CN103544712A - Method for automatically segmenting human lateral geniculate nucleus through prior knowledge - Google Patents

Abstract

The invention discloses a method for automatically segmenting the human lateral geniculate nucleus through prior knowledge. According to the method, field bias correction is performed on structural image data of the human brain, and then minor structures of the brain are segmented through the corrected image so as to obtain a template in the ventral diencephalon region. Next, the corrected structural image data and the template of the ventral diencephalon region are in registration into MNI standard space. The region limit in the MNI space of the LGN is obtained according to the prior knowledge of the anatomic structure of the lateral geniculate nucleus, then in the region, the LGN is segmented out though the region growing method, the k average value method, the Ostu method and the image segmenting method. Then, several segmenting results are fused to obtain an estimated value of the LGN region to serve as a segmenting result. Finally, the segmenting result is changed into original space of a structural image, and it is that the final segmenting result of the LGN is obtained.

Description

A kind of mankind's corpus geniculatum lateral automatic division method that utilizes priori

Technical field

The present invention relates to Computer Image Processing field, particularly a kind of dissection priori auto Segmentation in human brain structure picture that utilizes corpus geniculatum lateral (LGN) goes out the medical image cutting method of LGN.

Background technology

Image is cut apart and is referred to and utilize image information that " interested target object " in image extracted from complicated scene.Interested target object is commonly referred to as prospect, and in image, remainder is called as background.Generally speaking, image partition method can be divided into two kinds of systems: man-machine interactive is cut apart and computing machine full-automatic dividing.The former too relies on expertise and consuming time oversize, and the result that the latter is independently cut apart often can not can not be satisfactory.In practical application area, people select the method adapting to for different demands, up to the present also do not have a kind of general dividing method.

Corpus geniculatum lateral is the terminating point that leading portion is looked road afferent neurofibers, intersect with uncrossed nerve fibre finally after this place changes neuron to rear formation optic radiation.As the 4th neuron (first three is respectively intraretinal centrum cell, rod cells, intraretinal Beale's ganglion cells and intraretinal ganglion cell) on Yi Geshi road, corpus geniculatum lateral is being played the part of important role in pathways for vision.But because the LGN as a thalamus part is by temporal lobe with hippocampal gyrus blocks and LGN own vol is very little, LGN is difficult to be observed.By dissecting human brain, it is found that LGN is inevitable in veutro diencephalon district (VDC).The people such as Andrews in 1997 find that LGN is roughly in the cube region of a 5cm*5cm*5cm, and the people such as Kastner in 2004 find the position of LGN central point in Talairach space probably in left side (23.33 ,-21 ,-4.66) and right side (22.88,-21.3,4.63) near.

It is all that image is cut apart algorithm conventional in field that region growing method, k averaging method, Otsu algorithm and figure cut algorithm.The half-tone information that region growing method is utilized image approaches position and the similar pixel of gray scale is gathered into a region and obtains segmentation result.K averaging method is a kind of typical clustering algorithm, and it utilizes characteristics of image the voxel of image to be polymerized to the classification number of appointment.Otsu algorithm is that image is cut apart the optimal algorithm that middle threshold value is chosen, thereby the gray-scale value of its traversing graph picture is found, an optimal threshold is divided into background by image and prospect two parts make the inter-class variance between background and prospect maximum.Figure cuts algorithm and graph theory is introduced to image cuts apart field, the voxel of image is as the node of figure, correlativity between voxel is as the weight on connected node limit, and the mode of finding a cutting edge is divided into two parts by node, and the mode of cutting off-energy minimum is optimum Cutting mode.

In field of medical images, artificial segmentation object object is very consuming time, but automatic division method accuracy is not high, in order to solve the problem of low accuracy, we utilize the segmentation result of some different modes to calculate the probability estimate value of target object true segmentation, the segmentation result using this estimated value as target object.

Summary of the invention

(1) technical matters that will solve

Technical matters to be solved by this invention is to provide a kind of automatic LGN dividing method, with the LGN in human brain structure picture, carries out accurately, cuts apart automatically.

(2) technical scheme

For solving above-mentioned technical matters, the present invention proposes a kind of mankind's corpus geniculatum lateral automatic division method that utilizes priori, for the LGN of human brain structure picture is split.The method comprises the steps: step S1: the structure of human brain is carried out to nonuniform field correction as data; Step S2: corrected structure is looked like to be registrated to MNI normed space; Step S3: utilize the priori of LGN anatomical structure to obtain the approximate region of LGN in MNI space; Step S4: cut apart by region growing method, k averaging method, Otsu algorithm and the figure method of cutting respectively in this region; Step S5: four kinds of segmentation results are merged and obtain the estimation to the true segmentation of LGN; Step S6: the result estimating is transformed to the luv space of structure picture, obtain the final segmentation result of LGN.

(3) beneficial effect

The present invention utilizes the priori of LGN anatomical structure to make LGN priori template, and cut zone is limited in this template.This measure incorporates anatomical structure priori during LGN cuts apart, and has both avoided problem consuming time that man-machine interaction cuts apart problem, has also improved the accuracy of traditional auto Segmentation.The present invention is also merged several dividing methods to obtain the estimated value of LGN true segmentation, further improves the accuracy of auto Segmentation.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the mankind's corpus geniculatum lateral automatic division method based on priori of the present invention;

Fig. 2 is the variation diagram of image in LGN cutting procedure, and wherein (a) is the original MR structure picture of human brain; (b) be the MR structure picture of proofreading and correct through nonuniform field; (c) be the MR structure picture that is registrated to MNI normed space; (d) be the priori template of LGN under MNI normed space; (e) be to merge four kinds of estimated values that LGN is cut apart that dividing method obtains; (f) be the final segmentation result of LGN under luv space;

Fig. 3 is the 3-D display figure of LGN segmentation result.

Embodiment

For making the object, technical solutions and advantages of the present invention clearer, below in conjunction with specific embodiment, and with reference to accompanying drawing, the present invention is described in more detail.

Core concept of the present invention is that the priori of LGN anatomical structure is incorporated in the auto Segmentation of LGN, thereby has solved the problem consuming time of man-machine interaction and the accuracy of raising auto Segmentation.The true segmentation that several dividing methods are estimated LGN is also merged in the present invention, thereby further improves the accuracy of auto Segmentation.For making the object, technical solutions and advantages of the present invention clearer, below in conjunction with specific embodiment, and with reference to accompanying drawing, the present invention is described in further detail.Fig. 1 has shown the process flow diagram of corpus geniculatum lateral automatic division method of the present invention.As shown in Figure 1, method of the present invention comprises the steps:

Step S1: the structure of human brain is carried out to nonuniform field correction as data.

Utilizing structure that nuclear magnetic resonance technique obtains human brain as data, then utilize N3(nonparametric nonuniform intensity normalization, the inhomogeneous gray correction of nonparametric) algorithm proofreaies and correct nonuniform field.Concrete grammar is referring to Sled, J.G., A.P.Zijdenbos, et al. (1998). " A nonparametric method for automatic correction of intensity nonuniformity in MRI data. " IEEE Trans Med Imaging17 (1): 87-97.

Step S2: corrected structure is looked like to be registrated to MNI normed space.

Image registration is referred to image registration on MNI template to MNI normed space.MNI template is made by Montreal neurology research institution (Montreal Neurological Institute).This mechanism utilizes 305 normal persons' brain image to obtain last average template by modes such as manual markings key point and linear registrations, and this template is called as MNI template.Concrete MNI template construct method is referring to Evans, A.C., D.L.Collins, et al. (1993). " 3d Statistical Neuroanatomical Models from305Mri Volumes. " Nuclear Science Symposium & Medical Imaging Conference, Vols1-3:1813-1817.

First utilize the affined transformation of 12 parameters that the image registration after proofreading and correct is arrived to MNI template, this conversion process has solved the Main Differences problem of image and MNI template brain size and position.Then by non-linear registration method, process some small scale structures differences of image and MNI template.Concrete grammar is referring to Friston, K.J., J.Ashburner, et al. (1995). " Spatial registration and normalization of images. " Hum Brain Mapp3 (3): 165-189.

Step S3: utilize the priori of LGN anatomical structure to obtain the approximate region of LGN in MNI space.

In traditional medical science dividing method, the accuracy of full-automatic dividing is lower, man-machine interaction is cut apart because relying on expertise too consuming time, the present invention is in order to solve this two problems, the priori that is LGN anatomical structure by expertise changes into the manipulable step of computing machine, thereby makes the LGN priori template under MNI space.Concrete operations are as follows:

Step S3a:LGN is inevitable in veutro diencephalon district (VDC), so thereby the image through overcorrect is carried out obtaining VDC template cutting apart of brain minor structure, and VDC template is registrated to MNI normed space.

The brain map that the present invention utilizes the people such as Bruce Fischl to make is divided into 45 minor structures by brain, then chooses VDC region, left and right and makes VDC template.Then the method for utilizing step S2 to mention, is registrated to MNI normed space by VDC template.Utilize method that brain map cuts apart brain referring to Fischl, B., D.H.Salat, et al. (2002). " Whole brain segmentation:automated labeling of neuroanatomical structures in the human brain. " Neuron33 (3): 341-355.

Near step S3b: the center position of LGN is probably in left side (23.33 ,-21 ,-4.66) and right side (22.88 ,-21.3,4.63) in Talairach space.It is left side P that these two coordinate conversion are become to MNI volume coordinate _l(23 ,-22 ,-7) and right side P _r(26 ,-22 ,-8);

Step S3c: left and right sides LGN roughly, in the cubic space of a 5cm*5cm*5cm, considers the error that conversion brings, and builds and thinks P respectively _land P _rcentered by, the cube template that 7cm is the length of side;

Step S3d: VDC template and cube template are sought common ground, build LGN priori template.

Step S4: cut apart by region growing method, k averaging method, Otsu algorithm and the figure method of cutting respectively in this region.

Step S4a: in region growing method, with P _land P _rfor starting point, radius is that 5cm carries out region growing, and region growing result and LGN priori template are sought common ground and obtain segmentation result.

I voxel P in LGN priori template _i(i=1,2 ..., gray-scale value N) is expressed as I _i(i=1,2 ..., N), its coordinate vector is expressed as V _i, the constraint condition of region growing is candidate point and starting point P _sdistance be no more than radius R ₀, and the difference of gray-scale value is no more than I ₀.As follows:

$\mathrm{Region}={\∪}_{i=1}^N{P}_i,(i=\mathrm{1,2},...,N)$

if||V _i-V _s||≤R ₀，|I _i-I _s|≤I ₀

Here Region is the result of region growing method, and N is the number of voxel in LGN priori template, V _sand I _scoordinate vector and the gray-scale value of starting point.R in the present invention ₀get 5cm, I ₀be 30.

Step S4b: in k averaging method, voxel feature consists of gray-scale value and its coordinate in MNI space of voxel, utilizes k averaging method and voxel feature that the voxel in LGN priori template is polymerized to two classes, and the class that voxel number is many is segmentation result.

Step S4c: in Otsu algorithm, thereby the gray-scale value in traversal LGN priori template find the optimal threshold of cutting apart LGN and make to cut apart that to obtain the inter-class variance of result maximum, the class that voxel number is many is segmentation result.

Otsu algorithm is that image is cut apart the optimal algorithm that middle threshold value is chosen.It is the gamma characteristic by image, and image is divided into background and prospect two parts.Inter-class variance between background and prospect is larger, illustrates that the two-part difference of composing images is larger, when part prospect mistake is divided into background or part background mistake, is divided into prospect and all can causes two parts difference to diminish.Therefore, make to mean cutting apart of inter-class variance maximum that misclassification probability is minimum.

If the gray level of image is L in LGN priori template localized area, tonal range is [0, L-1], if threshold value is t, the inter-class variance of prospect and background is defined as:

f(t)=w ₀(t)*[u ₀(t)-u] ²+w ₁(t)*[u ₁(t)-u] ²,t∈[0,L-1]

W wherein ₀for background proportion, u ₀for the gray average of background, w ₁for prospect proportion, u ₁for the gray average of prospect, u is the average of all voxels in LGN priori template localized area.Travel through all threshold value t, make the t of above transition formula evaluation maximum, be the optimal threshold of cutting apart image.Utilize optimal threshold that image is divided into two classes, the class that voxel number is many is segmentation result.

Step S4d: cut in algorithm at figure, voxel in LGN priori template is as the node of figure, utilize voxel coordinate information and voxel intensity information to calculate the similarity between voxel, the weight on the limit using similarity between node, utilize figure to cut algorithm figure is divided into two parts, make to cut apart rear Minimal energy loss, the class that voxel number is many is segmentation result.

The present invention defines similarity between voxel i and voxel j

$\mathrm{similarity}(i,j)=w_c\·\exp (-\frac{{\left|\right|f_{\mathrm{ci}}-f_{\mathrm{cj}}\left|\right|}^2}{{2\mathrm{\σ}}_c^2})+w_g\·\exp (-\frac{{\left|\right|f_{\mathrm{gi}}-f_{\mathrm{gj}}\left|\right|}^2}{{2\mathrm{\σ}}_g^2})$

F wherein _ci=[x _i, y _i, z _i] and f _cj=[x _j, y _j, z _j] be respectively the feature that voxel i and voxel j coordinate information form, f _gi=[grayvalue _i, 6-neighborvalue _i, meanvalue _i, medvalue _i, stdvalue _i] and f _gj=[grayvalue _j, 6-neighborvalue _j, meanvalue _j, medvalue _j, stdvalue _j] be respectively near the feature that voxel i and voxel j, half-tone information forms, grayvalue is the gray-scale value that this voxel is corresponding, 6-neighborvalue is this voxel gray-scale value of six voxels in front and back up and down, meanvalue is this voxel and the average gray of six voxels around, medvalue is this voxel and the gray scale intermediate value of six voxels around, and stdvalue is this voxel and the gray standard deviation of six voxels around.W _c, w _grespectively the weight that coordinate feature and gray feature are corresponding, σ _c, σ _grespectively that the core that coordinate feature and gray feature are corresponding is wide;

Definition energy loss function E is

$E=\frac{\mathrm{cut}(A,B)}{\mathrm{assov}(A,V)}+\frac{\mathrm{cut}(A,B)}{\mathrm{assov}(B,V)}$

Wherein V is original connected graph, and A and B are the two parts that obtain after figure cuts, and cut (A, B) is the weight sum that connects all limits of node in A and B in figure V, and assov (A, V) is the weight sum that has all limits of a node in A.

Segmentation result vector is designated as to S=[s ₁, s ₂..., s _n] ^t, each component s _irepresent the segmentation result of each node in V figure: if s _ivalue is 1, and corresponding node belongs to a subgraph, if s _ivalue is corresponding node belongs to another subgraph.The similarity matrix of V figure is designated as

, use

statement degree matrix (each diagonal entry is respective nodes weight sum, and off diagonal element is 0),

and energy loss function E converts to

In order to calculate the binary set S that makes E minimum, first calculate the real number vector S that makes E minimum, then by being set, threshold value determines which limit each node belongs to.Through conversion, can know that the minimum value of E should be matrix

the second little eigenwert (because minimal eigenvalue is 0, and E can not be 0), obtain characteristic of correspondence vector and be real number vector S.Then travel through value in S as threshold value, by S change into only have 1 and-binary set of b, calculate corresponding loss function E, therefrom select a threshold value to make E minimum, now corresponding S is divided into image two classes, and the class that voxel number is many is the segmentation result that figure cuts algorithm.

Step S5: merge estimated value that four kinds of dividing methods obtain LGN region as segmentation result.

In order further to mention the accuracy of automatic segmentation algorithm, the segmentation result of some different modes is merged to estimate the true segmentation of LGN here.

Step S5a: the likelihood estimation function that structure LGN is cut apart.

Step S4 independently gives class sign Label by all voxels in LGN priori template region (N) with four kinds of partitioning algorithms, Label ∈ 0,1}, and for certain voxel i (i=1,2 ..., N), T (i) is its true class sign, note D _m(i) be the class sign that m kind partitioning algorithm is given i voxel, four kinds of segmentation results of N voxel in priori template region formed to the decision matrix D of N * 4 dimension, note p=[p ₁, p ₂, p ₃, p ₄] ^tand q=[q ₁, q ₂, q ₃, q ₄] ^t, p _jand q _jrepresent that j (j=1,2,3,4) plants susceptibility and the sensitivity of partitioning algorithm, complete data and its probability distribution function of using respectively (D, T) and f (D, T|p, q) to come presentation video to cut apart, likelihood estimation function is ln[f (D, T|p, q)].

Step S5b: utilize EM algorithm to maximize and estimate to obtain the estimated value that LGN is cut apart.

In EM algorithm, the E step computing formula of the k time iteration is

${W_i}^{\left(k\right)}=\frac{a_i^{\left(k\right)}}{a_i^{\left(k\right)}+b_i^{\left(k\right)}}$

This formula is used for estimating that i voxel is marked as 1 probability W _i ^(k), wherein

$a_i^{\left(k\right)}=f(T_i=1)\underset{j:D_{\mathrm{ij}}=1}{\mathrm{\Π}}{p}_j^{\left(k\right)}\underset{j:D_{\mathrm{ij}}=0}{\mathrm{\Π}}(1-p_j^{\left(k\right)}),$

$b_i^{\left(k\right)}=f(T_i=0)\underset{j:D_{\mathrm{ij}}=0}{\mathrm{\Π}}{q}_j^{\left(k\right)}\underset{j:D_{\mathrm{ij}}=1}{\mathrm{\Π}}(1-q_j^{\left(k\right)}),$

Here f (T _i=1) and f (T _i=0) represent respectively i (i=1,2 ..., N) the true sign of individual voxel is 1 and 0 probability, D _ij(i=1,2 ..., N; J=1,2,3,4) be the element in the dimension decision matrix D of N * 4, k is iterations, loop iteration, until W _i ^(k), p ^(k)and q ^(k)convergence.

In EM algorithm, the M step computing formula of the k time iteration is

$p_j^{(k+1)}=\frac{{\mathrm{\Σ}}_{i:D_{\mathrm{ij}}=1}{W_i}^{\left(k\right)}}{{\mathrm{\Σ}}_i{W_i}^{\left(k\right)}},q_j^{(k+1)}=\frac{{\mathrm{\Σ}}_{i:D_{\mathrm{ij}}=0}(1-{W_i}^{\left(k\right)})}{{\mathrm{\Σ}}_i(1-{W_i}^{\left(k\right)})}$

These two formulas are used for estimating the susceptibility p using in the k+1 time iteration ^(k+1)with sensitivity q ^(k+1), W wherein _i ^(k)that i voxel is marked as 1 probability, D _ij(i=1,2 ..., N; J=1,2,3,4) be the element in the dimension decision matrix D of N * 4, k is iterations, loop iteration, until W _i ^(k), p ^(k)and q ^(k)convergence.

In this step first by W _i ^(k), p ^(k)and q ^(k)initialization, then utilizes the formula of E step and M step to calculate, and loop iteration, until convergence.Finally utilize threshold value 0.5 by W _i ^(k)binaryzation, if W _i ^(k)>0.5 i voxel in LGN region, otherwise i voxel be not in LGN region.

More details are referring to Warfield, S.K., K.H.Zou, et al. (2004). " Simultaneous truth and performance level estimation (STAPLE): an algorithm for the validation of image segmentation. " IEEE Trans Med Imaging23 (7): 903-921.

Step S6: the segmentation result estimating is transformed to the luv space of structure picture, obtain the final segmentation result of LGN.

Because the present invention is in order to utilize the dissection priori in MNI space, to the cutting procedure of LGN, be all to carry out under MNI space, finally to utilize the information converting that step S2 obtains the LGN segmentation result contravariant under MNI space to be changed to the luv space of structure picture, complete cutting apart of whole LGN.

Below by one, specifically should be used for further illustrating method of the present invention.At one, dispose Intel Xeon E5-2687W eight processors of core 3.10GHz, the internal memory of 64Gb, operating system are to use FreeSurfer and MATLAB software to complete on the workstation of Linux CentOS6.3.

A normal person's brain structure is looked like to carry out LGN and cut apart, the original image of brain structure picture as shown in Figure 2 (a) shows.

Step S1: utilize N3 algorithm to carry out nonuniform field correction to the structure of human brain as data, after proofreading and correct, image is as shown in Fig. 2 (b).

Step S2: utilize the radiation conversion of 12 parameters and non-linear registration method that the image registration after proofreading and correct is arrived to MNI normed space, images after registration as shown in Figure 2 (c).

Step S3: prototype structure is looked like to be partitioned into VDC region and VDC template is registrated to MNI space.Then utilize LGN center position structure cube template, VDC template and cube template are sought common ground and obtain the priori template of LGN, as shown in Figure 2 (d) shows.

Step S4: use respectively region growing method, k averaging method (kmeans), Otsu algorithm (otsu) and figure to cut method (Ncut) in the priori template of LGN and cut apart.

Step S5: merge in four kinds of segmentation results and obtain the estimation to the true segmentation of LGN, this estimated value is the segmentation result of LGN under MNI space.Fig. 2 (e) merges four kinds of estimated values that LGN is cut apart that dividing method obtains.

Step S6: the LGN segmentation result estimating under MNI space is transformed to the luv space of structure picture, obtain the final segmentation result of LGN, as shown in Fig. 2 (f) and figure (3).

By above-mentioned for the specific embodiment of the present invention, and the explanation that is aided with embodiment is visible, the present invention utilizes the priori of LGN anatomical structure to make LGN priori template, priori incorporated during LGN cuts apart, solved the accuracy that man-machine interaction is cut apart the problem consuming time of existence and improved traditional auto Segmentation.Then multiple dividing method is merged to the estimated value that obtains LGN true segmentation, further improve the accuracy of auto Segmentation.

Above-described specific embodiment; object of the present invention, technical scheme and beneficial effect are further described; be understood that; the foregoing is only specific embodiments of the invention; be not limited to the present invention; within the spirit and principles in the present invention all, any modification of making, be equal to replacement, improvement etc., within all should being included in protection scope of the present invention.

Claims (7)

1. the mankind's corpus geniculatum lateral LGN automatic division method that utilizes priori, the method comprises the steps:

Step S1: the structure of human brain is carried out to nonuniform field correction as data;

Step S2: corrected structure is looked like to be registrated to MNI normed space;

Step S3: utilize the priori of LGN anatomical structure to obtain the approximate region of LGN in MNI space;

Step S4: cut apart by region growing method, k averaging method, Otsu algorithm and the figure method of cutting respectively in described region;

Step S5: the estimated value that several partitioning algorithms fusions are obtained to LGN region is as segmentation result;

Step S6: the LGN segmentation result of estimation is transformed to the luv space of structure picture, obtain the final segmentation result of LGN.

2. the method for claim 1, is characterized in that, the priori of utilizing LGN anatomical structure in described step S3 obtains the approximate region of LGN in MNI space and comprises:

Step S3a:LGN is inevitable in veutro diencephalon district VDC, so thereby the image through overcorrect is carried out obtaining VDC template cutting apart of brain minor structure, and VDC template is registrated to MNI normed space;

Near step S3b: the center position of LGN is in left side (23.33 ,-21 ,-4.66) and right side (22.88 ,-21.3,4.63) in Talairach space, and it is left side P that these two coordinate conversion are become to MNI volume coordinate _l(23 ,-22 ,-7) and right side P _r(26 ,-22 ,-8);

Step S3c: left and right sides LGN, in the cubic space of a 5cm*5cm*5cm, builds and thinks P respectively _land P _rcentered by, the cube template that 7cm is the length of side;

Step S3d: VDC template and cube template are sought common ground, build LGN priori template.

3. the method for claim 1, is characterized in that, the step in described step S4 comprises:

Step S4a: in region growing method, with P _land P _rfor starting point, radius is that 5cm carries out region growing, and region growing result and LGN priori template are sought common ground and obtain segmentation result;

Step S4b: in k averaging method, voxel feature consists of gray-scale value and its coordinate in MNI space of voxel, utilizes k averaging method and voxel feature that the voxel in LGN priori template is polymerized to two classes, and the class that voxel number is many is segmentation result;

Step S4c: in Otsu algorithm, thereby the gray-scale value in traversal LGN priori template find the optimal threshold of cutting apart LGN and make the inter-class variance of two classes maximum, the class that voxel number is many is segmentation result;

Step S4d: cut in algorithm at figure, voxel in LGN priori template is as the node of figure, utilize voxel coordinate information and voxel intensity information to calculate the similarity between voxel, the weight on the limit using similarity between node, utilize figure to cut algorithm figure is divided into two parts, make to cut apart rear Minimal energy loss, the class that voxel number is many is segmentation result.

4. method as claimed in claim 3, is characterized in that, in step S4d, between voxel i and voxel j, the definition of similarity is

$\mathrm{similarity}(i,j)=w_c\·\exp (-\frac{{\left|\right|f_{\mathrm{ci}}-f_{\mathrm{cj}}\left|\right|}^2}{{2\mathrm{\σ}}_c^2})+w_g\·\exp (-\frac{{\left|\right|f_{\mathrm{gi}}-f_{\mathrm{gj}}\left|\right|}^2}{{2\mathrm{\σ}}_g^2})$

F wherein _ci=[x _i, y _i, z _i] and f _cj=[x _j, y _j, z _j] be respectively the feature that voxel i and voxel j coordinate information form, f _gi=[grayvalue _i, 6-neighborvalue _i, meanvalue _i, medvalue _i, stdvalue _i] and f _gj=[grayvalue _j, 6-neighborvalue _j, meanvalue _j, medvalue _j, stdvalue _j] be respectively near the feature that voxel i and voxel j, half-tone information forms, grayvalue is the gray-scale value that this voxel is corresponding, 6-neighborvalue is this voxel gray-scale value of six voxels in front and back up and down, meanvalue is this voxel and the average gray of six voxels around, medvalue is this voxel and the gray scale intermediate value of six voxels around, stdvalue is this voxel and the gray standard deviation of six voxels around, w _c, w _grespectively the weight that coordinate feature and gray feature are corresponding, σ _c, σ _grespectively that the core that coordinate feature and gray feature are corresponding is wide;

In step S4d, the definition of energy loss function E is

$E=\frac{\mathrm{cut}(A,B)}{\mathrm{assov}(A,V)}+\frac{\mathrm{cut}(A,B)}{\mathrm{assov}(B,V)}$

Wherein V is original connected graph, and A and B are the two parts that obtain after figure cuts, and cut (A, B) is the weight sum that connects all limits of node in A and B in figure V, and assov (A, V) is the weight sum that has all limits of a node in A.

5. the method for claim 1, is characterized in that, in described step S5, four kinds of partitioning algorithms is merged to estimate the true segmentation of LGN, and step comprises:

Step S5a: the likelihood estimation function that structure LGN is cut apart;

Step S5b: utilize EM algorithm to maximize and estimate to obtain the estimated value that LGN is cut apart.

6. method as described in claim 5, is characterized in that, constructs the likelihood estimation function process that LGN cuts apart to be in step S5a:

Step S4 independently gives class sign Label by all voxels in LGN priori template region (N) with four kinds of partitioning algorithms, Label ∈ 0,1}, and for certain voxel i (i=1,2 ..., N), T (i) is its true class sign, note D _m(i) be the class sign that m kind partitioning algorithm is given i voxel, four kinds of segmentation results of N voxel in priori template region formed to the decision matrix D of N * 4 dimension, note p=[p ₁, p ₂, p ₃, p ₄] ^tand q=[q ₁, q ₂, q ₃, q ₄] ^t, p _jand q _jrepresent that j (j=1,2,3,4) plants susceptibility and the sensitivity of partitioning algorithm, complete data and its probability distribution function of using respectively (D, T) and f (D, T|p, q) to come presentation video to cut apart, likelihood estimation function is ln[f (D, T|p, q)].

7. method as claimed in claim 5, is characterized in that, utilizes the E step computing formula of the k time iteration in EM algorithm to be in step S5b

${W_i}^{\left(k\right)}=\frac{a_i^{\left(k\right)}}{a_i^{\left(k\right)}+b_i^{\left(k\right)}}$

This formula is used for estimating that i voxel is marked as 1 probability W _i ^(k), wherein

$a_i^{\left(k\right)}=f(T_i=1)\underset{j:D_{\mathrm{ij}}=1}{\mathrm{\Π}}{p}_j^{\left(k\right)}\underset{j:D_{\mathrm{ij}}=0}{\mathrm{\Π}}(1-p_j^{\left(k\right)}),$

$b_i^{\left(k\right)}=f(T_i=0)\underset{j:D_{\mathrm{ij}}=0}{\mathrm{\Π}}{q}_j^{\left(k\right)}\underset{j:D_{\mathrm{ij}}=1}{\mathrm{\Π}}(1-q_j^{\left(k\right)}),$

Here f (T _i=1) and f (T _i=0) represent respectively i (i=1,2 ..., N) the true sign of individual voxel is 1 and 0 probability, D _ij(i=1,2 ..., N; J=1,2,3,4) be the element in the dimension decision matrix D of N * 4, k is iterations, loop iteration, until W _i ^(k), p ^(k)and q ^(k)convergence,

In step S5b, utilize the M step computing formula of the k time iteration in EM algorithm to be

$p_j^{(k+1)}=\frac{{\mathrm{\Σ}}_{i:D_{\mathrm{ij}}=1}{W_i}^{\left(k\right)}}{{\mathrm{\Σ}}_i{W_i}^{\left(k\right)}},q_j^{(k+1)}=\frac{{\mathrm{\Σ}}_{i:D_{\mathrm{ij}}=0}(1-{W_i}^{\left(k\right)})}{{\mathrm{\Σ}}_i(1-{W_i}^{\left(k\right)})}$

These two formulas are used for estimating the susceptibility p using in the k+1 time iteration ^(k+1)with sensitivity q ^(k+1), W wherein _i ^(k)that i voxel is marked as 1 probability, D _ij(i=1,2 ..., N; J=1,2,3,4) be the element in the dimension decision matrix D of N * 4, k is iterations, loop iteration, until W _i ^(k), p ^(k)and q ^(k)convergence,

In step S5b first by W _i ^(k), p ^(k)and q ^(k)initialization, then utilizes the formula of E step and M step to calculate, and loop iteration, until convergence finally utilizes threshold value 0.5 by W _i ^(k)binaryzation, if W _i ^(k)>0.5 i voxel in LGN region, otherwise i voxel be not in LGN region.

Images