixel of this image outer peripheral areas is 1, if training image belongs to Malignant mass, then the classification information of each pixel of this image outer peripheral areas is 2, the classification information composition outer peripheral areas classification information matrix of all pixels.

Step 5: to central area Neighborhood matrix G _inclassification information matrix application LDA (Linear Discriminant Analysis, linear discriminant analysis) with its correspondence, obtains central area neighborhood weight row vector ω _in.

Step 6: to outer peripheral areas Neighborhood matrix G _exwith the classification information matrix application LDA of its correspondence, obtain outer peripheral areas neighborhood weight row vector ω _ex.

Step 7: by central area neighborhood weight row vector ω _inwith Neighborhood matrix G _inevery a line corresponding point be multiplied, obtain central area and can distinguish Neighborhood matrix Gd _in.

Step 8: by outer peripheral areas neighborhood weight row vector ω _exwith Neighborhood matrix G _exevery a line corresponding point be multiplied, obtain outer peripheral areas and can distinguish Neighborhood matrix Gd _ex.

Step 9: to Gd _inand Gd _exapply k mean cluster respectively, obtain central area and can distinguish texture primitive dictionary Dic _inand outer peripheral areas can distinguish texture primitive dictionary Dic _ex.The initial center of k mean cluster adopts the mode of random selecting, for reducing the impact of random selecting, carries out repeatedly k mean cluster, and center corresponding when selecting training accuracy the highest is as Dic _inand Dic _ex.Each Dic _inand Dic _excomputation process specific as follows:

1) corresponding to benign tumors training image Gd _inapplication k mean cluster, obtains cluster centre Cb _in, the quantity of cluster centre is K _dic, each center is M ²the row vector of length.

2) corresponding to Malignant mass training image Gd _inapplication k mean cluster, obtains cluster centre Cm _in, the quantity of cluster centre is K _dic, each center is M ²the row vector of length.

3) Cb is combined _inand Cm _in, form 2K _dicrow M ²the matrix of row $\left[\begin{array}{c}{\mathrm{Cb}}_{\mathrm{in}}\\ {\mathrm{Cm}}_{\mathrm{in}}\end{array}\right],$ Remove with other row vector average Euclidean apart from close K _dicindividual row vector, finally obtains K _dicindividual center, it can be used as the texture dictionary Dic of central area _in.

4) corresponding to benign tumors training image Gd _exapplication k mean cluster, obtains cluster centre Cb _ex, the quantity of cluster centre is K _dic, each center is M ²the row vector of length.

5) corresponding to Malignant mass training image Gd _exapplication k mean cluster, obtains cluster centre Cm _ex, the quantity of cluster centre is K _dic, each center is M ²the row vector of length.

6) Cb is combined _exand Cm _ex, form 2K _dicrow M ²the matrix of row $\left[\begin{array}{c}{\mathrm{Cb}}_{\mathrm{ex}}\\ {\mathrm{Cm}}_{\mathrm{ex}}\end{array}\right],$ Remove with other row vector average Euclidean apart from close K _dicindividual row vector, finally obtains K _dicindividual center, it can be used as the texture dictionary Dic of outer peripheral areas _ex.

Step 10: to arbitrary width training image, uses G _in, ω _inwith Dic _in, obtain the feature F of central area _in.Specific as follows:

1) by G _inany a line G _in ⁱwith ω _incorresponding point are multiplied, and obtain central area and can distinguish neighborhood vector Gd _in ⁱ, be specially: ${\mathrm{Gd}}_{\mathrm{in}}^i\left(k\right)={\mathrm{\&omega;}}_{\mathrm{in}}\left(k\right)\&CenterDot;{G_{\mathrm{in}}}^i\left(k\right),$ k＝1,...,M ².

2) by Gd _in ⁱquantize to Dic _inin the texture primitive nearest with it, obtain quantizing mark value L _in(Gd _in ⁱ), it is specifically calculated as:

$L_{\mathrm{in}}\left({{\mathrm{Gd}}_{\mathrm{in}}}^i\right)=\underset{j}{\arg \min }\left\{d({{\mathrm{Gd}}_{\mathrm{in}}}^i,{{\mathrm{Dic}}_{\mathrm{in}}}^j)\right\},j=1,...,K_{\mathrm{dic}}.$

Wherein d is Euclidean distance, K _diccan distinguish the texture primitive quantity that texture dictionary comprises, with the K in step 9 _dicidentical.

3), after obtaining the quantification mark value of each pixel in central area, the histogram of all quantification mark value is calculated, as the feature F of central area _in, it is specifically calculated as:

Wherein, N _inbe the quantity of central area pixel, δ function is:

$\mathrm{\&delta;}[L_{\mathrm{in}}\left({{\mathrm{Gd}}_{\mathrm{in}}}^i\right),j]=\left\{\begin{array}{cc}1,& L_{\mathrm{in}}\left({{\mathrm{Gd}}_{\mathrm{in}}}^i\right)=j;\\ 0,& L_{\mathrm{in}}\left({{\mathrm{Gd}}_{\mathrm{in}}}^i\right)\&NotEqual;j.\end{array}\right..$

Step 11: to arbitrary width training image, uses G _ex, ω _exwith Dic _ex, obtain the feature F of outer peripheral areas _ex.Specific as follows:

1) by G _exany a line G _ex ⁱwith ω _excorresponding point are multiplied, and obtain outer peripheral areas and can distinguish neighborhood vector Gd _ex ⁱ, ${\mathrm{Gd}}_{\mathrm{ex}}^i\left(k\right)={\mathrm{\&omega;}}_{\mathrm{ex}}\left(k\right)\&CenterDot;G_{\mathrm{ex}}^i\left(k\right),$ k＝1,...,M ².

2) by Gd _ex ⁱquantize to Dic _exin the texture primitive nearest with it, obtain quantizing mark value L _ex(Gd _ex ⁱ), it is specifically calculated as: $L_{\mathrm{ex}}\left({{\mathrm{Gd}}_{\mathrm{ex}}}^i\right)=\underset{j}{\arg \min }\left\{d({{\mathrm{Gd}}_{\mathrm{ex}}}^i,{{\mathrm{Dic}}_{\mathrm{ex}}}^j)\right\},j=1,...,K_{\mathrm{dic}}..$

3), after obtaining the quantification mark value of each pixel of outer peripheral areas, the feature F of histogram as outer peripheral areas of all mark value is calculated _ex, it is specifically calculated as:

Wherein N _exthe quantity of outer peripheral areas pixel.

Step 12: central area feature F _inbe multiplied by factor lambda, and with outer peripheral areas feature F _exjoin end to end, as proper vector F=[the λ F of this training image _inf _ex].

Step 13: the combination of eigenvectors of all training images is become training sample eigenmatrix, combined training sample characteristics matrix and its classification information, training k nearest neighbor sorter.Be input to the distance dis (F of k nearest neighbor sorter ₁, F ₂) be specifically calculated as follows:

$\mathrm{dis}(F_1,F_2)=\frac{1}{2}\underset{j=1}{\overset{K_{\mathrm{dic}}}{\mathrm{\&Sigma;}}}\frac{{[F_1\left(j\right)-F_2\left(j\right)]}^2}{F_1\left(j\right)+F_2\left(j\right)}+\frac{1}{2}\underset{i=K_{\mathrm{dic}}+1}{\overset{2K_{\mathrm{dic}}}{\mathrm{\&Sigma;}}}\frac{{[F_1\left(i\right)-F_1\left(i\right)]}^2}{F_1\left(i\right)+F_2\left(i\right)}$

Above-mentioned F ₁, F ₂be respectively the proper vector of two training images.

Two, use the sorter trained to carry out the good pernicious assorting process of lump, specifically comprise the following steps:

Step 14: the normalized described in lump area image to be identified carry out step 1.

Step 15: the proper vector F obtaining this lump area image according to the process described by step 10 to step 12 _q.

Step 16: by described training image feature extracting method, extract the proper vector F of image to be identified _q, by the proper vector F of image to be identified _qbe input to the k nearest neighbor sorter trained, determine in the proper vector of the training image in described k nearest neighbor sorter with F _qnearest front K width lump image, belongs to the quantity L of benign tumors in calculating K width lump image ₁with the quantity L belonging to Malignant mass in K width lump image ₂if, L ₁>=L ₂, then lump image to be identified is judged as optimum, if L ₁<L ₂, then lump image to be identified is judged as pernicious.

For ease of the understanding to the embodiment of the present invention, be further explained explanation below in conjunction with accompanying drawing for a specific embodiment, this embodiment does not form the restriction to the embodiment of the present invention.

Embodiment one

It is a kind of in conjunction with many coaxial form Region dividing and the good pernicious sorting technique of lump can distinguishing texture primitive that this embodiment provides, and carries out concrete enforcement by following steps:

Step 1: brightness normalization is carried out to training lump area image, the gray-scale value of image is all normalized between [0,1], trains the sample instantiation of lump area image in the present embodiment as shown in Figure 2.

Step 2: to arbitrary width training image, builds the rectangle template with two coaxial layers, as shown in Figure 3.Use this template that training image is divided into central area and outer peripheral areas.In this example, α=0.5.

Step 3: to arbitrary width training image, extracts M × M neighborhood gray scale of each pixel in central area, neighborhood gray scale is adjusted to length M ²row vector, all row vector organization center region Neighborhood matrix G _in, then computing center's area classification information matrix.In this example, M=5.

Step 4: to arbitrary width training image, extract M × M neighborhood gray scale of each pixel of outer peripheral areas, neighborhood gray scale being adjusted to length is M ²row vector, all row vectors composition outer peripheral areas Neighborhood matrix G _ex, then calculate outer peripheral areas classification information matrix.

Step 5: to central area Neighborhood matrix G _inwith the classification information matrix application LDA of its correspondence, obtain central area neighborhood weight row vector ω _in.

Step 6: to outer peripheral areas Neighborhood matrix G _exwith the classification information matrix application LDA of its correspondence, obtain outer peripheral areas neighborhood weight row vector ω _ex.

Step 7: by central area neighborhood weight row vector ω _inwith Neighborhood matrix G _inevery a line corresponding point be multiplied, obtain central area and can distinguish Neighborhood matrix Gd _in.

Step 8: by outer peripheral areas neighborhood weight row vector ω _exwith Neighborhood matrix G _exevery a line corresponding point be multiplied, obtain outer peripheral areas and can distinguish Neighborhood matrix Gd _ex.

Step 9: to Gd _inand Gd _exapply k mean cluster respectively, obtain central area and can distinguish texture primitive dictionary Dic _inand outer peripheral areas can distinguish texture primitive dictionary Dic _ex.In this example, the texture primitive quantity K that texture dictionary comprises can be distinguished _dic=100.

Step 10: to arbitrary width training image, uses G _in, ω _inwith Dic _in, obtain the feature F of central area _in.

Step 11: to arbitrary width training image, uses G _ex, ω _exwith Dic _ex, obtain the feature F of outer peripheral areas _ex.

Step 12: central area feature F _inbe multiplied by factor lambda, and with outer peripheral areas feature F _exjoin end to end, as the proper vector F of this training image.In this example, λ=0.8.

Step 13: the combination of eigenvectors of all training images is become training sample eigenmatrix, combined training sample characteristics matrix and its classification information, training k nearest neighbor sorter.In this example, the parameter K=3 of k nearest neighbor sorter.

Step 14: the normalized described in lump area image to be identified carry out step 1, the lump area image to be identified that this embodiment provides as shown in Figure 4.

Step 15: the proper vector F obtaining this lump area image according to the process described by step 10 to step 12 _q.

Step 16: by the proper vector F of image to be identified _qbe input to the k nearest neighbor sorter trained, obtain classification results.

In sum, the embodiment of the present invention is solve the problem that texture primitive does not consider pixel space distribution information, from tumor growth Track character, designs many coaxial layers template, by region and outer peripheral areas centered by lump Region dividing.Consider the feature otherness of central area and outer peripheral areas, respectively texture dictionary built to central area and outer peripheral areas and extract feature.Lose the problem of classification information when building texture dictionary for solving texture primitive, linear discriminant analysis is incorporated into texture primitive dictionary and builds by the present invention, proposition can realize the good malignant characteristics of lump and extract by subregion texture primitive, achieves and perniciously effectively to classify to image lump is good based on texture primitive.

The embodiment of the present invention, without the need to carrying out Accurate Segmentation to lump edges of regions, directly carries out good pernicious identification to the region comprising lump.Be applied to the clinical doctor of can be and objectively auxiliary suggestion is provided, reduce unnecessary live tissue puncturing.

One of ordinary skill in the art will appreciate that: accompanying drawing is the schematic diagram of an embodiment, the module in accompanying drawing or flow process might not be that enforcement the present invention is necessary.

As seen through the above description of the embodiments, those skilled in the art can be well understood to the mode that the present invention can add required general hardware platform by software and realizes.Based on such understanding, technical scheme of the present invention can embody with the form of software product the part that prior art contributes in essence in other words, this computer software product can be stored in storage medium, as ROM/RAM, magnetic disc, CD etc., comprising some instructions in order to make a computer equipment (can be personal computer, server, or the network equipment etc.) perform the method described in some part of each embodiment of the present invention or embodiment.

Each embodiment in this instructions all adopts the mode of going forward one by one to describe, between each embodiment identical similar part mutually see, what each embodiment stressed is the difference with other embodiments.Especially, for device or system embodiment, because it is substantially similar to embodiment of the method, so describe fairly simple, relevant part illustrates see the part of embodiment of the method.Apparatus and system embodiment described above is only schematic, the wherein said unit illustrated as separating component or can may not be and physically separates, parts as unit display can be or may not be physical location, namely can be positioned at a place, or also can be distributed in multiple network element.Some or all of module wherein can be selected according to the actual needs to realize the object of the present embodiment scheme.Those of ordinary skill in the art, when not paying creative work, are namely appreciated that and implement.

The above; be only the present invention's preferably embodiment, but protection scope of the present invention is not limited thereto, is anyly familiar with those skilled in the art in the technical scope that the present invention discloses; the change that can expect easily or replacement, all should be encompassed within protection scope of the present invention.Therefore, protection scope of the present invention should be as the criterion with the protection domain of claim.

Claims (8)

1., based on the good pernicious sorting technique of image lump of texture primitive, it is characterized in that, comprising:

Training lump area image is normalized, lump area image after normalized is divided into central area and outer peripheral areas, described central area and outer peripheral areas are built respectively and can distinguish texture dictionary, then obtain the feature of described central area and outer peripheral areas respectively;

The feature of described central area and outer peripheral areas is merged to the proper vector of the lump area image after obtaining described normalized;

The combination of eigenvectors of the training lump area image after described normalized is become training sample eigenmatrix, training k nearest neighbor sorter;

Lump area image to be identified is normalized, the proper vector F of lump image-region after extraction normalized _q, by the proper vector F of lump area image _qbe input to the k nearest neighbor sorter trained, obtain the good pernicious classification results of described lump area image to be identified.

2. the good pernicious sorting technique of the image lump based on texture primitive according to claim 1, it is characterized in that, described is normalized training lump area image, training image after normalized is divided into central area and outer peripheral areas, the central area of described training image and outer peripheral areas are built respectively and can distinguish texture dictionary, then obtain the central area of described training image and the feature of outer peripheral areas respectively, comprising:

Brightness normalization is carried out to the training image of image lump, image I after normalization _nrepresent, build the rectangle template with two coaxial layers, Mask _inand Mask _ex, described Mask _inand Mask _exlength and be widely respectively the length of described lump training image and wide setting multiple, use described rectangle template that described lump training image is divided into central area R _inwith outer peripheral areas R _ex, R _in=I × Mask _in, R _ex=I × Mask _ex;

Extract the central area R of described training image _inm × M neighborhood gray scale of each pixel, it is M that described M × M neighborhood gray scale is adjusted to length ²row vector, all row vector organization center region Neighborhood matrix G _in; Extract outer peripheral areas R _exm × M neighborhood gray scale of each pixel, it is M that this M × M neighborhood gray scale is adjusted to length ²row vector, all row vectors composition outer peripheral areas Neighborhood matrix G _ex;

To described central area Neighborhood matrix G _inwith the classification information matrix application linear discriminant analysis of its correspondence, obtain central area neighborhood weight row vector ω _in; To described outer peripheral areas Neighborhood matrix G _exwith the classification information matrix application linear discriminant analysis of its correspondence, obtain outer peripheral areas neighborhood weight row vector ω _ex;

By described central area neighborhood weight row vector ω _inwith described central area Neighborhood matrix G _inevery a line corresponding point be multiplied, obtain central area and can distinguish Neighborhood matrix Gd _in; By described outer peripheral areas neighborhood weight row vector ω _exwith described outer peripheral areas Neighborhood matrix G _exevery a line corresponding point be multiplied, obtain outer peripheral areas and can distinguish Neighborhood matrix Gd _ex;

To described Gd _inand Gd _exapply k mean cluster respectively, obtain central area and can distinguish texture primitive dictionary Dic _inand outer peripheral areas can distinguish texture primitive dictionary Dic _ex;

According to described training image central area Neighborhood matrix G _in, ω _inwith Dic _inobtain training image central area feature F _in; According to described training image outer peripheral areas Neighborhood matrix G _ex, ω _exwith Dic _exobtain training image outer peripheral areas feature F _ex.

3. the good pernicious sorting technique of the image lump based on texture primitive according to claim 2, is characterized in that, described to training image Gd _inand Gd _exapply k mean cluster respectively, obtain central area and can distinguish texture primitive dictionary Dic _inand outer peripheral areas can distinguish texture primitive dictionary Dic _ex, comprising:

The Gd corresponding to benign tumors training image _inapplication k mean cluster, obtains cluster centre Cb _in, the quantity of cluster centre is to distinguish the primitive quantity K that texture primitive dictionary comprises _dic, each center is M ²the row vector of length; The Gd corresponding to Malignant mass training image _inapplication k mean cluster, obtains cluster centre Cm _in, the quantity of cluster centre is K _dic, each center is M ²the row vector of length; Combine described Cb _inand Cm _in, form 2K _dicrow M ²the matrix of row $\left[\begin{array}{c}C{b}_{\mathrm{in}}\\ {\mathrm{Cm}}_{\mathrm{in}}\end{array}\right],$ Remove with other row vector average Euclidean apart from close K _dicindividual row vector, finally obtains K _dicindividual center, by this K _dicindividual center is as the texture dictionary Dic of central area _in;

The Gd corresponding to benign tumors training image _exapplication k mean cluster, obtains cluster centre Cb _ex, the quantity of cluster centre is to distinguish the primitive quantity K that texture primitive dictionary comprises _dic, each center is M ²the row vector of length; The Gd corresponding to Malignant mass training image _exapplication k mean cluster, obtains cluster centre Cm _ex, the quantity of cluster centre is K _dic, each center is M ²the row vector of length; Combine described Cb _exand Cm _ex, form 2K _dicrow M ²the matrix of row $\left[\begin{array}{c}{\mathrm{Cb}}_{\mathrm{ex}}\\ {\mathrm{Cm}}_{\mathrm{ex}}\end{array}\right],$ Remove with other row vector average Euclidean apart from close K _dicindividual row vector, finally obtains K _dicindividual center, by this K _dicindividual center is as the texture dictionary Dic of outer peripheral areas _ex.

4. the good pernicious sorting technique of the image lump based on texture primitive according to claim 2, is characterized in that, described according to described G _in, ω _inwith Dic _inobtain central area feature F _in, comprising:

By described G _inany a line G _in ⁱwith ω _incorresponding point are multiplied, and obtain central area and can distinguish neighborhood vector Gd _in ⁱ, Gd _i ⁱ _n(k)=ω _in(k) G _in ⁱ(k), k=1 ..., M ².; By described Gd _in ⁱquantize to Dic _inin the texture primitive nearest with its Euclidean distance quantize mark value, obtain quantize after mark value L _in(Gd _in ⁱ), it is specifically calculated as:

$L_{\mathrm{in}}\left({{\mathrm{Cd}}_{\mathrm{in}}}^i\right)=\underset{j}{\arg \min }\left\{d({{\mathrm{Gd}}_{\mathrm{in}}}^i,D{{\mathrm{ic}}_{\mathrm{in}}}^j)\right\},j=1,...,K_{\mathrm{dic}},$

Wherein d is Euclidean distance, K _dicto distinguish the texture primitive quantity that texture dictionary comprises;

After obtaining the quantification mark value of each pixel in central area, calculate the histogram of all quantification mark value, as the feature F of central area _in, it is specifically calculated as:

Wherein, N _inbe the quantity of central area pixel, δ function is:

$\mathrm{\&delta;}[L_{\mathrm{in}}\left({{\mathrm{Gd}}_{\mathrm{in}}}^i\right),j]=\left\{\begin{array}{cc}1,& L_{\mathrm{in}}\left({{\mathrm{Gd}}_{\mathrm{in}}}^i\right)=j;\\ 0,& L_{\mathrm{in}}\left({{\mathrm{Gd}}_{\mathrm{in}}}^i\right)\&NotEqual;j.\end{array}\right..$

5. the good pernicious sorting technique of the image lump based on texture primitive according to claim 2, is characterized in that, described according to described G _ex, ω _exwith Dic _exobtain outer peripheral areas feature F _ex, comprising:

By described G _exany a line G _ex ⁱwith ω _excorresponding point are multiplied, and obtain outer peripheral areas and can distinguish neighborhood vector Gd _ex ⁱ, ${\mathrm{Gd}}_{\mathrm{ex}}^i\left(k\right)={\mathrm{\&omega;}}_{\mathrm{ex}}\left(k\right)\&CenterDot;G_{\mathrm{ex}}^i\left(k\right),k=1,...,M^2.$

By described Gd _ex ⁱquantize to Dic _exin the texture primitive nearest with it quantize mark value, obtain the mark value L after quantizing _ex(Gd _ex ⁱ), $L_{\mathrm{ex}}\left({{\mathrm{Gd}}_{\mathrm{ex}}}^i\right)=\underset{j}{\arg \min }\left\{d({{\mathrm{Gd}}_{\mathrm{ex}}}^i,{{\mathrm{Dic}}_{\mathrm{ex}}}^j)\right\},j=1,...,K_{\mathrm{dic}}.;$

After obtaining the quantification mark value of each pixel of outer peripheral areas, calculate the feature F of histogram as outer peripheral areas of all mark value _ex, $F_{\mathrm{ex}}\left(j\right)=\underset{i=1}{\overset{N_{\mathrm{ex}}}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}[L_{\mathrm{ex}}\left({{\mathrm{Gd}}_{\mathrm{ex}}}^i\right),j]/N_{\mathrm{ex}};$

Wherein N _exthe quantity of outer peripheral areas pixel.

6. the good pernicious sorting technique of the image lump based on texture primitive according to claim 1, is characterized in that, described merges to the central area of described training image and the feature of outer peripheral areas the proper vector obtaining described training image, comprising:

By described central area feature F _inbe multiplied by factor lambda, and with described outer peripheral areas feature F _exjoin end to end, as proper vector F=[the λ F of described training image _inf _ex].

7. the good pernicious sorting technique of the image lump based on texture primitive according to claim 1,

It is characterized in that, the described proper vector training k nearest neighbor sorter utilizing described training image, comprising:

The combination of eigenvectors of all training images is become training sample eigenmatrix, combined training sample characteristics matrix and its classification information, training k nearest neighbor sorter, is input to the distance dis (F of described k nearest neighbor sorter ₁, F ₂) be specifically calculated as follows:

$\mathrm{dis}(F_1,F_2)=\frac{1}{2}\underset{j=1}{\overset{K_{\mathrm{dic}}}{\mathrm{\&Sigma;}}}\frac{{[F_1\left(j\right)-F_2\left(j\right)]}^2}{F_1\left(j\right)+F_2\left(j\right)}+\frac{1}{2}\underset{i=K_{\mathrm{dic}}+1}{\overset{2K_{\mathrm{dic}}}{\mathrm{\&Sigma;}}}\frac{{[F_1\left(i\right)-F_2\left(i\right)]}^2}{F_1\left(i\right)+F_2\left(i\right)}$

Above-mentioned F ₁, F ₂be respectively the proper vector of two training images.

8. the good pernicious sorting technique of the image lump based on texture primitive according to any one of claim 1 to 7, it is characterized in that, described is input to the proper vector of the lump area image to be identified after described normalized the k nearest neighbor sorter trained, obtain the good pernicious classification results of described lump area image to be identified, comprising:

By described training image feature extracting method, extract the proper vector F of image to be identified _q, by the proper vector F of image to be identified _qbe input to the k nearest neighbor sorter trained, determine in the proper vector of the training image in described k nearest neighbor sorter with F _qnearest front K width lump image, belongs to the quantity L of benign tumors in calculating K width lump image ₁with the quantity L belonging to Malignant mass in K width lump image ₂if, L ₁>=L ₂, then lump image to be identified is judged as optimum, if L ₁<L ₂, then lump image to be identified is judged as pernicious.