KR20190132832A - Method and apparatus for predicting amyloid positive or negative based on deep learning - Google Patents

Abstract

The present invention relates to a method for predicting amyloid positive or negative based on deep learning which improves accuracy of determination of amyloid positive or negative. According to an embodiment of the present invention, the method for predicting amyloid positive or negative based on deep learning comprises: a preprocessing step of generating a slice image from a first and a second brain photographing image and a mask for the slice image based on a determination reference area for predicting amyloid positive or negative; a step of generating input data based on the slice image and the mask generated in the preprocessing step by using deep learning or machine learning; a step of extracting and learning a score for predicting amyloid positive or negative from the input data generated by using the deep learning or the machine learning; and a step of generating an algorithm model for diagnosing amyloid positive or negative based on the score extracted and learned by using the deep learning or the machine learning, and determining amyloid positive or negative by the generated algorithm model.

Description

METHOD AND APPARATUS FOR PREDICTING AMYLOID POSITIVE OR NEGATIVE BASED ON DEEP LEARNING}

The present invention relates to a method and apparatus for predicting amyloid positive or negative on the basis of deep learning, and more particularly, deep learning using amyloid PET (PET) image, which is one of brain imaging images for diagnosing brain disease. The present invention relates to a method and system for predicting amyloid positive or negative from a similar point of view to human intuition.

Recent advances in science and imaging techniques have enabled computerized tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (positron) to diagnose and predict brain diseases. emission tomography (PET). Among them, amyloid PET test is known to be effective in diagnosing degenerative brain diseases including Alzheimer's disease and dementia.

Amyloid PET test is to detect whether abnormal form of amyloid, which is produced by abnormality in protein metabolism, is deposited in the brain and injected with special contrast agent that only reacts with amyloid. It is to determine whether it is positive or negative.

Conventionally, a clinician directly selects an image of a patient's photographed amyloid PET image that has a clear output or easy deposition to diagnose the amyloid PET image. Since the selection of amyloid PET images through such a process requires a clinician to verify and confirm a large number of amyloid PET images, there is a problem that accurate selection of amyloid PET images required for diagnosis is difficult.

In addition, since a conventional feature extraction methodology is used in extracting a feature of a pet image for analysis of a selected amyloid PET image, it does not properly process data including a new environment or condition, and thus appropriate amyloid positive or There is a problem that is difficult to predict whether the voice.

Japanese Patent Publication No. 5468905 (2014.02.07)

The present invention is to solve the problems described above, the features and scores of brain imaging images including amyloid PET using deep learning or machine learning in the conventional brain imaging image analysis process that was selected or identified by human judgment It is an object of the present invention to provide a method and apparatus that can be automated through a process such as extraction.

In addition, the present invention improves the accuracy of amyloid positive or negative judgment through continuous learning process for the region-specific features of brain imaging images including amyloid PET using deep learning or machine learning, thereby increasing the reliability of amyloid positive or negative It is an object of the present invention to provide a method and apparatus for enabling high prediction to be performed.

A method for predicting amyloid positive or negative based on deep learning according to an embodiment of the present invention, the first brain imaging image and the second based on the reference region for predicting amyloid positive or negative A preprocessing step of generating a slice image and a mask for the slice image from the brain imaging image, and generating input data based on the slice image and the mask generated in the preprocessing step using deep learning or machine learning. Extracting and learning a score for predicting amyloid positive or negative from input data generated using deep learning or machine learning, and based on the extracted and learned score using deep learning or machine learning. Create an algorithmic model for diagnosing positive or negative, and generate the It may include determining the amyloid positive or negative through.

The preprocessing step according to an embodiment of the present invention may include generating a split image of the first brain image, generating a matched image of the first brain image and the second brain image, and generating the split image. Selecting a first slice image based on an output state of the determination base region, extracting a second slice image corresponding to the first slice image from the registration image, and generating a mask for the determination reference region of the first slice image The determination reference area may include a determination basic area.

According to an embodiment of the present invention, the first brain imaging image is an MRI T1 volume image, the second brain imaging image is an amyloid PET volume image, and the judgment basis region is a temporal region or a frontal lobe. And a frontal region, a posterior cingulate-precuneus region, and a parietal region, wherein the judgment reference region is a white matter region of each of the left side and the right side and the left side or the right side of each of the basal regions. white matter (wm) and gray matter (gm).

The generating of input data according to an embodiment of the present disclosure may include extracting feature data of a determination reference region based on a slice image in which a mask is generated using deep learning or machine learning, and deep learning. Alternatively, the method may include extracting label data for assigning scores based on the extracted feature data using machine learning.

A deep convolutional network (FCN) algorithm may be included in deep learning or machine learning used in the step of extracting feature data according to an embodiment of the present invention.

Deep learning or machine learning used in the step of extracting and learning the score according to an embodiment of the present invention may include a regression algorithm and a support vector machine (SVM) algorithm.

In generating an algorithm model according to an embodiment of the present invention and determining amyloid positive or negative based on the generated algorithm model, an algorithm model may be generated based on a score extracted and learned using a support vector machine algorithm. Can be.

The apparatus for predicting amyloid positive or negative based on deep learning according to an embodiment of the present invention, the first brain imaging image and the second based on the reference region for predicting amyloid positive or negative Preprocessing unit for generating a slice image and a mask for the slice image from the brain image, the first data generation using deep learning to generate input data based on the slice image and the mask generated in the preprocessing step An input data generator comprising a module and a second data generation module using machine learning; a first scoring module using deep learning to extract and learn a score for predicting amyloid positive or negative from the generated input data And a scoring unit and a second scoring module using machine learning; Generating an algorithmic model for diagnosing amyloid positive or negative based on the wet score, and using the first positive and negative judgment module and machine learning using deep learning to determine amyloid positive or negative through the generated algorithm model. 2 may include an amyloid positive voice prediction unit including a positive voice determination module.

The preprocessor according to an embodiment of the present invention, the image segmentation unit for generating a split image for the first brain image, the image registration unit for generating a registration image for the first brain image and the second brain image, the segmentation A slice image generator for selecting a first slice image based on the output state of the determination base region for the image, and extracting a second slice image corresponding to the first slice image from the matched image, and a determination reference region of the first slice image. The image mask generator may be configured to generate a mask, and the determination reference region may include a determination basic region.

According to an embodiment of the present invention, the first brain imaging image is an MRI T1 volume image, the second brain imaging image is an amyloid PET volume image, and the judgment basis region is a temporal region or a frontal lobe. And a frontal region, a posterior cingulate-precuneus region, and a parietal region, wherein the judgment reference region is a white matter region of each of the left side and the right side and the left side or the right side of each of the basal regions. white matter (wm) and gray matter (gm).

The first data generation module according to an embodiment of the present disclosure may include a first feature data extractor and a deep extracting feature data for a determination reference region based on a slice image in which a mask is generated using deep learning. A first label data extracting unit extracting label data for assigning scores based on the feature data extracted using running;

The second data generation module according to an embodiment of the present invention may include a second feature data extractor and a machine for extracting feature data of a determination reference region based on a slice image in which a mask is generated using machine learning. The apparatus may include a second label data extractor configured to extract label data for assigning scores based on the feature data extracted using the learning.

In the deep learning used in the first feature data extractor or the machine learning used in the second feature data extractor according to an embodiment of the present invention, a FCN (Fully Convolutional Network) algorithm may be included.

Deep learning or machine learning used in the scoring unit according to an embodiment of the present invention may include a regression algorithm and a support vector machine (SVM) algorithm.

In the amyloid positive-negative prediction unit according to an embodiment of the present invention, an algorithm model may be generated based on the scores extracted and learned using a support vector machine algorithm.

Meanwhile, according to an embodiment of the present invention, a computer-readable recording medium having recorded thereon a program for executing the above-described method on a computer can be provided.

According to the prediction method and apparatus provided as an embodiment of the present invention, the entire process for predicting amyloid positive or negative using deep learning or machine learning can be automated to further improve user convenience.

Also, rather than diagnosing amyloid positive or negative according to the human judgment on the brain imaging image as in the prior art, it is determined by analyzing features of the region of the brain imaging image including amyloid PET using deep learning or machine learning. As such, the reliability and accuracy of prediction results for amyloid positive or negative can be further improved.

In addition, since deep learning or machine learning can be used to predict amyloid positive or negative from a viewpoint similar to human intuition, it can be easily applied to multi-sensors or multi-devices regardless of the source of brain imaging data.

1 is a flowchart illustrating a method for predicting amyloid positive or negative based on deep learning according to an embodiment of the present invention.
2 is a flowchart illustrating a pretreatment step of a method for predicting amyloid positive or negative based on deep learning according to an embodiment of the present invention.
3 is a flowchart illustrating an input data generation step of a method for predicting amyloid positive or negative based on deep learning according to an embodiment of the present invention.
4 illustrates a process of predicting amyloid positive or negative by using deep learning according to an embodiment of the present invention.
5 illustrates a process of predicting amyloid positive or negative by using machine learning according to an embodiment of the present invention.
6 is a block diagram showing an apparatus for predicting amyloid positive or negative based on deep learning according to an embodiment of the present invention.
7 is a block diagram illustrating a preprocessor included in an apparatus for predicting amyloid positive or negative based on deep learning according to an embodiment of the present invention.
8 is a block diagram illustrating a prediction process of an apparatus for predicting amyloid positive or negative based on deep learning according to an embodiment of the present invention.

Terms used herein will be briefly described and the present invention will be described in detail.

The terms used in the present invention have been selected as widely used general terms as possible in consideration of the functions in the present invention, but this may vary according to the intention or precedent of the person skilled in the art, the emergence of new technologies and the like. In addition, in certain cases, there is also a term arbitrarily selected by the applicant, in which case the meaning will be described in detail in the description of the invention. Therefore, the terms used in the present invention should be defined based on the meanings of the terms and the contents throughout the present invention, rather than the names of the simple terms.

When any part of the specification is to "include" any component, this means that it may further include other components, except to exclude other components unless otherwise stated. In addition, the terms "... unit", "module", etc. described in the specification mean a unit for processing at least one function or operation, which may be implemented in hardware or software or a combination of hardware and software. .

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art may easily implement the present invention. As those skilled in the art would realize, the described embodiments may be modified in various different ways, all without departing from the spirit or scope of the present invention. In the drawings, parts irrelevant to the description are omitted in order to clearly describe the present invention, and like reference numerals designate like parts throughout the specification.

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

1 is a flowchart illustrating a method for predicting amyloid positive or negative based on deep learning according to an embodiment of the present invention.

Referring to FIG. 1, a method for predicting amyloid positive or negative based on deep learning according to an embodiment of the present invention may include: a first method based on a determination reference region for predicting amyloid positive or negative; Preprocessing step (S100) of generating a mask for the slice image and the slice image from the brain image and the second brain image, the slice image and the mask generated in the preprocessing step using deep learning or machine learning Generating input data based on the step S200, extracting and learning a score for predicting amyloid positive or negative from the input data generated using deep learning or machine learning (S300), and deep learning Or generate algorithmic models for diagnosing amyloid positive or negative based on the extracted and learned scores using machine learning. Through the, and the resulting algorithm model may include the step (S400) to determine the amyloid positive or negative.

In this case, the determination reference region may refer to regions of the brain in the brain imaging image which is a reference for generating input data, extracting scores, and learning required to predict amyloid positive or negative. In other words, the judgment reference region refers to regions of the brain that are output from the brain imaging image as an overall criterion in each process for predicting amyloid positive or negative, and these judgment reference regions are described in more detail together with the judgment basis region to be described later. Do it.

In the method for predicting amyloid positive or negative based on deep learning according to an embodiment of the present invention, a preprocessing step (S100) for a first brain image and a second brain image is performed in common. Subsequently, it may be broadly classified into a process using deep learning and a process using machine learning. In other words, after the preprocessing step S100 is performed, input data is generated using deep learning, an algorithm model is generated by extracting and learning a score, and a process of determining positive sound and input data is generated using machine learning. Extracting and learning the scores to generate an algorithmic model and determine the positive sound may be performed separately. For example, machine learning may not be used to generate input data, and machine learning may not be used to extract and learn scores.

That is, in the case of using deep learning, the process of using deep learning is integrally performed, and in the case of using machine learning, the process of using machine learning may be integrally performed. Therefore, the subjects performing each step using deep learning (ie, the first data generation module 210, the first scoring module 310, and the first positive voice determination module 410) and each step using machine learning The subjects (ie, the second data generation module 220, the second scoring module 320, and the second positive voice determination module 420) for performing the operation are separately formed, and perform the subject and machine learning using deep learning. Data transmission between performing subjects to be used may not be performed. More details will be described with reference to FIG. 8 to be described later.

2 is a flowchart showing a pre-processing step (S100) of a method for predicting amyloid positive or negative based on deep learning according to an embodiment of the present invention.

Referring to FIG. 2, the preprocessing step (S100) according to an embodiment of the present invention may include generating a split image for the first brain imaging image (S110), a first brain imaging image, and a second brain imaging image. Generating a matched image for the first slice image based on an output state of the determination base region for the split image (S130); and a second slice image corresponding to the first slice image from the matched image. And extracting a mask for the determination reference region of the first slice image (S150), and the determination reference region may include a determination basic region.

In this case, the first brain imaging image according to an embodiment of the present invention may be an MRI T1 volume image, and the second brain imaging image may be an amyloid PET volume image.

For example, if the first brain imaging image is an MRI T1 volume image, the segmented image is an MRI T1 segmented image, which is a ventricular image, a temporal image, a frontal precuneus image, and a parietal lobe. It may be a (parietal) image. As described above, the MRI T1 split image is generated to facilitate identification of the determination basis region.

The determination basis region according to an embodiment of the present invention may refer to regions of the brain output from the brain imaging image required for selecting and extracting slice images. More specifically, the judgment basis region refers to an area of the brain that can be used as a lesion basis for predicting positive or negative amyloid. For example, the judgment basis region may include a temporal region, a frontal region, a posterior cingulate-precuneus region, and a parietal region. Therefore, when the segmented image is an MRI T1 segmented image (ie ventricle, temporal lobe, anterior wedge lobe, parietal lobe image), the first slice image is a judgment base region (ie temporal lobe, frontal lobe, posterior capsular and parietal lobe region) of the MRI T1 segmented image. The output state of can be selected from the best split image.

The determination reference region according to an embodiment of the present invention includes the above-described determination basic region, which can be understood as an extension of the determination basic region. In other words, if the judgment basis region is a criterion for generating basic data used to predict amyloid positive or negative, the judgment reference region uses deep learning or machine learning to predict amyloid positive or negative based on the judgment basis region. As a criterion for use in the overall process, it may refer to areas of the brain for predicting amyloid positive or negative including a judgment basis area.

For example, the determination reference region may include a white matter region (wm) and a gray matter region (gm) of each of the left side portion and the right side portion of each of the determination basis regions, and the left side portion or the right side portion, respectively. If the judgment basis region is a total of four regions, the temporal lobe region, the frontal lobe region, the posterior target lingual region and the parietal lobe region, the judgment reference region may include the white and gray regions of each of the left, right, and left or right portions of each region. Therefore, there may be a total of 16 areas. Therefore, when a mask is generated in the first slice image according to the above-described example, the mask is generated in a total of 16 areas including white areas or gray areas of each of the left and right parts and the left weight or right part of each of the determination base areas. Can be.

3 is a flowchart illustrating an input data generation step S200 of a method for predicting amyloid positive or negative based on deep learning according to an embodiment of the present invention.

Referring to FIG. 3, in operation S200 of generating input data according to an embodiment of the present invention, a feature of a determination reference region based on a slice image in which a mask is generated using deep learning or machine learning is performed. The method may include extracting data (S210) and extracting label data for assigning scores based on the extracted feature data using deep learning or machine learning (S220). Therefore, the input data may include feature data and label data extracted through the above-described steps.

4 illustrates a process of predicting amyloid positive or negative using deep learning according to an embodiment of the present invention, and FIG. 5 illustrates predicting amyloid positive or negative using machine learning according to an embodiment of the present invention. Indicate the process.

The step of extracting feature data according to an embodiment of the present invention looks at the case of using deep learning and the case of using machine learning.

First, referring to FIG. 4, in the extracting of feature data using deep learning, generating a padding image as input data, and generating a padding image corresponding to a mask for a determination reference region generated in the first slice image. And extracting a corresponding region of the two-slice image and reducing and inserting the corresponding region of the extracted second slice image into the padding image. For example, in order to extract feature data using deep learning, first, an image padded with −2 of [50 × 50] may be generated as input data. Next, eight areas including the left part and the right part of each of the judgment basis areas generated in the first slice image (eg, left temporal lobe, left frontal lobe, left posterior prelingual, left parietal lobe, and right temporal lobe, right frontal lobe, right) A region of the second slice image corresponding to the mask of the posterior prefrontal and right parietal lobe region may be extracted. The region of the second slice image of the extracted eight regions may be reduced to [50x50] and inserted into the padded image. Through this process, feature data included in the input data can be extracted.

Referring to FIG. 5, in the extracting of feature data using machine learning, estimating an average standard uptake value ratio (SUVR) value of a mask for a determination reference region generated in a first slice image and an estimated average SUVR Performing normalization on the value. For example, in order to extract feature data using machine learning, eight mask areas (eg, left white temporal lobe, left white frontal lobe, left frontal lobe) corresponding to the white matter area among the above 16 total reference areas of the first slice image Mean SUVR values of the white matter region of the posterior epiglottis, the left parietal parietal lobe and the right parietal temporal lobe, the right parietal frontal lobe, the right parietal posterior capsule, and the right parietal parietal lobe) can be estimated. In addition, the average SUVR value for each area of the mask for the aforementioned 16 determination reference areas may be estimated. Next, the estimated SUVR values may be normalized, and the level of gray or white matter values for each reference reference region may be calculated. By using the result data and the second slice image through the normalization and level calculation process, the feature data included in the input data may be extracted.

Also, a deep convolutional network (FCN) algorithm may be included in deep learning or machine learning used in the step S210 of extracting feature data according to an embodiment of the present invention. FCN algorithm is a deep learning network algorithm that consists of several convolutional blocks, and there is no lastly fully connected layer. Since the FCN algorithm does not lose location information due to its characteristics, more accurate information can be obtained. By using the FCN algorithm, each step of extracting the aforementioned feature data can be effectively performed.

In the extracting the label data according to an embodiment of the present invention (S220), when using deep learning or machine learning, label data may be extracted by allocating a score corresponding to the feature data generated through each process. have. For example, the score can be a value from 0 to 3.

Referring to FIG. 1, after the input data generation process S200 is performed as described above, the step S300 of extracting and learning the score from the input data may be performed. The steps of extracting and learning the scores will also be described in detail by using deep learning and machine learning.

First, referring to FIG. 4, in the step of extracting and learning a score by using deep learning, the extraction and learning of the score may be performed by using a gradient gradient method (SGD) as a regression algorithm. In this case, the loss may be estimated through a mean square error, and the output may be linearly activated. Gradient descent is a method of directly adjusting the weight of a deep learning neural network by calculating an error for each learning data. Since the weight is updated every single learning data, the gradient learning is performed while the performance of the neural network changes. Therefore, by using the gradient descent method, it is possible to heuristically select a learning interruption point.

Referring to FIG. 5, in the step of extracting and learning a score using machine learning, a regression algorithm may be used when the score is extracted as a continuous value, and the score is extracted as a binary value. In this case, a support vector machine (SVM) may be used. In this case, the support vector machine algorithm is a traditional machine learning methodology with high accuracy. The support vector machine algorithm generates the most general model using the concept of the support vector. Using this machine learning (regression algorithm, sopot vector machine), it is possible to easily extract and learn the score for each reference reference region from a large number of SUVR values extracted using the FCN algorithm.

When the score is extracted and learned through the above-described process (S300), in the step (S400) of generating an algorithm model according to an embodiment of the present invention and determining amyloid positive or negative through the generated algorithm model, the support vector machine An algorithm may be used to generate an algorithm model based on the extracted and learned scores.

That is, a support vector machine algorithm may be used to generate an algorithm model for more general amyloid positive or negative judgment. Therefore, when data for amyloid positive or negative judgment is input to an algorithm model generated using a support vector machine algorithm, it is related to the origin of the data (ie, data generated based on brain imaging images of an arbitrary third party). Can predict amyloid positive or negative from a more objective and general perspective.

In addition, as learning about the brain imaging image is made, it is possible to continuously improve the accuracy of amyloid positive or negative judgment of the algorithm model.

FIG. 6 is a block diagram illustrating an apparatus 1000 for predicting amyloid positive or negative based on deep learning according to an embodiment of the present invention, and FIG. 7 is based on deep learning according to an embodiment of the present invention. It is a block diagram showing the preprocessing unit 100 included in the device 1000 for predicting amyloid positive or negative.

8 is a block diagram illustrating a prediction process of the apparatus 1000 for predicting amyloid positive or negative based on deep learning according to an embodiment of the present invention.

6 and 8, an apparatus for predicting amyloid positive or negative based on deep learning according to an embodiment of the present disclosure may include a first reference based on a determination reference region for predicting amyloid positive or negative; A preprocessing unit 100 generating a mask for the slice image and the slice image from the brain image and the second brain image, and using deep learning to generate input data based on the slice image and the mask generated in the preprocessing step. An input data generator 200 including a first data generation module 210 and a second data generation module 220 using machine learning, and extracting and learning a score for predicting amyloid positive or negative from the generated input data Including a first scoring module 310 using deep learning and a second scoring module 320 using machine learning. A first quantity using deep learning to generate an algorithm model for diagnosing amyloid positive or negative based on the scoring unit 300 and the extracted and learned scores, and to determine the amyloid positive or negative through the generated algorithm model It may include an amyloid positive voice predictor 400 including a voice judgment module 410 and a second positive voice determination module 420 using machine learning.

Referring to FIG. 8, except for the preprocessor 100, the input data generator 200, the scoring unit 300, and the amyloid positive-negative prediction unit 400 are each a performance module using deep learning and a performance module using machine learning. It may include, the performance module using the deep learning and the performance module using the machine learning may not be connected to each other. Accordingly, input data generated by the first data generation module 210 may not be transmitted to the second scoring module 320, and data regarding scores extracted and learned through the second scoring module 320 may be transmitted in a first amount. It cannot be delivered to the voice determination module 410.

According to an embodiment of the present disclosure, the first data generating module 210 or the second data generating module 220 may be configured as a preprocessor 100 according to an external input or data processing speed applied from a user, whether or not data processing is being performed. The basic data (eg, slice image, mask, etc.) selected by the preprocessing unit 100 and generated by the preprocessing unit 100 may be transferred to the selected data generation module. For example, according to a user's purpose for processing data using deep learning, the first data generation module 210 may be selected by an external input applied by the user to perform data processing using deep learning. In addition, when a new brain image is received but data processing using deep learning is already being performed, the second data generation module 220 is automatically selected by the preprocessor 100 to perform data processing using machine learning. Can be.

Referring to FIG. 6, an apparatus for predicting amyloid positivity or voice based on deep learning according to an embodiment of the present invention includes a communication unit 500 and a brain imaging image for transmitting and receiving brain imaging images and other data. It may further include a database 600 for storing and managing all data. That is, the apparatus 1000 for predicting amyloid positive or negative according to an embodiment of the present invention performs wired / wireless communication with brain imaging apparatuses including an MRI imaging apparatus, an amyloid PET imaging apparatus, and the like through the communication unit 500. The input data, score data, etc. generated in the data processing process as well as the brain imaging image received through the communication unit 500 may be generated and immediately stored in the database 600, and the data stored in the database 600. The deletion may be managed by an external input authorized by the user.

Referring to FIG. 8, the preprocessing unit 100 according to an embodiment of the present disclosure may include an image splitter 110 that generates a split image for a first brain image, a first brain image, and a second brain image. An image matching unit 120 generating a registration image for the image, selecting a first slice image based on an output state of the determination base region for the split image, and a second slice image corresponding to the first slice image from the registration image. And a slice image generation unit 130 for extracting the image and the image mask generation unit 140 for generating a mask for the determination reference region of the first slice image, and the determination reference region may include the determination basis region. . In addition, the preprocessing unit 100 may further include an image transmitting and receiving unit 150 for receiving a brain imaging image received from the communication unit 500, and transmits a slice image generated by the mask.

According to an embodiment of the present invention, the first brain imaging image is an MRI T1 volume image, the second brain imaging image is an amyloid PET volume image, and the judgment base region includes a temporal lobe region, a frontal lobe region, a posterior prefrontal region, and a parietal lobe region. The determination reference region may include a left portion, a right portion, a white matter region, or a gray matter region of each of the determination basis regions.

Referring to FIG. 8, the first data generation module 210 according to an embodiment of the present disclosure may extract first feature data of a determination reference region based on a slice image in which a mask is generated using deep learning. A feature data extractor and a first label data extractor for extracting label data for assigning scores based on feature data extracted using deep learning, the second data generation module according to an embodiment of the present invention; In operation 220, the second feature data extractor extracts feature data of the determination reference region based on the slice image in which the mask is generated using machine learning, and allocates a score based on the feature data extracted using machine learning. It may include a second label data extraction unit for extracting label data for.

In the deep learning used in the first feature data extractor or the machine learning used in the second feature data extractor according to an embodiment of the present invention, a FCN (Fully Convolutional Network) algorithm may be included.

Deep learning or machine learning used in the scoring unit 300 according to an embodiment of the present invention may include a regression algorithm and a support vector machine algorithm.

The amyloid positive-negative predictor 400 according to an embodiment of the present invention may generate an algorithm model based on the extracted and learned scores using a support vector machine algorithm.

With regard to the apparatus according to an embodiment of the present invention, the above description may be applied. Therefore, with respect to the apparatus, the description of the same contents as those of the above-described method is omitted.

Meanwhile, according to an embodiment of the present invention, a computer-readable recording medium having recorded thereon a program for executing the above-described method on a computer can be provided. In other words, the above-described method can be written as a program that can be executed in a computer, and can be implemented in a general-purpose digital computer for operating the program using a computer readable medium. In addition, the structure of the data used in the above-described method can be recorded on the computer-readable medium through various means. A recording medium for recording an executable computer program or code for performing various methods of the present invention should not be understood to include temporary objects, such as carrier waves or signals. The computer readable medium may include a storage medium such as a magnetic storage medium (eg, a ROM, a floppy disk, a hard disk, etc.), an optical reading medium (eg, a CD-ROM, a DVD, etc.).

The foregoing description of the present invention is intended for illustration, and it will be understood by those skilled in the art that the present invention may be easily modified in other specific forms without changing the technical spirit or essential features of the present invention. will be. Therefore, it should be understood that the embodiments described above are exemplary in all respects and not restrictive. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as distributed may be implemented in a combined form.

The scope of the present invention is shown by the following claims rather than the detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalents should be construed as being included in the scope of the present invention. .

100: preprocessing unit 110: image segmentation unit
120: image matching unit 130: slice image generation unit
140: image mask generator 150: image transceiver
200: input data generation unit 210: first data generation module
220: second data generation module
300: scoring unit 310: first scoring module
320: second scoring module
400: amyloid positive negative predictor 410: first positive negative determination module
420: second positive voice determination module
500: communication unit 600: database
1000: device for predicting amyloid positive or negative

Claims (15)

In the method for predicting amyloid positive or negative based on deep learning,
A preprocessing step of generating a slice image and a mask for the slice image from the first brain image and the second brain image based on the determination reference region for predicting the amyloid positive or negative;
Generating input data based on a slice image and a mask generated in the preprocessing step using the deep learning or machine learning;
Extracting and learning a score for predicting the amyloid positive or negative from the generated input data using the deep learning or machine learning; And
Generating an algorithm model for diagnosing the amyloid positive or negative based on the extracted and learned scores using the deep learning or machine learning, and determining the amyloid positive or negative based on the generated algorithm model A method for predicting amyloid positive or negative comprising.
The method of claim 1,
The pretreatment step,
Generating a segmented image of the first brain image;
Generating a registration image of the first brain imaging image and the second brain imaging image;
Selecting a first slice image based on an output state of the determination base region with respect to the divided image;
Extracting a second slice image corresponding to the first slice image from the registration image; And
Generating a mask for the determination reference region of the first slice image,
And the determination reference region comprises the determination basis region.
The method of claim 2,
The first brain image is an MRI T1 volume image, the second brain image is an amyloid PET volume image,
The judgment basis region includes a temporal region, a frontal region, a posterior cingulate-precuneus region, and a parietal region,
The determination reference region may include a white matter region (wm) and a gray matter region (gm) and a gray matter region (gm) of each of the left and right portions of each of the determination basic regions, and the left or right portion of each of the determination basic regions. Method for predicting speech.
The method of claim 1,
Generating the input data,
Extracting feature data for the determination reference region based on the slice image in which the mask is generated using the deep learning or machine learning; And
Extracting label data for assigning the scores based on the extracted feature data using the deep learning or machine learning method.
The method of claim 4, wherein
The deep learning or machine learning used in the step of extracting the feature data includes a fully convolutional network (FCN) algorithm, the method for predicting the amyloid positive or negative.
The method of claim 1,
The deep learning or machine learning used in the step of extracting and learning the score includes a regression algorithm and a support vector machine (SVM) algorithm for predicting amyloid positive or negative. Way.
The method of claim 1,
In the generating of the algorithm model and determining the amyloid positive or negative through the generated algorithm model,
And generating an algorithmic model based on the extracted and learned scores using a support vector machine algorithm.
Apparatus for predicting amyloid positive or negative based on deep learning,
A preprocessor configured to generate a slice image and a mask for the slice image from a first brain image and a second brain image based on the determination reference region for predicting the amyloid positive or negative;
An input data generation unit including a first data generation module using the deep learning and a second data generation module using the machine learning to generate input data based on the slice image and the mask generated in the preprocessing step;
A scoring unit including a first scoring module using the deep learning and a second scoring module using the machine learning to extract and learn a score for predicting the amyloid positive or negative from the generated input data; And
A first positive voice using the deep learning to generate an algorithmic model for diagnosing the amyloid positive or negative based on the extracted and learned scores, and to determine the amyloid positive or negative through the generated algorithm model Apparatus for predicting amyloid positive or negative comprising an amyloid positive negative predictor comprising a determination module and a second positive determination module using the machine learning.
The method of claim 8,
The preprocessing unit,
An image divider configured to generate a split image of the first brain image;
An image registration unit generating a registration image of the first brain imaging image and the second brain imaging image;
A slice image generation unit selecting a first slice image based on an output state of the determination base region with respect to the divided image, and extracting a second slice image corresponding to the first slice image from the matched image; And
An image mask generator configured to generate a mask for the determination reference region of the first slice image,
The apparatus for predicting amyloid positive or negative, wherein the determination reference region includes the determination basis region.
The method of claim 9,
The first brain image is an MRI T1 volume image, the second brain image is an amyloid PET volume image,
The judgment basis region includes a temporal region, a frontal region, a posterior cingulate-precuneus region, and a parietal region,
The determination reference region may include a white matter region (wm) and a gray matter region (gm) and a gray matter region (gm) of each of the left and right portions of each of the determination basic regions, and the left or right portion of each of the determination basic regions. Device for predicting speech.
The method of claim 8,
The first data generation module,
A first feature data extractor configured to extract feature data of the determination reference region based on the slice image in which the mask is generated using the deep learning; And
A first label data extracting unit extracting label data for allocating the scores based on the extracted feature data using the deep learning;
The second data generation module,
A second feature data extractor configured to extract feature data of the determination reference region based on the slice image in which the mask is generated using the machine learning; And
And a second label data extractor for extracting label data for allocating the scores based on the extracted feature data using the machine learning.
The method of claim 11,
The apparatus for predicting amyloid positive or negative is a deep convolutional network (FCN) algorithm is included in the deep learning used in the first feature data extractor or the machine learning used in the second feature data extractor.
The method of claim 8,
The deep learning or machine learning used in the scoring unit includes a regression algorithm and a support vector machine (SVM) algorithm for predicting amyloid positive or negative.
The method of claim 8,
In the amyloid positive negative prediction unit,
And generate an algorithmic model based on the extracted and learned scores using a support vector machine algorithm.
A computer-readable recording medium having recorded thereon a program for implementing the method of any one of claims 1 to 7.

Images