CN101609451A - Sort out the relevant feedback of identification based on the medical image fuzzy region feature and estimate method - Google Patents

Abstract

The invention discloses a kind of relevant feedback of sorting out identification based on the medical image fuzzy region feature and estimate method, it comprises the steps: 1) all medical images of choosing in the medical image databases cut apart; The hard feature in each zone after 2) extraction is cut apart; 3) hard Feature Conversion is become fuzzy characteristics, deposit in the property data base; 4) choose a medical image to be compared and extract its fuzzy characteristics, obtain the fuzzy similarity of the medical image in medical image to be compared and the property data base, according to the size of fuzzy similarity, the medical image in the property data base is sorted, export M width of cloth image from high to low; 5) fuzzy characteristics of this a M output image is brought into calculates in estimating based on the feedback of fuzzy similarity, recomputate the fuzzy similarity of all medical images in medical image to be compared and the property data base, order is from high to low exported N width of cloth image again.This feedback is estimated method can effectively sort out required medical image.

Description

Sort out the relevant feedback of identification based on the medical image fuzzy region feature and estimate method

Technical field

The present invention relates to a kind of image processing method, especially relate to a kind of relevant feedback and estimate method based on medical image fuzzy region feature classification identification.

Background technology

Along with the increasingly extensive application of medical digital image documentation equipment in clinical, the technology of electronic health record and Picture Archiving And Communication System (PACS) aspect constantly develops, and all can produce great amount of images data (bigger hospital every day view data have more than the 10G) clinically every day.How organizing, manage and export medical image effectively is the current problem that presses for solution.In clinical, to not making a definite diagnosis that clinical image is diagnosed and browse in the research,, the integrality of reliability of clinical diagnosis and Data acquisition, will be improved greatly if can find out and this essentially identical diagnostic image of kitchen range picture material in spite of illness by export technique in teaching.Tradition can't address that need gradually based on the data base administration mode of text mode, and (Content-Based ImageRetrieval, CBIR) technology becomes the research focus in this field in recent years in content-based for this reason image output.

The basic thought of CBIR is that low-level image features such as the visual signature of picture material such as color, texture, shape are exported.Though the CBIR technology has obtained using preferably in the output of natural image and video request program, but because medical image is more than the natural image complexity, it is bigger that difficult treatment and image collection relate to individual privacy reason difficulty, and the CBIR technology has run into very big challenge in the output of medical image is used.Compare successful CBIR system in the world, X-rabat CBIR system (ASSERT) as U.S. Purdue university, the pathological image CBIR system (Pathfinder) of Stanford Univ USA medical center, the cell CBIR system (PathMaster) of pathology system of Yale, the backbone x mating plate CBIR system (NHANES II) of American National Instrument Library of Medicine exploitation etc.The related medical image of these systems is single mode, and handles than being easier to.There is the CBIR system (IRMA) of medical college of German Aachen polytechnical university in the multi-modality medical image CBIR system of more complicated, but the current research of IRMA mainly concentrates on the automatic classification aspect of image with respect to the aspect at mode, anatomical structure and visual angle, also seldom relates in the research of kitchen range image output facet in spite of illness.Focus generally all is positioned at the regional area of medical image, be effectively must cut apart this image kitchen range medical image output in spite of illness.Limited by existing cutting techniques, be difficult to also at present realize that perfection cuts apart, the uncertainty of lesion segmentation has considerable influence to the output result.Export technique based on fuzzy region feature can reduce this influence to a certain extent.

In recent years, related feedback method also was used to improve the image export technique.Based on this method, system at first carries out preliminary screening according to the original output of user to database.Then, the user marks relevant and incoherent image on return results.These label informations can be combined in the whole follow-up search procedure according to the mode of iteration output, and continue to optimize output, return satisfied result until the user.The relevant feedback technology overwhelming majority among the present CBIR is based on hard feature.Based on the relevant feedback algorithm of fuzzy region feature also seldom.Have the researcher to utilize the AdaBoost method to realize fuzzy region feature relevant feedback algorithm, but this method can only be to realizing the weight adjustment between the different characteristic, and can not realize that the optimization feature in shines upon.

Summary of the invention

The purpose of this invention is to provide based on the relevant feedback of medical image fuzzy characteristics classification identification and estimate method, this method can effectively sort out required medical image.

Above-mentioned purpose of the present invention realizes by following technical solution: a kind of relevant feedback based on medical image fuzzy region feature classification identification is estimated method, and it comprises the steps:

1) each medical image of choosing in the medical image databases is cut apart, and is divided into several regions;

The hard feature in each zone after 2) extraction is cut apart;

3) the utilization index membership function is with step 2) the hard Feature Conversion extracted becomes fuzzy characteristics, deposits in the property data base;

4) choose a medical image to be compared, adopt above-mentioned steps 1) to 3) identical treatment step extracts the fuzzy characteristics of medical image to be compared, calculate the fuzzy similarity of the fuzzy characteristics of the fuzzy characteristics of medical image to be compared and the medical image in the property data base, size according to fuzzy similarity, the medical image of exporting in the property data base system is sorted, according to the high M width of cloth image of order output similarity from high to low, this M width of cloth image is referred to as output image one time, wherein fuzzy similarity is referred to as associated picture one time according to preceding 50% image from high to low, and the image of back 50% is referred to as once irrelevant image;

5) fuzzy characteristics of this a M output image is brought into calculates in estimating based on the feedback of fuzzy similarity, recomputate the fuzzy similarity of all medical images in medical image to be compared and the property data base, export N width of cloth image again according to fuzzy similarity order from high to low, this N width of cloth image is referred to as the secondary output image.

The processing procedure that the feedback based on fuzzy similarity in the described step 5) is estimated is as follows:

5a) obtain outer weight term and interior weight term by the weighted product that maximizes the fuzzy similarity of inquiring about an image to be compared and an associated picture,

5b) weight term u outside obtaining _m ⁱWith interior weight term W _m ⁱAfter, recomputate the fuzzy similarity between all medical images in image to be compared and the property data base, the medical image in the property data base is sorted again N width of cloth image before the output according to the size of fuzzy similarity.

The present invention can do following improvement: this method also comprises step 6):

Continuous repeating step 5) processing procedure, the fuzzy characteristics of the output image that step 5) is obtained is calculated in being brought into and estimating based on the feedback of fuzzy similarity, recomputate the fuzzy similarity of all medical images in medical image to be compared and the property data base, at last output and the high preceding L width of cloth medical image of medical image similarity to be compared.

Among the present invention, step 6) is the loop iteration process of step 5), the processing procedure of its processing procedure and step 5) is the same, fuzzy similarity is referred to as the secondary associated picture according to preceding 50% image from high to low in the secondary output image that step 5) obtains, and afterwards 50% image is referred to as the irrelevant image of secondary; The fuzzy characteristics of all secondary output images that step 5) is obtained is calculated in being brought into and estimating based on the feedback of fuzzy similarity, recomputate the fuzzy similarity of all medical images in medical image to be compared and the property data base, export multiple image again according to fuzzy similarity order from high to low, the loop iteration of step 6) can carry out once or twice or repeatedly iteration, output and the high L width of cloth medical image of medical image similarity to be compared at last as required.

Above-mentioned steps 1 of the present invention) be to utilize the EM algorithm of the limited gauss hybrid models of parameter that medical image is cut apart the present invention can do following improvement: before step 1), adopt gaussian filtering that input picture is carried out pre-service, to reduce the influence of noise, specifically by linear transformation (I-I to Flame Image Process _Min) * 255/ (I _Max-I _Min) all gray values of pixel points are transformed in 0～255 the scope, wherein, I is the gradation of image value, I _MinBe the minimum value of gradation of image, I _MaxMaximal value for gradation of image.

Hard feature described in the present invention comprises average gray feature, invariant moments shape facility, Gabor textural characteristics, direction histogram feature.

M value described in the present invention is 80～100, and the N value is 60～80, and the L value is 10～30.

Compared with prior art, the present invention has following remarkable result:

(1) this law invention utilizes the EM algorithm of the limited gauss hybrid models of parameter that each medical image in the medical data base is cut apart.This dividing method limits the dynamic range of each partitioning parameters when utilizing the EM algorithm that the limited gauss hybrid models parameter of parameter is done dynamic estimation.This restriction makes that when optimizing iteration each Gaussian function in the gauss hybrid models keeps stable.Thereby obtain reasonable segmentation result.

(2) this law invention adopts the index membership function to extract the fuzzy characteristics of each cut zone.On the basis of expressed fuzzy characteristics, the distance terms during the fuzzy similarity of proposition is estimated is arranged in the branch subitem of Index for Calculation, calculates simple and is easy to and expand.

(3) this law invention adopts fuzzy characteristics that medical image is expressed, and can effectively reduce to cut apart uncertain influence to output.

(4) the present invention optimizes output by the amassing of fuzzy similarity between maximization associated picture and the image to be compared, though exponential function is non-linear, but when multiplying each other owing to exponential function, its exponential term is a linear, additive, when fuzzy similarity adopted the mode that multiplies each other to make up, the array mode of exponential term was linear, and this makes us can adopt Lagrangian method commonly used to separate this optimization problem, and can obtain unique analytic solution, improve the speed of calculating.

Description of drawings

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

Embodiment

A kind of feedback based on medical image fuzzy region feature classification identification is estimated method, as shown in Figure 1, specifically comprises the steps:

Step 1 is read in medical image from PACS system of hospital, adopt gaussian filtering that input picture is carried out pre-service respectively, to reduce the influence of noise to Flame Image Process, then by linear transformation (I-I _Min) * 255/ (I _Max-I _Min) all gray values of pixel points are transformed in 0～255 the scope, wherein, I is the gradation of image value, I _MinBe the minimum value of gradation of image, I _MaxMaximal value for gradation of image;

Step 2 utilizes the EM algorithm of the limited gauss hybrid models of parameter that all medical images in the medical data base are cut apart, and is divided into several regions, and wherein the limited gauss hybrid models of parameter is as follows:

$f\left(x\right|\mathrm{\Θ})=\underset{i=1}{\overset{K}{\mathrm{\Σ}}}{\mathrm{\α}}_i{f}_i\left(x\right|{\mathrm{\θ}}_i),$ $f_i\left(x\right|{\mathrm{\θ}}_i)=\frac{1}{{\left(2\mathrm{\π}{\mathrm{\σ}}_i\right)}^{1/2}}\exp \{-\frac{1}{2}\frac{{(x-{\mathrm{\μ}}_i)}^2}{{\mathrm{\σ}}_i^2})\},{\mathrm{\σ}}_i\⇐{\mathrm{\σ}}_{i\max },$

μ wherein _iAnd σ _i ²Be respectively the average and the variance of i Gaussian function;

X represents the gradation of image feature of one dimension,

K is the number of Gaussian function,

α _iBe hybrid weight,

Based on the EM algorithm of the limited gauss hybrid models of parameter, utilize the EM algorithm that the limited gauss hybrid models parameter of parameter is done dynamic estimation, each parameter calculation formula is constant during iteration, upgrades σ at every turn _iAfter, if find σ _i＞σ _Imax, then force to make σ _i=σ _ImaxThis restriction makes that when optimizing iteration each Gaussian function in the gauss hybrid models keeps stable.

When specifically cutting apart, at first that the imagery exploitation parameter is limited EM algorithm is divided into two zones, distance in the average class spacing of difference computed image and the class, if distance is greater than average class spacing in certain regional class, just the specification area number is added one, again cut apart again, automatically determine several regional numbers of cutting apart by this method, cutting apart the limited EM algorithm of the parameter that is adopted is exactly the size that the parameter σ of each cut zone Gaussian function is represented in control when utilizing the EM algorithm that image is cut apart, if σ exceeds certain scope, just force to allow σ get back in this scope.This method makes the EM algorithm repeatedly keeping stable in the band process, can obtain reasonable segmentation result.

Step 3, average gray feature, invariant moments shape facility, Gabor textural characteristics, these hard features of direction histogram feature that after extraction is cut apart each is regional, the hard Feature Conversion that the utilization index membership function will extract becomes fuzzy characteristics, deposit in the property data base, this index membership function is as follows:

$E\left(x\right)=\exp (-\frac{{|x-\mathrm{\μ}|}^2}{2{\mathrm{\σ}}^2})$

Wherein, μ represents hard feature center, and σ and hard characteristic distribution mean square deviation are directly proportional;

Step 4, choose a medical image to be compared, adopt above-mentioned steps 1) to 3) identical treatment step extracts the fuzzy characteristics of medical image to be compared, and the fuzzy characteristics of medical image to be compared and the fuzzy characteristics of the medical image in the property data base are compared, calculate fuzzy similarity, size according to fuzzy similarity, the medical image of exporting out in the property data base system is sorted, according to 80～100 high width of cloth images of order output similarity from high to low, these images are referred to as output image one time, wherein fuzzy similarity is referred to as associated picture one time according to preceding 50% image from high to low, and the image of back 50% is referred to as once irrelevant image;

In this step, adopt following formula to calculate the fuzzy similarity of the fuzzy characteristics of the fuzzy characteristics of medical image to be compared and the medical image in the property data base:

Wherein, following fuzzy set A and the B that defines for above-mentioned index subordinate function:

$E_A\left(x\right)=\exp (-\frac{{|x-{\mathrm{\μ}}_A|}^2}{2{\mathrm{\σ}}_A^2}),$ $E_B\left(x\right)=\exp (-\frac{{|x-{\mathrm{\μ}}_B|}^2}{2{\mathrm{\σ}}_B^2})$

Fuzzy similarity between them is

$S(A,B)=\exp \{-\frac{{|{\mathrm{\μ}}_A-{\mathrm{\μ}}_B|}^2}{2{({\mathrm{\σ}}_A+{\mathrm{\σ}}_B)}^2}\}$

When the coupling of carrying out between each zone of image, adopt the UFM algorithm.

Step 5, the fuzzy characteristics of these 80～100 output images is brought into calculates in estimating based on the feedback of fuzzy similarity, recomputate the fuzzy similarity of all medical images in medical image to be compared and the property data base, export 60～80 width of cloth images again according to fuzzy similarity order from high to low, these images are referred to as the secondary output image;

The processing procedure that the feedback based on fuzzy similarity in the described step 5) is estimated is as follows:

5a) obtain outer weight term and interior weight term by the weighted product that maximizes the fuzzy similarity of inquiring about an image to be compared and an associated picture, detailed process is as follows:

After result's output first time, finished coupling between each zone among the image Q to be compared and each zone of associated picture, suppose that α mn represents n associated picture RX ⁿIn and m the regional Q in the output image _mThe subscript of institute's matching area, promptly α mn shows Q _mAnd RX _{α mn} ⁿCoupling, Q _mK region R X with n associated picture _k ⁿFuzzy similarity be:

$S_{\mathrm{\Π}}(Q_m,R{X}_k^n)=\exp \{-\frac{1}{2}\underset{i=1}{\overset{I}{\mathrm{\Σ}}}{u}_m^i\frac{{({\mathrm{\μ}}_m^i-{\mathrm{\μ}}_{\mathrm{nk}}^i)}^T{W}_m^i({\mathrm{\μ}}_m^i-{\mathrm{\μ}}_{\mathrm{nk}}^i)}{{({\mathrm{\σ}}_m^i+{\mathrm{\σ}}_{\mathrm{nk}}^i)}^2}\}$

Wherein, u _m ⁱBe outer weight term;

W _m ⁱBe interior weight term;

μ _m ⁱAnd σ _m ⁱBe regional Q _mThe fuzzy content characteristic parameter about the i category feature;

μ _Nk ⁱAnd σ _Nk ⁱBe region R X _k ⁿFuzzy content characteristic parameter about the i category feature;

Relevant feedback algorithm based on fuzzy characteristics is exactly to obtain outer weight term u by the weighted product that maximizes the fuzzy similarity of inquiring about an image to be compared and an associated picture _m ⁱWith interior weight term W _m ⁱ, constraint condition is $\begin{array}{cc}\underset{i=1}{\overset{I}{\mathrm{\Σ}}}\frac{1}{u_m^i}=1& \det \left(W_m^i\right)=1\end{array},$ Optimizing the result is:

$W_m^i={\left(\det \left(C_m^i\right)\right)}^{\frac{1}{K_i}}{\left(C_m^i\right)}^{-1}$

$C_m^{i_{\mathrm{rs}}}=\frac{{\mathrm{\Σ}}_{n=1}^N{\mathrm{\π}}_n{f}_m^{\mathrm{in}}({\mathrm{\μ}}_m^{\mathrm{ir}}-{\mathrm{\μ}}_{\mathrm{\αmn}}^{\mathrm{ir}})({\mathrm{\μ}}_m^{\mathrm{is}}-{\mathrm{\μ}}_{\mathrm{\αmn}}^{\mathrm{is}})}{{\mathrm{\Σ}}_{n=1}^N{\mathrm{\π}}_n{f}_m^{\mathrm{in}}}$

$f_m^{\mathrm{in}}=\frac{1}{{({\mathrm{\σ}}_m^i+{\mathrm{\σ}}_{\mathrm{\αmn}}^i)}^2}$

$u_m^i=\underset{j=1}{\overset{I}{\mathrm{\Σ}}}\sqrt{\frac{p_j}{p_i}}$

$p_i={\mathrm{\Σ}}_{n=1}^N{\mathrm{\π}}_n{f}_m^{\mathrm{in}}{({\mathrm{\μ}}_m^i-{\mathrm{\μ}}_{\mathrm{\αmn}}^i)}^T{W}_m^i({\mathrm{\μ}}_m^i-{\mathrm{\μ}}_{\mathrm{\αmn}}^i)$

Wherein, K _iRepresent i category feature average μ ⁱLength;

μ ^IrRepresent vectorial μ ⁱIn r element;

5b) weight term u outside obtaining _m ⁱWith interior weight term W _m ⁱAfter, recomputate the fuzzy similarity between all medical images in image Q to be compared and the property data base, according to the size of fuzzy similarity the medical image in the property data base is sorted again, export preceding 60～80 width of cloth images.

Step 6, continuous repeating step 5) processing procedure, the fuzzy characteristics of the output image that step 5) is obtained is calculated in being brought into and estimating based on the feedback of fuzzy similarity, recomputate the fuzzy similarity of all medical images in medical image to be compared and the property data base, at last output and high preceding 10～30 width of cloth medical images of medical image similarity to be compared.

Step 6 can be chosen according to actual needs, if the image of step 5) output just meets the demands, then need not step 6, and the loop iteration of step 6) can carry out once or twice or iteration repeatedly as required.

Be not limited thereto in embodiments of the present invention, the acquisition of more above-mentioned parameters also can obtain with similar computing formula, and Feature Selection, fuzzy characteristics parameter etc. are not enumerated one by one at this; According to foregoing of the present invention, ordinary skill knowledge and customary means according to this area, do not breaking away under the above-mentioned basic fundamental thought of the present invention prerequisite, the present invention can also make equivalent modifications, replacement or the change of other various ways, all can realize the object of the invention.

Claims (9)

1, a kind of relevant feedback based on medical image fuzzy region feature classification identification is estimated method, it is characterized in that comprising the steps:

1) each medical image of choosing in the medical image databases is cut apart, and is divided into several regions;

The hard feature in each zone after 2) extraction is cut apart;

3) with step 2) the hard Feature Conversion extracted becomes fuzzy characteristics, deposits in the property data base;

4) choose a medical image to be compared, adopt above-mentioned steps 1) to 3) identical treatment step extracts the fuzzy characteristics of medical image to be compared, calculate the fuzzy similarity of the fuzzy characteristics of the fuzzy characteristics of medical image to be compared and each medical image in the property data base, size according to fuzzy similarity, property data base traditional Chinese medicine image is sorted, according to fuzzy similarity order output M width of cloth image from high to low, this M width of cloth image is referred to as output image one time, wherein fuzzy similarity is referred to as associated picture one time according to preceding 50% image from high to low, and the image of back 50% is referred to as once irrelevant image;

5) fuzzy characteristics of this a M output image is brought into calculates in estimating based on the feedback of fuzzy similarity, recomputate the fuzzy similarity of obtaining all medical images in medical image to be compared and the property data base, export N width of cloth image again according to fuzzy similarity order from high to low, this N width of cloth image is referred to as the secondary output image.

2, feedback according to claim 1 is estimated method, it is characterized in that: the processing procedure that the feedback based on fuzzy similarity in the described step 5) is estimated is as follows:

A) obtain outer weight term and interior weight term by the weighted product that maximizes the fuzzy similarity of inquiring about an image to be compared and an associated picture;

B) outside obtaining after weight term and the interior weight term, recomputate the fuzzy similarity between all medical images in image to be compared and the property data base, size according to fuzzy similarity sorts again to the medical image in the property data base, N width of cloth image before the output.

3, feedback according to claim 1 is estimated method, it is characterized in that: this method also comprises step 6):

Continuous repeating step 5) processing procedure, the fuzzy characteristics of the output image that step 5) is obtained is calculated in being brought into and estimating based on the feedback of fuzzy similarity, recomputate the fuzzy similarity of obtaining all medical images in medical image to be compared and the property data base, at last output and the high preceding L width of cloth medical image of medical image similarity to be compared.

4, feedback according to claim 1 is estimated method, it is characterized in that: adopt gaussian filtering to carry out pre-service to the input medical image in the step 1).

5, feedback according to claim 1 is estimated method, it is characterized in that: above-mentioned steps 3) be that the utilization index membership function becomes fuzzy characteristics with hard Feature Conversion.

6, feedback according to claim 1 is estimated method, it is characterized in that: described M value is 80～100, and the N value is 60～80.

7, feedback according to claim 3 is estimated method, it is characterized in that: described L value is 10～30.

8, feedback according to claim 1 is estimated method, it is characterized in that: above-mentioned steps 1) be to utilize the EM algorithm of the limited gauss hybrid models of parameter that medical image is cut apart.

9, feedback is estimated method according to claim 1 or 5, and it is characterized in that: described hard feature comprises average gray feature, invariant moments shape facility, Gabor textural characteristics, direction histogram feature.

Images