Cardiac function estimation device, cardiac function estimation method, and program

Abstract

A cardiac function estimation device and the like are provided that can identify people who require medical care or treatment related to cardiac function without using test results obtained by echocardiography. [Solution] The cardiac function estimation device 1 includes a trained machine learning model 124 that outputs output data related to the contractile function and diastolic function of the left ventricle when an X-ray image obtained by imaging a person's chest is input, and an estimation unit 122 that inputs the X-ray image to the machine learning model 124, and the estimation unit 122 identifies and outputs the person's cardiac function corresponding to the output data output by the machine learning model 124.

Description

[Technical Field]

[0001] The present disclosure relates to a cardiac function estimation device and the like that estimates human cardiac function. [Background technology]

[0002] In recent years, the number of patients diagnosed with heart failure has been steadily increasing due to the aging of society. Currently, medical professionals working in general internal medicine provide the following medical care to patients suspected of having heart failure. The medical professional takes chest X-rays and echocardiograms of the patient to determine whether the patient has heart failure or another disease. Based on the results of the diagnosis, the medical professional then decides whether the patient needs to be examined by a cardiologist or whether other tests are necessary.

[0003] In the above-mentioned medical examinations, echocardiograms are necessary for accurate diagnosis, but echocardiography has various disadvantages compared to chest X-rays. For example, echocardiography requires a longer examination time than chest X-rays, resulting in longer consultation times for each patient. Furthermore, in Japan, chest X-rays are available at almost all medical institutions, while echocardiography is only available at a limited number of medical institutions. This means that patients suspected of heart failure tend to visit specific medical institutions, placing a heavy burden on those institutions. Therefore, there is a need to develop a diagnostic system that does not require echocardiography in order to reduce the burden on the entire medical system.

[0004] For example, a machine learning model has been developed that determines abnormalities in cardiac function from X-ray images obtained by capturing a person's chest (see Non-Patent Document 1). Specifically, the machine learning model estimates parameters for left ventricular contractile function and valvular disease from information contained in the X-ray images, and determines abnormalities in cardiac function. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Daiju Ueda et al., “Artificial intelligence-based model to classify cardiac functions from chest radiographs: a multi-institutional, retrospective model development and validation study,” The Lancet Digital Health, Volume 5, Issue 8, August 2023, Pages e525-e533. Summary of the Invention [Problem to be solved by the invention]

[0006] However, because the machine learning model disclosed in Non-Patent Document 1 estimates parameters of valvular disease, it may determine that a person has abnormal cardiac function even if they do not have cardiac dysfunction such as heart failure and are not experiencing any disruption in their daily lives.

[0007] In other words, the device having the above machine learning model is not a device for identifying people who need medical treatment or therapy related to cardiac function.

[0008] Therefore, the present disclosure has been made to solve such problems, and aims to provide a cardiac function estimation device, etc. that can identify people who require medical treatment or therapy related to cardiac function without using the test results obtained from echocardiography. [Means for solving the problem]

[0009] A cardiac function estimation device according to one embodiment of the present disclosure includes a trained machine learning model that, when an X-ray image obtained by imaging a person's chest is input, outputs output data related to the contractile function and diastolic function of the left ventricle, and an estimation unit that inputs the X-ray image into the machine learning model, and the estimation unit identifies and outputs the cardiac function of the person corresponding to the output data output by the machine learning model.

[0010] A cardiac function estimation method according to one embodiment of the present disclosure is a cardiac function estimation method executed by a cardiac function estimation device equipped with a trained machine learning model, in which an X-ray image obtained by imaging a person's chest is input into the machine learning model, and the machine learning model identifies and outputs the person's cardiac function corresponding to output data regarding the left ventricular systolic function and the left ventricular diastolic function based on the X-ray image.

[0011] A program according to one aspect of the present disclosure causes a computer to execute steps included in the cardiac function estimation method. [Effects of the Invention]

[0012] According to the present disclosure, a cardiac function estimation device and the like are provided that can identify people who require medical care or treatment related to cardiac function without using test results obtained from echocardiography. [Brief explanation of the drawings]

[0013] [Figure 1] FIG. 1 is a block diagram showing the configuration of a cardiac function estimation device according to an embodiment. [Figure 2] FIG. 2 is a schematic diagram showing an example of the structure of a convolutional neural network (CNN). [Figure 3] FIG. 3 is a flowchart showing the operations executed by the cardiac function estimation device according to the embodiment. [Figure 4] FIG. 4 is a diagram showing the correspondence between combinations of output data output by the machine learning model and cardiac functions identified by the estimation unit based on the combinations and output to the display unit. [Figure 5] FIG. 5 is a diagram showing an outline of a method for training a machine learning model and a method for testing the machine learning model after training. [Figure 6] Figure 6 is a graph showing the evaluation results for each of the six CNN models. [Figure 7A] FIG. 7A is a diagram showing a first example of a verification result obtained using Grad-CAM. [Figure 7B] FIG. 7B is a diagram showing a second example of the verification result obtained using Grad-CAM. [Figure 7C] FIG. 7C is a diagram showing a third example of the verification result obtained using Grad-CAM. DETAILED DESCRIPTION OF THE INVENTION

[0014] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. Note that each of the embodiments described below represents a specific example of the present disclosure. Therefore, the numerical values, shapes, materials, components, component placement positions, connection forms, etc. shown in the following embodiments are merely examples and are not intended to limit the present disclosure. Therefore, among the components in the following embodiments, components that are not recited in the independent claims of the present disclosure will be described as optional components.

[0015] Note that each figure is a schematic diagram and is not necessarily an exact illustration. Therefore, the scales and the like do not necessarily match in each figure. Furthermore, in each figure, the same reference numerals are used for substantially the same configurations, and redundant explanations will be omitted or simplified.

[0016] (Embodiment) [composition] First, the configuration of a cardiac function estimation device according to this embodiment will be described. Fig. 1 is a block diagram showing the configuration of a cardiac function estimation device 1 according to this embodiment.

[0017] The cardiac function estimation device 1 is a device that identifies and outputs a person's cardiac function. Specifically, the cardiac function estimation device 1 identifies and outputs a person's cardiac function by estimating the state of the left ventricle's systolic function and the state of the left ventricle's diastolic function.

[0018] As shown in FIG. 1, the cardiac function estimation device 1 includes an acquisition unit 11, a control unit 12, a storage unit 13, and a display unit .

[0019] The acquisition unit 11 acquires X-ray images or training data from an external device. The X-ray images and training data acquired by the acquisition unit 11 will be described.

[0020] The X-ray image for examination is image data captured in advance by an X-ray inspection device, and is image data obtained by imaging a person's chest. The X-ray image for examination is an X-ray image that is actually used when the cardiac function estimation device 1 identifies a person's cardiac function. The X-ray image for examination is image data expressed in grayscale, and, for example, each pixel value is expressed as a numerical value from 0 to 255. The X-ray image for examination also includes not only image data obtained by imaging a person's chest, but also character data such as the date and time the X-ray image was captured or the name of the person who underwent the examination using the X-ray inspection device.

[0021] The training data is data consisting of multiple X-ray images and multiple pieces of output data associated with each of the multiple X-ray images. Note that, hereinafter, the multiple X-ray images that make up the training data will be referred to as "X-ray images for training."

[0022] The training X-ray images, like the inspection X-ray images, are image data captured in advance by an X-ray inspection device, and are image data obtained by capturing an image of a person's chest. The training X-ray images are X-ray images used when the machine learning model 124, which will be described later, learns.

[0023] The output data is a set of a first value and a second value. The first value indicates whether the left ventricular systolic function is normal or abnormal. Specifically, if the left ventricular systolic function is normal, the first value indicates 0, and if the left ventricular systolic function is abnormal (i.e., left ventricular systolic dysfunction), the first value indicates 1. The second value indicates whether the left ventricular diastolic function is normal or abnormal. Specifically, if the left ventricular diastolic function is normal, the second value indicates 0, and if the left ventricular diastolic function is abnormal (i.e., left ventricular diastolic dysfunction), the second value indicates 1. In other words, the output data is data related to the left ventricular systolic function and the left ventricular diastolic function, and specifically, the output data is data indicating one of four patterns: [0,0], [1,0], [0,1], and [1,1]. Note that [0,0] indicates normal cardiac function, [1,0] indicates left ventricular systolic dysfunction, [0,1] indicates left ventricular diastolic dysfunction, and [1,1] indicates left ventricular systolic dysfunction and left ventricular diastolic dysfunction. The definitions of left ventricular systolic dysfunction and left ventricular diastolic dysfunction will be described later.

[0024] In addition, when there is no need to distinguish between X-ray images for inspection and X-ray images for learning, simply write "X-ray images."

[0025] The control unit 12 controls the entire cardiac function estimation device 1. The control unit 12 is realized by a microcomputer, a processor, or the like. That is, the functions of the control unit 12 are realized by the microcomputer or the processor executing a program stored in a memory, etc. The control unit 12 also includes an image conversion unit 121, an estimation unit 122, a learning unit 123, and a machine learning model 124.

[0026] The image conversion unit 121 converts the X-ray image for inspection or the X-ray image for learning acquired by the acquisition unit 11 into a data format that can be input to the machine learning model 124. Specifically, the image conversion unit 121 extracts only image data from the X-ray image. Next, the image conversion unit 121 converts the X-ray image from which only the image data has been extracted into an X-ray image (input data) of 224 pixels x 224 pixels (i.e., 50,176 pixels).

[0027] The estimation unit 122 inputs the X-ray image for examination (input data) converted by the image conversion unit 121 into the trained machine learning model 124. Furthermore, the estimation unit 122 identifies the cardiac function of the person corresponding to the output data output by the machine learning model 124, and outputs it to the display unit 14. That is, based on the output data output by the machine learning model 124, the estimation unit 122 identifies any one of "normal cardiac function," "left ventricular systolic dysfunction," "left ventricular diastolic dysfunction," or "left ventricular systolic dysfunction and left ventricular diastolic dysfunction," and outputs it to the display unit 14.

[0028] The learning unit 123 trains the machine learning model 124 using the training data acquired by the acquisition unit 11. Specifically, the learning unit 123 uses the training X-ray images (input data) converted by the image conversion unit 121 as explanatory variables, and further sets known output data associated with the training X-ray images as objective variables, thereby training the machine learning model 124 using a set of multiple explanatory variables and objective variables. Note that a specific training method for the machine learning model 124 will be described later.

[0029] The machine learning model 124 is a machine learning model that outputs output data when an X-ray image (input data) is input. The inventors have realized the machine learning model 124 using a trained convolutional neural network (CNN). Figure 2 is a schematic diagram showing an example of the structure of a convolutional neural network (CNN).

[0030] As shown in Figure 2, a CNN consists of three layers: an input layer, a hidden layer (intermediate layer), and an output layer.

[0031] The input layer is a layer that inputs a plurality of pixel values ​​2 that make up the input data. Figure 2 shows a schematic example in which seven pixel values ​​2 are input to the input layer, but in reality, 50,176 pixel values ​​2 are input to the input layer.

[0032] The hidden layer is a layer that exists between the input layer and the output layer. The hidden layer weights the values ​​output from the input layer and calculates feature quantities. Figure 2 shows a case where there are two hidden layers, but the number of hidden layers may be other than two.

[0033] The output layer is a layer that outputs a first value 3 and a second value 4 that constitute the output data.

[0034] In this specification, the inventors have realized the machine learning model 124 using at least one of six CNN models: VGG16, ResNet50, DenseNet121, EfficientNetB0, InceptionV3, and MobileNetV3Large. However, the machine learning model 124 may also be realized using a CNN model other than the above-listed CNN models.

[0035] Returning to the explanation of FIG. 1, the storage unit 13 is a memory or the like that stores the X-ray images for examination acquired by the acquisition unit 11, teacher data, programs executed by the control unit 12, and the like.

[0036] The display unit 14 is realized by, for example, a liquid crystal display or the like. The display unit 14 displays information output by the control unit 12 (particularly, the cardiac function identified by the estimation unit 122). For example, the display unit 14 displays information such as (i) whether the cardiac function of the heart is normal or abnormal, (ii) if the cardiac function of the heart is abnormal, which functional disorder it is classified as, and (iii) whether an examination by a cardiologist or a specialist in another field is necessary. Of the information (i) to (iii), (i) and (ii) are the cardiac functions identified by the estimation unit 122.

[0037] The acquiring unit 11 may communicate with an external device by wireless communication or by wired communication, and the standard of the communication performed by the acquiring unit 11 is not particularly limited.

[0038] The display unit 14 may also be realized by a device separate from the cardiac function estimation device 1 (for example, an external monitor, etc.).

[0039] [Definition of left ventricular systolic dysfunction and left ventricular diastolic dysfunction] The definitions of the left ventricular systolic dysfunction and left ventricular diastolic dysfunction described above will now be explained. Note that the left ventricular systolic dysfunction and left ventricular diastolic dysfunction defined in this specification are indicators of heart failure.

[0040] 1) Definition of left ventricular systolic dysfunction In this specification, the inventors defined a left ventricular systolic dysfunction when the left ventricular ejection fraction (LVEF) is less than 50%. That is, when the left ventricular ejection fraction is less than 50%, the first value 3 indicates 1, and when the left ventricular ejection fraction is 50% or more, the first value 3 indicates 0. The inventors measured the left ventricular ejection fraction using the modified Simpson's method.

[0041] 2) Definition of left ventricular diastolic dysfunction In this specification, the inventors have defined left ventricular diastolic dysfunction as occurring when three or more of the following four items are met. That is, when three or more of the following four items are met, the second value 4 indicates 1, and when two or less of the following four items are met, the second value 4 indicates 0. Details of the four items are shown below.

[0042] Mean ratio of early diastolic velocity measured by transmitral Doppler to early diastolic velocity measured by tissue Doppler (E / e') > 14 Mid-wall mitral annular early diastolic velocity (septal e') < 7 cm / s measured by tissue Doppler, or lateral mitral annular early diastolic velocity (lateral e') < 10 cm / s measured by tissue Doppler Tricuspid regurgitation velocity (TRV) > 2.8 m / s Left Atrial Volume Index (LAVI) > 34 mL / m2 measured by end-systolic apical four-chamber view 2

[0043] The inventors determined the above four items based on the algorithm recommended by the American Society of Echocardiography in 2016.

[0044] [Operation] Next, the operations performed by the cardiac function estimation device 1 when identifying a person's cardiac function will be described with reference to Fig. 3. Fig. 3 is a flowchart showing the operations performed by the cardiac function estimation device 1 according to the embodiment.

[0045] First, the acquisition unit 11 acquires an X-ray image for inspection from an external device (step S1).

[0046] The image conversion unit 121 converts the X-ray image for examination acquired by the acquisition unit 11 in step S1 into a data format that can be input to the trained machine learning model 124 (step S2).

[0047] The estimation unit 122 inputs the X-ray image for examination converted by the image conversion unit 121 in step S2 (that is, input data) to the machine learning model 124 (step S3).

[0048] The estimation unit 122 acquires the output data output by the machine learning model 124 (step S4).

[0049] The estimation unit 122 identifies the cardiac function of the person corresponding to the output data acquired in step S4 (step S5).

[0050] The estimation unit 122 outputs the result determined in step S5 to the display unit 14 (step S6).

[0051] FIG. 4 is a diagram showing the correspondence between the combination of output data output by the machine learning model 124 and the cardiac function identified by the estimation unit 122 based on the combination and output to the display unit 14.

[0052] As shown in FIG. 4 , for example, when the output data acquired by the estimation unit 122 is [0,0], the estimation unit 122 determines that “cardiac function is normal” and outputs this to the display unit 14. When the output data acquired by the estimation unit 122 is [1,0], the estimation unit 122 determines that “left ventricular systolic dysfunction” and outputs this to the display unit 14. When the output data acquired by the estimation unit 122 is [0,1], the estimation unit 122 determines that “left ventricular diastolic dysfunction” and outputs this to the display unit 14. When the output data acquired by the estimation unit 122 is [1,1], the estimation unit 122 determines that “left ventricular systolic dysfunction and left ventricular diastolic dysfunction” and outputs this to the display unit 14.

[0053] As described above, the cardiac function estimation device 1 can identify a person's cardiac function from an X-ray image.

[0054] [How to train a machine learning model and how to test the machine learning model after training] FIG. 5 is a diagram showing an outline of a method for training the machine learning model 124 and a method for testing the machine learning model 124 after training.

[0055] As shown in Fig. 5, the inventors prepared 1633 sets of training data (All samples) in advance. The inventors used 1473 of the 1633 sets as a training dataset (Training) and the remaining 160 sets as a test dataset (Test). One set is a pair of explanatory variables and response variables for one X-ray image.

[0056] The inventors used K-fold cross validation as a training method for the machine learning model 124 (i.e., the CNN model). K-fold cross validation is a technique in which a training dataset is divided into K groups, one of which is used as a validation dataset (Validation), and the remaining groups are used as training datasets (Training). In other words, there are K combinations of the evaluation dataset and the training dataset. A specific procedure for training the machine learning model 124 using the above K-fold cross validation will be described below.

[0057] First, the learning unit 123 prepares K identical CNN models. The learning unit 123 inputs a training dataset to each of the K CNN models and trains the K CNN models. Note that the learning unit 123 ensures that the training datasets input to each of the K CNN models are not identical. In this way, the learning unit 123 prepares K trained CNN models (denoted as model1-model5 in FIG. 5) that have been trained using different training datasets. Then, the learning unit 123 inputs a corresponding evaluation dataset to each of the K trained CNN models. The learning unit 123 averages the accuracy rates of the output data output by each of the K trained CNN models. The learning unit 123 evaluates the accuracy of the K trained CNN models from the result obtained by averaging the accuracy rates of the output data, adjusts the parameters set in the K trained CNN models (hyper parameter tuning), and performs training again. The learning unit 123 performs the above operation a predetermined number of times (for example, 100 epochs) to improve the accuracy of the output data output by the machine learning model 124.

[0058] Also, in Figure 5, an example is shown in which the CNN model is trained with K=5, but the value of K can be changed arbitrarily depending on the number of datasets used in the training dataset.

[0059] Furthermore, the learning unit 123 may perform K-fold cross-validation on each of the six CNN models described above.

[0060] The inventors also conducted a test using the machine learning model 124 after learning (i.e., the trained CNN model). Specifically, the learning unit 123 input a test dataset (Test) to each of the five trained CNN models (model1-model5) and obtained output data (prediction1-prediction5) output from each of the five trained CNN models. The learning unit 123 then averaged the accuracy rates of the output data output from each of the five trained CNN models to evaluate the accuracy of the CNN models. The evaluation results (i.e., the accuracy of the CNN models) evaluated by the learning unit 123 will be described below with reference to FIG. 6.

[0061] Figure 6 is a graph showing the evaluation results for each of the six CNN models. The graph shown in Figure 6 is a graph showing an ROC (Receiver Operating Characteristic) curve, with the vertical axis representing sensitivity and the horizontal axis representing 1-specificity. Note that the closer the ROC curve is to the point where the horizontal axis represents 0.0 and the vertical axis represents 1.0 (i.e., the upper left of the graph), the higher the accuracy (predictive ability) of the CNN model.

[0062] As shown in Figure 6, ResNet50, DenseNet121, EfficientNetB0, and InceptionV3 output results near the upper left corner of Figure 6, and were found to have relatively higher accuracy than VGG16 and MobileNetV3Large. In particular, DenseNet121 showed an AUC (Area Under the Curve) of 0.788, the highest accuracy among the six CNN models. Note that AUC is the area to the right of the ROC curve, and is an indicator that the closer the area value is to 1.0, the higher the accuracy (predictive ability) of the CNN model.

[0063] [Verification of CNN model] The inventors verified which data of the input X-ray image (i.e., input data) the machine learning model 124 (i.e., CNN model) focuses on when learning. Gradient-weighted Class Activation Mapping (Grad-CAM) was used for the verification. Grad-CAM is a technology that visualizes which part of the input image the CNN model focuses on when making predictions or learning.

[0064] The verification results obtained using Grad-CAM are shown in Figures 7A, 7B, and 7C. In the verification, DenseNet121 was used as the CNN model.

[0065] FIG. 7A shows a first example of verification results obtained using Grad-CAM. FIG. 7B shows a second example of verification results obtained using Grad-CAM. FIG. 7C shows a third example of verification results obtained using Grad-CAM. (a) of FIG. 7A, (a) of FIG. 7B, and (a) of FIG. 7C show X-ray images (i.e., input data) input to DenseNet121. (b) of FIG. 7A, (b) of FIG. 7B, and (b) of FIG. 7C are heat maps superimposed with verification results obtained using Grad-CAM on the input X-ray images. Note that the heat maps shown in (b) of FIG. 7A, (b) of FIG. 7B, and (b) of FIG. 7C are represented by dot density. Specifically, a region with a higher dot density indicates that the CNN model is focusing more on learning, and a region with a lower dot density indicates that the CNN model is focusing less on learning.

[0066] 7A(b), 7B(b), and 7C(b) show heat maps in which dot density is classified into four levels, but heat maps in which dot density is classified into levels other than four may also be shown. Heat maps may also be shown using a continuous change in color tone. For example, areas that the CNN model is learning with more emphasis may be shown in warm colors (e.g., red), and areas that the CNN model is learning with less emphasis may be shown in cool colors (e.g., blue).

[0067] Figures 7A(a), 7B(a), and 7C(a) are X-ray images obtained by imaging a person's chest. The X-ray images shown in Figures 7A(a), 7B(a), and 7C(a) show two black areas (areas surrounded by solid lines) around the chest and a white area (area surrounded by dashed lines) between the two black areas. The lungs are visible in the area surrounded by the solid lines, and the heart is visible in the area surrounded by the dashed line.

[0068] As shown in Figures 7A(b), 7B(b), and 7C(b), the CNN model learned by focusing on the area around the heart and the lungs rather than the heart itself. Generally, doctors diagnose whether or not a patient has heart failure by focusing on factors such as the size of the heart. The verification results shown in Figures 7A(b), 7B(b), and 7C(b), which show a focus different from that of a typical doctor's diagnosis, indicate that there is something unique about the way the CNN model learned.

[0069] From the above verification results, the inventors inferred the following: When the systolic function or diastolic function of the left ventricle declines, blood may stagnate in the lungs, resulting in pulmonary congestion. Generally, when comparing X-ray images of the chest of a person with pulmonary congestion with X-ray images of the chest of a person without disease, the lungs of a person with pulmonary congestion appear whiter than the lungs of a person without disease. From the above, the inventors inferred that the CNN model may be learning by focusing on differences in image shading around the heart and in the lungs, which cannot be distinguished by humans.

[0070] [effect] The cardiac function estimation device 1 according to this embodiment includes a trained machine learning model 124 that outputs output data relating to the contractile function and diastolic function of the left ventricle when an X-ray image (X-ray image for examination) obtained by imaging a person's chest is input, and an estimation unit 122 that inputs the X-ray image (X-ray image for examination) to the machine learning model 124, and the estimation unit 122 identifies and outputs the cardiac function of the person corresponding to the output data output by the machine learning model 124.

[0071] According to this, the cardiac function estimation device 1 identifies a person's cardiac function from an X-ray image (an X-ray image for examination), and therefore can identify people who require medical examination or treatment related to cardiac function without using the test results obtained from an echocardiogram.

[0072] Furthermore, in the cardiac function estimation device 1 according to this embodiment, the output data output by the machine learning model 124 is data indicating one of four patterns, each of which is a pair of a first value 3 indicating whether the contractile function of the left ventricle is normal or abnormal, and a second value 4 indicating whether the diastolic function of the left ventricle is normal or abnormal.

[0073] According to this, the machine learning model 124 outputs output data consisting of a pair of a first value 3 and a second value 4 from an X-ray image (an X-ray image for examination), so the cardiac function estimation device 1 can identify people who require medical treatment or therapy related to cardiac function, regarding the left ventricular contractile function and left ventricular diastolic function, without using the test results obtained from an echocardiogram.

[0074] Furthermore, the cardiac function estimation device 1 according to this embodiment further includes a learning unit 123 that trains the machine learning model 124 using X-ray images (learning X-ray images) as explanatory variables and output data as objective variables.

[0075] According to this, the learning unit 123 trains the machine learning model 124 in advance, so that the machine learning model 124 can output output data when an X-ray image (an X-ray image for examination) is input. Therefore, the cardiac function estimation device 1 can identify people who require medical examination or treatment related to cardiac function without using the examination results obtained by echocardiography.

[0076] Furthermore, the cardiac function estimation method according to this embodiment is a cardiac function estimation method executed by a cardiac function estimation device 1 equipped with a trained machine learning model 124, in which an X-ray image (an X-ray image for examination) obtained by imaging a person's chest is input into the machine learning model 124 (S3), and the machine learning model 124 identifies (S5) and outputs (S6) the cardiac function of the person corresponding to output data relating to the left ventricular contractile function and the left ventricular diastolic function, which data is output based on the X-ray image (an X-ray image for examination).

[0077] According to this method for estimating cardiac function, a person's cardiac function is identified from an X-ray image (an X-ray image for examination), and therefore it is possible to identify people who require medical examination or treatment related to cardiac function without using the test results obtained from an echocardiogram.

[0078] Furthermore, the program according to this embodiment causes a computer to execute the steps included in the cardiac function estimation method described above.

[0079] Such a program provides the same effects as the cardiac function estimation method according to the present embodiment.

[0080] (Other embodiments) Although the cardiac function estimation device and the like according to the present disclosure have been described based on the above-mentioned embodiment, they are not limited to the above-mentioned embodiment. As long as they do not deviate from the spirit of the present disclosure, various modifications conceivable by those skilled in the art to the above-mentioned embodiment and configurations constructed by combining components of different embodiments may also be included within the scope of one or more aspects.

[0081] For example, the machine learning model may store a plurality of CNN models, and the machine learning model may switch between CNN models in response to an input from a user.

[0082] Furthermore, the first value and the second value constituting the output data may be represented by any value other than 0 or 1.

[0083] In the above embodiments, each component may be configured with dedicated hardware, or may be realized by executing a software program suitable for each component. Each component may be realized by a program execution unit such as a CPU or processor reading and executing a software program recorded on a recording medium such as a hard disk or semiconductor memory.

[0084] In the above-described embodiments, some or all of the functions of the components may be realized by a processor such as a CPU executing a program.

[0085] In the above-described embodiment, the processing performed by a specific processing unit may be performed by another processing unit. The order of multiple processing operations may be changed, or multiple processing operations may be performed in parallel.

[0086] Furthermore, some or all of the components constituting each of the above devices may be configured as an IC card or a standalone module that can be attached to or detached from each device. The above IC card or module is a computer system composed of a microprocessor, ROM, RAM, etc. The above IC card or module may include a super multi-function LSI. The above IC card or module achieves its functions when the microprocessor operates in accordance with a computer program. The IC card or module may be tamper-resistant.

[0087] Furthermore, the general or specific aspects of the present invention may be realized as a system, an apparatus, a method, an integrated circuit, a computer program, or a computer-readable recording medium such as a CD-ROM. Alternatively, the present invention may be realized as a computer-readable non-transitory recording medium such as an optical disk, a HDD, or a semiconductor memory on which the computer program is stored. Furthermore, the present invention may be realized as any combination of a system, a method, an integrated circuit, a computer program, and a recording medium. [Industrial Applicability]

[0088] A cardiac function estimation device according to the present disclosure is useful, for example, as a screening device for estimating a person's cardiac function. [Explanation of symbols]

[0089] 1. Cardiac function estimation device 11 Acquisition Department 12 Control Unit 121 Image conversion unit 122 Estimation Department 123 Learning Department 124 Machine Learning Models 13 Storage section 14 Display section 2 Pixel Values 3 First Value 4 Second Value

Claims

1. a trained machine learning model that receives an X-ray image of a person's chest and outputs output data relating to the systolic function and diastolic function of the left ventricle; an estimation unit that inputs the X-ray image into the machine learning model; The estimation unit identifies and outputs a cardiac function of the person corresponding to the output data output by the machine learning model. Cardiac function estimation device.

2. The output data output by the machine learning model is data indicating one of four patterns, each of which is a pair of a first value indicating whether the systolic function of the left ventricle is normal or abnormal and a second value indicating whether the diastolic function of the left ventricle is normal or abnormal. The cardiac function estimation device according to claim 1 .

3. Further, a learning unit is provided that trains the machine learning model using the X-ray image as an explanatory variable and the output data as a target variable. The cardiac function estimation device according to claim 1 or 2.

4. A cardiac function estimation method executed by a cardiac function estimation device equipped with a trained machine learning model, comprising: An X-ray image obtained by imaging a person's chest is input into the machine learning model; Identifying and outputting cardiac function of the person corresponding to output data regarding left ventricular systolic function and left ventricular diastolic function output by the machine learning model based on the X-ray images. Cardiac function estimation method.

5. A program for causing a computer to execute the steps included in the cardiac function estimation method according to claim 4.

Images