CN101788656B - Method for recognizing function response signal under function nuclear magnetic resonance scan - Google Patents

Abstract

The invention relates to a method for recognizing a function response signal under function nuclear magnetic resonance scan, comprising the following steps of: (1) obtaining task-state data and resting-state data by utilizing a nuclear magnetic resonance apparatus, processing the data in space and time by utilizing a function nuclear magnetic resonance analysis software SPM (statistical parametric mapping) to obtain a data set; (2)reducing the dimensionality of the task-state data and the resting-state data by utilizing a current principal component analysis PCA method, conserving main information, i.e. conserving the eigenvectors of information energy of over 90 percent of the data, reestablishing data, and then respectively extracting independent components of the two kinds of data including a machine noise signal component, a non-neurogenic physiological noise component and a neurogenic function signal response component by utilizing an independent component analysis ICA method in current time domain; (3) finding out the range of the function signal component by carrying out corresponding traversal among independent components of the two kinds of data to obtain a data set containing the function signal component; and (4) carrying out spectral analysis on every signal in the data set, eliminating components without obvious energy peak values in frequency domain, and selecting a principal component from rest components which is the principal function respond signal.

Description

The discriminating method of function response signal under function nuclear magnetic resonance scan

Technical field

The present invention relates to the extraction of function nuclear magnetic resonance, NMR (fMRI) signal, particularly the discriminating method of a kind of function nuclear magnetic resonance, NMR (fMRI) scanning response signal under function, belong to the technical fields such as image processing and spectrum analysis.

Background technology

Function nuclear magnetic resonance, NMR (fMRI) signal sensitivity is very high, in the gatherer process of temporal image, the physiological signal that experimenter's small head is moving, heartbeat, eye movement and some incoherent muscular movements cause also is mingled in the fMRI signal, has consisted of the noise signal in the brain function signal.Want to extract the functional response signal under outside stimulus, these noise signals and brain function signal must be separated.The method that extensively adopts at present is to utilize the hematodinamics receptance function (HRF) of blood oxygen saturation (BOLD) signal, this method relies on existing fixed model and to the hypothesis of signal distributions, be difficult to find the difference between individuality, to the examination of non-functional signal also robust not.The method that has also developed successively in recent years data-driven obtains the functional response signal, and (the independent element fractal method with the Martin proposition is the most typical, see Martin J.M., Scott M., Greg G.B.et.al, Analysis of fMRI Data by Blind Separation into Independent SpatialComponents, Human Brain Mapping, 1998,6:160-188), but method itself can only decomposite each unordered signal component, and brain function signal wherein can only be distinguished by the shape of composition ripple.Given this, these methods are more applicable for the data of exemplary functions response signal waveform, as the experimental data under tile designs, and for nearest application more and more widely the experimental data under the event relevant design seem helpless.

Summary of the invention

For overcoming the deficiency of prior art, how to disobey the pattern type, the function signal composition in abstraction function nuclear magnetic resonance, NMR (fMRI) signal.A key technology difficult problem relates to: how to merge huge task attitude data and tranquillization attitude data 1.; 2. how to screen the function signal composition in without the typical waveform situation.

The invention provides the discriminating method of a kind of function nuclear magnetic resonance, NMR (fMRI) scanning response signal under function, technical scheme is as follows: a kind of discriminating method of function response signal under function nuclear magnetic resonance scan is characterized in that:

(1) utilize nuclear magnetic resonance analyser to obtain task attitude and tranquillization attitude data, then utilize function nuclear magnetic resonance spectroscopy software SPM to carry out pretreatment on room and time to task attitude and tranquillization attitude data, obtain data set;

(2) utilize existing principal component analytical method PCA, the dimension of reduction task attitude data and tranquillization attitude data, keep main information, namely keep formation task attitude and the tranquillization attitude data characteristic vector of 90% above information energy separately, rebuild task attitude and tranquillization attitude data; Then utilize the Independent Component Analysis ICA on existing time domain, extract respectively two kinds of data independent element separately, comprise machine noise signal component, non-neuropathic physiological noise composition and nerve function signal response component;

(3) by do the method for relevant traversal between each independent element of two kinds of data, find out the scope of function signal composition in task attitude data component, obtain comprising the set of signals of function signal composition;

(4) each signal in set of signals is carried out spectrum analysis, reject the composition that there is no remarkable energy peak on the frequency domain space, choose first composition in residual components, be the major function response signal.

Following concrete steps can be adopted in above-mentioned (2), (3), (4):

The first step is used principal component analytical method PCA, and the variable of task attitude data and tranquillization attitude data is reassembled into respectively one group of new irrelevant aggregate variable mutually;

(1) calculate two kinds of data matrix covariance matrix S separately;

(2) calculate the eigenvalue of the characteristic vector of each auto-covariance matrix S, choose front feature by large to little sequence according to eigenvalue, keep the information of two kinds of data more than separately 〉=90%;

(3) two kinds of data of projection to the new space that is comprised of the vector of character pair separately, realize data compression;

Second step, carry out independent component analysis ICA, first two kinds of data are carried out centralization and whitening pretreatment separately, solve from the albefaction sample optimization and separate mixed matrix, obtain independent base vector separately, be independent signal composition on time domain, actual measurement to time-domain signal be the linear combination of these independent elements; Task attitude data and tranquillization attitude data are carried out respectively above process, set and keep respectively 25 compositions, the new signal that comprises the main information of noise, non-function of nervous system physiological signal and function of nervous system's signal obtains the independent element set of two class data, is designated as respectively SE _i(t) (i=1,2 ..., 25) and SR _i(t) (i=1,2 ..., 25);

At the tranquillization attitude independent element SE that decomposes out _i(t) and task attitude independent element SR _i(t) correlation coefficient d is calculated in combination in twos between _ij

$d_{\mathrm{ij}}=\frac{\underset{t=1}{\overset{T}{\mathrm{\Σ}}}({\mathrm{SE}}_i\left(t\right)-{\stackrel{\‾}{\mathrm{SE}}}_i)({\mathrm{SR}}_j\left(t\right)-{\stackrel{\‾}{\mathrm{SR}}}_j)}{\sqrt{\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{({\mathrm{SE}}_i\left(t\right)-{\stackrel{\‾}{\mathrm{SE}}}_i)}^2\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{({\mathrm{SR}}_j\left(t\right)-{\stackrel{\‾}{\mathrm{SR}}}_j)}^2}}$

Wherein, i=1,2 ..., 25; J=1,2 ..., 25; SE _i(t) be task attitude data independent elements, SE _i(t) be the meansigma methods of this independent element on time shaft; SR _j(t) be tranquillization attitude data independent elements, SR _jBe the meansigma methods of this independent element on time shaft.Select front 3 independent elements of task attitude data by following formula:

${\mathrm{SE}}_i\left(t\right)=\arg \underset{i}{\min }\left[\arg \underset{j}{\max }\left(\right|d_{\mathrm{ij}}\left|\right)\right]$

The 3rd step, by observing spectrogram, noise, non-function of nervous system physiological signal and 3 task attitude data independent elements of function of nervous system's signal of choosing are carried out spectrum analysis, rejecting does not have the composition of remarkable energy peak on the frequency domain space, choose first composition in residual components, namely select tacticly according to the second step method, satisfying for the 3rd step, first signal of score value energy requirement is arranged is the major function response signal.All the other compositions are the composition that contains residue functional response signal energy, according to practical application request, do auxiliary the use.

Advantage of the present invention and remarkable result:

(1) the inventive method can combine task attitude data and the tranquillization attitude data under function nuclear magnetic resonance, NMR (fMRI) scanning, under the condition without any priori, extracts neururgic function signal composition, need not prior model;

(2) due to task attitude data and tranquillization attitude data in conjunction with same individuality, by the basal signal characteristic of tranquillization attitude data, compare the correlation coefficient d between two types of data heterogeneities _ijSize (i=1,2 ..., 25; J=1,2 ..., 25), can screen and contain the high composition of function signal energy comparison, thereby not need to screen it by the waveform characteristic of function signal itself.Make this technology can be widely used in the experimental data of non-tile designs, be not limited to the data analysis of tile designs;

(3) reflected from different angles the different modalities that same individuality is current due to task attitude data and tranquillization attitude data, the original method (method of model-based methods and employing single type signal) that relatively is compared between them can embody neururgic individual variation more; In addition, also can better eliminate effect of noise by spectral performance to the screening of composition in method.The functional response signal that extracts can embody individual instances more.

Description of drawings

Fig. 1 is the inventive method flow chart;

Fig. 2 is that a certain individuality calculates acquisition correlation coefficient d according to Fig. 1 flow process _ijFront 3 compositions after sequence;

Fig. 3 is spectrogram corresponding to each signal in Fig. 2;

Fig. 4 is that the application performance of the functional response signal of the whole bag of tricks extraction compares.

The specific embodiment

The experimenter is under Chinese Facial Expression Video System (Chinese Facial Expression Video System, CFEVS) stimulates, and scanning acquisition task attitude data scan afterwards and obtain tranquillization attitude data.Utilize fMRI analysis software SPM to carry out pretreatment on room and time to data, comprise a dynamic(al) correction, Spatial normalization, Gaussian smoothing filtering, low frequency noise is removed in time standard and high-pass filtering, after operation, obtains data set.Utilize principal component analysis PCA (Tipping M, BishopC.Mixtures of probabilistic principal component analyzers.NeuralComputation, 1999, the 11:443-482) dimension of reduction task attitude data and tranquillization attitude data keeps main information; Then utilize the independent component analysis (ICA) on time domain to extract respectively each 25 independent elements of two kinds of data; By do the method for relevant traversal between the independent element of two kinds of data, find out the scope of function signal composition again, obtain comprising the set of signals of function signal composition, as shown in Figure 1; Each signal in set of signals is carried out spectrum analysis, its spectrogram as shown in Figure 2, its transverse axis and the longitudinal axis represent respectively the energy under frequency and respective frequencies.According to the derivation of formula (2), first pickup electrode in Fig. 2 might comprise the function signal composition of a large amount of neural activity features, might be also the noise composition.The first width figure of Fig. 3 has shown the spectrogram of this signal, from the spectrogram of this signal, the energy of each frequency content is uneven distribution, can get rid of the probability that this signal is noise signal, thinks that this signal is the function signal composition that has comprised a large amount of neural activity features.Second signal in Fig. 2, in relevant traversal, also very little with the correlation coefficient absolute value of each independent element of tranquillization attitude, in Fig. 3, the second width figure is the spectrogram of this signal, as seen from the figure, the Energy distribution of this each frequency content of signal is evenly distributed, and may be to cause because mix more noise, for the signal guaranteeing to gather is the function signal composition, we reject it.The frequency spectrum of the 3rd signal also is uneven distribution, and as seen, this signal contains part residue functional response signal energy, can continue to employ to do and assist a ruler in governing a country application.

In order further to analyze the effectiveness of the function signal that is extracted, can extracting function signal as feature, by checking, the identification ability of dysfunction individuality under same task be verified.Collect 13 Depression in women patients and 9 healthy persons, repeat said process, obtain each individual functional response signal, as feature, enter Bayes classifier, observe classifying quality.In order to verify performance, the function signal that has also adopted other two typical methods to obtain is respectively as recognition feature.They are respectively 1) utilize function signal that Martin obtains for the analytical method of representative as characteristic of division (Martin J.M., Scott M., Greg G.B.et.al, Analysis of fMRI Data by Blind Separation intoIndependent Spatial Components, Human Brain Mapping, 1998,6:160-188); 2) utilize hematodinamics receptance function (HRF) to optimize the function signal of acquisition as characteristic of division (Cynthia H.Y.Fu, JanainaMourao-Miranda, Sergi G.Costafreda et.al.Pattern Classification of Sad Facial Processing:Toward the Development of Neurobiological Markers in Depression.BiologicalPsychiatry, 2008,63:656-662).We use whole accuracy rate, and patient's accuracy rate and healthy person accuracy rate are come the classification of assessment performance, classification results such as Fig. 4.In Fig. 4 form, method one is to utilize the function signal of this method as characteristic of division; Method two is to utilize function signal that Martin obtains for the analytical method of representative as characteristic of division (data of event related experiment do not have the typical function signal waveform can be for reference, therefore this way lost efficacy to this batch experimental data); Method three is to utilize hematodinamics receptance function (HRF) to optimize the function signal of acquisition as characteristic of division.

As seen from Figure 4, although the inventive method and based on the method for prior model done same thing (as the table in method three), all to extract the function signal composition relevant to neural activity in brain, but the method that the present invention proposes has but obtained classification results relatively preferably without any the participation of prior model.The superiority of result may come from, and has broken away from model, by the comparison to the polymorphic signal of individuality own, can well embody individual specificity, also can eliminate preferably effect of noise simultaneously.Data in literary composition are the data of event relevant design experiment, but not the data of tile designs experiment.The advantage of this stimulation design is single stimulation is come respectively, can better show the response process of brain to stimulating, but meanwhile, also increased the difficulty of data analysis, method two in table is difficult to distinguish that from waveform which signal component is the function signal composition, therefore can't complete identification mission.

Claims (2)

1. the discriminating method of a function response signal under function nuclear magnetic resonance scan is characterized in that:

(1) utilize nuclear magnetic resonance analyser to obtain task attitude data and tranquillization attitude data, then utilize function nuclear magnetic resonance spectroscopy software SPM to carry out pretreatment on room and time to task attitude data and tranquillization attitude data, obtain data set;

(2) utilize existing principal component analytical method PCA, the dimension of reduction task attitude data and tranquillization attitude data, keep main information, namely keep formation task attitude data and the tranquillization attitude data characteristic vector of 90% above information energy separately, rebuild two kinds of data of task attitude and tranquillization attitude; Then utilize the Independent Component Analysis ICA on existing time domain, extract respectively two kinds of data independent element separately, comprise machine noise, non-neuropathic physiological signal and nerve function signal;

(3) by do the method for relevant traversal between each independent element of two kinds of data, find out the scope of function signal composition in task attitude data component, obtain comprising the set of signals of function signal composition;

(4) each signal in set of signals is carried out spectrum analysis, reject the composition that there is no remarkable energy peak on the frequency domain space, choose first composition in residual components, be the major function response signal.

2. the discriminating method of function response signal under function nuclear magnetic resonance scan according to claim 1, it is characterized in that: (2), (3), (4) concrete steps are:

The first step is used principal component analytical method PCA, and the variable of task attitude data and tranquillization attitude data is reassembled into respectively one group of new irrelevant aggregate variable mutually;

(1) calculate two kinds of data matrix covariance matrix S separately;

(2) calculate the eigenvalue of the characteristic vector of each auto-covariance matrix S, choose front feature by large to little sequence according to eigenvalue, keep two kinds of data information more than 90% separately;

(3) two kinds of data of projection to the new space that is comprised of the vector of character pair separately, realize data compression;

Second step, carry out independent component analysis ICA, first two kinds of data are carried out centralization and whitening pretreatment separately, solve from the albefaction sample optimization and separate mixed matrix, obtain independent base vector separately, be independent element on time domain, actual measurement to time-domain signal be the linear combination of these independent elements; These two kinds of data settings are kept respectively 25 compositions, comprise the new signal of the main information of noise, non-function of nervous system physiological signal and function of nervous system's signal, obtain the independent element set of two class data, be designated as respectively SE _i(t) (i=1,2 ..., 25) and SR _i(t) (i=1,2 ..., 25);

At the tranquillization attitude independent element SR that decomposes out _i(t) and task attitude independent element SE _i(t) correlation coefficient d is calculated in combination in twos between _ij

$d_{\mathrm{ij}}=\frac{\underset{t=1}{\overset{T}{\mathrm{\Σ}}}({\mathrm{SE}}_i\left(t\right)-{\stackrel{\‾}{\mathrm{SE}}}_i)({\mathrm{SR}}_j\left(t\right)-{\stackrel{\‾}{\mathrm{SR}}}_j)}{\sqrt{\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{({\mathrm{SE}}_i\left(t\right)-{\stackrel{\‾}{\mathrm{SE}}}_i)}^2\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{({\mathrm{SR}}_j\left(t\right)-{\stackrel{\‾}{\mathrm{SR}}}_j)}^2}}$

Wherein, i=1,2 ..., 25; J=1,2 ..., 25; SE _i(t) be task attitude data independent elements, Be the meansigma methods of this independent element on time shaft; SR _j(t) be tranquillization attitude data independent elements,

Be the meansigma methods of this independent element on time shaft, select front 3 independent elements of task attitude data by following formula:

${\mathrm{SE}}_i\left(t\right)=\arg \underset{i}{\min }\left[\arg \underset{j}{\max }\left(\right|d_{\mathrm{ij}}\left|\right)\right]$

The 3rd step, by observing spectrogram, noise, non-function of nervous system physiological signal and 3 task attitude data independent elements of function of nervous system's signal of choosing are carried out spectrum analysis, rejecting does not have the composition of remarkable energy peak on the frequency domain space, choose first composition in residual components, namely select tacticly according to the second step method, satisfying for the 3rd step, first signal of peak energy requirement is arranged is the major function response signal.

Images