CN101706561A - Clustering method for functional magnetic resonance images - Google Patents

Abstract

The invention relates to a clustering method for functional magnetic resonance images, belonging to biological information technologies. The method comprises the following steps of pre-processing an original fMRI image, clustering an affine in regions, acquiring a new image and determining a corresponding bias parameter, acquiring clustering results corresponding to all bias parameters, determining an optimal clustering result and acquiring a finally clustered fMRI image. Because the hierarchical clustering, the affine clustering and the self-adapting affine clustering are organically combined, and the images are comprehensively processed, the invention overcomes the defects that the conventional clustering methods are affected by the subjective factor of people, heavily dependent on the personal experience and cannot effectively carry out the clustering processing on the images with large data quantity, and effectively solves the problem of objectively selecting the optimal clustering result. Therefore, the invention has the characteristics of having strong ability for processing the images with large data quantity and effectively improving the objectivity in the fMRI image processing procedure as well as the accuracy of selecting the optimal clustering result.

Description

The clustering method of functional magnetic resonance images

Technical field

The invention belongs to pattern, field of image recognition in the biology information technology, particularly a kind of Functional MRI (functional magnetic resonance imaging, fMRI) post processing of image technology.

Background technology

At present, the cerebral function imaging technology has obtained widespread use, (the functionalmagnetic resonance imaging of Functional MRI wherein, fMRI) be at magnetic resonance imaging (magnetic resonance imaging, MRI) grow up on the basis of technology, by the functional mri technology many physiology and biophysical parameters are measured, it is the no wound means of carrying out cerebration detection and imaging.Its formation method mainly comprises model-driven (Model-driven) analytical approach and data-driven (Data-driven) analytical approach two classes.Wherein:

Model-driven (Model-driven) analytical approach: this method will rely on experiment model (type) when handling the fMRI image, now usefulness many is that people such as Friston is based on generalized linear model (General Linear Model, GLM) (Statistical Parametric Mapping SPM) analyzes the fMRI image to the statistical parameter figure of Ti Chuing.These class methods are fairly simple comparatively speaking, directly perceived; But owing to need one to comprise the priori pattern relevant the fMRI image is analyzed, and in processing procedure, also to depend on statistical inference and obtain last result with factors such as experiment model, other artificial heartbeat that increases, breathing, head move; Thereby have that disturbing factor is many, noisiness is big, defective such as its net result depends on the accuracy of statistical inference, and the human factor influence is big.

Data-driven (Data-driven) analytical approach: this method mainly comprises independent component analytical approach and clustering method, wherein, people such as McKeown is at 1998 independent component analytical approach (the Independent ComponentAnalysis that propose, ICA), this method is the component that multidimensional data is decomposed into several independent, in the fMRI data analysis, can the task of experiment not carried out isolating all kinds of independently compositions under any hypothesis prerequisite, thereby these class methods are more and more being paid attention to aspect fMRI graphical analysis and the processing; But how composition being carried out objective selection in the independent component analysis, still is a difficult problem that does not have solution at present, and this method also rests on the basis of artificial judgement at present to the selection of composition, thereby still existence is subjected to people's subjective factor to influence big defective.

Clustering method comprises: first K-mean cluster (K-means) analytical approach, be that each classification all adopts the mean value of such all data to represent, though it can embody the meaning of cluster on geometry and statistics well, but the selection to classification number and initial cluster center in the K-mean cluster method does not have criterion to follow, human factor to cluster result good bad influence very big; It two is hierarchical clustering (Hierarchical clustering, HC) analytical approach, this method is by organizing data into some groups, and form a corresponding dendrogram and carry out cluster, after but the shortcoming of hierarchical clustering is data being merged or divide, just can't adjust again, be combined or the choice of location of split point can be followed owing to no objective standard in addition, mainly handle, therefore still exist to be subjected to people's subjective factor to influence big defective with the personal experience; It three is affine clustering method (Frey B J, Dueck D.Clustering by passing messagesbetween data points.Science, 2007,315 (5814): 972～976), this analytical approach is just proposition in nearly 2 years, it is the cluster centre that at first all sample points of data set all is considered as the candidate, and be the Attraction Degree parameter of each sample point foundation and other sample points, each sample point is competed final cluster centre in the loop iteration process, when iterative process restrains, promptly also determine with corresponding cluster centre of all categories thereupon, then each sample point to be distributed to nearest cluster centre, thereby determine the classification under it. the advantage of this method is embodied in does not need to specify in advance cluster numbers, and it is objective to classify; But because the deflection parameter of decision cluster centre then needs artificial setting, thereby its optimum cluster result subjective factor influence still under one's control, it is random big. to this, people such as Wang Kaijun have proposed a kind of by filter out the affine clustering method of self-adaptation of optimum cluster result from the pairing cluster result of whole deflection parameters, this method is under the less situation of data volume, compare with the affine clustering method of tradition, its cluster quality is better than traditional affine clustering method, but adaptive clustering scheme itself needs to determine all pairing cluster results of deflection parameter, its treatment capacity is big, also just limit the ability of its deal with data, thereby existed the processing power of data relatively poor, can't carry out defectives such as clustering processing to the big image of data volume.

Summary of the invention

The clustering method that the objective of the invention is a kind of functional magnetic resonance images of research and design, hierarchical clustering is combined in the fMRI Flame Image Process or the selection difficulty of split point, big to personal experience's dependence to overcome, and the disadvantage that the affine cluster of self-adaptation can't be handled the bigger image of data volume, reach in effective raising fMRI image processing process objectivity and to the purposes such as processing power of images with large data volume.

Solution of the present invention is by hierarchical clustering (HC) and the affine cluster of self-adaptation (APC) method are organically combined, fully utilize both characteristics, overcome its defective, thereby realize its goal of the invention.Therefore, the present invention includes:

A. original fMRI image is carried out pre-service: at first the fMRI image to input carries out the correction of space displacement, and utilize the low-frequency noise that Hi-pass filter will be breathed, heartbeat class physiological activity causes to filter, image after will proofreading and correct again carries out standardization, then the imagery exploitation gaussian kernel after the standardization is carried out space smoothing and handles;

B. the affine cluster of subregion: at first carry out subregion according to the brain subregion template of routine, again according to resulting each sectional image subclass, utilize related coefficient between each voxel as the element in the similarity matrix, and the intermediate value of getting matrix is as the deflection parameter, utilize traditional affine clustering method (APC) that the image subset of each subregion is carried out cluster then, obtain the cluster result of each subregion correspondence;

C. obtain new image and determine corresponding deflection parameter: the subregion cluster result that will obtain by the B step, utilize the disposal route of hierarchical clustering (HC), with data similar in each subregion as a polymerization voxel, and reconfigure according to space distribution, obtain new image, to new image utilize again between each polymerization voxel related coefficient as the element in the similarity matrix, get matrix intermediate value as the deflection parameter;

D. obtain the pairing cluster result of whole deflection parameters: at first to by the image that forms after the C step subregion cluster and deflection parameter thereof by traditional affine clustering method cluster again, obtain this step cluster result for the first time; And then the image that forms after the C step subregion cluster used by the deflection parameter value after setting stride and reducing it is carried out cluster again, obtain cluster result for the second time; So recycle after successively decreasing successively by stride the deflection parameter again cluster, be 2 to end until the classification number, and preserve the whole cluster results of gained, change step e;

E. determine optimum cluster result: step D gained cluster result is pressed the Silhouette index evaluation, and the pairing cluster result of wherein minimum Silhouette value is optimum cluster result;

F. obtain the fMRI image after the cluster: with the optimum cluster result of step e gained, throw back in the original fMRI image, promptly obtain the fMRI image after the final cluster.

Above-mentioned Hi-pass filter is the Hi-pass filter of frequency 1/128Hz.Described image being carried out standardization, is that the EPI template that adopts SPM to carry is carried out standardization.Described brain subregion template is carried out subregion, and its template is the AAL template.And the related coefficient between described each voxel is determined by following formula:

$r_i=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}(x_i-\stackrel{\‾}{x})(y_i-\stackrel{\‾}{y})}{\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(x_i-\stackrel{\‾}{x})}^2\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(y_i-\stackrel{\‾}{y})}^2}}$

Wherein: x _iAnd y _iBe any two samples, i is an i sample.

Described by the cluster once more of the deflection parameter value after setting stride and reducing, its stride is:

p _step＝0.1p～0.001p

Wherein: p _StepBe that stride, p are by C step gained deflection parameter.

Described cluster result is pressed the Silhouette index evaluation, and each desired value is provided by following formula:

$S_{\mathrm{ht}}\left(\mathrm{ii}\right)=\frac{1}{n}\underset{t=1}{\overset{n}{\mathrm{\Σ}}}\frac{b\left(t\right)-a\left(t\right)}{\max \{a\left(t\right),b\left(t\right)\}}$

Wherein a (t) is the average dissimilar degree or the distance of other all samples in the class at a sample in certain class and its place, and b (t) is this sample to the minimum average B configuration dissmilarity degree or the distance of all classes in addition.

The present invention is because with hierarchical clustering and affine cluster and the affine cluster combination of self-adaptation, carry out overall treatment to image, overcome that subjective factor that traditional clustering method is subjected to the people influences and big, and can't carry out defectives such as effective clustering processing the big image of data volume to personal experience's dependence; Efficiently solve to optimum cluster result objective selection problem.Thereby the present invention has the processing power of images with large data volume strong, and the characteristics such as accuracy that effectively improved the objectivity in the fMRI image processing process and optimum cluster result is selected.

Description of drawings

Fig. 1 is the inventive method schematic flow sheet (block scheme);

Fig. 2 is the 18th layer in the fMRI original image in the specific embodiment of the invention, among the figure representative of different gray scales different classes of, by finding out among the figure that the classification base originally equals the voxel number;

Fig. 3 is the result of the affine cluster of subregion and hierarchical clustering in the embodiment, is illustrated as the 18th layer, by finding out among the figure that the classification number reduces in a large number;

Fig. 4 is the 18th layer of fMRI image after the final cluster.

Embodiment

The present invention is described further below in conjunction with the drawings and specific embodiments, wherein according to the character of bimanual movements fMRI image, can reasonablely observe the active region for the 17th, 18,19 layer of the fMRI image, here we choose the 18th layer of original fMRI image and demonstrate corresponding results.

For the process and the effect of a kind of functional MRI image clustering method based on affine cluster that the present invention is mentioned are described, analyze adopting the fMRI image of more common bimanual movements, character according to bimanual movements fMRI image, we choose the 18th layer of original fMRI image and demonstrate corresponding results, and concrete steps are as follows;

A. original fMRI image is carried out pre-service: at first the fMRI image to input carries out the correction of space displacement, and utilize frequency to filter for the low-frequency noise that the Hi-pass filter of 1/128Hz will be breathed, heartbeat class physiological activity causes, the EPI template that image after will proofreading and correct again adopts SPM to carry is carried out standardization to each voxel resampling to 3 * 3 * 3mm3; Be that the gaussian kernel of 8 * 8 * 8mm3 (being respectively x, y, three directions of z) is carried out the space smoothing processing to the half high value (FWHM) of the imagery exploitation overall with after the standardization then, its result as shown in Figure 2;

B. the affine cluster of subregion: at first carry out subregion, according to resulting each sectional image subclass, utilize related coefficient again according to the brain subregion template (AAL template) of routine:

$r_i=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}(x_i-\stackrel{\‾}{x})(y_i-\stackrel{\‾}{y})}{\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(x_i-\stackrel{\‾}{x})}^2\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(y_i-\stackrel{\‾}{y})}^2}}$

As the corresponding first number in the similarity matrix, wherein x _iAnd y _iBe any two samples wherein, i is an i sample, and the intermediate value of getting matrix utilizes affine clustering method (APC) that the image subset of each subregion is carried out cluster as the deflection parameter then, obtains the cluster result of each subregion correspondence;

C. obtain new image and determine corresponding deflection parameter: the subregion cluster result that obtains by the B step, utilize the disposal route of hierarchical clustering (HC), with data similar in each subregion as a polymerization voxel, and reconfigure according to space distribution, obtain new image, as shown in Figure 3, to new image utilize again between each voxel related coefficient as the element in the similarity matrix, get matrix intermediate value as the deflection parameter p;

D. obtain the pairing cluster result of whole deflection parameters: at first to by the image that forms after the C step subregion cluster and corresponding deflection parameter by traditional affine clustering method cluster again, obtain this step cluster result for the first time; And then with the image that forms after the C step subregion cluster with in the deflection parameter value cluster of setting after stride (ratio) reduces once more, obtain cluster result for the second time; So recycle by stride p _StepDeflection parameter after=0.01p successively decreases successively again cluster, be 2 to end until the classification number, and preserve the whole cluster results of gained, change step e;

E. determine optimum cluster result: step D gained cluster result is pressed the Silhouette index:

$S_{\mathrm{ht}}\left(\mathrm{ii}\right)=\frac{1}{n}\underset{t=1}{\overset{n}{\mathrm{\Σ}}}\frac{b\left(t\right)-a\left(t\right)}{\max \{a\left(t\right),b\left(t\right)\}}$

Evaluation, wherein a (t) is the average dissimilar degree or the distance of other all samples in the class at a sample and its place in a certain class, b (t) is minimum average B configuration dissmilarity degree or the distance of this sample to other all classes; The S of minimum in all indexs _Ht(ii) be worth pairing cluster result and be optimum cluster result;

F. obtain the fMRI image: with the optimum cluster result of step e gained, throw back in the original fMRI image, promptly obtain final fMRI image, as shown in Figure 4.

Claims (7)

1. the clustering method of a functional magnetic resonance images comprises:

A. original fMRI image is carried out pre-service: at first the fMRI image to input carries out the correction of space displacement, and utilize the low-frequency noise that Hi-pass filter will be breathed, heartbeat class physiological activity causes to filter, image after will proofreading and correct again carries out standardization, then the imagery exploitation gaussian kernel after the standardization is carried out space smoothing and handles;

B. the affine cluster of subregion: at first carry out subregion according to the brain subregion template of routine, again according to resulting each sectional image subclass, utilize related coefficient between each voxel as the element in the similarity matrix, and the intermediate value of getting matrix is as the deflection parameter, utilize traditional affine clustering method that the image subset of each subregion is carried out cluster then, obtain the cluster result of each subregion correspondence;

C. obtain new image and determine corresponding deflection parameter: the subregion cluster result that will obtain by the B step, utilize the disposal route of hierarchical clustering, with data similar in each subregion as a polymerization voxel, and reconfigure according to space distribution, obtain new image, to new image utilize again between each polymerization voxel related coefficient as the element in the similarity matrix, get matrix intermediate value as the deflection parameter;

D. obtain the pairing cluster result of whole deflection parameters: at first to by the image that forms after the C step subregion cluster and deflection parameter thereof by traditional affine clustering method cluster again, obtain this step cluster result for the first time; And then the image that forms after the C step subregion cluster used by the deflection parameter value after setting stride and reducing it is carried out cluster again, obtain cluster result for the second time; So recycle after successively decreasing successively by stride the deflection parameter again cluster, be 2 to end until the classification number, and preserve the whole cluster results of gained, change step e;

E. determine optimum cluster result: step D gained cluster result is pressed the Silhouette index evaluation, and the pairing cluster result of wherein minimum Silhouette value is optimum cluster result;

F. obtain the fMRI image after the cluster: with the optimum cluster result of step e gained, throw back in the original fMRI image, promptly obtain the fMRI image after the final cluster.

2. by the clustering method of the described functional magnetic resonance images of claim 1, it is characterized in that described Hi-pass filter is the Hi-pass filter of frequency 1/128Hz.

3. by the clustering method of the described functional magnetic resonance images of claim 1, it is characterized in that described image is carried out standardization is that the EPI template that adopts SPM to carry is carried out standardization.

4. by the clustering method of the described functional magnetic resonance images of claim 1, it is characterized in that described brain subregion template carries out subregion, its template is the AAL template.

5. by the clustering method of the described functional magnetic resonance images of claim 1, it is characterized in that and related coefficient between described each voxel is determined by following formula:

$r_i=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}(x_i-\stackrel{\‾}{x})(y_i-\stackrel{\‾}{y})}{\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(x_i-\stackrel{\‾}{x})}^2\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(y_i-\stackrel{\‾}{y})}^2}}$

Wherein: x _iAnd y _iBe any two samples, i is an i sample.

6. by the clustering method of the described functional magnetic resonance images of claim 1, it is characterized in that describedly by the deflection parameter value cluster of setting after stride reduces once more, its stride is:

p _step＝0.1p～0.001p

Wherein: p _StepBe that stride, p are by C step gained deflection parameter.

7. by the clustering method of the described functional magnetic resonance images of claim 1, it is characterized in that described cluster result by the Silhouette index evaluation, each desired value is provided by following formula:

$S_{\mathrm{ht}}\left(\mathrm{ii}\right)=\frac{1}{n}\underset{t=1}{\overset{n}{\mathrm{\Σ}}}\frac{b\left(t\right)-a\left(t\right)}{\max \{a\left(t\right),b\left(t\right)\}}$

Wherein: a (t) is the average dissimilar degree or the distance of other all samples in the class at a sample in certain class and its place, and b (t) is this sample to the minimum average B configuration dissmilarity degree or the distance of all classes in addition.

Images