CN101515367A - Method for segmenting sulus regions on surface of pallium of a three-dimensional cerebral magnetic resonance image - Google Patents

Abstract

The invention relates to a method for segmenting a sulus region on the surface of a pallium of a three-dimensional cerebral magnetic resonance image. The technical characteristics lie in that firstly, the 3D cerebral nuclear magnetic resonance image is pretreated and the surface of the pallium is reconstructed, including removing a braincase and non-cerebral tissues; carrying out cerebral tissue segmenting on a cerebral image and reconstructing the surface of the pallium of geometrical accuracy and correct topological structure in the segmented cerebral image; the surface of the pallium is shown by a series of vertexes and triangles; secondly, the maximum principle curvature and the minimum principle curvature of each vertex on the surface of the pallium are estimated; and finally, sulus and gyrus regions are segmented on the surface of the pallium according to the maximum principle curvature and Hidden Markov Random Field expectation maximization framework and each sulus region is marked by connectivity analysis. Compared with other methods, the method has the advantages of simple and effective algorithm and high segmentation accuracy.

Description

Brain ditch region segmentation method on the brain cortex surface of three-dimensional brain magnetic resonance image

Technical field

The present invention relates to brain ditch region segmentation method on a kind of brain cortex surface of three-dimensional brain magnetic resonance image, belong to Medical Image Processing, calculate fields such as Nervous System Anatomy.Be applicable to cutting apart of brain ditch zone on the brain cortex surface of the trigonometric ratio that human three-dimensional brain nuclear magnetic resonance image reconstructs.

Background technology

The human brain cortex is an extremely complicated anatomical structure of curling, and mainly is made of brain ditch and gyrus, corresponds respectively to paddy and ridge on the cerebral cortex.Although the precise geometrical patterns of change of brain ditch between the different people and gyrus is very big, topmost several brain ditches and gyrus are the total anatomic landmarks on the cerebral cortex.Therefore, main brain ditch and gyrus have been widely used in auxiliary non-linear cerebral nucleus magnetic resonance image (MRI) registration, analyze the anatomical structure Changing Pattern of normal person's brain, and are used for distinguishing normal person and disease patient.But owing to manual cut apart and demarcates brain ditch and time-consuming, and be subjected to extraneous subjective the influence easily.

Brain ditch zone cut apart the hot subject that becomes research automatically in the last few years.Various brain ditch region segmentation method is suggested, but still has a lot of problems.Method 1: based on the dividing ridge method of brain trench depth on the brain cortex surface, this method at first utilizes the active surface method to find the gyrus zone, utilizes dividing ridge method to extract brain ditch zone then, utilizes heuristic rule to merge the brain ditch zone of over-segmentation at last.Its shortcoming is, the brain trench depth can not well be distinguished buried gyrus zone in brain ditch zone and the brain ditch, because the degree of depth in buried gyrus zone is all very big in brain ditch zone and the brain ditch, dividing ridge method is easy to produce the over-segmentation phenomenon, with a brain ditch Region Segmentation is a plurality of brain ditches zones, and heuristic rule merges the difficult control in over-segmentation brain ditch zone; Method 2: based on the figure cutting method of mean curvature on the brain cortex surface, this method is regarded brain cortex surface as a connected graph structure, according to mean curvature information, utilizes the figure cutting method that brain cortex surface is divided into brain ditch zone and gyrus zone.Characteristics are that figure cutting method efficient is very high, but mean curvature can not well be distinguished brain ditch and gyrus zone; Method 3: based on the bayes method of mean curvature and brain trench depth, this method at first utilizes bayes method that brain cortex surface is divided into brain ditch and gyrus zone, utilize the watershed region growing method to extract single brain ditch zone then, utilize heuristic rule to merge the brain ditch zone of over-segmentation at last.Shortcoming is that the defined Euclidean brain of this method trench depth is not real brain trench depth, and mean curvature can not well be distinguished brain ditch and gyrus zone simultaneously.

Brain ditch region segmentation method has following three main defectives on the present existing brain cortex surface: one, utilize average principal curvatures to distinguish brain ditch and gyrus zone, but average principal curvatures is the mean value of maximum principal curvatures and minimum principal curvatures, and the minimum principal curvatures in brain ditch and gyrus zone is all very little, so mean curvature can not well be distinguished brain ditch and gyrus zone.Two, utilize the brain trench depth to distinguish brain ditch and gyrus zone, but the degree of depth in buried gyrus zone is all very big in brain ditch zone and the brain ditch, so the brain trench depth can not well be distinguished brain ditch and gyrus zone.Three, utilizing dividing ridge method and the single brain ditch of heuristic merging Rule Extraction zone, is a plurality of brain ditches zones but dividing ridge method is easy to a brain ditch Region Segmentation, and the difficult control of heuristic merging rule.

Summary of the invention

The technical matters that solves

For fear of the weak point of existing method, the present invention proposes brain ditch region segmentation method on a kind of brain cortex surface of three-dimensional brain magnetic resonance image, can obtain geometry accurately and the correct brain cortex surface image of topological structure.

Technical scheme

Basic thought of the present invention is: the maximum principal curvatures on brain cortex surface summit be respectively in brain ditch zone and gyrus zone negative value and on the occasion of, we utilize two mixed Gauss models to come the histogram distribution of the maximum principal curvatures of modeling, in order simultaneously statistical information and spatial neighborhood information to be merged in the gauss hybrid models, utilize hidden Markov random field expectation maximization framework to carry out cutting apart of brain ditch zone.

Technical characterictic concrete steps of the present invention are as follows:

Step 1 pair three-dimensional brain magnetic resonance image carries out pre-service: utilize the changeability model method to remove skull, utilize method for registering to remove non-cerebral tissue, utilize the gauss hybrid models method that brain image is carried out tissue segmentation, obtain white matter, the image that three kinds of types of organizations of grey matter and celiolymph represent;

Step 2 brain cortex surface is rebuild: the brain cortex surface that utilizes Marching Cubes method reconstruct trigonometric ratio from the brain image after the tissue segmentation that above-mentioned steps obtains;

Step 3 utilizes finite difference method to estimate to obtain the maximum principal curvatures on each summit on the brain cortex surface:

Step 4 statistics obtains maximum principal curvatures distribution histogram, cuts apart brain ditch zone and gyrus zone by hidden Markov random field expectation maximization model:

May further comprise the steps:

A, according to maximum principal curvatures distribution histogram, utilize Otsu adaptive threshold method to obtain brain ditch initial on brain cortex surface zone and gyrus Region Segmentation X ⁽⁰⁾, X=(x wherein ₁..., x _n), n is total number of vertices on the brain cortex surface, x _i∈ 0, and 1} class mark, 0 expression brain ditch zone, initial gauss hybrid models parameter θ is estimated in 1 expression gyrus zone ⁽⁰⁾: described initial model parameter is average μ _l ⁽⁰⁾And variances sigma _l ⁽⁰⁾, wherein l represents brain ditch zone or gyrus zone;

B, utilize iterated conditional mould ICM, find the solution MRF-MAP,

$X^{\left(t\right)}=\underset{x}{\arg \max }\{\log P\left(Y\right|X,{\mathrm{\θ}}^{\left(t\right)})+\log P\left(X\right)\},$ Each summit is divided into brain ditch or gyrus, wherein Y=(y ₁..., y _n), y _iIt is the maximum principal curvatures of summit i;

C, basis adopt when forebrain ditch and gyrus segmentation result ${\mathrm{\μ}}_l^{(t+1)}=\frac{{\mathrm{\Σ}}_{i\∈S}{P}^{\left(t\right)}\left(l\right|y_i)y_i}{{\mathrm{\Σ}}_{i\∈S}{P}^{\left(t\right)}\left(l\right|y_i)}$ With ${\left({\mathrm{\σ}}_l^{(t+1)}\right)}^2=\frac{{\mathrm{\Σ}}_{i\∈S}{P}^{\left(t\right)}\left(l\right|y_i){(y_i-{\mathrm{\μ}}_l)}^2}{{\mathrm{\Σ}}_{i\∈S}{P}^{\left(t\right)}\left(l\right|y_i)},$ Upgrade gauss hybrid models parameter θ ^(t+1), wherein S represents summits all on the brain cortex surface, posteriority is distributed as $P^{\left(t\right)}\left(l\right|y_i)=\frac{g^{\left(t\right)}(y_i;{\mathrm{\θ}}_l)\·P^{\left(t\right)}\left(l\right|x_{N_i})}{p\left(y_i\right)},$

Be x _i=l and θ={ μ _l, σ _lMarkov local correlation probability, p (y _i) be prior probability, g (y _iθ _l) be Gaussian function: $g(y;{\mathrm{\θ}}_l)=\frac{1}{\sqrt{2\mathrm{\π}{\mathrm{\σ}}_l^2}}\exp (-\frac{{(y-{\mathrm{\μ}}_l)}^2}{2{\mathrm{\σ}}_l^2});$

D, repeating step b～c up to adjacent iteration result's variation less than total number of vertex purpose ten thousand/;

E, all gyrus being labeled as a kind of color, utilizing to be communicated with constituent analysis, is same color with the brain ditch zone marker of same connection, obtains gyrus zone and a series of brain ditches zone on the brain cortex surface.

Beneficial effect

Brain ditch region segmentation method on the brain cortex surface of the three-dimensional brain magnetic resonance image that the present invention proposes, at first, along with the precision of MR imaging apparatus improve constantly with the preprocess method of three-dimensional brain magnetic resonance image further ripe, obtain geometry accurately and the correct brain cortex surface of topological structure relatively easy; Simultaneously, maximum principal curvatures in brain ditch zone be negative value and in gyrus zone on the occasion of, utilize maximum principal curvatures to cut apart brain ditch zone and the gyrus zone is feasible.

The present invention has the following advantages with respect to other method: 1, utilize maximum principal curvatures can well distinguish brain ditch zone and gyrus zone, maximum principal curvatures can be distinguished brain ditch zone and buried gyrus zone in the brain ditch simultaneously; 2, utilize hidden Markov random field expectation maximization method to consider surperficial statistical information and the neighborhood information of going up the maximum principal curvatures on summit simultaneously.

Description of drawings

Fig. 1: the basic flow sheet of the inventive method.

Fig. 2: maximum principal curvatures distribution plan on 1 cerebral cortex left hemisphere inside surface.

Fig. 3: brain ditch Region Segmentation result on 12 true normal person's cerebral cortex left hemispheres.

Fig. 4: the segmentation result that has buried gyrus in 1 brain ditch.

Embodiment

Now in conjunction with the embodiments, accompanying drawing is further described the present invention:

The dividing method based on brain ditch zone on the brain cortex surface of maximum principal curvatures and hidden Markov random field expectation maximization model that proposes according to the present invention, we have realized a brain ditch Region Segmentation prototype system with C Plus Plus.The source of view data is: normal person's three-dimensional brain nuclear magnetic resonance image in the reality.

At first the three-dimensional brain nuclear magnetic resonance image is carried out pre-service and brain cortex surface reconstruction: comprise and remove skull and non-cerebral tissue, brain image is carried out brain tissue to be cut apart and (is divided into white matter, three types of grey matter and celiolymphs), the reconstruct geometric consequence is accurate the brain image after tissue segmentation, the brain cortex surface that topological structure is correct, this cortical surface is represented by a series of summits and triangle.Then, estimate the maximum principal curvatures on summit on the brain cortex surface.At last, will be divided into brain ditch zone and gyrus zone on the brain cortex surface according to maximum principal curvatures and hidden Markov random field expectation maximization framework, and utilize each brain ditch zone of connectivity analysis mark.

The method of estimation of the maximum principal curvatures on summit is on the described brain cortex surface: the Weingarten matrix that at first calculates each gore, the Weingarten matrix computations on each summit is the weighted mean of the gore Weingarten matrix that makes a circle in this summit week then, calculate the eigenwert and the proper vector of the Weingarten matrix on a summit at last, wherein the eigenwert of absolute value maximum is maximum principal curvatures.

Based on the brain ditch of hidden Markov random field expectation maximization framework and the principle of gyrus region segmentation method be: at first, according to maximum principal curvatures distribution histogram, obtain brain ditch and the initial segmentation result in gyrus zone and the parameter of mixed Gauss model on the brain cortex surface by Otsu adaptive threshold method, comprise average and variance; Then, estimate the class mark (brain ditch or gyrus) on each summit by hidden Markov random field expectation maximization model; Then, upgrade the parameter of mixed Gauss model according to current brain ditch and gyrus segmentation result; Carrying out second step and the 3rd by iteration goes on foot and obtains final brain ditch and gyrus segmentation result.At last, utilizing connectivity analysis that each is communicated with brain ditch zone marker is a value.

The whole flow process of the present invention can be with reference to the accompanying drawings 1, and concrete implementation step is as follows:

1. pre-service and brain cortex surface are rebuild:

The three-dimensional brain nuclear magnetic resonance image is removed skull and non-cerebral tissue, and cerebral tissue is cut apart and to rebuild geometric consequence accurate, the brain cortex surface that topological structure is correct.

2. maximum principal curvatures is estimated on the brain cortex surface:

At first calculate the Weingarten matrix of each gore, the gore Weingarten matrix that makes a circle in each summit week of weighted mean then calculates the Weingarten matrix on each summit, calculate the eigenwert and the proper vector of the Weingarten matrix on each summit at last, wherein the eigenwert of absolute value maximum is maximum principal curvatures.Accompanying drawing has shown the maximum principal curvatures that 1 cerebral cortex left hemisphere inside surface (interface of cerebral white matter and grey matter) is upward estimated.

3. based on the brain ditch and the gyrus Region Segmentation of hidden Markov random field expectation maximization framework:

Suppose at summit i, y _iBe maximum principal curvatures value, x _i∈ 0, and 1} class mark, wherein 0 represents brain ditch zone, 1 expression gyrus zone.Brain ditch zone and gyrus Region Segmentation problem can be formulated as maximization a posterior probability distribution P (X|Y), X=(x here ₁..., x _n), Y=(y ₁..., y _n), n is total number of vertices on the brain cortex surface of the trigonometric ratio rebuild.This segmentation problem can be formulated as again seeks real class mark $\hat{X}=({\hat{x}}_1,...,{\hat{x}}_n),$ It is satisfied:

$\hat{X}=\underset{X}{\arg \max }\left\{P\left(Y\right|X)P\left(X\right)\right\}$

This problem is specifically found the solution as follows:

1),, obtains the initial segmentation result X of brain ditch and gyrus zone on the brain cortex surface by the Otsu self-adaption thresholding method according to maximum principal curvatures distribution histogram ⁽⁰⁾, and estimate brain ditch and the gyrus zone parameter θ of Gauss model separately ⁽⁰⁾, comprise average and variance;

2) estimate the class mark on each summit by Markov random field model:

$X^{\left(t\right)}=\underset{x}{\arg \max }\{\log P\left(Y\right|X,{\mathrm{\θ}}^{\left(t\right)})+\log P\left(X\right)\}$

3), estimate brain ditch and the gyrus zone parameter θ of Gauss model separately according to current class mark ^(t+1):

${\mathrm{\μ}}_l^{(t+1)}=\frac{{\mathrm{\Σ}}_{i\∈S}{P}^{\left(t\right)}\left(l\right|y_i)y_i}{{\mathrm{\Σ}}_{i\∈S}{P}^{\left(t\right)}\left(l\right|y_i)},$ ${\left({\mathrm{\σ}}_l^{(t+1)}\right)}^2=\frac{{\mathrm{\Σ}}_{i\∈S}{P}^{\left(t\right)}\left(l\right|y_i){(y_i-{\mathrm{\μ}}_l)}^2}{{\mathrm{\Σ}}_{i\∈S}{P}^{\left(t\right)}\left(l\right|y_i)}$

Here S represents summits all on the brain cortex surface, and the posteriority Distribution calculation is:

$P^{\left(t\right)}\left(l\right|y_i)=\frac{g^{\left(t\right)}(y_i;{\mathrm{\θ}}_l)\·P^{\left(t\right)}\left(l\right|x_{N_i})}{p\left(y_i\right)}$

G (y _iθ _l) be a Gaussian function:

$g(y;{\mathrm{\θ}}_l)=\frac{1}{\sqrt{2\mathrm{\π}{\mathrm{\σ}}_l^2}}\exp (-\frac{{(y-{\mathrm{\μ}}_l)}^2}{2{\mathrm{\σ}}_l^2})$

4) carrying out the 2nd step and the 3rd by iteration goes on foot and to obtain final brain ditch and gyrus Region Segmentation result.

At last, utilizing connectivity analysis that each is communicated with brain ditch zone marker is a value.

In order to test the accuracy of this brain ditch region segmentation method, we are used for the brain cortex surface that 12 true normal persons' cerebral nucleus magnetic resonance image (MRI) is reconstructed with this method.Fig. 3 shows brain ditch Region Segmentation result on 12 true normal person's cerebral cortex left hemispheres, and wherein each brain ditch zone is denoted as a color.For this method of quantitative evaluation, we adopt over-segmentation and two kinds of measures of less divided.Over-segmentation is meant that comparing an original brain ditch Region Segmentation with expert's range estimation is a plurality of brain ditches zones.Less divided is meant not with adjacent a plurality of brain ditch suitable separating in zone.We utilize three main brain ditches, comprise central sulcus, and postcentral area and sulcus temporalis superior are verified.On above 12 normal person's left hemispheres, all central sulcus and sulcus temporalis superior are all correctly split, and therefore do not have over-segmentation and less divided mistake.There is a routine over-segmentation mistake in the postcentral area and do not have the less divided mistake.Find out from experimental result: our brain ditch region segmentation method has good performance.This dividing method can well handle gyrus buried in the brain ditch situation, accompanying drawing 4 has shown an example: the segmentation result of a buried example in the brain ditch of gyrus.(a) and (b) be that the maximum principal curvatures figure in zone and brain ditch Region Segmentation result's enlarged drawing are confined in red rectangular area among the figure (c).(c) be brain ditch Region Segmentation result on the whole brain cortex surface hemisphere, wherein the brain ditch zone of a connection of each color showing.

Claims (1)

1. brain ditch region segmentation method on the brain cortex surface of three-dimensional brain magnetic resonance image is characterized in that:

Step 1 pair three-dimensional brain magnetic resonance image carries out pre-service: utilize the changeability model method to remove skull, utilize method for registering to remove non-cerebral tissue, utilize the gauss hybrid models method that brain image is carried out tissue segmentation, obtain white matter, the image that three kinds of types of organizations of grey matter and celiolymph represent;

Step 2 brain cortex surface is rebuild: the brain cortex surface that utilizes Marching Cubes method reconstruct trigonometric ratio from the brain image after the tissue segmentation that above-mentioned steps obtains;

Step 3 utilizes finite difference method to estimate to obtain the maximum principal curvatures on each summit on the brain cortex surface:

Step 4 statistics obtains maximum principal curvatures distribution histogram, cuts apart brain ditch zone and gyrus zone by hidden Markov random field expectation maximization model:

May further comprise the steps:

A, according to maximum principal curvatures distribution histogram, utilize Otsu adaptive threshold method to obtain brain ditch initial on brain cortex surface zone and gyrus Region Segmentation X ⁽⁰⁾, X=(x wherein ₁..., x _n), n is total number of vertices on the brain cortex surface, x _i∈ 0, and 1} class mark, 0 expression brain ditch zone, initial gauss hybrid models parameter θ is estimated in 1 expression gyrus zone ⁽⁰⁾: described initial model parameter is average μ _l ⁰⁾And variances sigma _l ⁽⁰⁾, wherein l represents brain ditch zone or gyrus zone;

B, utilize iterated conditional mould ICM, find the solution MRF-MAP,

$X^{\left(t\right)}=\underset{x}{\arg \max }\{\log P\left(Y\right|X,{\mathrm{\θ}}^{\left(i\right)})+\log P\left(X\right)\},$ Each summit is divided into brain ditch or gyrus, wherein Y=(y ₁..., y _n), y _iIt is the maximum principal curvatures of summit i;

C, basis adopt when forebrain ditch and gyrus segmentation result ${\mathrm{\μ}}_l^{(t+1)}=\frac{{\mathrm{\Σ}}_{i\∈S}{P}^{\left(t\right)}\left(l\right|y_i)y_i}{{\mathrm{\Σ}}_{i\∈S}{P}^{\left(t\right)}\left(l\right|y_i)}$ With ${\left({\mathrm{\σ}}_l^{(t+1)}\right)}^2=\frac{{\mathrm{\Σ}}_{i\∈S}{P}^{\left(t\right)}{\left(l\right|y_i)(y_i-{\mathrm{\μ}}_l)}^2}{{\mathrm{\Σ}}_{i\∈S}{P}^{\left(t\right)}\left(l\right|y_i)},$ Upgrade gauss hybrid models parameter θ ^(t+1), wherein S represents summits all on the brain cortex surface, posteriority is distributed as $P^{\left(t\right)}\left(l\right|y_i)=\frac{g^{\left(t\right)}(y_i;{\mathrm{\θ}}_l)\·P^{\left(t\right)}\left(l\right|x_{N_l})}{p\left(y_i\right)},$

Be x _i=l and θ={ μ _l, σ _lMarkov local correlation probability, p (y _i) be prior probability, g (y _iθ _l) be Gaussian function:

$g(y;{\mathrm{\θ}}_l)=\frac{1}{\sqrt{2\mathrm{\π}{\mathrm{\σ}}_l^2}}\exp (-\frac{{(y-{\mathrm{\μ}}_l)}^2}{2{\mathrm{\σ}}_l^2});$

D, repeating step b～c up to adjacent iteration result's variation less than total number of vertex purpose ten thousand/;

E, all gyrus being labeled as a kind of color, utilizing to be communicated with constituent analysis, is same color with the brain ditch zone marker of same connection, obtains gyrus zone and a series of brain ditches zone on the brain cortex surface.

Images