KR101944069B1 - Method and apparatus for deciding Alzheimer's disease based on machine learning - Google Patents

Abstract

The method for determining Alzheimer's disease based on machine learning is a method for determining the shape of the first cerebral cortex extracted from the brain image of a normal group of the Alzheimer's disease apparatus and the image processing of the second cerebral cortex shape extracted from the brain image of the Alzheimer's disease patient group To generate vertex-sampled first cortical features and vertex-sampled second cortical features; A first feature vector is extracted from vertex-sampled first cerebral cortical features by performing a noise removal procedure on vertex-sampled first cortical features and vertex-sampled second cortical features of an Alzheimer's disease decision device, Extracting a second feature vector from a second cerebral cortex shape; Determining an evaluation criterion for the determination of Alzheimer's disease based on the first feature vector and the second feature vector; And determining whether the Alzheimer's disease determining device is capable of detecting the onset of Alzheimer's disease based on the criteria for the newly entered input cortical shape.

Description

Technical Field [0001] The present invention relates to a method and apparatus for determining Alzheimer's disease based on machine learning,

The present invention relates to a method and apparatus for determining Alzheimer's disease, and more particularly, to a method and apparatus for determining Alzheimer's disease based on machine learning.

Alzheimer's disease is the most common cause of dementia and accounts for about 60-80% of dementia. Approximately 33.9 million people worldwide have Alzheimer's disease, and the prevalence is expected to triple in 40 years, with an increase in life expectancy. At present, there are no drugs that can expect a definite improvement in Alzheimer's disease. After the onset of the disease, it has worsened for decades, and therefore there is a growing interest in research on prevention of Alzheimer's disease.

Alzheimer's disease is known to be caused by complex interactions of various risk factors. Sociodemographic risk factors, such as age, sex, and education, as well as genetic risk factors, as well as environmental risk factors such as smoking, drinking, nutrition, and social activity, are known. A representative genetic marker for Alzheimer's disease is APOE e4, but these genetic markers are not altered or modified. However, in addition to studies examining how well these factors (such as vascular risk factors, lifestyle) are modifiable factors and how these modifiable factors modulate dementia expression, these factors may be reversed to lower the expression of dementia, The timing of the expression can be delayed.

Factors currently known to be a risk factor for ameliorative Alzheimer's disease include diabetes, middle-aged hypertension, middle-aged obesity, depression, low physical activity, smoking, and low educational attainment.

One aspect of the present invention provides a method for determining Alzheimer's disease based on machine learning.

Another aspect of the present invention provides an apparatus for performing an Alzheimer's disease determination method based on machine learning.

The method for determining Alzheimer's disease based on machine learning according to one aspect of the present invention is characterized in that the Alzheimer's disease determination apparatus is divided into a first cerebral cortex shape extracted from a brain image of a normal group and a second cerebral cortex shape extracted from a brain image of a group of Alzheimer's disease patients Performing image processing on the cerebral cortex shape to generate vertex-sampled first cerebral cortex and apex-sampled second cerebral cortex shapes, and the Alzheimer's disease determination apparatus includes a first vertex-sampled first cerebral cortex shape and a vertex- A noise removal procedure for the sampled second cerebral cortex shape is performed to extract a first feature vector from the vertex-sampled first cortex shape, and a second feature vector from the vertex-sampled second cortex shape is extracted Determining whether the Alzheimer's disease determining device is based on the first feature vector and the second feature vector; Determining a determination criterion for the determination of the Alzheimer's disease and determining whether the Alzheimer's disease determinant determines the onset of the Alzheimer's disease based on the criterion for the newly inputted input cerebral cortex shape .

Meanwhile, the image processing may generate a cerebral cortical surface model for each of the first cortical shape and the second cortical shape, and may downsample a plurality of vertices contained in the cerebral cortical surface model.

The noise removal procedure may include converting the vertex-sampled first cortex shape and the vertex-sampled second cortex shape into a frequency domain, removing a high frequency band in the frequency domain, Off dimension, and the cut-off dimension may be determined by a cut-off dimension determination factor (G).

In addition, the cut-off dimension determination coefficient G is determined based on the following equation,

≪ Equation &

는 j번째 대상의 k번째 주파수 요소이고, N은 트레이닝 대상의 개수이고, F는 대뇌 피질 두께 데이터를 위한 상기 컷-오프 차원이고,

는 상기 컷-오프 차원을 고려하여 상기 노이즈 제거 절차에 의해 노이즈 필터링된 j번째 대상의 대뇌 피질 두께 데이터이고, 상기 고 주파수 대역의 제거를 위한 상기 컷-오프 차원은 상기 G를 0.025로 설정하여 결정될 수 있다.Wherein c ^j is vertex-sampled cerebral cortical thickness data of the jth training subject in the training set including the normal group and the Alzheimer's disease patient group,

Is the kth frequency element of the jth object, N is the number of training objects, F is the cut-off dimension for the cerebral cortical thickness data,

Is the cortical thickness data of the j-th object subjected to noise filtering by the noise removal procedure in consideration of the cut-off dimension, and the cut-off dimension for elimination of the high frequency band is determined by setting G to 0.025 .

The determination criterion may further include reducing a dimension of the first feature vector and the second feature vector based on a principal component analysis (PCA) transformation, reducing the dimension of the first feature vector of the reduced dimension and the feature vector of the reduced dimension Generating a new training set based on the two feature vectors and using linear discriminant analysis (LDA) to maximize the variance between the normal group and the Alzheimer's patient group based on the new training set, And determining the criterion that minimizes variance and dispersion within the group of Alzheimer's patients.

According to another aspect of the present invention, there is provided an apparatus for determining Alzheimer's disease based on machine learning, the apparatus comprising a processor, wherein the processor comprises: a first cortical shape extracted from a brain image of a normal group, To generate vertex-sampled first cerebral cortex and vertex-sampled second cerebral cortex shapes by performing image processing on the second cerebral cortex shape extracted from the brain image of the first cerebral cortex, A noise removal procedure for a vertex-sampled second cortex shape is performed to extract a first feature vector from the vertex-sampled first cortex shape, and a second feature vector is extracted from the vertex-sampled second cortex shape And determining, based on the first feature vector and the second feature vector, the determination of the Alzheimer's disease Determining a criterion, and for the newly entered input cortical shape can be implemented to determine whether the onset of Alzheimer's disease based on the criteria.

Meanwhile, the image processing may generate a cerebral cortical surface model for each of the first cortical shape and the second cortical shape, and may downsample a plurality of vertices contained in the cerebral cortical surface model.

The noise removal procedure may include converting the vertex-sampled first cortex shape and the vertex-sampled second cortex shape into a frequency domain, removing a high frequency band in the frequency domain, Off dimension, and the cut-off dimension may be determined by a cut-off dimension determination factor (G).

In addition, the cut-off dimension determination coefficient G is determined based on the following equation,

≪ Equation &

는 j번째 대상의 k번째 주파수 요소이고, N은 트레이닝 대상의 개수이고, F는 대뇌 피질 두께 데이터를 위한 상기 컷-오프 차원이고,

는 상기 컷-오프 차원을 고려하여 상기 노이즈 제거 절차에 의해 노이즈 필터링된 j번째 대상의 대뇌 피질 두께 데이터이고, 상기 고 주파수 대역의 제거를 위한 상기 컷-오프 차원은 상기 G를 0.025로 설정하여 결정될 수 있다.Wherein c ^j is vertex-sampled cerebral cortical thickness data of the jth training subject in the training set including the normal group and the Alzheimer's disease patient group,

Is the kth frequency element of the jth object, N is the number of training objects, F is the cut-off dimension for the cerebral cortical thickness data,

Is the cortical thickness data of the j-th object subjected to noise filtering by the noise removal procedure in consideration of the cut-off dimension, and the cut-off dimension for elimination of the high frequency band is determined by setting G to 0.025 .

The determination criterion may further include reducing a dimension of the first feature vector and the second feature vector based on a principal component analysis (PCA) transformation, reducing the dimension of the first feature vector of the reduced dimension and the feature vector of the reduced dimension Generating a new training set based on the two feature vectors and using linear discriminant analysis (LDA) to maximize the variance between the normal group and the Alzheimer's patient group based on the new training set, And generating the criterion that minimizes variance and dispersion within the Alzheimer's disease patient group.

The method and apparatus for determining Alzheimer's disease based on machine learning according to an embodiment of the present invention generate a criterion for determining the shape of a cerebral cortex of a normal person and a cerebral cortex shape of a patient based on machine learning, The determination of Alzheimer ' s disease can be performed accurately according to the judgment criterion.

1 is a flowchart illustrating a method for determining Alzheimer's disease according to an embodiment of the present invention.
Figure 2 illustrates an image preprocessing step according to an embodiment of the present invention.
3 shows a method for generating a feature vector according to an embodiment of the present invention.
4 is a conceptual diagram illustrating a method for determining a criterion for distinguishing between a normal group and an Alzheimer's disease patient group according to an embodiment of the present invention.
FIG. 5 is a conceptual diagram illustrating an apparatus for determining Alzheimer's disease that performs a method for determining Alzheimer's disease based on machine learning according to an embodiment of the present invention.
6 is a flowchart illustrating a method for determining Alzheimer's disease based on machine learning according to an embodiment of the present invention.

The following detailed description of the invention refers to the accompanying drawings, which illustrate, by way of illustration, specific embodiments in which the invention may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the invention. It should be understood that the various embodiments of the present invention are different, but need not be mutually exclusive. For example, certain features, structures, and characteristics described herein may be implemented in other embodiments without departing from the spirit and scope of the invention in connection with an embodiment. It is also to be understood that the position or arrangement of the individual components within each disclosed embodiment may be varied without departing from the spirit and scope of the invention. The following detailed description is, therefore, not to be taken in a limiting sense, and the scope of the present invention is to be limited only by the appended claims, along with the full scope of equivalents to which such claims are entitled, if properly explained. In the drawings, like reference numerals refer to the same or similar functions throughout the several views.

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

Alzheimer's disease is a long-term disease and pathology accumulates before it appears to be a clinical manifestation. Thus, if Alzheimer's disease is determined early, appropriate intervention may delay the onset of dementia.

Early identification of dementia is being studied competitively worldwide. The concept of mild cognitive impairment (MCI) appeared for the early determination of Alzheimer's disease causing dementia and thereafter the concept of subjective memory impairment (SMI) or subjective memory complaint . In addition, the early decision of Alzheimer's disease is now expanding to a normal person who does not have any symptoms. Recently, studies on normal people who have no symptoms have been increasing, and attempts have been made to find out an indicator of significant brain changes even in normal people.

On the decisive point, if early diagnosis or pre-emergence of dementia is possible before the appearance of dementia through techniques such as blood tests, brain imaging, etc., it can be used as a conservative drug that alleviates the symptoms of dementia or inhibits progression It is possible not only to maximize the effect of dementia treatment but also to control the dementia risk factors more precisely, thereby delaying the onset of dementia symptoms as much as possible.

Although recent molecular biologic advances have accelerated the development of therapeutic agents for Alzheimer's disease, there has been no significant clinical progress in clinical trials of these drugs. The most important reason for this is that there is a lot of point that it is not meaningful to treat after much damage to the brain. Therefore, it is important to determine this at the earliest point of Alzheimer ' s disease, to predict the onset and to intervene early.

The early detection of early detection of dementia by delaying the onset of dementia by two years, the prevalence of dementia after 40 years is 80% lower than the case of not delaying the onset of dementia. It is also reported that delaying symptoms by one year not only reduces the prevalence of Alzheimer's disease by 9 million people over the next 40 years, but also reduces the average severity of dementia patients.

A number of epidemiological studies have been conducted to predict future risk from current risk factors, but most of them have been derived from studies in Western populations.

Efforts are needed to identify new risk factors, including blood tests and quantification of brain cortical thickness via brain magnetic resonance imaging, including known risk factors. It is also necessary to develop a risk index for early detection of Alzheimer's disease. It may be a more reasonable preventive measure to identify risk factors for Alzheimer's disease in Asians, to assess individual risk factors, to identify high-risk groups, and to adjust adjustable risk factors according to the level of risk. Since the onset and progression of the disease progresses slowly over a long period of time, it is necessary to develop a risk index that integrates MRI brain images in order to analyze the risk factors in connection with pre- Time.

It is necessary to identify the risk factors that may be a risk factor for dementia in Korean adults and to analyze their effects by using various tests and MRI data including blood tests. Combining these risk factors, it is very urgent to find indicators to predict disease outbreaks early.

In addition, it is necessary to develop brain age, which is a personal analysis as a basis for early decision of dementia. Most of the current studies are divided into three groups (normal, dementia, mild cognitive impairment, and dementia). However, the primary purpose of early decision of dementia is to determine early on whether a person has Alzheimer's disease. In other words, it is necessary to show the degree to which the individual's brain health is at a certain position and the probability of going to dementia when an actual person images the brain.

Hereinafter, embodiments of the present invention disclose a method and apparatus for determining Alzheimer's disease based on machine learning.

Patients with Alzheimer's disease are known to experience atrophy of the cerebral cortex years before the onset of Alzheimer's disease. In the embodiment of the present invention, when the method and apparatus for determining Alzheimer's disease based on machine learning are used, prediction of the presence or absence of Alzheimer's disease based on image only is possible.

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

FIG. 1 shows an example of a brain image (brain magnetic resonance imaging (MRI) image) of a cognitive normal (CN) group of patients with Alzheimer's disease as training data for learning using a machine learning classifier, A method for predicting a disease is disclosed.

Alzheimer's disease prediction based on machine learning on the cerebral cortical thickness data can be used to generate Alzheimer's disease classifier through map learning using training data as well as general machine learning and to determine the presence or absence of Alzheimer's disease based on Alzheimer's disease classifier Can be performed.

An image preprocessing process for the brain image is performed (step S100).

Based on image preprocessing of brain images of the Alzheimer's disease patient group and brain images of the cognitive normal (CN) group, the pre-processed cerebral cortical features of the Alzheimer's disease patient group and the normal group can be extracted. The preprocessed cerebral cortex shape may be a vertex-sampled (or resampled) shape. Vertex sampling is described below.

A feature vector for the vertex-sampled cerebral cortex shape is generated (step S110).

Feature vectors for vertex-sampled cerebral cortical features can be generated based on a manifold harmonic transform (MHT).

Feature vectors can be generated based on clinical information that patients with Alzheimer's disease begin to thin from a specific area of the cerebral cortex at an early stage of disease. The resulting feature vectors should reflect the differences between the two groups (normal and Alzheimer's disease groups) for maximum classification performance. However, the generated feature vector can be advantageous as it maximizes the difference between the two groups for high classification performance, while the size is small for efficiency of calculation.

In the embodiment of the present invention, on the basis of the vertex sampling procedure and the noise removal procedure for the cerebral cortical shape described above, the size of the feature vector is reduced to 1/20 level, and the resampled cerebral cortex- Features can be removed.

Based on the dimensional reduction procedure and the grouping procedure, a search is performed (step S120) in which the difference between the two groups (the normal group and the Alzheimer's disease group) is most apparent.

For example, the dimension reduction procedure may be a procedure for reducing the dimension of the feature vector determined in step S110 to one dimension based on principal component analysis (PCA).

The group classification procedure is based on linear discriminant analysis (LDA), and is used to generate a decision criterion for classifying the cerebral cortex shape by judging whether the cortical shape used in the machine learning is closer to a group through the step S110) Procedure. LDA is an analytical method that finds the axis that minimizes intra-group variance and maximizes inter-group variability. If LDA is used, the criterion by which the difference between the two groups is most apparent can be determined.

Based on the criterion that most clearly shows the difference between the two groups, classification of the input cerebral cortex shape is performed, and whether the users matching the inputted cerebral cortex shape are included in the normal group or in the Alzheimer's disease group (Step S130).

In the embodiment of the present invention, a specific method for each step will be described below.

Figure 2 illustrates an image preprocessing step according to an embodiment of the present invention.

FIG. 2 shows a method of extracting a cerebral cortex shape (cerebral cortical image) from a brain image of a group of Alzheimer's disease patients and a brain image of a normal group and performing vertex sampling (or resampling) on the cerebral cortex shape. do.

Referring to FIG. 2, the cortical thickness can be extracted from brain imaging and / or cerebral cortical shape of the collected Alzheimer's disease patient group and normal group. The process of extracting cortical thickness from the brain image and / or the cerebral cortex shape may be performed based on an image or surface model registration process.

Since different cortical features between patients are divided into sub-regions based on characteristic wrinkles of the cortex, all vertices of the cortical model extracted through the matching process can be assumed to belong to similar sub-regions of the brain.

For example, a cerebral cortical surface model (200) composed of more than 100,000 vertices in each of the left and right hemispheres of the cerebrum can be generated, and the generated cerebral cortex surface model (200) has 40,962 vertices of the left and right hemispheres (Vertex-sampled cerebral cortical shape, vertex-sampled cerebral cortical image) 220. The downsampled down-sampled down-sampled cerebral cortical surface model The number of vertices constituting the cerebral cortical surface model 200 and the number of vertices of the vertex-sampled cerebral cortical surface model 220 may be varied by way of example.

3 shows a method for generating a feature vector according to an embodiment of the present invention.

FIG. 3 discloses a method for generating a feature vector based on a noise removal procedure for vertex-sampled cortical features.

As described above, the feature vectors for vertex-sampled cerebral cortex shapes can be generated based on a manifold harmonic transform (MHT). In the embodiment of the present invention, the size of the feature vector can be reduced to 1/20 level based on the vertex sampling procedure and the noise removal procedure, and the noise in the cerebral cortex shape can be removed. In this way, the efficiency of the computation can be improved while maximizing the difference between the normal group and the Alzheimer's disease group.

Vertex sampling is a resampling procedure performed in the image preprocessing procedure. For example, the initially generated surface model can consist of more than 100,000 vertices per left / right hemisphere. Through the resampling procedure, the surface model can consist of a total of 81,924 vertices. That is, it can be about 1/3 to 1/4 of the original model having more than 100,000 vertices per left / right hemisphere with 40,962 vertices in each of the left and right hemispheres.

Noise removal can be performed based on MHT. When MHT is used, the scalar function defined for the surface may be convertible to the frequency domain. And high frequency band components may represent portions with noise or high individual variation. Therefore, after the vertex sampling, 81,924 cortical thickness values are not all used, but the average model is made from the surface models to be discriminated. By applying MHT, only the low frequency band components of 81,924 cortical thickness values are constructed, Can be greatly reduced.

Specifically, the high-frequency components of the cerebral cortical thickness data can represent individual characteristics independent of noise and group characteristics. Thus, it is possible to reduce the dimension of the cerebral cortical thickness data by removing this high frequency component f _i based on the determined cutoff dimension 300.

The following is a formula for calculating the coefficient G for determining the high frequency component to be removed.

≪ Equation &

는 j번째 대상의 k번째 주파수 요소이고, N은 트레이닝 대상의 개수이다. F는 대뇌 피질 두께 데이터를 위한 컷-오프 차원(cut-off dimension)(300)일 수 있다.

는 컷-오프 차원(300)을 고려하여 노이즈 필터링된 j번째 대상의 대뇌 피질 두께 데이터이다. h^k는 이산 LB(Laplace-Beltrami) 연산자의 k번째 아이젠 벡터(eigen vector)이다.Where c ^j is the original cortical thickness data of the jth training subject in a training set comprising a normal group and a group of Alzheimer's patients,

Is the kth frequency element of the jth object, and N is the number of training objects. F may be a cut-off dimension 300 for cerebral cortical thickness data.

Is the cerebral cortical thickness data of the noise filtered j-th object in consideration of the cut-off dimension (300). h ^k is the kth eigenvector of the discrete LB (Laplace-Beltrami) operator.

In the embodiment of the present invention, the cut-off dimension 300 can be determined by setting G to 0.025. However, it is needless to say that the value of G is not limited to the above value but can be determined as an appropriate value. The high frequency band (or high frequency component) may be eliminated based on the determined cut-off dimension 300. Based on the elimination of these high frequency bands, noise and individual characteristics independent of group characteristics can be eliminated.

4 is a conceptual diagram illustrating a method for determining a criterion for distinguishing between a normal group and an Alzheimer's disease patient group according to an embodiment of the present invention.

In Fig. 4, a method for determining a determination criterion that most clearly indicates a difference between two groups based on a dimension reduction procedure and a group classification procedure is disclosed.

Referring to FIG. 4, a principal component analysis (PCA) (400) -LDA (linear discriminant analysis) 410 method among various statistical methods is sequentially applied in order to facilitate incremental classifier update according to the accumulation of image data The final classification can be performed by reducing the dimension of the input feature vector to one dimension finally.

The dimension of the feature vector may be reduced based on the PCA 400 transformation first. For example, the number of eigenvectors selected during the PCA 400 transform may be empirically selected to maximize the performance of the classifier during 10-fold cross validation. For example, the number of selected eigenvectors may be 64. The LDA 410 is an analytical method that finds axes that minimize intra-group variance and maximize inter-group variance, and can be used to find a criterion that most clearly shows the difference between the two groups.

중 하나의 그룹에 포함된 주어진 특징 벡터

,

에 대하여 그룹 분류기는 PCA(400) 및 LDA(410)를 기반으로 순차적으로 트레이닝될 수 있다. 아래의 수학식 2를 기반으로 PCA(400)를 수행하기 위해 트레이닝 데이터 집합

의 공분산 매트릭스 V가 유도될 수 있다. The PCA 400 -LDA 410 can be performed in the following manner.

≪ / RTI > a given feature vector < RTI ID = 0.0 >

,

The group classifier may be trained sequentially based on the PCA 400 and the LDA 410. [ In order to perform the PCA 400 based on the following Equation 2,

Lt; RTI ID = 0.0 > V < / RTI >

&Quot; (2) "

는 모든 특징 벡터의 평균이다. 각

는 특징 벡터

를 나타내고, V는

매트릭스이기 때문이다. 공분산 매트릭스 V가 일반적으로 싱귤러(singular)이기 때문에 LDA(410)를 수행함에 있어서 싱귤러리티 문제를 가지고 올 수 있다. 이러한 문제를 방지하기 위해 PCA 변환 매트릭스(

)를 생성하기 위해 가장 큰 k개의 0이 아닌 아이젠벨류(eigen value)와 관련된 V의 아이젠벡터가 선택될 수 있다. 선택된 아이젠벡터의 개수는 감소된 PCA 공간의 차원을 결정할 수 있다. here,

Is the average of all feature vectors. bracket

Is a feature vector

, V represents

This is because it is a matrix. Since the covariance matrix V is typically a singular, it may have a singularity problem in performing the LDA 410. [ To avoid this problem, the PCA transformation matrix (

The eigenvectors of V associated with the largest k " non-zero eigenvalues may be selected. The number of selected EigenVectors can determine the dimension of the reduced PCA space.

를 기반으로 특징 벡터 x는 PCA 공간 상의 벡터 y로 전환될 수 있다. A given PCA transformation matrix < RTI ID = 0.0 >

The feature vector x can be converted to a vector y on the PCA space.

&Quot; (3) "

이 생성될 수 있다. By applying the PCA 400, a new training set on the PCA space for all feature vectors of X

Can be generated.

LDA can be performed on the new training set.

The LDA 410 can find a coordinate axis for separating the group of data sets as much as possible. In other words, the LDA 410 can find coordinate axes to minimize the variance of each group. The LDA 410 may maximize the variance among the groups and minimize the variance of each group.

According to an embodiment of the present invention, the LDA 410 may find the axis w for minimizing the energy function of Equation 4 below.

&Quot; (4) "

는 축 w로 투영된 그룹 간 분산이고,

는 축 w로 투영된 그룹 내 분산일 수 있다. 그룹 간 스캐터 행렬 S_B와 그룹 내 스캐터 행렬 S_W는 아래의 수학식5와 같이 정의될 수 있다. here,

Is the inter-group variance projected on the axis w,

May be in-group variance projected on axis w. The inter-group scatter matrix S _B and the intra-group scatter matrix S _W can be defined as Equation (5) below.

Equation (5)

는 그룹 i의 평균,

는 그룹i의 공분산 행렬,

는 그룹i의 크기일 수 있다. y_ik는 그룹 G_i의 k번째 특징 벡터일 수 있다. here

Is the average of group i,

Is the covariance matrix of group i,

May be the size of group i. y _ik may be the kth feature vector of group G _i .

It is possible to determine whether the cerebral cortex shape of the normal cerebral cortex shape or the cerebral cortex shape of the Alzheimer's disease patient is inputted after using the extracted criterion based on the PCA 400-LDA 410 as described above.

FIG. 5 is a conceptual diagram illustrating an apparatus for determining Alzheimer's disease that performs a method for determining Alzheimer's disease based on machine learning according to an embodiment of the present invention.

5, the apparatus for determining Alzheimer's disease includes at least one of an image input unit 500, an image preprocessing unit 510, a feature vector generating unit 520, a determination criterion determining unit 530, an Alzheimer's disease determining unit, and a processor 540 .

The image input unit 500 may be implemented to receive a brain image of a group of Alzheimer's disease patients and a brain image of a normal group.

The image preprocessing unit 510 may be implemented to perform a preprocessing procedure for extracting the cerebral cortical thickness from the acquired brain image of the Alzheimer's disease patient group and the brain image of the normal group. The image preprocessing unit 510 can generate a cerebral cortical surface model composed of more than 100,000 vertices in each of the left and right hemispheres of the cerebrum based on the input image and generate the cerebral cortex surface model with the left and right hemispheres 40,962 (Or resampling) vertices to the vertices of the vertex. The number of vertices constituting the cerebral cortical surface model and the number of vertex samples to be vertex-sampled may vary as an example.

The feature vector generation unit 520 may be implemented to generate a feature vector by removing noise from vertex-sampled cortical features. As described above, the high frequency band components of the cerebral cortical thickness data may represent individual characteristics independent of noise and group characteristics. Thus, the dimensionality of the cerebral cortical thickness data can be reduced by removing these high frequency band elements

The feature vector generation unit 520 may be implemented to convert cortical thickness data into a frequency band, remove a specific frequency band, and configure only low frequency band components to greatly reduce the size of the feature vector.

The determination criterion determining unit 530 may reduce the dimension of the feature vector to one dimension and determine a criterion for classifying the normal group and the Alzheimer's disease group based on the machine learning.

The Alzheimer's disease judgment unit 540 can be implemented to judge Alzheimer's disease based on the judgment criteria determined by the judgment reference generation unit.

The processor 550 may be implemented to control the operations of the image input unit 500, the image preprocessing unit 510, the feature vector generation unit 520, the determination criterion determination unit 530, and the Alzheimer's disease determination unit 540 have.

6 is a flowchart illustrating a method for determining Alzheimer's disease based on machine learning according to an embodiment of the present invention.

Referring to FIG. 6, a first cerebral cortex shape can be extracted from a brain image of a normal group of the Alzheimer's disease determination apparatus, and a second cerebral cortex shape can be extracted from a brain image of the Alzheimer's disease patient group (step S600).

The Alzheimer's disease apparatus can perform image processing on the first cerebral cortex shape and the second cerebral cortex shape to generate vertex-sampled first cerebral cortex and vertex-sampled second cerebral cortex shapes (step S610).

Image processing can generate a cerebral cortical surface model for each of the first and second cerebral cortex shapes and downsample multiple vertices contained in the cerebral cortical surface model.

A first feature vector is extracted from vertex-sampled first cerebral cortical features by performing a noise removal procedure on vertex-sampled first cortical features and vertex-sampled second cortical features of an Alzheimer's disease decision device, The second feature vector can be extracted from the second cerebral cortex shape (step S620).

The noise removal procedure transforms vertex-sampled first cortical and vertex-sampled second cerebral cortex shapes to the frequency domain and removes high frequency bands in the frequency domain, wherein the high frequency bands are cut-off dimensions dimension, and the cut-off dimension can be determined by the cut-off dimension determination coefficient G.

The Alzheimer's disease determination apparatus can determine a determination criterion for the determination of Alzheimer's disease based on the first feature vector and the second feature vector (step S630).

The decision criteria include decreasing the dimensions of the first feature vector and the second feature vector based on the PCA transformation, generating a new training set based on the first feature vector of the reduced dimension and the second feature vector of the reduced dimension Based on the steps to maximize the variance between the normal and Alzheimer's patient groups based on the new training set using the LDA and determining the criterion that minimizes dispersion within the normal group and dispersion within the Alzheimer's patient group Can be determined.

The Alzheimer's disease determination device can determine whether or not the newly input input cerebral cortex shape is onset based on the judgment criteria (step S640).

The method of determining Alzheimer's disease based on machine learning can be implemented in an application or implemented in the form of program instructions that can be executed through various computer components and recorded on a computer-readable recording medium. The computer-readable recording medium may include program commands, data files, data structures, and the like, alone or in combination.

The program instructions recorded on the computer-readable recording medium may be those specially designed and configured for the present invention and may be those known and used by those skilled in the computer software arts.

Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tape, optical recording media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, media, and hardware devices specifically configured to store and execute program instructions such as ROM, RAM, flash memory, and the like.

Examples of program instructions include machine language code such as those generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter or the like. The hardware device may be configured to operate as one or more software modules for performing the processing according to the present invention, and vice versa.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it will be understood by those skilled in the art that various changes and modifications may be made therein without departing from the spirit and scope of the invention as defined in the appended claims. It will be possible.

Claims (10)

A method of determining Alzheimer's disease based on machine learning,
The apparatus for determining Alzheimer's disease extracts a cortical thickness from a first cerebral cortex shape extracted from a brain image of a normal group and a second cerebral cortex shape extracted from a brain image of a group of Alzheimer's disease patients, Performing image processing on the second cerebral cortex shape to generate vertex-sampled first cortical features and vertex-sampled second cortical features;
Wherein the Alzheimer's disease determining device converts the vertex-sampled first cortex shape and the vertex-sampled second cortex shape into a frequency domain, and then removes high frequency band components of the cortex thickness data based on the cutoff dimension A noise removal procedure is performed to extract a first feature vector composed of only low frequency band components from the cortex thickness values in the vertex-sampled first cortex shape, and a second feature vector composed of only low frequency band components Extracting from the vertex-sampled second cerebral cortex shape;
Determining the criteria for the determination of the Alzheimer's disease based on the first feature vector and the second feature vector; And
Wherein said Alzheimer's disease determination device comprises determining whether the newly entered input cerebral cortex shape is due to Alzheimer's disease based on said criteria. The method according to claim 1,
Wherein the image processing generates a cerebral cortical surface model for each of the first cerebral cortex shape and the second cerebral cortex shape and downsamples a plurality of vertices contained in the cerebral cortical surface model. 3. The method of claim 2,
The noise cancellation procedure may include transforming the vertex-sampled first cortex shape and the vertex-sampled second cortex shape into a frequency domain, removing a high frequency band in the frequency domain,
The high frequency band is determined based on a cut-off dimension,
Wherein the cut-off dimension is determined by a cut-off dimension determination coefficient (G). The method of claim 3,
The cut-off dimension determination coefficient G is determined based on the following equation,
≪ Equation &

Wherein c ^j is vertex-sampled cerebral cortical thickness data of the jth training subject in the training set including the normal group and the Alzheimer's disease patient group,

Is the kth frequency element of the jth object, N is the number of training objects, F is the cut-off dimension for the cerebral cortical thickness data,

Is cortical thickness data of the j-th object subjected to noise filtering by the noise removal procedure in consideration of the cut-off dimension,
Wherein the cut-off dimension for removal of the high frequency band is determined by setting G to 0.025. 5. The method of claim 4,
Reducing a dimension of the first feature vector and the second feature vector based on a principal component analysis (PCA) transformation;
Generating a new training set based on the first feature vector of the reduced dimension and the second feature vector of the reduced dimension; And
Using linear discriminant analysis (LDA) to maximize variance between the normal group and the Alzheimer's disease patient group based on the new training set and to minimize variance within the normal group and to minimize variance within the Alzheimer's disease patient group And generating a decision criterion based on the decision criterion. An apparatus for determining Alzheimer's disease based on machine learning,
Wherein the apparatus for determining Alzheimer's disease comprises a processor,
The processor comprising:
The cortical thickness is extracted from the second cerebral cortex shape extracted from the brain image of the first cerebral cortex shape and the Alzheimer's disease patient group extracted from the brain image of the normal group, and the first cerebral cortex shape and the second cerebral cortex shape To generate vertex-sampled first cortical features and vertex-sampled second cortical features,
A noise removal procedure for removing the high frequency band components of the cortical thickness data based on the cutoff dimension after transforming the vertex-sampled first cortex shape and the vertex-sampled second cortex shape into a frequency domain, Extracting a first feature vector composed of only low frequency band components from the vertex-sampled first cortex shape, calculating a second feature vector composed of low frequency band components only from the vertex-sampled second cortex shape, Extraction from the cortical shape,
Determining a criterion for determination of the Alzheimer's disease based on the first feature vector and the second feature vector,
Wherein the controller is configured to determine whether the newly entered input cerebral cortex shape is the onset of Alzheimer's disease based on the determination criterion. The method according to claim 6,
Wherein the image processing generates a cerebral cortical surface model for each of the first cortical shape and the second cortical shape and downsamples a plurality of vertices contained in the cerebral cortical surface model, Device. 8. The method of claim 7,
The noise cancellation procedure may include transforming the vertex-sampled first cortex shape and the vertex-sampled second cortex shape into a frequency domain, removing a high frequency band in the frequency domain,
The high frequency band is determined based on a cut-off dimension,
Wherein the cut-off dimension is determined by a cut-off dimension determination coefficient (G). 9. The method of claim 8,
The cut-off dimension determination coefficient G is determined based on the following equation,
≪ Equation &

Wherein c ^j is vertex-sampled cerebral cortical thickness data of the jth training subject in the training set including the normal group and the Alzheimer's disease patient group,

Is the kth frequency element of the jth object, N is the number of training objects, F is the cut-off dimension for the cerebral cortical thickness data,

Is cortical thickness data of the j-th object subjected to noise filtering by the noise removal procedure in consideration of the cut-off dimension,
Wherein the cut-off dimension for removal of the high frequency band is determined by setting the G to 0.025. 10. The method of claim 9,
Reducing a dimension of the first feature vector and the second feature vector based on a principal component analysis (PCA) transformation;
Generating a new training set based on the first feature vector of the reduced dimension and the second feature vector of the reduced dimension; And
Using linear discriminant analysis (LDA) to maximize variance between the normal group and the Alzheimer's disease patient group based on the new training set and to minimize variance within the normal group and to minimize variance within the Alzheimer's disease patient group Wherein the determination is based on the step of generating a criterion.

Images