KR20200124827A - Alzheimer's disease classification based on multi-feature fusion - Google Patents

Abstract

Disclosed is a method for classifying Alzheimer′s disease based on multi-feature fusion, which combines morphological and texture features as diagnostic biomarkers of magnetic resonance imaging and using machine learning techniques so as to classify Alzheimer′s disease, mild cognitive impairment, and normal people. This classification method comprises: an MR image acquisition step of acquiring a scanned magnetic resonance (MR) image; an image processing step of performing noise removal, non-homogeneous correction, contrast normalization, and patch-based hippocampal region segmentation for the MR image; a feature extraction step of extracting volume factor-based morphological features and extracting texture features using a gray-level co-occurrence matrix (GLCM) and a gabor filter for the hippocampus image segmented in the image processing step; a feature selection step of selecting and classifying a certain number of feature sets from feature vectors of morphological and texture features extracted from the feature extraction step; and a result analysis step of calculating a difference among Alzheimer′s disease, mild cognitive impairment, and normal group by using machine learning and measuring accuracy of classification.

Description

Alzheimer's disease classification method based on multiple trait fusion {ALZHEIMER'S DISEASE CLASSIFICATION BASED ON MULTI-FEATURE FUSION}

The present invention relates to a method for classifying Alzheimer's disease, and more particularly, combining morphological characteristics and texture characteristics as diagnostic biomarkers of MR images, and classifying Alzheimer's disease, mild cognitive impairment, and normal people using machine learning techniques. The present invention relates to a method for classifying Alzheimer's disease based on risk multi-characteristic fusion.

In the past decades, many methods have been studied to quantify disease-dependent hippocampal atrophy using magnetic resonance imaging (MRI). In particular, much effort has been made for early detection and prediction of Alzheimer's Disease (AD) and Mild Cognitive Impairment (MCI).

AD and MCI are progressive neurodegenerative disorders in which memory, cognitive and ability to perform daily tasks are slowly destroyed. The main symptoms of AD are amnesia, impaired planning and problem solving, visual cognitive impairment, inappropriate social behavior, and depression. AD and MCI are the most common forms of dementia in the elderly and cause many problems in terms of social health.

Another characteristic of AD and MCI is that beta-amyloid and tau protein, which are made up of proteins, accumulate inside/outside of neurons and damage neurons.

Accordingly, brain MRI is a medical imaging technology used to evaluate the impression of AD and MCI patients, and can be effectively used in the evaluation stage because it can observe changes in areas such as the hippocampus and the Entorhinal Cortex.

These changes can be used as biomarkers for early detection of neurodegenerative diseases. Also, in order to efficiently diagnose a large amount of medical images, an excellent diagnosis method is essential.

However, in previous studies, AD classification was not used in combination with visual and morphological features.

For this, a technology for efficiently diagnosing images based on morphology and visual content is required.

As the background technology of the present invention, a brain disease diagnosis service apparatus and method are known in Korean Patent Publication No. 10-2015-0057045 (2015.05.28.).

However, the conventional technique is that when a brain lesion is extracted using images acquired from an MRI device, an individual brain region of the subject to be diagnosed is generated by applying the image of the subject to be diagnosed and a transform matrix estimated using the standard image to a standard brain region map. A brain disease diagnosis service apparatus and method capable of determining the type of disorder based on the result of calculating the proportion of brain lesions for each brain region on the map, and the texture characteristics of the hippocampus of the AD patient as in the present invention. It is different from the technology that uses morphological characteristics by fusion, classifies them, and measures accuracy.

The present invention was derived to solve the problems of the prior art, and an object of the present invention is to combine the morphological characteristics of the extracted hippocampal region of the MRI image and the texture characteristics in 2D and 3D, and use a machine learning technique. Using Alzheimer's Disease (AD), Mild Cognitive Impairment (MCI), and Normal Control (NC), a multi-characteristic fusion-based Alzheimer's disease classification method can be classified.

Another object of the present invention is to provide a multi-characteristic fusion-based Alzheimer's disease classification method capable of classifying AD, MCI and NC and measuring accuracy based on texture analysis and morphological characteristics of an image.

Alzheimer's disease classification method according to an aspect of the present invention for solving the above technical problem, MR image acquisition step of obtaining a magnetic resonance imaging (MRI, Magnetic Resonance Imaging) scan image; An image processing step of performing noise removal, non-homogeneity correction, contrast normalization, and patch-based hippocampal region segmentation for the MR image; A feature extraction step of performing volume factor-based morphological feature extraction on the divided hippocampal image and texture feature extraction using a gray-level co-occurrence matrix (GLCM) and a Gabor filter; A feature selection step of selecting and classifying a predetermined number of feature sets from feature vectors extracted from morphological features and texture feature vectors extracted in the feature extraction step; And a result analysis step of training differences between Alzheimer's disease (AD), mild cognitive impairment (MCI) and normal people (NC) groups using machine learning and measuring classification accuracy. have.

In addition, the image processing step of the present invention is an intensity non-uniformity correction step of correcting the local entropy (Local Entropy bicubic Spline Models)-based local entropy (LEMS) of the MR image to a certain level or less, and the standard brain It may include a linear registration step for the ICBM152 template matching the template and a patch-based hippocampal region segmentation step of segmenting the hippocampus region of the MR image into 21 right and 21 left regions based on a patch.

In addition, the feature extraction step of the present invention includes the step of extracting morphological features for the volume and surface using the divided hippocampal image, and the texture using the Gray-Level Co-occurrence Matrix (GLCM) method and the Gabor Filter method. It may be characterized by comprising performing a feature extraction step.

In addition, the characteristic selection step of the present invention includes the volume, area, sum average, cluster shade, cluster tendency, and local energy of the image. Fisher's correlation coefficient analysis step for selecting six features and three types of Gabor texture analysis, hippocampus morphometric, and 2D and 3D gray level co-occurrence determinants (GLCM) for the six features. It may consist of a feature set classification step of classifying the features identified from the MR image using.

In addition, the result analysis step of the present invention trains differences between groups by using a supervised learning-based support vector machine (SVM), a kind of machine learning, and Gabor texture analysis. ), a morphological hippocampus (hippocampus morphometric), and a 2D and 3D gray level co-occurrence determinant (GLCM).

In an MR image-based Alzheimer's disease classification apparatus according to another aspect of the present invention for solving the above technical problem, an MR image acquisition unit for acquiring a magnetic resonance imaging (MRI) scan image; An image processing unit that performs noise removal, non-homogeneity correction, contrast normalization, and patch-based hippocampal region segmentation for the MR image; A feature extraction unit that extracts morphological features based on volume elements and texture features using a gray-level co-occurrence matrix (GLCM) and a Gabor filter for the divided hippocampal image; A feature selector for selecting a predetermined number of feature sets from feature vectors extracted from morphological features and texture feature vectors extracted in the feature extraction step through Fisher discriminant analysis; And a result analysis unit measuring classification accuracy between Alzheimer's disease (AD), mild cognitive impairment (MCI) and normal people (NC) groups using a support vector machine (SVM).

Alzheimer's disease classification method with improved performance by independently extracting features from texture features and morphological features in 3DGLCM and selecting features by combining them with the above-described method of classifying Alzheimer's disease based on multi-property fusion. It is possible to provide.

According to the present invention, by additionally using a patch-based texture analysis system instead of the whole brain data in the algorithm, classification can be performed by focusing only around the hippocampus according to the patch-based strategy, whereby the hippocampus and Alzheimer's dementia ( AD: Alzheimer's Disease) and Mild Cognitive Impairment (MCI) have the advantage of being able to effectively select different trait values according to the differences between groups.

In addition, according to the present invention, AD diagnosis by combining the morphological characteristics of the hippocampal region extracted from the MRI image and the texture characteristics in 2D and 3D, and classifying AD, MCI, and normal person (NC) using a machine learning technique. In the process for hippocampal detection and analysis, it has the effect of obtaining high accuracy.

1 is a flowchart of a method for classifying Alzheimer's disease based on multi-characteristic fusion according to an embodiment of the present invention.
FIG. 2 is an exemplary view showing a corrected image obtained by performing image non-uniformity correction using LEMS according to an embodiment of the present invention.
3 is an exemplary diagram showing a pre-processing step for an MR image according to an embodiment of the present invention,
4 is an exemplary diagram of AD-NC classification performance classification of 5-fold cross-validation according to an embodiment of the present invention;
5 is an exemplary diagram of MCI-NC performance classification of 5-fold cross-validation according to an embodiment of the present invention;
6 is an exemplary diagram showing the AD-MCI performance classification of 5-fold cross-validation according to an embodiment of the present invention,
7 is a block diagram showing the configuration of an Alzheimer's disease classification apparatus based on multi-characteristic fusion according to an embodiment of the present invention.

In describing the present invention, when it is determined that a detailed description of a related known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted. Even if the terms are the same, if the displayed parts are different, the reference numerals do not coincide.

In addition, terms to be described later are terms set in consideration of functions in the present invention, and since they may vary according to the intention or custom of operators such as experimenters and measurers, their definitions should be made based on the contents throughout the present specification.

The terms used in the present specification are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs. Terms as defined in a commonly used dictionary should be interpreted as having substantially the same meaning as the meaning in the context of the related technology, and unless explicitly defined in this application, interpreted as an ideal or excessively formal meaning. It doesn't.

In addition, when a part "includes" a certain component, it means that other components may be further included rather than excluding other components unless specifically stated to the contrary.

Accordingly, in the present invention, various transformations may be applied and various embodiments may be provided, and specific embodiments will be illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the present invention to a specific embodiment, it should be understood to include all conversions, equivalents, or substitutes included in the spirit and scope of the present invention. In describing the present invention, when it is determined that a detailed description of a related known technology may obscure the subject matter of the present invention, a detailed description thereof will be omitted.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a flow chart of a method for classifying Alzheimer's disease based on multi-characteristic fusion according to an embodiment of the present invention.

Referring to Figure 1, the Alzheimer's disease classification method of the present invention includes an MR image acquisition step (S100), an image processing step (S200), a feature extraction step (S300), a feature selection step (S400), and a result analysis step (S500). It can be made including.

In the MR imaging step S100, a magnetic resonance imaging (MRI) scan image (hereinafter, also referred to as “MR image”) is acquired. This step S100 corresponds to the step of receiving the MR image or may be omitted when using an MR image input from a separate scanning device or a database.

In the image processing step S200, noise removal, non-homogeneity correction and contrast normalization, patch-based segmentation, and hippocampal region segmentation are performed on the MR image.

In the feature extraction step (S300), volume factor-based morphological features are extracted from the hippocampus image divided in the image processing step, and texture features are extracted using a gray-level co-occurrence matrix (GLCM) and a Gabor filter. Perform.

In the feature selection step S400, a predetermined number of feature sets are selected and classified from feature vectors of morphological features and texture features extracted in the feature extraction step.

In the result analysis step (S500), differences between Alzheimer's disease (AD), mild cognitive impairment (MCI) and normal people (NC) groups are calculated using machine learning, and classification accuracy is measured.

A more detailed description of the Alzheimer's disease classification method based on multi-characteristic fusion is as follows.

First, the data used in the present invention was the Alzheimer's Disease Neuroimaging Initiative Database (ADNI).

All subjects performed a protocol consisting of an interview and a comprehensive cognitive evaluation, and the evaluation included a physical examination, an MRI brain scan, and a blood test.

In addition, participants were randomly selected from the subject database. All subjects were aged 60 or older and were subjectively classified into AD, MCI and NC categories by neurologists, radiologists and psychiatrists. 150 final subjects were selected from the ADNI-1 database, and details of the subjects used in the present invention are shown in Table 1 below. Table 1 is a table showing information on a subject in the present invention.

AD MCI NC Sample size 50 50 50 Age(years, mean ±SD) 71.3 ± 2.9 69.5 ± 3.0 69.5 ± 3.0 Gender(Male: Female) 25:25 22:28 25/25

The MR image acquisition step (S100) of the present invention is a step of acquiring a magnetic resonance imaging (MRI) scan image. This is an MR scan image of 150 subjects selected from the ADNI-1 database, 1.5 Tesla of TE/TI/TR (2.98/900/2300 ms), and was performed using a T1-scanner, and slice thickness Is 1.2 mm, and the size of rows, columns and slices is 256 × 256 × 166, respectively. The voxel size is 1.0 × 1.0 × 1.2 mm 3.

The image processing step (S200) is a pre-processing of the MR image, which includes a noise removal and intensity non-uniformity correction step (S210), an ICBM152 template matching step, a patch-based segmentation step (S220), and a hippocampal region segmentation step (S230). Step.

The noise removal and intensity non-uniformity correction step (S210) is a step of removing noise and correcting non-uniformity in an image, and the International Consortium for Brain Mapping (ICBM) 152 template (1×1×1 mm³) for all MR images It may be a pretreatment process that minimizes local entropy based on Local Entropy bicubic Spline Models (LEMS) for inhomogeneity and consistency.

In the patch-based segmentation step (S220), the MR image is segmented based on a preset patch. In addition, the hippocampal region subdividing step S230 is a step of subdividing the hippocampal region in the MR image, and may be performed as an image processing step for morphological analysis and morphological feature extraction.

Detailed configuration of the subsequent steps will be described later with reference to FIGS. 2 to 6.

FIG. 2 is an exemplary view showing a corrected image obtained by performing image non-uniformity correction using the LEMS according to an embodiment of the present invention, where (a) is an original image, (b) is a non-uniform image, and (c) Shows the unevenly corrected image.

Subsequently, the MNI-ICBM152 template matching step is a step of matching the subject's image with the standard brain templates of 152 people of the Montreal Neurological Institute (MNI) and The International Consortium for Brain Mapping (ICBM).

Additionally, before dividing the object of interest in the MR image, a step of normalizing the contrast with other images may be performed.

Thereafter, the patch-based hippocampal region segmentation step (S220) is a step of dividing 42 hippocampal regions (right: 21, left: 21) in the MR image based on a patch.

All pixels in the patch in the present invention are used for texture analysis, and the geometric regions of the left and right hippocampus can be set manually.

3 is an exemplary diagram showing a pre-processing step for an MR image according to an embodiment of the present invention, and shows a pre-processing workflow used for segmenting a hippocampal patch.

In the present embodiment, the feature extraction step (S300) includes a volume and surface morphological feature extraction step (S310) using a segmented hippocampal image, a gray-level co-occurrence matrix (GLCM) method, and a Gabor filter method. It may include a texture feature extraction step (S320) using.

The morphological feature extraction step (S310) is a step of extracting characteristic values for the volume and surface of the hippocampal image segmented in the patch-based hippocampal region segmentation step (S230). In the morphological feature extraction step (S320), the morphological features are extracted from the hippocampus image segmented through the morphological analysis step (S310) using a volume factor-based morphological measurement (VBM) technology.

Volume factor-based morphological measurement (VBM) is a brain image morphology measurement technology that statistically measures local brain deformation and projects it into a three-dimensional space, and volume factor-based morphology is very practical in the field of brain medical imaging. In the invention, it is also used to obtain characteristic values by performing morphological analysis on the region of interest.

Through this ROI-based hippocampal data analysis, useful information such as volume, surface shape analysis, and structural characteristics such as tissue thickness can be obtained, and as an advantage, point-to-point analysis for each model is possible.

In the GLCM (Gray-Level Co-occurrence Matrix) texture characteristic analysis method and the texture characteristic extraction step (S340) using Gabor texture analysis, MR for all pixels in the patch analyzed in the texture analysis step (S330). You can extract the texture characteristics of the image.

The GLCM is also referred to as a gray level co-occurrence determinant or a contrast co-occurrence matrix, and the Gabor filter and the gray level co-occurrence determinant are one of widely used texture feature extraction methods.

The GLCM texture characteristic analysis method will be described in more detail as follows.

First, the GLCM texture characteristic analysis method is a texture analysis method based on the spatial dependence of the gray level, and generates a texture matrix expression for an image by using the distance between pixels and the direction between pixels as parameters.

At this time, the 2D texture characteristics are calculated using all the pixels of each slice.

On the other hand, if 3D volume data is processed as a 2D slice, information between slices is ignored and there is a possibility of data loss. 3D texture analysis was applied to solve this problem. Despite the simplicity of extending the existing determinant-based algorithm to 3D, this approach can show significantly improved results.

2DGLCM considers the spatial dependence of each slice pixel in the data set, while 3DGLCM quantifies the 3D spatial dependence of volumetric element data on the volume of entities that exist across multiple slices.

3DGLCM can be expressed as n×n×m, where n is the gray level (gray level or light and dark level), and m is the number of slices.

In the GLCM calculation of the present invention, a direction and a distance between pixels may be set as parameters.

In addition, in the existing 2DGLCM, a total of 8 directions were used, but in the present invention, texture characteristics obtained by applying GLCM in 4 directions of 0°, 45°, 90°, and 135° are calculated and normalized to avoid overlapping. I did. A total of 26 directions can be calculated in 3DGLCM, but 13 directions can be used to prevent redundant calculations.

Accordingly, a new determinant can be generated by calculating a matrix for each data set and then calculating the average. Based on this, 19 GLCM texture characteristics can be extracted from 2DGLCM and 3DGLCM.

Gabor texture analysis was developed by Gabor in 1946 and refers to the analysis of texture characteristics using the Gabor Filter, one of the most well-known texture analysis methods. Gabor filter is basically a Gaussian function modulated by a complex sine wave, and is used to extract features of the domain of an image. In addition, the Gabor filter can be performed in the spatial and frequency domains of several dimensions. The Gabor filter can be used to determine the optimal joint position in the spatial and frequency domains.

In the present invention, texture characteristics may be extracted using 2D Gabor filters having different frequencies and directions. In this case, the Gabor filter may decompose the image into constituent elements in different proportions and directions. Through this, in the system of the present embodiment, visual characteristics according to spatial position, direction selection, and spatial frequency can be captured.

That is, in this embodiment, the MR image is filtered by a plurality of Gabor filters, and the feature vector may be composed of an absolute value of the response.

In summary, if the given image is (x, y), the 2D Gabor filter bank can be defined by Equation 1 as follows.

는 담체(carrier)라고 불리는 복잡한 형태의 사인곡선이고,

는 엔베로프로 불리며 2-D 가우시안 형태의 함수이다. 여기서 2D 가버필터는 주파수 및 방향의 복잡한 사인파에 의해 변조된 가우스 함수로 구성되며, 다음과 같이 수학식 2로 계산될 수 있다.In Equation 1,

Is a complex sinusoidal curve called a carrier,

Is called an envelope and is a 2-D Gaussian form of function. Here, the 2D Gabor filter is composed of a Gaussian function modulated by a complex sine wave of frequency and direction, and can be calculated by Equation 2 as follows.

In Equation 2, ω _x0 and ω _y0 are center frequencies in the x and y directions, and the filter produces the largest response at this frequency (center frequency in the x and y directions. ), and σ ² _x and σ ² _y are the standard deviation of the Gaussian function in the x and y directions. x and y are the position of a pixel in the image.

과 같이 정의할 수 있으며,

이 된다. 실험에서 사용된 θ 값은

이며 중심 주파수 i 는 다음과 같이 수학식 3으로 정의될 수 있다.In the present invention, several parameters are defined for the Gabor filter. It was set as L=8, where L is the number of directions. Also, the rotation bandwidth is

Can be defined as

Becomes. The θ value used in the experiment is

And the center frequency i can be defined by Equation 3 as follows.

At this time, the size of the input image is 96×96, and a total of 16 Gabor filters are applied to each data set by applying four different directions and four frequencies. Finally, the local energy, average amplitude, and variance of the amplitude are calculated as the texture characteristics extracted by each filter. The characteristic values reflect the concentration tendency and discrete distribution of the local energy spectrum of the image.

The Gabor filter can extract features of an image using various frequencies and directions, and has an advantage of using an optimal cavity space and frequency resolution.

In the feature selection step (S400), a certain number of feature sets are selected through Fisher discriminant analysis in order to select an optimal feature from feature vectors extracted from the GLCM texture feature analysis and the Gabor filter texture feature analysis. This is the step.

The reason for taking such a characteristic selection step is that in general, a linear combination of morphological characteristics and texture characteristics may not be effective in classification, and relevance and redundancy may not be verified. Therefore, it is because a separate criterion for selecting an optimal feature from all feature vectors is required.

In this case, the feature selection step may include a Fisher correlation coefficient analysis step (S410) and a feature set classification step (S420).

First, in the Fisher correlation coefficient analysis step (S410), a Fisher discriminant analysis method is used in order to reduce data complexity in classification, and to remove redundant data with irrelevant data in the linear combination of morphological characteristics and texture characteristics. This is the step of selecting an appropriate number of characteristics.

Therefore, due to the duplication and complexity of data, it is possible to perform an experiment by selecting an appropriate number of characteristics through Fisher discriminant analysis, which is widely used in multivariate analysis, without using all the calculated characteristics.

Equation 4 below is an equation for calculating the Fisher coefficient according to the Fisher discriminant analysis approach, and in the present invention, the Fisher coefficient is based on the automated feature selection method of MaZda. (Fisher coefficient) was analyzed.

및

는 클래스 i 의 평균 및 분산,

는 클래스 i 의 확률을 뜻한다.In Equation 4, F is the Fisher coefficient, D is the variance between classes, V is the variance inside the class,

And

Is the mean and variance of class i,

Means the probability of class i.

In this embodiment, six characteristics according to Fisher coefficient can be selected through the Mazda program. The six features can be classified as Volume, Area, Sum Average, Cluster Shade, Cluster Tendency, and Local Energy of the image. have.

In addition, the optimal feature set classification step (S420) includes Gabor texture analysis, hippocampus morphometric, and 2D and 3D gray levels for the six features selected through the Fisher correlation coefficient analysis step (S410). This is the step of classifying the features identified from MR images using three types of co-occurrence determinants (GLCM). GM can mean a combination of Gabor texture analysis + hippocampus morphometric analysis.

Table 2 below is an exemplary diagram showing calculated hippocampal features according to the six features for each of the three types of categories.

3DGLCM + Gabor + Morphometric features (3DGLCM+GM) Features AD MCI NC Volume (cm ³ ) 5.13 ±0.44 5.36 ±0.30 5.48 ±0.28 Area (cm ² ) 24.82 ±1.22 26.55 ±1.54 26.85 ±1.55 Sum Average 50.35 ±5.59 50.22 ±4.30 49.92 ±4.49 Cluster Shade 2.76×10 ⁵ ±0.53×10 ⁴ 2.63×10 ⁵ ±0.47×10 ⁴ 2.59×10 ⁵ ±0.43×10 ⁴ Cluster Tendency 5.96×10 ⁸ ±0.32×10 ⁸ 5.45×10 ⁸ ±0.35×10 ⁸ 5.33×10 ⁸ ±0.29×10 ⁸ Local Energy 5.78×10 ⁵ ±0.26×10 ⁵ 6.19×10 ⁵ ±0.18×10 ⁵ 6.41×10 ⁵ ±0.32×10 ⁵ 2DGLCM + Gabor + Morphometric features (2DGLCM+GM) Volume (cm ³ ) 5.13 ±0.44 5.36 ±0.30 5.48 ±0.28 Area (cm ² ) 24.82 ±1.22 26.55 ±1.54 26.85 ±1.55 Sum Average 51.16 ±4.94 49.85 ±4.55 48.32 ±4.61 Cluster Shade 2.43×10 ⁵ ±0.49×10 ⁵ 2.37×10 ⁵ ±0.38×10 ⁵ 2.35×10 ⁵ ±0.37×10 ⁵ Cluster Tendency 5.65×10 ⁸ ±0.28×10 ⁸ 5.25×10 ⁸ ±0.21×10 ⁸ 5.03×10 ⁸ ±0.19×10 ⁸ Local Energy 4.98×10 ⁵ ±0.26×10 ⁵ 5.55×10 ⁵ ±0.18×10 ⁵ 6.13×10 ⁵ ±0.32×10 ⁵

Here, 3DGLCM is a 3-Dimensional Gray Level Co-occurrence Matrix, and 2DGLCM is a 2-Dimensional Gray Level Co-occurrence Matrix.

The result analysis step (S500) is a step of training differences between groups to be classified using a support vector classification (SVM), which is one of the supervised learning-based classification methods.

In the present invention, C-SVM (a variant of the support vector machine (SVM)) and a linear and non-linear radial basis kernel were used, and the kernel was Matlab's'libsvm' toolbox. It can be implemented using

In addition, the present invention calculates 2DGLCM and 3DGLCM using a Gabor filter for classifying subjects of Alzheimer's Disease (AD), Mild Cognitive Impairment (MCI) and Normal Control (NC), and morphological characteristics It can be combined with to compose texture characteristics. Here, 75 of the 150 data may be used as training data and 75 data may be used as verification data.

Table 3 below is a nonlinear SVM classification matrix of AD, MCI and NC groups using 2DGLCM and GM features, and Table 4 is an exemplary diagram showing nonlinear SVM classification matrix of AD, MCI and NC groups using 3DGLCM and GM features. .

AD MCI NC Total Accuracy (%) AD 19 4 2 25 76.00 MCI 4 17 4 25 68.00 NC One 3 20 25 80.00 Total/Average 24 24 26 75 74.67

AD MCI NC Total Accuracy (%) AD 20 3 2 25 80.00 MCI 3 19 3 25 76.00 NC 2 3 20 25 80.00 Total/Average 25 25 25 75 78.67

According to Tables 3 and 4, it can be seen that when the texture characteristics of 2DGLCM and 3DGLCM are combined with Gabor and morphological characteristics, classification accuracy of 74.67% and 78.67% is achieved. As a result, it was found that the 3DGLCM method provides better classification accuracy than the 2DGLCM method.

In addition, through binary classification, four items can be calculated: True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN). Through this, the following five indicators can be calculated.

(I)

(Ⅱ)

(Ⅲ)

(Ⅳ)

(Ⅴ)

The above five indicators can be viewed as accuracy, sensitivity, specificity, positive predictive value, and negative predictive value.

Table 5 below is an exemplary view showing the performance evaluation of five indicators for the classification of the combining method of 2DGLCM and GM and the combining method of 3DGLCM and GM according to an embodiment of the present invention.

Methods AD-NC (%) ACC SEN SPEC PPV NPV 2DGLCM+GM 81.05 ± 1.34 83.33 ± 2.15 80.95 ± 1.52 78.95 ± 1.34 85.00 ± 1.42 3DGLCM+GM 86.61 ± 1.25 84.21 ± 1.42 85.00 ± 1.24 84.21 ± 2.08 85.00 ± 1.14 MCI-NC (%) 2DGLCM+GM 76.92 ± 1.27 77.78 ± 1.31 76.19 ± 1.42 73.68 ± 1.18 80.00 ± 1.32 3DGLCM+GM 82.05 ± 1.26 83.33 ± 1.42 80.95 ± 1.36 78.94 ± 1.04 85.00 ± 1.16 AD-MCI (%) 2DGLCM+GM 76.31 ± 2.18 75.00 ± 1.34 77.78 ± 1.26 78.95 ± 1.14 73.68 ± 1.04 3DGLCM+GM 78.95 ± 2.36 76.19 ± 1.84 82.35 ± 1.34 84.21 ± 1.62 73.68 ± 1.48

Referring to Table 5, it can be seen that when 3DGLCM and GM methods are used together through 5-fold cross-validation, a classification accuracy of 86.61% was achieved in terms of accuracy.

In this example, the data set was divided into five classifications and 5-fold cross validation was performed, and accuracy, sensitivity, specificities, and positive results obtained from 10 runs were performed. Positive predictive and negative predictive are as shown in FIGS. 4 to 6.

Figure 4 is a 5-fold cross-validation AD-NC classification performance classification, Figure 5 is a 5-fold cross-validation MCI-NC performance classification and Figure 6 is an exemplary diagram showing the AD-MCI performance classification of 5-fold cross-validation , It can be seen that FIGS. 4 to 6 compare and show the performance when classified AD-NC, MCI-NC, and AD-MCI using 2DGLCM+GM and 3DGLCM+GM.

For example, referring to FIG. 4, the confirmed AD versus NC (NC) group was classified showing an accuracy between 83.00% and 80.00% as a result of using the 2DGLCM+GM cross-validation method. In addition, it was obtained with a sensitivity of 83.33±2.15% and a specificity of 80.95±1.52%, and an average accuracy of 81.05±1.34% was achieved. The 3DGLCM+GM method was also classified using the same data set. As a result, AD and NC groups were correctly classified into categories between 88.50% and 84.00% by cross-validation method, achieving an average accuracy of 86.61±1.25%, sensitivity of 84.21±1.42%, and specificity of 85.00±1.24%. It can be seen that the 3DGLCM+GM method is observed to be superior to the 2-GLCM+GM method in terms of classification accuracy.

Referring to FIG. 5, the subjects of the confirmed MCI versus NC (Confirmed MCI versus NC) group were correctly classified with an accuracy between 78.0% and 73.0% by the 2DGLCM+GM cross-validation method, respectively. With a sensitivity of 77.78±1.31% and a specificity of 76.19±1.42%, the average accuracy was 76.92±1.27%. The same data set was classified using the 3DGLCM+GM method. As a result, the MCI and NC groups were correctly classified into appropriate diagnostic categories with an average estimate of 80.00% and 74.00%, respectively, using a cross-validation method. An average accuracy of 82.05±1.26% was achieved, with a corresponding sensitivity of 83.33±1.42% and a specificity of 80.95±1.36%. The MCI-NC classification results show that 3DGLCM+GM is superior to the 2DGLCM+GM method.

According to FIG. 6, subjects in the confirmed AD versus MCI (Confirmed AD versus MCI) group were correctly classified with an accuracy between 76.40% and 72.00% when using the 2DGLCM+GM cross-validation method. A sensitivity of 75.00±1.34% and a specificity of 77.78±1.26% were obtained, and an average accuracy of 76.31±2.18% was achieved. The same data set was classified using the 3DGLCM+GM method. As a result, AD and MCI groups were correctly classified into categories between 78.20% and 75.60% by cross-validation method. Here, the sensitivity was 76.19±1.84%, and the specificity was 82.35±1.34%. The average accuracy achieved 78.95±2.36%. It can be seen that the 3DGLCM+GM method acquires more accuracy than the 2DGLCM+GM method and is proven to be superior in the classification of MCI-AD.

As in the present invention, the approach used by combining multiple characteristics using three types of Gabor texture analysis, hippocampus morphometric, and 2D and 3D gray level co-occurrence determinants (GLCM) It can be seen that it is superior to the conventional technique, and the accuracy was calculated as the average value of all multiple rows, and the classification performance was all performed five times.

As a result, in the prior art, only the texture characteristics or morphological characteristics are limited to use, but it can be seen that higher accuracy can be obtained when the texture characteristics and morphological characteristics are combined and used as in the present invention.

7 is a block diagram illustrating an apparatus for classifying Alzheimer's disease based on an MR image according to an embodiment of the present invention.

Referring to FIG. 7, the apparatus 100 for classifying Alzheimer's disease based on an MR image of the present invention includes a control unit 110a and a storage unit 110b, and an MR image acquisition unit 110 and an image processing unit mounted on the control unit 110a. 120, a characteristic extraction unit 130, a characteristic selection unit 140, and a result analysis unit 150. MR image acquisition unit 110, image processing unit 120, characteristic extraction unit 130, characteristic The selection unit 140 and the result analysis unit 150 are stored in the storage unit 110b in the form of a program or software module, and functions of the corresponding module may be performed by the control unit 110a.

The MR image acquisition unit 110 acquires a magnetic resonance image scan image, and is a device capable of receiving and storing an MR scan image of a subject selected from the ADNI-1 database.

The image processing unit 120 is a device that removes noise, corrects inhomogeneity, standardizes contrast, and subdivides a patch-based hippocampus region for an MR image. You can run a workflow that includes a region segmentation module.

The feature extraction unit 130 performs volume factor-based morphological feature extraction on the divided hippocampal image and texture feature extraction using a gray-level co-occurrence matrix (GLCM) and a Gabor filter. A structural feature extraction module and a texture feature extraction module may be provided, and a program environment for calculating and extracting features may be included in interworking with them.

The feature selection unit 140 selects a certain number of feature sets through Fisher discriminant analysis from feature vectors extracted from morphological features and texture feature vectors extracted from the feature extraction section.

The result analysis unit 150 performs a function of measuring the classification accuracy between groups of Alzheimer's disease (AD), mild cognitive impairment (MCI), and normal people (NC). The result analysis unit may use a support vector machine (SVM), which is one of the supervised learning-based classification methods.In the present invention, the C-SVM, which is a variant of the support vector machine (SVM), and a linear and nonlinear radial basis kernel ), and the kernel can be implemented by using the Matlab'libsvm' Toolbox program.

In addition, the Alzheimer's disease classification apparatus 100 of the present embodiment is implemented through programs and memory, input/output means, and display unit 160 processed for the present invention in a computer system environment having a microcontroller device or a central processing unit (CPU). Can be.

Although the above has been described with reference to the preferred embodiments of the present invention, those skilled in the art can variously modify and change the present invention without departing from the spirit and scope of the present invention described in the claims. You can understand.

This thesis is a basic research project conducted with the support of the National Research Foundation of Korea as a source of information (Ministry of Education) in 2015 (2015-059319)

Claims (6)

In the Alzheimer's disease classification method,
MR image acquisition step of obtaining a magnetic resonance imaging (MRI) scan image;
An image processing step of performing noise removal, non-homogeneity correction, contrast normalization, and patch-based hippocampal region segmentation on the MR image;
A feature extraction step of performing volume factor-based morphological feature extraction and texture feature extraction using a gray-level co-occurrence matrix (GLCM) and a Gabor filter on the hippocampal image segmented in the image processing step;
A feature selection step of selecting and classifying a predetermined number of feature sets from feature vectors of morphological and texture features extracted in the feature extraction step; And
Based on multi-feature fusion including a result analysis step of calculating differences between Alzheimer's disease (AD), mild cognitive impairment (MCI) and normal people (NC) groups using machine learning and measuring classification accuracy. How to classify Alzheimer's disease. The method according to claim 1,
The image processing step,
Intensity non-uniformity correction step of removing noise of MR images and correcting local entropy based on LEMS (Local Entropy bicubic Spline Models) below a certain level,
A linear registration step for the ICBM152 template that matches the standard brain template,
Alzheimer's disease classification method based on multi-characteristic fusion including a patch-based hippocampal region segmentation step of segmenting the hippocampus region of the MR image into 21 right and 21 left regions based on a patch. The method according to claim 1,
The feature extraction step includes a step of extracting morphological features for a volume and a surface using a divided hippocampal image, and a step of extracting a texture feature using a Gray-Level Co-occurrence Matrix (GLCM) method and a Gabor filter. Alzheimer's disease classification method based on fusion. The method according to claim 1,
The characteristic selection step,
Fisher's correlation coefficient analysis to select six characteristics of Volume, Area, Sum Average, Cluster Shade, Cluster Tendency, and Local Energy of the image Step and,
For the above six features, three types of Gabor texture analysis, hippocampus morphometric, and 2D and 3D gray level co-occurrence determinants (GLCM) are used to classify features identified from MR images. Alzheimer's disease classification method based on multiple feature fusion comprising a feature set classification step. The method according to claim 1,
In the result analysis step, differences between the groups are calculated using a support vector machine (SVM) based on machine learning or supervised learning, and Gabor texture analysis, morphological hippocampus Alzheimer's disease classification method based on multi-feature fusion that measures classification accuracy for three types of (hippocampus morphometric), 2D and 3D gray level co-occurrence determinants (GLCM). In the Alzheimer's disease classification apparatus based on MR image,
An MR image acquisition unit that acquires a magnetic resonance imaging (MRI) scan image;
An image processor for removing noise, correcting non-homogeneity, normalizing contrast, and subdividing a hippocampus region based on a patch on the MR image;
A feature extraction unit that extracts morphological features based on volume elements and texture features using a gray-level co-occurrence matrix (GLCM) and a Gabor filter for the divided hippocampal image;
A feature selector for selecting a certain number of feature sets from feature vectors extracted from morphological features and texture feature vectors extracted in the feature extraction step through Fisher discriminant analysis; And
Alzheimer's based on multi-characteristic fusion, characterized in that it includes a result analyzer that measures the classification accuracy between Alzheimer's disease (AD), mild cognitive impairment (MCI) and normal people (NC) groups using a support vector machine (SVM). Bottle sorting device.

Images