CN100484474C - Galactophore cancer computer auxiliary diagnosis system based on galactophore X-ray radiography - Google Patents

Abstract

The invention discloses a breast cancer computer auxiliary diagnosis method and system based on galactophore X-ray radiograph. A galactophore X-ray radiograph for diagnosis is input into the system of the invention firstly, and is treated through an extracting module in a region of interest, a partitioning module and a feature extracting module in a region of doubtful lump, hereby a series of relative feature values about the doubtful lump region; then the feature values are input into a trained classifier which classifies and identifies the doubtful lump region and lastly the computer automatic examined final result of portioning the doubtful lump region is located on the input galactophore X-ray radiograph for diagnosis and the calculated relative feature values of the region are displayed to a roentgenologist according to requirements to indicate the roentgenologist about the region needing special attention and relative important parameters of the region. The invention can improve the accurateness and efficiency of the breast cancer diagnosis by the roentgenologist to some extent and help the roentgenologist to bring forward a diagnostic opinion and a therapeutic schedule more objectively and effectively.

Description

Breast carcinoma computer-aided diagnosis system based on picture of mammary gland x-ray radiography

Technical field

The invention belongs to the computer analysis The Application of Technology field of medical image, be specifically related to a kind of breast carcinoma computer-aided diagnosis system based on picture of mammary gland x-ray radiography.

Background technology

Breast carcinoma is one of modal malignant tumor of middle and aged women, has occupied the first place of women's malignant tumor in China, and the trend that rises is year by year arranged.At present the prevention of breast carcinoma is not still had good plan, early diagnosis is the effective way that reduces M ﹠ M and improve the breast carcinoma cure rate.Studies show that, picture of mammary gland x-ray radiography (breast molybdenum target x ray technology) is a kind of inspection method of effectively carrying out early stage clinical asymptomatic breast carcinoma, by FDA (Food and Drug Adminstration) (Food and Drug Administrator, FDA) Ren Ke conventional breast disease screening method.

Limited by image-forming principle, normal galactophore tissue at the picture of mammary gland x-ray radiography middle-high density also can present the high brightness close with abnormal structure, young Asia women its focus fine and close for mammary gland tissue then is difficult for being found more, cause easily and fail to pinpoint a disease in diagnosis and mistaken diagnosis, and the early diagnosis of breast carcinoma depends on radiologist's experience, professional ability, degree of fatigue to a great extent, the subjective factors influence is bigger, especially in the generaI investigation activity of mastopathy, be difficult to guarantee efficient and the accuracy diagnosed.

The computer-aided diagnosis of breast carcinoma (computer-aided diagnosis, CAD) method and system thereof combine the computer powerful computing ability in the automatic analysis of picture of mammary gland x-ray radiography, effectively offer radiologist's valuable " second suggestion ", alleviating the diagnostic work amount, improving and the advantage that is difficult to compare to be arranged aspect diagnosis efficiency and the objectivity.The result of study of T.W.Freer etc. shows, the use of breast carcinoma computer aided diagnosing method and system thereof in clinical, the verification and measurement ratio that has obviously improved early stage malignant tumor is (specifically referring to T.W.Freer, et al. ＂ Screening Mammography withComputer-aided Detection:Prospective Study of 12,860 Patients in aCommunity Breast Center, ＂ Radiology, 220 (3): 781-786 (2001) .).Just because of this, the computer aided diagnosing method of breast carcinoma and system thereof are accepted and are applied in the clinical practice by worldwide medical institutions just gradually.

Usually, breast carcinoma computer aided diagnosing method and system thereof comprise that the auxiliary diagnosis of mammary gland tumor and the auxiliary diagnosis of microcalciffcation kitchen range are (specifically referring to David Gur., Bin Zheng, Yuan HsiangChang, Computerized detection of masses and microcalcifications in digitalmammograms, United States Patent No.5,627,907 and Zhejiang University's master thesis, calendar year 2001 December 31 days, mammary X-ray image computer-aided diagnosis algorithm research, Jiang Yuefeng).Auxiliary diagnosis is by computer galactophore image automatically to be analyzed and handled, identify the lesion region of suspicious lump in the image or suspicious microcalciffcation kitchen range, and obtain a series of relevant feature parameters of marking unsuspicious lump or suspicious microcalciffcation kitchen range zone, return as required and be shown to the radiologist, the advisory opinion of auxiliary its diagnosis is provided.At present, computer aided diagnosing method and system thereof have the characteristics of high sensitivity and low false positive rate in microcalciffcation kitchen range suspicious region context of detection, accepted and adopt (specifically referring to R.F.Brem by the radiologist, J.W.Hoffmeister, G.Zisman, et al. " Acomputer-aided detection system for evaluation of breast cancer bymammographic appearance and lesion size; " AJR, Am.J.Roentgenol184:893-896 (2005)), but the detection effect in the mammary gland tumor auxiliary diagnosis is very low, is not trusted by the radiologist.So, how designing a breast carcinoma computer aided diagnosing method and system thereof accurately locatees breast carcinoma lump zone suspicious in the picture of mammary gland x-ray radiography, and extract the effective regional relevant feature parameters of series, assist the doctor to carry out the focus that breast cancer diagnosis becomes our research exactly and effectively.

Summary of the invention

The object of the present invention is to provide the breast carcinoma computer-aided diagnosis system based on picture of mammary gland x-ray radiography, this system offers the advisory opinion of radiologist's diagnosis, helps to improve the efficient and the precision of breast carcinoma lump location and diagnosis.

Breast carcinoma computer-aided diagnosis system based on picture of mammary gland x-ray radiography provided by the invention comprises that input module, area-of-interest extraction module, suspicious lump cut apart module, suspicious lump provincial characteristics extraction module, classification diagnosis module and output module;

Input module is used to receive the picture of mammary gland x-ray radiography to be diagnosed of user's input, and sends the area-of-interest extraction module to;

The area-of-interest extraction module is used for extracting the area-of-interest of the picture of mammary gland x-ray radiography of input, obtains initial suspicious lump zone position information, and sends suspicious lump to and cut apart module;

Suspicious lump is cut apart the suspicious lump that module is used for area-of-interest that the area-of-interest extraction module is extracted and is split, and obtains the boundary information of suspicious lump, and sends suspicious lump provincial characteristics extraction module to;

Suspicious lump provincial characteristics extraction module calculates associated eigenvalue, and sends the classification diagnosis module to according to the boundary information of the suspicious lump that receives;

The classification diagnosis module is used for the eigenvalue input grader in each the suspicious lump zone that will calculate, computer automatic sorting identification is carried out in initial suspicious lump zone, determine final suspicious lump zone, and send the result to output module, the segmentation result in the final lump zone that will be detected by output module is positioned on the picture of mammary gland x-ray radiography to be diagnosed of input, and the regional associated eigenvalue that will calculate is shown to the user as required;

It is characterized in that:

Described area-of-interest extraction module is handled according to following step (2.1) to the process of (2.5):

Step (2.1) utilizes two-dimentional hyperbolic secant function to obtain template, calculates the degree of association of above-mentioned input picture of mammary gland x-ray radiography and this template, obtains the degree of association image;

The degree of association numerical value of the some pixels of step (2.2) degree of association image is 0 with this assignment then less than selected threshold, is 1 otherwise compose; In binaryzation degree of association image, extract all connected regions, and have the center of the point of maximal correlation degree in the corresponding region in the degree of association image that obtains of extraction step (2.1) as this connected region;

Step (2.3) utilizes two-dimentional hyperbolic secant function to obtain and the middle different template of template size of step (2.1); Each connected region for step (2.2) extraction, each yardstick template center is moved on to the position of center on original input picture of mammary gland x-ray radiography of connected region, calculate the degree of association of corresponding subimage in each yardstick template and the original input picture of mammary gland x-ray radiography image respectively, with the final degree of association of the maximum in all degree of association that calculate, choose a threshold value final degree of association is got rid of as the false positive zone less than the connected region of threshold value as this connected region;

Step (2.4) is got rid of the connected region of 5-30 pixel size and the connected region of strip again for the connected region that is not excluded in the step (2.3);

Step (2.5) to the connected region that remains, is center with its geometric center through the zone screening of (2.4), intercepting square in original input galactophore image, and this square area of extraction is area-of-interest or is called initial suspicious lump zone;

Described suspicious lump is cut apart module and is cut apart suspicious lump according to following step in area-of-interest:

Step (3.1) is obtained the relevant contour line group of gradient characteristic of region of interest area image, be end points afterwards with the region of interest centers, from zero angle, outwards draw some rays successively by counter clockwise direction at interval with equal angles, try to achieve the intersection point of each bar ray and contour line group, obtain boundary candidates point on every ray with this;

Step (3.2) obtains many boundary candidate lines by the candidate boundary point on each the bar ray that obtains in (3.1), selects wherein optimum boundary candidate line as suspicious lump boundary line again.

The present invention at first imports a picture of mammary gland x-ray radiography to be diagnosed to system, by area-of-interest extract, suspicious lump is cut apart and the feature extraction in suspicious lump zone (zone that comprises suspicious lump), obtain a series of associated eigenvalue of suspicious lump, then eigenvalue is imported grader classification diagnosis is carried out in initial suspicious lump zone, the segmentation result in the final suspicious lump zone that computer is detected automatically is positioned on the picture of mammary gland x-ray radiography to be diagnosed of input at last, and the serial associated eigenvalue that will calculate is shown to the user as required.In a word, the inventive method automatically detects suspicious mammary gland tumor by a breast carcinoma computer-aided diagnosis system based on picture of mammary gland x-ray radiography, suspicious lump position and shape are provided, and provide a series of relevant feature parameters of suspicious lump as required, thereby accuracy and the efficient of radiologist to breast cancer diagnosis has been improved in zone that the prompting radiologist need pay close attention to and the relevant emphasis parameter in zone to a certain extent.

Description of drawings

Fig. 1 is the flow chart that the present invention is based on the breast carcinoma computer aided diagnosing method of picture of mammary gland x-ray radiography;

Fig. 2 is the structural representation that the present invention is based on the breast carcinoma computer-aided diagnosis system of picture of mammary gland x-ray radiography;

Fig. 3 is for obtaining the process sketch map of candidate boundary point in the embodiment of the invention;

Fig. 4 is the feature list of the embodiment of the invention to suspicious lump region extraction;

Fig. 5 is an embodiment of the invention grader training process sketch map.

The specific embodiment

The present invention is further detailed explanation below in conjunction with accompanying drawing and example.

As shown in Figure 1, the inventive method may further comprise the steps:

(1) input one width of cloth picture of mammary gland x-ray radiography to be diagnosed.

(2) on the picture of mammary gland x-ray radiography of input, extract area-of-interest, obtain initial suspicious lump regional location.

Entire image is analyzed, not only had bulk redundancy information, also introduce mistake easily.For speed and the accuracy that improves processing, process object need be narrowed down to some zonules from entire image, i.e. area-of-interest, the position of these area-of-interests also is the initial suspicious lump regional location of subsequent treatment.

For area-of-interest extract existing research adopt about two width of cloth breast image control methods (specifically referring to F.F.Yin, M.L.Giger, Doi Kunio, and et al. " Computerized detection of massesin digital mammograms:analysis of bilateral subtraction images; " MedicalPhysics, 18:955-963 (1991) .), the Laplce of Gaussian (Laplacian of Gaussian, LoG) method, poor (the Difference of Gaussian of Gaussian function, DoG) method is (specifically referring to B.Zheng, Y.H.Chang, D.Gu. " Computerized detection of masses in digitizedmammograms using single-image segmentation and a multilayer topographicfeature analysis; " Acad.Radiol., 2:959-966 (1995) .) etc.

Consider lump have high brightness, be similar to circle and owing to and the features such as contrast that produce of the gray difference of surrounding tissue, the inventive method adopts template matching method location area-of-interest, promptly initial suspicious lump zone.Be implemented as follows:

(2.1) utilize two-dimentional hyperbolic secant (sech) function to obtain template T, calculate the degree of association of above-mentioned input picture of mammary gland x-ray radiography and template T, obtain the degree of association image;

Adopt two-dimentional hyperbolic secant (sech) function to generate the template T of size for (2L+1) * (2L+1)

$\sec h(x,y)=\frac{2\×\mathrm{\α}}{e^{-\mathrm{\β}\×(x^2+y^2)}+e^{\mathrm{\β}\×(x^2+y^2)}}---\left(1\right)$

Wherein, be initial point with the template center, x, y represent horizontal stroke, the vertical coordinate in the template, and span is respectively [L, L], and L is a positive integer.α=input picture of mammary gland x-ray radiography maximum gray scale-1, β=ln (2 * α)/L * L.

Template T is moved by pixel on the picture of mammary gland x-ray radiography of input, utilize the subimage S of formula (2) calculation template T and template T covering degree of association cor (T, S)

$\mathrm{cor}(T,S)=\frac{{\mathrm{\μ}}_{\mathrm{st}}-{\mathrm{\μ}}_s{\mathrm{\μ}}_t}{{\mathrm{\σ}}_s\×{\mathrm{\σ}}_t}---\left(2\right)$

Wherein, μ _StBe the meansigma methods of subimage S and template T respective pixel gray scale product, μ _tAnd μ _sBe the meansigma methods of pixel grey scale in template T and the subimage S, σ _tAnd σ _sVariance yields for pixel grey scale in template T and the subimage S.Use such method, each pixel all can calculate a degree of association numerical value relevant with template T in the input picture of mammary gland x-ray radiography, and the degree of association numerical range is [1,1].Afterwards all degree of association are demarcated, obtained the degree of association image.Wherein, the degree of association less than 0 is changed to 0 with it; Otherwise, multiply by input picture of mammary gland x-ray radiography maximum gray scale with degree of association numerical value.

(2.2) utilize selected threshold that above-mentioned degree of association image is carried out binary conversion treatment, in binaryzation degree of association image, extract the center of all connected regions and definite each connected region;

Choose an appropriate threshold T _Low, the degree of association image that step (2.1) is obtained carries out binary conversion treatment, and even the degree of association numerical value of the some pixels of degree of association image is less than T _Low, be 0 then with this assignment, be 1 otherwise compose.In the degree of association binary image, extract several connected regions.For each connected region, get in the degree of association image that step (2.1) obtains and have the center of the point of maximal correlation degree in the corresponding region as this connected region.

(2.3) redefine and the middle different template of template size of step (2.1) with the described definition template method of step (2.1), each connected region that step (2.2) is extracted is carried out multiscale analysis;

Define template method described in (2.1) set by step,, redefine and obtain 3 templates according to 200%, 33%, 66% of original template size 2L+1.Each connected region for step (2.2) extraction, each yardstick template center is moved on to the position of center on original input picture of mammary gland x-ray radiography of connected region, calculate the degree of association of corresponding subimage in each yardstick template and the original input picture of mammary gland x-ray radiography image respectively, maximum in the degree of association that 3 degree of association calculating and step (2.1) are obtained is chosen an appropriate threshold T as the final degree of association of this connected region _High(T _HighGreater than T _Low), with final degree of association less than T _HighConnected region get rid of as false positive zone (the non-lump zone that computer is thought).

(2.4) utilizing area and regular shape that step (2.3) is handled back residue connected region screens;

Connected region for through not being excluded in step (2.3) multiscale analysis method adopts simple area and regular shape further to screen.The connected region of area less (being generally 5-30 pixel size) generally speaking, usually corresponding is mammary gland calcification kitchen range, and the zone of strip usually corresponding be normal gland tissue, will satisfy the connected region of both of these case and get rid of as the false positive zone.

(2.5) area-of-interest extracts;

Zone screening through step (2.4), to the connected region that remains, with its geometric center is the center, intercepting (being generally 125 * 125 pixel sizes) square in original input galactophore image, and this square area of extraction is area-of-interest or is called initial suspicious lump zone.

(3) in the area-of-interest that step (2) is extracted, be partitioned into suspicious lump, determine the border of suspicious lump.

After extracting area-of-interest, need cut apart the suspicious lump that comprises in the area-of-interest, accurately determine its border.

Cutting apart existing research for suspicious lump adopts multilamellar landform region growing (specifically referring to B.Zheng, Y.H.Chang, D.Gu. " Computerized detection of masses in digitizedmammograms using single-image segmentation and a multilayer topographicfeature analysis; " Acad.Radiol., 2:959-966 (1995) .), movable contour model is (specifically referring to N.Karssemeijer, Segmentation of suspicious densities in digital mammograms, Med.Phys, vol.28:259-266 (2001) .), based on cutting apart of multiresolution analysis (specifically referring to Liu, C.Babbs, and E.Delp, " Multiresolution Detection of Spiculated Lesions inDigital Mammograms; " IEEE Transactions on Image Processing, 10 (6): 874884 (2001) .), threshold value based on fuzzy entropy is cut apart (specifically referring to S.Amr R.Abdel-Dayem, Mahmoud R.El-Sakka, " Fuzzy entropy based detection of suspicious masses indigital mammogram; " Images Proceedings of the 2005 IEEE Engineering inMedicine and Biology 27th Annual Conference, 4017-4022 (September, 2005) .) etc.

The inventive method adopts a kind of dividing method based on image gradient and dynamic programming method.Be implemented as follows:

(3.1) utilize the gradient correlated characteristic of region of interest area image, determine the borderline candidate boundary point of suspicious lump;

Shown in left figure among Fig. 3, for obtaining an area-of-interest and a given gray threshold Thres in the step (2) ₁, this area-of-interest can be done binary conversion treatment.Use the boundary tracking method to obtain corresponding contour line 1 in the area-of-interest after binaryzation; Adopt threshold value Thres ₂, repeat said process, will obtain contour line 2; ... adopt threshold value Thres _i, repeat said process, will obtain contour line I; ... adopt threshold value Thres _n, repeat said process, will obtain contour line N.Select a plurality of threshold values for use, obtain a plurality of contour lines (contour line group), the dense degree of contour line is relevant with image gradient, and the intensive place of contour line image gradient is bigger, and the sparse place of contour line image gradient is less.

With the region of interest centers is end points, from zero angle, at interval by counterclockwise outwards drawing R bar ray successively, tries to achieve the intersection point of each bar ray and contour line group with equal angles.If the Euclidean distance between two adjacent intersection points on the same ray is less than D _Min, then claim these two intersection points to be communicated with.To existing greater than S _MinThe point set of individual connection is represented this connected set of points with their center as a candidate boundary point.On a ray, have a plurality of candidate boundary point, may there be candidate boundary point yet.Right figure is the candidate boundary point of labelling in original area-of-interest among Fig. 3.

(3.2) obtain many boundary candidate lines by the candidate boundary point on each the bar ray that obtains in the step (3.1), utilize dynamic programming method to select wherein optimum boundary candidate line, promptly determine suspicious lump boundary line;

Ideally,, then connect the candidate boundary point on each bar ray in turn, will form unique boundary candidate line, promptly suspicious lump boundary line if having and only have a candidate boundary point on each the bar ray in the step (3.1).But under the practical situation, article one, have a plurality of candidate boundary point on the ray, on every ray, choose a candidate boundary point at every turn, connect the candidate boundary point of choosing on each bar ray in turn and form a boundary candidate line, can form many boundary candidate lines with this, can change greatly the position of (be image gradient value greatly) by the gradation of image value and have the characteristics of certain slickness according to the actual boundary line, the inventive method is utilized dynamic programming method, and set the cost that cost function is determined each bar boundary candidate line, in all boundary candidate lines, select the boundary candidate line of an optimum cost as final suspicious lump boundary line.Be implemented as follows:

If boundary candidate line S:{n ₁, n ₂... n _R, variable n _iBe illustrated in the candidate boundary point of choosing on the i bar ray.If i bar ray do not have candidate boundary point, then carry out interpolation and obtain according to the candidate boundary point on i-1 bar and the i+1 bar ray and the distance between region of interest centers.The cost function C of boundary candidate line S is the local cost sum of all candidate boundary point correspondences on the boundary candidate line, that is:

$C=\underset{i=1}{\overset{R}{\mathrm{\Σ}}}C\left(n_i\right)---\left(3\right)$

Local cost C (n _i) by inner cost C _Int(n _i) and outside cost C _Ext(n _i) form:

C(n _i)＝αC _int(n _i)+C _ext(n _i) (4)

Wherein α is a constant, is used to regulate the smooth degree of boundary line.

Inner cost C _Int(n _i) be defined as candidate boundary point n _iAnd n _I-1Between the standardization distance:

$C_{\mathrm{int}}\left(n_i\right)=\frac{2\mathrm{dist}(n_i,n_{i-1})}{\mathrm{dist}(O,n_i)+\mathrm{dist}(O,n_{i-1})}---\left(5\right)$

Dist (n wherein _i, n _I-1), dist (O, n _i), dist (O, n _I-1) represent candidate boundary point n respectively _iAnd n _I-1, center O interested and n _i, center O interested and n _I-1Distance, standardization distance is more little, promptly the boundary candidate line is smooth more, cost is also more little.

Outside cost C _Ext(n _i) be defined as candidate boundary point n _iThe connectivity points number at place, the connectivity points of a candidate point representative is many more, shows that gradient herein is big more, might be the real border point more.

C _Ext(n _i(the n of)=- _iThe connectivity points number at place) (6)

Obtain the cost of each bar boundary candidate line according to above-mentioned cost function, adopt dynamic programming to determine that in all boundary candidate lines the boundary candidate line of an optimum is as final suspicious lump boundary line.

(4) the suspicious lump of having cut apart is calculated its serial associated eigenvalue, the feature of selecting for use generally can be divided into several classes such as geometric properties, morphological feature, gray feature and textural characteristics.

The eigenvalue of selecting should be followed following characteristics:

1. recognizability: the eigenvalue of inhomogeneity object has notable difference;

2. reliability: the eigenvalue of homogeneous object applications similar;

3. independence: strong correlation does not have between the eigenvalue;

According to above rule, extract 25 features in suspicious lump zone altogether, specifically describe and see that Fig. 4 tabulates.

(5) with in the eigenvalue input grader that calculates, computer automatic analysis is carried out in initial suspicious lump zone, determine final suspicious lump zone.

Suspicious lump classification is the last stage that mammary gland tumor detects automatically.The suspicious lump that is partitioned into the border is extracted after the eigenvalue of reflection lump characteristic, the utilization grader is judged this suspicious lump positive (the true lump zone that computer is thought) or false positive.The selection of grader and design have determined the accuracy that lump detects to a great extent.Classification is the important component part of pattern recognition theory, can adopt methods such as linear classification, heuristic rule, statistical classification, fuzzy classification, artificial neural network that feature is classified usually.The inventive method adopts a kind of improved k nearest neighbour method to classify to extracting feature in the step (4).

The fundamental rule of k nearest neighbour method is: find out k the sample to test sample book characteristic vector nearest (or the most similar) in whole samples (removing test sample book), voted by these samples, test sample book is included in the maximum classification of sample votes.The k nearest neighbour method grader that the inventive method adopts at first defines characteristic vector similarity function and decision function (decision index is called for short DI).

1. the definition of similarity function

Test sample book Y _QCharacteristic vector be designated as V (Y _Q), sample X removes test sample book) characteristic vector be designated as V (X), then define similarity function and be the inverse of Euclidean distance between two characteristic vectors square, promptly

$\mathrm{Sim}(Y_Q,X)=\frac{1}{{\left|\right|V\left(Y_Q\right)-V\left(X\right)\left|\right|}^2}---\left(7\right)$

2. the definition of decision function

To test sample book Y _QWhen making a strategic decision, characteristic vector with it should be big more apart from the influence of near more sample participative decision making in principle, and in primary k nearest neighbour method grader, simply ballot method is difficult to embody the influence of this vector distance difference, and the inventive method defines according to distance weighted decision function.Following is that lump class and normal class two classes are example with the decision-making test sample book.

Preceding k sample with the descending arrangement of similarity calculates test sample book Y as shown in Equation (8) _QDecision value.With test sample book Y _QIn preceding k the nearest sample, the lump class is designated as Mass, and its number of samples is M, and normal class is designated as Norm, and its number of samples is N, Sim (Y _Q, X) be the 1. middle test sample book Y that defines _QWith the similarity of sample X, the order of Rnk (X) representative sample X in similarity is arranged,

Represent j lump class sample, the span of j is [1, M],

Represent l normal class sample, the span of l is [1, N].

$\mathrm{DI}\left(Y_Q\right)=\frac{{\mathrm{\Σ}}_{j=1}^M\{\mathrm{Sim}(Y_Q,X_j^{\mathrm{Mass}})\×(K+1-\mathrm{Rnk}\left(X_j^{\mathrm{Mass}}\right))\}}{{\mathrm{\Σ}}_{j=1}^M\{\mathrm{Sim}(Y_Q,X_j^{\mathrm{Mass}})\×(K+1-\mathrm{Rnk}\left(X_j^{\mathrm{Mass}}\right))\}+{\mathrm{\Σ}}_{l=1}^N\{\mathrm{Sim}(Y_Q,X_l^{\mathrm{Norm}})\×(K+1-\mathrm{Rnk}\left(X_l^{\mathrm{Norm}}\right)\}}---\left(8\right)$

This decision value computational methods had both been considered the classification of closing on sample with test sample book as the simple vote method, considered the order of they and test sample book similarity again, experimental results show that this decision function is better than the computational methods of original k neighbour decision function.

For the utilization of grader, at first to train it with training data, obtain being applicable to the classifier parameters of particular problem, the grader training process comprises to be collected the grader training data and obtains two steps of classifier parameters, as shown in Figure 5:

(5.1) collect the grader training data;

At first import one group of known diagnosis result's picture of mammary gland x-ray radiography, use above-mentioned area-of-interest extraction, suspicious lump Region Segmentation, processing with step in the feature extraction in suspicious lump zone, obtain segmentation result and a series of associated eigenvalue of suspicious lump, so far finish the collection of grader training data.

(5.2) obtain classifier parameters;

Train the grader that has designed with the associated eigenvalue of the suspicious lump that calculates and the actual diagnostic result of this suspicious lump, obtain sorting parameter, write the classifier parameters file, so far finish the training process of grader.

(6) segmentation result in final suspicious lump zone is positioned on the picture of mammary gland x-ray radiography to be diagnosed of input in the step (1), and, is shown to the user as required the eigenvalue in the lump zone that calculates in the step (4).

As shown in Figure 2, assistant diagnosis system of the present invention comprises that input module 100, area-of-interest extraction module 200, suspicious lump cut apart module 300, suspicious lump provincial characteristics extraction module 400, classification diagnosis module 500 and output module 600.

Input module 100 is used to receive the picture of mammary gland x-ray radiography to be diagnosed of user's input, and sends area-of-interest extraction module 200 to.

The area-of-interest that area-of-interest extraction module 200 extracts in the picture of mammary gland x-ray radiography of importing obtains initial suspicious lump zone position information according to the described step of above-mentioned steps (2), sends suspicious lump to and cuts apart module 300.

Suspicious lump is cut apart module 300 according to the described process of above-mentioned steps (3), suspicious lump in the area-of-interest of area-of-interest extraction module 200 extractions is split, obtain the boundary information of suspicious lump, send suspicious lump provincial characteristics extraction module 400 to.

Suspicious lump provincial characteristics extraction module 400 is according to the boundary information of the suspicious lump that receives, calculate a series of regional associated eigenvalue, as several classes such as geometric properties, morphological feature, gray feature and textural characteristics, and send classification diagnosis module 500 to.

In the eigenvalue input grader in each suspicious lump zone that classification diagnosis module 500 will calculate, computer automatic sorting identification is carried out in initial suspicious lump zone, determine final suspicious lump zone, and send the result to output module 600, be positioned on the picture of mammary gland x-ray radiography to be diagnosed of input by the segmentation result of output module 600, and the regional associated eigenvalue that will calculate is shown to the user as required the automatic final lump zone of detecting of computer.

Example:

Relate to several parameters in the breast carcinoma computer-aided diagnosis system that the present invention proposes based on picture of mammary gland x-ray radiography, these parameters will be carried out the comprehensive adjustment setting at the data characteristics of concrete processing to reach the superperformance of total system, list the parameter of setting at deal with data set of the present invention herein:

The original template size relevant parameter L=25 that step (2.1) utilizes two-dimentional hyperbolic secant (sech) function to obtain;

Step (2.2) is carried out the threshold value T that binary conversion treatment is chosen to the degree of association image _Low=0.5 * input picture of mammary gland x-ray radiography maximum gray scale;

The threshold value T that chooses in step (2.3) multiscale analysis _High=0.6 * input picture of mammary gland x-ray radiography maximum gray scale;

Step (3.1) is an end points with the region of interest centers, from zero angle, at interval by counterclockwise outwards drawing R=64 bar ray successively, tries to achieve the intersection point of each bar ray and contour line group with equal angles.If the Euclidean distance between two adjacent intersection points on the same ray is less than D _Min=3, then claim these two intersection points to be communicated with.To existing greater than S _MinThe point set of=10 connections;

The local cost C (n of step (3.2) _i) by inner cost C _Int(n _i) and outside cost C _Ext(n _i) form:

C(n _i)＝αC _int(n _i)+C _ext(n _i)

Wherein be used to regulate constant parameter α=110 of the smooth degree of boundary line.

The inventive method is automatically analyzed and is handled suspicious mammary gland tumor zone in the picture of mammary gland x-ray radiography by a breast carcinoma computer-aided diagnosis system based on picture of mammary gland x-ray radiography, lump position and lump shape are provided, and provide the zone relevant series of features parameter as required, thereby accuracy and the efficient of radiologist to breast cancer diagnosis has been improved in zone that the prompting radiologist need pay close attention to and the relevant emphasis parameter in zone to a certain extent.Realization of the present invention is not limited to the disclosed scope of above-mentioned example, can adopt the mode that is different from above-mentioned example to realize technique scheme.

Claims (2)

1, a kind of breast carcinoma computer-aided diagnosis system based on picture of mammary gland x-ray radiography comprises that input module (100), area-of-interest extraction module (200), suspicious lump cut apart module (300), suspicious lump provincial characteristics extraction module (400), classification diagnosis module (500) and output module (600);

Input module (100) is used to receive the picture of mammary gland x-ray radiography to be diagnosed of user's input, and sends area-of-interest extraction module (200) to;

Area-of-interest extraction module (200) is used for extracting the area-of-interest of the picture of mammary gland x-ray radiography of input, obtains initial suspicious lump zone position information, and sends suspicious lump to and cut apart module (300);

Suspicious lump is cut apart the suspicious lump that module (300) is used for area-of-interest that area-of-interest extraction module (200) is extracted and is split, and obtains the boundary information of suspicious lump, and sends suspicious lump provincial characteristics extraction module (400) to;

Suspicious lump provincial characteristics extraction module (400) calculates associated eigenvalue, and sends classification diagnosis module (500) to according to the boundary information of the suspicious lump that receives;

Classification diagnosis module (500) is used for the eigenvalue input grader in each the suspicious lump zone that will calculate, computer automatic sorting identification is carried out in initial suspicious lump zone, determine final suspicious lump zone, and send the result to output module (600), the segmentation result in the final lump zone that will be detected by output module (600) is positioned on the picture of mammary gland x-ray radiography to be diagnosed of input, and the regional associated eigenvalue that will calculate is shown to the user as required;

It is characterized in that:

Area-of-interest extraction module (200) is handled according to following step (2.1) to the process of (2.5):

Step (2.1) utilizes two-dimentional hyperbolic secant function to obtain template, calculates the degree of association of above-mentioned input picture of mammary gland x-ray radiography and this template, obtains the degree of association image;

The degree of association numerical value of the some pixels of step (2.2) degree of association image is 0 with this assignment then less than selected threshold, is 1 otherwise compose; In binaryzation degree of association image, extract all connected regions, and have the center of the point of maximal correlation degree in the corresponding region in the degree of association image that obtains of extraction step (2.1) as this connected region;

Step (2.3) utilizes two-dimentional hyperbolic secant function to obtain and the middle different template of template size of step (2.1); Each connected region for step (2.2) extraction, each yardstick template center is moved on to the position of center on original input picture of mammary gland x-ray radiography of connected region, calculate the degree of association of corresponding subimage in each yardstick template and the original input picture of mammary gland x-ray radiography image respectively, with the final degree of association of the maximum in all degree of association that calculate, choose a threshold value final degree of association is got rid of as the false positive zone less than the connected region of threshold value as this connected region;

Step (2.4) is got rid of the connected region of 5-30 pixel size and the connected region of strip again for the connected region that is not excluded in the step (2.3);

Step (2.5) to the connected region that remains, is center with its geometric center through the zone screening of (2.4), intercepting square in original input galactophore image, and this square area of extraction is area-of-interest or is called initial suspicious lump zone;

Suspicious lump is cut apart module (300) and is cut apart suspicious lump according to following step in area-of-interest:

Step (3.1) is obtained the relevant contour line group of gradient characteristic of region of interest area image, be end points afterwards with the region of interest centers, from zero angle, outwards draw some rays successively by counter clockwise direction at interval with equal angles, try to achieve the intersection point of each bar ray and contour line group, obtain boundary candidates point on every ray with this;

Step (3.2) obtains many boundary candidate lines by the candidate boundary point on each the bar ray that obtains in (3.1), selects wherein optimum boundary candidate line as suspicious lump boundary line again.

2, breast carcinoma computer-aided diagnosis system according to claim 1 is characterized in that: the grader in the classification diagnosis module (500) adopts following method that characteristic vector similarity function and decision function in the original k nearest neighbour method are defined again:

Test sample book Y _QCharacteristic vector be designated as V (Y _Q), the characteristic vector of the sample X except that test sample book is designated as V (X), then defines similarity function Sim (Y _Q, X) be the inverse of Euclidean distance square between two characteristic vectors, promptly

$\mathrm{Sim}(Y_Q,X)=\frac{1}{{\left|\right|V\left(Y_Q\right)-V\left(X\right)\left|\right|}^2}$

With similarity function Sim (Y _Q, preceding k sample of the descending arrangement of value X) adopts following formula to calculate test sample book Y _QDecision function DI (Y _Q) value; With test sample book Y _QIn preceding k the nearest sample, the lump class is designated as Mass, and its number of samples is M, and normal class is designated as Norm, and its number of samples is N, the order of Rnk (X) representative sample X in similarity is arranged, Represent j lump class sample, the span of j is [1, M], Represent l normal class sample, the span of l is [1, N], decision function DI (Y _Q) be defined as:

$\mathrm{DI}\left(Y_Q\right)=\frac{{\mathrm{\Σ}}_{j=1}^M\{\mathrm{Sim}(Y_Q,X_j^{\mathrm{Mass}})\×(K+1-\mathrm{Rnk}\left(X_j^{\mathrm{Mass}}\right))\}}{{\mathrm{\Σ}}_{j=1}^M\{\mathrm{Sim}(Y_Q,X_j^{\mathrm{Mass}})\×(K+1-\mathrm{Rnk}\left(X_j^{\mathrm{Mass}}\right)\}+{\mathrm{\Σ}}_{l=1}^N\{\mathrm{Sim}(Y_Q,X_l^{\mathrm{Norm}})\×(K+1-\mathrm{Rnk}\left(X_l^{\mathrm{Norm}}\right)\}}.$

Images