CN103425999A - Brain cognitive state judgment method based on non-negative tensor projection operator decomposition algorithm - Google Patents

Abstract

Disclosed is a brain cognitive state judgment method based on a non-negative tensor projection operator decomposition algorithm. The method includes the steps of S1, collecting brain functional magnetic resonance images under different cognitive tasks to form a data sample set, carrying out preprocessing, and forming a sample set according to tensor modes, wherein the sample set is divided into a training set and a testing set according to the cognitive tasks, and the training set comprises functional magnetic resonance data in similar proportion of different cognitive states, S2, computing non-negative tensor projection operator decomposition of the training sample set to solve out a non-negative feature transformation matrix, projecting training samples to a non-negative tensor feature sub-space for dimensionality reduction to obtain a non-negative feature tensor set of the training set, S3, using lower-dimension non-negative feature tensor data after dimensionality reduction as input of an STM for training to solve out the optimum projection direction of the STM, and S4, projecting brain functional magnetic resonance data of tested samples to the non-negative tensor feature sub-space obtained through training to obtain non-negative feature tensors of the brain functional magnetic resonance data in the sub-space, and inputting the non-negative feature tensors of the tested samples to the trained STM to judge cognitive state types of the non-negative feature tensors.

Description

Brain cognitive state decision method based on non-negative tensor projection operator decomposition algorithm

Technical field

The invention belongs to brain function MRI feature extraction and brain cognitive state classification field, tensor machine identification and classification is supported in the pre-service and the expression of tensor pattern, the non-negative dimensionality reduction of the brain function MR data based on the tensor pattern and the utilization that relate to the function NMR imaging data of brain blood oxygen level, is based on feature extraction algorithm and the identification and classification algorithm of tensor pattern.

Background technology

For a long time, people are doing all effort always, hope can be opened the fan of large brain cognitive function, but still continuing to this research today, particularly in recent years, along with the develop rapidly of cerebral function imaging technology, Cognitive Neuroscience, computational neuroscience etc., formed the new upsurge of brain cognitive function research at home and abroad.

Use at present Functional MRI (functional magneticresonanceimages in the cognition neural research field more, be called for short fMRl) technology, the researchist thinks that brain function MRI can be used as the important means that the brain cognitive function is explored.Particularly use the brain function MR data based on task to study the more of large brain cognitive function.Because it can not only show the activation in brain district, can also directly show position and degree that the brain district activates, be considered to crack the hope place of a large brain cognitive function difficult problem.In the last few years research and the exploration of functional MRI data under different brain cognitive states were become to a focus, make this field constantly produce a large amount of data, how reasonably to utilize machine learning and algorithm for pattern recognition efficiently to process and analyze these brain function MR data, further probe into the brain Mechanism of Cognition, become gradually new scientific research focus.

The experimental data that the brain function magnetic resonance gathers is the data that a class dimension is high, data volume is huge, noise is very strong, structure is complicated especially.Therefore, the analysis of brain function MR data is a difficulty and far reaching work, and it has directly determined to adopt the brain function magnetic resonance method to carry out the success or failure of brain cognitive function research.At machine learning and area of pattern recognition, the form meaned as data object, pattern be pattern-recognition basic operation to the picture, its expression can present various ways, the data object representation that selection is conducive to the problem solution most is one of key of classification learning system success, in the conventional statistics pattern-recognition, data generally adopt vector pattern to mean, it is not always effective that yet research shows that vectorization means, be readily appreciated that the vectorization meeting destroys brain function magnetic resonance subject original multistage structure and the correlativity between raw data and very easily causes serious small sample problem, and may lose more succinct, the more useful form of expression that can obtain from primitive form.Therefore selecting suitable brain function magnetic resonance expression pattern of crucial importance, in view of the brain function MR data was exactly the characteristics on three rank originally, be applicable to very much the tensor modal representation, tensor is exactly multi-dimensional matrix.The tensor pattern is as the expansion of traditional vector pattern and supplement, and has caused in recent years the extensive concern in the fields such as machine learning, pattern-recognition.

Yet existing tensor resolution algorithm is such as CP decomposes, Tucker decomposes, the MPCA algorithm all could not keep the nonnegativity of protocerebrum archicerebrum functional MRI object, and the NTF(Non-negative Tensor Factorization of existing non-negative tensor resolution algorithm as decomposed based on CP) algorithm, although retained the nonnegativity of original object, but the sparse property that can not enough keep original object after projecting to non-negative tensor space, the NTD(Non-negative Tucker Decomposition decomposed based on Tucker) algorithm is in fact a kind of non-negative PCA algorithm of multidimensional, although be a kind of non-negative algorithm, but could not enough eliminate the redundancy between major component, be that feature representation is sparse not.

In view of the foregoing this patent proposes a kind of new non-negative tensor projection operator decomposition (Projective Non-negative Tensor Factorization is called for short PNTF) method and is applied to the analysis of brain function MR data.Non-negative tensor projection operator decomposes (PNTF) method.This method is directly processed the tensor data, from multiple directions tensor pattern brain function MR data object is carried out to non-negative dimensionality reduction and feature extraction, overcome simple the carrying out dimensionality reduction and destroyed structure and the correlativity of raw image data of traditional NMF, can not keep the deficiency of information and structure in raw image data fully.The great majority that it not only allows affine tensor to catch to appear on original tensor change simultaneously due to the non-negative orthogonality between extracted basic ingredient, project to after non-negative tensor space the sparse property that can keep original brain function magnetic resonance subject.

On this basis, we combine again the characteristics of existing STM, support that tensor machine algorithm (Support Tensor Machines is called for short STM) is the expansion of a kind of traditional algorithm of support vector machine (Support Vector Machines SVM) on the tensor pattern.Be the have supervise algorithm of a kind of classics based on the tensor mode data proposed in recent years, it can retain the original structure of brain function magnetic resonance and the correlativity between raw data effectively.The method that has finally formed PNTF-STM is applied in the discriminatory analysis of Different Cognitive state hypencephalon functional MRI data.

The research direction of this paper is based on the tensor schema object, to (the functional Magnetic Resonance Imaging of cerebral function magnetic resonance imaging under the Different Cognitive state, be called for short functional MRI data and carry out pre-service, Feature Dimension Reduction, feature extraction, and carry out on this basis the Classification and Identification of brain cognitive state.

Summary of the invention

The present invention is directed to the problem existed in the functional MRI data feature extraction that gathers under current brain Different Cognitive task and discriminant classification field, proposed a kind of analytical approach of the different brain cognitive state functional MRI datas based on the tensor pattern.In view of the deficiency that existing tensor resolution algorithm is applied on the brain function magnetic resonance subject, we propose a kind of brain cognitive state decision method based on non-negative tensor projection operator decomposition algorithm, comprise the steps:

S1 gathers the cerebral function magnetic resonance image (MRI) under the Different Cognitive task, form brain cognitive function MR data sample set, to described set of data samples, use the cerebral function imaging analysis software to carry out pre-service, the pre-service result data is made into sample set by the tensor modal sets, and described sample set is divided into to training set and test set two parts by Cognitive task, the training set part should comprise the suitable functional MRI data of Different Cognitive state ratio;

The non-negative tensor projection operator of S2 calculation training sample set decomposes, and obtains non-negative eigentransformation matrix, and training sample is projected to non-negative tensor property subspace dimensionality reduction, obtains the non-negative feature tensor set of training set;

The input of S3 using the non-negative feature tensor data of the low-dimensional after the training set dimensionality reduction as training STM, obtain the optimum projecting direction of STM, obtains weights and the constant offset of STM;

The non-negative tensor property subspace that S4 projects the training gained by test sample book brain function MR data obtains its non-negative feature tensor in subspace, and then the STM that the non-negative feature tensor input of test sample book is trained differentiates cognitive state classification under it.

On the basis of technique scheme, to non-negative feature projection matrix Carry out random initializtion, and it is done to iteration optimization, when upgrading U ⁽ⁿ⁾The time, keep original { U ⁽¹⁾, U ⁽²⁾... U ^(n-1), U ⁽ⁿ⁺¹⁾U ^(N-1), U ^(N)Constant, get successively n=1,2 ..., N, it is right to complete

An optimizing, determine whether also needing iteration optimization by the judgement end condition, if do not meet iteration optimization, until meet end condition, finish to optimize.

Last training sample X _m(m=1 ..., M) to non-negative tensor projection operator, decompose projection on (PNTF) non-negative feature tensor subspace afterwards, the feature tensor set that obtains training sample is { Y _m, m=1 ..., M}, $Y_m\&Element;{\mathrm{<figure-callout\; id="1"\; label="R\; P"\; filenames="HDA0000372913770000011.png"\; state="\{\{state\}\}">R</figure-callout>}}^{P_{}}.$

On the basis of technique scheme, obtain one by training sample feature tensor set and support the tensor machine

Wherein define 1 mark one class brain cognitive state, other classification brain cognitive state of-1 mark.

On the basis of technique scheme, test sample book brain function MR data is projected to the non-negative proper subspace that non-negative tensor projection operator decomposes gained, obtain the non-negative feature tensor of test sample book, be input to afterwards the training gained and support tensor machine STM

Completing the test sample book cognitive state differentiates.

On the basis of technique scheme, suppose

Be a test sample book, test sample book projects non-negative proper subspace: $\tilde{S}=S{\&times;}_1{U}^{{\left(1\right)}^T}{\&times;}_2{U}^{{\left(2\right)}^T}\&CenterDot;\&CenterDot;\&CenterDot;{\&times;}_N{U}^{{\left(N\right)}^T}$ Obtain the feature tensor of test sample book at the non-negative proper subspace of PNTF

Test sample book feature tensor afterwards

As supporting the tensor machine

Input: $g\left(\tilde{S}\right)=\mathrm{sign}(<\tilde{S},W>+b)$

Can judge the differentiation of test sample book cognitive state according to the labeling method of definition before, wherein the mathematical expression of whole testing process is as follows:

$g\left(\tilde{S}\right)=\mathrm{sign}(<S{\&times;}_1{U}^{{\left(1\right)}^T}{\&times;}_2{U}^{{\left(2\right)}^T}\&CenterDot;\&CenterDot;\&CenterDot;{\&times;}_N{U}^{{\left(N\right)}^T},W>+b).$

Characteristics and advantages of the present invention is:

1, the present invention is the three-dimensional even more feature of higher-dimension according to original brain function MR data, utilize non-negative tensor projection operator decomposition algorithm (PNTF) directly the brain function MR data of multistage tensor pattern to be carried out to dimensionality reduction and feature extraction, traditional dimension reduction method based on arrow pattern has destroyed raw image data structure and correlativity because carrying out dimensionality reduction have been overcome, can not keep redundancy in original image and the deficiency of structure fully, retain the space structure information of Functional MRI data.

2, experimental result shows that the non-negative Projection Character matrix of non-negative tensor projection operator decomposition algorithm (PNTF) gained has very strong orthogonality, when general iteration repeatedly restrains with end condition, orthogonality reaches more than 95%, so after projection, the feature tensor of gained can be good at retaining the non-negative sparse property of original brain function MR data.

3, in view of the feature tensor after dimensionality reduction, appointing is so the high-order object, the present invention has used existing based on tensor pattern classifier support tensor machine STM, the feature tensor not being carried out to vectorized process is classified, kept whole identifying not go to destroy the structure of brain function MR data, the higher-dimension disaster of simultaneously having avoided whole assorting process to cause due to vector quantization.

4, because whole analytic process is directly by the original structure of full brain function MR data, to carry out Feature Dimension Reduction and discriminator, so it is succinct not need the participant to grasp in advance a large amount of priori operations.

The accompanying drawing explanation:

Fig. 1 schematic flow sheet of the present invention.

Embodiment:

Below in conjunction with specific embodiment, the present invention is described in detail.

Step 1,1 module 1 of describing by reference to the accompanying drawings: the brain function MR data preprocessing part gathered under the Different Cognitive task.At first, gather certain sample size Different Cognitive task hypencephalon functional MRI experimental data, the brain function MR data is carried out to pre-service.In the experiment of Functional MRI, if the data of directly using the most original machine imaging device to obtain can be brought such as signal to noise ratio (S/N ratio) lowlyer, the image stabilization degree is not high, different differences between samples are excessive etc. series of problems.These all can become interference and affect the unfavorable factor of obtaining accurate rule from view data.So before analytic statistics, must first to preliminary image, be proofreaied and correct, registration, standardization, level and smooth, remove linear drift and filtering etc. and anticipate, for the effect that improves mathematical modeling and analysis is laid a solid foundation.

Its step comprises:

1) timeslice is proofreaied and correct, and purpose is to become the difference on the acquisition time point between layers of voxel for correction group.Because it is multiple that the scan mode of timeslice has, each section is to obtain on different time points, and this species diversity can be brought certain impact to statistical study.

2) free-air correction, being also referred to as head proofreaies and correct, although can fix being had a fling at head before experiment, but because experimental period is very long, tested anxiety even volunteer is breathed, heartbeat, the physiological factors such as blood flow pulsation all can cause head movement, this is difficult to be avoided, because the BOLD signal itself of measuring is just very small, generally in 5% left and right, even so slight movement, all may have influence on measurement result, even the result of wanting is flooded, and it is also more to measure number of times, factor on outside environment causes the situation of head movement also inevitable, will cause like this data that gather in the different moment skew to occur, destroyed the corresponding relation between image.SPM is used first photo in a same tested scanning sequence as reference standard, and other images in sequence will align with reference standard, and this process is called registration.The registration of SPM has been used principle of least square method.

3) image standardization, in view of there is very big-difference in the brain between the subject on dissecting, such as the big or small volume of brain is not quite similar, we need to be tested to difference mind map do standardization, different subjects' brain image is mapped in a unified standard form space, make the size of all tested brains with towards all consistent.Like this could be under same standard, studied and compared, SPM is used ICBM152 people's standard brain map of Montreal neuromedicine research institute (Montre Neurologieal Institute, MNI) issue as standard, and the brain function magnetic resonance generally adopts the EPI template.

4) space smoothing, spatially use view data a smooth function (normally gaussian kernel function) to remove to carry out convolution algorithm exactly.Following advantage is smoothly arranged: can improve signal to noise ratio (S/N ratio), what Functional MRI detected is the Hemodynamics signal that neuronal activation produces, main corresponding lower frequency region part in image reconstruction, noise is corresponding the HFS of signal, after space smoothing, noise will obtain very large inhibition, so just can improve the signal to noise ratio (S/N ratio) of image; Make the requirement of data fit Gaussian random field, this is very important to utilizing the gaussian random field theory to do statistical inference, because just can improve the accuracy of statistical inference like this; Can eliminate the difference of different subject's diencephalon structures, use the experiment of different subject's average results for needs, these differences also can be brought very large impact, and this otherness will, by obfuscation, can not filter out the information of significant HFS after space smoothing simultaneously.

5) remove linear drift and filtering, along with can there be a linear trend in the lasting of machine working time, linear drift is noise, may be moving etc. relevant with machine or head, and can affect the authenticity of signal, therefore the general linear regression method that adopts is removed linear drift, and adopts the 0.01-0.08Hz frequency range to carry out low frequency filter frequency range to the gained signal, remove the interference of high-frequency signal, to reduce physiological noise (as breathing, heartbeat etc.).

Now the brain function MR data after pre-service is divided into to training dataset and test data two parts according to individuality, training dataset should comprise the Different Cognitive status function MR data that ratio is suitable, in order to avoid the situation of serious deflection one party occurs in training process.Test figure after forming.

Step 2,1 module 2 of describing by reference to the accompanying drawings: dimensionality reduction and feature extraction part.

This module mainly illustrates: non-negative tensor projection operator decomposes (PNTF) algorithm and produces optimum non-negative feature projection matrix stack $\{U^{\left(n\right)}\&Element;R^{I_n\&times;P_n},n=1,\&CenterDot;\&CenterDot;\&CenterDot;,N\}$ Process.

For the training sample of choosing, define M sample here

Set { the X formed ₁, X ₂..., X _MTraining sample, the tensor space that sample is corresponding is

I wherein _n(1≤n≤N) is the n-mode(pattern of tensor) dimension, N means the order of a tensor number.The target that non-negative tensor projection operator decomposes (PNTF) algorithm is to find the linear non-negative eigentransformation matrix stack of multidimensional

U ⁽ⁿ⁾Represent the n-mode(pattern) non-negative feature projection matrix on direction, P _nBe the user according to the self-defining value of practical application, each mode(pattern here) P on direction _nValue generally smaller, the value of all directions is all between 5-15.Non-negative transformation matrix collection $\{U^{\left(n\right)}\&Element;R^{I_n{\&times;P}_n},n=1,\&CenterDot;\&CenterDot;\&CenterDot;,N\}$ By original tensor space

From N mode(pattern) direction simultaneously dimensionality reduction project non-negative tensor subspace

$Y_m=X_m{\&times;}_1{U}^{{\left(1\right)}^T}{\&times;}_2{U}^{{\left(2\right)}^T}\&CenterDot;\&CenterDot;\&CenterDot;{\&times;}_N{U}^{{\left(N\right)}^T},m=1,\&CenterDot;\&CenterDot;\&CenterDot;,M.Y_m\&Element;{\mathrm{<figure-callout\; id="1"\; label="R\; P"\; filenames="HDA0000372913770000011.png"\; state="\{\{state\}\}">R</figure-callout>}}^{P_{}}$

Wherein, X _m* _nU ⁽ⁿ⁾Mean tensor X _mWith matrix

In the n-mode(pattern) taking advantage of on direction, wherein X _mThe dimension of n mould direction must and matrix

Row equate.

Specific implementation process:

2.1) initialization.To non-negative feature projection matrix stack

Carry out random initializtion, meet non-negative requirement.

2.2) algorithm iteration optimization.It is right to need

Do further to optimize, meet objective function, non-negative feature projection matrix More sparse and approach quadrature.

The concrete steps of iteration optimization are as follows:

(a) the non-negative feature projection matrix of random initializtion

(b) mean iterations for k=1:K(K)

Mean the tensor exponent number for n=1:N(N)

Order ${\mathrm{\&Phi;}}^{\left(n\right)}=X_{\left(n\right)}{U_{{\mathrm{\&Phi;}}_n}}^T{U}_{{\mathrm{\&Phi;}}_n},1\&le;n\&le;N,$

${U_{\mathrm{\&Phi;}}}_n=U^{\left(N\right)}\&CircleTimes;U^{(N-1)}\&CircleTimes;\&CenterDot;\&CenterDot;\&CenterDot;\&CircleTimes;U^{(n+1)}\&CircleTimes;U^{(n-1)}\&CircleTimes;\&CenterDot;\&CenterDot;\&CenterDot;U^{\left(2\right)}\&CircleTimes;U^{\left(1\right)},$

${{U_{\mathrm{\&Phi;}}}_n}^T=U^{{\left(N\right)}^T}\&CircleTimes;U^{{(N-1)}^T}\&CircleTimes;\&CenterDot;\&CenterDot;\&CenterDot;\&CircleTimes;U^{{(n+1)}^T}\&CircleTimes;U^{{(n-1)}^T}\&CircleTimes;\&CenterDot;\&CenterDot;\&CenterDot;U^{{\left(2\right)}^T}\&CircleTimes;U^{{\left(1\right)}^T,}$

According to ${\tilde{U}}^{\left(n\right)}=U^{\left(n\right)}\frac{{2X}_{\left(n\right)}{\mathrm{\&Phi;}}^{\left(n\right)}{U}^{\left(n\right)}}{U^{\left(n\right)}{U}^{{\left(n\right)}^T}{\mathrm{\&Phi;}}^{\left(n\right)}{\mathrm{\&Phi;}}^{{\left(n\right)}^T}{U}^{\left(n\right)}+{\mathrm{\&Phi;}}^{\left(n\right)}{\mathrm{\&Phi;}}^{{\left(n\right)}^T}{U}^{\left(n\right)}{U}^{{\left(n\right)}^T}{U}^{\left(n\right)}},$ Ask for

And its is worth to assignment to the characteristic of correspondence matrix U ⁽ⁿ⁾Complete the renewal to the non-negative feature projection matrix of n-mode,

Mean Kronecker product.

Calculate { Y _m, m=1 ..., M},

If

(η is user-defined smaller threshold value) or meet k=K, jump out circulation, obtains optimum non-negative feature projection matrix stack $\{U^{\left(n\right)}\&Element;R^{I_n{\&times;P}_n},n=1,\&CenterDot;\&CenterDot;\&CenterDot;,N\},$ Enter into next step.

2.3) the feature tensor set of calculation training sample.Training sample X _m(m=1 ..., M) to non-negative tensor projection operator, decompose projection on (PNTF) non-negative feature tensor subspace afterwards, the feature tensor set that obtains training sample is $\{Y_m=X_m{\&times;}_1{U}^{{\left(1\right)}^T}{\&times;}_2{U}^{{\left(2\right)}^T}\&CenterDot;\&CenterDot;\&CenterDot;{\&times;}_N{U}^{{\left(N\right)}^T},m=1,\&CenterDot;\&CenterDot;\&CenterDot;,M\},Y_m\&Element;{\mathrm{<figure-callout\; id="1"\; label="R\; P"\; filenames="HDA0000372913770000011.png"\; state="\{\{state\}\}">R</figure-callout>}}^{P_{}}.$

Step 3,1 module 3 of describing by reference to the accompanying drawings: discriminator part.

This module comprises the content of two aspects: tensor machine STM and test sample book classification are supported in training.

First, tensor machine STM is supported in training:

Support the input of tensor machine STM using the feature tensor set of training sample set after non-negative tensor projection operator decomposition algorithm (PNTF) dimensionality reduction as training, ask for optimum STM projecting direction W and constant offset b.

Support the mathematical function g of tensor machine to be expressed as follows:

g(Y _m)=sign(<Y _m，W>+b)，

Y wherein _mRepresent the feature tensor of m training sample, tensor Represent optimum STM projecting direction, b is constant.

Optimum STM projecting direction W tries to achieve by being calculated as follows optimization problem:

$\underset{W,\mathrm{\&zeta;},b\&GreaterEqual;0}{\min }<W,W>+C\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}{\mathrm{\&zeta;}}_i$

S.t.y _i[<Y _i，W>+b]≥1-ζ _i，1≤i≤M，ζ _i≥0。

Wherein C is the penalty factor in similar and support vector machine, ζ=[ζ ₁, ζ ₂... ζ _M] be equal to the relaxation factor in support vector machine, y _i∈ [1 ,-1] is self-defining training sample brain cognitive state classification mark, and we define 1 mark one class brain cognitive state, other classification brain cognitive state of-1 mark usually.

Use afterwards existing support tensor machine optimized algorithm to calculate W and constant offset b, we obtain one by training sample feature tensor set and support the tensor machine like this

Second portion, the test sample book classification:

Test sample book brain function MR data is projected to the non-negative proper subspace of PNTF, obtain the non-negative feature tensor of test sample book, non-negative feature tensor is input to the training gained and supports tensor machine STM afterwards

Completing test sample book brain cognitive state differentiates.Concrete steps are:

Suppose

Be a test sample book, at first test sample book projects the non-negative proper subspace of PNTF: $\tilde{S}=S{\&times;}_1{U}^{{\left(1\right)}^T}{\&times;}_2{U}^{{\left(2\right)}^T}\&CenterDot;\&CenterDot;\&CenterDot;{\&times;}_N{U}^{{\left(N\right)}^T}$

Obtain the feature tensor of test sample book brain function MR data at the non-negative proper subspace of PNTF $\tilde{S}\&Element;{\mathrm{<figure-callout\; id="1"\; label="R\; P"\; filenames="HDA0000372913770000011.png"\; state="\{\{state\}\}">R</figure-callout>}}^{P_{}}.$

Afterwards by test sample book feature tensor

As supporting the tensor machine

Input:

$g\left(\tilde{S}\right)=\mathrm{sign}(<\tilde{S},W>+b)$

Can judge that according to the labeling method of definition before test sample book is affiliated brain cognitive state classification.

For a person skilled in the art, can be improved according to the above description or be converted, but should be understood that, a kind of brain cognitive state decision method based on non-negative tensor projection operator decomposition algorithm of the present invention is not limited in the description in instructions, within the spirit and principles in the present invention all, any modification of making, equal replacement, improvement etc., within all being included in claim scope of the present invention.

Claims (4)

1. the brain cognitive state decision method based on non-negative tensor projection operator decomposition algorithm is characterized in that:

Comprise the steps:

S1 gathers the cerebral function magnetic resonance image (MRI) under the Different Cognitive task, form brain cognitive function MR data sample set, to described set of data samples, use the cerebral function imaging analysis software to carry out pre-service, the pre-service result data is made into sample set by the tensor modal sets, and described sample set is divided into to training set and test set two parts by Cognitive task, the training set part should comprise the suitable functional MRI data of Different Cognitive state ratio;

The non-negative tensor projection operator of S2 calculation training sample set decomposes, and obtains non-negative eigentransformation matrix, and training sample is projected to non-negative tensor property subspace dimensionality reduction, obtains the non-negative feature tensor set of training set;

The input of S3 using the non-negative feature tensor data of the low-dimensional after the training set dimensionality reduction as training STM, obtain the optimum projecting direction of STM, obtains weights and the constant offset of STM;

The non-negative tensor property subspace that S4 projects the training gained by test sample book brain function MR data obtains its non-negative feature tensor in subspace, and then the STM that the non-negative feature tensor input of test sample book is trained differentiates cognitive state classification under it.

2. the brain cognitive state decision method based on non-negative tensor projection operator decomposition algorithm as claimed in claim 1 is characterized in that:

To non-negative feature projection matrix stack

Carry out random initializtion, and it is done to iteration optimization, when upgrading U ⁽ⁿ⁾The time, keep original { U ⁽¹⁾, U ⁽²⁾... U ^(n-1), U ⁽ⁿ⁺¹⁾U ^(N-1), U ^(N)Constant, get successively n=1,2 ..., N, it is right to complete

An optimizing, determine whether also needing iteration optimization by the judgement end condition, if do not meet iteration optimization, until meet end condition, finish to optimize;

Last training sample X _m(m=1 ..., M) to non-negative tensor projection operator, decompose projection on (PNTF) non-negative feature tensor subspace afterwards, the feature tensor set that obtains training sample is { Y _m, m=1 ..., M},

3. the brain cognitive state decision method based on non-negative tensor projection operator decomposition algorithm as claimed in claim 1,, it is characterized in that:

Obtain one by training sample feature tensor set and support the tensor machine

Wherein define 1 mark one class brain cognitive state, other classification brain cognitive state of-1 mark.

4. the brain cognitive state decision method based on non-negative tensor projection operator decomposition algorithm as claimed in claim 1, it is characterized in that: test sample book brain function MR data is projected to the non-negative proper subspace that non-negative tensor projection operator decomposes gained, obtain the non-negative feature tensor of test sample book, be input to afterwards the training gained and support tensor machine STM

Completing the test sample book cognitive state differentiates;

Suppose

Be a test sample book, test sample book projects non-negative proper subspace:

Obtain the feature tensor of test sample book at the non-negative proper subspace of PNTF

Test sample book feature tensor afterwards

As supporting the tensor machine

Input:

Can judge the differentiation of test sample book cognitive state according to the labeling method of definition before, wherein the mathematical expression of whole testing process is as follows:

。

Images