JPWO2019172181A1 - Diagnostic support devices, programs, trained models, and learning devices - Google Patents

Abstract

The diagnostic support device (1) stores a trained model constructed by learning the weighting coefficient of the neural network using the data of the nuclear medicine images of the brains of a large number of subjects and the data of the diseases of the subjects as teacher data. The trained model storage unit (18), the input unit (10) for inputting the nuclear medicine image of the brain of the person to be diagnosed, and the anatomy for anatomically standardizing the nuclear medicine image input in the input unit (10). A standardization unit (12), a brain surface image generation unit (13) that generates a brain surface image from an anatomically standardized image, and a normalization unit (14) that normalizes the signal strength of the brain surface image. Based on the normalized brain surface image and the brain surface images of a plurality of healthy subjects, the Z score image generation unit (15) that generates the Z score image of the diagnosis target and the normalized brain surface image are learned. It is provided with an inference unit (16) that infers the disease of the person to be diagnosed by applying it to the completed model, and an output unit (17) that outputs the data of the disease inferred by the inference unit (16).

Description

The present invention relates to diagnostic support devices, programs, trained models, and learning devices.

Conventionally, nuclear medicine tests have been performed in which a radiopharmaceutical that emits a small amount of radiation is administered into the body and the state of the body is captured by an image. In the nuclear medicine test, it is possible to obtain useful information for diagnosing a disease, confirming the stage and prognosis, determining a therapeutic effect, etc. by imaging how radiopharmaceuticals administered in the body gather in organs and tissues.

Dementia diseases include Alzheimer's disease, Lewy body dementias, frontotemporal dementia, etc., and diagnosis using nuclear medicine images is also performed to classify dementia. Conventionally, a doctor interprets a nuclear medicine image and distinguishes a disease based on other clinical information.

Yoshimi Takamura "Dementia Imaging" Medical Examination Vol.66 No.J-STAGE-2 Dementia Prevention Examination Special Feature 2017

In recent years, machine learning techniques such as AI have been developed, and research is being conducted to utilize machine learning for interpreting medical images. However, a large amount of teacher data is required to generate a model that can perform machine learning and make an appropriate diagnosis. However, the number of facilities that take nuclear medicine images is smaller than that of facilities that can take X-rays, CT, etc., and it is not easy to collect a large amount of teacher data.

In view of the above background, it is an object of the present disclosure to provide a diagnostic support device capable of classifying diseases based on images.

The diagnostic support device according to the present disclosure is a storage unit that stores a trained model configured by learning the weighting coefficient of the neural network using the data related to the brain images of a large number of subjects and the data of the diseases of the subjects as teacher data. An input unit for inputting the brain image of the person to be diagnosed, a standardization unit for anatomically standardizing and normalizing the brain image input in the input unit, and anatomically standardized and normalized. It includes an inference unit that infers the disease of the diagnosis target person by applying the data related to the brain image to the trained model, and an output unit that outputs the data of the disease inferred by the inference unit. The brain image may be a nuclear medicine image, in which case a brain surface image or a tomographic image may be used as an input to the trained model, and in place of or in addition to the brain surface image. An image obtained by statistical analysis such as a Z-score image of a person to be diagnosed may be used, and further, clinical information may be used. The data related to the brain image includes the image obtained by statistical analysis such as the brain surface image, the tomographic image, and the Z-score image of the person to be diagnosed described here. Further, the brain image may be an MRI image, and in that case, statistically including a gray matter brain image extracted from the MRI image and a Z-score image of the person to be diagnosed as input to the trained model. The analyzed image and clinical information may be used.

The program according to the present disclosure is a program for assisting in differentiating a disease of a diagnosis target person based on data related to the brain image of the diagnosis target person, and the data on the brain images of a large number of subjects and the said Using the subject's disease data as teacher data, the trained model constructed by learning the weighting coefficient of the neural network is stored, and the step of inputting the brain image of the diagnosis target and the input brain image are stored. Steps to anatomically standardize and normalize, and steps to infer the disease of the diagnosed subject by applying data on the anatomically standardized and normalized brain image to the trained model, and inference. Perform the steps to output the data of the disease. The brain image may be a nuclear medicine image, in which case a brain surface image or a tomographic image may be used as an input to the trained model, and in place of or in addition to the brain surface image. An image obtained by statistical analysis such as a Z-score image of a person to be diagnosed may be used, or clinical information may be used. Further, the brain image may be an MRI image, and in that case, statistically including a gray matter brain image extracted from the MRI image and a Z-score image of the person to be diagnosed as input to the trained model. The analyzed image and clinical information may be used.

The trained model according to the present disclosure is a trained model for making a computer function so as to support the discrimination of the disease affected by the diagnosed subject based on the data on the brain image of the diagnosed subject. The weighting coefficient of the neural network is learned and constructed using the data related to the brain image obtained by anatomically standardizing and the signal strength normalization for the brain images of a large number of subjects and the data of the disease of the subject as teacher data. The data related to the brain image of the diagnosed subject input to the input layer of the neural network is calculated based on the learned weighting coefficient, and the data of the disease affected by the diagnosed subject is output. It is a trained model for making a computer work. The brain image may be a nuclear medicine image, in which case the trained model may be constructed by being trained using a brain surface image or a tomographic image, and may be substituted for the brain surface image or In addition, it may be learned by using an image obtained by statistical analysis such as a Z-score image of a person to be diagnosed, or may be learned by using clinical information. Further, the brain image may be an MRI image, and in that case, the trained model performs statistical analysis such as a gray matter brain image extracted from the MRI image and a Z-score image of the person to be diagnosed. It may be an image that has been learned and constructed using clinical information.

The learning device according to the present disclosure includes an input unit for inputting data on brain images of a large number of subjects and data on diseases of the subjects as teacher data, and anatomical standardization and normalization of brain images input by the input unit. Data on a large number of anatomically standardized and normalized brain images are sequentially input to the input layer of the neural network, and the data of the subject's disease corresponding to the input brain images. Is input to the output layer of the neural network, and a learning unit that repeatedly performs a process of updating the weighting coefficient of the neural network by the inverse error propagation method and a storage unit that stores the neural network learned by the learning unit are provided. The brain image may be a nuclear medicine image, in which case the learning device may perform learning using a brain surface image or a tomographic image, and in place of or in addition to the brain surface image, Learning may be performed using an image obtained by statistical analysis such as a Z-score image of a person to be diagnosed, and further, learning may be performed using clinical information. Further, the brain image may be an MRI image, and in that case, the learning device performs statistical analysis such as a gray matter brain image extracted from the MRI image and a Z-score image of the subject to be diagnosed. Learning may be performed using images and clinical information.

According to the present disclosure, the diagnosis is made based on the brain image of the subject to be diagnosed by using a trained model generated by using the data on the brain surface image obtained by anatomically standardizing and normalizing the brain image of the subject. The disease of the subject can be appropriately differentiated.

FIG. 1 is a diagram showing a configuration of a diagnostic support device according to the first embodiment. FIG. 2 is a diagram showing an example of the configuration of the trained model used in the first embodiment. FIG. 3 is a flowchart showing the operation of the diagnosis support device according to the first embodiment. FIG. 4 is a diagram showing a configuration of a learning device according to the first embodiment. FIG. 5 is a flowchart showing the operation of the learning device of the first embodiment. FIG. 6 is a flowchart showing the operation of the learning device according to the modified example. FIG. 7 is a diagram showing the configuration of the diagnostic support device of the second embodiment. FIG. 8 is a diagram showing an example of the configuration of the trained model used in the second embodiment. FIG. 9 is a flowchart showing the operation of the diagnosis support device according to the second embodiment. FIG. 10 is a diagram showing the configuration of the learning device of the second embodiment.

Hereinafter, the diagnostic support device, the learning device, the learned model, and the like according to the embodiment of the present invention will be described.

(First Embodiment)
FIG. 1 is a diagram showing a configuration of a diagnosis support device 1 according to the first embodiment. The diagnosis support device 1 includes an input unit 10, a preprocessing unit 11, an inference unit 16, and an output unit 17. The input unit 10 has a function of accepting an input of a nuclear medicine image of the brain of the person to be diagnosed. As the nuclear medicine image, an image taken by the nuclear medicine examination device may be input from the nuclear medicine examination device, or a nuclear medicine image taken at another facility may be received via a network. Examples of the nuclear medicine examination device include a PET device and a SPECT device. Further, the nuclear medicine image may be a PET image captured by the PET apparatus or a SPECT image captured by the SPECT apparatus.

The pretreatment unit 11 has a function of performing processing such as normalization so that the disease can be appropriately differentiated based on the nuclear medicine image which is different depending on the imaging conditions and individual differences. The pretreatment unit 11 includes an anatomical standardization unit 12, a brain surface image generation unit 13, a normalization unit 14, and a Z-score image generation unit 15.

The anatomical standardization unit 12 performs a process of matching the image of the brain of the person to be diagnosed with the standard brain image. The brain surface image generation unit 13 generates a brain surface image from a tomographic image of the brain of the diagnosis target person. Specifically, by looking at the pixel values in the cortex from the brain surface pixels determined by the stereotactic brain coordinate system, extracting the peak of the pixel value in the cortex as the pixel value of the brain surface, and projecting it to the brain surface. Generates a brain surface image.

The normalization unit 14 is a process for normalizing signal intensities that differ depending on the imaging conditions. The normalization unit 14 normalizes the nuclear medicine image by dividing it by the average pixel value of a predetermined reference site in the brain. As the reference site, for example, the whole brain, thalamus, cerebellum, pons, and sensorimotor cortex can be used.

The Z-score image generation unit 15 generates a Z-score image using the data of a plurality of healthy subjects. Specifically, a Z-score image can be generated by obtaining a Z-score for each pixel using the following equation.
Z score = (mean value of healthy subjects-pixel value of diagnosed subjects) / (standard deviation of healthy subjects)

The Z score is the degree to which the number of standard deviations is above or below the population mean, and as shown in the above formula, statistical analysis is performed on the data of the subject and the data of a plurality of healthy subjects. Can be obtained from. Although the example of using the data of a plurality of healthy subjects is given here, the data of the plurality of healthy subjects is unnecessary when the average value and the standard deviation of the data of the plurality of healthy subjects are obtained. When the nuclear medicine image is a cerebral blood flow image, processing by the preprocessing unit 11 described above may be performed using software called "3D-SSP".

The reasoning unit 16 has a function of inferring the disease of the diagnosed subject by applying the brain surface image and the Z score image of the diagnosed subject to the learned model stored in the learned model storage unit 18.

FIG. 2 is a diagram showing an example of a trained model stored in the trained model storage unit 18. The trained model is a model of deep learning, which is a kind of neural network. It has a plurality of sets of a Convolution layer and a Pooling layer, a fully connected layer in the subsequent stage, and a Softmax layer. A plurality of sets of the Convolution layer and the Pooling layer extract the features of the input image by repeating convolution and pooling when the normalized image of the person to be diagnosed is input. The fully connected layer combines the extracted features into one node. The softmax layer outputs the probability of each disease based on the output from the fully connected layer.

In this trained model, the weighting coefficient is trained in advance using the data of a large number of subjects. Specifically, the parameters of the filter used in the Convolution layer and the parameters of the node weights of the fully connected layer are learned. The generation of the trained model will be described later.

The trained model of this embodiment is expected to be used as a program module that is a part of artificial intelligence software. The trained model of this embodiment is used in a computer including a CPU and a memory. Specifically, the CPU of the computer inputs the input data (brain surface image and Z-score image) input to the input layer (first Convolution layer) of the neural network according to the command from the trained model stored in the memory. On the other hand, the feature amount of the image is obtained by convolution and pooling, the calculated feature amount is calculated based on the learned weighting coefficient, and the result (probability value of having each disease) is obtained from the output layer (Softmax layer). Works to output.

The output unit 17 outputs the disease data inferred by the inference unit 16. The output unit 17 may display, for example, a brain image of the person to be diagnosed on the monitor together with the inferred disease name.

FIG. 3 is a flowchart showing the operation of the diagnosis support device 1. When a nuclear medicine image of the brain of the diagnosis target is input to the diagnosis support device 1 (S10), the diagnosis support device 1 performs anatomical standardization of the input brain image of the diagnosis target (S11). Subsequently, the diagnostic support device 1 generates a brain surface image using the anatomically standardized image (S12), and normalizes the signal intensity of the generated brain surface image (S13). Next, the diagnostic support device 1 generates a Z-score image using the data of a healthy person (S14). After that, the diagnostic support device 1 applies the brain surface image and the Z-score image to the trained model, infers the disease (S15), and outputs the inferred result (S16).

Next, the generation of the trained model stored in the trained model storage unit 18 will be described. FIG. 4 is a diagram showing a configuration of a learning device 2 that generates a trained model. The learning device 2 has an input unit 20, a preprocessing unit 21, a learning unit 26, and a model storage unit 27.

The input unit 20 has a function of accepting input of nuclear medicine images of the brains of a large number of subjects as learning teacher data. The subjects include those with dementia diseases as well as healthy subjects. The teacher data to be input is data of a subject having a disease to be classified using a trained model, and is, for example, Alzheimer's dementia, Lewy body dementia, and frontotemporal dementia. The input unit 20 receives input of data for identifying the disease of the subject together with the data of the brain image of the subject.

The preprocessing unit 21 preprocesses the input nuclear medicine image of the subject's brain. The configuration and processing contents of the preprocessing unit 21 are the same as those of the preprocessing unit 11 of the diagnosis support device 1.

The learning unit 26 has a function of learning the model stored in the model storage unit 27 by using the teacher data. The model shown in FIG. 2 is stored in the model storage unit 27. The trained model is the one in which the model stored in the model storage unit 27 is trained. That is, the model before the training and the model after the training differ in the weighting of the parameters, but the basic configuration is the same.

The learning unit 26 applies the brain surface image and the Z-score image generated from the subject's brain image data of the teacher data to the input layer of the model to be learned, and applies the subject's disease to the Softmax layer. The Softmax layer has nodes for Alzheimer's disease, Lewy body dementias, frontotemporal dementia, and healthy subjects. When inferring the disease of the person to be diagnosed, the probability value is output to each disease or the node of the healthy person, but when learning, the node corresponding to the disease of the subject, which is the teacher data. "1" is applied to, and "0" is applied to other nodes. The learning unit 26 learns the parameter values of the model by using the inverse error propagation method, and stores the learned model in the model storage unit 27.

FIG. 5 is a flowchart showing the operation of the learning device 2. The learning device 2 inputs nuclear medicine images of the brains of a large number of subjects and data on the diseases of the subjects as teacher data (S20). Next, the learning device 2 performs anatomical standardization of the brain image of the subject used for learning (S21), and generates a brain surface image from the anatomically standardized brain image (S22). Subsequently, the signal intensity of the brain surface image is normalized (S23), and a Z-score image is generated based on the normalized brain surface image and the brain surface images of a plurality of healthy subjects (S24).

The learning device 2 applies the brain surface images and Z-score images of a plurality of subjects to the input layer of the model to be trained, and applies the disease data of the subjects to the output layer of the model to perform an inverse error propagation method. By doing so, the model is trained (S25). The learning device 2 determines whether or not there is other teacher data (S26), and if there is other teacher data (YES in S26), trains the model using the other teacher data. .. Specifically, the process returns to step S21 of the anatomical standardization of the subject's brain. When the processing of all the input teacher data is completed and there is no other teacher data (NO in S26), the learning device 2 ends the processing of learning the model.

The configurations of the diagnostic support device 1 and the learning device 2 of the present embodiment have been described above. However, the hardware examples of the diagnostic support device 1 and the learning device 2 described above include a CPU, RAM, ROM, hard disk, display, and keyboard. , Mouse, communication interface, etc. The above-mentioned diagnosis support device 1 and learning device 2 are realized by storing a program having a module that realizes each of the above-mentioned functions in a RAM or ROM and executing the program by a CPU. Such programs are also included in the scope of the present invention.

The diagnostic support device 1 of the present embodiment has a trained model for inferring a disease from a nuclear medicine image. Using this trained model, it is possible to support the discrimination of diseases by inferring the disease of the subject from the nuclear medicine image of the diagnosed subject. Further, when generating the trained model, the model is trained using the brain surface image obtained by normalizing the signal strength generated from the nuclear medicine image of the subject, so that an appropriate model can be generated with a small amount of teacher data. ..

FIG. 6 is a flowchart showing the operation of the learning device 2 according to the modified example of the first embodiment. The operation according to the modified example is basically the same as the operation described with reference to FIG. 5, but the signal strength normalization process (S23a) is different.

When normalizing the signal strength, the normalization process is performed by dividing the brain surface image of the subject by the average pixel value of each of the subject's whole brain, thalamus, cerebellum, pons, and sensorimotor cortex. As a result, five normalized images can be obtained from the brain image of one subject. Since these five normalized images can be used for learning, teacher data will increase. As a result, even when the data of the subjects used for learning is small, the teacher data can be increased to generate an appropriate model.

(Second Embodiment)
FIG. 7 is a diagram showing the configuration of the diagnostic support device 3 of the second embodiment. The basic configuration of the second diagnostic support device 3 is the same as that of the first embodiment, but the second diagnostic support device 3 includes the nuclear medicine image of the brain of the subject to be diagnosed and the MMSE result of the subject. The difference is that the input of the above data is accepted and the disease is differentiated using the MMSE result as well. The MMSE is a test for diagnosing dementia called the Mini-Mental State Examination, and diagnoses dementia by answering a plurality of question items.

FIG. 8 is a diagram showing a trained model used in the diagnostic support device 3 of the second embodiment. The trained model of the second embodiment has a configuration in which the MMSE result is input to the fully connected layer in addition to the trained model of the first embodiment.

FIG. 9 is a flowchart showing the operation of the diagnosis support device 3. When the nuclear medicine image of the diagnosis target person's brain and the MMSE result are input to the diagnosis support device 3 (S30), the diagnosis support device 3 performs anatomical standardization of the input diagnosis target person's brain image (S30). S31). Subsequently, the diagnostic support device 3 generates a brain surface image using the anatomically standardized image (S32), and normalizes the signal intensity of the generated brain surface image (S33). Next, the diagnostic support device 3 generates a Z-score image using the data of a healthy person (S34). After that, the diagnostic support device 3 applies the brain surface image, the Z score image, and the MMSE result to the trained model, infers the disease (S35), and outputs the inferred result (S36).

Next, the generation of the trained model stored in the trained model storage unit 27 will be described. FIG. 10 is a diagram showing a configuration of a learning device 4 that generates a trained model. The basic configuration of the learning device 4 of the second embodiment is the same as that of the first embodiment, and includes an input unit 10, a preprocessing unit 21, a learning unit 26, and a model storage unit 27. are doing. The model to be learned by the learning device 4 of the second embodiment is different in that it is the model shown in FIG.

The input unit 20 has a function of accepting input of nuclear medicine images of the brains of a large number of subjects and data of MMSE results as learning teacher data. The input unit 20 receives input of data for identifying the disease of the subject together with the data of the brain image of the subject and the MMSE result.

The preprocessing unit 21 preprocesses the input nuclear medicine image of the subject's brain. The configuration and processing contents of the preprocessing unit 21 are the same as the processing of the preprocessing unit 21 of the diagnosis support device 3.

The learning unit 26 has a function of learning the model stored in the model storage unit 27 by using the teacher data. The model shown in FIG. 8 is stored in the model storage unit 27. The learning unit 26 applies the brain surface image and Z score image generated from the subject's brain image data of the teacher data and the MMSE result to the input layer of the model to be learned, and applies the subject's disease to the Softmax layer. The learning unit 26 learns the parameter values of the model by using the inverse error propagation method, and stores the learned model in the model storage unit 27.

Since the diagnosis support device 3 of the second embodiment uses the MMSE result in addition to the nuclear medicine image of the person to be diagnosed to infer the disease, the accuracy of disease discrimination can be improved. In the present embodiment, an example in which the MMSE is used as clinical information of the subject and the subject to be diagnosed has been given, but clinical information other than the MMSE may be used.

(Modification example)
In the above-described embodiment, both the brain surface image and the Z-score image are used as the brain images of the subject and the diagnosis target, but either the brain surface image or the Z-score image may be used. .. Further, a tomographic image of the brain may be used instead of the brain surface image. That is, by generating a tomographic image from an anatomically standardized nuclear medicine image of the brain and using the tomographic image generated by normalizing the signal intensity of the tomographic image, as in the above embodiment, with a small amount of teacher data. Appropriate models can be generated. Further, in the above-described embodiment, the Z-score image is given as an example of the image subjected to the statistical analysis, but an image subjected to the statistical analysis processing other than the Z-score image can also be used as the teacher data. Is.

In the above-described embodiment, an example of differentiating Alzheimer's dementia, Levy body dementia, and frontotemporal dementia has been given, but the present invention has described other dementias such as brain tumor and chronic hard dementia. It can also be used to distinguish subdural blood species, normal pressure hydrocephalus, and aftereffects of head trauma.

In the above-described embodiment, the nuclear medicine image has been described as an example, but the diagnostic support device of the present invention can also be applied to an MRI image, an fMRI image, and the like. For example, when MRI images are used to analyze morphological information of medial temporal atrophy characteristic of early Alzheimer's dementia (AD), anatomy of grayish images extracted from MRI images. By using the standardized and normalized images, it is possible to increase the teacher data and generate an appropriate model even when the data of the subject used for learning is small, as in the above-described embodiment. Further, when using an MRI image, an image obtained by statistical analysis may be used. As an example, the MRI image of the anatomically standardized and normalized subject may be compared with the MRI image of a plurality of healthy subjects to generate a Z-score image, and the model may be trained using the Z-score image. ..

In the above-described embodiment, an example of using a deep learning model has been given, but a neural network model may be used, or another learning model may be used.

This application claims priority on the basis of Japanese application Japanese Patent Application No. 2018-043242 filed on March 9, 2018, and incorporates all of its disclosures herein.

Claims (10)

A storage unit that stores a trained model constructed by learning the weighting coefficient of the neural network using the data related to the brain images of a large number of subjects and the data of the disease of the subject as teacher data.
An input unit for inputting the brain image of the person to be diagnosed,
A standardization unit that anatomically standardizes and normalizes the brain image input by the input unit,
An inference unit that infers the disease of the diagnosed subject by applying data on the anatomically standardized and normalized brain image to the trained model.
An output unit that outputs the disease data inferred by the inference unit, and an output unit.
Diagnostic support device equipped with. The diagnostic support device according to claim 1, wherein the brain image is a nuclear medicine image or an MRI image. The brain image is a nuclear medicine image and is
The standardization unit
An anatomical standardization unit that anatomically standardizes the nuclear medicine image input in the input unit,
A brain surface image generator that generates a brain surface image from the anatomically standardized image,
A normalization unit that normalizes the signal strength of the brain surface image,
Have,
The diagnostic support device according to claim 1 or 2, wherein the reasoning unit applies the normalized brain surface image to the trained model to infer the disease of the person to be diagnosed. A Z-score image generation unit that generates a Z-score image of a diagnosis target based on the normalized brain surface image and the brain surface images of a plurality of healthy subjects is provided.
The inference unit according to claim 3, wherein the inference unit infers the disease of the diagnosis target person by applying the Z score image to the trained model in place of the brain surface image or in addition to the brain surface image. Diagnosis support device. The brain image is a nuclear medicine image and is
The standardization unit
An anatomical standardization unit that anatomically standardizes the nuclear medicine image input in the input unit,
A tomographic image generator that generates a tomographic image from the anatomically standardized image,
A normalization unit that normalizes the signal strength of the tomographic image,
Have,
The diagnostic support device according to claim 1 or 2, wherein the inference unit applies the normalized tomographic image to the trained model to infer the disease of the diagnosis target person. In the trained model stored in the storage unit, in addition to the data on the brain images of a large number of subjects and the data on the disease of the subject, the weighting coefficient of the neural network is learned by using the clinical information of the subject as the teacher data. It is a constructed trained model and
The input unit inputs clinical information of the diagnosis target in addition to the nuclear medicine image of the diagnosis target's brain.
The diagnostic support device according to any one of claims 3 to 5, wherein the inference unit infers the disease of the diagnosis target person by applying the input clinical information to the learned model. The brain image is an MRI image and is
The standardization unit
A tissue division processing unit that extracts gray matter from the MRI image input by the input unit and generates a gray matter brain image, and a tissue division processing unit.
An anatomical standardization unit that anatomically standardizes the gray matter brain image,
A normalization unit that normalizes the signal intensity of the gray matter brain image,
Have,
The diagnostic support device according to claim 1 or 2, wherein the reasoning unit applies the normalized gray matter brain image to the trained model to infer the disease of the person to be diagnosed. It is a program to support the discrimination of the disease of the diagnosed subject based on the data related to the brain image of the diagnosed subject, and it is a computer.
Using the data related to the brain images of a large number of subjects and the data of the disease of the subject as teacher data, a trained model constructed by learning the weighting coefficient of the neural network is stored.
Steps to input the brain image of the person to be diagnosed,
Steps to anatomically standardize and normalize the input brain image,
The steps of applying data on anatomically standardized and normalized brain images to the trained model to infer the disease of the diagnosed subject,
Steps to output inferred disease data,
A program that executes. A trained model that allows a computer to function to assist in differentiating the disease that the diagnosed person is suffering from, based on data on the brain image of the diagnosed person.
The weighting coefficient of the neural network is learned and constructed using the data related to the brain image obtained by anatomically standardizing and the signal strength normalization for the brain images of a large number of subjects and the data of the disease of the subject as teacher data. And
The computer functions to output the data of the disease that the diagnosed subject is suffering from by performing the calculation based on the learned weighting coefficient on the data related to the brain image of the diagnosed subject input to the input layer of the neural network. Trained model to let. An input unit for inputting data on brain images of a large number of subjects and data on the diseases of the subjects as teacher data,
A standardization unit that performs anatomical standardization and normalization of the brain image input by the input unit,
Data on a large number of anatomically standardized and normalized brain images are sequentially input to the input layer of the neural network, and the disease data of the subject corresponding to the input brain images is input to the output layer of the neural network. A learning unit that repeatedly inputs and updates the weighting coefficient of the neural network by the inverse error propagation method,
A storage unit that stores the neural network learned by the learning unit, and a storage unit.
A learning device equipped with.

Images