KR20160016357A - Pattern Classification Apparatus and Method for fMRI - Google Patents

Abstract

Disclosed are a pattern classification device of functional magnetic resonance images and a method thereof. The present invention comprises: a brain condition signal acquisition unit for acquiring multiple signals of the conditions of a brain in which the patterns are classified beforehand for the machine learning of the pattern classification device; a characteristic extracting unit for extracting multiple characteristics in each of the multiple brain condition signals to acquire multiple characteristic vectors; a normalization unit for deducting averages of the characteristics included in each of the multiple characteristic vectors to remove drifting noise; an SVM classification unit for acquiring at least one linear classification function as a classifier having the maximum areas of a space for each of the multiple characteristics vectors using an SVM classification method and for inputting the multiple characteristic vectors into at least one classifier to compute a classifier output value; an output correction unit for removing time-series drifting noise from at least one classifier output value; and a classification determination unit for analyzing the classifier output value corrected by the output correction unit to classify the patterns of the brain condition signals into a corresponding category.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern classification apparatus and method for classifying functional MRI images,

The present invention relates to an apparatus and method for pattern classification of functional magnetic resonance images, and more particularly, to an apparatus and method for pattern classification of functional magnetic resonance images for multi-category brain state decoding.

Functional Magnetic Resonance Imaging (fMRI) is a method of measuring the activation pattern of the brain using an MRI device, which measures blood oxygenation level-dependent (BOLD) To measure the degree of activation. In other words, the part that is activated and functions in the brain or nervous system is illuminated and shown.

FIG. 1 is a view showing a comparison of a BOLD signal measured by fMRI in response to a stimulus and a stimulus applied to a subject, FIG. 2 shows an example of a two-dimensional sectional image obtained using a BOLD signal, and FIG. 2 shows an example of a three-dimensional image reconstructed from the two-dimensional image.

Increased activity in certain areas of the brain increases metabolism, which increases blood flow to the capillaries of certain tissues, which in turn increases the rate of oxygen-bound hemoglobin in the bloodstream. This hemoglobin has a higher signal intensity than the hemoglobin of surrounding tissues deprived of oxygen. The fMRI can detect this difference and generate the BOLD signal as shown in FIG. As shown in FIG. 2, the BOLD signal is used to construct a plurality of two-dimensional sectional images. A plurality of 2D cross-sectional images are reconstructed as a three-dimensional image so that activation patterns of the desired brain regions can be measured.

Therefore, fMRI can display the state of the brain as a two-dimensional and three-dimensional image signal (hereinafter referred to as "brain state signal") without the incision of the skull, so that it enables the noninvasive method of studying the relationship between brain and cognitive activity , And has set a great momentum for the development of brain research.

Currently, fMRI is being actively studied for diagnosis and treatment in various applications. In addition, fMRI technology has evolved from the brain activation signal to a level that can infer the brain state (mind action). In order to apply it to various application fields such as diagnosis of diseases using fMRI and neurofeedback to normalize the function of the body by controlling the brain activation pattern by itself, a basic task of pattern classification of the brain state signal acquired through fMRI should precede do. The fMRI pattern classification classifies the pattern of the brain signal detected by the fMRI to classify the state of the subject's brain. However, the pattern of the brain signal detected by the fMRI can vary widely depending on various conditions (gender, age, experience, personal preference, environment, etc.) of the subject, it's difficult. Therefore, pattern classification for brain state signals for brain state analysis is mostly performed using pattern classifiers learned through machine learning techniques.

The performance of the fMRI signal pattern classification using the pattern classifier greatly depends on the machine learning state of the pattern classifier performed prior to the application of the brain state signal to be classified. The machine learning of the pattern classifier is performed through a process of optimizing the parameters of the pattern classifier such that a pattern of a plurality of brain state signals obtained in advance is assigned to a corresponding category among a plurality of already known categories. After that, the pattern classifier transforms the newly applied brain state signal using the optimized parameters and classifies the patterns so that the output value of the pattern classifier of the category is maximized. Here, the pattern of the brain state signal represents the signal value of the voxel in the active region of the entire brain region.

However, there are various problems to be solved by the pattern classification technique of the brain state signal pattern using the machine learning technique of the pattern classifier, and the problem of the reduction of the classification accuracy due to the drift noise is the first problem to be solved . Flooding noise is considered to be caused by heat due to the main magnetic field of fMRI or other mechanical reason when performing long time fMRI measurement, but it can also be caused by physiological adaptation or change of the subject.

The floating noise makes the spatial pattern shape between the brain state signal pattern used for the machine learning and the newly obtained state signal pattern of the brain different from each other so that the pattern classifier generates the brain state signal pattern So that the performance of the pattern classifier is largely deteriorated.

Conventionally, a technique of dividing each of a plurality of brain state signal patterns into a spatial average value and normalizing it has been used as a method of removing floating noise. Recently, a linear tendency of a voxel based on time series information for each voxel trend) is generally used. However, the technique of removing the tendency may cause an estimation error in a real-time pattern classification environment in which insufficient and noisy data can not be used in estimating the drift trend for each voxel. And this estimation error is an obstacle to the machine learning of brain state analysis pattern classifier in which spatial pattern is important. This is due to the fact that the spatial pattern is likely to be distorted with the passage of time because the voiced noise is estimated and removed differently for each voxel. Particularly, when the pattern classification is classified into a plurality of categories, the distortion of the spatial pattern causes a significant performance degradation to the pattern classifier.

Therefore, in order to improve the performance of the pattern classifier of the machine learning method, there is a demand for a technology capable of minimizing spatial distortion by effectively removing the floating noise when classifying the acquired brain state signal pattern into various categories.

Korean Patent Laid-Open No. 10-2013-0051084 discloses a technique for reducing artifacts such as noise in MRI images in an apparatus and method for artifact reduction of high-order diffuse magnetic resonance imaging (published on May 20, 2013) As a technique for correction, it is not suitable for improving the performance of a pattern classifier for classifying brain state signal patterns into multiple categories.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a pattern classification apparatus for functional magnetic resonance images capable of improving the performance of a pattern classifier for classifying brain signal patterns into multiple categories.

It is another object of the present invention to provide a method of classifying functional MRI images to achieve the above objects.

According to an aspect of the present invention, there is provided an apparatus for classifying functional MRI images according to an exemplary embodiment of the present invention includes a brain state signal acquisition unit for obtaining a plurality of brain state signals, ; A feature point extracting unit for extracting a plurality of feature points from each of the plurality of brain state signals and obtaining a plurality of feature vectors; A normalization unit for subtracting an average of the feature points included in each feature vector from each of the plurality of feature vectors to remove floating noise; At least one linear discriminant function having a maximum margin for each of the plurality of feature vectors is obtained as at least one classifier using an SVM classification technique and the plurality of feature vectors are substituted into the obtained at least one classifier An SVM classifier for calculating a classifier output value; An output correction unit for removing time-series floating noise from at least one of the classifier output values; A classification discrimination unit for analyzing the classifier output value corrected by the output correcting unit and classifying the pattern of the brain state signal into a corresponding category; .

Wherein the SVM classifier obtains the linear discriminant function in a number corresponding to the number of categories to be classified and acquires a plurality of linear discriminant functions using a one-to-many residual approach when the number of categories to be classified is plural .

Wherein the classifying and discriminating unit analyzes the plurality of corrected classifier output values and analyzes the classifier output value having a maximum value if the classifying unit has a plurality of categories, And classifying the pattern of the brain state signal into categories.

Wherein the output correction unit obtains a slope and an offset using the least squares method for each of the at least one classifier output values, searches the time-series floating noise using the obtained slope and the offset, And the time-dependent floating noise is subtracted and removed.

The feature point extraction unit may divide a predetermined brain structure image by predetermined regions, match the image of the divided region with the brain state signal, and extract the feature points of the brain state signal using the matched image as a mask .

Wherein the fMRI pattern classification apparatus comprises: a preprocessing unit for performing spatial alignment and smoothing on the plurality of brain state signals acquired by the brain state signal acquisition unit and transmitting the spatial state alignment and smoothing to the feature point extraction unit; And further comprising:

According to another aspect of the present invention, there is provided a method for classifying a functional magnetic resonance image according to an embodiment of the present invention. The method includes classifying a fMRI including a brain state signal obtaining unit, a feature point extracting unit, a normalizing unit, an SVM classifying unit, A method for classifying an fMRI pattern of a pattern classification apparatus, the method comprising: acquiring a plurality of brain state signals in which categories of patterns are classified in advance for machine learning of the pattern classification apparatus; Extracting a plurality of feature points from each of the plurality of brain state signals to obtain a plurality of feature vectors; Normalizing the normalization by subtracting the average of the feature points included in each feature vector from each of the plurality of feature vectors by removing the floating noise; Acquiring, as at least one classifier, at least one linear discriminant function having a maximum margin for each of the plurality of feature vectors using the SVM classifier; Calculating a classifier output value by substituting the plurality of feature vectors into the at least one classifier obtained by the SVM classifier; The output correction unit corrects the classifier output value by removing time series floating noise from at least one classifier output value; And classifying the pattern of the brain condition signal into a corresponding category by analyzing the classifier output value classified and corrected; .

Therefore, the apparatus and method for classifying functional magnetic resonance images of the present invention not only improves the performance of the pattern classifier by removing the floating phenomenon of the output value of the multi-category pattern classifier at the time of pattern classification of the fMRI, And the pattern of the brain state signal is classified into a category corresponding to the classifier which outputs the maximum output value after eliminating the linear floating value individually. Therefore, stable pattern classification can be performed with high accuracy.

FIG. 1 is a graph comparing BOLD signals measured by fMRI in response to a stimulus and a stimulus applied to a subject.
2 shows an example of a two-dimensional sectional image obtained using the BOLD signal.
FIG. 3 shows an example of a three-dimensional image reconstructed from the two-dimensional image of FIG.
FIG. 4 illustrates a pattern classification apparatus for a functional magnetic resonance image according to an embodiment of the present invention.
FIG. 5 illustrates a pattern classification method of a functional magnetic resonance image according to an embodiment of the present invention.

In order to fully understand the present invention, operational advantages of the present invention, and objects achieved by the practice of the present invention, reference should be made to the accompanying drawings and the accompanying drawings which illustrate preferred embodiments of the present invention.

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, the present invention will be described in detail with reference to the preferred embodiments of the present invention with reference to the accompanying drawings. However, the present invention can be implemented in various different forms, and is not limited to the embodiments described. In order to clearly describe the present invention, parts that are not related to the description are omitted, and the same reference numerals in the drawings denote the same members.

Throughout the specification, when an element is referred to as "including" an element, it does not exclude other elements unless specifically stated to the contrary. The terms "part", "unit", "module", "block", and the like described in the specification mean units for processing at least one function or operation, And a combination of software.

FIG. 4 illustrates a pattern classification apparatus for a functional magnetic resonance image according to an embodiment of the present invention.

4, the fMRI pattern classification apparatus 100 includes a brain state signal acquisition unit 110, a preprocessor 120, a feature point extraction unit 130, a normalization unit 140, an SVM classification unit 150, an output correction unit 160, and a classification / discrimination unit 170. The brain state signal acquisition unit 110 acquires a brain state signal indicating an activated state of the brain as a two-dimensional and three-dimensional image through an MR scanner (not shown) of the fMRI. Here, since the three-dimensional image can be obtained by combining two-dimensional images, it is assumed that the brain state signal acquisition unit 110 acquires a brain state signal composed of two-dimensional images in the present invention. The brain state signal acquisition unit 110 may not acquire a brain state signal photographed by the MR scanner directly but may receive a previously captured and stored brain state signal. The previously captured and stored brain state signal can be used as data for machine learning of the pattern classifying apparatus 100.

The preprocessing unit 120 performs a spatial alignment operation to reduce an error caused by head movement of the subject in the brain state signal acquired by the brain state signal acquisition unit 110, A preprocessing operation is performed to smoothing the signal-to-noise ratio of the voxels. The spatial alignment and smoothing operations are well known in the art and will not be described in detail here.

When the preprocessing unit 120 completes the preprocessing operation on the brain state signal, the feature point extracting unit 130 extracts feature points from the brain state signal. The feature point extraction unit 130 divides the previously stored high-resolution T1-weighted MRI into regions and aligns the images of the divided regions with brain state signals applied by the preprocessing unit 120 . Then, using the matched image as a mask, voxels of region of interest (ROI) are extracted as feature points in brain signal. Here, T1-weighted images are acquired quickly by shortening the repetition time (TR) and echo time (TE) of fMRI.

Although all the extracted feature points can be used for machine learning of the pattern classification apparatus 100, in this case, since the number of feature points to be analyzed is large, pattern classification efficiency and accuracy are low and real-time processing is difficult. The feature point selection operation is performed to select feature points that are important to the brain state analysis among the extracted feature points. Feature point selection may be performed based on a known detection algorithm (searchlight) or an ANOVA algorithm.

The feature point extraction unit 130 extracts a feature vector X _t = (X ₁ , ..., X _d ) _t corresponding to each brain state signal t with feature points of d (where d is a natural number) .

The normalization unit 140 performs a normalization operation to subtract the average of the minutiae points from each feature vector to obtain a normalized feature vector x ', as shown in Equation (1), to remove the floating noise.

The SVM classifier 150 acquires a linear discriminant function having a maximum margin for the feature vector x 'using a linear support vector machine (SVM) classification technique. Here, the linear discriminant function is obtained in the form of Equation (2).

(Where w is the weight vector of the decision hyperplane, b is a bias value for moving the decision hyperplane of the feature point from the origin. Characteristic j of the vector x _'j is a natural number as the sequence of brain state signal.)

Since the feature vector x ' _j is a value already obtained in the normalization unit 140 and the feature vector x' _j for the brain state signal collected for machine learning is subjected to pattern classification in advance, (150) can easily obtain the linear discriminant function f (x ' _j ) in which the margin is maximized among the classified patterns from the distances of the feature values of the classified pattern. Then, a weight vector w and a deflection value b are derived from the obtained linear discriminant function f (x ' _j ).

The SVM classification scheme is a classification scheme designed to maximize the generalization ability by maximizing the margin, unlike the conventional classification scheme for minimizing the classification error rate. The SVM classification technique obtains a hyperplane of the feature that divides the feature points into two regions with the largest margin in the form of a linear discriminant function (f (x ' _j )) as shown in Equation (2). Then, the machine learning is performed by obtaining the weight vector w and the deflection value b in the derived linear discriminant function f (x ' _j ). And substituting, for _(j linear discriminant function (f (x in equation (2) the feature vectors x) 'to be applied to the subsequent pattern classification to _j)), and the feature vectors linearly (x' _j) is assigned to determine the function (f (x ' _j )) has a positive value or a negative value. That is, a plurality of feature points of the feature vector x 'are binarized into feature points included in one side and the other side with respect to the crystal hyperplane.

If the SVM classifier 150 performs multi-category classification, if there are M categories to be classified, the SMV classifier 150 applies the rest of the approach. When using the remainder approach, the weight vector w and the deflection value b are obtained in M number of ways. Here, each of the M linear discriminant functions f (x ') to which the M weight vectors (w) and the deflection values (b) are applied can be regarded as an individual classifier.

The rest of the approach is to classify the categories into two categories, one for the category and the other for the classification, even if there are multiple categories of patterns to classify. That is, the multi-category classification is replaced with the binary classification method.

)을 계산한다.The SVM classifier 150 substitutes the feature vector x 'into M classifiers to which M weight vectors w and b are applied and outputs M classifier output values

).

The output correction unit 160 then obtains the slope alpha and the offset beta using the least squares method for each of the output values of the M classifiers, The flood noise is linearly removed from the output of the classifier by searching the flood noise and subtracting the flood noise searched from the output value of the corresponding classifier.

For example, the output correction unit 160 may linearly remove the floating noise of the output value f _c (x ' _t ) of the classifier for classifying the c category (where c is a natural number equal to or less than M) among the M categories (3) is used.

(Where g _c (x ' _t ) is the corrected feature vector for category _c, and α _c and β _c are the slope and offset for category _c , respectively).

In Equation (3), α _c and β _c can be calculated by equations (4) and (5) according to the least squares method, respectively.

The classifying / discriminating unit 170 compares the output values of the M classifiers from which the noise noise has been removed by the output unit 160 to determine a classifier having the maximum output value, and outputs the characteristic vector x to a pattern corresponding to the discriminated classifier .

The classification discriminating section 170 classifies the pattern according to the expression (6).

(Where M _n means the n th classifier of M classifiers).

In other words, according to Equation (6), the classification discrimination unit 170 classifies patterns by acquiring a classifier having a maximum of the corrected feature vectors g _c (x ' _t ) among M classifiers.

When the SVM classifier 150 performs the multi-category classification, the classification discriminator 170 discriminates the classifier having the maximum output value and classifies the pattern into a plurality of binary classifiers, ') Are classified into categories. In other words, since the feature vector x 'is included in a plurality of categories by the plurality of binary classifiers, the most suitable one of the patterns must be discriminated. The class discrimination unit 170 analyzes the classifier having the maximum output value, And determines one pattern in which the vector (x ') should be included.

In this case, the normalization unit 140 for removing the floating noise by the normalization technique removes the linear floating noise from the output correction unit 160, and thus may be omitted. However, in order to increase the accuracy of the pattern classification, the flooding noise should be removed as much as possible, so that the normalization unit 140 is included in the present invention.

FIG. 5 illustrates a pattern classification method of a functional magnetic resonance image according to an embodiment of the present invention.

Referring to FIG. 4, the fMRI pattern classification method of FIG. 5 will be described. First, the brain state signal acquisition unit 110 acquires a brain state signal (S10). The brain state signal obtained here is a brain state signal previously stored for the machine learning of the pattern classifier, and a pattern to be classified is designated.

When the brain state signal is obtained, the preprocessing unit 120 performs a preprocessing operation on the obtained brain state signal (S20). The pre-processing may include smoothing to increase the signal-to-noise ratio of a plurality of voxels in a spatially aligned and aligned brain state signal to reduce an error between brain state signals due to head movement of a subject in a plurality of brain state signals .

The feature point extraction unit 130 extracts feature points from the preprocessed brain state signal (S30). The feature point extraction is performed by matching the T1 - weighted brain structure image with the brain state signal and extracting the voxels of the region of interest as feature points using the image of the matched brain state signal as a mask. Then, feature vectors (x _t = (x ₁ , ..., x _d ) _t ) are obtained by applying the detection algorithm or the analysis algorithm to the extracted feature points.

)을 계산한다(S60).The normalization unit 140 obtains the normalized feature vector x 'by subtracting the average of the feature points from the obtained feature vector x _t = (x ₁ , ..., x _d ) _{t as} shown in Equation (1). Thereafter, the SVM classifier 150 classifies the feature points of the normalized feature vector x 'according to a predetermined predetermined number (here, M) of categories, and sets a linear discriminant function Is obtained by the linear SVM technique (S50). Here, if M is a plurality of categories, the SVM classifier 150 acquires M linear discriminant functions using the one-to-many residual approach. Then, each of the M linear discriminant functions obtained is used as M classifiers which are used as criteria for pattern classification. Then, the SVM classifier 150 substitutes the normalized feature vector x 'for each of the obtained M classifiers and outputs the output value of the classifier

(Step S60).

The output correction unit 160 linearly removes the floating noise with respect to each of the output values calculated in each of the M classifiers, and corrects it (S70). Here, the correction of the output value is performed by obtaining a slope (α) and an offset (β) by using a least squares method on the output value of the classifier, and searching for the time-dependent floating noise using the obtained slope (α) , And subtracting the detected floating noise from the output value of the classifier.

When the output value is corrected by the output correction unit 160, the classification determination unit 170 determines the maximum output value among the M output values output from each of the M classifiers (S80). Then, the pattern is classified into a category corresponding to the classifier in which the obtained brain state signal is output as the maximum output value discriminated among M classifiers (S90).

FIG. 5 is a machine learning method for a fMRI pattern classifier substantially. In the case of classifying a pattern of brain state signals applied after the end of the machine learning, except for the linear discriminant function acquisition step (S50) in the method of FIG. 5, By performing the same procedure for the step, the pattern can be classified. Here, the linear discriminant function acquisition step (S50) is excluded because the linear discriminant function has already been obtained in the machine learning method.

As described above, in the present invention, not only the performance of the pattern classifier is improved by removing the floating phenomenon of the output value of the multi-category pattern classifier at the time of pattern classification of the fMRI, but also the multi-category classification is replaced with a plurality of binary classification methods, And classifies the pattern of the brain state signal into a category corresponding to the classifier that outputs the maximum output value. Therefore, stable pattern classification can be performed with high accuracy.

The method according to the present invention can be implemented as a computer-readable code on a computer-readable recording medium. A computer-readable recording medium includes all kinds of recording apparatuses in which data that can be read by a computer system is stored. Examples of the recording medium include a ROM, a RAM, a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, and the like, and a carrier wave (for example, transmission via the Internet). The computer-readable recording medium may also be distributed over a networked computer system so that computer readable code can be stored and executed in a distributed manner.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is evident that many alternatives, modifications and variations will be apparent to those skilled in the art.

Accordingly, the true scope of the present invention should be determined by the technical idea of the appended claims.

Claims (12)

A brain state signal acquisition unit for acquiring a plurality of brain state signals in which categories of patterns are classified in advance for machine learning of the pattern classification apparatus;
A feature point extracting unit for extracting a plurality of feature points from each of the plurality of brain state signals and obtaining a plurality of feature vectors;
A normalization unit for subtracting an average of the feature points included in each feature vector from each of the plurality of feature vectors to remove floating noise;
At least one linear discriminant function having a maximum margin for each of the plurality of feature vectors is obtained as at least one classifier using an SVM classification technique and the plurality of feature vectors are substituted into the obtained at least one classifier An SVM classifier for calculating a classifier output value;
An output correction unit for removing time-series floating noise from at least one of the classifier output values; And
A classification discriminating unit for analyzing the classifier output value corrected by the output correcting unit and classifying the pattern of the brain condition signal into a corresponding category; And the fMRI pattern classification device. The apparatus of claim 1, wherein the SVM classification unit
Wherein the linear discriminant function is obtained by a number corresponding to the number of categories to be classified and a plurality of linear discriminant functions are obtained using a one-to-many residual approach if the number of categories to be classified is plural Pattern classification device. 3. The apparatus of claim 2, wherein the classification /
Wherein when the number of categories to be categorized is plural, the plurality of corrected classifier output values are compared to analyze the classifier output value having a maximum value, and the classifier corresponding to the classifier corresponding to the classifier output value having the maximum value, And classifying the pattern of the status signal. 2. The apparatus of claim 1, wherein the output correction unit
Obtaining a slope and an offset by using a least squares method for each of the at least one classifier output values, searching for the time-series floating noise using the obtained slope and the offset, And removing the floating noise by subtracting the noise from the fMRI pattern. The apparatus of claim 1, wherein the feature point extracting unit
Wherein the method comprises the steps of: dividing a predetermined brain structure image by predetermined regions, matching an image of the divided region with the brain state signal, and extracting the feature points of the brain state signal using the matched image as a mask fMRI pattern classification device. The apparatus of claim 1, wherein the fMRI pattern classification apparatus
A pre-processing unit for performing spatial alignment and smoothing on the plurality of brain state signals acquired by the brain state signal acquisition unit and transmitting the spatial state alignment and smoothing to the feature point extraction unit; Wherein the fMRI pattern classification apparatus further comprises: A fMRI pattern classification method of an fMRI pattern classification apparatus including a brain state signal acquisition unit, a feature point extraction unit, a normalization unit, an SVM classification unit, an output correction unit, and a classification discrimination unit,
Obtaining a plurality of brain state signals in which the category of the pattern is classified in advance for the machine learning of the pattern classification apparatus;
Extracting a plurality of feature points from each of the plurality of brain state signals to obtain a plurality of feature vectors;
Normalizing the normalization by subtracting the average of the feature points included in each feature vector from each of the plurality of feature vectors by removing the floating noise;
Acquiring, as at least one classifier, at least one linear discriminant function having a maximum margin for each of the plurality of feature vectors using the SVM classifier;
Calculating a classifier output value by substituting the plurality of feature vectors into the at least one classifier obtained by the SVM classifier;
The output correction unit corrects the classifier output value by removing time series floating noise from at least one classifier output value; And
Analyzing the classification-corrected output of the classifier to classify the pattern of the brain condition signal into a corresponding category; / RTI > a method for classifying an fMRI pattern. 8. The method of claim 7, wherein obtaining the feature vector comprises:
Dividing a predetermined brain structure image into predetermined areas;
Matching an image of the divided region with the brain state signal;
Extracting feature points of the brain state signal using the matched image as a mask;
Selecting some of the extracted minutiae using one of a predetermined detection algorithm or a variance analysis algorithm; And
Obtaining the selected feature points with the feature vector; Wherein the fMRI pattern classification method comprises: 8. The method of claim 7, wherein acquiring as the classifier comprises:
And if the number of categories to be categorized is plural, a plurality of the linear discriminant functions are obtained using the one-to-many residual approach. 10. The method of claim 9, wherein the correcting the classifier output value comprises:
Comparing the plurality of corrected classifier output values to analyze the classifier output value having a maximum value if the number of categories to be classified is plural; And
Classifying the pattern of the brain condition signal into a category corresponding to the classifier corresponding to the classifier output value having the maximum value; Wherein the fMRI pattern classification method comprises: 8. The method of claim 7, wherein the step of correcting the output value comprises:
Obtaining a slope and an offset using a least squares method for each of the at least one classifier output values;
Searching the time series floating noise using the obtained slope and the offset; And
Subtracting the time-lapse floating noise detected in the classifier output value; Wherein the fMRI pattern classification method comprises: 8. The method of claim 7, wherein the fMRI pattern classification method comprises:
A pre-processing step of spatially aligning the plurality of brain state signals acquired by the brain state signal acquisition unit and smoothing to increase a signal-to-noise ratio of a plurality of voxels included in each of the plurality of spatially aligned brain state signals; The method of claim 1, further comprising:

Images