JP7638626B2 - Information processing device and program - Google Patents

Description

The embodiments disclosed in this specification and the drawings relate to an information processing device, a learning method, and a program.

In recent years, research into technology using AI (Artificial Intelligence) has been progressing in the field of medical image diagnosis. For example, research is being conducted into automating the entire process from image generation to image diagnosis using AI functions.

In processing using AI functions such as those described above, it is assumed that the network used for image generation is optimized according to the purpose of image diagnosis. For example, it is believed that an optimal network exists depending on the purpose of diagnosis, such as image generation network A being optimal when wanting to see a small tumor, and image generation network B being optimal when metal artifacts are a concern. In this case, it is believed that the workflow will be one in which images are generated using multiple image generation networks optimized according to the purpose. However, when generating images using multiple image generation networks, processing is required for each image generation network, which takes time.

Special Publication No. 2019-520655

The problem that the embodiments disclosed in this specification and the drawings attempt to solve is to shorten the time required to generate information according to a purpose. However, the problem that the embodiments disclosed in this specification and the drawings attempt to solve is not limited to the problem described above. Problems corresponding to the effects of each configuration shown in the embodiments described below can also be positioned as other problems.

The information processing device of the embodiment includes an acquisition unit and a processing unit. The acquisition unit acquires input data. The processing unit generates multiple output data by inputting the input data acquired by the acquisition unit to a model including a common network based on learning using multiple teacher data, and multiple individual networks located after the common network and based on individual learning using each of the multiple teacher data.

FIG. 1 is a block diagram showing an example of the configuration of a medical image processing apparatus 1 according to a first embodiment. FIG. 1 is a diagram for explaining the configuration of a trained model in the prior art. FIG. 1 is a diagram illustrating an example of the configuration of a trained model M according to the first embodiment. 4 is a flowchart showing an example of processing during operation of the medical image processing apparatus 1 according to the first embodiment. 4 is a flowchart showing an example of processing during learning of the medical image processing apparatus 1 according to the first embodiment. FIG. 4 is a diagram for explaining a first learning of a learning function 43 according to the first embodiment. FIG. 4 is a diagram for explaining second learning of the learning function 43 according to the first embodiment. FIG. 13 is a diagram illustrating a trained model M1 according to the second embodiment.

The following describes the information processing device, learning method, and program of the embodiment with reference to the drawings.

First Embodiment
[Example of configuration of medical information processing device]
1 is a block diagram showing an example of a configuration of a medical information processing device 1 according to a first embodiment. The medical information processing device 1 is connected to external devices, such as one or more medical imaging devices 3 and one or more medical databases 5, via a communication network NW so as to be able to transmit and receive data. The medical imaging device 3 is, for example, an X-ray diagnostic device, an MRI (Magnetic Resonance Imaging) device, or the like, that captures medical images of a subject. The medical database 5 is a storage device that stores medical information such as electronic medical records. The medical information processing device 1 is an example of an "information processing device".

The communication network NW refers to information and communication networks in general that use electrical communication technology. The communication network NW includes wireless/wired LANs such as hospital backbone LANs (Local Area Networks) and Internet networks, as well as telephone communication line networks, optical fiber communication networks, cable communication networks, and satellite communication networks.

The medical information processing device 1 includes, for example, a communication interface 10, an input interface 20, a display 30, a processing circuit 40, and a memory 50. The communication interface 10 includes, for example, a communication interface such as a NIC (Network Interface Card). The communication interface 10 communicates with external devices such as a medical image capturing device 3 and a medical database 5 via a communication network NW. The communication interface 10 receives medical data such as images captured by the medical image capturing device 3 and raw data of the images from the medical image capturing device 3. The communication interface 10 receives medical data such as electronic medical records from the medical database 5. The communication interface 10 outputs the received medical data to the processing circuit 40. The communication interface 10 may communicate with other external devices connected via the communication network NW under the control of the processing circuit 40.

The input interface 20 accepts various input operations from an operator, converts the accepted input operations into electrical signals, and outputs the electrical signals to the processing circuit 40. For example, the input interface 20 includes a mouse, a keyboard, a trackball, a switch, a button, a joystick, a touch panel, etc. The input interface 20 may be, for example, a user interface that accepts audio input from a microphone or the like. When the input interface 20 is a touch panel, the input interface 20 may also have the display function of the display 30 described below.

In this specification, the input interface is not limited to an interface equipped with physical operating parts such as a mouse and a keyboard. For example, an example of an input interface also includes an electrical signal processing circuit that receives an electrical signal corresponding to an input operation from an external input device provided separately from the device and outputs this electrical signal to a control circuit.

The display 30 displays various types of information. For example, the display 30 displays images generated by the processing circuit 40, a GUI (Graphical User Interface) for receiving various input operations from an operator, and the like. For example, the display 30 is an LCD (Liquid Crystal Display), a CRT (Cathode Ray Tube) display, an organic EL (Electro Luminescence) display, etc.

The processing circuit 40 includes, for example, an acquisition function 41, a learning function 43, a processing function 45, and an output control function 47. The processing circuit 40 realizes these functions by, for example, a hardware processor (computer) executing a program stored in a memory 50 (storage circuit). The acquisition function 41 is an example of an "acquisition unit." The learning function 43 is an example of a "learning unit." The processing function 45 is an example of a "processing unit." Details of each function realized in the processing circuit 40 will be described later.

The hardware processor refers to a circuit such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), an Application Specific Integrated Circuit (ASIC), a programmable logic device (e.g., a Simple Programmable Logic Device (SPLD) or a Complex Programmable Logic Device (CPLD), or a Field Programmable Gate Array (FPGA)). Instead of storing the program in the memory 50, the program may be directly incorporated into the circuit of the hardware processor. In this case, the hardware processor realizes the function by reading and executing the program incorporated in the circuit. The above program may be stored in the memory 50 in advance, or may be stored in a non-transient storage medium such as a DVD or CD-ROM, and may be installed in the memory 50 from the non-transient storage medium by mounting the non-transient storage medium in a drive device (not shown) of the medical information processing device 1. The hardware processor is not limited to being configured as a single circuit, but may be configured as a single hardware processor by combining multiple independent circuits to realize each function. Also, multiple components may be integrated into a single hardware processor to realize each function.

The memory 50 is realized, for example, by a semiconductor memory element such as a RAM (Random Access Memory), a flash memory, a hard disk, or an optical disk. These non-transient storage media may be realized by other storage devices connected via a communication network NW, such as a NAS (Network Attached Storage) or an external storage server device. The memory 50 may also include non-transient storage media such as a ROM (Read Only Memory) or a register. The memory 50 stores, for example, a trained model M generated by the learning function 43 of the processing circuit 40. In addition, the memory 50 stores programs, parameter data, and other data used by the processing circuit 40.

[Example of processing circuit configuration]
Next, a configuration example of the processing circuit 40 will be described. The acquisition function 41 acquires medical data from an external device such as the medical imaging device 3 via the communication interface 10. The medical data includes, for example, information such as an image captured by the medical imaging device 3, raw data of the image, and character string information of an electronic medical record. The acquisition function 41 may store the acquired medical data in the memory 50. The medical data is an example of "input data".

The learning function 43 performs deep learning based on the training data to generate a trained model M. For example, when medical data is input, the trained model M outputs medical information according to a specified purpose. The training data includes multiple pairs of input data and output data. For example, the input data used is information such as an image captured by the medical imaging device 3, raw data of the image, and character string information in an electronic medical record. For example, the output data used is information according to a specified purpose.

The specified purpose means, for example, the purpose of diagnosis. Examples of diagnostic purposes include diagnosing a relatively small tumor using an image with few artifacts, diagnosing using an image with little noise, diagnosing using an image with high contrast, diagnosing using an image with high resolution, diagnosing a specific disease such as cancer or cirrhosis, diagnosing a specific body part such as the head or chest, etc. Artifacts include metal artifacts, ring artifacts, and other artifacts caused by the X-ray device.

The information according to the specified purpose output by the trained model M is, for example, an image captured by the medical imaging device 3 that has been modified to make certain information easier to see, information indicating a disease or lesion location estimated based on character string information in an electronic medical record, etc. For example, if the purpose of diagnosis is "to make a diagnosis using an image with few artifacts," the information according to the specified purpose obtained as a result of inputting an image into the trained model M is an image with reduced artifacts. Also, for example, if the purpose of diagnosis is "to diagnose a specific disease such as cancer or cirrhosis," the information according to the specified purpose obtained as a result of inputting an electronic medical record or an image into the trained model M is information indicating the name of the disease, such as "cancer" or "cirrhosis."

Here, when there are multiple diagnostic purposes, it is necessary to prepare a trained model for each diagnostic purpose. FIG. 2 is a diagram explaining the configuration of a trained model in the conventional technology. As shown in FIG. 2, for example, when obtaining two patterns of images from medical data D, a first image D1 corresponding to a first diagnostic purpose (e.g., diagnosis using an image with reduced noise) and a second image D2 corresponding to a second diagnostic purpose (e.g., diagnosis using an image with reduced artifacts), it is necessary to prepare a trained model m1 corresponding to the first diagnostic purpose and a trained model m2 corresponding to the second diagnostic purpose, respectively. Here, for example, assuming that each of the trained model m1 and the trained model m2 has an eight-layer network structure, in order to obtain the first image D1 and the second image D2, it is necessary to input the medical data D into the trained model m1 and perform arithmetic processing for eight layers, and further input the medical data D into the trained model m2 and perform arithmetic processing for eight layers. That is, arithmetic processing for a total of 16 layers is required. In this way, when multiple trained models are used, calculations are performed for each trained model, which takes time.

On the other hand, the learning function 43 of this embodiment separates the trained model into an earlier stage and a later stage, and generates a trained model M in which the earlier stage is a network corresponding to processing common to multiple diagnostic purposes (hereinafter referred to as a "common network") and the later stage is a network corresponding to processing according to each diagnostic purpose (hereinafter referred to as an "individual network"). The common network corresponds, for example, to processing for creating a rough image, and the individual networks correspond, for example, to processing for creating images subdivided according to diagnostic purpose. The common network is located on the input layer side of the trained model, and the multiple individual networks are located on the output layer side of the trained model.

Figure 3 is a diagram illustrating an example of the configuration of the trained model M according to the first embodiment. As shown in Figure 3, the trained model M includes a common network CN and a plurality of individual networks UN. The plurality of individual networks UN include, for example, a first individual network UN1 including one or more layers corresponding to processing according to a first diagnostic purpose, and a second individual network UN2 including one or more layers corresponding to processing according to a second diagnostic purpose. The common network CN has, for example, a four-layer network structure, and each of the plurality of individual networks UN has, for example, a four-layer network structure.

When the trained model M of this embodiment is used as described above, medical data D is input to the trained model M, and arithmetic processing for four layers is performed using the common network CN, and further arithmetic processing for four layers is performed using the first individual network UN1 to obtain a first image D1 (first output data), and arithmetic processing for four layers is performed using the second individual network UN2 to obtain a second image D2 (second output data). In this case, arithmetic processing for a total of 12 layers is performed. Therefore, compared to the case where each of the multiple trained models m1 and m2 in the conventional technology shown in FIG. 2 is used, when the trained model M of this embodiment is used, it is possible to reduce the time required for calculation. The number of layers of the common network CN and the individual network UN is arbitrary.

The learning function 43 performs learning by updating the parameter data of the network included in the model each time teacher data is input so that the result of processing the input data with the model approaches the output data. The parameter data of the network is adjusted and updated, for example, by a method such as backpropagation of a convolution neural network (CNN). The learning function 43 may determine that learning is complete when the rate of change in the parameter data converges within a threshold value. The learning function 43 generates a learned model M, for example, in the form of a deep neural network (DNN), based on the parameter data after learning.

The trained model M is stored in memory 50 and is read from memory 50 during operation. The trained model M may be constructed using an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

The processing function 45 generates information according to a predetermined purpose by inputting the medical data of the subject acquired by the acquisition function 41 into the trained model M. That is, the processing function 45 generates multiple pieces of medical information by inputting the medical data acquired by the acquisition function 41 into the trained model M, which includes a common network CN based on learning using multiple pieces of teacher data, and multiple individual networks UN located downstream of the common network CN and based on individual learning using each of the multiple pieces of teacher data.

The processing function 45 performs control to switch the network to be used in the trained model M in response to, for example, an instruction of the diagnostic purpose input by the operator via the input interface 20. For example, when an instruction to acquire an image according to a first diagnostic purpose is received, the processing function 45 performs control to use the common network CN and the first individual network UN1 in the trained model M. On the other hand, when an instruction to acquire an image according to a second diagnostic purpose is received, the processing function 45 performs control to use the common network CN and the second individual network UN2 in the trained model M.

The output control function 47 causes the information corresponding to the specified purpose generated by the processing function 45 to be displayed on the display 30. The operator of the medical information processing device 1 can check the information corresponding to the specified purpose displayed on the display 30 and perform the desired diagnosis.

[Operational Processing]
The processing during operation of the medical information processing device 1 configured as above will be described below. Fig. 4 is a flowchart showing an example of processing during operation of the medical information processing device 1 according to the first embodiment. In the following, a case will be described as an example in which the medical information processing device 1 acquires and processes an image captured by the medical imaging device 3 as medical data, and outputs an image according to the purpose of diagnosis. In addition, in the following, a case will be described as an example in which the purpose of the first diagnosis is to perform a diagnosis using an image with reduced noise.

The acquisition function 41 in the processing circuit 40 acquires medical data D, which is an image of the subject, from the medical imaging device 3 via the communication interface 10 (step S100).

Next, the processing function 45 sets the diagnostic objective based on the operator's instruction input via the input interface 20 (step S102). For example, when the operator operates the input interface 20 to input the first diagnostic objective (diagnosis using a noise-reduced image), the processing function 45 sets the first diagnosis as the diagnostic objective.

Next, the processing function 45 generates an image according to the diagnostic purpose by inputting the medical data D acquired by the acquisition function 41 into the trained model M (step S104). When the first diagnosis is set as the diagnostic purpose, the processing function 45 controls the trained model M to use the common network CN and the first individual network UN1.

Next, the output control function 47 outputs the image corresponding to the diagnostic purpose generated by the processing function 45 (step S106). For example, the output control function 47 causes the image corresponding to the diagnostic purpose generated by the processing function 45 to be displayed on the display 30. The operator can check the image corresponding to the diagnostic purpose displayed on the display 30 (an image with reduced noise) and perform the desired diagnosis. This completes the processing of this flowchart.

[Learning process]
The process during learning of the medical information processing device 1 configured as above will be described below. Fig. 5 is a flowchart showing an example of the process during learning of the medical information processing device 1 according to the first embodiment. In the following, an example will be described in which an image captured by the medical image capturing device 3 is used as input data, and an image modified according to the purpose of diagnosis is used as output data to learn teacher data. In addition, in the following, an example will be described in which the purpose of the first diagnosis is to perform a diagnosis using an image with reduced noise, and the purpose of the second diagnosis is to perform a diagnosis using an image with reduced artifacts.

As the teacher data, both the first teacher data related to the first diagnostic purpose and the second teacher data related to the second diagnostic purpose are used. The first teacher data includes, for example, a set (k sets) of medical data Dk, which is an image including noise and artifacts as input data, and an image Atk (k is an integer equal to or greater than 1) obtained by removing noise from the medical data Dk as output data. The second teacher data includes, for example, a set (k sets) of medical data Dk, which is an image including noise and artifacts as input data, and an image Btk (k is an integer equal to or greater than 1) obtained by removing artifacts from the medical data Dk as output data. The teacher data may be provided from an external device, or may be provided via the input interface 20 based on the operation of the operator.

First, the learning function 43 in the processing circuit 40 extracts one pair (k=1, i.e., medical data D1 and image At1) from among the k pairs included in the first teacher data, and performs first learning using teacher data with the medical data D1 as input and the image At1 as output (step S200). FIG. 6 is a diagram illustrating the first learning of the learning function 43 according to the first embodiment. For example, the learning function 43 updates the parameter data of the network included in the model so that the result of processing the medical data D1 with the model approaches the image At1. Here, the learning function 43 updates the parameter data for the layer P1 including the common network CN and the first individual network UN1. The learning function 43 may extract two or more pairs (for example, two pairs of k=1 and k=2) from among the k pairs included in the first teacher data, and perform the first learning for two or more pairs at once. That is, in the first learning, learning may be performed using at least one data included in the first teacher data.

Next, the learning function 43 extracts one pair (k=1, i.e., medical data D1 and image Bt1) from among the k pairs included in the second teacher data, and performs second learning using teacher data with the medical data D1 as input and the image Bt1 as output (step S202). FIG. 7 is a diagram illustrating the second learning of the learning function 43 according to the first embodiment. For example, the learning function 43 updates the parameter data of the network included in the model so that the result of processing the medical data D1 with the model approaches the image Bt1. Here, the learning function 43 updates the parameter data for the layer P2 including the common network CN and the second individual network UN2. The learning function 43 may extract two or more pairs (for example, two pairs of k=1 and k=2) from among the k pairs included in the second teacher data, and perform second learning for two or more pairs at once. In the second learning, learning may be performed using at least one data included in the second teacher data.

Next, the learning function 43 determines whether k has reached N (step S204). If the learning function 43 determines that k has not reached N, it increments the value of k by one (step S206) and repeats the above steps S200, S202, and S204.

On the other hand, if the learning function 43 determines that k has reached N, it ends the learning process. This completes the processing of this flowchart.

In the above learning method, the learning function 43 alternates between a first learning corresponding to the purpose of the first diagnosis and a second learning corresponding to the purpose of the second diagnosis. That is, the learned model M is generated by alternately performing a first learning using at least one data included in the first teacher data and a second learning using at least one data included in the second teacher data. The learned model M generated in this manner includes a common network CN based on the first learning and the second learning, a first individual network UN1 based on the first learning, and a second individual network UN2 based on the second learning.

In the prior art, instead of alternating between the first learning and the second learning as described above, the first learning using the first teacher data is completed, and then the second learning using the second teacher data is performed. However, in such a prior art learning method, the common network CN is heavily influenced by the second learning performed later, i.e., the effect of reducing artifacts, which is the second objective, becomes greater, making it unsuitable for the first objective of noise reduction processing. On the other hand, the learning function 43 of this embodiment alternates between the first learning and the second learning, making the common network CN suitable for both objectives.

In the above example, the same medical data Dk is used as input data for the first learning and the second learning, but different input data may be used. In the above example, the number of data included in the first training data and the number of data included in the second training data are the same, but the numbers of data may be different. In addition, images are used as training data, but data in other data formats, such as raw data, may be used.

As described above, according to the first embodiment, the processing circuit 40 is provided with a model including an acquisition function 41 that acquires medical data, a common network based on learning using multiple teacher data, and multiple individual networks that are located after the common network and based on individual learning using each of the multiple teacher data, and a processing function 45 that generates multiple pieces of medical information by inputting the medical data acquired by the acquisition function 41, thereby making it possible to reduce the time required to generate information according to the purpose.

Second Embodiment
The second embodiment will be described below. The medical information processing device 1 of the second embodiment uses a trained model M1 having a different configuration from the trained model M of the first embodiment. The trained model M1 in the first embodiment includes a common network CN and an individual network UN. In contrast, the trained model M1 in the second embodiment has a further hierarchical network structure. Below, the differences from the first embodiment will be mainly described, and the points in common with the first embodiment will be omitted. In the description of the second embodiment, the same parts as those in the first embodiment will be described with the same reference numerals.

Figure 8 is a diagram illustrating a trained model M1 according to the second embodiment. As shown in Figure 8, the trained model M1 has one or more sub-networks SN in addition to a common network CN and an individual network UN. The sub-networks SN are arranged, for example, between the common network CN and the individual network UN.

In the example shown in FIG. 8, three first individual networks UN1-1 to UN1-3 and two second individual networks UN2-1 and UN2-2 are provided depending on the purpose of diagnosis. A first sub-network SN1 is arranged between the common network CN and the first individual networks UN1-1 to UN1-3. A second sub-network SN2 is arranged between the common network CN and the second individual networks UN2-1 to UN2-2.

For example, if the similarity between the eleventh image D1-1 to the first image D1-3, which correspond to the output data of the first individual networks UN1-1 to UN1-3, and the second images D2-1 and D2-2, which correspond to the output data of the second individual networks UN2-1 and UN2-2, is low, a first sub-network SN1 is placed between the common network CN and the first individual network UN1 group, and a second sub-network SN2 is placed between the common network CN and the second individual network UN2 group. In other words, based on the similarity between images generated according to the purpose of diagnosis, individual networks that output images with high similarity are grouped together into a predetermined individual network group, and a sub-network is provided that connects the individual network group to the common network.

In the learning process, the learning function 43 generates a trained model M1 having the above-mentioned network structure. Then, the processing function 45 inputs medical data, such as images captured by the medical imaging device 3, to the trained model M1 having the above-mentioned network structure, thereby obtaining highly accurate information according to the purpose.

As described above, according to the second embodiment, the processing circuit 40 is provided with a processing function 45 that generates multiple pieces of medical information by inputting the medical data acquired by the acquisition function 41 into a model that includes an acquisition function 41 that acquires medical data, a common network based on learning using multiple pieces of teacher data, and multiple individual networks that are located after the common network and based on individual learning using each of the multiple pieces of teacher data, thereby making it possible to reduce the time required to generate information according to the purpose. Furthermore, by further hierarchizing the configuration of the trained model, it is possible to obtain highly accurate information according to the purpose.

The above-described embodiment can be expressed as follows.
A storage device storing a program;
a hardware processor;
The hardware processor executes the program to acquire input data, and generates a plurality of output data by inputting the acquired input data to a model including a common network based on learning using a plurality of teacher data, and a plurality of individual networks located in a subsequent stage of the common network and based on individual learning using each of the plurality of teacher data.
Medical information processing equipment.

According to at least one of the embodiments described above, the processing circuit 40 is provided with a model including an acquisition function 41 for acquiring medical data, a common network based on learning using multiple teacher data, and multiple individual networks located after the common network and based on individual learning using each of the multiple teacher data, and a processing function 45 for generating multiple pieces of medical information by inputting the medical data acquired by the acquisition function 41, thereby making it possible to reduce the time required to generate information according to the purpose.

Although several embodiments have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are within the scope of the invention and its equivalents as set forth in the claims, as well as the scope and spirit of the invention.

REFERENCE SIGNS LIST 1 Medical information processing device 3 Medical image capturing device 5 Medical database 10 Communication interface 20 Input interface 30 Display 40 Processing circuit 50 Memory 41 Acquisition function 43 Learning function 45 Processing function 47 Output control function M, M1 Learned model

Claims (6)

an acquisition unit for acquiring input data which is a medical image;
a processing unit that generates a plurality of output data, which are medical images, by inputting the input data acquired by the acquisition unit into a model including a common network based on learning using a plurality of teacher data, and a plurality of individual networks located in a subsequent stage of the common network and based on individual learning using each of the plurality of teacher data;
Equipped with
the plurality of individual networks are provided in accordance with a plurality of diagnostic purposes,
the model further comprising a sub-network located between the common network and the plurality of individual networks;
the plurality of individual networks are grouped together into a predetermined individual network group based on a similarity between images generated by each of the plurality of individual networks, and the sub-network is disposed between the common network and the plurality of individual networks so as to connect the individual network group and the common network ;
Information processing device. The plurality of teacher data include first teacher data and second teacher data,
The model is generated by alternately performing a first learning process using at least one data included in the first teacher data and a second learning process using at least one data included in the second teacher data.
The information processing device according to claim 1 . the model includes the common network based on the first learning and the second learning, a first individual network based on the first learning, and a second individual network based on the second learning;
the processing unit inputs the input data acquired by the acquisition unit to the model, thereby generating first output data output from the first individual network and second output data output from the second individual network.
The information processing device according to claim 2 . The common network is located on the input layer side of the model,
The plurality of individual networks are located on the output layer side of the model.
The information processing device according to claim 1 . A learning unit that learns the plurality of teacher data and generates the model,
The information processing device according to claim 1 . On the computer,
Obtaining input data which is a medical image;
a model including a common network based on learning using a plurality of pieces of teacher data, and a plurality of individual networks located in a subsequent stage of the common network and based on individual learning using each of the plurality of pieces of teacher data, and inputting the acquired input data into the model to generate a plurality of output data, which are medical images;
A program,
the plurality of individual networks are provided in accordance with a plurality of diagnostic purposes,
the model further comprising a sub-network located between the common network and the plurality of individual networks;
the plurality of individual networks are grouped together into a predetermined individual network group based on a similarity between images generated by each of the plurality of individual networks, and the sub-network is disposed between the common network and the plurality of individual networks so as to connect the individual network group and the common network ;
program.

Images