JP7252035B2 - disease estimation system - Google Patents

Description

The present disclosure relates to disease estimation systems.

Conventionally, a diagnosis support device including a neural network is known (Patent Document 1).

JP-A-2008-36068

Such diagnostic support devices are required to have improved reliability.

A disease estimation system according to one embodiment of the present invention is a system capable of estimating a disease from an image. A disease estimation system comprises an input and an approximator. The input unit is capable of inputting input data. The approximator is subjected to learning processing using learning data including data of the same type as the input data and teacher data including diagnostic results related to the learning data. The approximator can retrieve the disease estimation result estimated from the input data input to the input unit and reference data similar to the estimation result, and output the retrieval result.

The user can confirm the reliability of the estimation result. Therefore, it is possible to improve the reliability of the estimation result.

It is a figure which shows an example of a structure of a disease prediction system. It is a figure which shows an example of a structure of one part of a disease prediction system. It is a figure which shows an example of a structure of one part of a disease prediction system. It is a figure which shows an example of a structure of one part of a disease prediction system. It is a figure which shows an example of a structure of one part of a disease prediction system. FIG. 10 illustrates an example of the operation of part of the disease prediction system; FIG. 10 illustrates an example of the operation of part of the disease prediction system;

The disease estimation system 1 of the present disclosure uses input data I to AI (Artificial Intelligence)
can be used to estimate disease. The input data I may be image data, for example. A disease estimated by the present invention may be a disease that can be diagnosed by a doctor based on image data. For example, cervical cancer, skin disease, and so on.

FIG. 1 conceptually shows the configuration of a disease estimation system 1 of the present disclosure.

A disease estimation system 1 of the present disclosure has a terminal device 2 and an estimation device 3 . The terminal device 2 acquires patient input data I used for disease diagnosis. The estimating device 3 can estimate the type and site of disease based on the data acquired by the terminal device 2 .

The terminal device 2 can acquire image data. Input data I acquired by the terminal device 2 is input to the estimation device 3 . In the case of cervical cancer, the terminal device 2 may be a device capable of taking colposcopy images, for example. In the case of skin diseases, any device capable of taking a dermoscopy image may be used. The input data I may be, for example, a colposcopy image in the case of cervical cancer, and a dermoscopy image in the case of a skin disease.

The terminal device 2 has a communication unit 21 and may transfer the input data I to the estimation device 3 after acquiring the input data I.

Note that the terminal device 2 does not have to transfer the input data I directly to the estimation device 3 . In this case, for example, the input data I acquired by the terminal device 2 may be stored in a storage medium, and the input data I may be input to the estimation device 3 via the storage medium.

FIG. 2 conceptually shows the configuration of the estimation device 3 according to this embodiment.

The estimating device 3 can estimate the patient's disease from the input data I input to the estimating device 3 . For example, when the disease estimation system 1 estimates cervical cancer, the estimation device 3 estimates the patient's cervical cancer type and lesion site from the input data I acquired by the terminal device 2, and estimates the estimated result O can be output.

The estimation device 3 has an input unit 31 , an approximator 32 and an output unit 33 . The input unit 31 receives input data I from the terminal device 2 . The approximator 32 can perform predetermined calculations from the input data I to classify the disease and estimate the lesion site. The output unit 33 can output the estimation result O estimated by the approximator 32 .

The input data I is input to the input unit 31 as described above. The input unit 31 of the present disclosure may be any communication device that can be connected to the communication 21 of the terminal device 2 via a network. Also, the input unit 31 may be a port such as a USB port capable of receiving the input data I via a storage medium. Also, the input unit 31 may include an input device capable of inputting information other than the input data I. FIG. The input device may be, for example, a keyboard, touch panel, mouse, USB port, or the like.

The approximator 32 can predict the patient's disease from the input data I input to the input unit 31 . The approximator 32 is a so-called AI (Artificial Intelligence). Specifically, the approximator 32 has hardware composed of a plurality of electronic components and circuits for functioning as AI.

As noted above, approximator 32 includes a number of electronic components and circuits. In other words, part of the approximator 32 is composed of a plurality of electronic components and circuits. The plurality of electronic components may be, for example, active elements such as transistors or diodes, or passive elements such as capacitors, and may be formed by conventionally known methods.

In addition, the learned approximator 32 performs calculations according to the learned model that has already undergone learning processing when estimating a disease. That is, the disease estimating system 1 acquires in advance the parameters and the like necessary for the calculation of the approximator 32 through the approximator 32 using the teacher data. As a result, the approximator 32 can calculate the estimation result O from the input data I. The learning data or teacher data may be data corresponding to the input data I to be input to the estimating device 3 and the estimation result O to be output from the estimating device 3 .

FIG. 3 conceptually illustrates the processing performed by the approximator 32 of the present disclosure.

The approximator 32 according to this embodiment is, for example, a CNN. The approximator 32 includes an input layer 3a, a hidden layer 3b, and an output layer 3c as functional units. The hidden layer 3b is also called an intermediate layer, for example. Each of the input layer 3a, the hidden layer 3b, and the output layer 3c has a plurality of nodes 3d and a plurality of edges 3e connecting the nodes 3d, forming a so-called neural network as a whole. Each functional unit can be implemented by a plurality of electronic components, circuits, etc. included in the approximator 32, as described above.

By applying the trained model to the approximator 32, the connection relationships between the plurality of nodes 3d and the plurality of edges 3e are optimized for disease estimation. That is, the trained approximator 32 can perform operations using parameters suitable for disease estimation.

The input layer 3a is a layer to which input data I is input. Data constituting the input data I is input to each node 3d of the input layer 3a. For example, when the input data I is image data, a numerical value indicating the color or shade of each pixel of the image data is input to each node 3d.

4 and 5 conceptually illustrate the configuration of part of the hidden layer 3b of the present disclosure.

The hidden layer 3b can perform operations based on the information of the image data input to the input layer 3a. The hidden layer 3b has a convolutional layer 3b1, a pooling layer 3b2, and a fully connected layer 3b3. The hidden layer 3b can extract features from the data input to the input layer 3a by the convolutional layer 3b1, the pooling layer 3b2, and the fully connected layer 3b3.

The convolution layer 3b1 can convolve so-called data by filtering the data input to each node 3d of the immediately preceding layer. In other words, the data can be reduced while maintaining the characteristics of the input data I. Specifically, the numerical value input to each node 3d of the previous layer can be calculated based on the filter coefficients obtained in the learning process, and the data can be reduced while maintaining the feature amount. Note that each of the plurality of convolution layers 3b1 performs arithmetic processing based on the respective filter coefficients.

The pooling layer 3b2 is positioned after the convolutional layer 3b1. The pooling layer 3b2 can reduce the dimensionality of the data while maintaining the characteristics of the data of the convolutional layer 3b1 based on the numerical value of each node 3d calculated in the immediately preceding convolutional layer 3b1. For example, the pooling layer 3b2 may extract the largest value within an arbitrary range from the numerical values of the nodes 3d of the immediately preceding convolutional layer 3b1.

One convolutional layer 3b1 and one pooling layer 3b2 constitute one set. Also, the hidden layer 3b may have a plurality of convolution layers 3b1 and a plurality of pooling layers 3b2. In this case, the plurality of convolutional layers 3b1 and the plurality of pooling layers 3b2 are alternately arranged to form a plurality of sets of the convolutional layers 3b1 and the pooling layers 3b2 in the hidden layer 3b. , the features of the input data I can be narrowed down gradually. The number of the plurality of convolutional layers 3b1 and the plurality of pooling layers 3b2 may be any number that allows the approximator 32 to function with required accuracy.

The fully connected layer 3b3 can extract the features of the input data I from the data input to each node 3d of the immediately preceding pooling layer 3b2. Specifically, it is possible to perform calculations on the numerical values input to each node 3d of the previous layer based on the weighting coefficients obtained in the learning process, and to extract feature amounts.

The output layer 3c can output the estimation result O based on the numerical value of each node 3d calculated by the fully connected layer 3b3. Specifically, based on the numerical pattern of each node 3d of the fully connected layer 3b3, for example, the degree of possibility of disease can be output.

The output unit 33 can display the estimation result O estimated in the output layer 3c. The output unit 33 is, for example, a liquid crystal display or an organic EL display. The output unit 33 can display various information such as characters, symbols, and graphics. The output unit 33 can display numbers or images, for example.

Note that the estimation result O is output as, for example, a character string.

The estimation device 3 further has a control unit 34 . The control unit 34 can centrally manage the operation of the estimating device 3 by controlling other components of the estimating device 3 . That is, the control unit 34 can control the calculation of the approximator 32 and the like. The control unit 34 has a plurality of electronic components and circuits, and constitutes, for example, a CPU (Central Processing Unit). The plurality of electronic components may be, for example, active elements such as transistors or diodes, or passive elements such as capacitors, and may be formed by conventionally known methods.

In addition, the estimation device 3 further has a storage unit 35 . The storage unit 35 can store programs and the like for controlling the estimation device 3 . The storage unit 35 may be, for example, ROM (Read-Only Memory) or HDD (Hard-Disc-Drive). Note that the storage unit 35 may also store the input data I, the calculation result of the approximator for the learning data, the teacher data, the learned model (learned parameters), and the like.

<Examples of input data, learning data, and teacher data>
The input data I may be data such as patient image data used for diagnosis of the disease. The input data I may be colposcopy images, for example. From colposcopy images, for example, the classification of cervical cancer can be deduced. Also, if the input data I is, for example, a dermoscopy image, a skin disease can be estimated.

In the input data I, the part to be diagnosed is photographed. For example, when estimating cervical cancer disease, the input data I may be the image of the cervix. Moreover, when estimating a skin disease, it is sufficient that the skin including the lesion site is imaged.

The learning data has the same kind of data as the first input data I1. For example, if the input data I1 is a colposcopy image, the learning data may also have a colposcopy image.

The teacher data has diagnosis results of diseases corresponding to the learning data. For example, when estimating cervical cancer, the teacher data are diagnosis results corresponding to each of a plurality of learning image data. It is sufficient that the diagnosis result is evaluated at approximately the same time as when the learning image data was captured.

For example, in the case of cervical cancer, the teacher data may be a doctor's diagnosis result.

<Neural network learning example>
The approximator 32 is optimized by machine learning using learning data and teacher data so that the estimation result O can be calculated from the input data I. That is, the approximator 32 before the learning process calculates a pseudo estimation result from the learning data based on the unlearned mathematical model, and adjusts the difference between the pseudo estimation result and the teacher data. is machine-learned by adjusting the parameters of As a result, the approximator 32 can perform calculations on the input data I based on the learned model and output the estimation result O. FIG.

As a parameter adjustment method, for example, an error backpropagation method is adopted. The parameters may be, for example, filter coefficients applied to calculations performed in a plurality of convolution layers 3b1, weighting coefficients applied to calculations performed in fully connected layer 3b3, and the like.

FIG. 6 shows the operation order of the approximator 32 according to this embodiment.

In the disease estimating system 1, the disease can be estimated by performing calculations in the approximator 32 based on the input data I. FIG. Further, in the disease estimation system 1 according to the present invention, the approximator 32 searches for the estimation result O of the estimated disease and the reference data R similar to the estimation result O as the basis of the estimation result O, and outputs the search result. can do.

Here, for example, a conventional diagnosis support device uses a neural network to support the determination of a disease. I couldn't do it. In other words, it is difficult for humans to understand the process of deriving results in a diagnostic support device using a neural network (so-called black box). is required. In addition, when data not subject to diagnosis is input, it is also required to detect and notify the user of it.

On the other hand, in the disease estimation system 1 according to the present invention, as described above, the approximator 32 retrieves the estimated disease estimation result O and reference data R similar thereto, and the output unit O outputs the estimation result O The search results can be output together with That is, the disease estimation system 1 according to the present invention can output the reference data R, for example, when there is similar reference data R, and on the other hand, when there is no similar reference data R, the reference data It becomes possible to output that there is no R. Therefore, the reliability of the estimation result O of the disease estimation system 1 can be confirmed.

FIG. 7 more specifically shows the calculation order after calculating the estimation result O of the approximator 32 according to the present embodiment.

In the disease estimation system 1 according to this embodiment, the approximator 32 may search the reference data R from the learning data. Specifically, the approximator 32 compares the numerical value of each node 3d of the fully connected layer 3b3 at the time of calculation of the input data I with the numerical value of each node 3d of the fully connected layer 3b3 of the learning data at the time of learning. By doing so, learning data similar to the estimation result O may be retrieved.

That is, in this case, in the disease estimation system 1 according to the present embodiment, after the learning process is completed, the approximator 32 performs arithmetic processing again using the learning data, and the numerical value and teacher data calculated in the process are Information may be stored in the storage unit 35 . As a result, the teacher data having a high matching rate between the numerical value calculated by the approximator 32 in the process of arithmetic processing for the data input to the input unit 31 and the numerical value stored in the storage unit 35 is combined with the estimation result O. can be shown to the user. The user can visually compare the displayed teacher data and the input data I to confirm the reliability of the estimation result.

It should be noted that whether the nodes are similar or not can be determined by comparing the numerical values of the respective nodes 3d. for example,
When each node 3d is compared, if there are 80% or more of the plurality of nodes 3d to be compared, the node 3d having a matching rate of ±20% or less for each numerical value of each node 3d is judged to be matched. You may

Also, the approximator 32 compares the numerical values of the nodes 3d of the fully connected layer 3b3 and the pooling layers 3b2 at the time of calculation of the input data I with the numerical values of the learning data at the time of learning, thereby approximating the estimation result O. You may search for learning data that

The approximator 32 may output learning data as the reference data R, and may also output teacher data related to the output learning data.

The approximator 32 may output learning data and teacher data as reference data R, and may mainly use personal data accompanying the learning data. The personal data may be age, sex, or the like.

Also, the approximator 32 may output learning data and teacher data as reference data R, as well as diagnostic information of the teacher data. Note that the diagnostic information may be information such as diagnostic person and diagnostic equipment.

The approximator 32 may output the same type of data as the input data I as the estimation result O. That is, for example, when the input data I is image data, the image data may be output as the estimation result O. In this case, in order to output the image data as the estimation result O, the approximator 32 needs to be learned. In this case, for example, a ConvLSTM network that combines a CNN (Convolutional Neural Network) and an LSTM (Long short-term memory) may be applied as the approximator 32 .

The approximator 32 may output the image data of the learning data as the search result O'. As a result, it becomes easier to compare the input data I and the learning data.

The approximator 32 may emphasize the difference between the estimation result O and the search result O' when the estimation result O and the search result O' are output as image data.

It should be noted that the present invention is not limited to the examples of the embodiments described above, and includes various modifications as long as there is no contradiction in the content thereof. Further, each embodiment according to the present invention can be appropriately combined.

For example, in the above example, when a neural network is used as the approximator 32, an example in which CNN is applied has been described, but the present invention is not limited to this. For example, the approximator 32 may use a ConvLSTM network combining CNN (Convolutional Neural Network) and LSTM (Long short-term memory), RNN (Recurrent Neural Network), or GAN (Generative Adversarial Network). The approximator 32 may also combine multiple neural networks. Specifically, it may be a composite neural network in which a ConvLSTM network and a convolutional neural network are combined.

1 disease estimation system 2 terminal device 21 communication unit 3 estimation device 31 input unit 32 approximator 33 output unit 34 control unit 35 storage unit 3a input layer 3b hidden layer 3b1 convolution layer 3b2 pooling layer 3b3 fully connected layer 3c output layer I input data O Estimation result O' Search result O'
R Reference data

Claims (6)

A disease estimation system capable of estimating a disease from an image,
an input unit capable of inputting input data;
an approximator that has undergone learning processing using learning data including data of the same type as the input data and teacher data including diagnostic results related to the learning data;
The approximator outputs an estimation result of the disease estimated from the input data input to the input unit, and reference data similar to the estimation result, which is a search result searched as a basis for the estimation result. , disease estimation system. The disease estimation system according to claim 1,
The disease estimation system, wherein the approximator searches the learning data for the reference data. The disease estimation system according to claim 2,
The disease estimation system, wherein the approximator outputs the learning data and the teacher data related to the learning data to be output. The disease estimation system according to claim 2 or 3,
The disease estimation system, wherein the approximator outputs the estimation result as image data of the same type as the input data. The disease estimation system according to claim 4,
The disease estimation system, wherein the approximator outputs the image data of the learning data as the search result. The disease estimation system according to claim 5,
The disease estimation system, wherein the approximator emphasizes a difference between the estimation result and the search result when the estimation result and the search result are output as the image data.

Images