CN101968851A - Medical image processing method based on dictionary studying upsampling - Google Patents

Abstract

The invention discloses a medical image processing method based on dictionary studying upsampling, and belongs to the technical field of image processing. The medical image processing method is realized by the following steps of: inputting an original medical image; cutting and reinforcing the original medical image; extracting the characteristics of the cut and reinforced medical image to obtain a training sample set and a testing sample set according the extracted characteristics; finding out boundary points of weakness samples from the training sample set to obtain the number of new samples to be generated according to the situation of the neighborhoods of the boundary points; generating required new samples by utilizing a sparse ligature and point getting method; adding the new samples into the training sample set to form a new training sample set; carrying out classifying diagnosis to the new training sample set to get a classifier; and diagnosing the testing sample set by adopting the classifier to get a final diagnosis result. The medical image processing method has the advantages of high recognition rate and strong generalization capability on medical image diagnosis, and can be used for medical workers to evaluate disease prognosis and therapeutic effect.

Description

The medical image processing method of sampling in study based on dictionary

Technical field

The invention belongs to technical field of image processing, particularly relate to medical image processing, can be used for monitoring of diseases and distribute, study pathogenesis and disease auxiliary diagnosis.

Background technology

The fast development of computer science and technology has produced tremendous influence to medical domain, and people attempt to allow computing machine replace the human automatic diagnosis disease etc. of realizing to have challenging work gradually, and medical image plays an important role in clinical diagnosis.Since roentgen in 1895 finds X ray, particularly occur the computer tomography CT technology in 1979, greatly promoted the development of medical imaging.Since nearly 30 years, new iconography technology emerges in an endless stream especially.

Yet because problems such as required value height in source and individual privacies, making has the medical image of pathology more less than normal picture, has caused the data imbalance problem, has finally caused the medical image recognition difficulty.How solving this difficulty effectively is problem demanding prompt solution in the field of medical images.At present, method in common is exactly to utilize the method for up-sampling, increases the data number that the pathology medical image is arranged, and changes to distribute, to reduce the uneven degree of data.

The most original top sampling method is the sample that duplicates rare class, but has done like this identical data repetitive learning, has expended the time, and to improving the not too big help of rare class discrimination.

Higher top sampling method then adopts some heuristic skills, duplicates the weak tendency sample selectively, perhaps generates new weak tendency sample.The SMOTE algorithm that people such as Chawla propose is a kind of simple and effective top sampling method, this method is at first selected several adjacent sample at random for each weak tendency sample, and get at random a little on the line of this sample and these contiguous samples, generating does not have the new weak tendency sample that repeats.2008, the ADASYN method that people such as Haibo He propose was that borderline weak tendency sample is generated new weak tendency sample at random according to the method that self weighted value utilizes the line of SMOTE to get a little.Yet the study that these methods are too much the weak tendency sample, cause the study of crossing easily to the weak tendency sample, make a little less than the generalization ability own, cause the overall discrimination of test sample book low.

Summary of the invention

The objective of the invention is to overcome the deficiency of above-mentioned prior art, at problem such as a little less than medical image data imbalance, the low and generalization ability of diagnosis discrimination, a kind of medical image processing method of sampling in study based on dictionary has been proposed, to improve the medical imaging diagnosis discrimination and to strengthen generalization ability.

For achieving the above object, the invention provides the medical image processing method of sampling in study, comprise the steps: based on dictionary

(1) input primitive medicine image;

(2) adopt histogram equalization and mean square deviation standardized method, the primitive medicine image is cut and enhancement process;

(3) medical image after cutting and the enhancement process is extracted its gray level co-occurrence matrixes or Hu moment characteristics; Definition has the training sample set T of sign ₁The test sample book of sign is not gathered T ₂:

T ₁＝{(x ₁，u ₁)..(x _i，u _i).，(x _l，u _l)}

T ₂＝{v ₁，..v _i，..v _m}

Wherein, x _iRepresent i the feature that the sign training sample is arranged, u _iRepresent i the sign that the sign training sample is arranged, V _iRepresent i the not feature of the test sample book of sign, l is the number that the sign training sample is arranged; M is the number of test sample book of sign not;

(4) to the training sample T of sign is arranged ₁Carry out the classification diagnosis of sampling in study, obtain sorter C based on dictionary:

4a) from the training sample T of sign is arranged ₁In select the minimum class of number of samples as weak tendency sample T ₃

4b) find out weak tendency sample T ₃In frontier point set B={ b ₁..b _i..b _s, calculate each frontier point b _iThe number n that needs the new samples of generation _i, b wherein _iRepresent i frontier point, s is the quantity of frontier point;

4c) weak tendency sample T to selecting ₃Adopt the method training of KSVD to generate a dictionary D;

4d) with each frontier point b _i, obtain n by sparse line point sampling method _iIndividual new samples is combined into new samples set T with these new samples ₄

4e) new samples is gathered T ₄Add weak tendency sample T ₃The current new weak tendency sample T of middle composition ₃

4f) with new weak tendency sample T ₃' with training set in other samples are common forms the current new training sample T that sign is arranged ₁';

4g) with support vector machine method to the new training sample T that sign is arranged ₁' carry out classification diagnosis, the sorter C after obtaining diagnosing;

(5) adopt the sorter C that obtains after the diagnosis, to not identifying test sample book T ₂Diagnose, obtain not identifying test sample book T ₂The last diagnostic result.

The present invention has the following advantages compared with prior art:

1, the present invention adopt frontier point in the ADASYN algorithm choose with the SMOTE algorithm in line get a strategy, in conjunction with the strong point of the two, improved the discrimination of weak tendency sample;

2, the present invention utilizes the rarefaction representation method to produce the newly-increased sample of virtual point conduct of boundary sample, has overcome ADASYN algorithm and the SMOTE algorithm mistake problem concerning study to the weak tendency sample, has improved overall discrimination;

3, the present invention adopts all weak tendency samples to carry out dictionary study, has taken into full account the overall performance of weak tendency sample, has strengthened the generalization ability to medical image processing.

Description of drawings

Fig. 1 is a process flow diagram of the present invention;

Fig. 2 is the original mammary gland X striograph that emulation of the present invention is adopted;

To be the present invention remove mammary gland X striograph after the redundancy to Fig. 2 to Fig. 3;

To be the present invention adopt mammary gland X striograph after histogram equalization and the mean square deviation standardized method to Fig. 3 to Fig. 4.

Embodiment

With reference to Fig. 1, the present invention is based on the medical image processing method that dictionary is sampled in study, comprise the steps:

Step 1: adopt histogram equalization and mean square deviation standardized method, the medical image in the primitive medicine image set is cut and enhancement process, obtain visual effect medical image collection preferably.

1a) input primitive medicine image, its size is M * N, the width of cloth image in this example selection original mammary gland X image set as shown in Figure 2, its size 1024 * 1024;

1b) to the primitive medicine image of input, adopt the automatic cutting method of computing machine, excise the artificial marking that exists in the background of its image and the image, the mammary gland X image after obtaining cutting, as shown in Figure 3;

1c) adopt histogram equalization and mean square deviation standardized method to remove noise to the medical image after the cutting, the mammary gland X image that obtains having better visual effect, as shown in Figure 4.

Step 2: the mammary gland X image after cutting and the enhancement process is carried out the Hu Moment Feature Extraction.

2a) in the visual effect that obtains preferably on the mammary gland X image, calculation level (x, (p+q) rank square m that y) locates _PqReach (p+q) rank central moment μ _Pq:

$m_{\mathrm{pq}}=\underset{x=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{N-1}{\mathrm{\Σ}}}{x}^p{y}^{q}f(x,y)$

${\mathrm{\μ}}_{\mathrm{pq}}=\underset{x=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{N-1}{\mathrm{\Σ}}}{(x-x_c)}^p{(y-y_c)}^{q}f(x,y)$

In the formula, x represents that (y represents point (x, abscissa value y) to point for x, abscissa value y);

M represents that the row of mammary gland X image is high, and N represents the row height of mammary gland X image;

P ∈ 0,1,2,3}, q ∈ 0,1,2,3}, x ^pThe p power of expression x, y ^qThe q power of expression y;

F (x, y) expression point (x, the pixel value of y) locating;

(x _c, y _c) the expression visual effect barycentric coordinates of mammary gland X image preferably, x _cBe abscissa value, y _cBe ordinate value;

(x-x _c) ^pBe (x-x _c) the p power, (y-y _c) ^qBe (y-y _c) the q power;

2b) according to the point (x, (p+q) rank square m that y) locates that obtain _PqReach (p+q) rank central moment μ _Pq, obtain point (x, the computing formula of normalization central moment y):

${\mathrm{\η}}_{\mathrm{pq}}={\mathrm{\μ}}_{\mathrm{pq}}/m_{\mathrm{pq}}^{\mathrm{\γ}}$

In the formula,

Expression m _PqThe γ power, γ=(p+q)/2+1;

2c) with 2a) in the value substitute point of p and q (x in the computing formula of normalization central moment y), obtains point (x, normalization central moment η y) ₀₂, η ₀₃, η ₁₁, η ₁₂, η ₂₀, η ₂₁, η ₃₀

(x, the normalization central moment of y) locating are extracted visual effect seven Hu moment characteristics of mammary gland X image preferably, are defined as φ respectively 2d) to utilize point ₁, φ ₂..., φ ₇, that is:

φ ₁＝η ₂₀+η ₀₂

${\mathrm{\φ}}_2={({\mathrm{\η}}_{20}-{\mathrm{\η}}_{02})}^2+{4\mathrm{\η}}_{11}^2$

φ ₃＝(η ₃₀-3η ₁₂)2 ⁺(3η ₂₁-η ₀₃) ²

φ ₄＝(η ₃₀+η ₁₂) ²+(η ₂₁+η ₀₃) ²

φ ₅＝(η ₃₀-3η ₁₂)(η ₃₀+η ₁₂)φ _x+(η ₀₃-3η ₂₁)(η ₂₁+η ₀₃)φ _y

φ ₆＝(η ₂₀-η ₀₂)[(η ₃₀+η ₁₂) ²-(η ₀₃+η ₂₁) ²]+4η ₁₁(η ₃₀+η ₁₂)(η ₀₃+η ₂₁)

φ ₇＝(3η ₂₁-η ₀₃)(η ₃₀+η ₁₂)φ _x+(η ₃₀-3η ₁₂)(η ₀₃+η ₂₁)φ _y，

Wherein, $\begin{array}{c}{\mathrm{\φ}}_x={({\mathrm{\η}}_{30}+{\mathrm{\η}}_{12})}^2-3{({\mathrm{\η}}_{03}-3{\mathrm{\η}}_{21})}^2\\ {\mathrm{\φ}}_y={({\mathrm{\η}}_{03}+{\mathrm{\η}}_{21})}^2-3{({\mathrm{\η}}_{30}-{\mathrm{\η}}_{12})}^2\end{array}.$

Step 3:, obtain the training sample set T of sign according to feature to visual effect mammary gland X image extraction preferably ₁Biao Shi test sample book collection T not ₂:

T ₁＝{(x _i，u _i)|x _i∈R ⁿ，u _i∈{-1，1}，i＝1，...，l}

T ₂＝{v ₁，..v _i，..v _m}

In the formula, x _iRepresent i the feature that the sign training sample is arranged, u _iRepresent i the sign that the sign training sample is arranged, v _iRepresent i the not feature of the test sample book of sign, l is the number that the sign training sample is arranged, and m be the number of test sample book of sign not, and the diagnostic result of the mammary gland X image that the present invention adopts comprises normally and cancer patient's two classes, be labeled as-1 normally, the cancer patient is labeled as 1.

Step 4: from the training sample T of sign is arranged ₁In select the less relatively class of number of samples as weak tendency sample set T ₁₁, an other class is as surging sample set T ₁₂:

T ₁₁＝{(x _i，u _i)|x _i∈R ⁿ，u _i∈{1，2}，i＝1，…l ₁}

t ₁₂＝{(x _i，u _i)|x _i∈R ⁿ，u _i∈{1，2}，i＝1，…l ₂}

In the formula, l ₁Be weak tendency sample set t ₁₁Number of samples, l ₂Be surging sample set T ₁₂Number of samples, R ⁿReal number set for the n dimension.

Step 5: find out weak tendency sample T ₁₁In the frontier point set B, calculate the number n that each frontier point B (i) needs the new samples that generates _i, wherein i is an integer, and 1≤i≤s, s is the quantity of frontier point, B and n _iCalculate as follows:

5.1) calculating weak tendency sample set T ₁₁In the weight set w of all samples ₁₁:

w ₁₁＝{w _i|i＝1，…l ₁}

In the formula, w _i=num _i/ K, expression sample (x _i, u _i) corresponding weight, wherein, K is sample x _iThe neighborhood number, K=5; Num _iRepresentative sample x _iThe number of surging sample in the neighborhood;

5.2) according to weight set w ₁₁From weak tendency sample set T ₁₁In select frontier point set B and its weight set w _B:

B＝{(x _i，u _i)|0＜w _i＜1，w _i∈w ₁₁}

w _B＝{w _i|0＜w _i＜1，w _i∈w ₁₁}

Wherein, (x _i, u _i) for the training sample set T of sign is arranged ₁In i sample, w _iBe weight set w ₁₁In i weight;

5.3) weight is gathered w _BNormalization obtains w _B':

${w_B}^{\′}=\{w_i/\underset{i=1}{\overset{l_3}{\mathrm{\Σ}}}{w}_i|i=1,...l_3,w_i\∈w_B\}$

In the formula, l ₃Be the number of frontier point set B, w _BBe the set of the weight after the normalization of frontier point set B;

5.4) calculate the number n that each frontier point B (i) needs the new samples that generates _i:

n _i＝w _B(i)*(l ₂-l ₁)

Wherein, w _B' (i) be w _B' in i element, l ₁Be weak tendency sample set T ₁₁Number of samples, l ₂Be surging sample set T ₁₂Number of samples.

Step 6: to the weak tendency sample T that selects ₃Adopt the method training of KSVD to generate a dictionary D.

Step 7:, obtain n by sparse line point sampling method with each frontier point B (i) _iIndividual new samples, the generation step of each new samples is as follows:

7.1) the dictionary D that utilizes step 6 to obtain obtains sparse factor alpha to B (i) rarefaction representation ₁

7.2) a weak tendency sample among picked at random B (i) neighbour, the dictionary D that utilizes step 6 to obtain obtains sparse factor alpha to this sample rarefaction representation ₂

7.3) pass through α ₁And α ₂Calculate the sparse factor alpha that will generate new samples:

α＝α ₁+(α ₂-α ₁)*rand

Wherein, rand is a random number of 0 to 1;

7.4) dictionary D and sparse factor alpha are multiplied each other a new samples T who obtains generating _Nwe

Step 8: all new samples that step 7 is generated add weak tendency sample T ₁₁The current new weak tendency sample T of middle composition ₁₁'.

Step 9: with new weak tendency sample T ₁₁' with surging sample set T ₁₂The current new training sample T that sign is arranged of common composition ₁

Step 10: with support vector machine method to training sample T ₁' carry out classification diagnosis, the sorter after obtaining diagnosing.

Step 11: adopt the sorter that obtains after the diagnosis, to not identifying test sample book T ₂Diagnose, obtain class label h ₁, h ₂..., h _l(h _i{ 1,1}), if class label is 1, diagnostic result is canceration to ∈; If class label is-1, diagnostic result is normal.

Step 12: output does not identify test sample book T ₂The last diagnostic result.

Effect of the present invention can further specify mammary gland X image simulation data by following:

1. experiment condition and content

The experiment simulation environment is: windows XP, SPI, CPU Pentium (R) 4, basic frequency 2.4GHZ, software platform are the Matlab7.0.4 operation.The original mammary gland X image that emulation is selected for use derives from common data sets MIAS, obtains 322 original mammary gland X images altogether.

Experiment content comprises: adopt top sampling method COPY, SMOTE, ADASYN and four kinds of methods of the present invention of reproduction copies that original mammary gland X image is tried experiment respectively, detect the image of suffering from cancer.In the checking of experimental result, we have listed four kinds of evaluation indexes and have proved superiority of the present invention, are respectively the total discrimination T_Ratio of sample, group sample discrimination Recall, total evaluation index F_measure and G_mean.

In order to define these evaluation indexes, need use basic evaluation index---the confusion matrix of machine learning method, as shown in table 1.Wherein, represent the weak tendency sample with positive class, negative class is represented surging sample.

Table 1 liang class confusion matrix

	Predict positive class	Predict anti-class
			Actual positive class	TP(true?positive)	FN(false?negative)；
Actual anti-class	FP(false?positive)；	TN(true?negative)

In one two class confusion matrix, reality is positive class, and prediction also is called correct positive class TP for the sample size of positive class; Reality is positive class, and what be predicted as anti-class is called the anti-class FN of mistake; Reality is anti-class, and what be predicted as positive class is called the positive class FP of mistake; Reality is anti-class, is predicted as correct anti-class TN of being called of anti-class.

Utilize four kinds of evaluation indexes of confusion matrix definition as shown in table 2:

The definition of four kinds of evaluation indexes of table 2

2, simulation result

In l-G simulation test, what the core mapping method of support vector machine the inside was selected for use is gaussian kernel, and what nuclear parameter was got is 0.005.What dictionary optimization of the present invention was used is the method for KSVD.The atom number of KSVD gets 10, and iterations also is 10, and the rarefaction representation error is 0.01.Experiment guarantees that the whole bag of tricks moves under same data.

Table 3 is the diagnostic results of four kinds of methods to 322 width of cloth integral image.Wherein, the image that canceration is arranged is 115 pairs, and the normal sample that does not have canceration is 207 pairs, and degree of unbalancedness is 1: 1.8.Experimental result is the average results of 20 experiments, and each experiment all is to take 70% sample as markd training sample at random, and remaining is 30% as unmarked test sample book.

Four kinds of methods of table 3 are to the diagnostic result of 322 width of cloth integral image

	Ratio	Recall	F_measure	G_mean
					COPY	0.48434	0.48727	0.46739	0.52909
SMOTE	0.49219	0.59545	0.52316	0.55133
					ADASYN	0.47656	0.51000	0.47478	0.52506
D_SMOTE	0.52344	0.66636	0.56835	0.59570

As can be seen from Table 3, the present invention is optimum in all evaluation indexes, and this proves absolutely that the present invention is in the superiority of handling on the medical image imbalance problem.

Be cut into 4 parts with the medical image of handling well is parallel, will obtain 1288 sub-pictures.According to the position of canceration, to this 1288 sub-picture mark classification, 1 representative has canceration, and-1 representative is normal.Remove 420 secondary normal pictures at random, finally obtain image 120 pairs of canceration, normal picture 748 pairs, totally 968 pairs, degree of unbalancedness is 1: 6.23.Four kinds of methods are as shown in table 4 to the experimental result of 968 width of cloth integral image.Experimental result is the average results of 20 experiments, and each experiment all is to take 50% sample as markd training sample at random, and remaining is 50% as unmarked test sample book.

Four kinds of methods of table 4 are to the experimental result of 968 width of cloth integral image

Ratio

Recall

F_measure

G_mean

COPY	0.82855	0.49764	0.49688	0.65083
					SMOTE	0.81235	0.59516	0.50001	0.70453
ADASYN	0.80210	0.60386	0.49782	0.70336
					D_SMOTE	0.90012	0.56518	0.62115	0.73446

As can be seen from Table 4, the present invention in degree of unbalancedness than under the condition with higher, relatively poor relatively to the discrimination of weak tendency sample.This is because the present invention prevents study and the training abundant to the weak tendency sample, when but degree of unbalancedness is big, have only and to improve up weak tendency sample discrimination, but influenced the total discrimination and the generalization ability of sorter like this enough big study of weak tendency sample and training.The present invention just has remarkable advantages on total discrimination and generalization ability.

Claims (3)

1. a medical image processing method of sampling in study based on dictionary comprises the steps:

(1) input primitive medicine image;

(2) adopt histogram equalization and mean square deviation standardized method, the primitive medicine image is cut and enhancement process;

(3) medical image after cutting and the enhancement process is extracted its gray level co-occurrence matrixes or Hu moment characteristics; Definition has the training sample set T of sign ₁The test sample book of sign is not gathered T ₂:

T ₁＝{(x ₁，u ₁)..(x _i，u _i).，(x _l，u _l)}

T ₂＝{v ₁，..v _i，..v _m}

Wherein, x _iRepresent i the feature that the sign training sample is arranged, u _iRepresent i the sign that the sign training sample is arranged, v _iRepresent the not feature of the test sample book of sign of i, l is the number that the sign training sample is arranged, and m be the number of the test sample book that do not identify;

(4) to the training sample T of sign is arranged ₁Carry out the classification diagnosis of sampling in study, obtain sorter C based on dictionary:

4a) from the training sample T of sign is arranged ₁In select the minimum class of number of samples as weak tendency sample T ₃

4b) find out weak tendency sample T ₃In frontier point set B={ b ₁..b _i..b _s, calculate each frontier point b _iThe number n that needs the new samples of generation _i, b wherein _iRepresent i frontier point, s is the quantity of frontier point;

4c) weak tendency sample T to selecting ₃Adopt the method training of KSVD to generate a dictionary D;

4d) with each frontier point b _i, obtain n by sparse line point sampling method _iIndividual new samples is combined into new samples set T with these new samples ₄

4e) new samples is gathered T ₄Add weak tendency sample T ₃The current new weak tendency sample T of middle composition ₃';

4f) with new weak tendency sample T ₃' with training set in other samples are common forms the current new training sample T that sign is arranged ₁';

4g) with support vector machine method to the new training sample T that sign is arranged ₁' carry out classification diagnosis, the sorter C after obtaining diagnosing;

(5) adopt the sorter C that obtains after the diagnosis, to not identifying test sample book T ₂Diagnose, obtain not identifying test sample book T ₂The last diagnostic result.

2. according to claims 1 described method, wherein step 4b) the described weak tendency sample T that finds out ₃In frontier point set B={ b ₁..b _i..b _s, calculate each frontier point b _iThe number n that needs the new samples of generation _i, calculate as follows:

(2a) calculate weak tendency sample set T ₃In the weight set w of all samples ₁₁:

w ₁₁＝{w _i|i＝1，…l ₁}

In the formula, l ₁Be weak tendency sample set T ₃Number, w _i=num _i/ K, expression sample (x _i, u _i) corresponding weight, wherein, K is sample (x _i, u _i) the neighborhood number, K=5; Num _iRepresentative sample (x _i, u _i) number of surging sample in the neighborhood;

(2b) according to weight set w ₁₁From weak tendency sample set T ₃In select frontier point set B and its weight set w _B:

$w_B=\{w_b^1,..w_b^i,..w_b^s\}$

B＝{b ₁，..b _i，..b _s}

Wherein, b _iRepresent i frontier point, Expression b _iWeight, s is the quantity of frontier point;

(3c) weight is gathered w _BNormalization obtains w _B':

${w_B}^{\′}=\{w_b^i/\underset{i=1}{\overset{s}{\mathrm{\Σ}}}{w}_b^i|i=1,...s\}$

Wherein, w _B' be the weight set after the normalization of frontier point set B;

(3d) calculate each frontier point b _iThe number that needs the new samples of generation:

n _i＝w _B(i)*(l ₂-l ₁)

Wherein, w _B' (i) be w _B' in i element, l ₁Be training sample set T ₁The number of middle weak tendency sample, l ₂Be training sample set T ₁In the number of other samples.

3. according to claims 1 described method, wherein step 4d) described with each frontier point b _i, obtain n by sparse line point sampling method _iIndividual new samples, the generative process of each new samples is calculated as follows:

(3a) utilizing step 4c) the dictionary D that obtains obtains sparse factor alpha to B (i) rarefaction representation ₁

(3b) a weak tendency sample among picked at random B (i) neighbour utilizes step 4c) the dictionary D that obtains obtains sparse factor alpha to this sample rarefaction representation ₂

(3c) pass through α ₁And α ₂Calculate the sparse factor alpha that will generate new samples:

α＝α ₁+(α ₂-α ₁)*rand

Wherein, rand is a random number of 0 to 1;

(3d) dictionary D and sparse factor alpha are multiplied each other a new samples T who obtains to generate _New

Images