JP2022050089A - Determination prediction system, determination prediction method, and determination prediction program - Google Patents

Abstract

To provide a determination prediction system, determination prediction method, determination prediction program for supporting determination on diseases.SOLUTION: In a determination prediction system, a control section 21 of a support server performs: generating a plurality of division images obtained by dividing an original image; generating division prediction models for predicting a determination result from the division images with the use of first teacher information recorded by association with a determination result in the original image with respect to the plurality of division images; recording the models in a learning result storage section 24; generating second teacher information obtained by summarizing feature information which is generated by inputting the respective division images of the first teacher information to the division prediction models, based on the positions of the division images; generating an integrated prediction model for predicting a determination result with the use of the second teacher information; and recording the model in the learning result storage section 24. When an evaluation object image is acquired, the control section 21 predicts a determination result with the use of the division prediction models and the integrated prediction model recorded in the learning result storage section 24.SELECTED DRAWING: Figure 1

Description

The present invention relates to a judgment prediction system, a judgment prediction method, and a judgment prediction program that support judgment of a disease.

In recent years, a technique for determining an image using a computer has been studied. Recently, in the medical field as well, support for diagnosis using images is being promoted, and an information processing device that presents a region that contributes to the judgment together with the judgment result regarding the diagnosis of a disease is also being studied (for example, Patent Document 1). See.). The information processing apparatus described in this document inputs an endoscopic image into a first model that outputs a diagnostic standard prediction regarding a diagnostic standard of a disease, and outputs a diagnostic standard prediction. In this case, the region that has influenced the diagnostic criterion prediction is extracted, and the diagnostic criterion prediction, the index indicating the extracted region, and the diagnostic prediction regarding the state of the disease are output in association with each other.

Japanese Unexamined Patent Publication No. 2020-89712

However, depending on the image, it may be difficult to detect the disease with high accuracy. For example, traditionally, an anterior chest X-ray image is determined by an image interpreter to determine the possibility of illness. Here, the X-ray image is a two-dimensional image having less information than the three-dimensional CT image, and it may be difficult to determine by the image processing technique.

The judgment prediction system that solves the above problems includes a learning result storage unit that records a division prediction model and an integrated prediction model, and a control unit that evaluates an image. Then, the control unit generates a plurality of divided images obtained by dividing the original image, and generates first teacher information recorded in association with the determination result in the original image with respect to the plurality of divided images, and the first teacher information is generated. Using the teacher information, a split prediction model that predicts the determination result is generated from the split image, recorded in the learning result storage unit, and each divided image of the first teacher information is input to the split prediction model. The second teacher information is generated by summarizing the feature information generated in the above process based on the position of the divided image, and the second teacher information is used to generate an integrated prediction model for predicting the determination result. When the image to be evaluated is acquired by recording in the learning result storage unit, the determination result is predicted by using the divided prediction model and the integrated prediction model recorded in the learning result storage unit.

According to the present invention, it is possible to support the determination of a disease.

Explanatory drawing of the judgment prediction system of embodiment. An explanatory diagram of a hardware configuration of an embodiment. Explanatory drawing of image segmentation of embodiment. The explanatory diagram of the processing procedure at the time of learning of an embodiment. Explanatory drawing of the division learning processing procedure of an embodiment. Explanatory drawing of the division prediction model of an embodiment. An explanatory diagram of the integrated learning processing procedure of the embodiment. An explanatory diagram of creating a heat map of an embodiment. Explanatory drawing of the heat map of an embodiment. Explanatory drawing of the integrated prediction model of an embodiment. Explanatory drawing of the processing procedure at the time of prediction of an embodiment. Explanatory drawing of the processing procedure of another embodiment. Explanatory drawing of the processing procedure of another embodiment. Explanatory drawing of the processing procedure of another embodiment.

An embodiment of the judgment prediction system, the judgment prediction method, and the judgment prediction program will be described with reference to FIGS. 1 to 11. In this embodiment, learning is performed using teacher data in which the examination results are associated with the X-ray image (organ image) of the lung cancer examination, and a prediction model for predicting the possibility of lung cancer is generated. Then, when a new X-ray image is acquired in the lung cancer screening, the diagnosis result is predicted using the prediction model.
As shown in FIG. 1, a user terminal 10 and a support server 20 connected via a network are used.

(Hardware configuration example)
FIG. 2 is a hardware configuration example of the information processing apparatus H10 that functions as a user terminal 10, a support server 20, and the like.

The information processing device H10 includes a communication device H11, an input device H12, a display device H13, a storage device H14, and a processor H15. Note that this hardware configuration is an example, and may have other hardware.

The communication device H11 is an interface that establishes a communication path with another device and executes data transmission / reception, such as a network interface card or a wireless interface.

The input device H12 is a device that receives input from a user or the like, and is, for example, a mouse, a keyboard, or the like. The display device H13 is a display, a touch panel, or the like that displays various information.

The storage device H14 is a storage device (for example, a teacher information storage unit 23 and a learning result storage unit 24, which will be described later) for storing data and various programs for executing various functions of the user terminal 10 and the support server 20. An example of the storage device H14 is a ROM, RAM, hard disk, or the like.

The processor H15 controls each process (for example, a process in the control unit 21 described later) in the user terminal 10 and the support server 20 by using the programs and data stored in the storage device H14. Examples of the processor H15 include a CPU, an MPU, and the like. The processor H15 expands a program stored in a ROM or the like into a RAM and executes various processes corresponding to various processes. For example, when the application program of the user terminal 10 and the support server 20 is started, the processor H15 operates a process for executing each process described later.

The processor H15 is not limited to the one that performs software processing for all the processing executed by itself. For example, the processor H15 may include a dedicated hardware circuit (for example, an integrated circuit for a specific application: ASIC) that performs hardware processing for at least a part of the processing executed by the processor H15. That is, the processor H15 is (1) one or more processors that operate according to a computer program (software), (2) one or more dedicated hardware circuits that execute at least a part of various processes, or ( 3) It can be configured as a circuitry including a combination thereof. The processor includes a CPU and a memory such as a RAM and a ROM, and the memory stores a program code or a command configured to cause the CPU to execute a process. Memory or computer readable media includes any available medium accessible by a general purpose or dedicated computer.

(Functions of each information processing device)
The user terminal 10 in FIG. 1 is a computer terminal used by a person in charge of lung cancer screening.
The support server 20 is a computer system for supporting lung cancer screening. The support server 20 includes a control unit 21, an image information storage unit 22, a teacher information storage unit 23, a learning result storage unit 24, and an evaluation target storage unit 25.

The control unit 21 performs a process described later (a process including an acquisition stage, a learning stage, a prediction stage, and the like). By executing the determination prediction program for this purpose, the control unit 21 functions as an acquisition unit 211, a learning unit 212, a prediction unit 213, and the like.

The acquisition unit 211 executes a process of acquiring an X-ray image to be learned or evaluated. The acquisition unit 211 divides the roentgen image including the entire lung field (the entire lung field image as the original image) into four lung field region images (divided lung field images) from the upper left to the lower right by pattern recognition. In this pattern recognition, the anatomical feature points of the lung field region are detected from the X-ray image, and the lung fields are divided according to the lung field region based on the feature points.

The learning unit 212 executes a process of generating a prediction model for predicting the determination result (negative probability and positive probability) of lung cancer screening using the teacher data.
The prediction unit 213 executes a process of predicting the determination result of interpretation of the X-ray image to be evaluated by using the prediction model.

An X-ray image taken in a lung cancer screening is recorded in the image information storage unit 22. The X-ray image is associated with the examiner's identifier and the determination result.
The examiner identifier is an identifier for identifying the examiner of the lung cancer examination who acquired this X-ray image.

As the judgment result, the diagnostic information by the image interpreting doctor is recorded. This diagnostic information includes (a) the presence or absence of abnormal findings, (b) the position of abnormal shadows, and (c) the contents of the abnormal findings. Here, depending on the presence or absence of (a) abnormal findings, judgment results such as "no abnormal findings", "suspected lung cancer", "abnormal findings are observed, and a condition requiring treatment for a disease other than lung cancer may be considered" are obtained. included. In this embodiment, the case where the presence or absence (positive or negative) of "suspected lung cancer" is used will be described. Further, in the present embodiment, (b) when the position of the abnormal shadow is determined to be positive, the lung field region (upper right, upper left, lower right, lower left) corresponding to the position of the abnormal shadow is used. explain.

In the teacher information storage unit 23, teacher data used for learning processing for generating a prediction model for predicting a determination result (negative probability and positive probability) of lung cancer screening is recorded. In this embodiment, teacher data for split learning and teacher data for integrated learning are recorded.

The teacher data for division learning is used to generate a division prediction model for making a positive judgment for each lung field region (upper right lung field, upper left lung field, lower right lung field, lower left lung field).
The teacher data for integrated learning is used to generate an integrated prediction model that integrates and judges the prediction results of the divided prediction model for each lung field.

As shown in FIG. 3, the original image 500 taken by roentgen is divided into divided lung field images 510 (upper right lung field image 511, upper left lung field image 512, lower right lung field image 513, lower left lung field image 514). Further, depending on the position of the abnormal shadow, a positive flag is associated with the divided lung field image (here, the lower left lung field image 514). On the other hand, a negative flag is associated with other divided lung field images. Then, it is recorded in the teacher information storage unit 23 as the teacher data for divided learning.

In the learning result storage unit 24, a prediction model (a split prediction model for each lung field and an integrated prediction model) generated by machine learning using teacher data is recorded.
An evaluation target X-ray image and a prediction result are recorded in the evaluation target storage unit 25. The X-ray image to be evaluated is recorded when it is newly acquired from the user terminal 10, and the prediction result is recorded at the prediction time stage described later.

Next, in the system configured as described above, a processing procedure for supporting a determination in lung cancer screening will be described.

(Processing during learning)
The learning process will be described with reference to FIG.
First, the control unit 21 of the support server 20 executes the image acquisition process (step S101). Specifically, the acquisition unit 211 of the control unit 21 acquires an X-ray image of lung cancer screening from the user terminal 10. This X-ray image (whole lung field image) includes a examiner's identifier (anonymous) and a positive judgment result by an image interpreter. If it is determined to be positive, information for identifying the determined lung field is included. The acquisition unit 211 records the acquired whole lung field image in the image information storage unit 22 in association with the examiner identifier and the determination result.

Next, the control unit 21 of the support server 20 executes a division process for each image of the entire lung field recorded in the image information storage unit 22 (step S102). Specifically, the acquisition unit 211 of the control unit 21 obtains the entire lung field image by pattern recognition, depending on the lung field region, the upper left lung field image, the lower left lung field image, the upper right lung field image, and the lower right lung field. Divide into images. Then, the acquisition unit 211 temporarily stores the divided lung field image from the upper left to the lower right in the memory in association with the examiner identifier. In this case, in the whole lung field image, a positive flag is recorded in the divided lung field image according to the position of the abnormal shadow. In addition, a negative flag is recorded in the divided lung field images of other lung fields.
The above process is repeated until the whole lung field image is completed.

Next, the control unit 21 of the support server 20 executes a process of generating first teacher information for each lung field for each lung field region (step S103). Specifically, the acquisition unit 211 of the control unit 21 collects the divided lung field images temporarily stored in the memory for each lung field region and generates teacher data for division learning (first teacher information). Then, the acquisition unit 211 records the generated teacher data for division learning in the teacher information storage unit 23 in association with the division lung field from the upper left to the lower right.

Next, the control unit 21 of the support server 20 executes the division learning process for each division lung field from the upper left to the lower right (step S104). This process will be described later with reference to FIG.
Next, the control unit 21 of the support server 20 executes the integrated learning process (step S105). This process will be described later with reference to FIG. 7.

(Divided learning process)
The division learning process will be described with reference to FIG.
First, the control unit 21 of the support server 20 executes the acquisition process of the first teacher information (step S201). Specifically, the learning unit 212 of the control unit 21 acquires the divided learning teacher data recorded in the teacher information storage unit 23.

Next, the control unit 21 of the support server 20 executes a feature amount map image generation process in a plurality of layers (step S202). Specifically, the learning unit 212 of the control unit 21 uses "DenseNet" as the machine learning method. This "Dense Net" is composed of a network in which "Dense Block" and "Transition Layer" are alternately sandwiched between a convolution layer and a pooling layer. Using "DenseNet", a divided lung field image is input as an input layer, and a predictive model ("DenseNet" model) that outputs positive / negative in the output layer is generated. In this case, the "DenseNet" model in which the divided image is input generates a feature map image in an intermediate layer composed of a plurality of layers according to the depth. The feature map image is an image obtained by applying various filters to each layer of the divided lung field image. Further, in the "DenseNet" model, the representative value of the feature amount map image of each layer is calculated. Then, in the "DenseNet" model, a weighted value (variable value in the model) for each layer that outputs a positive / negative probability is calculated from this representative value.

Next, the control unit 21 of the support server 20 executes the acquisition process of the representative value for each layer of the feature amount map image (step S203). Specifically, the learning unit 212 of the control unit 21 acquires representative values for each feature amount map image for each layer of the “DenseNet” model. In the present embodiment, the maximum value of the image value in the feature amount map image is used as the representative value.

Next, the control unit 21 of the support server 20 executes a weighted value acquisition process for each layer that outputs a positive / negative probability by inputting a representative value for each layer (step S204). Specifically, the learning unit 212 of the control unit 21 inputs a representative value of the feature amount map image of each layer in the "DenseNet" model, and calculates a positive and negative probability (probability) by a weighted value (weighted). Information) is acquired. This weighted value weights the feature map image that affects the positive and negative probabilities.

As shown in FIG. 6, the learning unit 212 outputs a negative probability to the input of the representative values (vd1 to vd4) of the feature amount map image 520 of the divided lung field image 510, weighted values (W01 to W04). A weighted value (W11 to W14) that outputs a positive probability is calculated. In this case, each divided lung field image 510 is input to the "DenseNet" model to generate a feature amount map image of each layer, and a representative value of this feature amount map image is calculated. In addition, the positive and negative probabilities output by the "DenseNet" model are acquired. Next, the learning unit 212 acquires weighted values (W01 to W14) from the “DenseNet” model. In FIG. 6, four layers are assumed, but when the "DenseNet" model has n layers, the number of representative values is n and the number of weighted values is 2n.

Then, the learning unit 212 records in the learning result storage unit 24 a network (division prediction model) including the “DenseNet” model and the weighted value for the divided lung field identifier and the hierarchical identifier.

(Integrated learning process)
The integrated learning process will be described with reference to FIG. 7.
Here, the control unit 21 of the support server 20 executes a heat map creation process using the feature amount map image and the weighted value for each lung field image of each examiner (step S301). Specifically, the learning unit 212 of the control unit 21 inputs each divided lung field image into the "DenseNet" model for each examiner, and the feature amount map image of each layer generated, and the positive probability in this layer. Use the weighted value to calculate the heat map according to the lung field region.

Here, as shown in FIG. 8, a weighted image is generated by multiplying each feature amount map image 520 of each layer by a weighted value (W11 to W14) for calculating a positive probability. Further, the learning unit 212 generates the heat map 600 by superimposing the weighted images for each layer at the same pixel position.

FIG. 9 is an example of a heat map 600 in which the divided lung field images of each examiner are combined with the original image 500 of the entire lung field image. In this heat map 600, a plurality of feature amount maps are superimposed, but the feature area of the feature amount map image to which a high weighted value is given is emphasized and displayed.

Next, the control unit 21 of the support server 20 executes the registration process of the second teacher information (for integrated learning) (step S302). Specifically, the learning unit 212 of the control unit 21 generates teacher data for integrated learning (second teacher information) using the divided learning result. Here, the heat map for each lung field region and the determination result prepared for each examiner created in step S301 are used as teacher data for integrated learning. Then, the learning unit 212 records the heat map and the determination result for each lung field region associated with the examiner identifier in the teacher information storage unit 23 as integrated learning teacher data.
The above processing is repeated for all the examiners recorded in the image information storage unit 22.

Next, the control unit 21 of the support server 20 executes the process of creating the integrated prediction model (step S303). Specifically, the learning unit 212 of the control unit 21 generates an integrated prediction model that predicts the negative probability and the positive probability from the teacher data for integrated learning generated by using the result of the division learning by machine learning.

Here, as shown in FIG. 10, the feature amounts (integrated input feature amounts vi1 to vi4) for each lung field region of each feature amount map image 520 are calculated, and each integrated input feature amount is used as an input to obtain a negative probability and a positive value. An integrated prediction model that outputs probabilities is generated by machine learning. In the present embodiment, machine learning is performed using the heat map for each lung field region created in step S301 as the input and the judgment result of the image interpreter as the output as the integrated input feature quantities vi1 to vi4. In FIG. 10, since the lung field image is divided into four from the upper left to the lower right, the feature amount is four, but the number of feature amounts is determined according to the number of divisions of the lung field.
Then, the learning unit 212 records the generated integrated prediction model in the learning result storage unit 24.

(Processing at the time of prediction)
Next, the prediction processing will be described with reference to FIG.

First, the control unit 21 of the support server 20 executes the image acquisition process (step S401). Specifically, the acquisition unit 211 of the control unit 21 acquires a lung field image in a new lung cancer screening from the user terminal 10 and records it in the evaluation target storage unit 25.

Next, the control unit 21 of the support server 20 executes the division process in the same manner as in step S102 (step S402). This creates a split lung field image.
Next, the control unit 21 of the support server 20 executes a prediction process using the divided lung field image (step S403). Specifically, the prediction unit 213 of the control unit 21 inputs the created divided lung field image into the division prediction model recorded in the learning result storage unit 24, and predicts the division determination result. Here, in the division determination result, the negative probability and the positive probability are calculated for each divided lung field.

Next, the control unit 21 of the support server 20 executes a prediction process using the entire lung field image (step S404). Specifically, the prediction unit 213 of the control unit 21 integrates the feature quantities for each lung field region generated in the prediction process (step S403) using the divided lung field image recorded in the learning result storage unit 24. Input to the prediction model to calculate the integrated prediction result. In the present embodiment, the prediction unit 213 generates a heat map in which a map obtained by multiplying the feature amount map calculated in step S403 by a weighted value is superimposed. Then, the prediction unit 213 inputs the created heat map into the integrated prediction model recorded in the learning result storage unit 24, and predicts the integrated prediction result. As the integrated determination result, the negative probability and the positive probability are calculated.

Next, the control unit 21 of the support server 20 executes the output processing of the determination result (step S405). Specifically, the prediction unit 213 of the control unit 21 outputs the determination result screen to the user terminal 10. The determination result screen includes the determination results (division determination result, integration determination result) predicted in steps S403 and S404, and the heat map.

According to this embodiment, the following effects can be obtained.
(1) In the present embodiment, the control unit 21 of the support server 20 executes the image acquisition process (step S101). As a result, the determination result of the image interpreting doctor can be obtained. Since this image also includes noise that affects the determination of abnormalities on the image, such as bones and organs, machine learning can be performed in consideration of the noise.

(2) In the present embodiment, the control unit 21 of the support server 20 performs division processing (step S102) for each lung field image recorded in the image information storage unit 22 and generates first teacher information for each lung field. The processing (step S103) and the division learning process (step S104) are executed for each divided lung field. This makes it possible to determine the presence or absence of local abnormalities in the divided lung field region. For example, lung cancer has different appearances and points to be noted for each area, and by performing specialized learning for each area, it is possible to realize a judgment close to that of an image interpreter. Furthermore, it is possible to improve the prediction accuracy by performing machine learning after dividing the lung field by aligning the anatomical positions.

(3) In the present embodiment, the control unit 21 of the support server 20 executes the feature amount map image generation process in a plurality of layers using the “Dense Net” (step S202). Here, by alternately sandwiching the "Dense Block" and the "Transition Layer", it is possible to deal with the disappearance of the gradient when the layer is deepened. Then, it is possible to generate a feature map image with various filters applied to each layer.

(4) In the present embodiment, the control unit 21 of the support server 20 creates a heat map using the feature map image and the weighted value (step S301), and registers the second teacher information (for integrated learning). (Step S302), the process of creating the integrated prediction model (step S303) is executed. The judgment in the divided area can be adjusted throughout. For example, when an abnormality is detected in an organ that is symmetrically arranged on the human body, a predictive model that adjusts the judgment can be generated according to the left-right arrangement of the abnormal region. Specifically, when a part suspected to be abnormal is detected on the upper left lung side, the difference between the left and right is confirmed by comparing with the upper right lung side, and the negative probability and the positive probability are calculated based on the presence or absence of a similar part on the right side. can do. Here, if a similar part is detected on the right side, it may not be abnormal.

(5) In the present embodiment, the control unit 21 of the support server 20 has an image acquisition process (step S401), a prediction process using a divided lung field image (step S403), and a prediction process using an entire lung field image (step S403). Step S404) is executed. As a result, it is possible to improve the quality of diagnosis, reduce the burden on the interpreting doctor, and shorten the working hours of the doctor. As a result, the burden on the doctor can be shifted from the primary interpretation to the secondary interpretation, and the quality of the comparative interpretation can be improved. In addition, by systematizing the interpretation of lung cancer screening, it is possible to contribute to the elimination of regional disparities due to the uneven distribution of highly specialized interpretation doctors.

(6) In the present embodiment, the control unit 21 of the support server 20 executes the output processing of the determination result (step S405). In this case, by outputting a heat map or the like, it is possible to visualize an object that can be a feature of discrimination such as an abnormality.

This embodiment can be modified and implemented as follows. The present embodiment and the following modified examples can be implemented in combination with each other within a technically consistent range.
-In the above embodiment, it is applied to the detection of lung cancer in the X-ray image of lung cancer screening. The applicable diseases are not limited to lung cancer. It can also be applied to organs other than the lungs (for example, the brain and kidneys having a shape close to symmetry).
-In the above embodiment, the positive / negative probability of lung cancer is predicted. The prediction target is not limited to two types, positive and negative. For example, in the presence or absence of (a) abnormal findings in the diagnostic information, judgments such as "no abnormal findings", "suspected lung cancer", "abnormal findings are found, and a condition requiring treatment for a disease other than lung cancer is considered". The results may be predicted individually. In this case, the first and second teacher information including the respective judgment results are generated, and the probability of "no abnormal findings", the probability of "suspected lung cancer", and the probability of "suspicion of lung cancer" are obtained by the split prediction model and the integrated prediction model. Predict the probability of "abnormal findings and possible conditions requiring treatment for diseases other than lung cancer."

-In the above embodiment, "Dense Net" is used as a machine learning method. The machine learning method is not limited to "Dense Net", and for example, "Residual Network (ResNet)" may be used.
-In the above embodiment, the control unit 21 of the support server 20 executes the process of creating the integrated prediction model (step S303). In this case, a heat map for each lung field region is used as the integrated input feature amount. The integrated input features are not limited to the heat map itself.

For example, as shown in FIG. 12, the heat map 600 for each lung field region and the positive and negative probabilities calculated by the split prediction model may be used as the integrated input feature amount.
Further, as shown in FIG. 13, an image group 530 obtained by multiplying each feature amount map image 520 for each lung field region by a weighted value (W11 to W14) may be used as an integrated input feature amount. In this case, it is not necessary to use all the feature amount map images, and the feature amount map image whose weight value is equal to or larger than the reference value may be specified and used.

Further, as shown in FIG. 14, an image group 530 obtained by multiplying each feature amount map image 520 for each lung field region by a weighted value (W11 to W14), and a positive probability and a negative probability are used as integrated input feature amounts. You may. In this case as well, it is not necessary to use all the feature amount map images, and the feature amount map image whose weight value is equal to or larger than the reference value may be specified and used.

-In the above embodiment, the acquisition unit 211 divides the X-ray image (the entire lung field image) into four lung field images from the upper left to the lower right by pattern recognition. In this pattern recognition, the anatomical feature points of the lung field region are detected from the X-ray image, and the lung field region is divided based on the feature points. The division method is not limited to the case of using feature points. For example, machine learning may be used to recognize the shape of an organ and divide each lung field region.

-In the above embodiment, the control unit 21 of the support server 20 executes the division process (step S102). Here, the acquisition unit 211 of the control unit 21 divides the entire lung field image into an upper left lung field image, a lower left lung field image, an upper right lung field image, and a lower right lung field image by pattern recognition. Here, the number of divisions is not limited to four. Further, the shape of the lung field may be deformed so as to be aligned with the average shape, and the lung field region may be divided from the aligned image.

10 ... user terminal, 20 ... support server, 21 ... control unit, 211 ... acquisition unit, 212 ... learning unit, 213 ... prediction unit, 22 ... image information storage unit, 23 ... teacher information storage unit, 24 ... learning result storage unit , 25 ... Evaluation target storage unit.

Claims (8)

A learning result storage unit that records split prediction models and integrated prediction models,
It is a judgment prediction system that has a control unit for evaluating images and predicts judgment results.
The control unit
Generate multiple split images by splitting the original image,
The first teacher information recorded by associating the determination result in the original image with the plurality of divided images is generated.
Using the first teacher information, a division prediction model that predicts the determination result from the division image is generated, recorded in the learning result storage unit, and recorded.
The second teacher information is generated by inputting each divided image of the first teacher information into the divided prediction model and summarizing the feature information generated based on the position of the divided image.
Using the second teacher information, an integrated prediction model for predicting the determination result is generated, recorded in the learning result storage unit, and recorded.
A judgment prediction system characterized in that when an evaluation target image is acquired, the judgment result is predicted using the division prediction model and the integrated prediction model recorded in the learning result storage unit. The determination prediction system according to claim 1, wherein the control unit uses the feature information generated in the intermediate layer from the input layer to the output layer of the division prediction model in the generation of the second teacher information. .. The determination prediction system according to claim 2, wherein the control unit acquires a plurality of feature amount map images generated in the intermediate layer. The control unit
Weighted information that links the feature amount map image to the determination result is calculated.
The determination prediction system according to claim 3, wherein the second teacher information is generated by using the weighted information. The determination prediction system according to claim 4, wherein the control unit generates a heat map using the feature amount map image and the weighted information. The control unit
As the original image and the evaluation target image, an organ image obtained by photographing an organ is used.
The determination prediction system according to any one of claims 1 to 5, wherein the diagnosis result relating to the organ image is used as the determination result. A learning result storage unit that records split prediction models and integrated prediction models,
It is a method of predicting a judgment result by using a judgment prediction system equipped with a control unit for evaluating an image.
The control unit
Generate multiple split images by splitting the original image,
The first teacher information recorded by associating the determination result in the original image with the plurality of divided images is generated.
Using the first teacher information, a division prediction model that predicts the determination result from the division image is generated, recorded in the learning result storage unit, and recorded.
The second teacher information is generated by inputting each divided image of the first teacher information into the divided prediction model and summarizing the feature information generated based on the position of the divided image.
Using the second teacher information, an integrated prediction model for predicting the determination result is generated, recorded in the learning result storage unit, and recorded.
A determination prediction method characterized in that when an evaluation target image is acquired, the determination result is predicted using the division prediction model and the integrated prediction model recorded in the learning result storage unit. It is a judgment prediction program for making a computer function so as to output a judgment result based on an image.
It consists of a split predictive model and an integrated predictive model that is combined so that the output from the split predictive model is input.
Generate multiple split images by splitting the original image,
The first teacher information recorded by associating the determination result in the original image with the plurality of divided images is generated.
A division prediction model generated so as to predict the determination result from the division image using the first teacher information, and
The second teacher information is generated by inputting each divided image of the first teacher information into the divided prediction model and summarizing the feature information generated based on the position of the divided image.
Including an integrated prediction model generated to predict the determination result using the second teacher information.
A judgment prediction program for functioning as a means for predicting the judgment result by using the division prediction model and the integrated prediction model when the evaluation target image is acquired.

Images