Method and apparatus for training and predicting of body composition prediction model

Abstract

A training and prediction device for a body composition prediction model according to one embodiment comprises, as an electronic device, a memory in which commands are stored; and a processor that executes the commands. When the commands are executed by the processor, the electronic device may perform a training and prediction method for a body composition prediction model comprising: a step of acquiring a plurality of medical images of a patient’s chest taken based on X-rays and body composition data individually corresponding to the plurality of medical images; a step of preprocessing the plurality of medical images by restricting the pixel value of a pixel corresponding to a non-interest portion among the pixels in each of the plurality of medical images to a reference value; a step of generating temporary output data based on a machine learning-based body composition prediction model from the preprocessed plurality of medical images; a step of training the machine learning-based body composition prediction model based on true value data determined from body composition data corresponding to the medical image input to the machine learning-based body composition prediction model and the temporary output data; and a step of predicting a body composition result corresponding to the input medical image from the input medical image using the trained machine learning-based body composition prediction model.

Description

Technology Field

[0001] Below, a technique for training a body composition prediction model and predicting body composition using the trained body composition prediction model is provided. Background Technology

[0002] Body composition prediction involves forecasting specific body composition values ​​(e.g., body fat mass, skeletal muscle mass, body water content, etc.) and can play a crucial role in various aspects, such as maintaining health, preventing disease, and evaluating treatment effectiveness. Accordingly, various technologies for predicting body composition values ​​have recently been proposed. For instance, Bioelectric Impedance Analysis (BIA) is a method that measures internal body water content through electrical resistance to predict skeletal muscle mass and body fat mass. While BIA is widely used due to its simple method of predicting body composition, it has the disadvantage that prediction accuracy can only be maintained if strict conditions, such as restricting water intake and maintaining a stable physical state, are strictly adhered to. Prior art literature

[65535] (Patent Document 0001) US 20220335605 A1 means of solving the problem

[0003] A method for training and predicting a body composition prediction model according to one embodiment may include: a step of acquiring a plurality of medical images of a patient’s chest taken based on X-rays and body composition data individually corresponding to the plurality of medical images, which is performed by a processor; a step of preprocessing the plurality of medical images by restricting the pixel value of a pixel corresponding to a non-interest portion among the pixels in each of the plurality of medical images to a reference value; a step of generating temporary output data based on a machine learning-based body composition prediction model from the preprocessed plurality of medical images; a step of training the machine learning-based body composition prediction model based on true value data determined from the body composition data corresponding to the medical image input to the machine learning-based body composition prediction model and the temporary output data; and a step of predicting a body composition result corresponding to the input medical image from the input medical image using the trained machine learning-based body composition prediction model.

[0004] In a training and prediction method for a body composition prediction model according to one embodiment, the output data of the trained machine learning-based body composition prediction model may include probabilities for each label indicating body composition values ​​in which a predetermined body composition range is discretely divided into predetermined units.

[0005] In a training and prediction method for a body composition prediction model according to one embodiment, the interval between the labels may be 0.1 kg.

[0006] In a training and prediction method for a body composition prediction model according to one embodiment, the training step may include a step of updating parameters of the machine learning-based body composition prediction model based on objective function values ​​between the true value data and the temporary output data.

[0007] In a training and prediction method for a body composition prediction model according to one embodiment, the step of updating the parameters may include the step of calculating the objective function value based on the difference between the mean of the temporary output data and the true value data, and the variance of the temporary output data.

[0008] In a method for training and predicting a body composition prediction model according to one embodiment, the step of preprocessing the plurality of medical images may include the step of determining the portion corresponding to 0.1% of the total pixels in order of decreasing brightness, starting from the pixel having the brightest pixel value among the total pixels in each of the plurality of medical images, as the portion of non-interest.

[0009] In a method for training and predicting a body composition prediction model according to one embodiment, the step of acquiring the body composition data includes acquiring additional true value data including at least one of the patient's age, gender, height, and weight together with the body composition data; the step of generating the temporary output data includes generating the temporary output data including additional prediction data including at least one of the patient's age, gender, height, and weight together with the body composition value predicted for the patient based on the machine learning-based body composition prediction model; and the step of training the machine learning-based body composition prediction model may include training the machine learning-based body composition prediction model based on the temporary output data, the true value data, and the additional true value data.

[0010] A training and prediction device for a body composition prediction model according to one embodiment comprises, as an electronic device, a memory in which commands are stored; and a processor that executes the commands. When the commands are executed by the processor, the electronic device may perform a training and prediction method for a body composition prediction model comprising: a step of acquiring a plurality of medical images of a patient’s chest taken based on X-rays and body composition data individually corresponding to the plurality of medical images; a step of preprocessing the plurality of medical images by restricting the pixel value of a pixel corresponding to a non-interest portion among the pixels in each of the plurality of medical images to a reference value; a step of generating temporary output data based on a machine learning-based body composition prediction model from the preprocessed plurality of medical images; a step of training the machine learning-based body composition prediction model based on true value data determined from body composition data corresponding to the medical image input to the machine learning-based body composition prediction model and the temporary output data; and a step of predicting a body composition result corresponding to the input medical image from the input medical image using the trained machine learning-based body composition prediction model.

[0011] In a training and prediction device for a body composition prediction model according to one embodiment, the output data of the trained machine learning-based body composition prediction model may include probabilities for each label indicating body composition values ​​in which a predetermined body composition range is discretely divided into predetermined units.

[0012] In a training and prediction device for a body composition prediction model according to one embodiment, the interval between the labels may be 0.1 kg.

[0013] In a training and prediction device for a body composition prediction model according to one embodiment, the training step may include a step of updating parameters of the machine learning-based body composition prediction model based on objective function values ​​between the true value data and the temporary output data.

[0014] In a training and prediction device for a body composition prediction model according to one embodiment, the step of updating the parameter may include the step of calculating the objective function value based on the difference between the mean of the temporary output data and the true value data, and the variance of the temporary output data.

[0015] In a training and prediction device for a body composition prediction model according to one embodiment, the step of preprocessing the plurality of medical images may include the step of determining the portion corresponding to 0.1% of the total pixels in order of decreasing brightness, starting from the pixel having the brightest pixel value among the total pixels in each of the plurality of medical images, as the portion of non-interest.

[0016] In a training and prediction device for a body composition prediction model according to one embodiment, the step of acquiring the body composition data includes acquiring additional true value data including at least one of the patient's age, gender, height, and weight together with the body composition data; the step of generating the temporary output data includes generating the temporary output data including additional prediction data including at least one of the patient's age, gender, height, and weight together with the predicted body composition value for the patient based on the machine learning-based body composition prediction model; and the step of training the machine learning-based body composition prediction model may include training the machine learning-based body composition prediction model based on the temporary output data, the true value data, and the additional true value data. Brief explanation of the drawing

[0017] FIG. 1 illustrates an overview of a body composition prediction system according to one embodiment. FIG. 2 illustrates an example of a neural network according to one embodiment. FIG. 3 is a flowchart illustrating the training and prediction operations of a body composition prediction model according to one embodiment. FIG. 4 illustrates the training operation of a body composition prediction model according to one embodiment. FIG. 5 illustrates the output of body composition data according to one embodiment. FIG. 6 illustrates an example of an objective function according to one embodiment. FIGS. 7 to 9 illustrate a pretreatment according to one embodiment. FIG. 10 is a block diagram illustrating a training and prediction device for a body composition prediction model according to one embodiment. Specific details for implementing the invention

[0018] Specific structural or functional descriptions of the embodiments are disclosed for illustrative purposes only and may be modified and implemented in various forms. Accordingly, actual implementations are not limited to the specific embodiments disclosed, and the scope of this specification includes modifications, equivalents, or substitutions included in the technical concept described by the embodiments.

[0019] Terms such as "first" or "second" may be used to describe various components, but these terms should be interpreted solely for the purpose of distinguishing one component from another. For example, the first component may be named the second component, and similarly, the second component may be named the first component.

[0020] The singular expression includes the plural expression unless the context clearly indicates otherwise. In this specification, terms such as "comprising" or "having" are intended to specify the existence of the described features, numbers, steps, actions, components, parts, or combinations thereof, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0021] In this document, each of the phrases such as "A or B", "at least one of A and B", "at least one of A or B", "A, B or C", "at least one of A, B and C", and "at least one of A, B, or C" may include any one of the items listed together in the corresponding phrase, or all possible combinations thereof.

[0022] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by those skilled in the art. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an ideal or overly formal sense unless explicitly defined in this specification.

[0023] Hereinafter, embodiments will be described in detail with reference to the attached drawings. In the description with reference to the attached drawings, identical components are given the same reference numeral regardless of the drawing number, and redundant descriptions thereof will be omitted.

[0025] FIG. 1 illustrates an overview of a body composition prediction system according to one embodiment.

[0026] A body composition prediction system according to one embodiment can predict body composition values ​​through medical imaging. For example, the body composition prediction system can predict body composition values ​​without computed tomography (CT) and dual-energy X-ray absorptiometry (DEXA).

[0027] An electronic device according to one embodiment can predict body composition data (130) and additional data (140) corresponding to a medical image (110) from a medical image (110) using a body composition prediction model (120).

[0028] A medical image (110) is an image that visually represents at least some organs or tissues inside the body, captured by medical imaging equipment, and may include, for example, an image in which at least some part of the human body is taken based on X-rays. In this specification, a chest X-ray image is primarily described as an example of a medical image (110). A chest X-ray image may be a plain X-ray image of the chest. A plain X-ray image is an image obtained by radiation generated through an X-ray generator passing through a target part of the human body and projecting it onto a flat X-ray detector, and the internal structure of the target part of the human body (e.g., chest) may appear in the plain X-ray image.

[0029] The body composition prediction model (120) may be a machine learning model designed and trained to predict body composition data (130) indicating body composition values ​​from a medical image (110). The body composition prediction model (120) may be trained by an electronic device based on a plurality of medical images (110) and true value data indicating body composition values ​​that correspond individually to the plurality of medical images (110). The body composition prediction model (120) may be a machine learning model designed and trained to predict not only the body composition data (130) but also data (e.g., additional data (140)) indicating at least one factor among age, gender, height, and weight of the human body captured in the input medical image (110). A body composition prediction model (120) can be trained by an electronic device based on true value data indicating a plurality of medical images (110) and body composition values ​​individually corresponding to the plurality of medical images (110), as well as data indicating at least one factor among age, gender, height, and weight individually corresponding to the plurality of medical images (110) (e.g., additional true value data). The body composition prediction model (120) may include a neural network. An exemplary neural network structure is shown in FIG. 3 below, and the training of the body composition prediction model (120) is described later in FIG. 3 to FIG. 6 below.

[0030] Body composition data (130) may be data indicating body composition values ​​(e.g., body fat mass, skeletal muscle mass) of a human body appearing in a medical image (110). Body composition data (130) may be expressed as probability values ​​that the human body appearing in the medical image (110) has specific body composition values. Body composition data (130) may include at least one of data indicating skeletal muscle mass and data indicating body fat mass. For example, body composition data (130) may include probabilities for each label indicating body composition values ​​in which a predetermined body composition range is discretely divided into predetermined units. For example, the body composition range (e.g., skeletal muscle range) of body composition data (130) indicating skeletal muscle mass (e.g., skeletal muscle data) may be 20 kg or more and 40 kg or less, and the unit may be 0.1 kg. Labels for skeletal muscle data may indicate skeletal muscle masses of 20.0 kg, 20.1 kg, 20.2 kg, 39.9 kg, and 40.0 kg. If the probability of a label indicating 30 kg among these labels is 0.9, it can be interpreted that there is a 90% probability that the skeletal muscle mass of the human body corresponding to the medical image (110) is 30 kg. Body composition data (130) indicating body fat mass (e.g., body fat data) may be configured in a similar manner. Labels for body fat data may indicate corresponding body fat masses. For reference, while the present specification mainly describes skeletal muscle mass and body fat mass as examples of body composition, it is not limited thereto, and various body compositions including human body water content and body fat mass may be configured as body composition data. Additionally, the body composition data (130) may include the mean and variance of the probabilities of the labels in addition to the probability values ​​for each of the aforementioned labels. According to one embodiment, the body composition range (e.g., skeletal muscle range, body fat range) of the body composition data can be determined based on the maximum and minimum values ​​of body composition values ​​(e.g., skeletal muscle mass, body fat mass) that individually correspond to a plurality of medical images (110).For example, if the maximum value of skeletal muscle mass corresponding individually to multiple medical images (110) is 40 kg and the minimum value is 20 kg, the skeletal muscle range can be determined to be between 20 kg and 40 kg. Additionally, for example, if the maximum value of skeletal muscle mass corresponding individually to multiple medical images (110) is 40 kg and the minimum value is 20 kg, the skeletal muscle range can be determined to be between 18 kg and 42 kg by setting the range between the maximum and minimum values ​​to 2 kg outside, which is 10% of the difference between the maximum and minimum values. A body composition range can be set similarly for body fat mass. Although the above methods for setting the body composition range have been described, they are not limited to the above methods, and the body composition range can be determined by various methods based on the body composition values ​​corresponding individually to multiple medical images (110). An example of body composition data (130) is described later in Fig. 5 below.

[0031] A body composition prediction model (120) according to one embodiment may be designed and trained to predict not only body composition data (130) but also additional data (140). The additional data (140) may indicate the value of a factor that has a correlation with body composition. For example, the factor of the additional data (140) may include at least one of the age, gender, height, and weight of a human body captured in a medical image (110). The present specification primarily describes an example in which the body composition prediction model (120) is designed to predict body composition data (130) and additional data (140) together, but is not limited thereto.

[0032] An electronic device according to one embodiment can predict more accurate body composition values ​​from medical images with a more economical and lower exposure dose.

[0033] For reference, in this specification, an electronic device that trains a body composition prediction model (120) and performs a prediction using the body composition prediction model (120) may also be referred to as a training and prediction device for a body composition prediction model.

[0034] FIG. 2 illustrates an example of a neural network according to one embodiment.

[0035] In a neural network (200) (e.g., a body composition prediction model) according to one embodiment, when input data (240) is input to an input layer (210), the input data (240) is sequentially propagated from the input layer (210) through a hidden layer (220) to an output layer (230), and output data (270) can be output from the output layer (230).

[0036] A neural network (200) may include multiple layers composed of multiple nodes (250). Additionally, the neural network (200) may include edges (260) that connect multiple nodes (250) included in each of the multiple layers to each node included in another layer, and connection weights assigned to each edge (260). For example, the neural network (200) may represent a model utilizing multiple nodes (250) connected by edges (260). For example, the neural network (200) may include an input layer (210), a hidden layer (220), and an output layer (230). For reference, for convenience of explanation, FIG. 2 mainly describes examples of layers, such as fully connected layers (FC layers), but is not limited thereto, and convolutional layers, residual layers, and various layer structures may be used.

[0037] The input layer (210) is a layer that receives input data (240) in the neural network (200) and can be composed of multiple nodes (250). The multiple nodes (250) included in the input layer (210) can be connected to each node included in the neighboring hidden layer (220) by a connection line (260). The hidden layer (220) is a layer located between the input layer (210) and the output layer (230) in the neural network (200) and can be composed of multiple nodes (250). The number of hidden layers (220) included in the neural network (200) may be one or more. The multiple nodes (250) included in the hidden layer (220) can be connected to each node included in the neighboring input layer (210), output layer (230), and other hidden layers (220) by a connection line (260). The output layer (230) is a layer that outputs output data (270) from the neural network (200) and can be composed of a plurality of nodes (250). The plurality of nodes (250) included in the output layer (230) can be connected to each node included in the neighboring hidden layer (220) by a connection line (260).

[0038] Input data (240) is data input to the neural network (200) and may include, for example, pixel values ​​of a medical image (e.g., a preprocessed chest X-ray image). Input data (240) may be input to a plurality of nodes (250) of the input layer (210). Each of the plurality of nodes (250) included in the plurality of layers may have a node value calculated based on a weighted sum of values ​​propagated from the previous layer and connection weights, a specified bias, and an activation function. Values ​​calculated based on the node values ​​of the plurality of nodes (250) included in any layer and the connection weights of the connection lines (260) may be propagated to each node included in another layer connected through the connection lines (260).

[0039] The node values ​​of a plurality of nodes (250) included in the output layer (230) can be output as output data (270). As will be described later, a body composition prediction model including a neural network (200) can be trained to provide output data (270) (e.g., body composition data) corresponding to body composition from a medical image. The output data (270) may include not only data indicating body composition (e.g., body composition data) but also data indicating at least one factor among age, gender, height, and weight (e.g., additional data). For example, the output data (270) may include additional data predicted from a medical image based on the body composition prediction model.

[0040] Multiple nodes (250) are elements constituting each layer of the neural network (200) and may include biases, which are values ​​uniquely assigned to each of the multiple nodes (250). Multiple nodes (250) may be connected to each node of an adjacent layer by a connection line (260). A connection line (260) is a line connecting multiple nodes (250), and the connection line (260) may include a connection weight, which is a value uniquely assigned to each connection line (260). For reference, in this specification, the parameters of the neural network (200) may include connection weights and biases.

[0041] A neural network (200) (e.g., a body composition prediction model) according to one embodiment may be trained by an electronic device based on data for training (e.g., a plurality of medical images and body composition data individually corresponding to the plurality of medical images). Training operations will be described later in FIGS. 3 to 6 below.

[0042] FIG. 3 is a flowchart illustrating the training and prediction operations of a body composition prediction model according to one embodiment.

[0043] An electronic device according to one embodiment can train a body composition prediction model through a training operation (300) and predict a body composition result based on the trained body composition prediction model from an input medical image through a prediction operation (305). The training operation (300) may include, for example, an operation to update parameters included in the body composition prediction model so that the body composition prediction model can predict data indicating a body composition value corresponding to an input medical image (e.g., a chest X-ray image). The prediction operation (305) may include, for example, an operation to predict data indicating a body composition value corresponding to an input medical image.

[0044] In step (310), the electronic device can acquire multiple medical images. Additionally, the electronic device can acquire body composition data that corresponds individually to the acquired multiple medical images. For example, the electronic device can acquire multiple medical images and acquire data (e.g., true value data) indicating at least one of the skeletal muscle mass and body fat mass of the human body captured in each of the acquired multiple medical images. For example, the electronic device can receive body composition values ​​(e.g., skeletal muscle mass, body fat mass) corresponding to the human body captured in any medical image. The body composition values ​​can be measured by separate equipment for measuring body composition. The medical images of the same person and the measured body composition values ​​can be mapped by an expert. For example, the chest X-ray images of the same person taken through a health checkup and at least one of the measured skeletal muscle mass and body fat mass can be mapped by an expert. The electronic device can acquire the medical images and body composition values ​​mapped to the same person. However, this is not limited thereto, and mapping between medical images and body composition values ​​may be performed automatically through the integration of medical imaging equipment and body composition measurement equipment, or through a single device (or single system) that performs medical image capture and body composition measurement. Additionally, the electronic device may acquire data indicating at least one factor among age, gender, height, and weight as additional data for the same person.

[0045] In step (320), the electronic device can preprocess multiple medical images by limiting the pixel value of a pixel corresponding to a non-interest part among the pixels in each of the multiple medical images to a reference value. A pixel may be a unit that makes up an image. The pixels included in the image may each include pixel values ​​indicating their respective brightness levels. The preprocessing may include, for example, a technique (e.g., max-min normalization) that transforms the data used for training (e.g., multiple medical images) to be suitable for training a neural network (e.g., a body composition prediction model). Examples of preprocessing are described later in FIGS. 7 to 9.

[0046] In step (330), the electronic device may obtain temporary output data based on a body composition prediction model from a plurality of preprocessed medical images. The temporary output data is output data obtained based on a body composition prediction model from a plurality of medical images input for training in the training operation (300), and may include predicted body composition data. The predicted body composition data may include body composition data obtained based on a body composition prediction model from a plurality of medical images input for training. The temporary output data may include additional prediction data in addition to the predicted body composition data. The additional prediction data is additional data obtained based on a body composition prediction model from a plurality of medical images input for training, and may indicate the value of a factor having a correlation with body composition. For example, the factor of the additional prediction data may include at least one of the age, gender, height, and weight of the human body captured in each of the acquired plurality of medical images.

[0047] In step (340), the electronic device can train a body composition prediction model based on true value data and temporary output data. The true value data may include data determined from body composition data that individually corresponds to a plurality of medical images acquired in step (310). An electronic device according to one embodiment may be designed to train a body composition prediction model based not only on true value data and temporary output data but also on additional true value data. The additional true value data may include additional data determined from the value of a factor that has a correlation with body composition. For example, the factor of the additional true value data may include at least one of the age, gender, height, and weight of the human body captured in the medical image acquired in step (310). The description of the training operation (300) of the body composition prediction model is described later in FIG. 4 below.

[0048] In step (350), the electronic device can predict a body composition result corresponding to the input medical image from the input medical image based on a trained body composition prediction model. The input medical image may include a medical image used in the prediction operation (305). For example, the input medical image may be a chest X-ray image input to the electronic device including the body composition prediction model in the prediction operation (305). The body composition result is body composition data predicted from the input medical image by the electronic device using the trained body composition prediction model, and may indicate a body composition value of the human body captured in the input medical image. For example, the body composition result may indicate at least one of the skeletal muscle mass and body fat mass of the human body captured in the input medical image. The body composition result may include probabilities for each label indicating body composition values ​​that are discretely divided into predetermined units within a predetermined body composition range. The body composition result may also include the mean and variance of the probabilities of the labels. In step (350) according to one embodiment, the electronic device may be designed to predict not only the body composition result corresponding to the input medical image but also additional results corresponding to the input medical image from the input medical image using a trained body composition prediction model. The additional results are additional data predicted from the input medical image by the electronic device using the trained body composition prediction model, and may indicate the value of a factor having a correlation with the body composition. For example, the factor of the additional results may include at least one of the age, gender, height, and weight of the human body captured in the input medical image.

[0049] FIG. 4 illustrates the training operation of a body composition prediction model according to one embodiment.

[0050] In a training operation according to one embodiment, an electronic device may acquire a plurality of medical images (110) and preprocess the acquired plurality of medical images (110). The electronic device may input the preprocessed plurality of medical images (110) into an electronic device including a body composition prediction model (120), and may predict temporary output data (430) (e.g., predicted body composition data (440), predicted additional data (450)) from the input plurality of medical images (110) based on the body composition prediction model (120). The electronic device may train the body composition prediction model (120) based on true value data (410) and temporary output data (430).

[0051] An electronic device according to one embodiment can train a body composition prediction model (120) through supervised learning. Supervised learning may include a technique of inputting a medical image (110) into the body composition prediction model (120) and updating the parameters of the body composition prediction model (120) so that body composition data corresponding to true value data (410) determined from body composition data mapped to the medical image (110) is predicted.

[0052] A training operation according to one embodiment can be performed by updating the parameters of a body composition prediction model (120) through supervised learning based on true value data (410) and temporary output data (430) (e.g., predicted body composition data (440)). The updating of parameters can be performed, for example, based on an objective function between the true value data (410) and the temporary output data (430). The updating of parameters based on the objective function can be performed, for example, through a backpropagation method that sequentially updates the parameters of each layer from the output layer to the input layer of the neural network so that the value of the objective function is minimized. Additionally, the training of the neural network can be performed by alternating backpropagation with forward propagation, in which input data for training is input into the input layer of the neural network, and each node value from each node of the input layer is sequentially computed and propagated through connecting lines to the output layer, and each node value of the nodes of the output layer is output as temporary output data (430). Although a training method based on supervised learning and backpropagation has been described, the training method is not limited to this, and the body composition prediction model (120) can be trained through other methods such as unsupervised learning. A training operation according to one embodiment can be performed based on true value data (410) and temporary output data (430), as well as true value additional data (420). An example of an objective function is described later in FIG. 6 below.

[0053] FIG. 5 illustrates the output of body composition data according to one embodiment.

[0054] The output of body composition data according to one embodiment may include probabilities for each label indicating body composition values ​​that are discretely divided into predetermined units of a predetermined body composition range, as illustrated in FIG. 5.

[0055] In FIG. 5, the horizontal axis represents body composition values ​​(e.g., skeletal muscle mass (510), body fat mass), and the vertical axis may represent output values ​​(520) of body composition data (e.g., probability for each label) for each label indicating body composition values. Body composition values ​​may include body composition values ​​of the human body (e.g., skeletal muscle mass (510), body fat mass) captured in medical images (e.g., chest X-ray images). For example, as shown on the horizontal axis of FIG. 5, the body composition range (e.g., skeletal muscle range) of body composition data (e.g., skeletal muscle data) indicating body composition values ​​(e.g., skeletal muscle mass (510)) may be 20 kg or more and 40 kg or less, and the unit may be 0.1 kg. The labels of body composition data may indicate body composition values ​​of 20.0 kg, 20.1 kg, 20.2 kg, to 39.9 kg, and 40.0 kg.

[0056] The output value (520) may include the value that the output data of the body composition prediction model has for each label. The output value (520) may include, for example, a value from 0 to 1. The output value (520) may be interpreted as representing the probability that the body composition of the human body captured in the medical image has a value indicated by a specific label (e.g., body composition value). For example, as shown in FIG. 5, if the output value (520) of the label indicating 20 kg among the labels indicating skeletal muscle mass (510) indicates 0.1, it may be interpreted as the probability that the skeletal muscle mass of the human body captured in the medical image is 20 kg being 10%. The output of body composition data may be configured similarly for body fat mass. The output values ​​(520) may be interpreted as a distribution of probabilities, as shown by the dotted line in FIG. 5.

[0057] FIG. 6 illustrates an example of an objective function according to one embodiment.

[0058] An objective function according to one embodiment may include a function expression for the error between true value data (410) and temporary output data (e.g., predicted body composition data (440)). The objective function may include a loss function. For example, the objective function may include a cross-entropy function between true value data (410) and temporary output data (e.g., predicted body composition data (440)). The cross-entropy function may be a function that serves as an indicator for evaluating the distance between the distributions of two probabilities.

[0059] An objective function according to one embodiment can be calculated based on mean loss (610) and variance loss (620), as illustrated in FIG. 6.

[0060] The mean loss (610) may be the difference between the mean (630) of the predicted body composition data (440) and the true value data (410). For example, the probabilities for each label of the temporary output data (e.g., predicted body composition data (440)) output from a medical image based on a body composition prediction model may be interpreted (e.g., approximated) as the probability values ​​for each interval of the probability density function (PDF). The mean (630) of the predicted body composition data (440) is the mean of the continuous random variable and can be calculated using the probability density function interpreted above. For example, the mean (630) of the predicted body composition data (440) may be represented as the sum of the products of each label and the probability for each label for all labels of the predicted body composition data (440). However, it is not limited to this, and the mean (630) of the predicted body composition data (440) may be the label having the highest probability value. The true value of the true value data (410) is true value data mapped to a medical image, and may be an actual body composition value measured for a human body corresponding to the time of capture of the medical image.

[0061] The variance loss (620) may include the variance of the temporary output data (e.g., predicted body composition data (440)). For example, the probabilities for each label of the temporary output data (e.g., predicted body composition data (440)) output from a medical image based on a body composition prediction model may be interpreted (e.g., approximated) as probability values ​​for each interval of the probability density function. The variance of the predicted body composition data (440) is the variance of the corresponding continuous random variable and can be calculated using the probability density function interpreted above. For example, the variance of the predicted body composition data (440) may be expressed as the sum of the products of the squared deviations of each label and the probabilities for each label for all labels of the predicted body composition data (440). The squared deviations of the labels may be expressed, for example, as the squared difference between the mean of the labels and each label.

[0062] For reference, the predicted body composition data (440) in FIG. 6 of this specification is shown in the form of a normal distribution for convenience of explanation, but is not limited thereto.

[0063] FIGS. 7 to 9 illustrate a pretreatment according to one embodiment.

[0064] Figure 7 is a histogram illustrating preprocessing according to min-max normalization.

[0065] In the step of preprocessing a plurality of medical images of a training operation according to one embodiment, the electronic device may preprocess a plurality of medical images according to min-max normalization. Min-max normalization may include a preprocessing technique that subtracts the minimum value of the pixel values ​​of the image from each pixel value of the image and then divides by the difference between the maximum value and the minimum value of the pixel values ​​of the image. For example, min-max normalization for each pixel of a medical image may be expressed as shown in the following Equation 1.

[0066]

[0067] In the aforementioned mathematical formula 1, Xi represents the i-th value among the pixel values ​​of the original image, and Zi may represent the i-th value among the pixel values ​​of the image preprocessed by min-max normalization. min(X) may represent the minimum value among the pixel values ​​of the original image, and max(X) may represent the maximum value among the pixel values ​​of the original image. The pixel values ​​transformed by the aforementioned mathematical formula 1 may have only values ​​from 0 to 1. For example, as shown in the histogram of FIG. 7, the pixel values ​​of the image (710) before min-max normalization may have values ​​from 2000 to 8000, whereas the pixel values ​​of the image (720) after min-max normalization may have only values ​​from 0 to 1. In FIG. 7, the horizontal axis may represent the pixel value of each pixel in the image, and the vertical axis may represent the number of pixels having the corresponding pixel value in each image.

[0068] Figures 8 and 9 illustrate preprocessing based on clipping.

[0069] In the step of preprocessing a plurality of medical images of a training motion according to one embodiment, the electronic device may preprocess the plurality of medical images by clipping. The electronic device may clip by limiting the brightness (e.g., pixel value) of a non-interest region to the maximum brightness of a region of interest.

[0070] Clipping may be an action that restricts the pixel value of a pixel corresponding to a non-interest area among the pixels of an image to a reference value. Through clipping, a neural network can be trained more effectively. A non-interest area may be a part of the image that contains information with a lower level of interest compared to other parts. For example, a non-interest area may include at least one of a part marked with an L / R mark in a chest X-ray image and a part where an artifact is captured. The L / R mark may include an L or R-shaped mark that is conveniently displayed to indicate the left and right sides of the human body in a chest X-ray image. For example, the L mark may be displayed in an L shape, as shown in the upper right corner of the 0.1% clipped image (810) of FIG. 9. An artifact may include an object with low radiation penetrability, such as an implant or precious metal jewelry.

[0071] Pixel values ​​in an X-ray image are expressed as absorbance; for example, as absorbance increases (e.g., as transmittance decreases), pixel values ​​increase and are represented as bright pixels, and as absorbance decreases, pixel values ​​decrease and are represented as dark pixels. The aforementioned non-interest area is a part with high absorbance other than the human skeleton and organs, and may have brighter pixel values ​​than most of the human body. In an X-ray image containing a non-interest area, pixel values ​​of the interest area (e.g., parts corresponding to the human organs and skeleton) may be relatively underestimated. For example, in a chest X-ray image containing 10,000 pixels, if the pixel values ​​of the non-interest area containing 100 pixels are 2,000 to 4,000 and the pixel values ​​of the interest area containing the remaining 9,900 pixels are 0 to 1,000, the pixel values ​​of the interest area having values ​​smaller than the pixel values ​​of the non-interest area may be relatively underestimated during the training process.

[0072] The clipping threshold value may include the brightest pixel value among all pixels. For example, as shown in the histogram of FIG. 8, when the brightest pixel value among all pixels of the image is 1, the clipping threshold value may be 1.

[0073] In clipping according to one embodiment, a portion corresponding to 0.1% of the total pixels can be determined as a non-interest portion in order of decreasing brightness starting from the pixel having the brightest pixel value among all the pixels of the image, and such clipping may be referred to as 0.1% clipping in this specification. For example, if the image contains 10,000 pixels, the 10 pixels starting from the pixel having the brightest pixel value can be normalized to have the same intensity value (e.g., 1, which is the maximum intensity value).

[0074] A medical image preprocessed according to max-max normalization according to one embodiment may be clipped by 0.1% or 1% as shown in the histogram of FIG. 8, and the reference value may be the brightness value of the brightest pixel among all pixels of the image. In FIG. 8, the horizontal axis may be the brightness value of the pixel preprocessed according to max-max normalization, and the vertical axis may be the number of pixels in the image having that brightness value.

[0075] According to one embodiment, a 0.1% clipped image (810) can be displayed such that anatomical structures within the image are better distinguished compared to a 1% clipped image (820), as illustrated in FIG. 9. In 1% clipping, each pixel value of an area having a high pixel value among anatomical structures of the human body (e.g., organs and skeletons of the human body) within the area of ​​interest, as well as non-interest areas, can be fixed to a reference value. In one embodiment, clipping can be performed on 0.1% to 1% of the total pixels of the image so that the pixel values ​​of the areas corresponding to anatomical structures within the image are not fixed to a reference value.

[0076] Although a preprocessing method for the step of preprocessing multiple medical images of abnormal training movements has been described, preprocessing is not limited to preprocessing based on min-max normalization and clipping, and multiple medical images can be preprocessed by various preprocessing techniques.

[0077] FIG. 10 is a block diagram of a training and prediction device for a body composition prediction model according to one embodiment.

[0078] An electronic device (1010) according to one embodiment may include a processor (1020) and a memory (1030). The electronic device (1010) may perform training and prediction operations based on a body composition prediction model (e.g., a machine learning model).

[0079] The memory (1030) is a device capable of storing data temporarily or permanently, and can store, for example, input data and output data. For example, the memory (1030) can store at least one of medical images, body composition data and additional data. Additionally, the memory (1030) can store a body composition prediction model (e.g., a machine learning model) and can store instructions for the implementation, training, and inference using the body composition prediction model.

[0080] The processor (1020) can execute commands to perform operations on input data and output the calculated data. The processor (1020) can predict body composition data and additional data from medical images based on, for example, a body composition prediction model (e.g., a machine learning model). Additionally, the processor (1020) can train a body composition prediction model based on temporary output data output from multiple medical images and true value data corresponding individually to multiple medical images based on the body composition prediction model.

[0081] The embodiments described above may be implemented as hardware components, software components, and / or combinations of hardware and software components. For example, the devices, methods, and components described in the embodiments may be implemented using a general-purpose computer or a special-purpose computer, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing unit may execute an operating system (OS) and software applications executed on said operating system. Additionally, the processing unit may access, store, manipulate, process, and generate data in response to the execution of the software. For ease of understanding, the processing unit may be described as being used as a single unit, but those skilled in the art will understand that the processing unit may include multiple processing elements and / or multiple types of processing elements. For example, the processing unit may include multiple processors or one processor and one controller. In addition, other processing configurations, such as parallel processors, are also possible.

[0082] Software may include computer programs, code, instructions, or a combination of one or more of these, and may configure a processing unit to operate as desired or command the processing unit independently or collectively. Software and / or data may be permanently or temporarily embodied in any type of machine, component, physical device, virtual equipment, computer storage medium or device, or transmitted signal wave so as to be interpreted by the processing unit or to provide instructions or data to the processing unit. Software may be distributed over networked computer systems and may be stored or executed in a distributed manner. Software and data may be stored on computer-readable recording media.

[0083] The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., either alone or in combination, and the program instructions recorded on the medium may be those specifically designed and configured for the embodiment or those known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc.

[0084] The hardware device described above may be configured to operate as one or more software modules to perform the operation of the embodiment, and vice versa.

[0085] Although the embodiments have been described above with reference to the limited drawings, those skilled in the art can apply various technical modifications and variations based thereon. For example, suitable results may be achieved even if the described techniques are performed in a different order than described, and / or if the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from described, or replaced or substituted by other components or equivalents.

[0086] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims set forth below.

Claims

Claim 1 A method for training and predicting a body composition prediction model performed by a processor, comprising: acquiring a plurality of medical images of a patient’s chest taken based on X-rays and body composition data individually corresponding to said plurality of medical images; preprocessing said plurality of medical images by restricting the pixel values ​​of pixels corresponding to non-interest regions, which correspond to parts with high absorbance outside the human skeleton and organs among the pixels in each of said plurality of medical images, to reference values; generating temporary output data based on a machine learning-based body composition prediction model from said preprocessed plurality of medical images; and training said machine learning-based body composition prediction model based on true value data determined from body composition data corresponding to the medical image input to said machine learning-based body composition prediction model and said temporary output data. A method for training and predicting a body composition prediction model, comprising the step of predicting a body composition result corresponding to an input medical image from an input medical image using the above-mentioned machine learning-based body composition prediction model, wherein the step of preprocessing the plurality of medical images includes: determining a portion corresponding to 0.1% of the total pixels in each of the plurality of medical images in order of decreasing brightness starting from the pixel having the brightest pixel value among the total pixels; and limiting the pixel value of the pixel corresponding to the portion of interest to a reference value corresponding to the maximum brightness of the portion of interest corresponding to the skeleton and organs of the human body, wherein the portion of interest corresponds to at least one of an implant with low radiation permeability or a left / right marker. Claim 2 A method for training and predicting a body composition prediction model according to claim 1, wherein the output data of the trained machine learning-based body composition prediction model includes probabilities for each label indicating body composition values ​​in which a predetermined body composition range is discretely divided into predetermined units. Claim 3 A method for training and predicting a body composition prediction model, wherein, in paragraph 2, the interval between the labels is 0.1 kg. Claim 4 A method for training and predicting a body composition prediction model according to claim 1, wherein the training step comprises updating the parameters of the machine learning-based body composition prediction model based on the objective function value between the true value data and the temporary output data. Claim 5 A method for training and predicting a body composition prediction model according to claim 4, wherein the step of updating the parameters comprises the step of calculating the objective function value based on the difference between the mean of the temporary output data and the true value data and the variance of the temporary output data. Claim 6 delete Claim 7 A method for training and predicting a body composition prediction model according to claim 1, wherein the step of acquiring the body composition data includes acquiring additional true value data including at least one of the patient's age, gender, height, and weight together with the body composition data; the step of generating the temporary output data includes generating the temporary output data including additional prediction data including at least one of the patient's age, gender, height, and weight together with predicted body composition data indicating a predicted body composition value for the patient based on the machine learning-based body composition prediction model; and the step of training the machine learning-based body composition prediction model includes training the machine learning-based body composition prediction model based on the temporary output data, the true value data, and the additional true value data. Claim 8 A computer-readable recording medium having a program recorded thereon for executing any one of the methods of paragraphs 1 through 5 and paragraph 7. Claim 9 An electronic device comprises: a memory in which instructions are stored; and a processor that executes said instructions, wherein, when the instructions are executed by said processor, the electronic device comprises: a step of acquiring a plurality of medical images of a patient’s chest based on X-rays and body composition data individually corresponding to said plurality of medical images; a step of preprocessing said plurality of medical images by restricting the pixel value of a pixel corresponding to a non-interest region corresponding to a part of the pixels in each of said plurality of medical images that has high absorbance other than the skeletal and organ parts of the human body to a reference value; a step of generating temporary output data based on a machine learning-based body composition prediction model from said preprocessed plurality of medical images; and a step of training said machine learning-based body composition prediction model based on true value data determined from body composition data corresponding to the medical image input to said machine learning-based body composition prediction model and said temporary output data. A training and prediction method for a body composition prediction model comprising the step of predicting a body composition result corresponding to an input medical image from an input medical image using the above-mentioned trained machine learning-based body composition prediction model, wherein the step of preprocessing the plurality of medical images comprises: determining a portion corresponding to 0.1% of the total pixels in each of the plurality of medical images in order of decreasing brightness starting from the pixel having the brightest pixel value among the total pixels as the portion of non-interest; and limiting the pixel value of the pixel corresponding to the portion of non-interest to the reference value corresponding to the maximum brightness of the portion of interest corresponding to the skeleton and organs of the human body, wherein the portion of non-interest corresponds to at least one of an implant with low radiation transmittance or a left and right marker. Claim 10 In claim 9, the output data of the trained machine learning-based body composition prediction model comprises probabilities for each label indicating body composition values ​​in which a predetermined body composition range is discretely divided into predetermined units, a training and prediction device for a body composition prediction model. Claim 11 A training and prediction device for a body composition prediction model, wherein the interval between the above labels is 0.1 kg in item 10. Claim 12 In claim 9, the training step comprises a step of updating parameters of the machine learning-based body composition prediction model based on the objective function value between the true value data and the temporary output data, in a training and prediction device for a body composition prediction model. Claim 13 A training and prediction device for a body composition prediction model according to claim 12, wherein the step of updating the parameter comprises the step of calculating the objective function value based on the difference between the mean of the temporary output data and the true value data and the variance of the temporary output data. Claim 14 delete Claim 15 In claim 9, the step of acquiring the body composition data includes acquiring additional true value data including at least one of the patient's age, gender, height, and weight together with the body composition data; the step of generating the temporary output data includes generating the temporary output data including additional prediction data including at least one of the patient's age, gender, height, and weight together with predicted body composition data indicating a predicted body composition value for the patient based on the machine learning-based body composition prediction model; and the step of training the machine learning-based body composition prediction model includes training the machine learning-based body composition prediction model based on the temporary output data, the true value data, and the additional true value data.

Images