KR102373992B1 - Method and apparatut for alzheimer's disease classification using texture features - Google Patents

Abstract

The present invention relates to a method and apparatus for classifying Alzheimer's disease using texture features in medical image data. Identifies a region of interest including Based on the texture characteristics of the images, the spatial dependence is identified through the gray level of each image to classify the texture characteristics, and the correlation coefficient between the variables used for the texture characteristics is selected and the optimal texture for classification of each image in the data set It involves a series of steps to select a feature.

Description

Method and apparatus for classifying Alzheimer's disease using texture features

The present invention relates to Alzheimer's disease classification technology, and more particularly, to a method and apparatus for classifying Alzheimer's disease using texture features in medical image data.

Due to advances in medical technology, the world's population is steadily increasing, and the number of elderly people is increasing rapidly. This trend is expected to continue to accelerate in the coming decades, increasing the incidence of geriatric diseases and the social costs of aging.

Geriatric diseases include Alzheimer's Disease (AD), cerebral hemorrhage, and Parkinson's disease. Among them, AD is the most common degenerative brain disease. A symptom of AD is the progressive deterioration of cognitive function, including memory, that begins with the onset of the disease.

Among Alzheimer's Disease Neuroimaging Initiative (ADNI) data, mild cognitive impairments (MCI) are an intermediate stage between dementia and normal aging, where patients have lower cognitive function than people of the same age in the normal aging group, but perform daily activities. does not affect

Advances in current medical technology have led to the development of early diagnosis, which can treat some diseases, such as cancer or heart disease, through surgery or drugs. However, AD cannot be cured even if diagnosed at an early stage of the disease. It is only possible to control the rate of disease progression. Therefore, early detection and diagnosis are very important.

Magnetic resonance imaging (MRI) is a major tool used in modern medicine and research to look inside the body. Recently, MRI technology has been remarkably developed in recent years, and although accurate images can be obtained, it is difficult to evaluate diseases with only visual data.

For example, acquired magnetic resonance (MR) images are different for each patient because they all have brains of different sizes, and doctors can diagnose the same MR image differently. In order to solve these problems, objective criteria for the diagnosis of AD patients should be established.

On the other hand, prior art studies focus on showing various early detection methods and biomarkers for diagnosing the onset of AD, so there is a limitation in providing objective criteria for diagnosing AD patients.

Muhammad Ovais, Nashmia Zia, rshad Ahmad et al. Phyto-Therapeutic and Nanomedicinal Approaches to Cure Alzheimer's Disease: Present Status and Future Opportunities. Front Aging Neurosci. 2018;10.

An object of the present invention is a method and apparatus for classifying Alzheimer's disease based on MR image texture of the hippocampus, which plays an important role in learning, remembering and recognizing new things and is the organ most affected in the pathogenesis of Alzheimer's disease (AD). is to provide

Another object of the present invention is to provide an Alzheimer's disease classification method and apparatus capable of objectively diagnosing Alzheimer's disease based on image processing and texture analysis applying a gray-level co-occurrence matrix (GLCM) technology. is to provide

In a method for classifying Alzheimer's disease according to an aspect of the present invention for solving the above technical problem, each anatomical region in a magnetic resonance (MR) image is labeled using a gray scale value, thereby providing a region of interest (region of interest) including the hippocampus interest); obtaining texture characteristics through a plurality of different sub-images of each image according to a plurality of preset types by applying a filter to each image of the data set obtained in the identifying and including the region of interest; classifying the texture feature by identifying spatial dependence through the gray level of each image based on the texture feature of the plurality of sub-images; and selecting an optimal texture feature for classification of each image in the data set by selecting a correlation coefficient between the variables used for the texture feature.

In one embodiment, the step of obtaining the texture characteristic uses a Gabor filter that extracts the characteristic in the spatial and frequency domain using a Gaussian function, wherein the Gabor filter includes a wavelength for each image, a direction of an edge, At least some of the five parameters: sigma, theta, lambda, phi, and gamma for the iteration duration of an operation applied to surrounding pixels relative to the center of the image, the distance from the center, and the ratio of filter width to height is adjusted

In an embodiment, the classifying may include evaluating a relationship between the plurality of sub-images or slices corresponding thereto by applying a preset characteristic coefficient.

In an embodiment, in the classifying step, contrast, entropy, variance, correlation, sum, mean, inverse variance, cluster shading, clustering tendency, Variables can be selected including homogeneity, maximum probability, inertia, energy, or a combination thereof.

In an embodiment, the selecting may use a Fisher coefficient to select a correlation coefficient between the variables to identify an optimal feature for each image group in the data set.

In an embodiment, the Alzheimer's disease classification method further includes learning each image classified as a texture feature in the selecting. In the learning step, image classification may be performed on a training data set or a test data set through a multi-layer perceptron (MLP) model based on a Keras deep learning library. In addition, the learning step may use a rectification linear unit (ReLu) which is used as a substitute for an S-shaped function and has a slope set to zero for a negative input and is designed to avoid a fixed slope of 0 or 1.

In one embodiment, the Alzheimer's disease classification method classifies the texture feature selected in the selecting step into three classes: AD-MCI, AD-NC, and MCI-NC to determine the accuracy or reliability of image classification for a data set It further includes the step of verifying.

In an apparatus for classifying Alzheimer's disease according to another aspect of the present invention for solving the above technical problem, each anatomical region in a magnetic resonance (MR) image is labeled using a gray scale value, thereby providing a region of interest including the hippocampus. an identification unit for identifying interest); a filtering unit obtained by the identification unit and applying a filter to each image of the data set including the region of interest to obtain texture characteristics through a plurality of different sub-images of each image according to a plurality of preset types; a classification unit for classifying the texture characteristics by identifying spatial dependence through the gray level of each image based on the texture characteristics of the plurality of sub-images; and a selection unit configured to select an optimal texture feature for classification of each image in the data set by selecting a correlation coefficient between the variables used for the texture feature.

In an embodiment, the apparatus for classifying Alzheimer's disease further includes a learning unit for learning each image classified by the texture feature in the selection unit.

In an embodiment, the Alzheimer's disease classification apparatus further includes a verification unit that classifies the texture feature selected by the selection unit into three classes, AD-MCI, AD-NC, and MCI-NC, and verifies the image classification result for the data set. include

In the case of using the Alzheimer's disease classification method and apparatus using the texture characteristics described above, Alzheimer's disease is objectively diagnosed based on image processing and texture analysis applying gray-level co-occurrence matrix (GLCM) technology. can be effectively diagnosed.

In addition, according to the present invention, when compared with the results obtained using a confusion matrix, a support vector machine (SVM) and a K-nearest neighbor (KNN) classifier, magnetic resonance imaging (MRI) image processing, texture analysis, and deep learning on image data can classify Alzheimer's disease in objective criteria with relatively higher accuracy.

1 is a flowchart of a method for classifying Alzheimer's disease according to an embodiment of the present invention.
FIG. 2 is an exemplary diagram of image data that may be employed in the Alzheimer's disease classification method of FIG. 1 .
FIG. 3 is an exemplary diagram for explaining a pre-processing process for reslicing a magnetic resonance (MR) image that can be employed in the Alzheimer's disease classification method of FIG. 1 .
4 is another exemplary diagram of an image pre-processing process that may be employed in the Alzheimer's disease classification method of FIG. 1 .
FIG. 5 is an exemplary diagram for explaining an image registration process that may be employed in the Alzheimer's disease classification method of FIG. 1 .
FIG. 6 is a diagram for explaining a process of analyzing a region of interest based on the ICBM template of FIG. 5 .
7 is an exemplary diagram of a type of a two-dimensional (2D) Gabor filter that can be employed in the Alzheimer's disease classification method of FIG. 1 .
8 is a diagram showing the performance of a proposed model of deep learning that can be employed in the Alzheimer's disease classification method of FIG. 1 , and is a graph for the case of using AD-MCI data.
9 is a diagram showing the performance of a proposed model of deep learning that can be employed in the Alzheimer's disease classification method of FIG. 1 , and is a graph for the case of using AD-NC data.
10 is a diagram showing the performance of a proposed model of deep learning that can be employed in the Alzheimer's disease classification method of FIG. 1 , and is a graph for the case of using MCI-NC data.
11 is a schematic block diagram of an apparatus for classifying Alzheimer's disease according to another embodiment of the present invention.

Since the present invention can have various changes and can have various embodiments, specific embodiments are illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the present invention to specific embodiments, and it should be understood to include all modifications, equivalents, and substitutes included in the spirit and scope of the present invention. In describing each figure, like reference numerals are used for like elements.

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

1 is a flowchart of a method for classifying Alzheimer's disease according to an embodiment of the present invention.

Referring to FIG. 1 , the Alzheimer's disease classification method according to the present embodiment includes an image preprocessing step (first step, Step 1) and a texture analysis step (second step, Step 2). and may further include a deep learning step (third step, Step 3).

In the image preprocessing stage, images of datasets obtained from magnetic resonance imaging (MRI), positron emission tomography (PET), etc. and stored and used such as the ADNI dataset (alzheimer's disease neuroimaging initiative dataset) are preprocessed ( S10a, S11), after mapping the preprocessed image through a template such as an international consortium for brain mapping template (ICBM), the image may be registered through a predetermined registration system (S12).

More specifically, in the image preprocessing step, a region of interest including the hippocampus can be identified by labeling each anatomical region using a gray scale value in a magnetic resonance (MR) image. In the registered image having the region of interest, the periphery of the image may be cut around the region of interest through image cropping (S13).

In addition, the image preprocessing step may further include a step of applying a filter, for example, a Gabor filter (Gabor filter adapt, S14). In this step, a filter is applied to each image of the data set including the region of interest to obtain a plurality of different sub-images of each image according to a plurality of preset types, and texture characteristics can be obtained through the plurality of sub-images. there is.

As such, in the step S14 of obtaining the texture characteristic using the Gabor filter, a Gabor filter that extracts the characteristic in the spatial and frequency domains using a Gaussian function is used. Here, the Gabor filter corresponds to the wavelength for each image, the direction of the edge, the repetition period of the operation applied to neighboring pixels relative to the center of the image, the distance from the center, and the corresponding in the order described for each of the filter width and height ratios. It can be used in a way that adjusts at least some of the five parameters: sigma, theta, lambda, phi, and gamma.

Next, in the texture analysis step, based on the texture characteristics of the plurality of sub-images, spatial dependence may be identified through the gray level of each image to classify the texture characteristics ( S15 ). In the texture feature classification step S15, statistical texture analysis in consideration of spatial relationships between pixels of each image may be performed using a gray-level co-occurrence matrix (GLCM).

Also, in the texture analysis step, an optimal feature for each image group in the data set may be identified by using a fisher coefficient to select a correlation coefficient between variables (S16).

That is, in the texture analysis step, it is possible to select an optimal texture feature for classification of each image in the data set by selecting a correlation coefficient between variables used for the texture feature. To this end, in the texture analysis step, a relationship between a plurality of sub-images or corresponding slices may be evaluated by applying a preset characteristic coefficient. In addition, in the texture analysis step, contrast, entropy, variance, correlation, sum, mean, inverse variance, cluster shading, clustering tendency, homogeneity, maximum probability, Variables can be selected including inertia, energy, or a combination thereof.

Next, the Alzheimer's disease classification method according to the present embodiment may further include learning each image classified as a texture feature in the texture analysis step.

In the learning step, image classification may be performed on a training data set or a test data set through a multi-layer perceptron (MLP) model based on a Keras deep learning library. In addition, the learning phase may use a commutation linear unit (ReLu) which is used as a substitute for a sigmoid function and is designed to avoid fixed slopes of 0 or 1 with the slope set to zero for negative inputs.

In addition, the Alzheimer's disease classification method according to this embodiment classifies the texture feature selected in the texture analysis step into three classes, AD-MCI, AD-NC, and MCI-NC, to increase the accuracy or reliability of image classification for a data set. It may further include the step of verifying. This step may be performed after the learning step.

Hereinafter, each component of the above-described Alzheimer's disease classification method will be described in more detail.

Materials

In this example, the DICOM (Digital Imaging and Communications in Medicine) format of the ADNI data set was used. Data were taken from the ADNI database (http://adni.loni.usc.edu). ADNI was established in 2003 by the National Institute on Aging (NIA), the National Institute of Biomedical Imaging and Bioengineering Institute (NIBIB), the Food and Drug Administration (FDA), a private pharmaceutical company, and a non-profit organization for $60 in 2003 in a 5-year public-private partnership started with

The main goal of ADNI is to test whether a combination of a series of MRIs, positron emission tomography (PET), other biological markers, and/or clinical and neuropsychological assessments can measure the progression of MCI and early AD. ADNI is the result of the efforts of many researchers in a wide range of educational institutions and private companies.

In this example, 180 MR images of three different classes were used in the ADNI data set. Each classification consisted of 60 subjects each with Alzheimer's disease patients (AD), mild cognitive impairments (MCI), or normal controls (NC). Each image of the three groups of datasets was composed of a gray scale of 256 × 256 × 161 pixels. The International Consortium for Brain Mapping (ICBM) template used in the registration process is a standard model for visualizing the human brain and is symmetrical. In addition, since the identifier of each organ in the brain is displayed as a gray scale value, data can be obtained for a specific organ. The size of the image was 217 × 181 × 181 pixels, and the interval between each slice was set to 1.0 mm.

2 is a template of the international consortium for brain mapping (ICBM) for image registration, in which (a) is an ICBM template and (b) is an anatomically labeled ICBM template.

Development Environment

As the development environment, a GeForce 1070 graphics card equipped with an Intel Core i5-4690 central processing unit (CPU) and 16GB RAM was used. Visual Studio Community 2015 and Matlab were used for software development. C++, Python, and MFC were used as major development languages, and Visualization Toolkit (VTK) 7.1, OpenCV 3.2.0, Insight Toolkit (ITK) 4.11 and Keras libraries were used.

Method and Experimental

The entire process of the Alzheimer's disease classification method according to the present embodiment can be divided into image processing, texture analysis, and deep learning (see FIG. 1 ).

In the first stage of image processing, an image is acquired for use in texture analysis, and convergence, region of interest, and heuristic filters are applied to the preprocessed image. And image processing library and 3D slicer program were used for data processing.

In the second stage of texture analysis, 3D-GLCM was applied to the image to perform texture analysis to calculate texture features, and then important features were extracted based on Fisher's coefficients.

In the third stage of deep learning, the accuracy of the proposed method of this embodiment is measured by dividing the acquired texture features into three classes, AD-MCI, AD-NC, and MCI-NC, and applying them to a Multi-Layer Perceptron (MLP) model. did

Each step of the proposed method and experiment will be described in more detail as follows.

Image Processing

In this embodiment, the region of interest (hippocampus) was segmented in the MR image using image processing libraries such as the insight segmentation and registration toolkit (ITK), an open source library for medical image processing, and the open source computer vision library (open CV). .

Image processing proceeded in two steps. First, slicing and registration were used to equalize the spacing between the ADNI dataset and the ICBM template. Then, using the ICBM template label, the region of interest including the hippocampus was segmented and Gabor filters were applied.

Specifically, the image preprocessing process is as follows.

ADNI dataset and ICBM template have different number of slices due to different spacing. Applying image registration without proper preprocessing steps can lead to the following problems: In other words, sorting images with objects based only on the background will not produce reliable results. This is not only because the comparison between objects is not accurate, but also because it is difficult to extract a reference point when applying a mask. In order to solve this problem, several pre-processing steps were applied in this embodiment. First, a brain image with a predetermined width and height was obtained from a DICOM file using the ITK library. Next, a determinant was applied to an affine transform to change the direction, and re-slice was applied to match the spacing of the ADNI dataset with the spacing of the ICBM template. In this process, spline interpolation can be applied to minimize losses during the slicing process.

3 is a diagram for explaining a pre-processing process for reslicing a magnetic resonance (MR) image that can be employed in the Alzheimer's disease classification method of FIG. 1 , wherein (a) is an original image with a spacing of 1.0, (b) is a The resliced image, and (c) shows the resliced image with an interval of 1.2, respectively.

In addition, in the image pre-processing according to the present embodiment, a process for equalizing the intensity between images was applied. In an analysis process that relies on voxel gray values, it is important to correct the signal intensity in the MR image. That is, the main cause of deterioration in the MR image is spatial non-uniformity with respect to the sensitivity of specially designed surface coils. Therefore, in this example, local entropy minimization based on LEMS (local entropy minimization with a bicubic spline model) was used to make the intensity uniform in all images used in the experiment. This is shown in FIG. 4 .

4 is another example of an image preprocessing process that can be employed in the Alzheimer's disease classification method of FIG. 1, and shows an image non-uniformity correction process based on local entropy minimization, (a) is the original image, (b) is the corrected image A bias field, and (c) represent the corrected image, respectively.

Registration

In this example, the pre-processed ADNI dataset was registered in the ICBM template using 3D Slicer software, which is a special imaging software developed by the National Institutes of Health (NIH) in the United States. With this, we can obtain robust results for the elasticity and distortion of rigid bodies in images using registration functions designed for brain imaging. It also supports automated functions, minimizing errors and time lost associated with manual tasks. The registration process uses the 3D Slicer's Transform and Brain (BRAINS) Fit module. The transformation module performs horizontal/vertical movement or rotation on the input images in the sagittal, coronal and axial planes. The Brain Fit module performs matching between the input image and the template using automatic masking and various options. The registration process is illustrated in FIG. 5 .

FIG. 5 exemplifies an image registration process that may be employed in the Alzheimer's disease classification method of FIG. 1 . 5 shows a process of registering an ADNI image and an ICBM template, (a) is the original image, (b) is the registration in the ICBM image, and (c) is the registered result image, respectively.

Region of Interest Segmentation

In this embodiment, a region of interest (ROI) including the hippocampus is identified from the MR image and used for the experiment. FIG. 6 is a diagram for explaining a process of analyzing a region of interest based on the ICBM template of FIG. 5 . As shown in Fig. 6(a), the ICBM template functions to distinguish and designate the hippocampus and peripheral regions of the hippocampus by labeling each anatomical region of the brain using grayscale values. Then, as shown in FIG. 6(b), ROIs are segmented in all MR images. Here, red and green represent the hippocampus and amygdala, respectively. In this embodiment, it can be defined as bundling the ROI with a yellow rectangle. The ROI may be 49×33 pixels. ROIs can be acquired from a total of 40 images, including 20 from the left and 20 from the right. 6 (c) shows an example of applying the ROI to the ADNI dataset, confirming that the hippocampus is included in the rectangle.

Gabor Filter

The Gabor filter uses a Gaussian function to extract various features in the spatial and frequency domains. The Gabor filter is suitable for texture analysis, so it is possible to obtain the texture characteristics of the image.

The formula of [Equation 1] above defines a two-dimensional (2D) Gabor filter. According to this formula, various types of images can be obtained by adjusting five parameters: sigma, theta, lambda, phi, and gamma. Here, sigma represents a wavelength, and as sigma increases, an area to which a kernel is applied increases. When theta is adjusted, you can set the direction of the edges detected by the filter. Lambda sets the period in which the operation applied to the surrounding pixels is repeated based on the center of the image, and phi represents the distance from the center. Finally, gamma is the ratio of filter width to height. In this example, only lambda and theta were adjusted, and all other parameters were left untouched. Four sets of values were applied to each parameter and 16 different images were obtained. 7 is a two-dimensional (2D) type of Gabor filter (see frequency: F1,F2,F3,F4) and theta (0°, 45°, 90°, 135°) as a type of Gabor filter. It shows the shape of the Gabor filter according to the value.

Texture Analyzes

In the texture analysis step according to the present embodiment, texture features were extracted for 16 types of images obtained from the Gabor filter. On the other hand, since the hippocampus appears continuously in the image, the relationship between each slice was evaluated using 3D-GLCM in this example. Among the 12 texture features obtained using 3D-GLCM, the most important features were selected using Fisher coefficients.

The texture analysis process will be described in more detail as follows.

Gray-Level Co-occurrence Matrix (GLCM)

GLCM is a statistical texture analysis method that considers the spatial relationship between pixels. This is also called the gray level spatial dependence matrix. It identifies spatial dependencies based on the gray level of the image and is suitable for classifying textures made up of small items. The co-occurrence matrix of GLCM is a 2D histogram of dimension G x G, where G is the number of grayscales. When I is a discrete image array and d is a displacement vector, the co-occurrence matrix H _ij ^CO is defined by the following [Equation 2].

In [Equation 2] above, the (i, j)-th element of the matrix is the number of gray levels i and j, and the distance and direction are specified as a displacement vector d, and the number of elements in the set and x=(x, y ) image array It passes through I [26, 27]. 3D-GLCM has 13 orientations: 0°, 45°, 90°, 135°, 180°, 225°, 270° and 315° from the origin of the top image. In this example, 12 functions used for texture analysis were selected as contrast, entropy, variance, correlation, sum, mean, inverse variance, cluster shading, cluster tendency, homogeneity, maximum probability, inertia, and energy.

Fisher’s Coefficient Selection

A correlation coefficient is a statistic that indicates the relationship between standardized variables. If the correlation between variables is applied correctly, more accurate classification results can be obtained. On the other hand, the ability to classify or select only some or specific functions may be effective in some applications, but cannot be guaranteed to be effective in all applications. Therefore, in this embodiment, an optimal feature for each group is identified using Fisher's coefficient.

The Fisher coefficient is based on Fisher's analysis of the process of transforming a correlation distribution into a normal distribution. It can be used to find correlations between distributions that are already well known in statistical analysis with Pearson's variables. The correlation between variables is expressed as a rank, and the higher the rank, the higher the significance. The Fisher coefficient may be defined by the following [Equation 3].

In Equation 3 above, D and V represent inter-class variance and intra-class variance. The representation probability for a feature is the deviation and means of the feature of a given class. Twelve texture features were calculated through Fisher's coefficient, and according to the calculation results, five with a grade of 8 or higher were selected as feature data for classifying Alzheimer's patients.

Deep Learning

A multilayer perceptron (MLP) model consists of an input layer with X input nodes, a hidden layer with H nodes, and an output layer with Y input nodes. The MLP model can solve more complex problems by adding an intermediate layer called a hidden layer that does not exist in the existing perceptron model. The input data and output values generated by the artificial neural network during the learning process are adjusted using activation functions and weights, so they are similar to the correct answer. The MLP model is applied to various problems and is effective for pattern recognition problems.

The implementation of the MLP model in this embodiment was based on the Keras deep learning library. The input data included texture features obtained from 180 patients with AD. Data selected for the training dataset and trial dataset were collected from 40 and 20 patients, respectively.

In this example, the commutation linear unit (ReLu) and Softmax activation functions were used. ReLu can be used instead of a sigmoid function and has the advantage of avoiding a fixed slope of 0 or 1 since the slope is set to 0 for negative inputs. If the input is greater than zero, the slope is guaranteed. Softmax is used to determine the significance of the kth input value in the n inputs. Classification results are displayed as values between 0 and 1.

In addition, overfitting, which frequently occurs when applying deep learning models, is an error that occurs when the learning process is not properly tuned or executed excessively, which affects the accuracy of the model. To address these unwanted effects, Keras' early stop and dropout features can be applied.

Deep Learning Classification

[Table 1] below shows the case of using the MLP model (Our Method) implemented in this example, the case of using the support vector machine (SVM) of the first comparative example, and the k-recent of the second comparative example. When the K-nearest neighbor (KNN) algorithm is used, the classification results of AD patients in each are shown. In the case of AD-MCI, AD-NC, and MCI-NC of the MLP model of this example, the accuracies were 72.5%, 85.0%, and 75.0%, respectively. The highest accuracy was obtained when the dropout ratio was 0.2 in overfitting evaluation for AD-NC. It was confirmed that the optimal ratios of AD-MCI and MCI-NC were 0.4 and 0.5.

8 to 10 show graphs of the performance of the MLP model at the optimal ratio. The graph shows that optimal learning was achieved by preventing overfitting.

8 is a diagram showing the performance of a proposed model of deep learning that can be employed in the Alzheimer's disease classification method of FIG. 1 , and shows the performance of the MLP model for the AD-MCI group (see Table 2). 9 is a diagram showing the performance of a proposed model of deep learning that can be employed in the Alzheimer's disease classification method of FIG. 1 , and shows the performance of the MLP model for the AD-NC group (see Table 3). And FIG. 10 is a diagram showing the performance of a proposed model of deep learning that can be employed in the Alzheimer's disease classification method of FIG. 1, and shows the performance of the MLP model for the MCI-NC group (see Table 4).

Evaluation of the model based on a confusion matrix

In this embodiment, the performance of the MLP model was verified using the fusion matrix. Confusion matrices can be used to evaluate the performance of machine learning algorithms or to solve statistical classification problems. The results of evaluating the performance of the MLP model using the following [Equations E4 to E9] are shown in [Table 2] to [Table 4].

[Equations E4 to E9]

On the other hand, accuracy and error rate are the most commonly used indicators for evaluating models generated by machine learning. This indicator shows the ratio of correctly predicted data to incorrectly estimated data. Sensitivity is a positive predictor variable of the total positive predictive data, and specificity refers to the number of correctly predicted negative results from the total number of negative results predicted. Precision is the actual number of positive predictors output by the model, and the false-positive rate is the number of negative classifications incorrectly classified as positive. Table 5 shows the AD classification results using the MLP model according to this embodiment.

To summarize the configuration and effects of the present embodiment described above, in this embodiment, a region of interest (ROI) including the hippocampus is extracted from an MR image and a texture function is used to facilitate early diagnosis of Alzheimer's disease (AD). A learning-based classification method is provided.

In addition, the following methods can be applied to increase the accuracy of the results. First, it is possible to implement a matching process, including an automated method for preprocessing image data and extracting ROI. It is almost impossible to control the output MR image so that the standard MR is constant under various conditions such as the patient's condition, equipment, and imaging standard. Therefore, in this embodiment, high objectivity and accuracy are achieved by matching all images to the ICBM template, which is a standard model of the brain. All images are reprocessed to the same standard.

Also, although the hippocampus is located inside the temporal lobe, the shape of the brain varies from person to person. Because of this, it is not easy to find the hippocampus in the MR image. However, in the present embodiment, this problem is solved through the above matching process and automatic segmentation of the ROI including the hippocampus is possible. The automated processing steps implemented to facilitate segmentation can greatly improve the productivity of studies or experiments using large ADNI data sets.

Next, in the second step of this embodiment, a Gabor filter is used. To extract hippocampus-based texture features, the hippocampus must be highlighted within the ROI. Gabor filters can generate a large number of images depending on the combination of parameters used and can identify different features. However, it takes a lot of processing time to use all the features of the Gabor filter. Accordingly, in this embodiment, only the lambda and theta of the Gabor filter are changed and used, thereby simplifying the process and minimizing the time required for calculation.

Next, in the third step of this embodiment, an optimal texture feature extraction method may be developed and applied. When conducting a study on object classification based on texture features, it is very important to extract optimal features. In general texture analysis, using all features or specific features collectively may be effective only for some groups. The same function can cause errors in different groups. To solve this problem, it is possible to statistically analyze the calculated features of each group and extract only the optimal features for classification. In this way, in the present embodiment, it is possible to minimize errors and processing time required to construct an optimal data combination from the beginning.

As a result, in this example, high training/test accuracy was obtained in the AD-MCI (76/72), AD-NC (90/85) and MCI-NC (80/75) groups. The highest training accuracy of 90% was obtained when analyzing the AD-NC group with a test accuracy of 85%. Also, by verifying the classification result of the proposed model using the fusion matrix, it was confirmed that the classification result of Alzheimer's disease was significant. These comparisons indicated that the AD-NC model was unbiased and balanced. A satisfactory classification result was obtained in the classification of the AD-NC group. On the other hand, the classification performance of the model incorporating MCI was lower. This is because MCI is an intermediate stage between dementia and is accompanied by multiple symptoms. MCI does not have a separate exact standard for diagnosis, and it is known that different diagnostic results for the same patient in clinical trials are known. In addition, the results of this example were objectively verified by comparative analysis of the results of the comparative examples obtained using the SVM and KNN classifiers (see Table 1).

In addition, the classification result of the classification method proposed in this example was 6.11% more accurate than SVM and 14.19% more accurate than KNN. The deep learning model (MLP model) of this embodiment implemented in the test phase was 3.7% more accurate than SVM and 9.16% more accurate than KNN. This means that the MLP, which is the proposed model of this embodiment, has better performance than a model using SVM or a model using KNN. This indicates that MLP can be appropriately used as a classifier for use cases.

The superiority of the method according to this example can be confirmed by comparing the results with the results of related studies. [Table 6] compares the results presented in this example with those of previously published papers (Liu et al.; Gray et al.) as comparative examples.

In the above [Table 6], Liu et al. used positron emission tomography (PET) for early diagnosis of AD, and divided the gray matter of the brain into 83 regions based on the ICBM template. Then, cerebral metabolic rate (CMRGlc) patterns for glucose consumption were extracted from each region of gray matter and Alzheimer's disease was classified using a deep learning model consisting of sparse automatic encoders and soft max regression layers. This model was valid for both binary classification and multi-class classification.

In addition, Gray et al. used fluorodeoxyglucose PET for AD diagnosis. This method uses the MNI template to divide the brain into 83 anatomical regions. It detects the signal strength in each domain and then classifies it using SVM. And we adopted a process to remove duplicate features using a statistical method, and obtained meaningful results for early diagnosis of AD.

It should be noted that the above Examples and Comparative Examples differ in both the method used for diagnosis and the data used. Nevertheless, it can be seen that the method of this example has better performance than the method of the comparative example. The results confirm the effectiveness of the proposed classification method and indicate that the onset of AD induces deformation of the hippocampus.

11 is a schematic block diagram of an apparatus for classifying Alzheimer's disease according to another embodiment of the present invention.

Referring to FIG. 11 , the apparatus 100 for classifying Alzheimer's disease according to the present embodiment includes a control unit 110 and a storage unit 120 , and may optionally further include a communication unit 130 . In addition, the control unit 110 may be connected to the database system 140 having a database, and may be connected to the input/output device 150 . Such a device 100 may be referred to as at least a portion of a computing device.

The control unit 110 may include a microprocessor and execute a program stored in the storage unit 120 such as a memory. The controller 110 may load a software module by a program and perform image processing, texture analysis, and deep learning for Alzheimer's disease classification.

The software module is acquired from the identification module 111, which identifies a region of interest (ROI) including the hippocampus by labeling each anatomical region in a magnetic resonance (MR) image using a gray scale value. A filtering module that obtains texture characteristics through a plurality of different sub-images of each image according to a plurality of preset types by applying a filter to each image of the data set including the region of interest The classification module 114, which classifies texture features by identifying spatial dependence through the gray level of each image with may include a selection module 115 for selecting

In addition, the software module may include a registration module 112 that registers an image using a template after identification of the ROI, and an image cropping module 113 that cuts the registered image to an appropriate size.

In addition, the software module further includes a learning module 116 for learning each image classified as a texture feature in the selection module, or selects the texture feature selected in the selection module in three ways: AD-MCI, AD-NC and MCI-NC. It may further include a verification module 117 for classifying into classes and verifying the image classification result for the dataset.

The learning module 116 may include a program for performing an operation of deep learning or machine learning. The learning module 116 may classify Alzheimer's disease from an input image based on attributes learned through training data using an artificial neural network (ANN), and big data stored in a database for such classification (big data) is available.

At least one of the above-described identification module 111 , registration module 112 , image cropping module 113 , filtering module, classification module 114 , selection module 115 , learning module 116 and verification module 117 . Some may be stored in the storage unit 120, and may be referred to as an identification unit, a registration unit, an image cropping unit, a filtering unit, a classification unit, a selection unit, a learning unit, and a verification unit, respectively, in the order described.

The database system 140 stores images for operation of the device of this embodiment, stores images and data generated during operation, or stores and manages data to support deep learning or machine learning learning or classification operations of the learning unit. can The database system 140 may be connected to an imaging medical device that generates an MRI or the like, or a server device that collects and stores data of the imaging medical device.

The input/output device 150 may include a keyboard, a touch panel, a mouse, a monitor, a speaker, etc. connected to the controller 110 .

As discussed above, this embodiment provides a novel deep learning-based method for classifying Alzheimer's patients using the hippocampal acquisition method and applying texture analysis to MR images. And according to this embodiment, it was confirmed that AD and NC can be accurately classified. As such, in the present embodiment, objective diagnostic criteria for diagnosing Alzheimer's patients may be provided, and a classification method that may be developed as an automatic classification system may be provided. In addition, more data and texture features can be used to improve the accuracy of classification, thereby more effectively performing Alzheimer's disease classification based on the combination of features of the 3D image of the hippocampus.

Although the above has been described with reference to preferred embodiments of the present invention, those skilled in the art can variously modify and change the present invention within the scope without departing from the spirit and scope of the present invention as set forth in the following claims. You will understand that there is

Claims (12)

identifying, by an identification unit, a region of interest including the hippocampus by labeling each anatomical region using a gray scale value in a magnetic resonance (MR) image;
obtaining, by a filtering unit, a texture characteristic through a plurality of different sub-images of each image according to a plurality of preset types by applying a filter to each image of the data set including the region of interest and obtained in the identifying step;
classifying, by a classification unit, a texture characteristic by identifying spatial dependence through a gray level of each image based on the texture characteristic of the plurality of sub-images; and
selecting, by a selection unit, a correlation coefficient between variables used for the texture feature to select a texture feature for classification of each image in a data set;
including,
The identifying step is
The identification unit slices the MR image of the data set to match the interval of the template, applies spline interpolation, and then applies LEMS (local entropy minimization with a bicubic spline model) to uniformly preprocess the intensity;
Alzheimer's disease classification method further comprising. The method according to claim 1,
In the step of obtaining the texture characteristic, the filtering unit uses a Gabor filter that extracts the characteristic in the spatial and frequency domain using a Gaussian function, wherein the Gabor filter includes a wavelength for each image, an edge direction, and an image At least some of the five parameters: sigma, theta, lambda, phi, and gamma for the iteration duration of operations applied to surrounding pixels relative to the center, the distance from the center, and the ratio of filter width to height are adjusted. How to classify Alzheimer's disease. The method according to claim 1,
In the step of classifying, the Alzheimer's disease classification method in which the personal interest classifier evaluates the relationship between the plurality of sub-images or the slices corresponding thereto by applying a preset characteristic coefficient. 4. The method according to claim 3,
In the classifying step, contrast, entropy, variance, correlation, sum, mean, inverse variance, cluster shade, clustering tendency, homogeneity, A method of classifying Alzheimer's disease by selecting variables including maximum probability, inertia, energy, or a combination thereof. The method according to claim 1,
wherein the selecting comprises identifying a texture feature for each image group in a data set using a fisher coefficient for the selector to select a correlation coefficient between the variables. The method according to claim 1,
Alzheimer's disease classification method further comprising the step of learning each image classified by the texture feature in the selecting step by the learning unit. 7. The method of claim 6,
The learning step is Alzheimer's disease classification in which the learning unit performs image classification on a training data set or a test data set through a multi-layer perceptron (MLP) model based on a Keras deep learning library. method. 8. The method of claim 7,
The step of learning is Alzheimer's disease classification method using a rectification linear unit (ReLu), in which the learning unit is used as a substitute for an S-shaped function, a slope is set to zero for a negative input, and a rectification linear unit (ReLu) designed to avoid a fixed slope of 0 or 1 . The method according to claim 1,
Alzheimer's disease further comprising the step of classifying the texture feature selected by the selection unit into three classes of AD-MCI, AD-NC, and MCI-NC in the selecting step and verifying the accuracy or reliability of image classification for the data set classification method. an identification unit for identifying a region of interest including the hippocampus by labeling each anatomical region using a gray scale value in a magnetic resonance (MR) image;
a filtering unit obtained by the identification unit and applying a filter to each image of the data set including the region of interest to obtain texture characteristics through a plurality of different sub-images of each image according to a plurality of preset types;
a classification unit for classifying the texture characteristics by identifying spatial dependence through the gray level of each image based on the texture characteristics of the plurality of sub-images; and
a selection unit for selecting an optimal texture feature for classification of each image in a data set by selecting a correlation coefficient between the variables used for the texture feature;
including,
The identification unit slices the MR image of the data set to match the interval of the template, applies spline interpolation, and then applies LEMS (local entropy minimization with a bicubic spline model) to uniformly preprocess the intensity. sorting device. 11. The method of claim 10,
Alzheimer's disease classification apparatus further comprising a learning unit for learning each image classified by the texture feature in the selection unit. 11. The method of claim 10,
The apparatus for classifying Alzheimer's disease further comprising a verification unit that classifies the texture feature selected by the selection unit into three classes, AD-MCI, AD-NC, and MCI-NC, and verifies the image classification result for the data set.

Images