Method for predicting biomarker by using fundus image

Abstract

The present invention relates to a method for predicting a biomarker by using a fundus image, the method comprising: an image acquisition step in which a user client acquires and transmits a fundus image of a subject; a step in which a server receives the fundus image and improves the quality of the fundus image; a step in which the server pre-trains a deep learning model by using training data; and a biomarker prediction step in which the server predicts a health biomarker by using the deep learning model on the basis of the fundus image.

Description

Method to predict biomarkers using fundus images

[0001] The present invention relates to a method for predicting biomarkers using fundus images, and more specifically, to a method for predicting a comprehensive set of health biomarkers using fundus images of a subject and mapping the biomarkers to related organs to predict potential chronic diseases.

[0002] Fundus photography is a medical imaging technique that captures detailed images of the internal surfaces of the eye, including the retina, optic disc, and blood vessels. This technique uses a specialized camera to obtain high-resolution color photographs of the retina and surrounding structures.

[0003] Fundus photography is generally performed using a digital camera equipped with a specialized lens that enables non-invasive imaging of the retina. This technique provides a comprehensive view of the retinal structure and can be used to diagnose and monitor various ophthalmic diseases such as diabetic retinopathy, macular degeneration, and retinal detachment.

[0004] Depending on the characteristics of the camera and imaging system (e.g., wide-field imaging, color or monochrome sensors), fundus photography can obtain high-resolution images of the retinal surface, aiding in the early detection and management of retinal diseases.

[0005] Recently, methods to process various medical images using deep learning to obtain necessary information are being actively developed.

[0006] Deep learning is defined as a series of machine learning algorithms that aim to perform high-dimensional abstraction (summarizing key content or extracting features from large amounts of data or complex materials) through a combination of multiple non-linear transformation techniques. In a broad sense, deep learning can be considered a field of machine learning that trains computers to think like humans.

[0007] The present invention was devised to solve the aforementioned problem and aims to provide a method for predicting health biomarkers using fundus images.

[0008] The problem that the present invention aims to solve is not limited to the problem mentioned above, and other unmentioned problems will be clearly understood by those skilled in the art from the following description.

[0009] A method for predicting biomarkers using a fundus image according to an embodiment of the present invention for solving the aforementioned problem comprises: an image acquisition step in which a user client acquires and transmits a fundus image of a target; a step in which a server receives the fundus image and improves the quality of the fundus image; a step in which a server pre-trains a deep learning model using training data; and a biomarker prediction step in which a server predicts a health biomarker using a deep learning model based on the fundus image.

[0010] According to another embodiment of the present invention, the biomarker may be composed of the patient's age, sex, waist circumference, height, weight, body mass index (BMI), systolic blood pressure (SBP), diastolic blood pressure (DBP), total cholesterol, triglycerides, blood glucose, hbA1c, albumin, estimated glomerular filtration rate (eGFR), creatinine, blood urea nitrogen (BUN), or uric acid.

[0011] According to another embodiment of the present invention, the step of improving the quality of the fundus image may be configured to include: a step in which the server applies CLAHE (Contrast Limited Adaptive Histogram Equalization) to the fundus image to improve the contrast of the fundus image; a step in which the server removes noise from the fundus image with improved contrast; and a step in which the server corrects the gamma of the fundus image from which noise has been removed.

[0012] According to another embodiment of the present invention, the step of pre-training the deep learning model may pre-train the deep learning model using labeled training data including health biomarkers that operate as labels on the server.

[0013] According to another embodiment of the present invention, the biomarker prediction step compares the predicted biomarker value with a normal value and an abnormal value, wherein the abnormal value may be higher or lower than the normal value and may be mapped to an organ associated with a potential chronic disease. According to another embodiment of the present invention, the biomarker prediction step compares the predicted biomarker value with a reference range indicating low, normal, and high and may be mapped to an organ associated with a potential chronic disease.

[0014] According to another embodiment of the present invention, the biomarker prediction step may involve the server receiving the fundus image and predicting binary biomarkers using a classification model, and the server receiving the fundus image and predicting continuous biomarkers using a regression model.

[0015] According to another embodiment of the present invention, the biomarker prediction step involves a server receiving the fundus image and predicting a health biomarker using a deep learning model that includes at least one residual block, wherein the residual block may be configured to transmit gradients from an intermediate layer of the network of the deep learning model to a subsequent layer.

[0016] A program for performing a method of predicting biomarkers using a fundus image according to an embodiment of the present invention may be stored on a computer-readable recording medium.

[0017] According to the present invention, a comprehensive set of health biomarkers of a subject can be predicted using fundus images, and potential chronic diseases can be predicted by mapping the predicted biomarkers to related organs.

[0018] The effects of the present invention are not limited to those mentioned above, and other unmentioned effects will be clearly understood by those skilled in the art from the following description.

[0019] FIG. 1 is a conceptual diagram of a fundus image analysis system according to an embodiment of the present invention.

[0020] FIG. 2 is a flowchart illustrating a method for predicting biomarkers using fundus images according to an embodiment of the present invention.

[0021] FIG. 3 is a diagram illustrating a method for improving the quality of a fundus image through image preprocessing according to an embodiment of the present invention.

[0022] FIG. 4 is a diagram illustrating a method for predicting biomarkers using a deep learning model according to an embodiment of the present invention.

[0023] FIG. 5 is a diagram illustrating the configuration of a deep learning model used for predicting biomarkers according to an embodiment of the present invention.

[0024] FIG. 6 is a diagram illustrating the correlation between a biomarker and a human organ according to an embodiment of the present invention.

[0025] The present invention is capable of various modifications and may have various embodiments, and specific embodiments are illustrated in the drawings and described in detail in the description of the invention. However, this is not intended to limit the present invention to specific embodiments, and it should be understood that it includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the invention.

[0026] However, in describing the embodiments, if it is determined that a detailed description of related known functions or configurations could unnecessarily obscure the essence of the invention, such detailed description is omitted. Furthermore, the sizes of each component in the drawings may be exaggerated for illustrative purposes and do not represent the actual sizes applied.

[0027] Furthermore, throughout the specification, when a component is referred to as being "connected" or "joined" with another component, it should be understood that the component may be directly connected or joined to the other component, but unless specifically stated otherwise, it may also be connected or joined through an intermediate component. Additionally, throughout the specification, when a part is described as "including" a component, unless specifically stated otherwise, this means that it may include additional components rather than excluding other components.

[0028] A detailed description of the embodiments of the present invention will be made with reference to the attached drawings.

[0029] According to one embodiment of the present invention, the term "image" refers to multidimensional data composed of discrete image elements, and means, for example, pixels in a 2D image and voxels in a 3D image.

[0030] Additionally, the term "object" according to one embodiment of the present invention may refer to a human or an animal, or a part or the whole thereof. For example, the object may include at least one of organs such as a liver, heart, kidney, brain, pancreas, and blood vessels.

[0031] In addition, the term "user" according to one embodiment of the present invention may refer to, but is not limited to, medical professionals such as doctors, nurses, clinical pathologists, medical imaging specialists, or technicians who repair medical devices.

[0032] FIG. 1 is a conceptual diagram of a fundus image analysis system according to an embodiment of the present invention.

[0033] A fundus image analysis system according to one embodiment of the present invention comprises a fundus image capturing device (10), a user client (20), and a server (100).

[0034] The fundus image capturing device (10) includes all types of devices capable of capturing fundus images of a target.

[0035] For example, the fundus image capturing device (10) may be a conventional fundus image capturing device or an ultra-wide-angle fundus image capturing device that captures a fundus image including the vitreous humor, retina, choroid, and optic nerve head through the pupil of the target.

[0036] Additionally, the fundus image capturing device (10) may be an OCT device that captures a cross-sectional image of the retina near the subject's eye. The OCT device can obtain information about the retinal nerve fiber layer, the optic nerve head, and the macula.

[0037] According to one embodiment of the present invention, a user client (20) refers to a user client of a computing device comprising at least one central processing unit, and a server (100) refers to a computing device comprising at least one processor and one graphics processing unit.

[0038] Additionally, according to one embodiment of the present invention, the fundus imaging device (10) and the user client (20) can be integrated into a single device, and the fundus imaging device (10) can be used physically or electrically with the user client (20).

[0039] The user client (20) obtains a fundus image of the target from the fundus image capturing device (10).

[0040] Additionally, the user client (20) transmits the acquired fundus image to the server (100).

[0041] The server (100) analyzes the fundus image obtained from the user client (20) and sends the analysis result back to the user client (20).

[0042] According to one embodiment of the present invention, the user client (20) and the server (100) are not separated, and the user client (20) can analyze the fundus image acquired and provide the analysis results to the user.

[0043] Analyzing fundus images showing the inside of the eye is a complex task typically performed by an ophthalmologist. Nevertheless, even in medical institutions without an ophthalmologist, if a fundus imaging device is available, the presence of disease in the subject can be determined through analysis by a user client (20) or server (100).

[0044] According to one embodiment of the present invention, the judgment result of the user client (20) or server (100) can be used to assist the judgment of a general physician or specialist.

[0045] Additionally, according to one embodiment of the present invention, if the judgment result of the user client (20) or server (100) determines that the predicted biomarker is within the normal range, the judgment result is provided to the user. If multiple biomarkers correspond to high-risk values, the fundus image of the subject may be classified for re-evaluation.

[0046] Fundus images classified for re-evaluation are evaluated for the presence of high-risk values ​​using the same or different algorithms used in the initial judgment, and the organs affected by them can also be determined. According to one embodiment of the present invention, fundus images classified for re-evaluation are interpreted by an expert to determine biomarkers of high-risk values, and the result is provided to a user through a user client (20) or a server (100).

[0047] According to the present invention, patients who require frequent fundus examinations, such as those with diabetes, high cholesterol, and high blood pressure, can easily receive frequent examinations without having to see a specialist every time, thereby improving convenience for both patients and doctors.

[0048] According to one embodiment of the present invention, angiography may be used to obtain an image of the eye of a subject.

[0049] Angiography is a method primarily used to observe a subject's blood vessels and involves the process of administering a contrast agent to the subject. For example, fluorescein angiography (FA) is a method of capturing fundus images by administering a fluorescent substance, a fluorescent dye, to the subject. This method provides information regarding changes in blood flow within the blood vessels of the eye and the extent of retinal vascular damage.

[0050] Alternatively, optical coherence tomography (OCT) angiography can be used to observe the movement of blood vessels and provide information on changes in blood flow within the blood vessels of the eye and the extent of retinal vascular damage. In addition, OCT angiography can capture changes in the cross-section of blood vessels, making it advantageous for detecting minute changes in blood flow and localized vascular damage.

[0051] In addition, OCT angiography is a completely non-invasive method that does not use contrast agents, so concerns regarding the side effects of contrast agents can be eliminated.

[0052] According to one embodiment of the present invention, a server (100) stores a pre-trained model using labeled health biomarkers. The server (100) can determine health biomarkers by inputting a fundus image of a subject into the model.

[0053] A biomarker generally refers to an indicator that can detect changes within the body using proteins, DNA, RNA (ribonucleic acid), metabolites, etc.

[0054] For example, the server (100) can predict comprehensive health biomarkers using the fundus image of the target.

[0055] Alternatively, the server (100) can determine whether the subject has organ failure by using an eye image obtained through angiography.

[0056] The server (100) can predict various health biomarkers of the subject using various types of fundus images taken by the method described above.

[0057] Hereinafter, as a preferred embodiment according to the disclosed example, a method for predicting a subject's health biomarker using a traditional fundus image (CFI) of the subject's eye will be described. However, the method for predicting health biomarkers according to the disclosed example is not limited to using fundus images and includes using various types of eye images.

[0058] FIG. 2 is a flowchart illustrating a method for predicting biomarkers using fundus images according to an embodiment of the present invention.

[0059] In addition, FIG. 3 is a diagram illustrating a method for improving the quality of a fundus image through image preprocessing according to an embodiment of the present invention, FIG. 4 is a diagram illustrating a method for predicting biomarkers using a deep learning model according to an embodiment of the present invention, and FIG. 5 is a diagram illustrating the configuration of a deep learning model used for predicting biomarkers according to an embodiment of the present invention.

[0060] In addition, FIG. 6 is a diagram illustrating the correlation between a biomarker and a human organ according to an embodiment of the present invention.

[0061] From now on, a method for predicting biomarkers using fundus images according to an embodiment of the present invention will be described with reference to FIGS. 2 to 6.

[0062] A method for predicting biomarkers using a fundus image according to one embodiment of the present invention performs an analysis of the fundus image in the order shown in FIG. 2.

[0063] All or at least some of the steps illustrated in FIG. 2 can be performed by a user client (20) or a server (100).

[0064] First, the user client (20) acquires an image of the target's fundus (S210). At this time, the user client (20) may acquire a CFI image, an OCT image, or an angiography image of the target's eye.

[0065] Accordingly, the server (100) can receive a CFI image, OCT image, or angiography image of the target's eye.

[0066] Afterwards, the server (100) receives the fundus image and improves the quality of the fundus image (S220).

[0067] Referring to Fig. 3, a method for improving the quality of the fundus image will be explained in more detail.

[0068] The fundus image (300) passes through an image enhancement model (400).

[0069] The model (400) receives an image (300) and uses Contrast Limited Adaptive Histogram Equalization (CLAHE) technology to enhance the contrast of the image. This prevents excessive amplification of noise while enhancing local details.

[0070] The contrast-enhanced image is then passed to a noise reduction model that averages pixels based on spatial proximity and intensity similarity to reduce image noise while preserving edges.

[0071] After contrast and noise removal, the image is passed to a brightness adjustment model that uses gamma correction, applying a non-linear transformation to adjust the image brightness so that details in dark or bright areas can be seen better.

[0072] After going through this process, the final improved image (410) can be obtained.

[0073] After the image preprocessing described above, we will explain the deep learning model processing method in more detail.

[0074] Several transformations are applied to the image to make it compatible with the model.

[0075] The transformation involves adjusting the image size to 256x256 pixels using bilinear interpolation. The transformation also involves centering the image to 224x224 pixels to focus on the center part of the image.

[0076] In addition, the final step of image transformation performs normalization to adjust pixel values ​​based on the mean and standard deviation values.

[0077] In the model according to one embodiment of the present invention, the mean [0.485, 0.456, 0.406] and standard deviation [0.229, 0.224, 0.225] derived from the ImageNet dataset were used.

[0078]

[0079] According to one embodiment of the present invention, the server (100) can obtain a model that predicts the health biomarker of a target by deriving a correlation between the fundus image and the biomarker (S230).

[0080] According to one embodiment of the present invention, the model may be a model trained using fundus images along with health biomarker labels in the form of electronic health records.

[0081] In addition, according to one embodiment of the present invention, a model can be trained using deep learning. Deep learning is defined as a set of machine learning algorithms that attempt high-dimensional abstraction by combining a number of non-linear transformation techniques. In a broad sense, deep learning can be viewed as a field of machine learning that aims to teach computers to think like humans.

[0082] When data is available, it is represented in a form that computers can understand (e.g., pixel information of an image is represented as a vector), and extensive research is conducted on these representation methods and how to build models for learning. As a result of these efforts, various deep learning technologies have been developed, such as Deep Neural Networks (DNN), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Deep Belief Networks (DBN).

[0083] A deep neural network (DNN) is an artificial neural network that has multiple hidden layers between the input layer and the output layer.

[0084] A model according to one embodiment of the present invention may be trained as a CNN using labeled data, but the training method of the model is not limited thereto.

[0085] In other words, a model that reflects the correlation between features appearing in fundus images and the prediction of health biomarkers can be trained using deep learning. The trained model can be used to determine the health biomarkers of a subject based solely on fundus images.

[0086]

[0087] The server predicts health biomarkers for such fundus images using a trained deep learning model (S240).

[0088] According to one embodiment of the present invention, in addition to directly associating the fundus with a health biomarker of a subject, various methods may be used to determine the presence of such biomarkers in an image of the subject's fundus directly or indirectly, or to determine them through alternative means.

[0089] By using various methods to determine the presence of health biomarkers in a subject, cross-checking the results or integrating multiple results can more accurately determine the presence of these diseases in the subject.

[0090] In addition, various testing methods can be used to determine biomarkers. As a more direct method than using fundus images, there are various techniques available to assess the presence of these biomarkers in subjects.

[0091] Furthermore, in the prediction process, biomarkers are divided into two categories based on their values. Binary biomarkers, such as gender, are binary because they have categorical values ​​of male or female, while the remaining biomarkers are all continuous, such as age, blood pressure, and cholesterol.

[0092] Furthermore, the selection of these health biomarkers is based on their close association with organs that may cause potential chronic diseases. Examples include diabetes, chronic kidney disease, chronic liver disease, cardiovascular disease, cardiometabolic disease, and neurodegenerative disease.

[0093] After prediction, these biomarkers are mapped to various organs associated with potential chronic diseases.

[0094] Furthermore, if various biomarkers associated with specific organs exhibit high-risk values, this may indicate a potential risk of organ failure. For example, systolic blood pressure (SBP) and diastolic blood pressure (DBP) affect blood vessels and the heart, and abnormal values ​​can lead to cardiovascular disease.

[0095] These aforementioned health biomarkers are often more complex, time-consuming, expensive, and require invasive procedures compared to fundus photography.

[0096] On the other hand, fundus photography can be performed without the need for an on-site expert, as long as fundus equipment is available. The captured images are evaluated via a server, and once health biomarkers are determined, they are mapped to the relevant organs. Additionally, expert consultations can be conducted based on the results. Furthermore, for the elderly who are more vulnerable to chronic diseases, even if they reside in areas with limited access to hospitals, having fundus equipment at local clinics or public health centers can provide significant benefits through regular and easy screenings for these conditions.

[0097] In addition, according to one embodiment of the present invention, the correlation between fundus images and health biomarkers can be evaluated, and these biomarkers can be mapped to related organs.

[0098] FIG. 4 describes in more detail a method for predicting health biomarkers in a fundus image of a subject using a deep learning model according to an embodiment of the present invention.

[0099] The enhanced image (410) is input into two different model architectures called classification (500) and regression (520).

[0100] The model (500) helps predict binary biomarkers (510), and the regression model (520) helps predict continuous biomarkers (530).

[0101] The model (500) is designed to classify data into discrete classes. In the present invention, it is used to predict binary biomarkers that can be classified into predefined classes (e.g., male or female), such as gender.

[0102] The model (520) is used to predict values ​​that can be specific values, such as age, blood pressure, and cholesterol levels. Instead of dividing into categories, this model estimates accurate values ​​based on input information.

[0103] These two models are combined to provide a total of 17 sets of biomarkers (600).

[0104] Deep learning models are primarily based on convolutional layers and include skip connections to increase learning speed and improve efficiency.

[0105] According to one embodiment of the present invention, a server (100) can train a model showing the correlation between a fundus image and a biomarker using labeled training data including an electronic health record containing a health biomarker that acts as a label.

[0106] As illustrated in FIG. 4, the server (100) can use a model that directly evaluates the presence of health biomarkers of the target in the fundus image.

[0107] As illustrated in FIG. 4, the server (100) can estimate health biomarkers from a fundus image of a target, and can predict health biomarkers using a model pre-trained with a larger dataset even if it is not a fundus image.

[0108] These pre-trained models have the ability to recognize patterns and features within fundus images that can represent various health biomarkers.

[0109] The server (100) may generate a model that predicts health biomarkers. This model may be a deep learning model, or an algorithm-based model that evaluates the presence or risk of such diseases according to predefined criteria.

[0110] In other words, the server (100) can estimate various markers from the fundus image of the subject and, based on this, evaluate the possibility (or risk) of various chronic diseases that may occur in organs associated with the predicted biomarkers.

[0111] The predicted biomarkers (600) are divided into two categories when output. The first is a biomarker that provides personal information such as age (601), gender (602), waist circumference (603), height (604), and weight (605).

[0112] The second is a biomarker that provides health information such as body mass index (BMI) (606), systolic blood pressure (SBP) (607), diastolic blood pressure (DBP) (608), total cholesterol (609), triglycerides (610), blood glucose (611), hbA1c (612), albumin (613), estimated glomerular filtration rate (eGFR) (614), creatinine (615), blood urea nitrogen (BUN) (616), and uric acid (617).

[0113] Figure 5 illustrates a method for predicting health biomarkers using fundus images and includes various model components designed to extract, process, and classify relevant features.

[0114] The components include Feature Extraction (FE), Dimensionality Reduction (DR), Residual Block (RB), Fully Connected Layer (FCL), Linear Layer (LL), and Softmax Layer (SL).

[0115] The Feature Extraction (FE) component is responsible for identifying and extracting key features from the input fundus image. These features may include patterns and structures important for determining health biomarkers.

[0116] A convolutional layer is used for feature extraction (FE), and features are extracted by applying filters of various sizes (1x1, 3x3, 5x5) to the input image.

[0117] Next, the Dimensionality Reduction (DR) component compresses features to reduce data complexity while preserving the most important information for the prediction task.

[0118] Max Pooling (MP) and Average Pooling (AP) layers are used for dimensionality reduction (DR). MP selects the maximum value from each segment of the feature map to emphasize key features, while AP calculates the average value to reduce noise and smooth the data.

[0119] The extracted features are connected according to the channel dimension. This connection merges various feature maps into a single output.

[0120] Residual blocks (RBs) enhance the model's ability to learn deeper and more complex patterns by allowing inputs to bypass specific layers and be added directly to the output.

[0121] In addition, a direct skip connection is established from the residual block (RB) to the fully connected layer (FCL), allowing the original features to bypass the intermediate layer and be integrated into the model later. This ensures that important information is preserved and directly influences the final prediction.

[0122] After the residual block, the model applies additional feature extraction (FE) and dimensionality reduction (DR) steps to further refine features, preserving only the information most relevant to the final prediction.

[0123] The refined features pass through the fully connected layer (FCL), where all previous features are combined into an integrated representation. This layer plays a crucial role in converting the spatial feature map into a form suitable for the final prediction task.

[0124] For regression tasks, the output of a fully connected layer is passed to a linear function representing continuous values, and then the values ​​are averaged for the final prediction.

[0125] For classification tasks, the model uses a softmax function after FCL to convert log probabilities into probabilities, enabling the model to perform category predictions such as gender. As shown in Figure 5, a value greater than or equal to 0.5 is considered as 1, where 1 represents male, and a value less than or equal to 0 is considered as 0, where 0 represents female.

[0126] The model ensures optimal performance and accuracy by using individual loss functions tailored to the characteristics of the prediction task.

[0127] For classification tasks, the model uses the Cross-Entropy Loss function. This loss function is particularly well-suited for category prediction and measures the discrepancy between the predicted probability distribution and the actual distribution, prompting the model to adjust parameters to minimize this error and improve the accuracy of category predictions.

[0128] For regression tasks, the Mean Squared Error (MSE) loss function is used. The MSE loss is designed to measure the mean square difference between the predicted continuous values ​​and the actual target values. By minimizing the MSE, the model is trained to make predictions as close as possible to the actual values, ensuring the reliability and precision of continuous output predictions.

[0129] The regression operation provides continuous values, and the mean error is calculated using MSE in the predicted loss group. In the example of Figure 5, various continuous values ​​such as 2.53, 2.62, 2.31, and 2.24 can be seen. All these values ​​are averaged, and the final mean squared error is calculated to be 2.497.

[0130] Both loss functions play a crucial role in the training process, and the model optimizes predictive performance for various health biomarkers by iteratively adjusting parameters based on the calculated loss.

[0131] When predicting gender (602), two classes, male and female, are considered. However, there is a class imbalance where males make up 60% of the sample and females make up 40%.

[0132] Oversampling techniques were used to address this class imbalance. This technique improves prediction accuracy by creating a balanced dataset through increasing the sample size of the minority class (women) to match that of the majority class (men).

[0133] Each health biomarker requires individual hyperparameter tuning to achieve optimal convergence. This tuning included adjusting the number of epochs, batch size, and learning rate. To accommodate these diverse requirements, separate models with the same overall architecture were used to predict each continuous health biomarker.

[0134] In this approach, 16 models with the same architecture were used for regression tasks, and each model was specifically tuned for continuous biomarkers using optimized hyperparameters. Additionally, a separate model with a different architecture was implemented for classification tasks.

[0135] In addition, various evaluation metrics were used to evaluate the model's performance.

[0136] Accuracy was used as an evaluation metric to evaluate the classification model that predicted gender (602).

[0137] The regression model used the mean absolute error (MAE) to evaluate the predictions of the remaining continuous biomarkers (601) and (603-617).

[0138] The evaluation results showed that the accuracy for gender (602) was 86%.

[0139] The following is the mean absolute error (MAE) of continuous biomarkers.

[0140] Age (601): 3.47, Waist Circumference (603): 6.30, Height (604): 6.05, Weight (605): 9.47, BMI (606): 2.60, Systolic Blood Pressure (SBP, 607): 10.47, Diastolic Blood Pressure (DBP, 608): 8.05, Total Cholesterol (609): 30.56, Triglycerides (610): 53.43, Blood Glucose (611): 11.75, Glycated Hemoglobin (HbA1c, 612): 0.40, Albumin (613): 0.23, Estimated Glomerular Filtration Rate (eGFR, 614): 11.05, Creatinine (615): 0.12, Blood Urea Nitrogen (BUN, 616): 2.55, Uric Acid (617): 0.99.

[0141] FIG. 6 is a diagram illustrating a method for mapping predicted health biomarkers (600) to organs.

[0142] According to one embodiment of the present invention, numbers 601 through 617 represent health biomarkers, each associated with one or more organs.

[0143] For example, SBP and DBP are related to blood pressure and the heart, respectively, so they are associated with these two organs. Similarly, uric acid affects the kidneys and the heart.

[0144] For example, if several biomarkers associated with a specific organ have high values, this can lead to organ failure or potential chronic disease.

[0145] More specifically, Figure 6 illustrates a total of six organs, namely the heart, pancreas, kidney, liver, blood vessels, and brain.

[0146] According to another embodiment of the present invention, there are pancreatic-related health biomarkers such as blood glucose and hbA1c, and high values ​​of these can cause diabetes.

[0147] According to another embodiment of the present invention, kidney-related biomarkers include creatinine, blood urea nitrogen (BUN), uric acid, estimated glomerular filtration rate (eGFR), and albumin. If one or more of these health biomarkers show abnormal values, it may lead to chronic kidney disease.

[0148] According to another embodiment of the present invention, abnormal values ​​of biomarkers such as SBP, DBP, total cholesterol, triglycerides, and blood glucose increase the risk of cerebrovascular disease and may cause conditions such as stroke or contribute to cognitive decline or neurodegenerative diseases such as Alzheimer's disease.

[0149] Additionally, the steps of the method or algorithm described in connection with the embodiments of the present invention may be implemented directly in hardware, as a software module executed in hardware, or as a combination of both. The software module may exist in the form of RAM (Random access memory), ROM (Read only memory), EPROM (Erasable Programmable Read Only Memory), EPROM (electrically erasable PROM), flash memory, hard disk, removable disk, CD-ROM, or other computer-readable media related to the art.

[0150] In the detailed description of the present invention as described above, specific embodiments have been described. However, various modifications are possible within the scope of the present invention. The technical concept of the present invention should not be limited to the aforementioned embodiments, but should be defined by the claims as well as equivalents thereof.

Claims

1. An image acquisition step in which a user client acquires and transmits a fundus image of a target; A step in which a server receives the fundus image and improves the quality of the fundus image; A step in which the server pre-trains the deep learning model using training data; and A biomarker prediction step in which a server predicts health biomarkers using a deep learning model based on the above fundus image; A method for predicting biomarkers using fundus images characterized by including 2. In Claim 1, The above biomarker is, A method for predicting biomarkers using fundus images characterized by including patient's age, sex, waist circumference, height, weight, body mass index (BMI), systolic blood pressure (SBP), diastolic blood pressure (DBP), total cholesterol, triglycerides, blood glucose, hbA1c, albumin, estimated glomerular filtration rate (eGFR), creatinine, blood urea nitrogen (BUN), or uric acid.

3. In Claim 1, The step of improving the quality of the above fundus image is, A step in which the server applies CLAHE (Contrast Limited Adaptive Histogram Equalization) to the fundus image to enhance the contrast of the fundus image; The above server removes noise from the contrast-enhanced fundus image; and A step in which the server corrects the gamma of the fundus image from which noise has been removed; A method for predicting biomarkers using fundus images characterized by including 4. In Claim 1, The step of pre-training the above deep learning model is, A method for predicting biomarkers using fundus images, characterized by the server pre-training the deep learning model using labeled training data containing health biomarkers that act as labels.

5. In Claim 1, The above biomarker prediction step is, The server identifies and extracts key features from the fundus image received by the server, classifies them, and predicts health biomarkers. A method for predicting biomarkers using fundus images characterized by mapping predicted biomarkers to organs associated with potential chronic diseases.

6. In Claim 5, The above biomarker prediction step is, A method characterized by comparing the predicted value of the biomarker with a normal value and an abnormal value, wherein the abnormal value is higher or lower than the normal value of the biomarker.

7. In Claim 1, The above biomarker prediction step is, The above server receives the fundus image and predicts binary biomarkers using a classification model, and A method for predicting biomarkers using a fundus image, characterized in that the server receives the fundus image and predicts continuous biomarkers using a regression model.

8. In Claim 1, The above biomarker prediction step is, The server receives the above fundus image and predicts health biomarkers using a deep learning model including at least one residual block, and A method for predicting biomarkers using fundus images, characterized in that the above residual block is configured to transmit gradients from an intermediate layer to a subsequent layer of the network of the deep learning model.

9. A computer-readable recording medium storing a program for performing a method of predicting biomarkers using fundus images according to claim 1.

Images