System and method for diagnosing osteopenia by using x-ray images

Abstract

A system for diagnosing osteopenia according to one embodiment comprises: an artificial intelligence model training device including a first artificial intelligence model training device for training a first artificial intelligence model that classifies an input image as osteoporosis or non-osteoporosis, and a second artificial intelligence model training device for training a second artificial intelligence model that classifies the non-osteoporosis image as normal or osteopenia; and an osteopenia prediction device for primarily classifying a to-be-read image as osteoporosis or non-osteoporosis on the basis of the first artificial intelligence model, and classifying the primarily classified non-osteoporosis image as normal or osteopenia on the basis of the second artificial intelligence model.

Description

Osteopenia Diagnostic System and Diagnostic Method Using X-ray Imaging

[0001] The present disclosure relates to a diagnostic system and method for osteopenia using X-ray images, and more particularly to a diagnostic system and method for osteopenia using whole chest X-ray images and segmented images.

[0002] South Korea entered an aging society in 2018 with the proportion of people aged 65 or older reaching 14%, and is expected to enter a super-aging society in 2025 with over 20%. Consequently, the number of patients with bone diseases is expected to surge, as are the medical and socioeconomic costs associated with these diseases.

[0003] Osteoporosis and osteopenia are representative bone diseases. Osteoporosis refers to a condition in which bone density and strength are reduced, increasing the likelihood of fractures, while osteopenia refers to a pre-stage of osteoporosis where bone density is lower than normal. Although osteoporosis and osteopenia often have no specific symptoms, the risk of secondary fractures and complications increases once a fracture occurs; therefore, it is necessary to prevent the risk of osteoporosis or osteopenia through screening for reduced bone density. Osteoporosis and osteopenia are diagnosed through bone density tests such as Dual-Energy X-ray Absorptiometry (DEXA), and are classified as normal, osteopenia, or osteoporosis based on the bone index (T-score) measured after the test.

[0004] With the recent rapid advancement of image analysis technology utilizing artificial intelligence models, AI models are also being utilized in the diagnosis of bone diseases. U.S. Patent No. 12,033,318 describes a technology for estimating bone density from X-ray images using an AI model. Specifically, according to U.S. Patent No. 12,033,318, human skeletal images are used as training data, and bone mineral density (BMD) measured from DEXA is used as supervised data to train an AI model. Through the trained AI model, continuous bone density or bone indices (such as T-scores) can be directly predicted from chest X-ray images, thereby allowing for the determination of the bone's condition or the presence of bone disease.

[0005] The technology described in U.S. Patent No. 12,033,318 has the advantage of providing detailed diagnostic information, such as the condition of the bone, through continuous predicted values ​​of bone density or bone indicators; however, training a highly accurate artificial intelligence model requires a large volume of medical images from various hospitals (or cohort groups) and paired DEXA extracted label information (bone indicator information such as bone density or T-score).

[0006] Recently, instead of using bone density as a training label, technology is being developed to train an AI model by constructing a training dataset consisting of full X-ray images of healthy individuals and osteoporosis patients along with their corresponding categorical labels (normal / osteoporosis), and then using the trained AI model to estimate osteoporosis from X-ray images. However, technology capable of diagnosing osteopenia from X-ray images using AI models has not yet been developed.

[0007] In order to train an artificial intelligence model using data (imaging data) from patients with osteopenia as training data, in addition to data from normal and osteoporotic patients, a large amount of training data is required. However, there is a problem in that it is not easy to secure the massive amount of medical data necessary for training the AI ​​model, and securing sufficient data incurs high costs.

[0008] The technical problem that the present invention aims to solve is to provide an osteopenia diagnosis system and a diagnosis method that additionally classifies osteopenia through a second artificial intelligence model for X-ray images classified as non-osteoporosis through a first artificial intelligence model.

[0009] In addition, the technical problem that the present invention aims to solve is to provide a diagnostic system and method for osteopenia that can accurately classify a condition as osteoporosis or non-osteoporosis even when the image is incomplete by synthesizing the output results of models learned from a plurality of extracted images using an ensemble algorithm to infer bone-related information.

[0010] An osteopenia diagnosis system according to one embodiment may include: an artificial intelligence model learning device comprising a first artificial intelligence model learning device for learning a first artificial intelligence model that classifies an input image as osteoporosis or non-osteoporosis, and a second artificial intelligence model learning device for learning a second artificial intelligence model that classifies a non-osteoporosis image as normal or osteopenia; and an osteopenia prediction device that classifies an image to be read into osteoporosis or non-osteoporosis based on the first artificial intelligence model, and classifies the first classified non-osteoporosis image into normal or osteopenia based on the second artificial intelligence model.

[0011] The above input image may be a chest X-ray image.

[0012] The above osteopenia prediction device may include a non-osteoporosis classification device that classifies the image to be read into osteoporosis or non-osteoporosis based on the first artificial intelligence model, and an osteopenia classification device that classifies the non-osteoporosis image classified by the non-osteoporosis classification device into osteopenia or normal based on the second artificial intelligence model.

[0013] The first artificial intelligence model includes a plurality of classification models that learn one or more segmented images, each of which is a whole image or a whole image divided by anatomical parts, and the first artificial intelligence model learning device can train a corresponding classification model among the plurality of classification models based on a learning label corresponding to the whole image or the segmented image.

[0014] The above-mentioned second artificial intelligence model learning device can train the second artificial intelligence model based on the entire image of a normal person or a patient with osteopenia and the corresponding learning labels.

[0015] The second artificial intelligence model learning device may include an image preprocessing unit that performs Grayscale Truncation preprocessing on a learning image, which is an entire image of a normal person or a patient with osteopenia; a learning dataset generation unit that generates a learning dataset based on the learning image preprocessed by the image preprocessing unit and a learning label of normal or osteopenia associated with the learning image; and a second artificial intelligence model learning unit that trains the second artificial intelligence model using the learning dataset generated by the learning dataset generation unit.

[0016] The above second artificial intelligence model may be a knowledge distillation model.

[0017] The second artificial intelligence model learning unit comprises: an augmentation unit that performs a first augmentation and a second augmentation stronger than the first augmentation on the training image preprocessed by the image preprocessing unit; a teacher model unit that receives the training image on which the first augmentation has been performed and outputs a first logit value regarding the osteopenia or normal class; a student model unit that receives the training image on which the second augmentation has been performed and outputs a second logit value regarding the osteopenia or normal class; a temperature scaling unit that applies temperature scaling to the first logit value and the second logit value output from the teacher model unit and the student model unit, respectively, and outputs a first corrected logit value and a second corrected logit value, respectively; a probability value conversion unit that performs probability conversion on the first corrected logit value and the second corrected logit value output from the temperature scaling unit, respectively, to output a first probability value and a second probability value, and performs probability conversion on the second logit value output from the student model unit to output a third probability value; and in the probability value conversion unit It may include a loss function calculation unit that calculates a loss function based on the outputted first probability value, second probability value, and third probability value, and a parameter update unit that updates the parameters of the student model unit based on the loss function output from the loss function calculation unit.

[0018] The above loss function calculation unit can calculate a cross entropy loss function based on the third probability value output from the probability value conversion unit, and calculate a Kullback-Leibler divergence loss function based on the first probability value and the second probability value output from the probability value conversion unit.

[0019] The parameter update unit can update the parameters of the student model unit by summing the cross-entropy loss function calculated by the loss function operation unit and the Kullback-Leibler Divergence loss function, and then performing backpropagation through an optimizer.

[0020] The above parameter update unit can calculate the exponential moving average value for the parameters of the teacher model unit and the parameters of the student model unit, and update the student model unit.

[0021] An osteopenia diagnosis system according to another embodiment may include: an artificial intelligence model learning device comprising a first artificial intelligence model learning device for learning a first artificial intelligence model that classifies an input image as osteoporosis or non-osteoporosis, and a second artificial intelligence model learning device for learning a second artificial intelligence model that classifies a non-osteoporosis image as normal, osteopenia, or osteoporosis; and an osteopenia prediction device that classifies an image to be read into osteoporosis or non-osteoporosis based on the first artificial intelligence model, and classifies the first classified non-osteoporosis image into normal, osteopenia, or osteoporosis based on the second artificial intelligence model.

[0022] The above osteopenia prediction device may include a non-osteoporosis classification device that classifies the image to be read into osteoporosis or non-osteoporosis based on the first artificial intelligence model, and an osteopenia classification device that classifies the non-osteoporosis image classified by the non-osteoporosis classification device into normal, osteopenia, or osteoporosis based on the second artificial intelligence model.

[0023] The second artificial intelligence model learning device may include an image preprocessing unit that performs Grayscale Truncation preprocessing on a learning image, which is a whole image of a normal person, a patient with osteopenia, or a patient with osteoporosis; a learning dataset generation unit that generates a learning dataset based on the learning image preprocessed by the image preprocessing unit and learning labels of normal, osteopenia, or osteoporosis associated with the learning image; and a second artificial intelligence model learning unit that trains the second artificial intelligence model using the learning dataset generated by the learning dataset generation unit.

[0024] A first artificial intelligence model learning device receives a source learning image, which is a whole image of a normal person or an osteoporosis patient, and divides the source learning image into images corresponding to each anatomical region; a first image division unit receives a target learning image, which is a whole image of a normal person, an osteopenia patient, or an osteoporosis patient, and divides the target learning image into images corresponding to each anatomical region; a learning dataset generation unit generates a plurality of source learning datasets—each source learning dataset includes the source learning image or the plurality of divided source images divided from the source learning image—and a source learning label labeled as normal or osteoporosis—based on the target learning image and a plurality of divided target images divided from the target learning image—each target learning dataset includes the target learning image or the plurality of divided target images and a target learning label labeled as non-osteoporosis or osteoporosis; and a plurality of source classification model learning units perform the first learning for the corresponding source classification model based on the plurality of source learning datasets. and may include a plurality of target classification model learning units that perform secondary learning on the corresponding target classification model based on the plurality of target learning datasets.

[0025] The first artificial intelligence model learning device may further include a transfer learning unit that performs transfer learning for the corresponding target classification model learning unit based on the hidden layer parameters, which are the learning results of the source classification model learning unit.

[0026] The first image segmentation unit crops and segments the entire image into images corresponding to each anatomical region, and the training image generated by the training dataset generation unit may include the entire chest image and one or more segmented images among the right clavicle-scapula, left clavicle-scapula, cervical spine, and thoracic / lumbar spine.

[0027] The above non-osteoporosis classification device may include: a second image segmentation unit that divides the image to be read into anatomical regions to generate a plurality of segmented images; a bone disease classification inference unit that infers a bone disease classification result for each input image based on a corresponding classification model among the plurality of target classification models, the entire image to be read and the plurality of segmented images divided by the second image segmentation unit; and a combination unit that combines the plurality of bone disease classification results inferred by the bone disease classification inference unit using an ensemble algorithm to diagnose whether the image to be read is an osteoporosis image or a non-osteoporosis image.

[0028] An artificial intelligence model learning device according to one embodiment includes: a first artificial intelligence model learning device for learning a first artificial intelligence model that classifies an input image as osteoporosis or non-osteoporosis; and a second artificial intelligence model learning device for learning a second artificial intelligence model that classifies a non-osteoporosis image as normal or osteopenia.

[0029] The above-mentioned first artificial intelligence model may include a plurality of classification models that each learn one or more segmented images obtained by dividing the entire image into anatomical regions.

[0030] The above-described first artificial intelligence model learning device can train a corresponding classification model among the plurality of classification models based on a learning label corresponding to the entire image or the segmented image.

[0031] The above-mentioned second artificial intelligence model learning device can train the second artificial intelligence model based on the entire image of a normal person or a patient with osteopenia and the corresponding learning labels.

[0032] The second artificial intelligence model learning device may include an image preprocessing unit that performs Grayscale Truncation preprocessing on a learning image, which is an entire image of a normal person or a patient with osteopenia; a learning dataset generation unit that generates a learning dataset based on the learning image preprocessed by the image preprocessing unit and a learning label of normal or osteopenia associated with the learning image; and a second artificial intelligence model learning unit that trains the second artificial intelligence model using the learning dataset generated by the learning dataset generation unit.

[0033] The above second artificial intelligence model may be a knowledge distillation model.

[0034] The above-mentioned second artificial intelligence model learning device can train the second artificial intelligence model to classify as normal, osteopenia, or osteoporosis.

[0035] A method for diagnosing osteopenia according to one embodiment is a method for diagnosing osteopenia from an image to be read, and may include: a step of training a first artificial intelligence model that classifies an input image as osteoporosis or non-osteoporosis; a step of training a second artificial intelligence model that classifies an input non-osteoporosis image as normal or osteopenia; a step of classifying the image to be read into osteoporosis or non-osteoporosis based on the first artificial intelligence model; and a step of classifying the first classified non-osteoporosis image into normal or osteopenia based on the second artificial intelligence model.

[0036] The above-mentioned first artificial intelligence model may include a plurality of classification models that each learn one or more segmented images obtained by dividing the entire image into anatomical regions.

[0037] The step of training the first artificial intelligence model may be characterized by training a corresponding classification model among the plurality of classification models based on a training label corresponding to the entire image or the segmented image.

[0038] The step of training the second artificial intelligence model may include: performing Grayscale Truncation preprocessing on a training image, which is a whole image of a normal person or a patient with osteopenia; generating a training dataset based on the preprocessed training image and a training label of normal or osteopenia associated with the training image; and training the second artificial intelligence model using the training dataset generated by the training dataset generation unit.

[0039] The step of training the second artificial intelligence model comprises: a step of performing a first augmentation and a second augmentation stronger than the first augmentation on the preprocessed training image; a step of receiving the training image on which the first augmentation was performed in the teacher model unit and outputting a first logit value regarding the osteopenia or normal class; a step of receiving the training image on which the second augmentation was performed in the student model unit and outputting a second logit value regarding the osteopenia or normal class; a step of applying temperature scaling to the first logit value and the second logit value output from the teacher model unit and the student model unit, respectively, to output a first corrected logit value and a second corrected logit value, respectively; a step of performing a probability transformation on the first corrected logit value and the second corrected logit value, respectively, to output a first probability value and a second probability value; a step of performing a probability transformation on the second logit value output from the student model unit to output a third probability value; and to the first probability value, the second probability value, and the third probability value It may include a step of calculating a loss function based on the above, and a step of updating the parameters of the student model part based on the above calculated loss function.

[0040] The step of calculating the loss function may include the step of calculating a cross-entropy loss function based on the third probability value, and the step of calculating a Kullback-Leibler divergence loss function based on the first probability value and the second probability value.

[0041] The step of updating the above parameters may include the step of summing the calculated cross-entropy loss function and the Kullback-Leibler Divergence loss function, and then performing backpropagation through an optimizer.

[0042] According to the embodiments, before inferring osteopenia using a second artificial intelligence model specialized for osteopenia classification, osteoporosis is first classified through a first artificial intelligence model, and osteopenia or normality is inferred from images classified as non-osteoporosis excluding osteoporosis, so a decline in the classification performance of osteopenia due to confusion in classifying osteoporosis and osteopenia can be prevented.

[0043] According to the embodiments, the output results of a model trained with a plurality of extracted images are combined with an ensemble algorithm to infer bone-related information, thereby increasing prediction accuracy and providing accurate analysis results even when the images are incomplete.

[0044] FIG. 1 is a diagram showing the configuration of a medical imaging system according to an embodiment of the present invention.

[0045] FIG. 2 is a drawing showing a computing device implementing a bone disease diagnosis system according to an embodiment of the present invention.

[0046] FIG. 3 is a diagram showing a bone disease diagnosis system according to an embodiment of the present invention.

[0047] FIG. 4 is a drawing showing an artificial intelligence model learning device and an osteopenia prediction device according to the first embodiment of the present invention.

[0048] FIG. 5 is a drawing showing a first artificial intelligence model learning device and a non-osteoporosis classification device according to a first embodiment of the present invention.

[0049] FIG. 6 is a drawing showing a first artificial intelligence model learning device according to a first embodiment of the present invention.

[0050] FIG. 7 is a diagram showing an example of a segmented CXR image according to a first embodiment of the present invention.

[0051] FIG. 8 is a diagram showing an example of transfer learning according to the first embodiment of the present invention.

[0052] FIG. 9 is a diagram showing a learning method of a first artificial intelligence model learning device according to a first embodiment of the present invention.

[0053] FIG. 10 is a drawing showing a non-osteoporosis classification device according to a first embodiment of the present invention.

[0054] FIG. 11 is a diagram showing a coupling process performed in a coupling part according to a first embodiment of the present invention.

[0055] FIG. 12 is a drawing showing a second artificial intelligence model learning device according to a first embodiment of the present invention.

[0056] FIG. 13 is a drawing showing a second artificial intelligence model learning unit according to a first embodiment of the present invention.

[0057] FIG. 14 is a diagram showing a second artificial intelligence model learning method according to the first embodiment of the present invention.

[0058] FIG. 15 is a drawing showing an osteopenia classification device according to the first embodiment of the present invention.

[0059] FIG. 16 is a drawing showing an artificial intelligence model learning device and an osteopenia prediction device according to a second embodiment of the present invention.

[0060] Embodiments of the present invention are described below with reference to the attached drawings so that those skilled in the art can easily implement them. However, the present invention may be embodied in various different forms and is not limited to the embodiments described herein. Furthermore, in order to clearly explain the present invention in the drawings, parts unrelated to the explanation have been omitted, and similar parts throughout the specification are denoted by similar reference numerals.

[0061] Throughout the specification, when a part is described as "including" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components.

[0062] Additionally, terms such as "...part," "...unit," and "...module" described in the specification refer to a unit that processes at least one function or operation, and this may be implemented in hardware, software, or a combination of hardware and software.

[0063] In this specification, "transmission or provision" may include not only direct transmission or provision but also indirect transmission or provision through another device or by using an alternative route.

[0064] Expressions described in the singular in this specification may be interpreted as singular or plural unless explicit expressions such as "one" or "single" are used.

[0065] In describing embodiments of the present invention, if it is determined that a detailed description of related known technology may obscure the essence of the present invention, such detailed description is omitted.

[0066] Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the attached drawings. In the drawings, the same reference numerals are used to indicate the same or similar components.

[0067] FIG. 1 is a diagram showing the configuration of a medical imaging system according to an embodiment of the present invention.

[0068] Referring to FIG. 1, the medical imaging system may include at least one user terminal (30), a medical imaging storage device (20), and a bone disease diagnosis system (10).

[0069] The user terminal (30) is composed of hardware and software that install programs executed by a processor and provide a computing environment and a network environment for performing operations according to an embodiment of the present invention. The user terminal (30) can be implemented in various types, such as, for example, a computing device within a workstation, a mobile device, etc.

[0070] The user terminal (30) can display medical image-related data stored in the medical image storage device (20) in conjunction with the medical image storage device (20). Additionally, the user terminal (30) can display bone disease results diagnosed by the bone disease diagnosis system (10). Such functions can be provided through a viewer, which is a dedicated program installed on the user terminal (30).

[0071] The medical image storage device (20) can store and manage captured medical images. The medical image storage device (20) can also store and manage analysis results for the medical images. The medical image storage device (20) may include a database (PACS database) of a medical image storage and transmission system (PACS, Picture Archiving and Communications System; hereinafter also referred to as 'PACS'). The medical image storage device (20) can store data according to a designated data format.

[0072] The medical image storage device (20) can obtain bone disease diagnostic information (e.g., osteoporosis / osteopenia / normal) from the bone disease diagnostic system (10). The medical images stored in the medical image storage device (20) may be X-ray images, MRI (magnetic resonance imaging) images, ultrasound images, CT (computed tomography) images, Digital MMG (Mammography) images, DBT (Digital breast tomosynthesis) images, etc.

[0073] In the embodiments of the present invention, a chest X-ray image (hereinafter also simply referred to as an "X-ray image") is described as an example of a medical image; however, the present invention is not limited thereto and may be applied according to the type of medical image.

[0074] A bone disease diagnosis system (10) analyzes a chest X-ray (hereinafter also referred to as 'CXR') image using a first artificial intelligence model (non-osteoporosis classification model) and a second artificial intelligence model (osteopenia classification model), and diagnoses whether there is a bone disease (e.g., whether there is osteopenia) based on the CXR analysis results.

[0075] A bone disease diagnosis system (10) according to an embodiment of the present invention can classify an input CXR image into osteoporosis or non-osteoporosis using a first artificial intelligence model, and classify a non-osteoporosis image classified by the first artificial intelligence model into osteopenia or normal using a second artificial intelligence model.

[0076] According to an embodiment of the present invention, the first artificial intelligence model may have a plurality of classification models specialized for the entire CXR image and a plurality of segmented CXR images, respectively. Specifically, according to an embodiment of the present invention, the input entire CXR image may be divided into a plurality of segmented CXR images reflecting anatomical features, and then the entire image and each segmented image may be classified as osteoporosis or non-osteoporosis based on the corresponding classification model of the first artificial intelligence model.

[0077] An artificial intelligence model according to an embodiment of the present invention is generated to perform medical inference from input medical images, and the model structure, training data configuration, training method, and medical inference target can be designed in various ways.

[0078] A bone disease diagnosis system (10) or a detailed device within the bone disease diagnosis system according to an embodiment of the present invention may be implemented as a computing device as shown in FIG. 2.

[0079] Referring to FIG. 2, a bone disease diagnostic system or a detailed device (computing device, 10) within the bone disease diagnostic system may include one or more processors (11), a memory (13) for loading a program executed by the processor (11), a storage (15) for storing the program and various data, a communication interface (17), and a bus (19) connecting them. In addition, the bone disease diagnostic system (10) may include various additional components.

[0080] The program may include instructions that cause the processor (11) to perform methods / operations according to various embodiments of the present invention when loaded into memory (13). That is, the processor (11) may perform methods / operations according to various embodiments of the present invention by executing the instructions. The program is composed of a series of computer-readable instructions grouped by function and indicates that it is executed by the processor.

[0081] The processor (11) controls the overall operation of each component of the computing device (10). The processor (11) may be composed of a single physical entity, but may also be composed of multiple entities. A processor (11) composed of multiple entities may process a single execution element by dividing it or process multiple execution elements by dividing them.

[0082] The processor (11) may be configured to include at least one of a CPU (Central Processing Unit), MPU (Micro Processor Unit), MCU (Micro Controller Unit), GPU (Graphic Processing Unit), or any form of processor well known in the art of the present invention. Additionally, the processor (11) may perform operations for at least one application or program for executing a method / operation according to various embodiments of the present invention.

[0083] Memory (13) stores various data, commands and / or information. Memory (13) may load one or more programs from storage (15) to execute a method / operation according to various embodiments of the present invention. Memory (13) may be implemented as volatile memory such as RAM, but the technical scope of the present invention is not limited thereto.

[0084] Storage (15) can store programs non-temporarily. Storage (15) may be configured to include non-volatile memory such as ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), flash memory, a hard disk, a removable disk, or any form of computer-readable recording medium well known in the art to which the present invention belongs.

[0085] According to an embodiment of the present invention, a storage may store a first artificial intelligence model (composed of a plurality of classification models) and a second artificial intelligence model, as well as learned parameters (weights or biases) of each artificial intelligence model. In this case, the first and second artificial intelligence models may be learning models having a neural network structure that learns the relationship between an input image and a label.

[0086] The communication interface (17) supports wired and wireless internet communication of the computing device (10). Additionally, the communication interface (17) may support various communication methods other than internet communication. To this end, the communication interface (17) may be configured to include a communication module well known in the technical field of the present invention.

[0087] The bus (19) provides communication functions between components of the computing device (10). The bus (19) can be implemented as various types of buses, such as an address bus, a data bus, and a control bus.

[0088] A computer program includes instructions executed by a processor (11) and is stored in a non-transitory computer-readable storage medium, wherein the instructions cause the processor (11) to execute an operation according to an embodiment of the present invention. The computer program may be downloaded over a network or sold as a product.

[0089] A computer program according to an embodiment of the present invention may include instructions for collecting a dataset for pre-training an artificial intelligence model and training the artificial intelligence model using the collected dataset. The computer program may include instructions for pre-processing input data and training the artificial intelligence model using the pre-processed input data.

[0090] The functional block (indicated as "...part") of the detailed device of the bone disease diagnosis system (10) described below can be implemented by the cooperation of the processor (11), which is a detailed configuration of the computing device shown in FIG. 2, and the program included in the memory (13) (or storage (15)).

[0091] Hereinafter, a bone disease diagnosis system (10) according to the first embodiment of the present invention will be described with reference to FIGS. 3 to 10.

[0092] FIG. 3 is a drawing showing a bone disease diagnosis system (10) according to an embodiment of the present invention.

[0093] Referring to FIG. 3, a bone disease diagnosis system (10) according to an embodiment of the present invention may include an artificial intelligence model learning device (100) and an osteopenia prediction device (200).

[0094] The artificial intelligence model learning device (100) includes a first artificial intelligence model that classifies an input CXR image as osteoporosis or non-osteoporosis, and a second artificial intelligence model that classifies a non-osteoporosis image as osteopenia or normal.

[0095] The osteopenia prediction device (200) classifies a CXR image to be read into osteoporosis or non-osteoporosis based on a first artificial intelligence model that has been trained by an artificial intelligence model learning device (100), and classifies the non-osteoporosis image classified in the first stage into normal or osteopenia based on a second artificial intelligence model that has been trained by an artificial intelligence model learning device (100).

[0096] That is, the osteopenia prediction device (200) outputs a prediction result of ‘osteoporosis’ when the CXR image to be read is classified as osteoporosis by the first artificial intelligence model of the artificial intelligence model learning device (100), and when the CXR image to be read is classified as non-osteoporosis by the first artificial intelligence model, it outputs a prediction result classified as ‘normal’ or ‘osteopenia’ based on the second artificial intelligence model that has been trained by the artificial intelligence model learning device (100) on the image classified as non-osteoporosis.

[0097] According to the embodiment, the first artificial intelligence model and the second artificial intelligence model may be included in the osteopenia prediction device (200).

[0098] FIG. 4 is a diagram showing the detailed configuration of an artificial intelligence model learning device (100) and an osteopenia prediction device (200) according to the first embodiment of the present invention.

[0099] Referring to FIG. 4, an artificial intelligence model learning device (100) according to the first embodiment of the present invention may include a first artificial intelligence model learning device (110) and a second artificial intelligence model learning device (120).

[0100] The first artificial intelligence model learning device (110) trains the first artificial intelligence model, which includes a plurality of classification models. Each classification model may be an artificial intelligence model corresponding to a whole CXR image (hereinafter also briefly referred to as "whole image") or a plurality of segmented CXR images (hereinafter also briefly referred to as "segmented images") that are segmented corresponding to each anatomical region (e.g., whole chest, left clavicle-scapula, right clavicle-scapula, cervical spine, thoracic / lumbar spine).

[0101] According to the first embodiment of the present invention, the first artificial intelligence model learning device (110) receives the entire CXR image (entire image) of a normal person, a patient with osteopenia, and a patient with osteoporosis as a learning image.

[0102] Specifically, the first artificial intelligence model learning device (110) receives the entire image of a normal person, a patient with osteopenia, or a patient with osteoporosis, and the corresponding learning label (osteoporosis / non-osteoporosis) as learning data. Here, the learning label of 'non-osteoporosis' is a learning label corresponding to the images of a normal person and osteopenia as a classification that is not osteoporosis, and the learning label of 'osteoporosis' is a learning label corresponding to the images of osteoporosis.

[0103] As described below, the first artificial intelligence model learning device (110) divides the entire input image into a plurality of segmented images corresponding to anatomical regions, generates a learning dataset based on the entire image or segmented images and the input learning labels (osteoporosis / non-osteoporosis), and then trains a corresponding classification model using each learning dataset.

[0104] The second artificial intelligence model learning device (120) generates a learning dataset based on the entire CXR image (entire image) of a normal person or a patient with osteopenia and the corresponding learning label (normal / osteopenia), and then trains the second artificial intelligence model using each learning dataset.

[0105] Thus, according to the first embodiment of the present invention, the first artificial intelligence model learning device (110) for learning osteoporosis or non-osteoporosis uses the entire image and the segmented image as learning data, whereas the second artificial intelligence model learning device (120) for learning osteopenia or normal uses only the entire image as learning data.

[0106] Since the difference between normal and osteopenia is smaller than the difference between non-osteoporosis and osteoporosis, there are limitations in precisely distinguishing even minute differences between normal and osteopenia based solely on information from specific areas derived from segmented images. Therefore, the second artificial intelligence model learning device (120) according to the first embodiment of the present invention utilizes the entire image as training data and considers the subtle contrast of bone brightness appearing throughout the image, the bone quality pattern of the entire body, and the overall conditioning of the image according to shooting conditions.

[0107] The osteopenia prediction device (200) according to the first embodiment of the present invention may include a non-osteoporosis classification device (210) and an osteopenia classification device (220).

[0108] The non-osteoporosis classification device (210) classifies the CXR image to be read into osteoporosis or non-osteoporosis based on the first artificial intelligence model that has been trained by the first artificial intelligence model learning device (110). When the CXR image to be read is classified as osteoporosis by the first artificial intelligence model of the artificial intelligence model learning device (100), the non-osteoporosis classification device (210) outputs a prediction result as 'osteoporosis', and when the CXR image to be read is classified as non-osteoporosis by the first artificial intelligence model, it transmits the image classified as non-osteoporosis to the osteopenia classification device (220).

[0109] The osteopenia classification device (220) classifies non-osteoporosis images classified by the non-osteoporosis classification device (210) as osteopenia or normal based on the second artificial intelligence model that has been trained by the second artificial intelligence model learning device (120), and outputs the classification result as a prediction result.

[0110] Next, with reference to FIG. 5, a first artificial intelligence model learning device (110) and a non-osteoporosis classification device (210) according to a first embodiment of the present invention will be described in detail.

[0111] The first artificial intelligence model learning device (110) may include a source classification model group (111) consisting of a plurality of source classification models (artificial intelligence models) and a target classification model group (112) consisting of a plurality of target classification models.

[0112] Each source classification model and each target classification model may be an artificial intelligence model corresponding to the entire CXR image or a plurality of segmented CXR images in which the entire CXR image is segmented corresponding to each anatomical region.

[0113] The first artificial intelligence model learning device (110) receives the entire CXR image (entire image) of a normal person, a patient with osteopenia, or a patient with osteoporosis as a learning image.

[0114] As described below, the first artificial intelligence model learning device (110) learns a source classification model group (111) based on learning images and learning labels (normal / osteoporosis) of normal people and osteoporosis patients, which are relatively easy to learn, and then learns the hidden layer parameters (weight or bias) that are the learning results of the source classification model group (111) into a target classification model group (112), and then learns a target classification model group (112) based on learning images and learning labels (non-osteoporosis / osteoporosis) of normal people, osteopenia, and osteoporosis patients, which are difficult to learn.

[0115] Specifically, the first artificial intelligence model learning device (110) according to the first embodiment of the present invention receives the entire CXR image (entire image) of a normal person or an osteoporosis patient and the corresponding learning label (normal / osteoporosis) as source learning data for learning a source classification model group (111). In addition, the first artificial intelligence model learning device (110) receives the entire CXR image (entire image) of a normal person, an osteopenia patient, or an osteoporosis patient and the corresponding learning label (non-osteoporosis / osteoporosis) as target learning data for learning a target classification model group (112). That is, the first artificial intelligence model learning device (110) according to the first embodiment of the present invention receives learning labels (non-osteoporosis / osteoporosis) classified into two categories for the entire CXR image of a normal person, an osteopenia patient, or an osteoporosis patient as target learning labels for learning a target classification model group (112).

[0116] Thus, according to the first embodiment of the present invention, since the first artificial intelligence model learning device (110) learns the target classification model group (112) through target learning labels (non-osteoporosis / osteoporosis) divided into two categories, there is an advantage that the target classification model group (112) can be sufficiently learned even with a relatively small number of images of normal people, osteopenia patients, and osteoporosis patients.

[0117] The source classification model and target classification model used in the first embodiment of the present invention may include a deep neural network structure. Deep neural network algorithms such as Deep Neural Network (DNN), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Restricted Boltzmann Machine (RBM), Deep Belief Network (DBM), Q Network, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Transformer, and Capsule Network may be used, but the present invention is not limited thereto and other artificial intelligence models may be used.

[0118] In this case, according to the first embodiment of the present invention, a plurality of source classification models and target classification models included in the source classification model group and the target classification model, respectively, may use an artificial intelligence model having the same neural network structure. However, the present invention is not limited thereto, and the source classification model group and the target classification model group may use classification models having different neural network structures.

[0119] Referring to FIG. 5, the non-osteoporosis classification device (210) receives a CXR image to be read (hereinafter also referred to as a 'target image') and classifies the received target image as osteoporosis or non-osteoporosis based on a target classification model group (112) that has been trained by the first artificial intelligence model learning device (110). Specifically, the non-osteoporosis classification device (210) divides the input target image into multiple segmented images corresponding to anatomical regions, and classifies the entire image or the segmented images as osteoporosis or non-osteoporosis based on the target classification model that has been trained for each of the multiple target classification models. Then, the non-osteoporosis classification device (210) finally classifies the inferred multiple bone disease classification results as osteoporosis or non-osteoporosis using an ensemble algorithm.

[0120] FIG. 6 is a diagram showing the detailed configuration of a first artificial intelligence model learning device (110) according to a first embodiment of the present invention.

[0121] Referring to FIG. 6, the first artificial intelligence model learning device (110) according to the first embodiment of the present invention may include an image segmentation unit (1110), a learning dataset generation unit (1120), a plurality of source classification model learning units (1130), a plurality of target classification model learning units (1140), and a transfer learning unit (1150).

[0122] The image segmentation unit (1110) receives a source learning image, which is a complete CXR image of a normal person and an osteoporosis patient, and a source learning label (normal / osteoporosis), and segments the source learning image into images corresponding to each anatomical region. Additionally, the image segmentation unit (1110) receives a target learning image, which is a complete CXR image of a normal person, an osteopenia patient, and an osteoporosis patient, and a target learning label (non-osteoporosis / osteoporosis) corresponding thereto, and segments the target learning image into images corresponding to each anatomical region.

[0123] FIG. 7 is a diagram showing an example of a CXR image divided by an image division unit (1110) and an entire CXR image according to a first embodiment of the present invention.

[0124] Referring to FIG. 7, the image segmentation unit (1110) according to the first embodiment of the present invention crops the entire CXR image by anatomical region to generate segmented CXR images. Specifically, the image segmentation unit (1110) crops the entire CXR image by anatomical region to generate segmented images corresponding to the entire chest image (E1), right clavicle-scapula (E2), left clavicle-scapula (E3), cervical spine (E4), and thoracic / lumbar spine (E5). As such, according to the first embodiment of the present invention, since a separate artificial intelligence model that detects bone regions or performs segmentation in the entire CXR image is not used, and the original CXR image is divided equally and the fragmented images containing each anatomical region are used as respective training images, there is an advantage of saving data preprocessing time and computing resources used in the classification model. In addition, while conventional methods reduce a high-resolution original image to a 1k (1024 pixel) level before inputting it, according to the first embodiment of the present invention, segmented images are generated by anatomical region without significantly reducing the resolution of the original image, thereby preventing degradation of the original image's resolution and consequently improving the accuracy of learning and classification. Furthermore, by cropping the entire image by anatomical region, the bone region can be intensively trained.

[0125] However, the present invention is not limited to the method of cropping by equally dividing the entire image into images corresponding to each anatomical region, and various image segmentation techniques, such as the use of a segmentation model, may be used in the present invention.

[0126] In addition, in the first embodiment of the present invention, a whole chest image (E1), a right clavicle-scapula (E2), a left clavicle-scapula (E3), a cervical spine (E4), and a thoracic / lumbar spine (E5) were used as learning images, but the present invention is not limited thereto and may use a whole CXR image and images corresponding to some anatomical regions (e.g., a right clavicle-scapula image and a left clavicle-scapula image), or may use only segmented images instead of the whole image.

[0127] The learning dataset generation unit (1120) generates multiple source learning datasets based on the entire CXR image input as a source learning image, the associated source learning label (normal / osteoporosis), and multiple segmented images segmented by the image segmentation unit (1110). Additionally, the learning dataset generation unit (1120) generates multiple target learning datasets based on the entire CXR image input as a target learning image, the associated target learning label (non-osteoporosis / osteoporosis), and multiple segmented images segmented by the image segmentation unit (310).

[0128] The source classification model training unit (1130) trains a source classification model group (111) based on source training images and source training labels (normal / osteoporosis) of normal people and osteoporosis patients, which are relatively easy to train. The source classification model training unit (1130) includes a plurality of source classification model training units (1130-1, 1130-2, ..., 1130_n) that train a plurality of source training datasets, respectively. For example, the first source classification model training unit (1130_1) trains a first source classification model using the entire source training image and source training label generated by the training dataset generation unit (1120). The second source classification model training unit (1130_2) trains a second source classification model (artificial intelligence model) using the first segmented image and source training label generated by the training dataset generation unit (1120).

[0129] The transfer learning unit (1150) transfers the hidden layer parameters, which are the learning results of the source classification model group (111), to the target classification model group (112).

[0130] FIG. 8 is a diagram illustrating that the transfer learning unit (1150) transfers the hidden layer parameters, which are the learning results of the source classification model group (111), to the target classification model group (112).

[0131] Referring to FIG. 8, the transfer learning unit (1150) transfers at least some of the parameters of the source classification model to the corresponding parameters of the target classification model. Specifically, the transfer learning unit (1150) transfers at least some of the parameters of the first source classification model to the parameters of the first target classification model and transfers at least some of the parameters of the second source classification model to the parameters of the second target classification model. In the same way, the transfer learning unit (1150) transfers at least some of the parameters of the nth source classification model to the parameters of the nth target classification model.

[0132] Referring again to FIG. 6, the target classification model training unit (1140) trains the target classification model group (112), in which the hidden layer parameters, which are the training results of the source classification model group (111), have been transferred and learned, based on target training images and target training labels (non-osteoporosis / osteoporosis) of normal people, patients with osteopenia, and patients with osteoporosis.

[0133] The target classification model training unit (1140) includes a plurality of target classification model training units (1140-1, 1140-2, …, 1140_n) that each train a plurality of target training datasets. For example, the first target classification model training unit (1140_1) trains the first target classification model using the entire target training image and target training label generated by the training dataset generation unit (1120). The second target classification model training unit (1140_2) trains the second target classification model using the first segmented image and target training label generated by the training dataset generation unit (1120).

[0134] Next, with reference to FIG. 9, a classification model learning method of the first artificial intelligence model learning device (110) according to the first embodiment of the present invention will be described.

[0135] The first artificial intelligence model learning device (110) receives source learning images / source learning labels and target learning images / target learning labels as learning data. (S110) That is, the first artificial intelligence model learning device (110) receives source learning images, which are the entire CXR images of a normal person and an osteoporosis patient, and the corresponding source learning labels (normal / osteoporosis), and receives target learning images, which are the entire CXR images of a normal person, an osteopenia patient, and an osteoporosis patient, and the corresponding target learning labels (non-osteoporosis / osteoporosis).

[0136] Subsequently, the image segmentation unit (1110) of the first artificial intelligence model learning device (110) divides the received source learning image and target learning image into a plurality of segmented images corresponding to anatomical regions, respectively. (S120) According to the first embodiment of the present invention, the image segmentation unit (310) crops the entire input CXR image by anatomical region to generate segmented CXR images.

[0137] The training dataset generation unit (1120) generates a plurality of source training datasets based on the entire CXR image input as a source training image, the corresponding source training label (normal / osteoporosis), and a plurality of segmented images segmented by the image segmentation unit (1110). (S130)

[0138] The source classification model learning unit (1130) learns the corresponding source classification model by using multiple source learning datasets, each consisting of training images of normal people and osteoporosis patients that are relatively easy to learn and source learning labels (normal / osteoporosis), as training data for the corresponding source classification model. (S140)

[0139] The training dataset generation unit (1120) generates a plurality of target training datasets based on the entire CXR image input as a target training image, the corresponding target training label (non-osteoporosis / osteoporosis), and a plurality of segmented images segmented by the image segmentation unit (1110). (S150)

[0140] The transfer learning unit (1150) performs transfer learning on the target classification model based on the source classification model that has been trained. (S160) That is, the transfer learning unit (1150) performs transfer learning on the corresponding target classification model based on the artificial intelligence neural network parameters of the source classification model that has been trained.

[0141] The target classification model learning unit (1140) learns the corresponding target classification model by using multiple target learning datasets, each consisting of target learning images of normal people, osteopenia patients, and osteoporosis patients, and target learning labels (non-osteoporosis / osteoporosis), as training data for the corresponding target classification model. (S170)

[0142] The target classification model training unit (1140) stores the training parameters (weights, biases, etc.), which are the training results of the target classification model based on the target training dataset, in memory or storage. (S180)

[0143] Next, with reference to FIG. 10, a non-osteoporosis classification device (210) according to the first embodiment of the present invention will be described.

[0144] FIG. 10 is a drawing showing the detailed configuration of a non-osteoporosis classification device (210) according to the first embodiment of the present invention.

[0145] Referring to FIG. 10, a non-osteoporosis classification device (210) according to the first embodiment of the present invention may include an image segmentation unit (211), a bone disease classification inference unit (212), and a combination unit (213).

[0146] The image segmentation unit (211) receives a CXR image to be read (hereinafter also referred to as a 'target image') and segments the entire target image into images corresponding to each anatomical part. At this time, the image segmentation unit (211) of the non-osteoporosis classification device (210) can segment the entire image into images corresponding to each anatomical part using the same segmentation method as the image segmentation unit (1110) of the first artificial intelligence model learning device (110).

[0147] The bone disease classification inference unit (212) infers a bone disease classification result for each image (whether the entire image or the segmented image) based on the entire target image and the multiple segmented images segmented by the image segmentation unit (211), and the corresponding target classification model among the multiple target classification model groups (112) that have completed learning performed by the first artificial intelligence model learning device (110). At this time, the classification result inferred by the bone disease classification inference unit (212) may be a logit or a probability value.

[0148] The combining unit (213) combines multiple bone disease classification results inferred by the bone disease classification inference unit (212) using an ensemble algorithm to diagnose whether there is a final bone disease (osteoporosis / non-osteoporosis).

[0149] FIG. 11 is a diagram showing the joining process performed in the joining part (213) according to the first embodiment of the present invention.

[0150] Referring to FIG. 11, a combining unit (213) can combine (synthesize) multiple bone disease classification results using an ensemble algorithm. For example, the combining unit (213) can average logit or probability values, which are bone disease classification result values ​​inferred from each classification model, and determine whether there is a bone disease (osteoporosis / non-osteoporosis) by comparing the average value with a reference value. In this case, according to the first embodiment of the present invention, when combining multiple bone disease classification results, the combining unit (213) can determine whether there is a bone disease (osteoporosis / non-osteoporosis) by excluding one or more results and calculating the average value only for the remaining ones. Meanwhile, according to the first embodiment of the present invention, various techniques other than averaging can be applied as ensemble techniques. For example, any one of voting, bagging, boosting, and stacking, or a combination thereof, can be used as an ensemble technique.

[0151] Next, with reference to FIGS. 12 to 14, a second artificial intelligence model learning device (120) and a second artificial intelligence model learning method according to a first embodiment of the present invention will be described in detail.

[0152] FIG. 12 is a drawing showing a second artificial intelligence model learning device (120) according to a first embodiment of the present invention, FIG. 13 is a drawing showing a detailed configuration of a second artificial intelligence model learning unit (1220) according to a first embodiment of the present invention, and FIG. 14 is a drawing showing a second artificial intelligence model learning method according to a first embodiment of the present invention.

[0153] Referring to FIG. 12, a second artificial intelligence model learning device (120) according to a first embodiment of the present invention may include an image preprocessing unit (1200), a learning dataset generation unit (1210), and a second artificial intelligence model learning unit (1220).

[0154] The image preprocessing unit (1200) performs preprocessing on a non-osteoporotic image (normal or osteopenia image) which is a training CXR image. (S210)

[0155] Since CXR imaging equipment has different image storage algorithms depending on the manufacturer or version, a dataset built using only images captured by equipment from a specific manufacturer may result in overfitting to images from that manufacturer during model training. Accordingly, in the first embodiment of the present invention, preprocessing is performed on training CXR images using Grayscale Truncation (GT), a technique that cuts out a portion of pixel values ​​from grayscale images.

[0156] According to the first embodiment of the present invention, the detailed method of preprocessing Grayscale Truncation (GT) performed in the image preprocessing unit (1200) is as follows.

[0157]

[0158] (1) Perform first normalization

[0159] After dividing the CXR image into matrix-shaped regions (pixels) with a predetermined height and width, each pixel value is normalized to between 0 and 1 using the following mathematical formula 1. Through this first-order normalization, the convergence speed of optimization algorithms such as gradient descent can be improved.

[0160]

[0161] Here, Max(X) represents the maximum pixel value and Min(X) represents the minimum pixel value.

[0162]

[0163] (2) Perform 2nd Normalization

[0164] After finding the center region in the image to which the first normalization algorithm has been applied using the following Equation 2, the minimum value Min(Area) obtained from the center region using the following Equation 3 c ) and maximum value MAX(Area cX', which is the total pixel value based on ) i,j Perform second normalization on it.

[0165]

[0166]

[0167] Referring to mathematical formula 3, the pixel value (X' i,j ) is Min(Area c If it is less than or equal to ), this pixel value is considered a very dark value, and the minimum value Min(Area c Adjust with ), and pixel value (X' i,j ) is Max(Area c If it is greater than or equal to ), it is considered excessively bright, and the corresponding pixel value is the maximum value Max(Area c Adjust to ).

[0168]

[0169] (3) Perform 3rd Normalization

[0170] MinMax scaling is performed on the 2nd normalized image by referring to the following mathematical formula 4.

[0171]

[0172] The training dataset generation unit (1210) generates a training dataset based on the training images (images of normal or osteopenia patients) that have undergone Grayscale Truncation (GT) processing performed by the image preprocessing unit (1200) and the associated training labels (normal / osteopenia).

[0173] The second artificial intelligence model learning unit (1220) learns the second artificial intelligence model using the training image processed by GT (Grayscale Truncation) performed in the image preprocessing unit (1200).

[0174] According to the first embodiment of the present invention, a Knowledge Distillation model was used as the second artificial intelligence model. A Knowledge Distillation model refers to a deep learning technique that improves the performance of a small model by transferring knowledge from a large model (teacher model) to a small model (student model).

[0175] Referring to FIG. 13, the second artificial intelligence model learning unit (1220) according to the first embodiment of the present invention may include an augmentation unit (1221), a teacher model unit (1222), a student model unit (1223), a temperature scaling unit (1224), a probability value conversion unit (1225), a label smoothing unit (1226), a loss function operation unit (1227), and a parameter update unit (1228).

[0176] The augmentation unit (1221) performs weak augmentation and strong augmentation on the GT-processed training CXR image. (S220)

[0177] In the first embodiment of the present invention, weak enhancement means a transformation that is relatively similar to the original, and strong enhancement means a transformation of the original more strongly. Techniques such as Horizontal flip, Resize, and Normalize may be used as weak enhancement methods, and techniques such as Blur, Motion Blur, Median Blur, Shift Scale Rotate, Grid Distortion, Elastic Transform, Optical Distortion, Random Gamma, and Grid Dropout may be used as strong enhancement methods.

[0178] The weak augmented image and strong augmented image produced in the augmentation unit (1221) are input into the teacher model unit (1222) and the student model unit (1223), respectively.

[0179] According to the first embodiment of the present invention, a teacher model unit (1222) that receives weak augmented data as input can learn relatively easily about the data, whereas a student model unit (1223) that receives strong augmented data as input may have difficulty learning because the data is in a heavily modified state. However, since the knowledge of the teacher model unit (1222) that learned easily about the same data is transferred to the student model unit (1223) through the Exponential Moving Average (EMA) technique described later, the student model unit (1223) can effectively learn even data containing complex and various modifications. Accordingly, robustness against data containing noise can be secured, generalization performance can be improved, and the feature representation ability of the model can also be strengthened.

[0180] The teacher model unit (1222) receives the weak augmentation image performed by the augmentation unit (1221) and outputs a Logit value for the class, and the student model unit (1223) receives the strong augmentation image performed by the augmentation unit (1221) and outputs a Logit value for the class. (S230)

[0181] Here, 'class' refers to a category or group to which the data may belong, and in the second artificial intelligence model according to the first embodiment of the present invention, two classes are used: osteopenia or normal. Logit refers to the 'denormalized value' for the class predicted by the artificial neural network. That is, logit corresponds to an intermediate value used to represent the probability of belonging to a specific class.

[0182] The teacher model unit (1222) and student model unit (1223) according to the first embodiment of the present invention can be implemented based on a CNN artificial intelligence neural network, and specifically, can be implemented using a ConvNeXt model.

[0183] The temperature scaling unit (1224) applies temperature scaling to the logit values ​​output from the teacher model unit (1222) and the student model unit (1223), respectively, as shown in Equation 5. (S240)

[0184]

[0185] Here, P i is the scaled predicted probability of Class i, z i is the output logit for class i of the model, and T is the temperature constant (temperature factor).

[0186] Through such temperature scaling, the logit values ​​output from the teacher model section (1222) and the student model section (1223), respectively, can be corrected. As the temperature constant (T) value increases, the logit can be corrected more strongly, and in the first embodiment of the present invention, T is set to be greater than 111.

[0187] In this way, when the temperature scaling unit (1224) applies temperature scaling to the logit values ​​output from the teacher model unit (1222) and the student model unit (1223), respectively, and corrects the logit, the probability distribution can be made smoother, thereby reducing the over-confidence of the model.

[0188] The probability value conversion unit (1225) applies the corrected logit value output from the temperature scaling unit (1224) (i.e., the value corrected by applying temperature scaling to the logit value output from the teacher model unit (1222) and the student model unit (1223), respectively) to a softmax function to convert it into a class-specific probability value between 0 and 1. (S250)

[0189] That is, the probability value conversion unit (1225) outputs a probability value (hereinafter also referred to as the 'first probability value') converted from a logit value corrected by applying temperature scaling to a logit value output from the teacher model unit (1222) and a probability value (hereinafter also referred to as the 'second probability value') converted from a logit value corrected by applying temperature scaling to a logit value output from the student model unit (1223). In addition, the probability value conversion unit (1225) converts a logit value output from the student model unit (1223) that has not passed through the temperature scaling unit (1224) (i.e., a logit value to which temperature scaling has not been applied) into a probability value (hereinafter also referred to as the 'third probability value') by applying a softmax function.

[0190] The label smoothing unit (1226) smooths the labels of the classes output from the student model unit (1223) using the following mathematical formula 6. (S260)

[0191]

[0192] Here, α is a smoothing factor with a value between 0 and 1, i represents a class, I is the total number of classes, and y i is the label of the class output from the student model section, y i LS represents the label smoothing value of the output class.

[0193] The loss function calculation unit (1227) calculates the Cross Entropy loss function (CE LOSS) and the Kullback-Leibler Divergence loss function (KLD LOSS) based on the probability value output from the probability value conversion unit (1225). (S270)

[0194] Specifically, the loss function calculation unit (1227) calculates a cross entropy loss function based on a third probability value output from the probability value conversion unit (1225) (i.e., a probability value obtained by applying the logit value output from the student model unit (1223) that has not passed through the temperature scaling unit (1224) to the softmax function).

[0195] At this time, the cross-entropy loss function (CE Loss ) can be calculated through mathematical formula 7.

[0196]

[0197] Here, is the label of class i, represents the probability predicted by the model for class i.

[0198] Additionally, the loss function calculation unit (1227) calculates the Kullback-Leibler Divergence loss function (KLD LOSS) based on the first probability value output from the probability value conversion unit (1225) (a probability value converted from a logit value corrected by applying temperature scaling to a logit value output from the teacher model unit (1222)) and the second probability value (a probability value converted from a logit value corrected by applying temperature scaling to a logit value output from the student model unit (1223)).

[0199] The KLD (Kullback-Leibler Divergence) loss function is an important concept used to measure the difference between probability distributions, and KL Divergence represents a value indicating "how much information is lost from P when approximated by Q" when there are two probability distributions P and Q.

[0200] The KLD loss function calculation process is applied to minimize the difference in distribution of probability values ​​obtained from the teacher model part (1222) and the student model part (1223), and can be expressed by the following mathematical formula 8.

[0201]

[0202] Here, is the value of class i predicted by the teacher model part by applying temperature scaling, and represents the value of class i predicted by the student model part by applying temperature scaling.

[0203] The parameter update unit (1228) updates the parameters of the student model unit (1223). (S280)

[0204] Specifically, the parameter update unit (1228) updates the parameters of the student model unit (1223) based on the cross-entropy loss function (CE Loss) and the KLD loss function (KLD Loss) output from the loss function operation unit (1227). That is, the parameter update unit (1228) updates the parameters of the student model unit (1223) by summing the cross-entropy loss function (CE Loss) and the KLD loss function (KLD Loss) output from the loss function operation unit (1227) and performing backpropagation through an optimizer.

[0205] An optimizer is a rule (algorithm) that determines how to update the weights of a model to minimize a loss function. Stochastic Gradient Descent (SGD), Adaptive Moment Estimation (AdamW), Adagrad, RMSprop, AdaMax, etc., can be used as optimizers, and in the first embodiment of the present invention, Stochastic Gradient Descent (SGD) was used as the optimizer.

[0206] Additionally, the parameter update unit (1228) calculates the Exponential Moving Average (EMA) value for the parameters (Weight or bias) of the teacher model unit (1222) and the parameters of the student model unit (1223) as shown in the following mathematical formula 9, and updates the student model unit (1223).

[0207]

[0208] Here, the Teacher Parameter is the parameter (Weight or bias) of the teacher model part, and the Student Parameter t - 1 is the parameter (Weight or bias) of the student model part, Student Parameter t represents the model parameters (Weight or bias) of the transferred (updated) student model part, and α represents the smoothing constant (Smoothing Factor).

[0209] Next, with reference to FIG. 15, an osteopenia classification device (220) according to the first embodiment of the present invention will be described.

[0210] Referring to FIG. 15, the osteopenia classification device (220) according to the first embodiment of the present invention may include an image preprocessing unit (221) and an osteopenia classification unit (222).

[0211] The image preprocessing unit (221) performs preprocessing on the non-osteoporosis image classified by the non-osteoporosis classification device (210). At this time, the image preprocessing unit (221) of the osteopenia classification device (220) performs preprocessing on the non-osteoporosis image classified by the non-osteoporosis classification device (210) using Grayscale Truncation (GT), similar to the image preprocessing unit (1200) of the second artificial intelligence model learning device (120).

[0212] The osteopenia classification unit (222) finally classifies the non-osteoporotic image, which has been GT processed by the image preprocessing unit (221), into normal or osteopenia based on the student model unit (1223) generated during the learning process of the second artificial intelligence model learning device (120), and according to the embodiment, the student model unit (1223) may be included in the osteopenia classification unit (222).

[0213] According to the first embodiment of the present invention described above, the osteopenia prediction device (200) classifies a CXR image to be read into osteoporosis or non-osteoporosis based on a first artificial intelligence model that has been trained by an artificial intelligence model learning device (100), and classifies the non-osteoporosis image classified in the first stage into normal or osteopenia based on a second artificial intelligence model that has been trained by an artificial intelligence model learning device (100).

[0214] Thus, according to the first embodiment of the present invention, before inferring osteopenia using a second artificial intelligence model specialized for osteopenia classification, osteoporosis is first classified through a first artificial intelligence model, and osteopenia or normality is inferred from images classified as non-osteoporosis excluding osteoporosis, thereby preventing a decline in the classification performance of osteopenia due to confusion in the classification of osteoporosis and osteopenia.

[0215] Meanwhile, according to the first embodiment of the present invention, since the training of the artificial intelligence model is insufficient, there may be cases where actual osteoporosis images are included among the images classified as non-osteoporosis through the first artificial intelligence model. In this case, the results inferred using the second artificial intelligence model are only two, such as osteopenia or normal, so the accuracy of the classification may be lowered.

[0216] Accordingly, in the second embodiment of the present invention, a training dataset is generated based on the entire CXR images of patients with osteoporosis in addition to normal people and patients with osteopenia, and the corresponding training labels (normal / osteopenia / osteoporosis), and then a second artificial intelligence model is trained using the training dataset, and the result is used to solve such problems.

[0217] Hereinafter, an artificial intelligence model learning device and an osteopenia prediction device according to a second embodiment of the present invention will be described with reference to FIG. 16. Among the components of the artificial intelligence model learning device and the osteopenia prediction device according to the second embodiment of the present invention, parts that perform the same or similar operations as those described in the first embodiment of the present invention are given the same reference numerals, and redundant descriptions are omitted below.

[0218] FIG. 16 is a diagram showing the detailed configuration of an artificial intelligence model learning device (100) and an osteopenia prediction device (200) according to a second embodiment of the present invention.

[0219] Referring to FIG. 16, an artificial intelligence model learning device (100) according to the first embodiment of the present invention may include a first artificial intelligence model learning device (110) and a second artificial intelligence model learning device (130).

[0220] The second artificial intelligence model learning device (130) generates a learning dataset based on the entire CXR images of a normal person, a patient with osteopenia, and a patient with osteoporosis, and the corresponding learning labels (normal / osteopenia / osteoporosis), and then learns an artificial intelligence model using each learning dataset. That is, unlike the artificial intelligence model learning device (120) according to the first embodiment of the present invention, the second artificial intelligence model learning device (130) according to the second embodiment of the present invention can generate a learning dataset by adding the entire CXR images of a patient with osteoporosis and the corresponding learning labels (osteoporosis).

[0221] In addition, the second artificial intelligence model learning device (130) according to the second embodiment of the present invention may include an image preprocessing unit (not shown), a learning dataset generation unit (not shown), and a second artificial intelligence model learning unit (not shown), similar to the second artificial intelligence model learning device (120) according to the first embodiment of the present invention shown in FIG. 12. However, unlike the first embodiment, the image preprocessing unit of the second artificial intelligence model learning device (130) according to the second embodiment of the present invention performs Grayscale Truncation preprocessing on the learning image, which is the entire image of a normal person, a patient with osteopenia, or a patient with osteoporosis, and the learning dataset generation unit generates a learning dataset based on the learning image preprocessed by the image preprocessing unit and the learning labels of normal, osteopenia, or osteoporosis associated with the learning image.

[0222] The detailed configuration and learning method of the second artificial intelligence model learning device (130) according to the second embodiment of the present invention are almost identical to the artificial intelligence model learning device (120) according to the first embodiment of the present invention, so redundant descriptions are omitted below.

[0223] The osteopenia prediction device (200) according to the second embodiment of the present invention may include a non-osteoporosis classification device (210) and an osteopenia classification device (230).

[0224] The non-osteoporosis classification device (210) classifies the CXR image to be read into osteoporosis or non-osteoporosis based on the first artificial intelligence model that has been trained by the first artificial intelligence model learning device (110). When the CXR image to be read is classified as osteoporosis by the first artificial intelligence model of the artificial intelligence model learning device (100), the non-osteoporosis classification device (210) outputs a prediction result as 'osteoporosis', and when the CXR image to be read is classified as non-osteoporosis by the first artificial intelligence model, the image classified as non-osteoporosis is input to the osteopenia classification device (230).

[0225] The osteopenia classification device (230) outputs a classification result as normal, osteopenia, or osteoporosis based on the second artificial intelligence model that has been trained by the second artificial intelligence model learning device (130) for the non-osteoporosis image classified by the non-osteoporosis classification device (210). Accordingly, according to the second embodiment of the present invention, even if the osteoporosis image is incorrectly classified as non-osteoporosis through the first artificial intelligence model, the second artificial intelligence model learning device (130) can classify the incorrectly classified non-osteoporosis image as osteoporosis.

[0226] The osteopenia prediction device (200) according to the second embodiment of the present invention can compare the prediction result output by the non-osteoporosis classification device (210) with the prediction result output by the osteopenia classification device (230). Subsequently, based on the comparison result, the prediction for the CXR image to be read can be repeated two or more times, or a message indicating that it corresponds to an uncertainty section or that classification is suspended can be output by the processor (11). For example, if the prediction result output by the non-osteoporosis classification device (210) is 'non-osteoporosis' and the prediction result output by the osteopenia classification device (230) is 'osteoporosis', the osteopenia prediction device (200) can repeat the prediction for the CXR image to be read two or more times.

[0227] The embodiments described above may be implemented in the form of a computer program that can be executed through various components on a computer, and such a computer program may be recorded on a computer-readable medium. In this case, the medium may 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.

[0228] Unless explicitly stated or contrary to the order of the steps constituting the method according to the embodiments of the present invention, said steps may be performed in a suitable order, and the present invention is not limited by the order in which said steps are described. The use of all examples or exemplary terms (e.g., etc.) in the present invention is merely for the purpose of describing the present invention in detail and does not thereby limit the scope of the present invention. Furthermore, a person skilled in the art will understand that various modifications, combinations, and changes may be made within the scope of the claims or their equivalents.

[0229] Although embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto, and various modified and improved forms by those skilled in the art to which the present invention pertains also fall within the scope of the present invention.

Claims

1. An artificial intelligence model learning device comprising a first artificial intelligence model learning device for learning a first artificial intelligence model that classifies input images as osteoporosis or non-osteoporosis, and a second artificial intelligence model learning device for learning a second artificial intelligence model that classifies non-osteoporosis images as normal or osteopenia; and An osteopenia diagnosis system comprising an osteopenia prediction device that primarily classifies an image to be read into osteoporosis or non-osteoporosis based on the first artificial intelligence model, and classifies the primarily classified non-osteoporosis image into normal or osteopenia based on the second artificial intelligence model.

2. In Paragraph 1, The above input image is a chest X-ray image, an osteopenia diagnostic system.

3. In Paragraph 1, The above osteopenia prediction device is A non-osteoporosis classification device that classifies the above-mentioned image to be read into osteoporosis or non-osteoporosis based on the above-mentioned first artificial intelligence model, and An osteopenia diagnostic system comprising an osteopenia classification device that classifies non-osteoporosis images classified by the above-mentioned non-osteoporosis classification device as osteopenia or normal based on the above-mentioned second artificial intelligence model.

4. In Paragraph 3, The first artificial intelligence model described above includes a plurality of classification models that each learn one or more segmented images obtained by dividing the entire image into anatomical regions, and The above-mentioned first artificial intelligence model learning device is an osteopenia diagnosis system that learns a corresponding classification model among the plurality of classification models based on a learning label corresponding to the entire image or the segmented image.

5. In Paragraph 4, The above-mentioned second artificial intelligence model learning device is an osteopenia diagnosis system that trains the second artificial intelligence model based on the entire image of a normal person or a patient with osteopenia and the corresponding learning labels.

6. In Paragraph 5, The above-mentioned second artificial intelligence model learning device is An image preprocessing unit that performs Grayscale Truncation preprocessing on a training image, which is the entire image of the above-mentioned normal person or osteopenia patient, A training dataset generation unit that generates a training dataset based on the training image preprocessed in the image preprocessing unit and the training labels of normal or osteopenia associated with the training image, and A osteopenia diagnosis system comprising a second artificial intelligence model learning unit that trains the second artificial intelligence model using the training data set generated by the training data set generating unit.

7. In Paragraph 6, The above second artificial intelligence model is a knowledge distillation model, and The above second artificial intelligence model learning unit An augmentation unit that performs a first augmentation and a second augmentation stronger than the first augmentation on the training image preprocessed in the image preprocessing unit, A teacher model unit that receives the training video on which the first augmentation has been performed and outputs a first logit value regarding the osteopenia or normal class, A student model unit that receives the training video on which the second augmentation has been performed and outputs a second logit value regarding the osteopenia or normal class, A temperature scaling unit that applies temperature scaling to the first logit value and the second logit value output from the teacher model unit and the student model unit, respectively, and outputs the first corrected logit value and the second corrected logit value, respectively. A probability value conversion unit that performs probability conversion on the first correction logit value and the second correction logit value output from the temperature scaling unit to output a first probability value and a second probability value, respectively, and performs probability conversion on the second logit value output from the student model unit to output a third probability value, A loss function calculation unit that calculates a loss function based on the first probability value, the second probability value, and the third probability value output from the probability value conversion unit, and A osteopenia diagnosis system comprising a parameter update unit that updates the parameters of the student model unit based on the loss function output from the loss function calculation unit.

8. In Paragraph 7, The above loss function operation unit Calculate the cross entropy loss function based on the third probability value output from the probability value conversion unit, and A osteopenia diagnosis system that calculates a Kullback-Leibler Divergence loss function based on the first probability value and the second probability value output from the probability value conversion unit.

9. In Paragraph 8, The above parameter update unit An osteopenia diagnosis system that sums the cross-entropy loss function calculated in the loss function calculation unit and the Kullback-Leibler Divergence loss function, and then performs backpropagation through an optimizer to update the parameters of the student model unit.

10. In Paragraph 7, The above parameter update unit A osteopenia diagnosis system that calculates an exponential moving average value for the parameters of the teacher model unit and the parameters of the student model unit, and updates the student model unit.

11. An artificial intelligence model learning device comprising a first artificial intelligence model learning device for learning a first artificial intelligence model that classifies input images as osteoporosis or non-osteoporosis, and a second artificial intelligence model learning device for learning a second artificial intelligence model that classifies non-osteoporosis images as normal, osteopenia, or osteoporosis; and An osteopenia diagnosis system comprising an osteopenia prediction device that primarily classifies an image to be read into osteoporosis or non-osteoporosis based on the first artificial intelligence model, and classifies the primarily classified non-osteoporosis image into normal, osteopenia, or osteoporosis based on the second artificial intelligence model.

12. In Paragraph 11, The above osteopenia prediction device is A non-osteoporosis classification device that classifies the above-mentioned image to be read into osteoporosis or non-osteoporosis based on the above-mentioned first artificial intelligence model, and An osteopenia diagnostic system comprising an osteopenia classification device that classifies non-osteoporosis images classified by the above-mentioned non-osteoporosis classification device into normal, osteopenia, or osteoporosis based on the above-mentioned second artificial intelligence model.

13. In Paragraph 11, The above-mentioned second artificial intelligence model learning device is An image preprocessing unit that performs Grayscale Truncation preprocessing on a training image, which is the entire image of the above-mentioned normal person, osteopenia patient, or osteoporosis patient, A training dataset generation unit that generates a training dataset based on the training image preprocessed in the image preprocessing unit and the training labels of normal, osteopenia, or osteoporosis associated with the training image, and It includes a second artificial intelligence model training unit that trains the second artificial intelligence model using the training data set generated by the training data set generation unit, and The above second artificial intelligence model is a knowledge distillation model, an osteopenia diagnosis system.

14. In any one of paragraphs 1 through 13, The first artificial intelligence model learning device is A first image segmentation unit that receives a source learning image, which is a whole image of a normal person or a patient with osteoporosis, and segments the source learning image into images corresponding to each anatomical region, receives a target learning image, which is a whole image of a normal person, a patient with osteopenia, or a patient with osteoporosis, and segments the target learning image into images corresponding to each anatomical region; A learning dataset generation unit that generates a plurality of source learning datasets—each source learning dataset includes the source learning image or the plurality of segmented source images segmented from the source learning image—and generates a plurality of target learning datasets—each target learning dataset includes the target learning image or the plurality of segmented source images and a source learning label labeled as normal or osteoporosis—based on the target learning image and a plurality of segmented target images segmented from the target learning image—each target learning dataset includes the target learning image or the plurality of segmented target images and a target learning label labeled as non-osteoporosis or osteoporosis; A plurality of source classification model learning units that perform primary learning on the corresponding source classification model based on the plurality of source learning datasets; and A osteopenia diagnosis system comprising a plurality of target classification model learning units that perform secondary learning on the corresponding target classification models based on the plurality of target learning datasets.

15. In Paragraph 14, The osteopenia diagnosis system further comprises a transfer learning unit that performs transfer learning for a corresponding target classification model learning unit based on hidden layer parameters, which are the learning results of the source classification model learning unit, in the above-mentioned first artificial intelligence model learning device.

16. In Paragraph 15, The first image segmentation unit crops and divides the entire image into images corresponding to each anatomical region, and The training image generated by the training dataset generation unit comprises a whole chest image and one or more segmented images among the right clavicle-scapula, left clavicle-scapula, cervical spine, and thoracic / lumbar spine, in an osteopenia diagnosis system.

17. In Paragraph 15, The above non-osteoporosis classification device is A second image segmentation unit that divides the above-mentioned image to be read by anatomical region to generate a plurality of segmented images; and A bone disease classification inference unit that infers a bone disease classification result for each input image based on a corresponding classification model among the plurality of target classification models, the above-mentioned entire image to be read and the plurality of segmented images segmented by the second image segmentation unit, and A osteopenia diagnosis system comprising a combination unit that combines multiple bone disease classification results inferred by the bone disease classification inference unit using an ensemble algorithm to diagnose whether the image to be read is an osteoporosis image or a non-osteoporosis image.

18. A first artificial intelligence model training device for training a first artificial intelligence model that classifies an input image as osteoporosis or non-osteoporosis; and It includes a second artificial intelligence model training device for training a second artificial intelligence model that classifies non-osteoporotic images as normal or osteopenia, and The first artificial intelligence model described above includes a plurality of classification models that each learn one or more segmented images obtained by dividing the entire image into anatomical regions, and The above-mentioned first artificial intelligence model learning device is an artificial intelligence model learning device that learns a corresponding classification model among the plurality of classification models based on a learning label corresponding to the entire image or the segmented image.

19. In Paragraph 18, The above second artificial intelligence model learning device is an artificial intelligence model learning device that trains the above second artificial intelligence model based on the entire image of a normal person or a patient with osteopenia and the corresponding learning label.

20. In Paragraph 19, The above-mentioned second artificial intelligence model learning device is An image preprocessing unit that performs Grayscale Truncation preprocessing on a training image, which is the entire image of the above-mentioned normal person or osteopenia patient, A training dataset generation unit that generates a training dataset based on the training image preprocessed in the image preprocessing unit and the training labels of normal or osteopenia associated with the training image, and An artificial intelligence model learning device comprising a second artificial intelligence model learning unit that trains the second artificial intelligence model using the training data set generated by the training data set generating unit.

21. In Paragraph 20, The above second artificial intelligence model is a knowledge distillation model, and The above second artificial intelligence model learning unit An augmentation unit that performs a first augmentation and a second augmentation stronger than the first augmentation on the training image preprocessed in the image preprocessing unit, A teacher model unit that receives the training video on which the first augmentation has been performed and outputs a first logit value regarding the osteopenia or normal class, A student model unit that receives the training video on which the second augmentation has been performed and outputs a second logit value regarding the osteopenia or normal class, A temperature scaling unit that applies temperature scaling to the first logit value and the second logit value output from the teacher model unit and the student model unit, respectively, and outputs the first corrected logit value and the second corrected logit value, respectively. A probability value conversion unit that performs probability conversion on the first correction logit value and the second correction logit value output from the temperature scaling unit to output a first probability value and a second probability value, respectively, and performs probability conversion on the second logit value output from the student model unit to output a third probability value, A loss function calculation unit that calculates a loss function based on the first probability value, the second probability value, and the third probability value output from the probability value conversion unit, and An artificial intelligence model learning device comprising a parameter update unit that updates the parameters of the student model unit based on the loss function output from the loss function calculation unit.

22. In Paragraph 21, The above loss function operation unit Calculate the cross entropy loss function based on the third probability value output from the probability value conversion unit, and An artificial intelligence model learning device that calculates a Kullback-Leibler Divergence loss function based on the first probability value and the second probability value output from the probability value conversion unit.

23. In Paragraph 22, The above parameter update unit An artificial intelligence model learning device that sums the cross-entropy loss function calculated in the loss function calculation unit and the Kullback-Leibler Divergence loss function, and then performs backpropagation through an optimizer to update the parameters of the student model unit.

24. In any one of paragraphs 18 through 23, The above second artificial intelligence model learning device is an artificial intelligence model learning device that trains the above second artificial intelligence model to classify as normal, osteopenia, or osteoporosis.

25. In a method for diagnosing osteopenia from images to be interpreted, A step of training a first artificial intelligence model that classifies input images as osteoporosis or non-osteoporosis; A step of training a second artificial intelligence model that classifies input non-osteoporotic images as normal or osteopenia; A step of primarily classifying the image to be read into osteoporosis or non-osteoporosis based on the first artificial intelligence model; and A method for diagnosing osteopenia comprising the step of classifying the above-mentioned first-classified non-osteoporotic images as normal or osteopenia based on the above-mentioned second artificial intelligence model.

26. In Paragraph 25, The first artificial intelligence model described above includes a plurality of classification models that each learn one or more segmented images obtained by dividing the entire image into anatomical regions, and The step of training the first artificial intelligence model is a method for diagnosing osteopenia in which a corresponding classification model among the plurality of classification models is trained based on a training label corresponding to the entire image or the segmented image.

27. In Paragraph 26, The step of training the above-mentioned second artificial intelligence model is A step of performing Grayscale Truncation preprocessing on training images, which are whole images of normal individuals or patients with osteopenia, A step of generating a training dataset based on the preprocessed training image and the training labels of normal or osteopenia associated with the training image, A method for diagnosing osteopenia comprising the step of training the second artificial intelligence model using the training data set generated by the training data set generating unit.

28. In Paragraph 27, The above second artificial intelligence model is a knowledge distillation model, and The step of training the above-mentioned second artificial intelligence model is A step of performing a first augmentation and a second augmentation stronger than the first augmentation on the preprocessed training image, A step of receiving the training video in which the first augmentation is performed in the teacher model section and outputting a first logit value regarding the osteopenia or normal class, A step of receiving the training video in which the second augmentation is performed in the student model section and outputting a second logit value regarding the osteopenia or normal class, A step of applying temperature scaling to the first logit value and the second logit value output from the teacher model unit and the student model unit, respectively, to output the first corrected logit value and the second corrected logit value, respectively. A step of performing a probability transformation on the first correction logit value and the second correction logit value, respectively, to output the first probability value and the second probability value, A step of performing a probability transformation on the second logit value output from the student model section to output a third probability value, A step of calculating a loss function based on the first probability value, the second probability value, and the third probability value. A method for diagnosing osteopenia comprising the step of updating the parameters of the student model part based on the above-calculated loss function.

29. In Paragraph 28, The step of calculating the above loss function is A step of calculating a cross-entropy loss function based on the above third probability value, and A method for diagnosing osteopenia comprising the step of calculating a Kullback-Leibler Divergence loss function based on the first probability value and the second probability value.

30. In Paragraph 29, The step of updating the above parameters is A method for diagnosing osteopenia comprising the step of summing the calculated cross-entropy loss function and the Kullback-Leibler Divergence loss function, and then performing backpropagation through an optimizer.

31. Includes at least one processor implemented to execute computer-readable instructions, and By the above at least one processor, Train a first artificial intelligence model that classifies input images as osteoporosis or non-osteoporosis, and Train a second artificial intelligence model that classifies input non-osteoporotic images as normal or osteopenia, and The image to be read is primarily classified as osteoporosis or non-osteoporosis based on the first artificial intelligence model, and A computer that classifies the above-mentioned first-classified non-osteoporotic images as normal or osteopenia based on the above-mentioned second artificial intelligence model.

Images