KR102125127B1 - Method of brain disorder diagnosis via deep learning - Google Patents

Abstract

The present invention relates to a method of diagnosing deep learning-based brain disease that performs premise diagnosis through diagnosis of each region of interest in a magnetic resonance image (MRI) of the brain.
Deep learning based brain disease diagnosis method according to an embodiment of the present invention comprises the steps of obtaining a brain volume map having the brain volume information from the input magnetic resonance image; Dividing a plurality of regions of interest by dividing an anatomical region from the brain volume map, and extracting a predetermined number of samples for each region of interest; Acquiring a disease probability as many as the predetermined number of samples of the extracted predetermined number of samples using a deep neural network different for each sample; Calculating a disease probability of a region of interest for each region of interest by integrating the disease probability acquired by the number of samples for each region of interest; And outputting a final diagnosis result using the calculated disease probability of all regions of interest as an input value of the final classifier.

Description

METHOD OF BRAIN DISORDER DIAGNOSIS VIA DEEP LEARNING

The present invention relates to a method for diagnosing brain disease, and more specifically, performs diagnosis through deep learning for each region of interest (ROI) in a magnetic resonance image (MRI) of the brain, and each interest It is related to a deep learning-based brain disease diagnosis method that performs overall diagnosis by integrating diagnosis by region.

With the development of society, aging is occurring, and accordingly, degenerative brain diseases are increasing. In the case of brain disease, it is necessary to detect and diagnose lesions without invasion.

The brain imaging test using MRI is a non-invasive test method that can test various diseases caused by structural and functional changes in the brain.

Although the appearance of the brain of each patient is different, the structure is identically changed with a uniform template to check the degree of change in common, and a method of checking the degree of change in each disease is used.

Recently, as a deep neural network technique for deep learning, which is a kind of machine learning, has been developed, it is known that a diagnostic accuracy of a medical image analysis system through a computer has high performance. Korean Patent Registration No. 10-1754291 is a related prior document.

However, the existing diagnostic system does not provide a diagnosis result for each part of the brain region because the entire brain is viewed and judged.

Particularly, a convolutional neural network (CNN), which is a type of deep neural network that has recently shown good performance, is a method of extracting features of a receptive field through a convolution kernel, a sliding window ( Sliding window) method extracts structural features of the input image.

However, if the progression of the bottle is insignificant, since there may be little morphological difference, it is difficult to detect by the method through the sliding window of the receiving area, and it requires a lot of data because there are many parameters to be learned.

Therefore, it is necessary to study a method of diagnosing a brain disease that can help a patient's understanding through visualization of abnormalities in each region, as well as assisting a doctor in making a decision using a magnetic resonance imaging.

An object of the present invention is to provide a deep learning-based brain disease diagnosis method capable of more reliably diagnosing detailed diagnostic information through diagnostic information for each region of interest, using a machine learning method utilizing deep learning that shows good performance in recent years. Has to do.

In order to achieve the above object, according to an embodiment of the present invention, obtaining a brain volume map having brain volume information from the input magnetic resonance image; Dividing a plurality of regions of interest by dividing an anatomical region from the brain volume map, and extracting a predetermined number of samples for each region of interest; Acquiring a disease probability as many as the predetermined number of samples of the extracted predetermined number of samples using deep neural networks different for each sample; Calculating a disease probability of a region of interest for each region of interest by combining disease probabilities obtained by the number of samples for each region of interest; And outputting a final diagnosis result using the calculated disease probability of all regions of interest as an input value of the final classifier.

The deep learning-based brain disease diagnosis method according to an embodiment of the present invention provides a method of more reliably diagnosing brain disease from a limited number of brain images and providing meaningful information for analysis of diagnosis for each region of interest.

According to an embodiment of the present invention, it is possible to not only help the doctor's decision making with the magnetic resonance image alone, but also to help the patient's understanding through visualization of abnormalities in each region.

1 is a block diagram of an apparatus for diagnosing brain disease according to an embodiment of the present invention.
2 is a flowchart illustrating a method for diagnosing a deep learning-based brain disease related to an embodiment of the present invention.
3 is a view for explaining a deep learning-based brain disease diagnosis method related to an embodiment of the present invention.
Figure 4 shows a framework of a model used in the diagnosis of brain disease based on magnetic resonance imaging according to an embodiment of the present invention.
5 is a brain map showing the probability of a patient with Alzheimer's disease by region classified by a deep learning-based brain disease diagnosis method related to an embodiment of the present invention.
6 is a weighted brain map related to an embodiment of the present invention.

Hereinafter, a method and apparatus for diagnosing a deep learning-based brain disease related to an embodiment of the present invention will be described with reference to the drawings.

As used herein, a singular expression includes a plural expression unless the context clearly indicates otherwise. In this specification, the terms "consisting of" or "comprising" should not be construed as including all of the various components, or various steps described in the specification, among which some components or some steps It may not be included, or it should be construed to further include additional components or steps.

1 is a block diagram of an apparatus for diagnosing brain disease according to an embodiment of the present invention.

As illustrated, the brain disease diagnosis apparatus 100 may include a pre-processing unit 110, a disease probability calculation unit 120 for each region of interest, and a final diagnosis unit 130.

The pre-processing unit 110 may pre-process the input magnetic resonance image to separate the magnetic resonance image into a plurality of regions of interest. The pre-processing unit 110 may extract a predetermined number of samples for each of the plurality of regions of interest.

The disease probability calculation unit 120 for each region of interest extracts features using a deep neural network different for each sample for a predetermined number of samples corresponding to each region of interest, and disease for each region of interest using the extracted features Probability can be calculated.

The final diagnosis unit 130 may integrate the disease probability of each region of interest as an input value and output the final diagnosis result of the disease. The final diagnosis result may determine whether the subject is a patient or normal.

2 is a flowchart illustrating a deep learning-based brain disease diagnosis method related to an embodiment of the present invention, and FIG. 3 is a diagram illustrating a deep learning-based brain disease diagnosis method related to an embodiment of the present invention.

First, the pre-processing unit 110 receives a magnetic resonance image of the entire brain of the subject to be diagnosed. The pre-processing unit 110 may perform pre-processing of the input magnetic resonance image (S210, S310). The pre-treatment process may be performed in the following process.

In order to check the volume change of the brain, the brains of different patients should be mapped to the same space and the changes must be confirmed. The pre-processing unit 110 may acquire a brain volume map mapped in the same space. The pre-processing unit 110 may obtain a brain volume map having volume information of the brain through a matching process and a brain tissue division process. Specifically, the pre-processing unit 110 first works to match the symmetry of all input images. Next, the pre-processing unit 110 removes regions (cranium, cerebellum, etc.) except for the brain region to be used for analysis. The pre-processing unit 110 separates the tissues (white matter, gray matter, and cerebrospinal fluid) of the brain after undergoing a shift correction operation to compensate for errors in the obtained brain images. The pre-processing unit 110 maps the brain tissue of all subjects to the same template space and obtains a volume map of the brain through a matching process. The obtained brain volume map can be divided into regions of interest through labeled templates for diagnosis by anatomical regions, and only gray matter regions can be used. Therefore, each voxel value of the 3D image represents the brain volume of the corresponding region.

In addition, the pre-processing unit 110 may extract anatomical regions from the obtained brain volume map and extract a plurality of samples having randomly selected voxel values for each region of interest. For example, the pre-processing unit 110 first selects candidate groups of voxels for each region of interest showing a large difference between groups through a statistical analysis method, and sets a sample having a predetermined number of voxels (a predetermined number corresponding to each region of interest) Of the sample).

Meanwhile, the number of samples to be extracted may be adjusted differently according to the size of the region of interest. For example, the larger the size of the region of interest, the larger the number of samples to be extracted can be adjusted. In addition, a number of samples having k random voxels may be extracted using a simple random reconstruction extraction (eg, bootstrap sampling) technique.

The disease probability calculation unit 120 for each region of interest may obtain a disease probability as many as the predetermined number of samples by using a deep neural network different for each sample of a predetermined number of samples extracted for each region of interest. In addition, the disease probability calculation unit 120 for each region of interest may calculate a disease probability of the region of interest for each region of interest by combining disease probabilities obtained by a predetermined number of samples (eg, average, maximum, minimum, etc.) Yes (S220, S320).

For example, the disease probability calculation unit 120 for each region of interest is a stacked denoising auto-encoder, which is one of deep neural networks, and is used as input from brain volume data. A model that learns the nonlinear relationship of a normal class can be extracted.

According to an embodiment of the present invention, a method of creating a model for each sample in each region of interest may be used to diagnose characteristics of a randomly extracted voxel sample for each region of interest. Hereinafter, in the embodiment, a stacked denoising auto-encoder (SDAE) is used as a method of extracting an inherent nonlinear feature. DAE is a type of autoencoder and is expressed as Equation 1 and Equation 2 below.

은 k차원을 갖는 표본 x∈R^k에 잡음이 추가된 벡터이다. Ψ함수는 은닉 계층 h의 유닛 값이 입력 벡터와 비선형적 관계를 갖도록 한다. 출력 계층을 위한 가중치 W′와 편향치 b′는 o가 원래의 표본 x값과 같아지도록 하는 파라미터이다. 이때, h의 유닛값은 입력 벡터로부터 원래의 표본을 잘 표현해 줄 수 있는 특징 벡터라고 볼 수 있다.In the above equation, W and b represent the weight vector and bias vector for the hidden layer, and W'and b'represent the weight vector and bias vector for the output layer. Ψ is a nonlinear activity function. In Equation 1 above

Is a vector with noise added to the sample x∈R ^k with k-dimensionality. The Ψ function causes the unit value of the hidden layer h to have a nonlinear relationship with the input vector. The weight W'and the bias value b'for the output layer are parameters such that o is equal to the original sample x value. At this time, the unit value of h can be regarded as a feature vector that can express the original sample well from the input vector.

The model used in the following examples is shown in FIG. 4.

Figure 4 shows a framework of a model used in the diagnosis of brain disease based on magnetic resonance imaging according to an embodiment of the present invention.

As illustrated, x having k voxels for each region of interest is created, and the probability can be derived through SDAE. Ensemble z that collects these probabilities and finally diagnose it as the input of the support vector machine. In this embodiment, the SDAE may use a softmax function for classification in the final output layer that stacks these DAEs. The output value of this Softmax is the probability that the sample belongs to each class, indicating the probability of being a patient and the probability of being a normal person. For quantitative analysis of each region of interest, N SDAEs of the same structure were trained in one region of interest, and the average z of the probability p of the patient can be obtained from the N models and used as a feature of one region of interest.

In addition, the feature and the overall feature vector of each region of interest may be expressed as Equations 3 and 4 below.

는 n번째 표본에서 r번째 관심 영역이 환자일 확률이고, z_r은 r번째 관심 영역에서 전체 N개 모델에서의 환자일 확률의 평균이다. v는 전체 관심 영역의 해당 질환(예: 알츠하이머병(Alzheimer’s Disease: AD))일 특징을 앙상블한 특징 벡터라고 볼 수 있다. 이 특징 벡터는 분류기 서포트벡터머신(Support Vector Machine: SVM)의 입력으로 사용될 수 있다. 여기서 서포트벡터머신은 최종 진단부(130)에서 사용하는 최종 분류기의 일례이다. That is, if the entire region of interest is R, an equation for obtaining the feature z _r of each region of interest can be expressed by Equation (3).

Is the probability that the r-th region of interest is a patient in the n-th sample, and z _r is the average of the probability of being patients in all N models in the r-th region of interest. v can be regarded as an ensemble feature vector that is a feature of the entire region of interest (eg, Alzheimer's Disease (AD)). This feature vector can be used as an input to a classifier support vector machine (SVM). Here, the support vector machine is an example of a final classifier used in the final diagnosis unit 130.

That is, the final diagnosis unit 130 may output the final diagnosis result by using all the probability of disease for the region of interest as an input value (S230, S330). In Figure 4, a support vector machine was used for the final diagnosis.

In addition, the final classifier may learn weights for each region of interest and classify patients and normal persons through the learning.

For example, when each region of interest is A, B, or C, the classifier determines that the regions of B and C are more important for the disease (e.g., Alzheimer's disease). Even if you have a very low probability at, if you are high at B, you can diagnose this person as a patient. Of course, the opposite is also possible. That is, it is possible to obtain a probability of a disease from a deep neural network and to compare the normal group and the patient group by combining the probabilities in the final classifier to diagnose by giving a weight to a significant area.

5 is a brain map showing the probability of a patient with Alzheimer's disease by region classified by a deep learning-based brain disease diagnosis method related to an embodiment of the present invention, and FIG. 6 is a weighted brain map related to an embodiment of the present invention .

In this example, for the quantitative analysis of the area, the randomized sample was reconstructed through SDAE to determine the probability of a patient with Alzheimer's disease, and characterized by the average of the probability to diagnose in the entire region of interest. The classification performance considering the entire region of interest showed an accuracy of 76.36%. As shown in FIG. 5, the probability of a patient with Alzheimer's disease for each region of interest can be visualized so that the diagnosis results can be visualized at a glance, and quantitative analysis for each region is possible. In addition, as shown in Figure 3 and S, it is possible to check which region can be used as a potential image biomarker to diagnose Alzheimer's disease by mapping and visualizing the weighted brain regions of the support vector machine.

As described above, the deep learning-based brain disease diagnosis method according to an embodiment of the present invention more reliably diagnoses brain disease from a limited number of brain images and provides significant information for analysis of diagnosis for each region of interest Provide a method.

According to an embodiment of the present invention, it is possible to not only help the doctor's decision making with the magnetic resonance image alone, but also to help the patient's understanding through visualization of abnormalities in each region.

The above-described deep learning-based brain disease diagnosis method is implemented in a program instruction form that can be executed through various computer means and can be recorded in a computer-readable recording medium. In this case, the computer-readable recording medium may include program instructions, data files, data structures, or the like alone or in combination. Meanwhile, the program instructions recorded on the recording medium may be specially designed and configured for the present invention or may be known to those skilled in computer software.

Computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs, DVDs, and magnetic media such as floppy disks. Included are hardware devices specifically configured to store and execute program instructions, such as magneto-optical media, and memory storage devices such as ROM, RAM, flash memory, and solid state drives (SSDs).

On the other hand, such a recording medium may be a transmission medium such as an optical or metal wire or waveguide including a carrier wave that transmits a signal specifying a program command, data structure, or the like.

In addition, the program instructions include not only machine language codes produced by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like. The hardware device described above may be configured to operate as one or more software modules to perform the operation of the present invention, and vice versa.

The deep learning-based brain disease diagnosis method and apparatus described above are not limited to the configuration and method of the above-described embodiments, and the above embodiments are all or part of each embodiment so that various modifications can be made. May be selectively combined.

100: brain disease diagnosis device
110: pre-processing unit
120: disease probability calculation unit for each region of interest
130: final diagnosis unit

Claims (5)

Obtaining a brain volume map having brain volume information from the input magnetic resonance image;
Dividing the anatomical region from the brain volume map into a plurality of regions of interest including a first region of interest and a second region of interest, and extracting a plurality of samples corresponding to each region of interest, respectively. The size of the region is larger than the size of the second region of interest, and the number of samples facing the first region of interest is greater than the number of samples corresponding to the second region of interest-;
Acquiring disease probabilities as many as the number of the plurality of samples extracted from the plurality of samples corresponding to the regions of interest using different deep neural networks for each of the samples;
Calculating disease probability of a region of interest for each region of interest by combining disease probabilities obtained by the number of samples; And
And outputting a final diagnosis result using the calculated disease probability of all regions of interest as an input value of the final classifier,
Calculating the probability of disease in the region of interest
A method of diagnosing a brain disease based on deep learning, comprising the step of integrating a plurality of disease probabilities obtained by the number of samples and calculating the probability of a disease. According to claim 1,
The final classifier is a deep learning-based brain disease diagnosis method characterized by learning weights for each region of interest and classifying patients and normal people through the learning. delete delete delete

Images