JP7723922B2 - X-ray age estimation learning device, X-ray age estimation device, image age estimation learning device, image age estimation device, X-ray imaging system, evaluation device, X-ray age estimation learning method, X-ray age estimation method and program - Google Patents

Description

The present invention relates to a X-ray age estimation learning device, a X-ray age estimation device, an image age estimation learning device, an image age estimation device, an X-ray imaging system, an evaluation device, a X-ray age estimation learning method, a X-ray age estimation method, and a program.

Conventionally, there has been technology for measuring a living body's vascular age from a subject's pulse wave (see, for example, Patent Document 1).

International Publication No. 2010/134233

Vascular age is currently widely used because it is intuitively easy to understand. However, vascular age often does not show significant changes due to treatment. For this reason, vascular age has rarely been used as information to help doctors change clinical treatment strategies.

In light of the above circumstances, the present invention aims to provide a X-ray age estimation learning device, X-ray age estimation device, image age estimation learning device, image age estimation device, X-ray imaging system, evaluation device, X-ray age estimation learning method, X-ray age estimation method, and program that can provide information useful for assessing health conditions and determining treatment plans.

One aspect of the present invention is a radiographic age estimation learning device that includes a learning unit that uses data on a subject's radiographic image and information on the subject's age at the time the radiographic image was taken to learn a radiographic age estimation model that indicates the correspondence between feature values obtained from the radiographic image data and the subject's age.

One aspect of the present invention is the above-mentioned X-ray age estimation learning device, in which the X-ray image is a chest X-ray image.

One aspect of the present invention is a radiographic age estimation device that includes an estimation unit that inputs feature amounts obtained from data on an estimation target radiographic image, which is an radiographic image of the subject whose age is to be estimated, into the radiographic age estimation model that is trained using data on a radiographic image of a subject and information on the subject's age at the time the radiographic image was taken, and that indicates the correspondence between feature amounts obtained from the radiographic image data and the subject's age, and obtains an estimated age of the subject whose age is to be estimated.

One aspect of the present invention is the above-mentioned X-ray age estimation device, further comprising a contribution information acquisition unit that acquires contribution information, which is information about portions of the X-ray image to be estimated that have feature values whose degree of contribution to age estimation of the subject to be age estimation satisfies predetermined conditions.

One aspect of the present invention is the above-mentioned X-ray age estimation device, further comprising an output unit that outputs one or more of information on the age estimated by the estimation unit, information on the difference between the actual age of the subject whose age is to be estimated and the age estimated by the estimation unit, and the contribution information.

One aspect of the present invention is the above-mentioned X-ray age estimation device, wherein the X-ray image and the estimation target X-ray image are X-ray images of the chest.

One aspect of the present invention is an image-based age estimation learning device that includes a learning unit that uses data on transmitted images, which are images obtained by transmitting images of the subject's body, and information on the subject's age at the time the transmitted images were taken, to learn an age estimation model that indicates the correspondence between feature values obtained from the transmitted image data and the subject's age.

One aspect of the present invention is an image-based age estimation device that includes an age estimation model trained using data on transmission images, which are images obtained by transmitting a transmission image of the subject's body, and information on the subject's age at the time the transmission image was captured, and an estimation unit that inputs feature amounts obtained from data on an estimation target image, which is a transmission image of the subject whose age is to be estimated, into the age estimation model that indicates a correspondence between feature amounts obtained from the transmission image data and the subject's age, and obtains an estimated age of the subject whose age is to be estimated.

One aspect of the present invention is an X-ray imaging system comprising the above-described X-ray age estimation device and an X-ray imaging device that performs X-ray imaging of a subject and inputs data of the X-ray image of the subject obtained by the X-ray imaging into the X-ray age estimation device as data of the X-ray image to be estimated.

One aspect of the present invention is an evaluation device that includes an evaluation unit that evaluates the subject's disease or pathological condition based on the results of a predetermined calculation using the result of the subject's age estimation by the above-mentioned X-ray age estimation device and information about the subject's biology or lifestyle or a numerical value representing that information.

One aspect of the present invention is a Roentgen age estimation learning method that includes a learning step of using data on a subject's Roentgen image and information on the subject's age at the time the Roentgen image was taken to learn a Roentgen age estimation model that indicates the correspondence between feature values obtained from the Roentgen image data and the subject's age.

One aspect of the present invention is a radiographic age estimation method that includes an estimation step of inputting feature values obtained from data on an estimation target radiographic image, which is an radiographic image of the subject whose age is to be estimated, into a radiographic age estimation model trained using data on the radiographic image of a subject and information on the subject's age at the time the radiographic image was taken, the radiographic age estimation model indicating the correspondence between feature values obtained from the radiographic image data and the subject's age, and obtaining an estimated age of the subject whose age is to be estimated.

One aspect of the present invention is a program for causing a computer to operate as any of the above-described X-ray age estimation learning devices.

One aspect of the present invention is a program for causing a computer to operate as any of the above-described X-ray age estimation devices.

This invention makes it possible to provide information useful for assessing health conditions and determining treatment plans.

1 is a block diagram showing the configuration of a X-ray age estimation system according to an embodiment of the present invention; FIG. 1 is a block diagram showing the configuration of a X-ray age estimation learning device. FIG. 1 is a block diagram showing the configuration of a X-ray age estimation device. FIG. 1 is a diagram illustrating the hardware configuration of a computer used as a Roentgen age estimation learning device or a Roentgen age estimation device. FIG. 10 is a flowchart showing the processing of the X-ray age estimation learning device. FIG. 2 is a flowchart showing the processing of the X-ray age estimation device. FIG. 1 is a diagram showing an overview of an embodiment using a radiographic age estimation system. FIG. 1 is a diagram showing the age structure shown in the data used in the X-ray age estimation system. FIG. 10 is a diagram showing the estimation accuracy of the X-ray age estimation model. FIG. 10 is a diagram showing the correspondence between chest X-ray age and actual age obtained using evaluation data as input. FIG. 10 is a diagram showing the correspondence between chest X-ray age and actual age obtained by using test data as input. FIG. 10 is a diagram showing a heat map. FIG. 10 is a diagram showing the relationship between age difference and odds ratio of findings. FIG. 1 shows the relationship between age differences and heart failure outcomes. FIG. 1 is a diagram showing the configuration of an X-ray imaging system. FIG. 1 is a diagram illustrating a configuration of an evaluation system.

The following describes in detail an embodiment of the present invention with reference to the drawings.

Figure 1 shows a radiographic age estimation system 1. The radiographic age estimation system 1 includes a radiographic age estimation learning device 10 and a radiographic age estimation device 50. The radiographic age estimation learning device 10 and the radiographic age estimation device 50 are examples of an image-based age estimation learning device and an image-based age estimation device, respectively, in which radiographic image data obtained by radiography is used as image data obtained by radiography of the inside of the human body. While only one radiographic age estimation device 50 is shown in the figure, any number of radiographic age estimation devices 50 may be used. The radiographic age estimation learning device 10 and the radiographic age estimation device 50 may also be implemented as a single integrated device.

The X-ray age estimation learning device 10 learns a X-ray age estimation model using learning data including X-ray image data and age information of a subject. The X-ray image data is data of X-ray images obtained by X-raying the subject. The age information indicates the subject's actual age at the time the X-ray was taken. The X-ray age estimation model is a model that shows the correspondence between feature amounts obtained from the X-ray image data and the X-ray age of the subject. The X-ray age is an age estimated based on the X-ray image. The X-ray age estimation device 50 obtains the X-ray age of the subject whose age is to be estimated from the X-ray image data of the subject using the X-ray age estimation model learned by the X-ray age estimation learning device 10. In this embodiment, chest X-ray image data is used as the X-ray image data. The chest X-ray image data is data of chest X-ray images obtained by X-raying the subject's chest. The X-ray age estimated based on the chest X-ray image is referred to as the chest X-ray age.

Chest X-rays are the most widely performed examination. The chest X-ray images obtained are quick and easy to obtain, and contain a wealth of useful information. For this reason, chest X-rays are used not only in routine medical practice but also for medical checkups. Chest X-rays are extremely important for diagnosing and monitoring cardiovascular and pulmonary diseases. However, interpreting chest X-rays requires specialized experience and knowledge, and the findings can be diverse. For this reason, chest X-rays have not been used as a consistent quantitative marker. This has made it difficult to incorporate chest X-ray information into recent healthcare reforms using big data.

On the other hand, it has long been known that chest X-ray images differ depending on the subject's age. However, it is not clear what specific characteristics these differences manifest as, and even experienced doctors have found it difficult to accurately estimate a subject's age from a chest X-ray image. The X-ray age estimation device 50 of this embodiment uses only the chest X-ray image data of the subject whose age is to be estimated as input, and obtains a chest X-ray age, which is an estimate of the subject's age.

The X-ray age estimation device 50 also presents the portions of the chest X-ray image that contributed to the estimation of the chest X-ray age. As described below, the chest X-ray age estimated by the X-ray age estimation device 50 is highly accurate. Therefore, if the estimated chest X-ray age is higher than the actual age and there is a large discrepancy between them, some abnormal finding is suggested. In such cases, the portions of the chest X-ray image that significantly contributed to the estimation of the chest X-ray age are estimated to be the areas of the abnormal finding. Furthermore, the difference between the actual age and the chest X-ray age (hereinafter also referred to as the age difference) is also associated with adverse outcomes such as heart failure. For these reasons, chest X-ray age is not only considered to be a new cardiovascular biomarker, but is also useful for clinicians to predict, prevent, and manage heart disease in patients.

Figure 2 is a block diagram showing the configuration of the Roentgen age estimation learning device 10. The Roentgen age estimation learning device 10 includes a model storage unit 11, a data storage unit 12, a conversion unit 13, a learning data generation unit 14, a learning unit 15, a selection unit 18, and an output unit 19. Note that the data storage unit 12 may be located outside the Roentgen age estimation learning device 10.

The model storage unit 11 stores a Roentgen age estimation model. For example, a DNN (Deep Neural Network) such as a CNN (Convolutional Neural Network) can be used as the Roentgen age estimation model. The model storage unit 11 may store multiple Roentgen age estimation models with different structures.

The data storage unit 12 stores training data and verification data. The training data is used to train the X-ray age estimation model. The verification data is used to verify the accuracy of the trained X-ray age estimation model. The training data and verification data include X-ray image data of the subject and information on the subject's age. The X-ray image data used in this embodiment is, for example, chest X-ray image data obtained by X-raying the subject's chest from the front. The age information included in the training data and verification data indicates the subject's actual age at the time the X-ray was taken. Information indicating the subject's date of birth and the date the X-ray was taken may also be used as age information.

The conversion unit 13 performs a predetermined data conversion on the X-ray image data so as to obtain feature quantities to be input into the X-ray age estimation model. Note that if data conversion of the X-ray image data is not required, the X-ray age estimation learning device 10 does not need to include the conversion unit 13. The training data generation unit 14 generates new training data using the training data stored in the data storage unit 12, and writes the generated training data to the data storage unit 12.

The learning unit 15 learns the Roentgen age estimation model using the training data. The learning unit 15 has an estimation unit 16 and an update unit 17. The estimation unit 16 acquires features from the Roentgen image data included in the training data. The features are, for example, pixel values. The estimation unit 16 inputs the acquired features into the Roentgen age estimation model stored in the model storage unit 11 to obtain an estimated Roentgen age. The update unit 17 updates the Roentgen age estimation model so as to reduce the error between the Roentgen age obtained based on the Roentgen image data included in the training data and the subject's actual age indicated by the training data. When the Roentgen age estimation model is a DNN, updating the Roentgen age estimation model means updating the weight values between neurons, i.e., the parameter values used in the function representing the connection relationships between the neurons that make up the DNN.

When there are multiple Roentgen age estimation models trained by the learning unit 15, the selection unit 18 inputs feature values obtained from the Roentgen age estimation model included in the verification data into each Roentgen age estimation model to obtain a Roentgen age estimation result. The selection unit 18 may cause the estimation unit 16 to calculate the Roentgen age. The selection unit 18 calculates the error between the Roentgen age obtained based on the Roentgen image data included in the verification data and the subject's actual age indicated by the verification data. The selection unit 18 selects the Roentgen age estimation model with the statistically smallest error from the multiple Roentgen age estimation models. Note that when there is only one Roentgen age estimation model, the Roentgen age estimation learning device 10 may not have the selection unit 18. Alternatively, the selection unit 18 may output statistical values of the errors calculated for each Roentgen age estimation model via the output unit 19, and the user may input information about the Roentgen age estimation model selected based on the output statistical values of the errors to the Roentgen age estimation learning device 10 via an input unit (not shown).

The output unit 19 outputs various types of data. The output may be displayed on a display, printed by a printer, written to a recording medium, or transmitted to another device connected to the X-ray age estimation learning device 10 via a network. The output unit 19 outputs the X-ray age estimation model selected by the selection unit 18 by transmitting it to the X-ray age estimation device 50 or another device, or by writing it to a recording medium.

Figure 3 is a block diagram showing the configuration of the X-ray age estimation device 50. The X-ray age estimation device 50 includes an input unit 51, a model storage unit 52, a conversion unit 53, an estimation unit 54, a contribution information acquisition unit 55, and an output unit 56.

The input unit 51 inputs data. The input may be received from another device connected to the Roentgen age estimation device 50 via a network, or may be read from a recording medium. The input unit 51 inputs the Roentgen age estimation model learned by the Roentgen age estimation learning device 10 and writes it to the model storage unit 52. The input unit 51 also inputs Roentgen image data to be used for Roentgen age estimation, and outputs it to the conversion unit 53.

The model storage unit 52 stores the Roentgen age estimation model. The model storage unit 11 included in the Roentgen age estimation learning device 10 may be used as the model storage unit 52. The conversion unit 53 performs data conversion on the Roentgen image data input by the input unit 51 in the same manner as the conversion unit 13 of the Roentgen age estimation learning device 10. Note that if data conversion of the Roentgen image data is not required, the Roentgen age estimation device 50 may not have the conversion unit 53. The estimation unit 54 acquires feature quantities from the Roentgen image data converted by the conversion unit 53, in the same way as the estimation unit 16 of the Roentgen age estimation learning device 10. The estimation unit 54 inputs the acquired feature quantities into the Roentgen age estimation model stored in the model storage unit 52 to obtain the Roentgen age.

The contribution information acquisition unit 55 acquires contribution information. The contribution information is information about portions of the X-ray image data from which feature values have been obtained, the degree of contribution to the estimation of the X-ray age using the X-ray age estimation model satisfying a predetermined condition. The predetermined condition may be a condition that classifies the degree of contribution into multiple stages. The contribution information may also be information that represents the degree to which feature values obtained from each portion of the X-ray image data (e.g., each pixel) contributed to the estimation of the X-ray age. If the X-ray age estimation model is a CNN, the contribution information acquisition unit 55 acquires, for example, a heat map generated by Grad-CAM as contribution information.

The output unit 56 outputs various data. The output may be displayed on a display, printed by a printer, written to a recording medium, or transmitted to another device connected to the X-ray age estimation device 50 via a network. The output unit 56 outputs information on the X-ray age estimated by the estimation unit 54 and the contribution information acquired by the contribution information acquisition unit 55.

Each of the X-ray age estimation learning device 10 and the X-ray age estimation device 50 is realized, for example, by a computer 80 shown in FIG. 4. FIG. 4 is a diagram showing the hardware configuration of the computer 80. The computer 80 is configured by connecting a processor 81, a memory unit 82, a user interface 83, and a communication interface 84 via a bus 85. The processor 81 performs calculations and control. The processor 81 is, for example, a CPU (central processing unit) or a GPU (graphics processing unit). The memory unit 82 stores programs for causing the computer 80 to function as the X-ray age estimation learning device 10 or the X-ray age estimation device 50. The processor 81 reads and executes the programs from the memory unit 82. The memory unit 82 also has a work area for the processor 81 to use when executing the programs. The user interface 83 is an input device such as a keyboard, pointing device (mouse, tablet, etc.), button, or touch panel, and a display device such as a monitor. The communication interface 84 connects the computer 80 to other devices so that they can communicate with each other. Note that all or part of the functions of the X-ray age estimation learning device 10 and the X-ray age estimation device 50 may be realized using hardware such as an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), or an FPGA (Field Programmable Gate Array).

When the X-ray age estimation learning device 10 and the X-ray age estimation device 50 are integrated into a single device (hereinafter referred to as the learning estimation device), the model memory unit 11 and the model memory unit 52 may be the same, the estimation unit 16 and the estimation unit 54 may be the same, and the output unit 19 and the output unit 56 may be the same.

Furthermore, the X-ray age estimation learning device 10, the X-ray age estimation device 50, and the learning estimation device may each be realized by multiple computer devices connected to a network. In this case, it is optional which of the multiple computer devices each functional unit of the X-ray age estimation learning device 10, the X-ray age estimation device 50, and the learning estimation device is realized by. Furthermore, one functional unit may be realized by multiple computer devices.

For example, the data memory unit 12, model memory unit 11, conversion unit 13, learning data generation unit 14, learning unit 15 and selection unit 18 of the X-ray age estimation learning device 10 may be realized by different computer devices, the data memory unit 12 and conversion unit 13 may be realized by different computer devices, and the data memory unit 12, conversion unit 13 and learning data generation unit 14 may be realized by different computer devices, and the model memory unit 11, learning unit 15 and selection unit 18 may be realized by different computer devices. Furthermore, for example, the model storage unit 52, the conversion unit 53, the estimation unit 54, and the contribution information acquisition unit 55 of the Roentgen age estimation device 50 may be implemented by different computer devices; the input unit 51, the conversion unit 53, and the output unit 56 may be implemented by different computer devices; or the input unit 51 and the output unit 56 may be implemented by different computer devices; the model storage unit 52, the conversion unit 53, the estimation unit 54, and the contribution information acquisition unit 55 may be implemented by different computer devices. Furthermore, some or all of the data storage unit 12, the conversion unit 13, the learning data generation unit 14, and the learning unit 15 of the Roentgen age estimation learning device 10 may be implemented by multiple computer devices. When the learning unit 15 is implemented by multiple computer devices, each computer device may learn a different type of Roentgen age estimation model in parallel.

Next, the learning process for the Roentgen age estimation model will be explained. The following explanation will be given using an example in which a DNN is used as the Roentgen age estimation model. The Roentgen age estimation learning device 10 learns the Roentgen age estimation model through transfer learning and fine tuning. This makes it possible to improve a DNN that has been trained in advance for another purpose, and to learn the Roentgen age estimation model in a shorter time and with less training data than if a DNN were to be trained as a Roentgen age estimation model from scratch.

Figure 5 is a flow diagram showing an example of the processing of the Roentgen age estimation learning device 10. The model storage unit 11 pre-stores a Roentgen age estimation model in which initial values are set for each parameter. This Roentgen age estimation model is generated by replacing the final layer, or the final layer and some of the layers, of a DNN that has been trained for another purpose. The trained DNN receives as input feature values obtained from image data. The output layer of the Roentgen age estimation model is a single neuron whose output value is the estimated age, i.e., the Roentgen age.

In this embodiment, the model storage unit 11 stores M types (M is an integer greater than or equal to 1) of Roentgen age estimation models (1) to (M). Neural networks with different structures are used for the Roentgen age estimation models (1) to (M). The mth (m is an integer greater than or equal to 1 and less than or equal to M) Roentgen age estimation model is referred to as Roentgen age estimation model (m).

The conversion unit 13 performs a predetermined data conversion on the X-ray image data of the training data and verification data stored in the data storage unit 12 so as to obtain feature quantities to be input into the X-ray age estimation model (step S105). For example, the conversion unit 13 converts the X-ray image data into image data of a predetermined data format and a predetermined data size. Note that if data conversion is not required, the X-ray age estimation learning device 10 does not need to perform the processing of step S105.

The training data generation unit 14 reads some or all of the training data from the data storage unit 12. The training data generation unit 14 generates new training data by replacing the image indicated by the X-ray image data included in the read training data with X-ray image data of an image randomly tilted within a predetermined angle. The training data generation unit 14 may generate multiple new training data with different image tilt angles from one training data. The training data generation unit 14 writes the generated training data to the data storage unit 12 (step S110). By performing training using the training data generated by the training data generation unit 14, the X-ray age estimation model can be made more general and overfitting can be prevented.

The learning unit 15 sets the variable m to an initial value of 1 (step S115). The learning unit 15 performs transfer learning of the Roentgen age estimation model (m) using the training data (step S120). That is, the estimation unit 16 selects one piece of training data and acquires features from the selected training data. The estimation unit 16 inputs the acquired features into the Roentgen age estimation model (m) to obtain an estimated Roentgen age. The update unit 17 updates the values of some of the parameters of the Roentgen age estimation model (m) so as to reduce the difference between the acquired Roentgen age and the actual age indicated by the selected training data. Specifically, the update unit 17 updates the values of the parameters to be updated in transfer learning using an optimization algorithm so as to minimize a loss function calculated using the actual age and the Roentgen age. The loss function is, for example, MSE (mean square error). In transfer learning, the parameters to be updated are parameters that are not included in the trained CNN from which the Roentgen age estimation model (m) is transferred, but may include some parameters included in the CNN from which the model is transferred. For each piece of training data, the learning unit 15 repeatedly estimates the Roentgen age using the estimation unit 16 and updates the values of the parameters to be updated using the update unit 17.

Next, the learning unit 15 fine-tunes the Roentgen age estimation model (m) using the training data (step S125). That is, the estimation unit 16 selects one piece of training data and acquires features from the selected training data. The estimation unit 16 inputs the acquired features into the Roentgen age estimation model (m) to obtain an estimated Roentgen age. The update unit 17 updates the values of all parameters of the Roentgen age estimation model (m) so as to reduce the difference between the acquired Roentgen age and the actual age indicated by the selected training data. That is, the update unit 17 updates the values of all parameters using an optimization algorithm so as to minimize a loss function calculated using the actual age and the Roentgen age. The loss function is, for example, MSE. The learning unit 15 repeats estimation by the estimation unit 16 and updating of all parameter values by the update unit 17 for each piece of training data.

The learning unit 15 determines whether the variable m has reached the number of models M of the X-ray age estimation model (step S130). If the learning unit 15 determines that the variable m has not reached the number of models M (step S130: NO), it adds 1 to the variable m (step S135). The learning unit 15 repeats the processing from step S120. Then, if the learning unit 15 determines that the variable m has reached the number of models M (step S130: YES), it proceeds to the processing of step S140.

The selection unit 18 inputs feature quantities obtained from the X-ray image data included in each verification data into each of the X-ray age estimation models (1) to (M) to obtain a X-ray age estimation result (step S140). The selection unit 18 calculates the difference between the X-ray age obtained based on the X-ray image data included in the verification data and the patient's actual age indicated by the verification data. The selection unit 18 calculates an evaluation index value for each of the X-ray age estimation models (1) to (M) using the calculated difference. Examples of evaluation indexes that can be used include MSE, RMSE (root mean squared error), Pearson's correlation coefficient r between actual age and X-ray age, and MAE (mean absolute error). The selection unit 18 selects the X-ray age estimation model (k) (k is an integer between 1 and M) from among the X-ray age estimation models (1) to (M) that exhibits the highest estimation accuracy in terms of the evaluation index value (step S145). The selection unit 18 may combine multiple evaluation indices to select the most accurate Roentgen age estimation model (k). The output unit 19 outputs the Roentgen age estimation model estimated by the selection unit 18.

The Roentgen age estimation learning device 10 may learn the Roentgen age estimation model without augmenting the learning data using the learning data generation unit 14. The Roentgen age estimation learning device 10 may also learn the values of all parameters of the Roentgen age estimation model based on the learning data without performing transfer learning. The Roentgen age estimation learning device 10 may also perform the learning process of multiple types of Roentgen age estimation models in parallel. If the number of models M = 1, the Roentgen age estimation learning device 10 does not need to perform the process of step S140. In this case, in step S145, the output unit 19 outputs the Roentgen age estimation model (1).

Next, the process of estimating the Roentgen age will be described. The Roentgen age estimation device 50 estimates the Roentgen age using the Roentgen age estimation model (k) learned by the Roentgen age estimation learning device 10. Figure 6 is a flow diagram showing an example of the process of the Roentgen age estimation device 50. The input unit 51 inputs the Roentgen age estimation model (k) output by the Roentgen age estimation learning device 10 and writes it to the model storage unit 52. The Roentgen age estimation device 50 performs the process shown in Figure 6 for each piece of Roentgen image data to be used for Roentgen age estimation.

The input unit 51 inputs X-ray image data and outputs it to the conversion unit 53 (step S205). The input X-ray image data is X-ray image data of a subject whose X-ray age is to be estimated. The input unit 51 may also input information about the subject's age. The conversion unit 53 performs a predetermined data conversion on the input X-ray image data so as to obtain feature values to be input into the X-ray age estimation model (k) (step S210). For example, the conversion unit 53 converts the X-ray image data into image data of a predetermined data format and a predetermined data size. Note that if data conversion is not required, the X-ray age estimation device 50 does not need to perform the processing of step S210.

The estimation unit 54 acquires features from the X-ray image data. The estimation unit 54 inputs the acquired features into the X-ray age estimation model (k) to obtain an estimated X-ray age (step S215). The contribution information acquisition unit 55 acquires contribution information indicating portions of the X-ray image data from which, among the features input into the X-ray age estimation model (k), feature values whose degree of contribution to the estimation of the X-ray age using the X-ray age estimation model satisfies a predetermined condition (step S220). Specifically, the contribution information is a heat map generated using Grad-CAM. Using Grad-CAM, the contribution information acquisition unit 55 acquires information on the positions of pixels from which feature values used in the estimation using the X-ray age estimation model (k) were obtained and the degree to which the feature values contributed to the estimation. The contribution information acquisition unit 55 generates a heat map by overlaying colors representing the degree to which each pixel contributed to the estimation on the X-ray image data. The contribution information acquisition unit 55 may use guided backpropagation to generate the heat map. The output unit 56 outputs the information on the X-ray age estimated by the estimation unit 54 and the heat map generated by the contribution information acquisition unit 55 as estimation results (step S225). When information on actual age is input, the output unit 56 may further output information on the age difference, which is the difference between the X-ray age and the actual age.

Next, we will explain an example of training the X-ray age estimation model and X-ray age estimation using the X-ray age estimation system 1. Figure 7 is a diagram showing an overview of an example using the X-ray age estimation system 1. In this example, 11 types of CNNs were used for the X-ray age estimation model: ResNet18, ResNet34, ResNet50, ResNet101, ResNet152, DenseNet121, DenseNet161, DenseNet169, DenseNet201, Inception-v4, and SENet154. To perform transfer learning, we obtained trained CNN models M0 for each of these 11 types from ImageNet (URL: https://github.com/Cadene/pretrained-models.pytorch). Model M0 consists of a convolutional section and a fully connected section, and classifies image data into 1,000 categories. In other words, the output layer of model M0 consists of 1,000 neurons corresponding to each category. The X-ray age estimation model M1 used in this embodiment uses the convolutional section of model M0 as is, but replaces the fully connected section of model M0 with two new fully connected sections. The two new fully connected sections after replacement consist of a batch-normalization layer, a fully connected layer consisting of 512 neurons using ReLU (Rectified Linear Unit) as the activation function, a batch normalization layer, and a final fully connected layer. The final fully connected layer is connected to a single neuron that outputs the chest X-ray age.

In this example, Database A of chest X-ray image data provided by the National Institutes of Health (https://cloud.google.com/healthcare/docs/resources/public-datasets/nih-chest) was used as the training data, validation data, and test data. The test data is data used to evaluate the accuracy of the X-ray age estimation model trained by the X-ray age estimation learning device 10, and corresponds to the input data input to the X-ray age estimation device 50. Database A contains data mainly from American patients. Of the 112,120 chest X-ray image data of 30,805 patients included in Database A, 63,328 are male. After removing 16 chest X-ray image data of patients over 100 years old from Database A, the remaining chest X-ray image data was randomly divided into training data, validation data, and test data. As a result, 102,029 chest X-ray image data from 28,029 patients were divided into training data D1, chest X-ray image data from 2,523 patients into verification data D2, and 613 chest X-ray image data from 250 patients into test data D3.

In addition, to evaluate the estimation accuracy in a population with different physiques, Database B (http://db.jsrt.or.jp/eng.php) of chest X-ray image data provided by the Japanese Society of Radiological Technology (JSRT), a public interest incorporated association, was used as further test data. Database B contains data from Japanese patients. Of the 247 chest X-ray image data from 247 patients included in Database B, two images for which the gender and age were unknown were excluded, leaving 245 image data from 245 patients as test data D4. Of the test data D4, 118 images were from men.

Furthermore, Database C, which contains chest X-ray image data of 1,562 patients hospitalized due to cardiac disease, was used as test data D5. Database C contains chest X-ray image data taken within two days of admission. Of the data in test data D5, 920 are from men. The chest X-ray image data in Databases A and B includes information on the age and gender at the time the chest X-ray image data was taken, as well as information on findings. In addition to age and gender, the chest X-ray image data in Database C includes more detailed medical data than Databases A and B. The medical data includes information on clinical background and events such as re-hospitalization due to cardiac disease and death.

Figure 8 shows the age structure shown in the data used by the X-ray age estimation system. Figures 8(a), 8(b), and 8(c) show histograms of ages for Database A, Database B, and Database C, respectively. The age range for Database A is 1 to 95 years, with a median of 49 years and an interquartile range of 35 to 39 years. The age range for Database C is 18 to 98 years, with a median of 78 years and an interquartile range of 69 to 84 years.

In this embodiment, the conversion unit 13 and the conversion unit 53 converted the chest X-ray image data into 320 x 320 pixel PNG image data (step S105 in FIG. 5 and step S210 in FIG. 6). The training data generation unit 14 generated additional training data by randomly padding and rotating the image represented by the chest X-ray image data included in the training data obtained from Database A (step S110 in FIG. 5). The rotation was performed randomly between -20 degrees and 20 degrees, and the image was not inverted. In steps S120 and S125 in FIG. 5, the update unit 17 updated the parameter values using the MSE (mean square error) shown in the following equation (1) as a loss function. Here, n is the number of training data. y _i is the correct age (chronological age) indicated by the training data, and y ^{^} _i is the chest X-ray age estimated from the chest X-ray image data included in the training data.

Furthermore, the update unit 17 used Adam (Adaptive Moment Estimation) and CLR (Cyclical Learning Rate) as optimization algorithms in steps S120 and S125 of FIG. 5.

Figure 9 shows the estimation accuracy of each X-ray age estimation model. Figure 9 shows the values of the evaluation indices obtained by performing the process of step S140 in Figure 5 using verification data D2 for each of the 11 types of X-ray age estimation models trained by the X-ray age estimation learning device 10 using training data D1. The evaluation indices are MSE, RMSE, Pearson's correlation coefficient r(R) between actual age and chest X-ray age, and MAE. Figure 9 shows that SENet154 has the highest estimation accuracy.

10 is a diagram showing the correspondence between chest X-ray ages obtained using verification data D2 and chronological ages. Figure 10 plots the correspondence between chest X-ray ages obtained by inputting feature quantities obtained from chest X-ray image data included in verification data D2 into a trained X-ray age estimation model using SENet154, and chronological ages indicated by the verification data D2. The Pearson correlation coefficient r value (R) was 0.9516, and the p value (P) was less than 2.2 × 10 ⁻³²³ .

Figure 11 shows the correspondence between chest X-ray ages obtained using test data D3 and D4 and actual ages. Figure 11(a) plots the correspondence between chest X-ray ages obtained by the X-ray age estimation device 50 using test data D3 as input and actual ages indicated by test data D3. Figure 11(b) plots the correspondence between chest X-ray ages obtained by the X-ray age estimation device 50 using test data D4 as input and actual ages indicated by test data D4. In Figures 11(a) and 11(b), a trained X-ray age estimation model using SENet154 was used to estimate chest X-ray ages.

As shown in Figures 11(a) and 11(b), a very strong correlation is observed between the chest X-ray age estimated by the X-ray age estimation device 50 and the actual age. For test data D3, the r value (R) of the Pearson correlation coefficient was 0.961, the p value (P) was less than 2.2 x ^10-323 , and the mean average average age (MAE) between the chest X-ray age and the actual age was 3.79 years. For test data D4, the r value (R) of the Pearson correlation coefficient was 0.916, the p value (P) was 1.51 x ^10-98 , and the mean average average age (MAE) between the chest X-ray age and the actual age was 4.56 years. Figure 11(c) plots the average ages estimated by four experienced physicians based on chest X-ray images and their correspondence with the actual ages. Although a correlation is observed between the physician-estimated ages and the actual ages, it can be seen that the accuracy of the estimation results by the X-ray age estimation device 50 is higher.

Figure 12 shows heat maps output by the radiographic age estimation device 50. Figure 12(a) is a heat map when Grad-CAM is used in the contribution information acquisition unit 55 of the radiographic age estimation device 50. Figure 12(b) is a heat map when Guided Grad-CAM is used in the contribution information acquisition unit 55 of the radiographic age estimation device 50. According to Figures 12(a) and 12(b), the upper part of the mediastinum and the circumference of the thorax contribute significantly to the estimation, and it is thought that the shape and calcification of the aorta affect the estimation of chest radiographic age.

Figure 13 shows the relationship between age difference and the odds ratio of findings. The age difference is the difference between the chest X-ray age obtained by the X-ray age estimation device 50 using test data D3 and D4 and the actual age. The X-ray age estimation device 50 used a trained X-ray age estimation model using SENet154 to estimate the chest X-ray age. Figure 13 also shows the 95% confidence interval on a log scale. Figure 13 shows that the larger the age difference, the higher the odds ratio of findings.

FIG. 14 shows the relationship between age difference and heart failure outcomes. The age difference is the difference between the chest X-ray age obtained by the X-ray age estimation device 50 using test data D5 and the actual age. The X-ray age estimation device 50 used a trained X-ray age estimation model using SENet154 to estimate the chest X-ray age. The estimation accuracy of the chest X-ray age obtained using test data D5 decreased. However, the Pearson correlation coefficient r value was 0.796 and the p value was 4.6 × 10 ⁻²⁹¹ , confirming a strong correlation.

Furthermore, test data D5 was separated into three groups based on age difference: group G1, whose chest X-ray age was younger; group G2, whose chest X-ray age was approximately the same as their chronological age; and group G3, whose chest X-ray age was older. Groups G1 and G3 each accounted for 20% of the total, and group G2 accounted for 60% of the total. Event-free survival rates were calculated for each group based on the medical data included in test data D5. Events were defined as a composite endpoint consisting of rehospitalization due to heart failure, heart transplant, and death from any cause. Figure 14 shows the change in event-free survival rates over time by group. Figure 14 indicates that the younger the chronological age was compared to the chest X-ray age, the higher the event-free survival rate, and the higher the chronological age was compared to the chest X-ray age, the lower the event-free survival rate. Thus, chest X-ray age can be used to stratify prognosis (presence or absence of cardiovascular events) in heart failure patients. Therefore, the age difference between chest X-ray age and chronological age is useful for preventing overlooking chest X-ray findings in healthy individuals. Furthermore, because the age difference indicates the severity of the disease, chest X-ray age is thought to be useful in determining disease treatment policies.

An example of a system to which the above-described X-ray age estimation device 50 is applied will be described.
15 is a diagram showing the configuration of an X-ray imaging system 100. The X-ray imaging system 100 has an input unit 110, an X-ray imaging device 120, a X-ray age estimation device 130, and an output unit 140. The X-ray age estimation device 130 is the X-ray age estimation device 50 described above.

The input unit 110 is a user interface for inputting instructions to the X-ray imaging system 100 and inputting various types of information. The input unit 110 is, for example, a button, a touch panel, a keyboard, a barcode reader, etc. The input unit 110 may receive information from other devices connected to the X-ray imaging system 100 via a network, or may read information from a recording medium. The various types of information input to the input unit 110 include subject information. The subject information includes, for example, the subject's identification information, name, date of birth, actual age, and other information.

The X-ray imaging device 120 is a conventional device for taking X-ray images of a subject. The X-ray imaging device 120 takes X-ray images in accordance with instructions input via the input unit 110. The X-ray imaging device 120 outputs X-ray image data of the subject obtained by the X-ray imaging to the X-ray age estimation device 130. Information on the date and time of imaging is added to the X-ray image data.

The Roentgen age estimation device 130 is the Roentgen age estimation device 50 described above. If the input unit 110 also serves as the input unit 51 of the Roentgen age estimation device 50, the Roentgen age estimation device 130 may be configured from the Roentgen age estimation device 50 without the input unit 51. Furthermore, if the output unit 140 also serves as the output unit 56 of the Roentgen age estimation device 50, the Roentgen age estimation device 130 may be configured from the Roentgen age estimation device 50 without the output unit 56. The Roentgen age estimation device 130 inputs the Roentgen image data output from the Roentgen imaging device 120 and the actual age information included in the subject information input by the input unit 110, and performs the Roentgen age estimation process shown in FIG. 6. The Roentgen age estimation device 130 may also calculate the actual age information using the date of birth information included in the subject information and the shooting date and time information added to the Roentgen image data. In step S225 of FIG. 6, the X-ray age estimation device 130 outputs the X-ray age information, age difference information, and heat map to the output unit 140.

The output unit 140 outputs various types of data. Examples of output include displaying on a display, printing on a printer, writing to a recording medium, and transmitting to another device connected to the X-ray imaging system 100 via a network. The output unit 140 outputs test result data including X-ray image data output from the X-ray imaging device 120, subject information input by the input unit 110, and X-ray age information, age difference information, and a heat map output from the X-ray age estimation device 130. For example, each time an X-ray image is taken using the X-ray imaging device 120, the output unit 140 displays all of the test result data, or part of the test result data including at least the X-ray age information or age difference information, on the display. Furthermore, the output unit 140 writes the test result data to a recording medium or transmits it to another device.

Note that actual age information does not need to be input to the X-ray age estimation device 130. In this case, age difference information is not output from the X-ray age estimation device 130. The output unit 140 calculates age difference information by calculating the difference between the actual age information indicated in the subject information and the X-ray age information output from the X-ray age estimation device 130.

Figure 16 is a diagram showing the configuration of the evaluation system 200. The evaluation system 200 has a Roentgen age estimation device 210 and an evaluation device 220. The Roentgen age estimation device 210 is the Roentgen age estimation device 50 of the embodiment described above. If the input unit 230 of the evaluation device 220 (described later) also serves as the input unit 51 of the Roentgen age estimation device 50, the Roentgen age estimation device 210 may be configured from the Roentgen age estimation device 50 without the input unit 51. Furthermore, if the output unit 250 of the evaluation device 220 (described later) also serves as the output unit 56 of the Roentgen age estimation device 50, the Roentgen age estimation device 210 may be configured from the Roentgen age estimation device 50 without the output unit 56. The Roentgen age estimation device 210 outputs information on the Roentgen age of the subject obtained by performing the processing of Figure 6 to the evaluation device 220. The Roentgen age estimation device 210 may output information on the age difference between the subject's actual age and Roentgen age to the evaluation device 220, instead of or in addition to the Roentgen age.

The evaluation device 220 evaluates the subject's disease or pathological condition based on the results of a predetermined calculation using the subject's X-ray age estimated by the X-ray age estimation device 210 and information about the subject's biology or lifestyle, or a numerical value representing that information. The evaluation device 220 may also use information about the subject's age difference instead of or in addition to the X-ray age to perform the evaluation. As an example, the evaluation device 220 is realized by the computer 80 shown in FIG. 4. In this case, the memory unit 82 of the computer 80 stores a program for causing the computer 80 to function as the evaluation device 220. All or part of the functions of the evaluation device 220 may be realized using hardware such as an ASIC or PLD.

Here, we will explain an example of a conventional assessment that does not use information on X-ray age or age difference. One example is the Suita Score, which predicts the future onset of angina pectoris, myocardial infarction, and other conditions (see, for example, http://www.ncvc.go.jp/pr/release/006484.html). The Suita Score uses risk factors, such as age, gender, current smoking, diabetes, blood pressure, LDL cholesterol, HDL cholesterol, and CKD (chronic kidney disease), as well as information about the subject's biological and lifestyle characteristics. A score is assigned to each subject for each type of risk factor, and the subject's 10-year probability of developing coronary artery disease is calculated based on the total score. The scoring method is determined according to the type of risk factor. For example, for age, a discrete score is assigned depending on whether the subject belongs to a tiered range of values categorized as 35-44 years, 45-54 years, etc. For blood pressure, a discrete score is assigned depending on the blood pressure status, such as optimal blood pressure, normal blood pressure, Stage 1 hypertension, or Stage 2 or higher hypertension. In the case of CKD, a discrete score is determined depending on which category the individual falls into, categorized based on the state according to the eGFR (estimated glomerular filtration rate) value, such as Stage 1 or 2 (eGFR≧60), Stage 3 (eGFR 30 to <60), or Stage 4 or 5 (eGFR<30). In this way, for some types of risk factors, the score is determined based on the category to which the value or state of the risk factor belongs. Furthermore, for some types of risk factors, the score differs depending on whether or not the individual corresponds to that risk factor. For example, if the individual currently smokes, a predetermined score is assigned, and if the individual currently does not smoke, no score is assigned; in other words, a score of 0 is assigned.

Other examples include the CHADS ₂ score and the CHA ₂ DS ₂ -VASc score (see, for example, https://www.jhf.or.jp/pro/hint/c3/hint005.html), which are used to assess the risk of cerebral infarction in non-valvular atrial fibrillation. The CHADS ₂ score is a risk assessment value obtained by adding up scores based on the presence or absence of each risk factor: heart failure, hypertension, age 75 or older, diabetes, and cerebral infarction/transient ischemic attack. The CHA ₂ DS ₂ -VASc score is a risk assessment value obtained by adding up scores based on the presence or absence of each risk factor, including age 65 to 74, vascular disease, and gender.

The evaluation value in the above example is calculated by substituting the value of each evaluation parameter into a predetermined calculation formula that adds up evaluation parameters corresponding to each risk factor related to the subject being evaluated. The value of each evaluation parameter is determined based on the risk factor corresponding to the evaluation parameter, i.e., information about the subject's biology or lifestyle.

The evaluation device 220 of this embodiment calculates an evaluation value for evaluating a disease or pathological condition by calculating a predetermined calculation formula using evaluation parameters corresponding to the subject's X-ray age estimated by the X-ray age estimation device 210 and evaluation parameters corresponding to information related to the subject's biology or lifestyle. Note that instead of or in addition to the X-ray age, an evaluation parameter corresponding to the age difference between the X-ray age and the actual age may be used to calculate the evaluation value.

The evaluation device 220 has an input unit 230, an evaluation unit 240, and an output unit 250. The input unit 230 inputs data. The input may be via a user interface such as a keyboard, mouse, or touch panel, or may be received from another device connected via a network, or read from a recording medium. The input unit 230 inputs information about the subject's biological or lifestyle habits according to the evaluation parameters, and further inputs one or both of the X-ray age information and age difference information output by the X-ray age estimation device 210. The input unit 230 outputs the input values to the evaluation unit 240.

When the evaluation unit 240 receives information about the Roentgen age from the Roentgen age estimation device 210, it sets a value corresponding to the Roentgen age for the evaluation parameter corresponding to the Roentgen age. Similarly, when the evaluation unit 240 receives information about the age difference from the Roentgen age estimation device 210, it sets a value corresponding to the age difference for the evaluation parameter corresponding to the age difference. The evaluation unit 240 may also calculate the age difference as the difference between the Roentgen age input from the Roentgen age estimation device 210 and the actual age input by the input unit 230. In this way, the age difference can also be information obtained by performing a predetermined calculation on the Roentgen age. Furthermore, the evaluation unit 240 sets a value corresponding to each piece of information about the subject's biological or lifestyle habits for each evaluation parameter. Examples of biological information include, but are not limited to, age, gender, test results from various clinical tests, conditions classified based on the test results from various clinical tests, various past diseases, and various current diseases. Examples of lifestyle information include, but are not limited to, smoking, sleep time, and diet.

The value set for the evaluation parameter may be, for example, a value indicated by the information corresponding to the evaluation parameter, a value corresponding to the state indicated by the information corresponding to the evaluation parameter, a value corresponding to the category to which the value or state indicated by the information corresponding to the evaluation parameter belongs, a value obtained by performing a predetermined calculation on the value indicated by the information corresponding to the evaluation parameter, or a value corresponding to whether or not the evaluation parameter corresponds to the information corresponding to the evaluation parameter. The evaluation unit 240 may use the value of the X-ray age as the evaluation parameter value directly, or may use the value as an evaluation parameter value corresponding to which of the categorized ranges the X-ray age falls (e.g., under a1 year old, a1 to a2 years old, a2 to a3 years old, etc.). Furthermore, for example, the evaluation unit 240 may use the value of the age difference as the evaluation parameter value directly, or may use the value as an evaluation parameter value corresponding to which of the categorized ranges the age difference falls (e.g., b years or older, b to -b years old, -b years or younger, etc.).

The evaluation unit 240 calculates an evaluation value, which quantitatively represents the evaluation of the evaluation target, by substituting the value of each evaluation parameter into a calculation formula that uses each evaluation parameter. An example of the calculation formula is the addition of evaluation parameters, but this is not limited to this; any calculation formula can be used. If the calculation formula is a weighted addition of evaluation parameters, the weight for other evaluation parameters may be determined according to the value of an evaluation parameter corresponding to the X-ray age or age difference. The evaluation target is information related to the disease or pathological condition, such as the severity of the disease or pathological condition, or a predicted prognosis. The evaluation unit 240 may also pre-store a correspondence between the evaluation value and an evaluation (e.g., low/medium/high risk, x1%/x2%/..., etc.) and obtain an evaluation corresponding to the calculated evaluation value. The output unit 250 outputs one or both of the evaluation value calculated by the evaluation unit 240 and the evaluation corresponding to the evaluation value as the evaluation result.

Artificial intelligence may be used to calculate the evaluation value using each evaluation parameter. In this case, the calculation formula for the evaluation value corresponds to, for example, a DNN. When a DNN is used, the input to the DNN is the value of each evaluation parameter, and the output is the evaluation value. For the value of the evaluation parameter corresponding to the Roentgen age or age difference, the value of the Roentgen age or age difference may be used directly, a discrete value corresponding to the numerical range to which the value of the Roentgen age or age difference belongs may be used, or a continuous value obtained by normalizing the Roentgen age or age difference may be used.

The types of evaluation parameters used to calculate the evaluation value and the calculation (calculation formula) for calculating the evaluation value depending on the evaluation target are statistically determined through learning using training data.

The output unit 250 outputs the evaluation results received from the evaluation unit 240. Output includes, for example, one or more of displaying on a display, printing on a printer, writing to a recording medium, and sending to another device connected to the evaluation device 220 via a network.

According to the above-described embodiment, artificial intelligence can be used to estimate a subject's age from their chest X-ray image data. The new concepts of X-ray age and chest X-ray age are intuitively easy for patients to understand and have the potential for widespread use. Furthermore, the difference between chest X-ray age and chronological age can assist in determining whether or not abnormal findings are present. This is expected to reduce the oversight of abnormal findings during general medical examinations and treatment. Furthermore, the ability to highlight detected abnormalities allows for more in-depth discussion of the cause of the same outcome. Furthermore, the difference between chest X-ray age and chronological age reflects the severity of the disease. This age difference stratifies the prognosis of heart failure patients after chest X-ray image data is taken. Therefore, it is expected that X-ray age and chest X-ray age will be useful in determining treatment strategies for patients with the disease, and they have the potential for widespread use by medical professionals. Furthermore, chest X-ray images have been difficult to utilize in digital medicine due to the variability in findings depending on the interpreter. In this embodiment, chest X-ray findings can be converted into a unique quantitative value known as chest X-ray age, which is thought to have the potential to contribute to the promotion of digital medicine.

Furthermore, when X-rays are taken of subjects during health checkups or medical treatment, it is possible to provide information on X-ray age and age differences. This allows for the rapid provision of useful clinical information. Furthermore, information on X-ray age and age differences can be incorporated as one of the evaluation parameters in disease/pathological condition severity assessments and prognosis prediction research. This means that X-ray age can be used as a new research tool.

In the above embodiment, an example was described in which chest X-ray image data was used as image data obtained by radiography of the inside of the human body, but images taken using other methods such as MRI (Magnetic Resonance Imaging) or CT (Computed Tomography) may also be used, or images of other parts of the body such as the brain.

According to the above-described embodiment, the X-ray age estimation learning device includes a learning unit. The learning unit uses data on the X-ray image of the subject and the subject's age at the time the X-ray image was taken to learn a X-ray age estimation model, which is a model that indicates the correspondence between feature amounts obtained from the X-ray image data and the subject's age.

The X-ray age estimation device includes an estimation unit. The estimation unit inputs feature quantities obtained from data on an X-ray image of the subject whose age is to be estimated into the X-ray age estimation model trained by the X-ray age estimation learning device, and obtains an estimated age of the subject whose age is to be estimated. The X-ray age estimation device may further include a contribution information acquisition unit. The contribution information acquisition unit acquires contribution information, which is information about portions of the X-ray image of the subject whose age is to be estimated, where feature quantities whose degree of contribution to the age estimation of the subject whose age is to be estimated meet predetermined conditions are obtained. The X-ray age estimation device may further include an output unit. The output unit outputs one or more of information on the age estimated by the estimation unit, information on the difference between the actual age of the subject whose age is to be estimated and the age estimated by the estimation unit, and contribution information.

For example, the X-ray image is an X-ray image of the subject's chest, and the estimation target X-ray image is an X-ray image of the subject's chest whose age is to be estimated.

The X-ray imaging system may also be equipped with a X-ray age estimation device. The X-ray imaging device in the X-ray imaging system inputs X-ray image data obtained by X-ray imaging of the subject into the X-ray age estimation device as data for the X-ray image to be estimated.

The evaluation device also includes an evaluation unit. The evaluation unit evaluates the subject's disease or pathological condition based on the results of a predetermined calculation using the subject's age estimation result obtained by the X-ray age estimation device and information about the subject's biology or lifestyle habits or numerical values representing that information.

The image-based age estimation learning device includes a learning unit. The image-based age estimation learning device corresponds, for example, to the X-ray age estimation learning device 10 of the embodiment. The learning unit uses data on radiographic images, which are images obtained by radiography of the subject's body, and information on the subject's age at the time the radiographic images were taken, to learn an age estimation model that indicates the correspondence between feature values obtained from the radiographic image data and the subject's age. The age estimation model corresponds, for example, to the X-ray age estimation model of the embodiment. Radiographic images include, for example, images obtained by X-ray photography and images obtained by MRI. The feature values are, for example, pixel values.

The image-based age estimation device includes an estimation unit. The image-based age estimation device corresponds, for example, to the X-ray age estimation device 50 of the embodiment. The estimation unit inputs feature quantities obtained from data on the estimation target image, which is a radiographic image of the subject whose age is to be estimated, into the age estimation model trained by the image-based age estimation learning device, and obtains an estimated age for the subject whose age is to be estimated.

The above describes an embodiment of the present invention with reference to the drawings, but it is clear that the above embodiment is merely an example of the present invention and that the present invention is not limited to the above embodiment. Therefore, components may be added, omitted, substituted, or otherwise modified without departing from the technical spirit and scope of the present invention.

1...Roentgen age estimation system, 10...Roentgen age estimation learning device, 11...Model memory unit, 12...Data memory unit, 13...Conversion unit, 14...Learning data generation unit, 15...Learning unit, 16...Estimation unit, 17...Update unit, 18...Selection unit, 19...Output unit, 50...Roentgen age estimation device, 51...Input unit, 52...Model memory unit, 53...Conversion unit, 54...Estimation unit, 55...Contribution information acquisition unit, 56...Output unit, 80...Computer, 81...Processor, 82...Memory unit, 83...User interface, 84...Communication interface, 85...Bus, 100...Roentgen imaging system, 110...Input unit, 120...Roentgen imaging device, 130...Roentgen age estimation device, 140...Output unit, 200...Evaluation system, 210...Roentgen age estimation device, 220...Evaluation device, 230...Input unit, 240...Evaluation unit, 250...Output unit

Claims (14)

a learning unit that uses training X-ray image data, which is data on an X-ray image of a subject and data on an image obtained by tilting the X-ray image within a predetermined angle, and information on the age of the subject at the time the X-ray image was taken , or that uses training converted image data , which is data on a converted X-ray image obtained by performing a predetermined conversion on the X-ray image data and data on an image obtained by tilting the X-ray image within a predetermined angle , and information on the age of the subject at the time the X-ray image was taken, and that inputs pixel values obtained from the X-ray image data or pixel values obtained from the converted X-ray image data, and outputs the age of the subject corresponding to the input pixel value information;
An X-ray age estimation learning device equipped with the device. The X-ray image is a chest X-ray image.
The X-ray age estimation learning device according to claim 1 . a radiological age estimation model trained using training radiological image data, which are radiological image data of a subject and data of an image obtained by tilting said radiological image within a predetermined angle, and information on the age of said subject at the time said radiological image was taken, or using training converted image data , which are converted radiological image data obtained by performing a predetermined conversion on said radiological image data and data of an image obtained by tilting said radiological image within a predetermined angle and performing the predetermined conversion on said data of said radiological image, and information on the age of said subject at the time said radiological image was taken, said radiological age estimation model receiving pixel values obtained from said radiological image data or pixel values obtained from said converted radiological image data as input and outputting the age of said subject corresponding to said input pixel values; an estimation unit which inputs pixel values obtained from data of an estimation target radiological image, which is an radiological image of a subject whose age is to be estimated, or pixel values obtained from data of a transformed estimation target radiological image obtained by performing the predetermined conversion on said data of said estimation target radiological image, and obtains an estimated result of the age of said subject whose age is to be estimated;
An X-ray age estimation device comprising: a contribution information acquisition unit that acquires contribution information, which is information about a portion of the X-ray image to be estimated or the converted X-ray image to be estimated, where the pixel value satisfies a predetermined condition as to the degree of contribution to the age estimation of the subject to be age estimation;
The X-ray age estimation device according to claim 3, further comprising: an output unit that outputs one or more of information on the age estimated by the estimation unit, information on a difference between the actual age of the subject whose age is to be estimated and the age estimated by the estimation unit, and the contribution information;
The X-ray age estimation device according to claim 4. The X-ray image and the estimation target X-ray image are X-ray images of the chest.
The X-ray age estimation device according to any one of claims 3 to 5. a learning unit that uses, as input, pixel values obtained from the transmitted image data or the converted transmitted image data, by using training transmitted image data, which are data of transmitted image obtained by transmitting image data of the human body of a subject and data of images obtained by tilting the transmitted image data within a predetermined angle , and information on the age of the subject at the time the transmitted image data was captured, or by using training converted transmitted image data, which are data of converted transmitted image data obtained by performing a predetermined conversion on the transmitted image data and data of images obtained by tilting the transmitted image data within a predetermined angle, and information on the age of the subject at the time the transmitted image data was captured; and learns an age estimation model that outputs the age of the subject corresponding to the input pixel values;
An image age estimation learning device comprising: an age estimation model trained using training transmission image data, which are transmission image data obtained by transmitting a transmission image of a human body of a subject and data of an image obtained by tilting the transmission image within a predetermined angle, and information on the age of the subject at the time the transmission image was captured , or using training transformed transmission image data, which are transformed transmission image data obtained by performing a predetermined transformation on the transmission image data and data of an image obtained by tilting the transmission image within a predetermined angle and performing the predetermined transformation on the data of the transmission image, and information on the age of the subject at the time the transmission image was captured, wherein the age estimation model receives as input pixel values obtained from the transmission image data or the transformed transmission image data, and outputs the age of the subject corresponding to the input pixel values, and an estimation unit inputs pixel values obtained from data of an estimation target image, which is a transmission image of the subject whose age is to be estimated, or pixel values obtained from data of a transformed estimation target image obtained by performing the predetermined transformation on the data of the estimation target image, to obtain an estimation result of the age of the subject whose age is to be estimated;
An image age estimation device comprising: The X-ray age estimation device according to any one of claims 3 to 6,
an X-ray imaging device that takes an X-ray of a subject and inputs data of the X-ray image of the subject obtained by the X-ray imaging into the X-ray age estimation device as data of an X-ray image to be estimated;
An X-ray imaging system comprising: an evaluation unit that performs an evaluation of a disease or pathological condition of the subject based on a result of a predetermined calculation using the result of the age estimation of the subject by the X-ray age estimation device according to any one of claims 3 to 6 and information on the subject's biological or lifestyle habits or a numerical value representing the information;
An evaluation device comprising: a learning step of learning a X-ray age estimation model using learning X-ray image data, which is X-ray image data of a subject and image data obtained by tilting said X-ray image within a predetermined angle, and information on the age of said subject at the time said X-ray image was taken, or using learning transformed image data, which is transformed X-ray image data obtained by performing a predetermined transformation on said X-ray image data and image data obtained by tilting said X-ray image within a predetermined angle and information on the age of said subject at the time said X-ray image was taken, and which takes pixel values obtained from said X-ray image data or pixel values obtained from said transformed X-ray image data as input, and outputs the age of said subject corresponding to the information on said input pixel values;
A learning method for X-ray age estimation. an estimation step of inputting pixel values obtained from data of an estimation target X-ray image, which is an X-ray image of a subject whose age is to be estimated, or pixel values obtained from data of a transformed X-ray image obtained by subjecting the X-ray image data to the transformation target X-ray image data, which is data of an X-ray image of a subject and data of an image obtained by tilting the X-ray image within a predetermined angle , into the X-ray age estimation model which is trained using data of training X-ray images, which are respectively data of an X-ray image of the subject and data of an image obtained by tilting the X-ray image within a predetermined angle and data of an image obtained by subjecting the X-ray image data to the transformation target X-ray image data, and information on the age of the subject when the X-ray image was taken;
A method for estimating the age of a person based on X-rays. Computer,
A program for causing the device to operate as the X-ray age estimation learning device according to claim 1 or 2. Computer,
A program for causing the X-ray age estimation device according to any one of claims 3 to 6 to operate.