CN104881687A - Magnetic resonance image classification method based on semi-supervised Gaussian mixed model - Google Patents

Abstract

The invention discloses a magnetic resonance image classification method based on a semi-supervised Gaussian mixed model, and relates to the technical field of image processing. First, an MR image slice is selected randomly from an original data set; then, clustering is performed by a k-means algorithm; and finally, classification is completed by use of statistics and the cluster label information of an obtained cluster. By adopting the method, the convergence rate is increased. Compared with two kinds of supervised Gaussian mixed models, a better segmentation result is obtained, there is no need to mark a training data set, the time of image processing is reduced, and the precision of image processing is improved.

Description

Based on the magnetic resonance image (MRI) sorting technique of semi-supervised gauss hybrid models

Technical field

The present invention relates to a kind of image processing method technical field, particularly relate to a kind of magnetic resonance image (MRI) sorting technique based on semi-supervised gauss hybrid models.

Background technology

At present, magnetic resonance image (MRI) (Magnetic Resonance Image, MRI) be widely used in the diagnosis of sacred disease, available programs guarantees safe, painless and Noninvasive researching human body, and it identifies abnormal before illness of being everlasting occurs for a long time and between the apparition.Especially, magnetic resonance image (MRI) is applicable to research the nervous system disease very much, due to the multispectral characteristic of its high spatial resolution, high soft tissue contrast and image, there is loose time (T1 and T2) and proton density (Proton density, Pd) information.

It is a complicated and task for difficulty that the MRI that undertaken by the human expert of training analyzes, because structures of interest shows complex edge configuration in image, the anatomical borders most of the time is unintelligible visible.In clinical testing, MR image set is usually all very large, thus human expert to carry out manual analysis very consuming time, and it be not immediately clear how expert merges the information from different channel when checking multispectral MR data.Therefore, with manually to split in relevant observer and between observer, changeability hinders the reappearance of result.For these reasons, robotization or the semi-automatic technology that can analyze the MR brain image segmentation of a large amount of 3D multispectral MR data are in a repeatable manner necessary.

About the quantitative measurement of brain anatomy, the key components of graphical analysis obtain the segmentation of accurate brain image, especially grey matter (GM), white matter (WM) and cerebrospinal fluid (CSF) from various anatomical structure or organization type.Brain image segmentation not only adapts to assessment for cortical surface mapping, cubing, tissue typing, function and figure and the nervous system disease characterizes, and it is the preliminary step needed for other image processing programs many, such as brain registration and the form based on voxel.Therefore, the Accurate Segmentation of brain image becomes one of most important tissue in MRI application, segmentation based on image voxel attribute, neighborhood information or geometrical property, obtain exact image segmentation difficulty because of noise, unevenness, local volume impact and cortex high complexity geometric configuration and increase.

Rely on training sample mark availability, Iamge Segmentation have supervision otherwise without supervision, generally speaking, based on the segmentation of supervised learning, such as neural network or support vector function produce good result, but it needs every class loading to there is a large amount of training datas (mark voxel), expensive and consuming time.

On the contrary, unsupervised learning method, such as k average or the method based on mixture model, possess how confirmed advantage than supervised classification method, such as user interactions is few.A kind of optimizing process is actually without supervision technology because nearly all, arranged by objective function, such as, total log-likelihood in hybrid modeling or the Euclidean distance summation in k average, this technology is inevitably subject to the impact (minimum or maximum) of local problem.Therefore, must make suitable adjustment to them, to produce gratifying result, namely without any priori, these method performances have limitation.

Semi supervise algorithm combines priori in unsupervised approaches, can improve the result of Data classification, and without the need to complete training dataset.Recently, scholars propose some technology of semi-supervised brain image segmentation, such as semi-supervised Maximun Posterior Probability Estimation Method (Semi-supervised Maximum A Posteriori, ssMAP).Document: Fan Wanshu. based on the medical image segmentation system [D] of semi-supervised fuzzy clustering. Dalian University of Technology, 2013. algorithms proposed improve segmentation result by exploring imperfect training dataset, wherein, flag data may, only to organizing subset to use, namely not be all types that marked tissue.Document: Bauer S, Tessier J, KrieterO, et al.Integrated spatio-temporal segmentation of longitudinal braintumor imaging studies [M] //Medical Computer Vision.Large Data in MedicalImaging.Springer International Publishing, 2014:74-83. propose a kind of MR brain image sorting technique, use semi-supervised spectral clustering, improve the result of segmentation image.They provide some data to constraint information for spectral clustering, replace voxel mark, distribute to obtain higher data.

Summary of the invention

Technical matters to be solved by this invention is to provide a kind of magnetic resonance image (MRI) sorting technique based on semi-supervised gauss hybrid models, described method accelerates speed of convergence, compare two kinds and have supervision gauss hybrid models, obtained better segmentation result, and without the need to marking training dataset, reduce image processing time, improve image procossing precision.

For solving the problems of the technologies described above, the technical solution used in the present invention is: a kind of magnetic resonance image (MRI) sorting technique based on semi-supervised gauss hybrid models, it is characterized in that: this method is realized by two modules: priori cluster module and sort module, priori cluster module has the class label information of class statistic tolerance for associating; Sort module is used for receiving prior imformation from priori cluster module and generates marking image;

Make Y={X ¹... X ^mit is the 3D magnetic resonance image (MRI) represented by one group of 2D section, wherein m is the number of slices in image, first, and Stochastic choice one width magnetic resonance image (MRI) section from Y, the voxel of this image slice of cluster, the cluster of each generation is labeled as grey matter GM, white matter WM or cerebrospinal fluid CSF by human expert; Then, label information and class statistic tolerance: average, variance matrix and prior probability are used for the assorting process of other sections of image; Finally, the class label information classification residual image section of statistics and gained cluster is utilized.

Further technical scheme is: make Y={X ¹... X ^mbe the 3D magnetic resonance image (MRI) represented by one group of 2D section, wherein m is the number of slices in image, Stochastic choice magnetic resonance image (MRI) section X from Y ^g, be clustered into three groups by k mean algorithm, result is the image of a width segmentation, is respectively grey matter GM image sections, white matter WM image sections, cerebrospinal fluid CFS image sections; Human expert assessment cluster result, if good segmentation image, human expert is by the section cluster c of each segmentation _jbe associated with one of them class l={GM, WM, CSF}, otherwise, again use different initial parameter ψ ⁰segmentation section X ^g.

Further technical scheme is: the magnetic resonance image (MRI) section X of Stochastic choice ^gevery section have a class label, suppose mixture model assembly and organize between class that there is man-to-man corresponding relation, thus use c _jrepresent a jth electric hybrid module, i.e. jth class, known tissue class and X ^gcorresponding relation between cluster, for the cluster voxel that it distributes calculates each class c _jhybrid parameter collection ψ _j={ μ _j, Σ _j, π _j,

${\mathrm{\π}}_j=\frac{\left|n_j\right|}{n}$

${\mathrm{\μ}}_j=\frac{{\mathrm{\Σ}}_{x^i\∈c_j}{x}^i}{\left|n_j\right|}$

${\mathrm{\Σ}}_j=\frac{{\mathrm{\Σ}}_{x^i\∈c_j}(x^i-{\mathrm{\μ}}_j){(x^i-{\mathrm{\μ}}_j)}^T}{\left|n_j\right|}$

In formula, l _jfor association class label, μ _jfor the average of class statistic tolerance, Σ _jfor the variance matrix of class statistic tolerance, π _jfor the prior probability of class statistic tolerance, | n _j| for belonging to the number of voxels of jth class, the cluster of each class and label are to, hybrid parameter set ψ _jrepresent the priori of data set.

Further technical scheme is: the priori of data set is used for the section of cluster residual image, and residual image set of slices Y* is made up of Y, does not comprise the section X for cluster analysis ^g, when Bayes classifier is submitted in the new section of Y*, what calculate in early days organizes class hybrid parameter collection for the initial parameter of semi-supervised gauss hybrid models, restrain with level and smooth expectation-maximization algorithm, residual image section is classified.

The beneficial effect adopting technique scheme to produce is: the improper initial selected produced problem that this method avoids mixture model parameter estimation period ψ when using expectation-maximization algorithm EM.This method, perform the classification of mankind's brain area in magnetic resonance image (MRI) section, without the need to marking training dataset, the human expert analysis of use is fewer than there being measure of supervision.Semi-supervised learning algorithm is used to compare the speed of convergence that other unsupervised learning with Bayes classifier accelerates gauss hybrid models GMM, the time of semi supervise algorithm consumption is less than the unsupervised learning with Bayes classifier, because when initial parameter is faster based on clustering algorithm convergence during priori.Because gauss hybrid models GMM is responsive to initial method, the priori of parameter improves its precision.

Accompanying drawing explanation

Below in conjunction with the drawings and specific embodiments, the present invention is further detailed explanation.

Fig. 1 is the overall flow figure of the method for the invention;

Fig. 2 is the process flow diagram of priori sorting procedure in the present invention;

Fig. 3 is the process flow diagram of classifying step in the present invention;

Fig. 4 represents the original BrainWeb image of slice numbers 50,93 and 120;

Fig. 5 Fig. 4 carries out the identical image after pre-treatment step;

Fig. 6 is the real image of Fig. 5 after this method process.

Embodiment

Below in conjunction with the accompanying drawing in the embodiment of the present invention, be clearly and completely described the technical scheme in the embodiment of the present invention, obviously, described embodiment is only a part of embodiment of the present invention, instead of whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art, not making the every other embodiment obtained under creative work prerequisite, belong to the scope of protection of the invention.

Set forth a lot of detail in the following description so that fully understand the present invention, but the present invention can also adopt other to be different from alternate manner described here to implement, those skilled in the art can when without prejudice to doing similar popularization when intension of the present invention, therefore the present invention is by the restriction of following public specific embodiment.

Cluster based on gauss hybrid models:

X is made to be by voxel collection X={x ¹... x ⁿthe image slice that represents, wherein, each voxel is that a d ties up random vector x ⁱ=(x ₁..., x _d), n is the number of voxels in section.Supposing that each region j of composition diagram picture follows probability density function is p _j(x| θ _j) the distribution of class condition, each has its oneself parameter vector θ _j, j=1 ..., k, wherein, k organizes number in image, and therefore, each voxel is independent of the hybrid density describing all class weighted sums:

$p\left(x^i\right|\mathrm{\ψ})=\underset{j=1}{\overset{k}{\mathrm{\Σ}}}{\mathrm{\π}}_j{p}_j\left(x^i\right|{\mathrm{\θ}}_j)---\left(1\right)$

In above formula, π _jbe mixing mixing ratio, be just, and equal 1, p _j(x| θ _j) be the component density relevant to region j, ψ={ θ ₁..., θ _k, π ₁..., π _kthe set of all hybrid parameters, mixing mixing ratio π _icorresponding to the prior probability of any voxel belonging to jth group.

In order to use the clustering method based on mixture model, suppose that data to be clustered belong to the mixing of designation number object k group in various ratio, each data point is independent of the hybrid density in formula (1), and wherein k assembly corresponds to k group, each component density p _j(x| θ _j) parameterized form illustrate after, parameter is by maximal possibility estimation.Can obtain the probability clustering of data after mixing, for data are equipped with the posterior probability of component functions, the data being clustered into k group are obtained by the assembly each data point being categorized into the highest estimated probability.

Suppose component density p _j(x| θ _j) normal state, there is parameter θ _j={ μ _j, Σ _j, variable x ⁱindependent and with distribution (i.i.d), make μ _jand Σ _jbe respectively unknown mean vector and the variance matrix of group j, have:

$p_j\left(x^i\right|{\mathrm{\θ}}_j)=\frac{1}{2{\mathrm{\π}}^{d/2}\left|{\mathrm{\Σ}}_j\right|}\exp [-\frac{1}{2}(x^i-{\mathrm{\μ}}_j){\mathrm{\Σ}}_j^{-1}{(x^i-{\mathrm{\μ}}_j)}^T]---\left(2\right)$

In formula, | Σ | be the determinant of Σ.

Estimate that outfit obtains blended data by maximum likelihood (Maximum Likelihood, ML) after formalized model, the target of ML obtains the parameter maximizing data available joint probability density function (or data likelihood).Use expectation maximization (Expectation Maximization, EM) algorithm to perform, conveniently, usually use log-likelihood to replace likelihood, be defined as:

Document Choi J, Kim D, Oh C, et al.An iterative reconstruction method ofcomplex images using expectation maximization for radial parallel MRI [J] .Physics in medicine and biology, 2013,58 (9): 2969-2976. describe a kind of EM algorithm, to hypothesis p _j(x| θ _j) situation of normal state carries out the ML parameter estimation of component density.In image clustering method, slice element is for upgrading hybrid parameter, and alternately following EM step is until lnp (X| ψ) convergence:

Desired step:

$z_j\left({\mathrm{\ψ}}^t\right|x^i)=\frac{{\mathrm{\π}}_j{p}_j\left(x^i\right|{\mathrm{\θ}}_j)}{p\left(x^i\right|{\mathrm{\ψ}}^t)},j=1,...,k---\left(4\right)$

Maximization steps:

${\mathrm{\π}}_j^{t+1}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{z}_j\left({\mathrm{\ψ}}^t\right|x^i),j=1,...,k---\left(5\right)$

${\mathrm{\μ}}_j^{t+1}=\frac{{\mathrm{\Σ}}_{i=1}^n{x}^i{z}_j\left({\mathrm{\ψ}}^t\right|x^i)}{n{\mathrm{\π}}_j^{t+1}},j=1,...,k---\left(6\right)$

With

$\underset{j}{\overset{t+1}{\mathrm{\Σ}}}=\frac{{\mathrm{\Σ}}_{i=1}^n(x^i-{\mathrm{\μ}}_j^{t+1}){(x^i-{\mathrm{\μ}}_j^{t+1})}^T{z}_j\left({\mathrm{\ψ}}^t\right|x^i)}{n{\mathrm{\π}}_j^{t+1}},j=1,...,k---\left(7\right)$

EM algorithm originates in initial value ψ ⁰, ψ ⁰improper selection hinder EM convergence, under certain situation, likelihood is to the limit unbounded of parameter space, if the ψ selected ⁰too close to border, the argument sequence generated by EM is estimated to disperse.Another problem of mixture model is that likelihood equation has many to correspond to locally maximum usually, and therefore EM algorithm should be applied to and extensively select initial value, and maximum to search for local, this is very consuming time.

Parameter ψ describes tolerance x ⁱglobal property, but they do not show how distribute labels is to voxel, are equipped with k-assembly mixture model with after the estimation obtaining ψ, distribute n voxel { x ¹..., x ⁿprobability clustering, be equipped with posterior probability their assembly member.For each x ⁱ, k Probability p ₁(ψ ₁| x ⁱ) ..., p _k(ψ _k| x ⁱ), there is ψ _j={ μ _j, Σ _j, π _j, provide the posterior probability of estimation, this observation belongs to the 1st, the 2nd of mixing respectively ... with a kth assembly, distribute each x ⁱthe hard cluster of these voxels can be provided to reply to the electric hybrid module with the highest posterior probability.Make r ⁱfor component label vector, if i-th voxel belongs to a jth group, then otherwise any voxel prior probability belonging to j group is π _j, probability voxel x ⁱbelong to the probability of j group by class conditional density function p _j(x ⁱ| θ _j) provide, i.e. electric hybrid module.If then corresponding voxel is by component density p _j(x ⁱ| θ _j) generate, that is:

$p\left(x^i\right|r_j^i=1)=p_j\left(x^i\right|{\mathrm{\θ}}_j)---\left(8\right)$

By Bayes rule, the posterior probability of assembly member is provided by following formula:

$p_j\left({\mathrm{\ψ}}_j\right|x^i)=\mathrm{Prob}(r_j^i=1|x^i,{\mathrm{\ψ}}_j)=\frac{{\mathrm{\π}}_j{p}_j\left(x^i\right|{\mathrm{\θ}}_j)}{p\left(x^i\right|\mathrm{\ψ})}---\left(9\right)$

In above formula, denominator p (x ⁱ| ψ) provided by formula (1), this is one and depends on tolerance x ⁱclass constant value.Generally speaking, x ⁱfor estimating utilization formula (9), if the parameter ψ of component density is known, Bayes classifier can be used for distributing class label to voxel, if parameter ψ is known:

$j*=\arg \underset{j}{\max }{p}_j\left({\mathrm{\ψ}}_j\right|x^i)---\left(10\right)$

Voxel i distributes to the class j* with the highest posterior probability.

Research shows, the clustering method based on gauss hybrid models can be advantageously applied to MR Iamge Segmentation, and shortcoming is to need to mark training dataset, and clustering algorithm convergence time is longer.

The present invention proposes a kind of magnetic resonance image (MRI) sorting technique based on semi-supervised gauss hybrid models, improve magnetic resonance MR Iamge Segmentation, accelerating convergence process, use the semisupervised classification based on cluster, do not need to mark training dataset, the human expert analysis used than there being measure of supervision is few.Compare other semi-supervised technology, priori (such as cluster label information and statistics are measured) is easy to obtain, without the need to any mark training data.

Semi supervise algorithm combines priori in unsupervised approaches, can improve the result of Data classification, without the need to marking training dataset, can overcome the defect of gauss hybrid models cluster, therefore, proposes gauss hybrid models and merges Novel semi-supervised.

Context of methods is based on four hypothesis of data available: (1) data are produced by mixture model; (2) man-to-man corresponding relation is had between assembly and brain tissue class; (3) electric hybrid module is the multinomial distribution in the multiple regions presented in image; (4) the identical brain tissue class presented in different images section distributes to same distribution.

Make Y={X ¹... X ^mbe the 3DMR image represented by one group of 2D section, wherein m is the number of slices in image.This method is made up of two modules: priori cluster module (module 1), and association has the class label information of class statistic tolerance (average, variance matrix and prior probability); Sort module (module 2), receives prior imformation from priori cluster module and generates marking image.The block diagram of whole process as shown in Figure 1, first, the voxel of cluster piece image section, the cluster of each generation is labeled as grey matter (GM), white matter (WM) or cerebrospinal fluid (CSF) by human expert; Then, label information and class statistic tolerance: average, variance matrix and prior probability are used for the assorting process of other sections of image; Finally, the class label information classification residual image section of statistics and gained cluster is utilized.

Priori cluster:

Priori cluster module as shown in Figure 2, Stochastic choice MR image slice X from Y ^g, be clustered into three groups by k mean algorithm, result is the image of a width segmentation, without any association between image sections, and brain tissue's class: grey matter (GM), white matter (WM), cerebrospinal fluid (CSF).In next step, human expert assessment cluster result, if good segmentation image, human expert is by the section cluster c of each segmentation _jbe associated with one of them class l={GM, WM, CSF}.Otherwise, again use different initial parameter ψ ⁰segmentation section X ^g, expert intervenes the supervision corresponded in semi-supervised learning method.

X ^gevery section of image has a class label, supposes mixture model assembly and organize between class to have man-to-man corresponding relation, thus uses c _jrepresent a jth electric hybrid module, i.e. jth class.Known tissue class and X ^gcorresponding relation between cluster, for the cluster voxel that it distributes calculates each class c _jhybrid parameter collection ψ _j={ μ _j, Σ _j, π _j,

${\mathrm{\π}}_j=\frac{\left|n_j\right|}{n}---\left(11\right)$

${\mathrm{\μ}}_j=\frac{{\mathrm{\Σ}}_{x^i\∈c_j}{x}^i}{\left|n_j\right|}---\left(12\right)$

${\mathrm{\Σ}}_j=\frac{{\mathrm{\Σ}}_{x^i\∈c_j}(x^i-{\mathrm{\μ}}_j){(x^i-{\mathrm{\μ}}_j)}^T}{\left|n_j\right|}---\left(13\right)$

In formula, | n _j| for belonging to the number of voxels of jth class, the cluster of each class and label are to, hybrid parameter set ψ _jrepresent the priori of data set, association class label information l in Fig. 2 _j(average μ is measured with class statistic _j, variance matrix Σ _jwith prior probability π _j).

Classification

As shown in Figure 3, the priori of data set is used for the section of cluster residue to classifying step flow process, and residue set of slices Y* is made up of Y, does not comprise the section X for module 1 ^g.When Bayes classifier is submitted in the new section of Y*, what calculate in early days organizes class parameter for the initial parameter of gauss hybrid models (GMM), restrain with level and smooth EM.Because extracting parameter from the image of good segmentation, so should be ψ _jeM is made to converge on the overall situation maximum.

This method is also supposed corresponding to organizing class c _jsection X ^gin voxel and correspond to c _jresidue section in voxel produced by same distribution.Therefore, zones of different is assigned to different hybrid parameter, each electric hybrid module p of formula (1) _j(x| θ _j) organize class corresponding to one, receive the label ψ distributing to hybrid parameter _j, as initial parameter.Therefore, after using the clustering method process GMM cluster of gauss hybrid models, each voxel x ⁱthat distributes to mixing has marker assemblies, and it has the highest posterior probability.Result is the image of segmentation, and wherein, all voxels belonging to cluster receive the class label distributing to the electric hybrid module of cluster, and namely the cluster of cutting procedure not necessarily will associate with each class of organizing by new classifying step.

Experiment

Data set

In order to the performance of quantitative evaluation context of methods, use the synthesis MR image of simulating brain data storehouse from BrainWeb, result is based on the synthesis normal brain multispectral image be made up of proton density, T1 and t2 weighted image, 181 dimensions are had to be the sagittal slices of 181 × 217 voxels, there is 1 cubic millimeter of resolution, 3% noise level, 0% strength non-uniformity.

Therefore, in order to assess Gaussian density function, need the inverse of variance matrix, the estimation of variance matrix is only useful to classification, if nonsingular (namely reversible).In 181 sagittal slices, select at least 10 to belong to each those sections organizing class, guarantee to estimate that Gaussian Distribution Parameters becomes possibility, the raw 133 width images of common property.Three classes are considered in this experiment: white matter (WM), grey matter (GM) and cerebrospinal fluid (CSF), and before Iamge Segmentation, pre-treatment step is applied to splits brain from non-brain tissue.Fig. 4 represents the original BrainWeb image of slice numbers 50,93 and 120, there is 3% noise and 0% strength non-uniformity, Figure 5 shows that the identical image after pre-treatment step, Figure 6 shows that corresponding ground truth, the RGB version of the multispectral image that in Fig. 4, image is made up of proton density, t1 weighted image and t2 weighted image.

Experimental design

Test semi-supervised gauss hybrid models (GMM this method proposed ^sSC) have and supervise gauss hybrid models with other two kinds and compare.

First method is a kind of Cluster Program with random initializtion parameter, cluster result be a width segmentation without label image, the then necessary labeled clusters of classifying step, this method is called GMM ^sUP1.In training step, Stochastic choice image slice, uses and has stray parameter initialized k mean algorithm cluster, and the cluster that handmarking produces is WM, GM or CSF, and calculates their statistics parameter.Therefore, each cluster and class label l _jassociation, also with hybrid parameter collection ψ _jassociation.Sorting phase utilizes after the section of stray parameter initialized gauss hybrid models segmentation residual image, the parametric classification that consequent cluster is calculated based on first step by parametrization Bayes classifier.

Another kind method be context of methods have measure of supervision, be called GMM ^sUP2.Because image ground truth comprises the class label of verification msg collection, use this available information to have monitor mode to calculate hybrid parameter set ψ herein.Chapter 2, in the module 1 described, eliminate cluster process, use formula (11)-(13) directly to calculate ψ from flag data, GMM ^sUP2use module 2.

Appraisal procedure

In order to assess classification results, use segmentation and the ground truth of dice index of similarity (DSI) more each class j, dice index of similarity S (j) ^[4]be defined as:

$\mathrm{DSI}=S\left(j\right)=\frac{2N_{p\∩g}\left(j\right)}{N_p\left(j\right)+N_g\left(j\right)}---\left(14\right)$

In formula, N _{p ∩ g}j () is that context of methods and ground truth are classified to the number of voxels of class j, N _p(j) and N _gj () represents and is classified to the number of voxels of class j by context of methods and ground truth respectively.If context of methods and ground truth match, index S (j) tends to 1, along with segmentation deterioration, is reduced to 0, represent there is a fabulous agreement between two segmentations.

Because two kinds realize all responsive to initialization, each 30 times of two kinds of methods are run herein, in each new execution, for module 1 Stochastic choice new images and residual image of classifying.

Use the performance of paired t test and comparison method, following agreement used to p value: "? " "=" represents that p value is less than or equal to 0.01, is strong evidence, and a kind of method produces the greater or lesser value of measure of effectiveness than another method." > " and " < " represents that p value is greater than 0.01 and be less than or equal to 0.05, is weak evidence, and a kind of method produces the greater or lesser value of measure of effectiveness than another kind of method." ~ " represents that p value is greater than 0.05, namely compares the performance of two kinds of methods, and it does not have the significance difference opposite sex.

Experimental analysis

Table 1 represents DSI value (average and standard deviation), use the average acquisition of class distribution of various method and all images, optimum is represented by runic, and table 2 is the subform of table 1, represents t test result, compares GMM respectively ^sSCand GMM ^sUP1, GMM ^sSCand GMM ^sUP2.

The DSI (means standard deviation) of three kinds of tissues in each method Iamge Segmentation of table 1

The comparison of table 2 three kinds of methods, GMM ^sSC× GMM ^sUP1and GMM ^sSC× GMM ^sUP2: t-test result

As can be seen from table 1, table 2, for CSF, GM and WM, GMM ^sSCcompare GMM ^sUP1have statistically significant superior function, and GMM cluster selects improper meeting to have a strong impact on performance by initial parameters, the priori of the relevant parameters that this method is provided by the semi-supervised method of use avoids these impacts.This method is also observed, and for CSF, GM and WM, compares GMM ^sUP2with GMM ^sSC, there is no statistically-significant difference, show that prior imformation that semi-supervised method provides is for module 1, without any training data, is similar to by the information having measure of supervision to calculate training dataset.

In addition, obtain the CPU time of all 132 test patterns of process, the performance of comparative approach, relative to assessing the cost.For GMM ^sUP1, GMM ^sUP2and GMM ^sSC, use Intel multinuclear 1.66GHz processor, 2G internal memory and MATLAB7.0 version, this method time-related second average and standard deviation be respectively 709 ± 994.72,564.65 ± 431.23,609.74 ± 204.37.

GMM ^sUP2compare GMM ^sUP1, GMM ^sSChurry up, because GMM ^sUP1and GMM ^sSCthe semi-supervised method based on cluster is used to calculate initial mixing parameter sets ψ, and (GMM ^sUP2) directly use the information of ground truth to calculate ψ, therefore, GMM ^sUP2calculate the time mainly direct application shown in formula (11)-(13) that ψ consumes, be significantly less than the clustering convergence time of same data set.

But, when training dataset is unavailable, can not GMM be performed ^sUP2, therefore GMM ^sSCavailable, because its implementation effect is better than GMM ^sUP1, create statistics and obtain remarkable improvement, i.e. the p value of t test is in pairs less than 0.01.This experiment shows, because GMM is responsive to initialization technique, so the priori of parameter makes clustering algorithm Fast Convergent.

This method is that mankind's brain MR image proposes a kind of magnetic resonance image (MRI) sorting technique based on semi-supervised gauss hybrid models, this method avoids the improper initial selected produced problem of mixture model parameter estimation period ψ when using EM algorithm.This method performs the classification of mankind's brain area in MR image slice, and without the need to marking training dataset, the human expert analysis of use is fewer than there being measure of supervision.

Semi-supervised learning algorithm is used to compare the speed of convergence that other unsupervised learnings with Bayes classifier accelerate GMM, the time of semi supervise algorithm consumption is less than the unsupervised learning with Bayes classifier, because when initial parameter is faster based on clustering algorithm convergence during priori.Because GMM is responsive to initial method, the priori of parameter improves its precision.

Claims (4)

1. based on a magnetic resonance image (MRI) sorting technique for semi-supervised gauss hybrid models, it is characterized in that: this method is realized by two modules: priori cluster module and sort module, priori cluster module has the class label information of class statistic tolerance for associating; Sort module is used for receiving prior imformation from priori cluster module and generates marking image;

Make Y={X ¹... X ^mit is the 3D magnetic resonance image (MRI) represented by one group of 2D section, wherein m is the number of slices in image, first, and Stochastic choice one width magnetic resonance image (MRI) section from Y, the voxel of this image slice of cluster, the cluster of each generation is labeled as grey matter GM, white matter WM or cerebrospinal fluid CSF by human expert; Then, label information and class statistic tolerance: average, variance matrix and prior probability are used for the assorting process of other sections of image; Finally, the class label information classification residual image section of statistics and gained cluster is utilized.

2. the magnetic resonance image (MRI) sorting technique based on semi-supervised gauss hybrid models according to claim 1, is characterized in that:

Make Y={X ¹... X ^mbe the 3D magnetic resonance image (MRI) represented by one group of 2D section, wherein m is the number of slices in image, Stochastic choice magnetic resonance image (MRI) section X from Y ^g, be clustered into three groups by k mean algorithm, result is the image of a width segmentation, is respectively grey matter GM image sections, white matter WM image sections, cerebrospinal fluid CFS image sections; Human expert assessment cluster result, if good segmentation image, human expert is by the section cluster c of each segmentation _jbe associated with one of them class l={GM, WM, CSF}, otherwise, again use different initial parameter ψ ⁰segmentation section X ^g.

3. the magnetic resonance image (MRI) sorting technique based on semi-supervised gauss hybrid models according to claim 2, is characterized in that the concrete steps of cluster are:

The magnetic resonance image (MRI) section X of Stochastic choice ^gevery section have a class label, suppose mixture model assembly and organize between class that there is man-to-man corresponding relation, thus use c _jrepresent a jth electric hybrid module, i.e. jth class, known tissue class and X ^gcorresponding relation between cluster, for x ⁱ, the cluster voxel that it distributes calculates each class c _jhybrid parameter collection ψ _j={ μ _j, Σ _j, π _j,

${\mathrm{\&pi;}}_j=\frac{\left|n_j\right|}{n}$

${\mathrm{\&mu;}}_j=\frac{{\mathrm{\&Sigma;}}_{x^i\&Element;c_j}{x}^i}{\left|n_j\right|}$

${\mathrm{\&Sigma;}}_j=\frac{{\mathrm{\&Sigma;}}_{x^i\&Element;c_j}(x^i-{\mathrm{\&mu;}}_j){(x^i-{\mathrm{\&mu;}}_j)}^T}{\left|n_j\right|}$

In formula, l _jfor association class label, μ _jfor the average of class statistic tolerance, Σ _jfor the variance matrix of class statistic tolerance, π _jfor the prior probability of class statistic tolerance, | n _j| for belonging to the number of voxels of jth class, the cluster of each class and label are to, hybrid parameter set ψ _jrepresent the priori of data set.

4. the magnetic resonance image (MRI) sorting technique based on semi-supervised gauss hybrid models according to claim 3, it is characterized in that: the priori of data set is used for the section of cluster residual image, residual image set of slices Y* is made up of Y, does not comprise the section X for cluster analysis ^g, when Bayes classifier is submitted in the new section of Y*, what calculate in early days organizes class hybrid parameter collection ψ _j={ μ _j, Σ _j, π _j, for the initial parameter of semi-supervised gauss hybrid models, restrain with level and smooth expectation-maximization algorithm, residual image section is classified.