JP2020089711A - Model generation method and program - Google Patents

Abstract

To provide a model generation method, etc. that generates a model for presenting a prediction regarding a diagnosis reference used for diagnosis of a disease.SOLUTION: The model generation method acquires plural sets of teacher data obtained by associating and recording an endoscope image 49 and a result of determination regarding a diagnosis reference used to diagnosis of a disease, and generates a first model 61 for outputting diagnosis reference prediction that is a prediction regarding the diagnosis reference of the disease in a case where the endoscope image 49 is input, by using the teacher data.SELECTED DRAWING: Figure 11

Description

The present invention relates to a model generation method and program.

An image processing apparatus has been proposed which performs texture analysis of endoscopic images and the like and performs classification corresponding to pathological diagnosis. By using such a diagnosis support technique, even a doctor who does not have highly specialized knowledge and experience can quickly make a diagnosis.

JP, 2017-70609, A

However, the classification by the image processing device of Patent Document 1 is a black box for the user. Therefore, it is not always possible for the user to understand and understand the reason for the output classification.

For example, in ulcerative colitis (UC), it is known that the judgment may be divided between specialists who have seen the same endoscopic image. In the case of such a disease, there is a possibility that the doctor who is the user may not be satisfied with the determination result obtained by the diagnosis support technology.

In one aspect, it is an object to provide a method of generating a model, which generates a model that presents predictions regarding diagnostic criteria used for diagnosing a disease.

The method for generating the model is to acquire a plurality of sets of teacher data recorded in association with the endoscopic image and a determination result determined for a diagnostic criterion used for diagnosing a disease, and use the teacher data to generate an internal A first model that outputs a diagnostic criterion prediction that predicts a diagnostic criterion of a disease when an endoscopic image is input is generated.

In one aspect, it is possible to provide a method of generating a model, which generates a model that presents predictions regarding diagnostic criteria used for diagnosing a disease.

It is an explanatory view explaining the outline of a diagnosis support system. It is explanatory drawing explaining the structure of a diagnostic support system. It is explanatory drawing explaining the structure of a 1st score learning model. It is explanatory drawing explaining the structure of a 2nd model. 6 is a time chart schematically explaining the operation of the diagnosis support system. It is a flow chart explaining the flow of processing of a program. It is explanatory drawing explaining the outline|summary of the diagnostic assistance system of a 1st modification. It is explanatory drawing explaining the screen display of a 2nd modification. It is explanatory drawing explaining the screen display of a 3rd modification. It is a time chart which illustrates operation of the 4th modification typically. It is explanatory drawing explaining the outline|summary of the process which produces|generates a model. It is explanatory drawing explaining the structure of a model generation system. It is explanatory drawing explaining the record layout of teacher data DB. It is explanatory drawing explaining a teacher data input screen. It is explanatory drawing explaining a teacher data input screen. It is a flow chart explaining a flow of processing of a program which generates a learning model. It is a flow chart explaining a flow of processing of a program which updates a learning model. It is a flow chart explaining the flow of processing of the program which collects teacher data. It is explanatory drawing explaining the outline|summary of the diagnostic assistance system of Embodiment 3. It is explanatory drawing explaining the feature-value acquired from a 2nd model. It is explanatory drawing explaining the conversion of a feature-value and a score. It is explanatory drawing explaining the record layout of feature amount DB. It is a flow chart explaining the flow of processing of the program which creates a converter. 9 is a flowchart illustrating a flow of processing of a program at the time of endoscopy according to the third embodiment. It is explanatory drawing explaining the outline|summary of the diagnostic assistance system of Embodiment 4. It is explanatory drawing explaining the conversion of the endoscopic image and score of Embodiment 4. 16 is a flowchart illustrating a flow of processing of a program that creates the converter according to the fourth embodiment. 16 is a flowchart illustrating a flow of processing of a program at the time of endoscopy according to the fourth embodiment. It is explanatory drawing explaining the outline|summary of the diagnosis assistance system of Embodiment 5. It is explanatory drawing explaining the structure of the 1st score learning model of Embodiment 6. It is explanatory drawing explaining the screen display of Embodiment 6. It is explanatory drawing explaining the screen display of Embodiment 7. It is explanatory drawing explaining the outline|summary of the diagnostic assistance system of Embodiment 8. It is a functional block diagram of the server of Embodiment 9. It is explanatory drawing explaining the structure of the model generation system of Embodiment 10.

[Embodiment 1]
In the present embodiment, the diagnosis support system 10 that supports the diagnosis of ulcerative colitis will be described as an example. Ulcerative colitis is one of the inflammatory bowel diseases in which the mucous membrane of the large intestine is inflamed. It is known that the affected area extends from the rectum to the entire circumference of the large intestine and progresses toward the oral side.

Since the active phase in which the symptoms appear strongly and the remission phase in which the symptoms subside may be repeated, and the risk of developing colon cancer may increase if inflammation continues, and so on. Follow-up with endoscopy is recommended.

The doctor inserts the tip of the colonoscope up to the cecum, for example, and then observes the endoscopic image while removing it. In the affected area, that is, in the area where inflammation has occurred, inflammation seems to spread over the entire endoscopic image.

Public institutions such as WHO (World Hearth Organization), medical associations, medical institutions, and the like define diagnostic criteria used when diagnosing various diseases. For example, in the case of ulcerative colitis, a plurality of items such as the degree of redness of the affected area, the degree of vascular see-through which means the appearance of blood vessels, and the degree of ulcer are listed as diagnostic criteria.

After examining each item of the diagnostic criteria, the doctor makes a comprehensive judgment and diagnoses the site under observation with the endoscope 14. The diagnosis includes determination of whether or not the site under observation is an affected part of ulcerative colitis, and determination of severity such as severe or mild in the case of the affected part. A skilled doctor examines each item of the diagnostic criteria while removing the colonoscope, and makes a real-time diagnosis of the position under observation. The doctor comprehensively diagnoses the process of removing the colonoscope to determine the range of the affected area in which inflammation is caused by ulcerative colitis.

FIG. 1 is an explanatory diagram illustrating an overview of the diagnosis support system 10. An endoscopic image 49 captured by using the endoscope 14 (see FIG. 2) is input to the first model 61 and the second model 62. The second model 62 outputs a diagnostic prediction regarding the state of ulcerative colitis when the endoscopic image 49 is input. In the example shown in FIG. 1, a diagnostic prediction is output that the probability of not being a normal, ie, non-affected part of ulcerative colitis, is 70%, and the probability of having mild ulcerative colitis is 20%. Details of the second model 62 will be described later.

The first model 61 includes a first score learning model 611, a second score learning model 612, and a third score learning model 613. In the following description, the first score learning model 611 to the third score learning model 613 may be simply referred to as the first model 61 if it is not necessary to distinguish them.

When the endoscopic image 49 is input, the first score learning model 611 outputs a predicted value of the first score that is a numerical evaluation of the degree of redness. When the endoscopic image 49 is input, the second score learning model 612 outputs a predicted value of the second score that is a numerical evaluation of the degree of vascular see-through. When the endoscopic image 49 is input, the third score learning model 613 outputs the predicted value of the third score that is a numerical evaluation of the degree of ulcer.

The degree of redness, the degree of vascular see-through, and the degree of ulcer are examples of the diagnostic criteria items included in the diagnostic criteria used by a doctor when diagnosing the condition of ulcerative colitis. The predicted values of the first score to the third score are examples of the diagnostic standard prediction regarding the diagnostic standard of ulcerative colitis.

In the example shown in FIG. 1, predicted values that the first score is 10, the second score is 50, and the third score is 5 are output. The first model 61 is a score learning model that outputs a predicted value of a score that numerically evaluates various diagnostic criteria items for ulcerative colitis, such as the degree of easy bleeding and the degree of secretion adhesion. May be included. Details of the first model 61 will be described later.

The outputs of the first model 61 and the second model 62 are acquired by the first acquisition unit and the second acquisition unit, respectively. The screen shown in the lower part of FIG. 1 is displayed on the display device 16 (see FIG. 2) based on the outputs acquired by the first acquisition unit and the second acquisition unit. The displayed screen includes an endoscopic image column 73, a first result column 71, a first stop button 711, a second result column 72, and a second stop button 722.

In the endoscopic image column 73, an endoscopic image 49 taken by using the endoscope 14 is displayed in real time. In the first result column 71, a list of diagnostic reference predictions output from the first model 61 is displayed. In the second result column 72, the diagnostic prediction output from the second model 62 is displayed.

The first stop button 711 is an example of a first reception unit that receives an operation stop instruction of the first model 61. That is, when the first stop button 711 is selected, the output of the predicted value of the score using the first model 61 is stopped. The second stop button 722 is an example of a second reception unit that receives an operation stop instruction for the second model 62. That is, when the second stop button 722 is selected, the output of the predicted value of the score using the second model 62 is stopped.

By referring to the diagnostic criterion prediction displayed in the first result column 71, the doctor confirms the basis of whether the diagnostic prediction displayed in the second result column 72 is appropriate in light of the diagnostic criterion, It is determined whether or not to adopt the diagnostic prediction displayed in the first result column 71.

FIG. 2 is an explanatory diagram illustrating the configuration of the diagnosis support system 10. The diagnosis support system 10 includes an endoscope 14, an endoscope processor 11, and an information processing device 20. The information processing device 20 includes a control unit 21, a main storage device 22, an auxiliary storage device 23, a communication unit 24, a display device I/F (Interface) 26, an input device I/F 27, and a bus.

The endoscope 14 includes an elongated insertion portion 142 having an image sensor 141 at the tip. The endoscope 14 is connected to the endoscope processor 11 via an endoscope connector 15. The endoscope processor 11 receives the video signal from the image sensor 141 and performs various image processings to generate an endoscopic image 49 suitable for observation by a doctor. That is, the endoscope processor 11 functions as an image generation unit that generates an endoscopic image 49 based on the video signal acquired from the endoscope 14.

The control unit 21 is an arithmetic and control unit that executes the program of this embodiment. For the control unit 21, one or more CPUs (Central Processing Units), GPUs (Graphics Processing Units), multi-core CPUs, etc. are used. The control unit 21 is connected to each of the hardware units configuring the information processing device 20 via a bus.

The main storage device 22 is a storage device such as SRAM (Static Random Access Memory), DRAM (Dynamic Random Access Memory), and flash memory. The main storage device 22 temporarily stores information required during the process performed by the control unit 21 and a program being executed by the control unit 21.

The auxiliary storage device 23 is a storage device such as an SRAM, a flash memory, or a hard disk. The auxiliary storage device 23 stores the first model 61, the second model 62, a program to be executed by the control unit 21, and various data necessary for executing the program. As described above, the first model 61 includes the first score learning model 611, the second score learning model 612, and the third score learning model 613. The first model 61 and the second model 62 may be stored in an external mass storage device connected to the information processing device 20.

The communication unit 24 is an interface that performs data communication between the information processing device 20 and the network. The display device I/F 26 is an interface that connects the information processing device 20 and the display device 16. The display device 16 is an example of an output unit that outputs the diagnostic reference prediction acquired from the first model 61 and the diagnostic prediction acquired from the second model 62.

The input device I/F 27 is an interface that connects the information processing device 20 to an input device such as the keyboard 17. The information processing device 20 is an information device such as a general-purpose personal computer, a tablet, or a smartphone.

FIG. 3 is an explanatory diagram illustrating the configuration of the first score learning model 611. The first score learning model 611 outputs the predicted value of the first score when the endoscopic image 49 is input.

The first score is a numerical value of the degree of redness determined based on the diagnostic criteria for ulcerative colitis when a skilled doctor views the endoscopic image 49. For example, a doctor determines a score with a maximum score of 100, such as "red deemed" is 0 point and "strong redness" is 100 points.

In addition, the doctor makes a judgment in four stages, such as "red deemed", "mild", "medium", and "severe", and "red deemed" is 0 point, "mild" is 1 point, and "medium" is You may score 2 points and “severity” as 3 points. The score may be set so that “severity” has a smaller numerical value.

The first score learning model 611 according to the present embodiment is a learning model generated by machine learning using CNN (Convolutional Neural Network), for example. The first score learning model 611 includes an input layer 531, an intermediate layer 532, an output layer 533, and a neural network model 53 having a convolutional layer (not shown) and a pooling layer. The method of generating the first score learning model 611 will be described later.

The endoscopic image 49 is input to the first score learning model 611. The input image is repeatedly processed by the convolutional layer and the pooling layer, and then input to the fully connected layer. The predicted value of the first score is output to the output layer 533.

Similarly, the second score is a numerical value for judging the degree of vascular see-through based on the diagnostic criteria for ulcerative colitis when a skilled specialist views the endoscopic image 49. The third score is a numerical value for judging the degree of ulcer based on the diagnostic criteria for ulcerative colitis when a skilled specialist views the endoscopic image 49. The configurations of the second score learning model 612 and the third score learning model 613 are the same as those of the first score learning model 611, and therefore illustration and description thereof are omitted.

FIG. 4 is an explanatory diagram illustrating the configuration of the second model 62. The second model 62 outputs a diagnostic prediction of ulcerative colitis when the endoscopic image 49 is input. The diagnostic prediction is a prediction about how an expert specialist diagnoses ulcerative colitis when he/she sees the endoscopic image 49.

The second model 62 of the present embodiment is a learning model generated by machine learning using CNN, for example. The second model 62 is composed of an input layer 531, an intermediate layer 532, an output layer 533, and a neural network model 53 having a convolutional layer and a pooling layer (not shown). The method of generating the second model 62 will be described later.

The endoscopic image 49 is input to the second model 62. The input image is repeatedly processed by the convolutional layer and the pooling layer, and then input to the fully connected layer. The diagnostic prediction is output to the output layer 533.

In FIG. 4, in the output layer 533, the judgment when the endoscopic image 49 is seen by a trained specialist is the probability of severe ulcerative colitis, the probability of moderate ulcerative colitis, and the mild ulcer. It has four output nodes that output the probability that it is inflammatory bowel disease and the probability that it is normal, that is, not the affected part of ulcerative colitis, respectively.

FIG. 5 is a time chart schematically explaining the operation of the diagnosis support system 10. FIG. 5A shows the timing of shooting by the image sensor 141. FIG. 5B shows the timing of generating the endoscopic image 49 by the image processing in the endoscope processor 11. FIG. 5C shows the timing at which the first model 61 and the second model 62 output prediction based on the endoscopic image 49. FIG. 5D shows the timing of display on the display device 16. The horizontal axis in each of FIGS. 5A to 5D represents time.

At time t ₀ , the image sensor 141 captures the frame “a”. The video signal is transmitted to the endoscope processor 11. The endoscopic processor 11 performs image processing to generate an endoscopic image 49 of “a” at time t ₁ . The control unit 21 acquires the endoscopic image 49 generated by the endoscopic processor 11 and inputs the endoscopic image 49 to the first model 61 and the second model 62. At time t ₂ , the control unit 21 acquires the predictions output from the first model 61 and the second model 62, respectively.

At time t ₃ , the control unit 21 outputs the endoscopic image 49 and the prediction of the frame “a” to the display device 16. This is the end of the processing of the image for one frame captured by the image sensor 141. Similarly, at time t ₆ , the image sensor 141 captures the frame “b”. At time t ₇ , the endoscopic image 49 of “b” is generated. The control unit 21 acquires the prediction at time t ₈ and outputs the endoscopic image 49 and the prediction of the frame “b” to the display device 16 at time t ₉ . Since the operations after the frame of “c” are similar, the description thereof will be omitted. As described above, the endoscopic image 49 and the predictions by the first model 61 and the second model 62 are displayed in synchronization with each other.

FIG. 6 is a flowchart explaining the flow of processing of the program. The program described with reference to FIG. 6 is executed every time the control unit 21 acquires the endoscopic image 49 of one frame from the endoscopic processor 11.

The control unit 21 acquires the endoscopic image 49 from the endoscopic processor 11 (step S501). The control unit 21 inputs the acquired endoscopic image 49 into the second model 62 and acquires the diagnostic prediction output from the output layer 533 (step S502). The control unit 21 inputs the acquired endoscopic image 49 into one of the score learning models that configure the first model 61, and acquires the predicted value of the score output from the output layer 533 (step S503). ..

The control unit 21 determines whether or not the processing of the score learning model forming the first model 61 has been completed (step S504). When it is determined that the processing is not completed (NO in step S504), the control unit 21 returns to step S503.

When it is determined that the processing is completed (YES in step S504), the control unit 21 generates the image described using the lower portion of FIG. 1 and outputs the image to the display device 16 (step S505). The control unit 21 ends the process.

According to the present embodiment, it is possible to provide the diagnosis support system 10 that displays the diagnostic reference prediction output from the first model 61 and the diagnostic prediction output from the second model 62 together with the endoscopic image 49. While observing the endoscopic image 49, the doctor can confirm the diagnostic prediction and the diagnostic reference prediction that predict the diagnosis when a skilled specialist views the same endoscopic image 49.

By referring to the diagnostic criterion prediction displayed in the first result column 71, the doctor confirms the basis of whether the diagnostic prediction displayed in the second result column 72 is appropriate in light of the diagnostic criterion, It is possible to determine whether to adopt the diagnostic prediction displayed in the first result column 71.

In the second result column 72, only the item having the highest probability and the probability thereof may be displayed. By reducing the number of characters displayed, the character size can be increased. The doctor can detect a change in the display of the second result column 72 while gazing at the endoscope image column 73.

The doctor can stop the prediction and display of the score by selecting the first stop button 711. The doctor can stop the diagnosis prediction and the display of the diagnosis prediction by selecting the second stop button 722. The doctor can restart the display of the diagnostic prediction and the diagnostic reference prediction by selecting the first stop button 711 or the second stop button 722 again.

The first stop button 711 and the second stop button 722 can be operated from any input device such as the keyboard 17, the mouse, the touch panel, or voice input. The first stop button 711 and the second stop button 722 may be operable using a control button or the like provided on the operation unit of the endoscope 14.

For example, when performing endoscopic treatment such as excision of a polyp or EMR (Endoscopic Mucosal Resection), the time lag from the imaging by the imaging device 141 to the display on the display device 16 is It is desirable to be as short as possible. The doctor can shorten the time lag by stopping the diagnosis prediction and the diagnosis reference prediction by selecting the first stop button 711 and the second stop button 722.

The diagnostic reference prediction using each score learning model forming the first model 61 and the diagnostic prediction using the second model 62 may be performed by parallel processing. By using parallel processing, it is possible to improve the real-time property of the display on the display device 16.

According to the present embodiment, it is possible to provide the information processing apparatus 20 and the like that presents the determination result and the determination result regarding a predetermined disease such as ulcerative colitis. Whether the doctor outputs a correct result based on the diagnostic standard by checking both the diagnosis probability of the disease output by the second model 62 and the score regarding the diagnostic standard output by the first model 61. Can be confirmed.

If there is a discrepancy between the output of the second model 62 and the output of the first model 61, the doctor suspects a disease other than ulcerative colitis, consults with the instructor, or adds necessary tests. Etc. can be dealt with. As described above, it is possible to avoid rare oversight of diseases.

The diagnostic reference prediction using the first model 61 and the diagnostic prediction using the second model 62 may be executed by different hardware.

The endoscopic image 49 may be an image recorded in an electronic medical chart system or the like. For example, by inputting each image captured during follow-up observation to the first model 61, it is possible to provide the diagnosis support system 10 that can compare the temporal changes of each score.

[First Modification]
FIG. 7: is explanatory drawing explaining the outline|summary of the diagnostic assistance system 10 of a 1st modification. Descriptions other than the differences from FIG. 2 are omitted. The display device 16 includes a first display device 161 and a second display device 162. The first display device 161 is connected to the display device I/F 26. The second display device 162 is connected to the endoscope processor 11. It is desirable that the first display device 161 and the second display device 162 be arranged adjacent to each other.

The endoscopic image 49 generated by the endoscopic processor 11 is displayed on the first display device 161 in real time. The second display device 162 displays the diagnosis prediction and the diagnosis reference prediction acquired by the control unit 21.

According to this modification, it is possible to provide the diagnosis support system 10 that displays the diagnostic prediction and the diagnostic reference prediction while reducing the time lag of displaying the endoscopic image 49.

The diagnosis support system 10 may have three or more display devices 16. For example, the endoscopic image 49, the first result column 71, and the second result column 72 may be displayed on different display devices 16.

[Second Modification]
FIG. 8 is an explanatory diagram illustrating screen display of the second modification. Descriptions are omitted except for differences from the lower part of FIG. In this modification, the CPU 21 outputs the first result column 71 and the second result column 72 in a graph format.

In the first result column 71, three diagnostic reference predictions are displayed in a triaxial graph format. In FIG. 8, the upward axis indicates the predicted value of the first score, that is, the score for redness. The lower right axis indicates the predicted value of the second score, that is, the score related to vascular see-through. The lower left axis shows the predicted value of the third score, the score for ulcers.

The predicted values of the first score, the second score, and the third score are displayed by the inner triangle. In the second result column 72, the diagnostic prediction output from the second model 62 is displayed as a bar graph. According to this modification, the doctor can intuitively understand the diagnostic criterion prediction by looking at the triangle and the bar graph.

[Third Modification]
FIG. 9 is an explanatory diagram illustrating screen display of the third modification. FIG. 9 is a screen displayed by the diagnosis support system 10 that supports the diagnosis of Crohn's disease. Crohn's disease is also a type of inflammatory bowel disease like ulcerative colitis. In FIG. 9, the first score is the degree of longitudinal ulcers extending in the lengthwise direction of the intestinal tract, the second score is the degree of cobblestone statues of dense mucosal ridges, and the third score is the degree of red spotted afterh. Show.

Diseases that the diagnosis support system 10 supports for diagnosis are not limited to ulcerative colitis and Crohn's disease. It is possible to provide the diagnosis support system 10 that supports the diagnosis of any disease for which the appropriate first model 61 and second model 62 can be created. The user may be able to switch which disease is assisted in the diagnosis during the endoscopic examination. Information that supports the diagnosis of each disease may be displayed on the plurality of display devices 16.

[Fourth Modification]
FIG. 10 is a time chart schematically explaining the operation of the fourth modified example. Descriptions of portions common to FIG. 5 are omitted. FIG. 10 shows an example of a time chart when a long time is required for the processing using the first model 61 and the second model 62.

At time t ₀ , the image sensor 141 captures the frame “a”. The endoscopic processor 11 performs image processing to generate an endoscopic image 49 of “a” at time t ₁ . The control unit 21 acquires the endoscopic image 49 generated by the endoscopic processor 11 and inputs the endoscopic image 49 to the first model 61 and the second model 62. At time t ₂ , the control unit 21 outputs the endoscopic image 49 of “a” to the display device 16.

At time t ₆ , the image sensor 141 captures the frame “b”. The endoscopic processor 11 performs image processing to generate an endoscopic image 49 of “b” at time t ₇ . The endoscopic image 49 of “b” is not input to the first model 61 and the second model 62. At time t ₈ , the control unit 21 outputs the endoscopic image 49 of “b” to the display device 16.

At time t ₉ , the control unit 21 acquires predictions based on the endoscopic image 49 of “a” output from the first model 61 and the second model 62, respectively. At time t ₁₀ , the control unit 21 outputs the prediction based on the endoscopic image 49 of “a” to the display device 16. At time t ₁₂ , the image sensor 141 captures the frame “c”. The subsequent processing is the same as that from the time t ₀ to the time t ₁₀ , and thus the description thereof is omitted. As described above, the endoscopic image 49 and the predictions by the first model 61 and the second model 62 are displayed in synchronization with each other.

According to this modification, even if it takes time to process using the first model 61 and the second model 62 by thinning out the endoscopic images 49 input to the first model 61 and the second model 62. Real-time display can be realized.

[Embodiment 2]
The present embodiment relates to a model generation system 19 that generates a first model 61 and a second model 62. Descriptions of portions common to the first embodiment will be omitted.

FIG. 11 is an explanatory diagram for explaining the outline of the process of generating the model. In the teacher data DB 64 (see FIG. 12), a plurality of sets of teacher data in which the endoscopic image 49 and the judgment result by an expert such as a skilled specialist are associated with each other are recorded. The judgment result by the expert is the diagnosis of ulcerative colitis based on the endoscopic image 49, the first score, the second score, and the third score.

The second model 62 is generated by machine learning using a set of the endoscopic image 49 and the diagnosis result as teacher data. A first score learning model 611 is generated by machine learning using a set of the endoscopic image 49 and the first score as teacher data. A second score learning model 612 is generated by machine learning using a set of the endoscopic image 49 and the second score as teacher data. A third score learning model 613 is generated by machine learning using a set of the endoscopic image 49 and the third score as teacher data.

FIG. 12 is an explanatory diagram illustrating the configuration of the model generation system 19. The model generation system 19 includes a server 30 and a client 40. The server 30 includes a control unit 31, a main storage device 32, an auxiliary storage device 33, a communication unit 34, and a bus. The client 40 includes a control unit 41, a main storage device 42, an auxiliary storage device 43, a communication unit 44, a display unit 46, an input unit 47, and a bus.

The control unit 31 is an arithmetic and control unit that executes the program of this embodiment. For the control unit 31, one or more CPUs, multi-core CPUs, GPUs, etc. are used. The control unit 31 is connected to each of the hardware units configuring the server 30 via a bus.

The main storage device 32 is a storage device such as SRAM, DRAM, or flash memory. The main storage device 32 temporarily stores information required during the process performed by the control unit 31 and a program being executed by the control unit 31.

The auxiliary storage device 33 is a storage device such as SRAM, flash memory, hard disk or magnetic tape. The auxiliary storage device 33 stores a program to be executed by the control unit 31, a teacher data DB 64, and various data necessary for executing the program. Further, the first model 61 and the second model 62 generated by the control unit 31 are also stored in the auxiliary storage device 33. The teacher data DB 64, the first model 61, and the second model 62 may be stored in an external mass storage device or the like connected to the server 30.

The server 30 is a general-purpose personal computer, a tablet, a large computer, a virtual machine operating on the large computer, a cloud computing system, or a quantum computer. The server 30 may be a plurality of personal computers that perform distributed processing.

The control unit 41 is an arithmetic and control unit that executes the program of this embodiment. The control unit 41 is an arithmetic and control unit that executes the program of this embodiment. For the control unit 41, one or a plurality of CPUs, a multi-core CPU, a GPU, or the like is used. The control unit 41 is connected to each of the hardware units configuring the client 40 via a bus.

The main storage device 42 is a storage device such as SRAM, DRAM, or flash memory. The main storage device 42 temporarily stores information required during the process performed by the control unit 41 and a program being executed by the control unit 41.

The auxiliary storage device 43 is a storage device such as an SRAM, a flash memory, or a hard disk. The auxiliary storage device 43 stores a program to be executed by the control unit 41 and various data required for executing the program.

The communication unit 44 is an interface that performs data communication between the client 40 and the network. The display unit 46 is, for example, a liquid crystal display panel, an organic EL (Electro Luminescence) display panel, or the like. The input unit 47 is, for example, the keyboard 17 and the mouse. The client 40 may have a touch panel in which a display unit 46 and an input unit 47 are stacked.

The client 40 is an information device such as a general-purpose computer, tablet, or smartphone used by a specialist who creates teacher data. The client 40 may be a so-called thin client that realizes a user interface under the control of the control unit 31. When a thin client is used, most of the processing described below performed by the client 40 is executed by the control unit 31 instead of the control unit 41.

FIG. 13 is an explanatory diagram illustrating a record layout of the teacher data DB 64. The teacher data DB 64 is a DB that records teacher data used to generate the first model 61 and the second model 62. The teacher data DB 64 has a site field, a disease field, an endoscopic image field, an endoscopic finding field, and a score field. The score field has a reddish field, a blood vessel see-through field and an ulcer field.

The region where the endoscopic image 49 is captured is recorded in the region field. In the disease field, the name of the disease judged by a specialist when creating the teacher data is recorded. An endoscopic image 49 is recorded in the endoscopic image field. In the endoscopic finding field, the state of a disease judged by a specialist or the like by looking at the endoscopic image 49, that is, the endoscopic finding is recorded.

In the reddish field, the first score relating to reddishness, which is judged by a specialist or the like by viewing the endoscopic image 49, is recorded. In the blood vessel see-through field, a second score regarding the blood vessel see-through, which is judged by a specialist or the like by looking at the endoscopic image 49, is recorded. In the ulcer field, a third score regarding blood vessels, which is judged by a specialist or the like by looking at the endoscopic image 49, is recorded. The teacher data DB 64 has one record for one endoscopic image 49.

14 and 15 are explanatory diagrams illustrating the teacher data input screen. FIG. 14 shows an example of a screen displayed on the display unit 46 by the control unit 41 when creating teacher data without using the existing first model 61 and second model 62.

The screen shown in FIG. 14 includes an endoscopic image field 73, a first input field 81, a second input field 82, a next button 89, a patient ID field 86, a disease name field 87, and a model button 88. The first input field 81 includes a first score input field 811, a second score input field 812, and a third score input field 813. In FIG. 14, the model button 88 is set to the “model not used” state.

An endoscopic image 49 is displayed in the endoscopic image column 73. The endoscopic image 49 may be an image taken by an endoscopic examination performed by a specialist who inputs teacher data or an image delivered from the server 30. The specialist or the like makes a diagnosis regarding “ulcerative colitis” displayed in the disease name column 87 based on the endoscopic image 49, and selects the check box provided at the left end of the second input column 82.

The “unsuitable image” means that a specialist or the like has determined that the image is unsuitable for use in diagnosis due to a large amount of residue or blurring. The endoscopic image 49 determined to be the “unsuitable image” is not recorded in the teacher data DB 64.

The specialist or the like determines the first score to the third score based on the endoscopic image 49, and inputs them into the first score input field 811 to the third score input field 813, respectively. After completing the input, the specialist or the like selects the next button 89. The control unit 41 transmits the endoscopic image 49, the input to the first input field 81, and the input to the second input field 82 to the server 30. The control unit 31 adds a new record to the teacher data DB 64 and records the endoscopic image 49, the endoscopic findings and each score.

FIG. 15 shows an example of a screen displayed on the display unit 46 by the control unit 41 when creating teacher data by referring to the existing first model 61 and second model 62. In FIG. 15, the model button 88 is set to the “model in use” state. When the existing first model 61 and second model 62 have not been generated, the model button 88 is set so that the state of “model in use” is not selected.

The result of inputting the endoscopic image 49 into the first model 61 and the second model 62 is displayed in the first input field 81 and the second input field 82. In the second input field 82, the check box at the left end of the item with the highest probability is checked by default.

The specialist or the like determines whether or not each score in the first input field 81 is correct based on the endoscopic image 49, and changes the score as necessary. The specialist or the like determines whether or not the check in the second input field 82 is correct based on the endoscopic image 49, and reselects the check box as necessary. After setting the first input field 81 and the second input field 82 to appropriate states, the specialist or the like selects the next button 89. The subsequent processing is the same as the case of “model not used” described with reference to FIG.

FIG. 16 is a flowchart illustrating the flow of processing of a program that generates a learning model. The program described with reference to FIG. 16 is used to generate each learning model forming the first model 61 and the second model 62.

The control unit 31 selects a learning model to be created (step S522). The learning model to be created is any one of the learning models forming the first model 61 or the second model 62. The control unit 31 extracts necessary fields from the teacher data DB 64 and creates teacher data composed of a pair of the endoscopic image 49 and output data (step S523).

For example, when the first score learning model 611 is generated, the output data is a score regarding redness. The control unit 31 extracts the endoscopic image field and the reddish field from the teacher data DB 64. Similarly, when generating the second model 62, the output data is endoscopic findings. The control unit 31 extracts the endoscopic image field and the endoscopic finding field from the teacher data DB 64.

The control unit 31 separates the training data created in step S523 into training data and test data (step S524). The control unit 31 uses the training data and adjusts the parameters of the intermediate layer 532 using the error back propagation method or the like to perform supervised machine learning and generate a learning model (step S525).

The control unit 31 verifies the accuracy of the learning model using the training data (step S526). The verification is performed by calculating the probability that the output matches the output data corresponding to the endoscopic image 49 when the endoscopic image 49 in the training data is input to the learning model.

The control unit 31 determines whether or not the accuracy of the learning model generated in step S525 is acceptable (step S527). When it is determined that the learning model is acceptable (YES in step S527), the control unit 31 records the learning model in the auxiliary storage device 33. (Step S528).

When it is determined that the process is not passed (NO in step S527), the control unit 31 determines whether to end the process (step S529). For example, when the processing from step S524 to step S529 is repeated a predetermined number of times, the control unit 31 determines to end the processing. When it is determined that the process is not finished (NO in step S529), the control unit 31 returns to step S524.

When it is determined that the processing is to be ended (YES in step S529), or after the end of step S528, the control unit 31 determines whether to end the processing (step S531). When it is determined that the process is not finished (NO in step S531), the control unit 31 returns to step S522. When it is determined that the process is to be ended (YES in step S531), the control unit 31 ends the process.

If a learning model that is determined to be acceptable is not generated, after reviewing each record recorded in the teacher data DB 64 and adding a record, the program described using FIG. 16 is used. Is executed again.

The first model 61 and the second model 62 updated by the program described with reference to FIG. 16 are, for example, via a network or a recording medium after procedures such as approval under the law of pharmaceuticals and medical devices are completed. Is distributed to the information processing device 20 via.

FIG. 17 is a flowchart illustrating the flow of processing of the program for updating the learning model. The program described using FIG. 17 is appropriately executed when an additional record is recorded in the teacher data DB 64. The additional teacher data may be recorded in a database different from the teacher data DB 64.

The control unit 31 acquires the learning model to be updated (step S541). The control unit 31 acquires additional teacher data (step S542). Specifically, the control unit 31 extracts the endoscopic image 49 recorded in the endoscopic image field and the output data corresponding to the learning model acquired in step S541 from the record added to the teacher data DB 64. get.

The control unit 31 sets the endoscopic image 49 as the input data of the learning model, and sets the output data associated with the endoscopic image 49 as the output of the learning model (step S543). The control unit 31 updates the parameters of the learning model by the error back propagation method (step S544). The control unit 31 records the updated parameter (step S545).

The control unit 31 determines whether or not the processing of the record added to the teacher data DB 64 has been completed (step S546). When it is determined that the processing has not ended (NO in step S546), the control unit 31 returns to step S542. When it is determined that the process is completed (YES in step S546), the control unit 31 ends the process.

The first model 61 and the second model 62 updated by the program described with reference to FIG. 17 are, for example, via a network or a recording medium after procedures such as approval under the law of pharmaceuticals and medical devices have been completed. Is distributed to the information processing device 20 via. As described above, the first model 61 and the second model 62 are updated. The learning models forming the first model 61 and the second model 62 may be updated at the same time or individually.

FIG. 18 is a flowchart illustrating the flow of processing of a program that collects teacher data. The control unit 41 acquires the endoscopic image 49 from a hard disk or the like mounted in the electronic medical chart system or the endoscope processor 11 (not shown) (step S551). The control unit 41 determines whether or not to use the model is selected via the model button 88 described using FIG. 14 (step S552).

When it is determined that the use of the model is not selected (NO in step S552), the control unit 41 displays the screen described using FIG. 14 on the display unit 46 (step S553). When it is determined that the use of the model is selected (YES in step S552), the control unit 41 acquires the first model 61 and the second model 62 from the server 30 (step S561).

The control unit 41 may temporarily store the acquired first model 61 and second model 62 in the auxiliary storage device 43. By doing so, the control unit 41 can omit the processing of step S561 from the second time.

The control unit 41 inputs the endoscopic image 49 acquired in step S551 into the first model 61 and the second model 62 acquired in step S561, respectively, and acquires the estimation result output from the output layer 533 (step). S562). The control unit 41 displays the screen described with reference to FIG. 15 on the display unit 46 (step S563).

After the end of step S553 or step S563, the control unit 41 acquires the determination result input by the user via the input unit 47 (step S564). The control unit 41 determines whether or not “inappropriate image” is selected in the second input field 82 (step S565). When it is determined that “inappropriate image” is selected (YES in step S565), the control unit 41 ends the process.

When it is determined that the “inappropriate image” is not selected (NO in step S565), the control unit 41 transmits the teacher record in which the endoscopic image 49 and the input result by the user are associated with each other to the server 30 (step). S566). The teacher record may be recorded in the teacher data DB 64 via a portable recording medium such as a USB (Universal Serial Bus) memory.

The control unit 31 creates a new record in the teacher data DB 64 and records the received teacher record. It should be noted that, for example, a plurality of experts may judge the same endoscopic image 49, and when the judgment by a predetermined number of experts coincides, it may be recorded in the teacher data DB 64. By doing so, the accuracy of the teacher data DB 64 can be improved.

According to the present embodiment, teacher data can be collected, and the first model 61 and the second model 62 can be generated and updated.

[Third Embodiment]
The present embodiment relates to a diagnosis support system 10 that outputs a score in accordance with a diagnosis criterion, based on the feature amount extracted from the intermediate layer 532 of the second model 62. Descriptions of portions common to the first or second embodiment will be omitted.

FIG. 19 is an explanatory diagram illustrating an overview of the diagnosis support system 10 according to the third embodiment. The endoscopic image 49 captured by using the endoscope 14 is input to the second model 62. The second model 62 outputs a diagnostic prediction of ulcerative colitis when the endoscopic image 49 is input. As will be described later, the feature amount 65 such as the first feature amount 651, the second feature amount 652, and the third feature amount 653 is acquired from the nodes forming the intermediate layer 532 of the second model 62.

The first model 61 includes a first converter 631, a second converter 632 and a third converter 633. The first feature amount 651 is converted by the first converter 631 into a predicted value of a first score indicating the degree of redness. The second feature amount 652 is converted by the second converter 632 into a predicted value of the second score indicating the degree of blood vessel see-through. The third feature amount 653 is converted by the third converter 633 into a predicted value of a third score indicating the degree of ulcer. In the following description, the first converter 631 to the third converter 633 will be referred to as the converter 63 unless otherwise specified.

The outputs of the first model 61 and the second model 62 are acquired by the first acquisition unit and the second acquisition unit, respectively. The screen shown in the lower part of FIG. 19 is displayed on the display device 16 based on the outputs acquired by the first acquisition unit and the second acquisition unit. The displayed screen is the same as the screen described in the first embodiment, and thus the description is omitted.

FIG. 20 is an explanatory diagram illustrating the characteristic amount acquired from the second model 62. The middle layer 532 includes a plurality of nodes connected to each other. When the endoscopic image 49 is input to the second model 62, various feature amounts of the endoscopic image 49 appear in each node. As an example, the respective characteristic amounts appearing in the five nodes are indicated by symbols from the characteristic amount A65A to the characteristic amount E65E.

The feature amount may be acquired from a node immediately before being input to the fully connected layer after being repeatedly processed by the convolutional layer and the pooling layer, or may be acquired from a node included in the fully connected layer.

FIG. 21 is an explanatory diagram illustrating conversion between the feature amount and the score. The upper part of FIG. 21 schematically shows the teacher data included in the teacher data DB 64. In the teacher data DB 64, teacher data that associates the endoscopic image 49 with the determination result by an expert such as a specialist is recorded. Since the record layout of the teacher data DB 64 is the same as that of the teacher data DB 64 of the first embodiment described with reference to FIG. 13, description thereof will be omitted.

As described above, the endoscopic image 49 is input to the second model 62, and a plurality of feature amounts such as the feature amount A65A are acquired. Correlation analysis is performed between the acquired feature amount and the first to third scores associated with the endoscopic image 49, and the feature amount having a high correlation with each score is selected. FIG. 21 shows a case where the correlation between the first score and the feature amount A65A, the correlation between the second score and the feature amount C65C, and the correlation between the third score and the feature amount D65D are high.

The first converter 631 is obtained by performing a regression analysis of the first score and the feature amount A65A. Similarly, the second converter 632 is obtained by the regression analysis of the second score and the feature amount C65C, and the third converter 633 is obtained by the regression analysis of the third score and the feature amount D65D. Linear regression or non-linear regression may be used for the regression analysis. The regression analysis may be performed using a neural network.

FIG. 22 is an explanatory diagram illustrating a record layout of the feature amount DB. The feature amount DB is a DB that records the teacher data and the feature amount acquired from the endoscopic image 49 in association with each other. The feature amount DB has a region field, a disease field, an endoscopic image field, an endoscopic finding field, a score field, and a feature amount field. The score field has a reddish field, a blood vessel see-through field and an ulcer field. The feature amount field has many subfields such as A field and B field.

The region where the endoscopic image 49 is captured is recorded in the region field. In the disease field, the name of the disease judged by a specialist when creating the teacher data is recorded. An endoscopic image 49 is recorded in the endoscopic image field. In the endoscopic finding field, the state of a disease judged by a specialist or the like by looking at the endoscopic image 49, that is, the endoscopic finding is recorded.

In the reddish field, the first score relating to reddishness, which is judged by a specialist or the like by viewing the endoscopic image 49, is recorded. In the blood vessel see-through field, a second score regarding the blood vessel see-through, which is judged by a specialist or the like by looking at the endoscopic image 49, is recorded. In the ulcer field, a third score regarding the ulcer, which is judged by a specialist or the like by looking at the endoscopic image 49, is recorded. In each subfield of the feature amount field, the feature amount such as the feature amount A64A acquired from each node of the intermediate layer 532 is recorded.

The feature amount DB has one record for one endoscopic image 49. The feature amount DB is stored in the auxiliary storage device 33. The feature amount DB may be stored in an external mass storage device or the like connected to the server 30.

FIG. 23 is a flowchart illustrating the flow of processing of a program that creates the converter 63. The control unit 31 selects one record from the teacher data DB 64 (step S571). The control unit 31 inputs the endoscopic image 49 recorded in the endoscopic image field to the second model 62, and acquires the feature amount from each node of the intermediate layer 532 (step S572). The control unit 31 creates a new record in the feature amount DB, and records the data recorded in the record acquired in step S571 and the feature amount acquired in step S572 (step S573).

The control unit 31 determines whether to end the process (step S574). For example, when the processing of the predetermined number of teacher data records is completed, the control unit 31 determines to end the processing. When it is determined that the process is not finished (NO in step S574), the control unit 31 returns to step S571.

When it is determined that the process is to be ended (YES in step S574), the control unit 31 selects one subfield from the score fields of the feature amount DB (step S575). The control unit 31 selects one subfield from the feature amount fields of the feature amount DB (step S576).

The control unit 31 performs a correlation analysis between the score selected in step S575 and the feature amount selected in step S576 to calculate a correlation coefficient (step S577). The control unit 31 temporarily records the calculated correlation coefficient in the main storage device 32 or the auxiliary storage device 33 (step S578).

The control unit 31 determines whether to end the process (step S579). For example, the control unit 31 determines to end the process when the correlation analysis of all the combinations of the score and the feature amount is completed. The control unit 31 may determine to end the process when the correlation coefficient calculated in step S577 is greater than or equal to a predetermined threshold.

When it is determined that the process is not finished (NO in step S579), the control unit 31 returns to step S576. When it is determined that the process is to be ended (YES in step S579), the control unit 31 selects the feature amount having the highest correlation with the score selected in step S575 (step S580).

The control unit 31 performs a regression analysis using the score selected in step S575 as an objective variable and the feature amount selected in step S580 as an explanatory variable, and calculates a parameter that specifies the converter 63 that converts the feature amount into a score ( Step S581). For example, when the score selected in step S575 is the first score, the converter 63 specified in step S581 is the first converter 631, and when the score selected in step S575 is the second score, The converter 63 specified in step S581 is the second converter 632. The control unit 31 stores the calculated converter 63 in the auxiliary storage device 33 (step S582).

The control unit 31 determines whether or not the processing of all score fields of the feature amount DB has been completed (step S583). When it is determined that the processing is not completed (NO in step S583), the control unit 31 returns to step S575. When it is determined that the process is completed (YES in step S583), the control unit 31 ends the process. By the above, each converter 63 which comprises the 1st model 61 is generated.

The first model 61 including the converter 63 created by the program described with reference to FIG. 23 is, for example, via a network or a recording medium after a procedure such as approval under the law of pharmaceutical medical equipment is completed. Is distributed to the information processing device 20 via.

FIG. 24 is a flowchart illustrating the flow of processing of a program during endoscopic examination according to the third embodiment. The program of FIG. 24 is executed by the control unit 21 instead of the program described using FIG.

The control unit 21 acquires the endoscopic image 49 from the endoscopic processor 11 (step S501). The control unit 21 inputs the acquired endoscopic image 49 into the second model 62 and acquires the diagnostic prediction output from the output layer 533 (step S502).

The control unit 21 acquires the characteristic amount from a predetermined node included in the middle layer 532 of the second model 62 (step S601). The predetermined node is the node from which the feature amount selected in step S580 described with reference to FIG. 23 is acquired. The control unit 21 converts the acquired feature amount by the converter 63 to calculate a score (step S602).

The control unit 21 determines whether or not the calculation of all scores has been completed (step S603). When it is determined that the processing has not ended (NO in step S603), the control unit 21 returns to step S601. When it is determined that the processing is completed (YES in step S603), the control unit 21 generates the image described using the lower part of FIG. 19 and outputs the image to the display device 16 (step S604). The control unit 21 ends the process.

According to the present embodiment, since the learning model generated by deep learning is only the second model 62, the diagnosis support system 10 can be realized with a relatively small amount of calculation.

By acquiring the feature amount from the middle layer 532 of the second model 62, it is possible to obtain the feature amount having a high correlation with the score, without being limited to the feature amount in the range that a person can normally think of. Therefore, each diagnostic reference prediction can be accurately calculated based on the endoscopic image 49.

Note that some of the first score, the second score, and the third score may be calculated by the same method as in the first embodiment.

[Embodiment 4]
The present embodiment relates to an information processing system that calculates a diagnostic reference prediction based on a method other than deep learning. Descriptions of portions common to the first or second embodiment will be omitted.

FIG. 25 is an explanatory diagram illustrating an overview of the diagnosis support system 10 according to the fourth embodiment. The endoscopic image 49 captured by using the endoscope 14 is input to the second model 62. The second model 62 outputs a diagnostic prediction of ulcerative colitis when the endoscopic image 49 is input.

The first model 61 includes a first converter 631, a second converter 632 and a third converter 633. The first converter 631 outputs a predicted value of the first score indicating the degree of redness when the endoscopic image 49 is input. The second converter 632 outputs the predicted value of the second score indicating the degree of blood vessel see-through when the endoscopic image 49 is input. The third converter 633 outputs the predicted value of the third score indicating the degree of ulcer when the endoscopic image 49 is input.

The outputs of the first model 61 and the second model 62 are acquired by the first acquisition unit and the second acquisition unit, respectively. The screen shown in the lower part of FIG. 25 is displayed on the display device 16 based on the outputs acquired by the first acquisition unit and the second acquisition unit. The displayed screen is the same as the screen described in the first embodiment, and thus the description is omitted.

FIG. 26 is an explanatory diagram illustrating conversion between the endoscopic image 49 and the score according to the fourth embodiment. Note that the illustration of the second model 62 is omitted in FIG.

In the present embodiment, various converters 63 such as the converter A 63A and the converter B 63B that output the feature amount when the endoscopic image 49 is input are used. For example, the converter A63A converts the endoscopic image 49 into a feature amount A65A.

The converter 63 converts the endoscopic image 49 into a feature amount based on, for example, the number or ratio of pixels satisfying a predetermined condition. The converter 63 may convert the endoscopic image 49 into a feature amount by classification using SVM (Support Vector Machine), random forest, or the like.

Correlation analysis is performed between the feature amount converted by the converter 63 and the first score to the third score associated with the endoscopic image 49, and the feature amount having high correlation with each score is obtained. To be selected. FIG. 26 shows a case where the correlation between the first score and the feature amount A65A, the correlation between the second score and the feature amount C65C, and the correlation between the third score and the feature amount D65D are high.

The regression analysis of the first score and the feature amount A65A is performed, and the first converter 631 is obtained by combining the first score and the feature A65A with the converter A63A. Similarly, regression analysis is performed on the second score and the feature amount C65C, and the second converter 632 is obtained by combining the second score and the feature amount C65C with the converter C63C.

FIG. 27 is a flowchart for explaining the processing flow of the program that creates the converter 63 according to the fourth embodiment. The control unit 31 selects one record from the teacher data DB 64 (step S611). The control unit 31 uses the plurality of converters 63 such as the converter A 63A and the converter B 63B to convert the endoscopic image 49 recorded in the endoscopic image field into a feature amount (step S612). The control unit 31 creates a new record in the feature amount DB, and records the data recorded in the record acquired in step S611 and the feature amount acquired in step S612 (step S613).

The control unit 31 determines whether to end the process (step S614). For example, when the processing of the predetermined number of teacher data records is completed, the control unit 31 determines to end the processing. When it is determined that the process is not finished (NO in step S614), the control unit 31 returns to step S611.

When it is determined that the process is to be ended (YES in step S614), the control unit 31 selects one subfield from the score fields of the feature amount DB (step S575). Since the processing from step S575 to step S581 is the same as the processing flow of the program described with reference to FIG. 23, description thereof will be omitted.

The control unit 31 combines the result obtained by the regression analysis with the converter 63 that has converted the endoscopic image 49 into the feature amount in step S612 to calculate a new converter 63 (step S620). The control unit 31 stores the calculated converter 63 in the auxiliary storage device 33 (step S621).

The control unit 31 determines whether or not the processing of all score fields of the feature amount DB has been completed (step S622). When it is determined that the processing has not ended (NO in step S622), the control unit 31 returns to step S575. When it is determined that the process is completed (YES in step S622), the control unit 31 ends the process. By the above, each converter 63 which comprises the 1st model 61 is generated.

The first model 61 including the converter 63 created by the program described with reference to FIG. 27 is, for example, via a network or a recording medium after procedures such as approval under the law of pharmaceuticals and medical devices are completed. Is distributed to the information processing device 20 via.

FIG. 28 is a flowchart illustrating the flow of processing of a program at the time of endoscopy according to the fourth embodiment. The program of FIG. 28 is executed by the control unit 21 instead of the program described using FIG.

The control unit 21 acquires the endoscopic image 49 from the endoscopic processor 11 (step S501). The control unit 21 inputs the acquired endoscopic image 49 into the second model 62 and acquires the diagnostic prediction output from the output layer 533 (step S502).

The control unit 21 inputs the acquired endoscopic image 49 to the converter 63 included in the first model 61 to calculate a score (step S631).

The control unit 21 determines whether or not the calculation of all scores has been completed (step S632). When it is determined that the processing has not ended (NO in step S632), the control unit 21 returns to step S631. When it is determined that the processing is completed (YES in step S632), the control unit 21 generates the image described using the lower portion of FIG. 25 and outputs the image to the display device 16 (step S633). The control unit 21 ends the process.

According to the present embodiment, since the learning model generated by deep learning is only the second model 62, the diagnosis support system 10 can be realized with a relatively small amount of calculation.

Note that some of the first score, the second score, and the third score may be calculated by the same method as in the first or third embodiment.

[Fifth Embodiment]
The present embodiment relates to a diagnosis support system 10 that supports the diagnosis of a localized disease such as cancer or polyp. Descriptions of portions common to the first or second embodiment will be omitted.

FIG. 29 is an explanatory diagram illustrating an overview of the diagnosis support system 10 according to the fifth embodiment. The endoscopic image 49 captured by using the endoscope 14 is input to the second model 62. In the second model 62, when the endoscopic image 49 is input, a region prediction that predicts a range of a lesion region 74 in which a lesion such as a polyp or a cancer is predicted, and the lesion are predicted. It outputs a diagnostic prediction whether it is benign or malignant. In FIG. 29, it is predicted that the probability that the polyp in the lesion area 74 is “malignant” is 5%, and the probability that it is “benign” is 95%.

The second model 62 uses an arbitrary object detection algorithm such as RCNN (Regions with Convolutional Neural Network), Fast RCNN, Faster RCNN, SSD (Single Shot Multibook Detector), or YOLO (You Only Look Once). This is the generated learning model. Since a learning model that receives an input of a medical image and outputs a region in which a lesion portion exists and diagnostic prediction has been conventionally used, a detailed description thereof will be omitted.

The first model 61 includes a first score learning model 611, a second score learning model 612, and a third score learning model 613. The first score learning model 611 outputs a predicted value of the first score indicating the degree of clarity of the boundary when an image in the lesion area 74 is input. The second score learning model 612 outputs the predicted value of the second score indicating the degree of irregularity of the surface when the image in the lesion area 74 is input. The third score learning model 613 outputs the predicted value of the third score indicating the degree of redness when an image in the lesion area 74 is input.

In the example shown in FIG. 29, predicted values that the first score is 50, the second score is 5, and the third score is 20 are output. The first model 61 may include a score learning model that outputs diagnostic criteria predictions regarding various diagnostic criteria items regarding polyps, such as the shape such as pedunculated and the degree of secretion attachment.

The outputs of the first model 61 and the second model 62 are acquired by the first acquisition unit and the second acquisition unit, respectively. The screen shown in the lower part of FIG. 29 is displayed on the display device 16 based on the outputs acquired by the first acquisition unit and the second acquisition unit. The displayed screen is the same as the screen described in the first embodiment, and thus the description is omitted.

When a plurality of lesion areas 74 are detected in the endoscopic image 49, each lesion area 74 is input to the first model 61, and the diagnostic reference prediction is output. By selecting the lesion area 74 displayed in the endoscopic image column 73, the user can browse the diagnostic prediction and the score regarding the lesion area 74. Note that the diagnostic predictions and scores relating to the plurality of lesion areas 74 may be displayed as a list on the screen.

The lesion area 74 may be surrounded by a circle, an ellipse, or any closed curve. In such a case, the peripheral area is masked with black or white, and an image corrected to have a shape suitable for input to the first model 61 is input to the first model 61. For example, when a plurality of polyps are close to each other, a region including one polyp can be cut out and the score can be calculated by the first model 61.

[Sixth Embodiment]
The present embodiment relates to a diagnosis support system 10 that outputs the probabilities that the first model 61 is each category defined in a diagnosis standard related to a disease. Descriptions of portions common to the first embodiment will be omitted.

FIG. 30 is an explanatory diagram illustrating a configuration of the first score learning model 611 according to the sixth embodiment. The first score learning model 611 described using FIG. 30 is used instead of the first score learning model 611 described using FIG.

In the first score learning model 611, when the endoscopic image 49 is input, the degree of redness is “determination 1”, “determination 2”, and “determination 3” based on the diagnostic criteria for ulcerative colitis. The output layer 533 has three output nodes that output the probabilities of each of the three stages. The "judgment 1" means that the degree of redness is "normal", the "judgment 2" is "erythema", and the "judgment 3" is "strong erythema".

Similarly, in the second score learning model 612, “judgment 1” indicates that the degree of blood vessel observation is “normal”, and “judgment 2” indicates that the blood vessel observation has “disappeared in a patchy manner”. "Judgment 3" means that "disappeared" over almost the entire area.

The number of nodes in the output layer 533 of the score learning model is arbitrary. In the present embodiment, the third score learning model 613 has four output nodes from “judgment 1” to “judgment 4” in the output layer 533. “Judgment 1” indicates that the ulcer degree is “none”, “Judgment 2” indicates “erosion”, and “judgment 3” indicates that the ulcer has a “medium” depth. Judgment 4 means a "deep" ulcer, respectively.

FIG. 31 is an explanatory diagram illustrating screen display according to the sixth embodiment. An endoscopic image field 73 is displayed in the upper left part of the screen. A first result column 71 and a first stop button 711 are displayed on the right side of the screen. Below the endoscope image column 73, the second result column 72 and the second stop button 722 are displayed.

According to the present embodiment, it is possible to provide the diagnosis support system 10 that displays the first result column 71 in an expression according to the definition defined in the diagnostic standard.

[Embodiment 7]
The present embodiment relates to a diagnosis support system 10 that displays an alert when there is a discrepancy between the output from the first model 61 and the output from the second model 62. Descriptions of portions common to the first embodiment will be omitted.

FIG. 32 is an explanatory diagram illustrating screen display according to the seventh embodiment. In the example shown in FIG. 32, the probability of being normal is 70%, the first score indicating the degree of redness is 70, the second score indicating the degree of vascular see-through is 50, and the third score indicating the degree of ulcer is 5. The diagnostic criterion prediction that there is is output.

A warning field 75 is displayed at the bottom of the screen. According to the diagnostic criteria, the warning column 75 should be determined to be not “normal” when the first score, which is the degree of “redness”, is high, so the first result column 71 and the second result Indicates that there is a discrepancy with the column 72. The presence or absence of discrepancies is determined by a rule base based on diagnostic criteria.

In this way, when there is a discrepancy between the output of the first model 61 and the output of the second model 62, the warning column 75 is displayed, so that the doctor who is the user can be alerted.

[Embodiment 8]
The present embodiment relates to a diagnosis support system 10 in which an endoscope processor 11 and an information processing device 20 are integrated. Descriptions of portions common to the first embodiment will be omitted.

FIG. 33 is an explanatory diagram illustrating an overview of the diagnosis support system 10 according to the eighth embodiment. Note that, in FIG. 33, the illustration and description of the configuration that realizes the basic functions of the endoscope processor 11 such as the light source, the air/water pump, and the control unit of the image sensor 141 are omitted.

The diagnosis support system 10 includes an endoscope 14 and an endoscope processor 11. The endoscope processor 11 includes an endoscope connection unit 12, a control unit 21, a main storage device 22, an auxiliary storage device 23, a communication unit 24, a display device I/F 26, an input device I/F 27, and a bus.

The control unit 21, the main storage device 22, the auxiliary storage device 23, the communication unit 24, the display device I/F 26, and the input device I/F 27 are the same as those in the first embodiment, and therefore description thereof will be omitted. The endoscope 14 is connected to the endoscope connecting portion 12 via an endoscope connector 15.

According to the present embodiment, the control unit 21 receives a video signal from the endoscope 14 via the endoscope connection unit 12 and performs various image processings to generate an endoscopic image 49 suitable for observation by a doctor. To generate. The control unit 21 inputs the generated endoscopic image 49 to the first model 61 and acquires the diagnostic reference prediction of each item according to the diagnostic reference. The control unit 21 inputs the generated endoscopic image 49 to the second model 62 and acquires the diagnosis prediction of the disease.

The first model 61 and the second model 62 may be configured to accept the video signal acquired from the endoscope 14 or an image in the process of generating the endoscopic image 49 based on the video signal. good. By doing so, it is possible to provide the diagnosis support system 10 in which information lost during the generation of an image suitable for observation by a doctor can be used.

[Ninth Embodiment]
FIG. 34 is a functional block diagram of the server 30 according to the ninth embodiment. The server 30 has an acquisition unit 381 and a generation unit 382. The acquisition unit 381 acquires a plurality of sets of teacher data in which the endoscopic image 49 and the determination result of the determination criterion used for the diagnosis of a disease are recorded in association with each other. The generation unit 382 uses the teacher data to generate a first model that outputs a diagnostic criterion prediction that predicts a diagnostic criterion of a disease when the endoscopic image 49 is input.

[Embodiment 10]
The present embodiment relates to a mode in which a general-purpose server computer 901, a client computer 902, and a program 97 are operated in combination to realize the model generation system 19 of the present embodiment. FIG. 35 is an explanatory diagram showing the structure of the model generation system 19 according to the tenth embodiment. Descriptions of parts common to the second embodiment will be omitted.

The model generation system 19 of this embodiment includes a server computer 901 and a client computer 902. The server computer 901 includes a control unit 31, a main storage device 32, an auxiliary storage device 33, a communication unit 34, a reading unit 39, and a bus. The server computer 901 is a general-purpose personal computer, a tablet, a large computer, a virtual machine operating on the large computer, a cloud computing system, or a quantum computer. The server computer 901 may be a plurality of personal computers that perform distributed processing.

The client computer 902 includes a control unit 41, a main storage device 42, an auxiliary storage device 43, a communication unit 44, a display unit 46, an input unit 47, and a bus. The client computer 902 is an information device such as a general-purpose personal computer, tablet, or smartphone.

The program 97 is recorded on the portable recording medium 96. The control unit 31 reads the program 97 via the reading unit 39 and stores it in the auxiliary storage device 33. The control unit 31 may also read the program 97 stored in the semiconductor memory 98 such as a flash memory installed in the server computer 901. Further, the control unit 31 may download the program 97 from another server computer (not shown) connected via the communication unit 24 and a network (not shown) and store the program 97 in the auxiliary storage device 33.

The program 97 is installed as a control program for the server computer 901, loaded into the main storage device 22 and executed. The control unit 31 distributes the part of the program 97 executed by the control unit 41 to the client computer 902 via the network. The distributed program 97 is installed as a control program for the client computer 902, loaded into the main storage device 42, and executed.

As a result, the server computer 901 and the client computer 902 function as the diagnosis support system 10 described above.

The technical features (constituent elements) described in the respective embodiments can be combined with each other, and by combining them, new technical features can be formed.
The embodiments disclosed this time are to be considered as illustrative in all points and not restrictive. The scope of the present invention is shown not by the above meaning but by the scope of the claims, and is intended to include meanings equivalent to the scope of the claims and all modifications within the scope.

(Appendix 1)
An image acquisition unit that acquires an endoscopic image,
When the endoscopic image is input, the endoscopic image acquired by the image acquisition unit is input to the first model that outputs the diagnostic standard prediction related to the diagnostic standard of the disease, and the output diagnostic standard prediction is acquired. A first acquisition unit that
An information processing apparatus, comprising: an output unit that outputs the diagnosis reference prediction acquired by the first acquisition unit in association with the diagnosis prediction related to the disease state acquired based on the endoscopic image.

(Appendix 2)
The information processing device according to appendix 1, wherein the first acquisition unit acquires the diagnostic reference prediction of each item from a plurality of first models that respectively output the diagnostic reference predictions of the plurality of items included in the diagnostic criteria for the disease. ..

(Appendix 3)
The first model is a learning model generated by machine learning. The information processing device according to supplementary note 1 or supplementary note 2.

(Appendix 4)
The information processing device according to supplementary note 1 or supplementary note 2, wherein the first model outputs a numerical value calculated based on the endoscopic image acquired by the image acquisition unit.

(Appendix 5)
The information processing apparatus according to any one of appendices 1 to 4, comprising a first accepting unit that accepts an operation stop instruction of the first acquisition unit.

(Appendix 6)
The diagnostic prediction is the output diagnostic prediction when the endoscopic image acquired by the image acquisition unit is input to a second model that outputs the diagnostic prediction of the disease when the endoscopic image is input. The information processing apparatus according to any one of Supplementary Notes 1 to 5.

(Appendix 7)
The information processing device according to attachment 6, wherein the second model is a learning model generated by machine learning.

(Appendix 8)
The second model is
An input layer to which an endoscopic image is input,
An output layer that outputs the diagnosis prediction of the disease,
A neural network model having an intermediate layer in which parameters are learned by a plurality of sets of teacher data recorded in association with endoscopic images and diagnostic predictions,
The information processing device according to supplementary note 6 or supplementary note 7, wherein the first model outputs a diagnostic reference prediction based on a feature amount acquired from a predetermined node of the intermediate layer.

(Appendix 9)
The second model outputs a region prediction regarding a lesion region including the disease when an endoscopic image is input,
The first model outputs a diagnostic criterion prediction regarding a diagnostic criterion of a disease when an endoscopic image of a lesion area is input,
The first acquisition unit inputs a portion of the endoscopic image acquired by the image acquisition unit, which corresponds to the region prediction output from the second model, to the first model, and outputs the diagnostic reference prediction. The information processing device according to supplementary note 6 or supplementary note 7.

(Appendix 10)
The information processing apparatus according to any one of appendixes 6 to 9, further comprising a second accepting unit that accepts an instruction to stop the acquisition of the diagnostic prediction.

(Appendix 11)
The information processing apparatus according to any one of appendixes 6 to 10, wherein the output unit also outputs the endoscopic image acquired by the image acquisition unit.

(Appendix 12)
The image acquisition unit acquires the endoscopic image taken during the endoscopy in real time,
The information processing apparatus according to any one of appendices 1 to 11, wherein the output unit outputs in synchronization with acquisition of the endoscopic image by the image acquisition unit.

(Appendix 13)
An endoscope connection unit to which an endoscope is connected, and an image generation unit that generates an endoscope image based on a video signal acquired from the endoscope connected to the endoscope connection unit,
When the endoscopic image is input, the endoscopic image generated by the image generation unit is input to the first model that outputs the diagnostic reference prediction related to the diagnostic reference of the disease, and the output diagnostic reference prediction is acquired. A first acquisition unit that
An output processor that outputs the diagnosis reference prediction acquired by the first acquisition unit in association with the diagnosis prediction related to the disease state acquired based on the endoscope image.

(Appendix 14)
Acquire an endoscopic image,
When the endoscopic image is input, the acquired endoscopic image is input to the first model that outputs the diagnostic standard prediction related to the diagnostic standard of the disease, and the diagnostic standard prediction that is output is acquired.
An information processing method for causing a computer to execute a process of outputting the acquired diagnosis reference prediction in association with the diagnosis prediction related to the disease state acquired based on the endoscopic image.

(Appendix 15)
Acquire an endoscopic image,
When the endoscopic image is input, the acquired endoscopic image is input to the first model that outputs the diagnostic standard prediction related to the diagnostic standard of the disease, and the diagnostic standard prediction that is output is acquired.
A program that causes a computer to execute a process of outputting the acquired diagnosis reference prediction in association with the diagnosis prediction related to the disease state acquired based on the endoscopic image.

10 Diagnostic Support System 11 Endoscope Processor 12 Endoscope Connection Section 14 Endoscope 141 Imaging Device 142 Insertion Section 15 Endoscope Connector 16 Display Device 161 First Display Device 162 Second Display Device 17 Keyboard 19 Model Generation System 20 information processing device 21 control unit 22 main storage device 23 auxiliary storage device 24 communication unit 26 display device I/F
27 Input device I/F
30 server 31 control unit 32 main storage device 33 auxiliary storage device 34 communication unit 381 acquisition unit 382 generation unit 39 reading unit 40 client 41 control unit 42 main storage device 43 auxiliary storage device 44 communication unit 46 display unit 47 input unit 49 introspection Mirror image 53 Neural network model 531 Input layer 532 Intermediate layer 533 Output layer 61 First model 611 First score learning model 612 Second score learning model 613 Third score learning model 62 Second model 63 Converter 631 First converter 632 Second converter 633 Third converter 64 Teacher data DB
65 feature amount 651 1st feature amount 652 2nd feature amount 653 3rd feature amount 71 1st result column 711 1st stop button 72 2nd result column 722 2nd stop button 73 endoscopic image column 74 lesion area 75 warning column 81 First input field 811 First score input field 812 Second score input field 813 Third score input field 82 Second input field 86 Patient ID field 87 Disease name field 88 Model button 89 Next button 901 Server computer 902 Client computer 96 Yes Portable recording medium 97 Program 98 Semiconductor memory

Claims (7)

Endoscopic images, to obtain a plurality of sets of teacher data recorded in association with the determination results determined for the diagnostic criteria used to diagnose the disease,
A model generating method for generating a first model that outputs a diagnostic criterion prediction that predicts a diagnostic criterion of a disease when an endoscopic image is input, using the teacher data. The teacher data includes a determination result determined for each of a plurality of diagnostic criteria items included in the diagnostic criteria,
The model generation method according to claim 1, wherein the first model is generated corresponding to each of the plurality of diagnostic reference items. The said 1st model is produced|generated by the deep learning which adjusts the parameter of an intermediate|middle layer so that the acquired determination result may be output from an output layer, when the acquired endoscopic image is input into an input layer. The method of generating a model according to claim 1 or claim 2. The first model is
To the neural network model that outputs the diagnosis prediction of the disease when the endoscopic image is input, input the endoscopic image in the acquired teacher data,
From the nodes constituting the intermediate layer of the neural network model, to obtain a plurality of feature quantities regarding the input endoscopic image,
From the plurality of acquired feature amounts, select a feature amount having a high correlation with the determination result associated with the endoscopic image,
The method according to claim 1 or 2, wherein a regression method of the selected feature amount and a score obtained by digitizing the determination result determines a calculation method for calculating the score based on the selected feature amount. How to generate the model. The first model is
Extracting multiple features from the acquired endoscopic image,
From the plurality of extracted feature amounts, select a feature amount having a high correlation with the determination result associated with the endoscopic image,
The method according to claim 1 or 2, wherein a regression method of the selected feature amount and a score obtained by digitizing the determination result determines a calculation method for calculating the score based on the selected feature amount. How to generate the model. The disease is ulcerative colitis,
The method for generating a model according to claim 1, wherein the diagnostic criterion prediction is prediction regarding redness of an endoscopic image, vascular see-through, or ulcer severity. Endoscopic images, to obtain a plurality of sets of teacher data recorded in association with the determination results determined for the diagnostic criteria used to diagnose the disease,
To the neural network model that outputs the diagnosis prediction of the disease when the endoscopic image is input, input the endoscopic image in the acquired teacher data,
From the nodes constituting the intermediate layer of the neural network model, to obtain a plurality of feature quantities regarding the input endoscopic image,
The acquired multiple feature quantities are recorded in association with the numerically converted score of the determination result associated with the input endoscopic image,
Based on the correlation between each of the plurality of recorded feature amount and the score, select a feature amount having a high correlation with the score,
By predicting the diagnostic criteria for the disease when an endoscopic image is input, by defining a calculation method for calculating the score based on the selected feature amount by regression analysis of the selected feature amount and the score. A program that causes a computer to execute a process of generating a first model that outputs the predicted diagnostic criterion.

Images