Training method for obesity type identification model and related product

Abstract

The invention discloses a training method for an obesity type identification model and a related product. The training method comprises the following steps: preprocessing a retina image and examination information of an examinee to obtain target data containing the retina image; inputting the target data into a recognition model to perform obesity type recognition operation so as to obtain a target thermodynamic diagram; performing target cutting operation based on the target thermodynamic diagram to cut out a key region of the retina image in the target data; and generating a training sample based on the target data and the image of the key area, and training the recognition model for recognizing obesity types in a comparative learning mode. According to the scheme, the obesity type can be recognized in a non-invasive imaging detection mode, and meanwhile the recognition accuracy is improved.

Description

Technical Field

[0001] The present application generally relates to the field of image recognition technology. More specifically, the present application relates to a training method and device for an obesity category recognition model, a device for obesity category recognition, and a computer-readable storage medium. Background Art

[0002] Obesity has become a major public health problem worldwide. It is closely related to a variety of metabolic diseases (such as type 2 diabetes, hypertension, and cardiovascular disease). However, not all obese individuals exhibit these metabolic disorders, which leads to metabolically healthy obesity (MHO) and metabolically unhealthy obesity (MUO). The identification of MHO and MUO involves distinguishing the differences in metabolic health status between obese individuals. Although MHO individuals are overweight or obese, they do not have health problems related to metabolic syndrome, such as hypertension, hyperglycemia, and abnormal lipid levels; MUO individuals are accompanied by these problems. Therefore, the identification of MUO is particularly important for the early diagnosis, prevention, and management of obesity-related metabolic diseases, which helps to reduce medical costs and improve the quality of life of patients. At present, the methods for identifying obesity types are mainly based on weight, body mass index (BMI), waist circumference measurement, and the evaluation of blood biochemical indicators, such as blood sugar, blood lipids, and blood pressure. Although these methods are effective to a certain extent, they have certain limitations. For example, BMI cannot accurately reflect the individual's fat distribution and body fat content. These tests also often require multiple invasive samplings, which is inconvenient for patients.

[0003] In recent years, with the advancement of medical imaging technology, non-invasive imaging detection methods have gradually attracted attention. Retinal images have become a research focus because they can reveal the health status of microvessels non-invasively. Among them, changes in retinal microvessels may be related to metabolic pathology, which provides a new way to use retinal images for obesity identification. By using machine learning or deep learning technology, using complex algorithm models to process a large amount of retinal image data and learn patterns from them, it is possible to reveal the subtle characteristics of metabolic abnormalities in obese people. In practical application scenarios, due to the limited retinal image data, it is often necessary to increase the training samples through data augmentation operations (such as cropping) when training the model. However, the existing data augmentation operation is to directly randomly crop the training data, which will cause key information to be missed during the learning process, resulting in the optimization of the distance between the target and the background in the embedding space. Misleading representation learning reduces the accuracy of the model.

[0004] In view of this, there is an urgent need to provide a training method for an obesity type recognition model, so that obesity types can be identified through non-invasive imaging detection methods while improving the accuracy of recognition. Summary of the invention

[0005] In order to at least solve one or more of the technical problems mentioned above, the present application proposes a training method for an obesity category recognition model and a solution of related products in multiple aspects.

[0006] In a first aspect, the present application provides a training method for a recognition model for obesity types, comprising: preprocessing a subject's retinal image and examination information to obtain target data containing the retinal image; inputting the target data into a recognition model to perform an obesity type recognition operation to obtain a target heat map; performing a target cropping operation based on the target heat map to crop out a key area of ​​the retinal image in the target data; and generating training samples based on the target data and images of the key area to train the recognition model for identifying obesity types in a comparative learning manner.

[0007] In some embodiments, performing a target cropping operation based on the target heat map to crop out the key area of ​​the retinal image in the target data includes: replacing the activation point values ​​in the target heat map that are greater than the target threshold with target indication values ​​to obtain a target matrix; extracting the minimum rectangular area containing all target indication values ​​from the target matrix as a target positioning frame; and using a random cropping function to obtain the positioning information of the cropping frame contained in the target positioning frame to obtain the key area of ​​the retinal image in the target data.

[0008] In other embodiments, using a random cropping function to obtain the positioning information of the cropping frame contained in the target positioning frame to obtain the key area of ​​the retinal image in the target data includes: increasing the variance of random sampling to obtain a U-shaped β random probability distribution with a low center and high surroundings; optimizing the random cropping function using the U-shaped β random probability distribution to reduce the probability that the generated cropping frame appears in the center of the image; and using the optimized random cropping function to obtain the positioning information of the cropping frame contained in the target positioning frame to obtain the key area of ​​the retinal image in the target data.

[0009] In some other embodiments, after using a random cropping function to obtain the positioning information of the cropping frame contained in the target positioning frame to obtain the key area of ​​the retinal image in the target data, it also includes: adjusting the size and position of the cropping frame, the adjustment is used to generate new positioning information to update the key area; and / or adaptively scaling the cropping frame, the scaling is used to generate new positioning information to update the key area.

[0010] In some other embodiments, the size and position of the cropping frame are adjusted, and the adjustment is used to generate new positioning information to update the key area, including: subtracting half of the target fill value from the first coordinate value and the second coordinate value of the center point of the cropping frame to correspondingly replace the first coordinate value and the second coordinate value; and increasing the width value and the height value of the cropping frame by the target fill value to correspondingly replace the width value and the height value.

[0011] In some other embodiments, the cropping frame is adaptively scaled, and the scaling is used to generate new positioning information to update the key area, including: multiplying the width value, height value, and first coordinate value and second coordinate value of the center point of the cropping frame by the target scaling factor respectively to correspondingly replace the width value, the height value, the first coordinate value, and the second coordinate value.

[0012] In some further embodiments, the target filling value is calculated by the following operation: obtaining a width value or a height value of the target positioning box, and the width value or the height value is used to multiply a preset proportionality coefficient to obtain the target filling value.

[0013] In some further embodiments, the target zoom factor is calculated by: dividing the width value of the retinal image by the width value of the target heat map to obtain the target zoom factor; or dividing the height value of the retinal image by the height value of the target heat map to obtain the target zoom factor.

[0014] In some other embodiments, the preprocessing of the retinal image and examination information of the subject includes: collecting the examination information of the subject for data cleaning; comparing the numerical values ​​in the examination information after data cleaning with one or more target thresholds and / or target ranges; and based on the results of the comparison, establishing binary target data labeled as metabolic unhealthy obesity and metabolic healthy obesity for the retinal image in the examination information.

[0015] In a second aspect, the present application provides a device for training an identification model for obesity types, comprising: a processor, a memory, and a system bus; wherein the processor and the memory are connected via the system bus; the memory is used to store one or more programs, and the one or more programs include instructions, which, when executed by the processor, enable the processor to execute any one of the methods described in the first aspect.

[0016] In a third aspect, the present application provides a device for identifying obesity types, comprising: a processor; and a memory on which computer instructions of an obesity type identification model are stored, and when the computer instructions are executed by the processor, the device performs the following operations: obtaining a retinal image to be identified; and inputting the retinal image into a recognition model trained according to the method described in the first aspect above to perform obesity type identification, and obtain an identification result of the obesity type.

[0017] In a fourth aspect, the present application provides a computer-readable storage medium, including a computer program instruction for training an identification model for obesity types stored thereon, which, when executed by one or more processors, implements the method according to any one of the first aspects; or a computer program instruction for obesity type identification stored thereon, which, when executed by one or more processors, implements the operations performed by the apparatus described in the third aspect.

[0018] Through a training scheme for an obesity type recognition model provided as above, the embodiment of the present application obtains a target heat map by inputting target data including retinal images into the recognition model to perform an obesity type recognition operation, and then directly crops out the key areas of the retinal image through a target cropping operation, and adopts a comparative learning method, thereby helping to learn the key area features more effectively and accurately to improve the accuracy of the recognition model, and recognize the obesity type based on the retinal image, which can non-invasively reveal the health status of the microvessels and thus determine the type of obesity. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] By reading the detailed description below with reference to the accompanying drawings, the above and other purposes, features and advantages of the exemplary embodiments of the present application will become easy to understand. In the accompanying drawings, several embodiments of the present application are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0020] Figure 1 An example diagram showing a method 100 for training a recognition model for obesity types according to some embodiments of the present application;

[0021] Figure 2An example diagram showing a data preprocessing method 200 according to some embodiments of the present application;

[0022] Figure 3 An example diagram showing the effect 300 of an existing random cropping strategy is shown;

[0023] Figure 4 An example diagram showing the effect 400 of a data clipping enhancement strategy based on a retinal image heat map according to some embodiments of the present application;

[0024] Figure 5 A schematic block diagram of an apparatus 500 according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0026] It should be understood that the terms "include" and "comprising" used in the specification and claims of the present application indicate the presence of described features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0027] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in this application specification and claims, unless the context clearly indicates otherwise, the singular forms of "a", "an" and "the" are intended to include plural forms. It should also be further understood that the term "and / or" used in this application specification and claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0028] As used in this specification and claims, the term "if" may be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" may be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.

[0029] The specific implementation of the present application is described in detail below with reference to the accompanying drawings.

[0030] Figure 1 An example diagram of a method 100 for training a recognition model for obesity types according to some embodiments of the present application is shown. Figure 1 As shown, at step S101, the retinal image and inspection information of the subject are preprocessed to obtain target data including the retinal image. It should be understood that the inspection information generally includes data such as individual physiological parameters, medical images and laboratory test results. This information is crucial for assessing the health status and disease risk of an individual. Specifically, the aforementioned inspection information may include, for example, the subject's age, gender, BMI, systolic blood pressure / SBP (mmHg), diastolic blood pressure / DBP (mmHg), triacylglycerol content / Triacylglycerols (mmol / l), fasting blood glucose / FPG (mmol / l), high-density lipoprotein cholesterol / HDL-C (mmol / l) and medical history or medication history (such as receiving antihypertensive treatment, taking lipid-lowering drugs, taking diabetes drugs). The aforementioned retinal image may include an image of a fundus color photograph, wherein the fundus color photograph is a medical imaging technology that obtains eye structure and lesion information by taking color photos of the fundus.

[0031] After obtaining the retinal image and examination information of the subject, data preprocessing operations are required. For retinal images, the original retinal images may contain noise due to factors such as shooting conditions or equipment. Therefore, image processing techniques such as median filtering and Gaussian filtering can be used for preprocessing operations to effectively remove random noise in the image. At the same time, in order to observe the details of the retina more clearly, the image can be enhanced. The enhancement processing may include methods such as histogram equalization and contrast adjustment. These methods can improve the visual effect of the image and highlight important detail features. For inspection information, preprocessing operations such as formatting can be performed to ensure that the format of all inspection information is unified, which is convenient for subsequent data processing and analysis. In addition, standardization processing can also be performed at the same time, such as converting the measurement values ​​of different scales into unified standard units to eliminate errors caused by different equipment or measurements.

[0032] In one embodiment, the aforementioned preprocessing operation may include collecting the examination information of the subject for data cleaning, comparing the values ​​in the examination information after data cleaning with one or more target thresholds and / or target ranges, and based on the comparison result, establishing binary classification target data labeled as metabolically unhealthy obesity and metabolically healthy obesity for the retinal image in the examination information. More details about the preprocessing operation will be discussed later in conjunction with Figure 2 Describe in detail.

[0033] Then, at step S102, the target data is input into the recognition model to perform obesity type recognition operation to obtain a target heat map. It should be understood that the heat map is obtained by quantifying the attention of the model to different areas in the retinal image. This process can be achieved, for example, by the forward propagation of the recognition model and the calculation of the feature activation map. Specifically, the class activation mapping ("CAM") technology can be used to generate a heat map. In an embodiment of the present application, it also includes but is not limited to a heat map generated by using methods such as convolutional feature maps, activation maps, saliency maps or attention maps.

[0034] After obtaining the target heat map of the recognition model on the retinal image in the target data, the cropping strategy of the retinal image can be optimized based on the heat map. Further, at step S103, a target cropping operation is performed based on the target heat map to crop out the key areas of the retinal image in the target data. It should be understood that in model training, random cropping is a commonly used data enhancement strategy that aims to increase the diversity and robustness of training data. Random cropping is usually applied to image data, and random cropping operations are performed on the input image during training to generate sub-images with certain changes for training. This helps the model learn features of different positions, scales, and angles, and improves its generalization ability. The random cropping strategy can be expressed by the following formula:

[0035]

[0036] Where s represents the size of the crop area (scale), r is the aspect ratio of the crop (aspectratio), and I is the input image. Represents the random cropping function, which returns a four-tuple (x, y, h, w), which defines the center coordinates (x, y) of the crop, as well as the height h and width w.

[0037] Random cropping is a basic sampling method that makes all individual crops be selected with equal probability. Although this method can simply achieve the sampling of images, it ignores the semantic information that may exist in the image. Especially when dealing with paired views, the random cropping strategy may cause some problems. This shows that in the absence of object prior information (such as scale, position, etc.), direct random cropping can easily lead to missed objects (such as Figure 3 ), so optimizing the distance between the target and the background in the embedding space may mislead the learning of representations. In addition, random cropping does not always carry enough variance to effectively represent the diversity of objects.

[0038] Based on the above reasons and factors such as limited data, the traditional random cropping strategy is not suitable for retinal images. Therefore, the embodiment of the present application defines a new data cropping enhancement strategy, which helps to learn better representations more efficiently and accurately. Specifically, the present application designs an enhanced cropping strategy based on the heat map of the retinal image, which performs a target cropping operation based on the target heat map to crop out the key areas of the retinal image in the target data (such as Figure 4 shown).

[0039] Specifically, semantic perception is performed on the retinal image. In one embodiment, the activation point values ​​in the target heat map that are greater than the target threshold can be replaced with the target indication value to obtain a target matrix. Then, the minimum rectangular area containing all the target indication values ​​is extracted from the target matrix as the target positioning box. In an implementation scenario, the above operation can be expressed by the following formula:

[0040] B=L(1[M>k])

[0041] Where M represents the heat map, k represents the threshold of the activation point, 1[*] is the indicator function (i.e., if the pixel value in M ​​is greater than k, the pixel value is returned, otherwise it returns 0), and L is the function for calculating the rectangular closure, which returns a positioning box B. Considering that the positioning box is not completely accurate, and in order to cover some background information around the object, only the center of the target crop area crop is limited to the box.

[0042] In an exemplary scenario, the aforementioned indication function can be set to return the pixel value itself, and the aforementioned function L for calculating the rectangular closure can adopt, for example, the MIN_RECT function, which represents the calculation of the minimum rectangular border. In other words, the aforementioned target positioning box B can be obtained by calculating the minimum rectangular bounding box. Therefore, the aforementioned B=L(1[M>k]) can be expressed as the following formula:

[0043] B=MIN_RECT({(x,y)|H(x,y)>τ})

[0044] Among them, MIN_RECT represents the function of calculating the minimum rectangular border, H(x,y) represents the pixel value of the heat map, τ represents the threshold of the activation point; (x,y)[*] represents the indicator function, when M(x,y)>τ, it returns the pixel value, otherwise it returns 0.

[0045] After obtaining the target positioning frame B, it is also necessary to use a random cropping function to obtain the positioning information of the cropping frame contained in the target positioning frame to obtain the key area of ​​the retinal image in the target data. In one embodiment, after obtaining the target positioning frame B, the sampling can be expressed as:

[0046]

[0047] Among them, s represents the size of the crop area, r is the aspect ratio of the crop, B is the target positioning box, and the four-tuple The center coordinates of the target cropping area crop are defined respectively and high He Kuan

[0048] In another embodiment, in order to suppress the center sampling to avoid the cropping box always being near the center of the target positioning box, after obtaining the target positioning box B, the variance of the random sampling can be first increased to obtain a U-shaped β random probability distribution with a low center and high surroundings. Then, the random cropping function is optimized using the U-shaped β random probability distribution to reduce the probability that the generated cropping box appears in the center of the image. Subsequently, the optimized random cropping function is used to obtain the positioning information of the cropping box contained in the target positioning box to obtain the key area of ​​the retinal image in the target data. The process can be expressed as:

[0049]

[0050] in, It is a function representing random sampling after center sampling suppression, s represents the size of the crop area crop, r is the aspect ratio of the crop, and B is the target positioning box obtained previously.

[0051] In another embodiment, in order to take into account the position and size of key areas in the image (such as optic cups, optic discs, blood vessels, lesions, etc.) and perform cropping near these areas while maintaining a certain degree of background information, it is necessary to implement an adaptive cropping window size strategy. That is, after using a random cropping function to obtain the positioning information of the cropping frame contained in the target positioning frame to obtain the key area of ​​the retinal image in the target data, it also includes: adjusting the size and position of the cropping frame, the adjustment is used to generate new positioning information to update the key area; and / or adaptively scaling the cropping frame, the scaling is used to generate new positioning information to update the key area. Therefore, the present application defines a function that can dynamically adjust the size and proportion of the cropping window according to the feature map and semantic information of the input image. The aforementioned adaptive strategy can be implemented by padding (p) and an adaptive scaling factor (σ).

[0052] Generally speaking, the padding (p) can be determined according to the size of the bounding box to ensure that the cropping window can cover the bounding box and certain background information around it. The adaptive scaling factor (σ) can be determined according to the size ratio of the heat map and the original image so that the size of the cropping window on the original image is suitable for extracting the area of ​​interest. The calculation of padding (p) and adaptive scaling factor (σ) can be adjusted according to specific needs and actual conditions.

[0053] In one implementation scenario, in order to adjust the size and position of the cropping window, the size and position of the cropping window are adjusted according to the bounding box, and a certain padding p can be added on the basis of the bounding box. The first coordinate value and the second coordinate value of the center point of the cropping box are respectively subtracted by half of the target padding value to correspondingly replace the first coordinate value and the second coordinate value, and the width value and the height value of the cropping box are respectively increased by the target padding value to correspondingly replace the width value and the height value. Specifically, the process can be expressed by the following formula:

[0054] C(x,y,w,h)=(B x -p / 2,B y -p / 2,B w +p,B h +p)

[0055] Among them, (B x ,B y ) is the center position of the cropping frame B, w and B h is the width and height of the cropping box B, C(x, y, w, h) is the center position (x, y) of the adaptively adjusted window and the width w and height h. The calculation method of the target padding value p includes obtaining the width value or height value of the target positioning box, and the width value or height value is used to multiply the preset scale factor to obtain the target padding value. Specifically, the calculation of padding (p) can be based on the width (B of the bounding box) w ) and height (B h ), you can choose a fixed ratio or adjust it according to the specific scenario. For example, you can set the padding to a multiple of the bounding box width or height. As an example, if you choose to fill half the width or height of the bounding box, the padding value is:

[0056]

[0057] Among them, B w Indicates the width of the bounding box, B h Indicates the height of the bounding box.

[0058] In another implementation scenario, considering that the heat map H may be different in size from the original image I, in order to achieve adaptive scaling, a scaling factor σ can be used to adjust the size of the cropping window to make it suitable for the original image. For example, the width value, height value, and first coordinate value and second coordinate value of the center point of the cropping box can be multiplied by the target scaling factor respectively to replace the width value, height value, first coordinate value, and second coordinate value accordingly. Specifically, the process can be expressed by the following formula:

[0059]

[0060] Where C(x, y, w, h) is the cropping window adjusted according to the original image size, C x ,C y ,C w ,C h The first coordinate value and the second coordinate value respectively correspond to the width value, the height value and the center point of the cropping box, and σ represents the scaling factor. In some embodiments, the method for calculating the target scaling factor includes: dividing the width value of the retinal image by the width value of the target heat map to obtain the target scaling factor; or dividing the height value of the retinal image by the height value of the target heat map to obtain the target scaling factor. Specifically, the calculation of the adaptive scaling factor σ can be determined based on the size ratio of the heat map and the size of the original image to ensure that the size of the cropping window on the original image is suitable for extracting the area of ​​interest. In one embodiment, the scaling factor can be calculated by the following formula:

[0061]

[0062] Among them, the width of the heat map is H w , height H h , the width of the original image is I w , height is I h .

[0063] Based on the method described in the above embodiment, the adaptive cropping window strategy can dynamically adjust the size and position of the cropping window according to the characteristics and complexity of the fundus image, so as to better adapt to different types of fundus images. This adaptive processing method can improve the applicability and performance of the algorithm when processing various fundus images.

[0064] In general, through this retinal image-based heat map and data cropping enhancement strategy including semantic perception, center sampling suppression and adaptive cropping window, the model can more effectively learn the features of key areas in fundus color photos, providing strong support for the automated diagnosis and monitoring of eye diseases.

[0065] It should be understood that the method of the present application was trained and tested on a data set containing retinal images, and by introducing a data cropping enhancement strategy, the feature learning effect of key areas was improved.

[0066] According to the above-mentioned cropped key areas, at step S104, training samples are generated based on the target data and the images of the key areas, and the recognition model for identifying the type of obesity is trained in a comparative learning manner. The embodiment of the present application adopts a comparative learning method, so that the model can learn a more robust and discriminative feature representation by comparing the features of different samples, thereby improving the classification performance and generalization ability of the model. In this process, the present application selects the EfficientNet model as the basic model. EfficientNet is an efficient convolutional neural network structure with a low parameter amount and high computational efficiency. Considering that obesity diagnosis is a two-classification problem, the present application designs the model as a two-classification model, that is, judging whether the sample belongs to the category of "metabolic healthy obesity" (MHO) or "metabolic unhealthy obesity" (MUO). In order to evaluate the performance of the model, the present application selects multiple classification indicators, including the area under the ROC curve (AUC), accuracy, recall, specificity, F1 score, and precision. These indicators help to comprehensively evaluate the performance of the model in diagnosing obesity, and can fully reflect the classification ability and stability of the model.

[0067] Figure 2 An example diagram of a data preprocessing method 200 of some other embodiments of the present application is shown. It should be understood that the process sequence in the figure is exemplary and not restrictive. This means that the process sequence of the data preprocessing method 200 involved is not limited to a specific form or design, but is presented as an example. In other words, the process sequence shown in the figure can be regarded as a way of expression to clearly describe related concepts and relationships, and does not exclude other forms of process sequence. Therefore, when interpreting the process sequence in the picture, it should be understood that the data preprocessing method 200 has flexibility and diversity, and its purpose is to provide an exemplary description rather than a restrictive provision on a specific form.

[0068] According to the above embodiments, before constructing an enhanced clipping strategy to train the obesity diagnosis model based on the contrastive learning method, it is necessary to collect physical examination information, which is used for data cleaning to obtain a data set containing retinal images. Specifically, first at step S201, the examination information of the subject is collected for data cleaning. It should be understood that before data cleaning, the collected examination information must first be sorted and standardized for subsequent processing and analysis. Data collection may include the multiple indicators mentioned above (such as BMI, systolic blood pressure, diastolic blood pressure, etc.).

[0069] Data cleaning is an important step in data preprocessing, and its purpose is to eliminate noise and inconsistencies in the data and make the data more suitable for subsequent analysis and modeling. In the collected physical examination information, there may be various types of errors, such as entry errors, outliers, and missing values. These erroneous data will affect the results of subsequent data processing and analysis, so they need to be strictly cleaned and processed. It can be understood that when performing data cleaning, it is necessary to process and filter this information to remove erroneous, missing or abnormal values ​​to ensure the accuracy and completeness of the data. Through data cleaning, a high-quality data set can be obtained, which contains the important information of retinal images.

[0070] When processing datasets containing retinal images, special attention should be paid to the quality and characteristics of the retinal images. Retinal images are usually acquired by devices such as fundus cameras or OCT scanners, which contain information about retinal vessels, macula, optic discs, and other structures. This information is important for diagnosing eye diseases and monitoring disease progression. Therefore, during data cleaning, it is necessary to ensure that the retinal images are clear, complete, and free of artifacts or distortion that affect diagnosis.

[0071] In the data cleaning process, missing values ​​and outliers need to be preprocessed first. Missing value processing refers to filling or deleting missing values ​​in the data to ensure the integrity and reliability of the data. For example, when BMI is missing, height and weight can be used for conversion. Outlier preprocessing is to identify and process abnormal values ​​in the data, which may affect subsequent analysis results. Fundus color photography quality control refers to the quality assessment and screening of the collected fundus color photography to exclude blurred, distorted or unclear photos. Fundus color photography black edge removal refers to the preprocessing of fundus color photography images to remove the black areas at the edges to reduce image noise and interference.

[0072] Secondly, at step S202, the values ​​in the inspection information after data cleaning are compared with one or more target thresholds and / or target ranges. Specifically, in one embodiment, the cleaned data is classified, and the classification labels and the standards of the target thresholds and / or target ranges are as follows:

[0073] Normal: 18.5≤BMI<24.0;

[0074] Overweight: 24.0≤BMI<28.0;

[0075] Obesity: BMI 28.0 ≤ BMI, within the obesity range, and if two or more of the following criteria are met, it is metabolically unhealthy obesity (MUO), otherwise it is metabolically healthy obesity (MHO):

[0076] (1) systolic blood pressure / SBP ≥ 130 mmHg and diastolic blood pressure / DBP ≥ 85 mmHg, or receiving antihypertensive treatment;

[0077] (2) Triacylglycerols ≥ 1.7 mmol / l, or currently taking lipid-lowering drugs;

[0078] (3) fasting blood glucose / FPG ≥ 5.6 mmol / l, or taking diabetes medication;

[0079] (4) High-density lipoprotein cholesterol (HDL-C) < 1.04 mmol / l for men and HDL-C < 1.04 mmol / l for women

[0080] <1.30mmol / l.

[0081] Finally, at step S203 and step S204, based on the comparison results, binary target data labeled as metabolic unhealthy obesity and metabolic healthy obesity are established for the retinal images in the examination information, and data labeled as non-obese data is established. Specifically, when the value after data cleaning is in the normal range or overweight range, the label is non-obese data. When the value after data cleaning is in the obese range, further binary classification is performed, wherein for the value in the fat intestine range, when it meets at least two of the above (1), (2), (3) and (4), the label is metabolic unhealthy obesity; otherwise, the label is metabolic healthy obesity. In one embodiment, the fundus photography data of the physical examination population can be used, and combined with the physical examination results to establish a binary classification data set based on metabolic unhealthy obesity (MUO) of the obese population, otherwise it is metabolic healthy obesity (MHO).

[0082] Figure 3 An example diagram of the effect of the existing random cropping strategy 300 is shown. Figure 3 The left and right figures of are schematic diagrams of existing random cropping, wherein the two dashed boxes in each case represent the results of two random croppings. Figure 3 As can be seen in the examples shown, random cropping can lead to the following problems: in the case of the left image, attention is focused on non-important areas, while in the case of the right image, there may be redundant repetition of information.

[0083] Figure 4 An example diagram of the data cropping enhancement strategy effect 400 based on the retinal image heat map of some other embodiments of the present application is shown. In combination with the effects shown in the figure, the applicant found that the model can focus on the key areas of the fundus color photo during the learning period, such as the optic cup, optic disc, blood vessels, lesions, etc. Therefore, the heat map in the learning process can be used to obtain the object cropping border for comparison in the learning process. Heat represents the importance of different areas in the fundus image. By weighting these importances, the key areas of the fundus image can be effectively captured and learned. At the same time, Figure 4 It is shown that through the guidance of the heat map, the cropping box can effectively appear in the key areas of the retinal image, especially near the blood vessels and pupil positions, and the cropping enhancement strategy shows a relatively reliable improvement effect. This means that the cropping enhancement strategy can effectively improve the accuracy of feature learning in key areas and improve the robustness of the algorithm in complex situations. In short, the embodiment of the present application can significantly improve the contrastive learning effect by applying the cropping enhancement strategy to the existing algorithm. The experimental results of this method on the retinal image dataset show the potential and effectiveness of the cropping enhancement strategy, and provide a valuable reference for research and application in the field of contrastive learning.

[0084] Figure 5 FIG. 5 is a schematic block diagram of a device 500 for training a recognition model of obesity types according to an embodiment of the present application. Figure 5 As shown, the device 500 may include a processor 501 and a memory 502. The memory 502 stores computer instructions for obesity diagnosis model training, and when the computer instructions are executed by the processor 501, the device 500 executes the method described above.

[0085] In some embodiments, the present application also provides a device for identifying obesity types, the device comprising a processor; and a memory, on which computer instructions of an obesity type identification model are stored, and when the computer instructions are executed by the processor, the device performs the following operations: obtaining a retinal image to be identified; and inputting the retinal image into the identification model trained by the method of the present application embodiment to identify the obesity type, and obtaining an identification result of the obesity type. The identification result includes non-obesity, metabolically unhealthy obesity, and metabolically healthy obesity.

[0086] In some embodiments, various aspects of the present application can also be implemented in the form of a program product, which includes a program code, and when the program code is executed by a processor, the program code is used to cause the processor to perform the method described above. The program product can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can be, for example, - but not limited to - an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of readable storage media (non-exhaustive list) include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0087] Although multiple embodiments of the present application have been shown and described herein, it is obvious to those skilled in the art that such embodiments are provided only by way of example. Those skilled in the art can think of many changes, modifications and alternatives without departing from the thought and spirit of the present application. It should be understood that in the process of practicing the present application, various alternatives to the embodiments of the present application described herein can be adopted. The attached claims are intended to limit the scope of protection of the present application, and therefore cover equivalents or alternatives within the scope of these claims.

[0088] The collection and acquisition of various data in this application complies with relevant laws and regulations and is authorized by the data provider. Any organization or individual that needs to obtain external data must obtain authorization in accordance with the law and ensure data security. It is not allowed to illegally collect, use, process, or transmit unauthorized or unprotected data, or to illegally buy, sell, provide, or disclose unauthorized or unprotected data.

Claims

1. A training method for an obesity type recognition model, characterized in that: include: Preprocessing the retinal image and examination information of the subject to obtain target data including the retinal image; Inputting the target data into a recognition model to perform an obesity type recognition operation to obtain a target heat map; Performing a target cropping operation based on the target heat map to crop out a key area of ​​the retinal image in the target data; as well as A training sample is generated based on the target data and the image of the key area, and the recognition model for identifying obesity types is trained in a comparative learning manner.

2. The training method according to claim 1, characterized in that: The target cropping operation is performed based on the target heat map to crop out the key area of ​​the retinal image in the target data, including: Replacing the activation point values ​​in the target heat map that are greater than the target threshold with the target indication values ​​to obtain a target matrix; Extracting a minimum rectangular area containing all target indication values ​​from the target matrix as a target positioning frame; and The positioning information of the cropping frame contained in the target positioning frame is acquired by using a random cropping function to obtain the key area of ​​the retinal image in the target data.

3. The training method according to claim 2, characterized in that: The random cropping function is used to obtain the positioning information of the cropping frame contained in the target positioning frame to obtain the key area of ​​the retinal image in the target data, including: Increase the variance of random sampling to obtain a U-shaped β random probability distribution with a low center and high surroundings; Optimizing the random cropping function using the U-shaped β random probability distribution to reduce the probability that the generated cropping box appears in the center of the image; and The optimized random cropping function is used to obtain the positioning information of the cropping frame contained in the target positioning frame to obtain the key area of ​​the retinal image in the target data.

4. The training method according to claim 2, characterized in that: in, After obtaining the positioning information of the cropping frame contained in the target positioning frame by using the random cropping function to obtain the key area of ​​the retinal image in the target data, the method further includes: Adjusting the size and position of the cropping frame, wherein the adjustment is used to generate new positioning information to update the key area; and / or The cropping frame is adaptively scaled, and the scaling is used to generate new positioning information to update the key area.

5. The training method according to claim 4, characterized in that: The adjusting of the size and position of the cropping frame, wherein the adjusting is used to generate new positioning information to update the key area, includes: Subtract half of the target fill value from the first coordinate value and the second coordinate value of the center point of the cropping frame, respectively, to replace the first coordinate value and the second coordinate value accordingly; and The width value and the height value of the cropping frame are respectively increased by the target filling value to correspondingly replace the width value and the height value.

6. The training method according to claim 4, characterized in that: The step of adaptively scaling the cropping frame, wherein the scaling is used to generate new positioning information to update the key area, includes: The width value, the height value, and the first coordinate value and the second coordinate value of the center point of the cropping frame are respectively multiplied by the target scaling factor to correspondingly replace the width value, the height value, the first coordinate value, and the second coordinate value.

7. The training method according to claim 5, characterized in that: The target filling value is calculated by the following operations: The width value or the height value of the target positioning frame is obtained, and the width value or the height value is used to multiply the preset proportional coefficient to obtain the target filling value.

8. The training method according to claim 6, characterized in that: The target scaling factor is calculated by the following operations: Dividing the width value of the retinal image by the width value of the target heat map to obtain the target scaling factor; or The height value of the retinal image is divided by the height value of the target heat map to obtain the target scaling factor.

9. The training method according to claim 1, characterized in that: The preprocessing of the subject's retinal image and examination information includes: Collect the examination information of the examinee for data cleaning; Comparing the values ​​in the cleaned inspection information with one or more target thresholds and / or target ranges; and Based on the comparison result, binary classification target data labeled as metabolically unhealthy obesity and metabolically healthy obesity are established for the retinal image in the inspection information.

10. A device for training a recognition model for obesity types, characterized in that: include: A processor, a memory, and a system bus; wherein the processor and the memory are connected via the system bus; The memory is used to store one or more programs, wherein the one or more programs include instructions, and when the instructions are executed by the processor, the processor executes the method according to any one of claims 1 to 9.

11. A device for identifying obesity types, characterized in that: include: processor; as well as A memory having computer instructions for an obesity category identification model stored thereon, wherein when the computer instructions are executed by a processor, the device performs the following operations: acquiring a retinal image to be identified; and The retinal image is input into a recognition model trained by the method according to any one of claims 1 to 9 to perform obesity type recognition to obtain a recognition result of the obesity type.

12. A computer-readable storage medium, characterized in that: It includes computer program instructions stored thereon for training an identification model for obesity types, which, when executed by one or more processors, implement the method according to any one of claims 1 to 9; or it includes computer program instructions stored thereon for identifying obesity types, which, when executed by one or more processors, implement the operations performed by the device according to claim 11.

Images