Methods and related products for training classification models for metabolic diseases
By training a metabolic disease classification model using techniques such as serpentine convolution modules and high-order continuous constraints, the problem of incomplete retinal image feature capture in existing technologies has been solved, achieving higher classification accuracy and precision.
Patent Information
- Application Number
- CN202410421934.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-09
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-04-09
AI Technical Summary
Existing methods for identifying metabolic diseases are inefficient, especially classification models based on retinal images, which struggle to capture complex fundus vascular features, resulting in poor accuracy.
The classification model is trained using a snake convolution module, a fusion module, and a classifier module. Through snake convolution operations, multi-view fusion operations, and classification operations, combined with high-order continuous constraints and morphological loss functions, the shape of the snake convolution kernel is dynamically adjusted, and confidence and adaptive weight layers are set to improve the accuracy and precision of the model.
It significantly improves the accuracy and precision of metabolic disease classification models, better captures complex features in retinal images, reduces overfitting, and enhances model performance.
Smart Images

Figure CN118196535B_ABST
Abstract
Description
Technical Field
[0001] This application generally relates to the field of artificial intelligence technology. More specifically, this application relates to a method, apparatus, device, and computer-readable storage medium for training a classification model for metabolic diseases. Background Technology
[0002] The complete identification of metabolic diseases (i.e., endocrine and nutritional metabolic diseases) should include etiological examination, pathological examination (qualitative and localization examination), and functional examination. First, the identification process needs to consider whether it is an endocrine disorder. Second, it's necessary to differentiate between diseases caused by the endocrine glands or tissues themselves and secondary endocrine syndromes, such as dwarfism secondary to liver or kidney disease or malnutrition. This identification process often requires analyzing multiple aspects, including the patient's clinical manifestations, laboratory data, glandular function tests, imaging examinations, and pathological examinations, to determine the location of the lesion, its pathological nature, the presence of a tumor, and whether it is benign or malignant. The location of the primary lesion should also be identified. This makes the identification efficiency of metabolic diseases relatively low. Furthermore, the identification of endocrine and metabolic diseases is more difficult because the etiology of many endocrine diseases is still unknown or lacks diagnostic methods. Therefore, early identification is crucial to facilitate treatment.
[0003] Currently, fundus images (or retinal images) have been found to have significant clinical value in the detection and screening of certain diseases. For example, fine segmentation of fundus vessels in retinal images can ensure the accuracy of subsequent predictions or classifications. Furthermore, artificial intelligence, especially machine learning and deep learning, has provided better solutions for disease detection and prevention using fundus images. However, due to the tubular structure of fundus vessels, the main challenge stems from the combination of slender, weak local structural features and complex, variable global morphological features. Existing methods utilize the slender, continuous tubular structure in retinal images to simultaneously enhance perception in the three stages of a neural network model: feature extraction, feature fusion, and loss constraint. However, the convolutional kernel used in feature extraction only considers first-order continuity (i.e., continuity between front and back positions). This makes it difficult for the model to capture more complex structural features, especially when the target structure exhibits high-frequency changes, resulting in poor model accuracy.
[0004] Therefore, there is an urgent need to provide a scheme for training classification models for metabolic diseases in order to better capture complex features in retinal images and improve the accuracy of classification models. Summary of the Invention
[0005] In order to at least address one or more of the technical problems mentioned above, this application proposes a scheme for training classification models for metabolic diseases in several aspects.
[0006] In a first aspect, this application provides a method for training a classification model for metabolic diseases, wherein the classification model includes a serpentine convolution module, a fusion module, and a classifier module, and the method includes: inputting a raw retinal image and a retinal grayscale image into the classification model respectively; sequentially performing a serpentine convolution operation, a multi-view fusion operation, and a classification operation via the serpentine convolution module, the fusion module, and the classifier module to obtain corresponding first classification results and second classification results; determining a final classification result for the metabolic disease based on the first classification result and the second classification result, and calculating an initial loss function for the classification model; introducing higher-order continuous constraint terms from the serpentine convolution operation process into the initial loss function to form a first loss function; and training the classification model for metabolic diseases based on the first loss function.
[0007] In one embodiment, the serpentine convolution operation includes serpentine convolution in the horizontal direction and serpentine convolution in the vertical direction, and the higher-order continuity constraint term includes a first higher-order continuity constraint term for the serpentine convolution in the horizontal direction and a second higher-order continuity constraint term for the serpentine convolution in the vertical direction.
[0008] In another embodiment, the first higher-order continuous constraint term and the second higher-order continuous constraint term are obtained by calculating the higher-order derivatives of the position changes of the contour points of the serpentine convolution kernel in the horizontal direction and the vertical direction, respectively.
[0009] In yet another embodiment, the method further includes: adding a morphological loss function from the serpentine convolution operation to the first loss function to form a second loss function; and training the classification model for metabolic diseases based on the second loss function.
[0010] In another embodiment, the serpentine convolution module is followed by a shape generator, and the shape loss function is constructed by: using the shape generator to predict the shape of the serpentine convolution kernel based on the features extracted by the serpentine convolution module, to obtain a shape prediction result; and constructing the shape loss function based on the difference between the shape prediction result and the target shape.
[0011] In yet another embodiment, the method further includes setting a weight layer before the multi-view fusion operation to adjust the weight of each view feature.
[0012] In yet another embodiment, determining the final classification result of a metabolic disease based on the first classification result and the second classification result includes: setting corresponding confidence levels for the first classification result and the second classification result, respectively; and determining the final classification result of the metabolic disease based on the first classification result and the second classification result after setting the corresponding confidence levels.
[0013] In a second aspect, this application provides an apparatus for training a classification model for metabolic diseases, comprising: a processor; and a memory storing program instructions for training the classification model for metabolic diseases, wherein when the program instructions are executed by the processor, the apparatus implements the various embodiments of the first aspect described above.
[0014] In a third aspect, this application provides an apparatus for classifying metabolic diseases, comprising: a processor; and a memory storing program instructions for classifying metabolic diseases, wherein when the program instructions are executed by the processor, the apparatus performs the following operations: acquiring a raw retinal image and a retinal grayscale image; and inputting the raw retinal image and the retinal grayscale image into a classification model trained according to several embodiments of the first aspect above for classifying metabolic diseases, so as to obtain a final classification result of the metabolic disease.
[0015] In a fourth aspect, this application provides a computer-readable storage medium having computer-readable instructions stored thereon for training a classification model for metabolic diseases, wherein when executed by one or more processors, the computer-readable instructions implement the various embodiments of the first aspect described above; or having computer-readable instructions stored thereon for classifying metabolic diseases, wherein when executed by one or more processors, the computer-readable instructions implement the operations performed by the apparatus in the third aspect described above.
[0016] Using the above-described scheme for training a classification model for metabolic diseases, this embodiment of the application, based on the original retinal image and retinal grayscale image, sequentially performs serpentine convolution operations, multi-view fusion operations, and classification operations using a serpentine convolution module, a fusion module, and a classifier module to obtain first and second classification results and determine the initial loss function of the classification model. Then, by introducing higher-order continuous constraint terms from the serpentine convolution operation into the initial loss function, a first loss function is obtained to train the classification model. Based on this, by adding higher-order continuous constraint terms to the serpentine convolution operation, the classification model can better adapt to and capture complex features in the retinal image, thereby facilitating more accurate classification results and improving the accuracy of the classification model. Furthermore, this embodiment of the application also adds a morphological loss function to the aforementioned first loss function, enabling dynamic adjustment of the morphology of the serpentine convolution kernel to ensure the performance and accuracy of the classification model.
[0017] Furthermore, embodiments of this application also set a weight layer before the multi-view fusion operation to adaptively select features from important views, thereby avoiding overfitting and improving model performance. In addition, embodiments of this application also set confidence levels to fully utilize the advantages of each classification result, thereby obtaining better classification results and greatly improving model accuracy. Attached Figure Description
[0018] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily understood by reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of this application are illustrated by way of example and not limitation, and the same or corresponding reference numerals denote the same or corresponding parts, wherein:
[0019] Figure 1 This is an exemplary schematic diagram showing the original retinal image and the grayscale image of the retina;
[0020] Figure 2 This is an exemplary flowchart illustrating a method for training a classification model for metabolic diseases according to an embodiment of this application;
[0021] Figure 3 This is an exemplary schematic diagram illustrating the overall training of a classification model for metabolic diseases according to an embodiment of this application;
[0022] Figure 4 This is an exemplary schematic diagram illustrating a serpentine convolution kernel and a standard convolution kernel according to embodiments of this application;
[0023] Figure 5 This is an exemplary structural block diagram illustrating an apparatus for classifying metabolic diseases according to an embodiment of this application;
[0024] Figure 6 This is an exemplary structural block diagram illustrating an apparatus for training a classification model for metabolic diseases according to an embodiment of this application. Detailed Implementation
[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0026] It should be understood that the terms "comprising" and "including" used in the specification and claims of this application indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.
[0027] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this specification and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this specification and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.
[0028] As used in this specification and claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."
[0029] As described in the background section above, the identification of metabolic diseases (i.e., endocrine and nutritional metabolic diseases) often requires analysis from multiple aspects, including the patient's clinical presentation, laboratory data, glandular function tests, imaging examinations, and pathological examinations. Laboratory data includes, for example, biochemical changes, hormone levels, immunological and genetic tests. Glandular function tests include stimulation tests, inhibition tests, provocation tests, and antagonism tests. These tests not only determine endocrine dysfunction but some can also help identify the location and cause of the lesion. For example, the thyrotropin-releasing hormone ("TRH") stimulation test can help differentiate whether the lesion in hypothyroidism is located in the hypothalamus, pituitary gland, or thyroid gland, while the high-dose dexamethasone suppression test can help differentiate whether adrenal lesions are hyperplasia or tumors. Imaging examinations, including X-ray examinations, radionuclide scans, computed tomography ("CT"), B-mode ultrasound, and magnetic resonance imaging ("MRI"), are significant for lesion localization and etiological determination, but their identification efficiency is relatively low, leading to missed opportunities for optimal treatment.
[0030] Existing methods for identifying metabolic diseases include genomics-based predictive models, biomarker-based predictive models, separate models designed for diabetes, dyslipidemia, obesity, metabolic syndrome (MS), hyperuricemia, and deep learning models, each with its own limitations. For example, genomics-based predictive models use an individual's genomic information to predict their disease risk. These models can use different algorithms, such as support vector machines and neural networks, to predict whether an individual has endocrine and metabolic diseases, but they suffer from the difficulty and high cost of obtaining genomic data. Biomarker-based predictive models use biomarkers (such as blood glucose, cholesterol, and thyroid hormones) to predict whether an individual has endocrine and metabolic diseases. These models can use statistical methods and machine learning algorithms to predict disease risk, but obtaining biomarkers is difficult, often requiring invasive medical examinations, which is not conducive to large-scale screening.
[0031] Separate models have been designed for diabetes, dyslipidemia, obesity, metabolic syndrome (MS), and hyperuricemia, with outputs being classification results. These models can only predict whether a patient has a specific disease, not multiple diseases simultaneously. Furthermore, while deep learning models are trained using data such as fundus images, coronal capillary imaging, glow images of the ten fingers, total cholesterol, low-density lipoprotein, systolic blood pressure, C-reactive protein, and erythrocyte sedimentation rate, some image data is difficult to obtain and expensive.
[0032] Currently, fundus images (or retinal images) have been found to have significant clinical value in the detection and screening of certain diseases. For example, fine segmentation of fundus vessels in retinal images can ensure the accuracy of subsequent predictions or classifications. However, because fundus vessels are tubular structures, the main challenge stems from the delicate and subtle local structural features and the complex and variable global morphological features. Regarding delicate and subtle local structural features, the vascular structures that reflect lesion characteristics only occupy a small portion of the entire image (e.g., Figure 1 As shown in the image, the composition of pixels is finite. Furthermore, these structures are easily affected by complex backgrounds, making it difficult for the model to accurately distinguish subtle changes in the target. Regarding complex and variable global morphological features, the morphological changes of targets located in different regions depend on the number of branches, the location of forks, the path length, and their position in the image. Therefore, when data exhibits unseen morphological features, the model tends to overfit to previously seen features, failing to recognize unseen morphological features, resulting in weak generalization.
[0033] Figure 1 This is an exemplary schematic diagram showing the original retinal image and the retinal grayscale image. (Example...) Figure 1In Figure (a), the left and right sides respectively show the standard raw retinal image and the corresponding grayscale retinal image. Figure 1 Figure (b) shows magnified views of the original retinal image and its corresponding grayscale image on the left and right sides, respectively. As can be seen, the vascular structure occupies only a small portion of the entire image and displays a complex and varied morphology of slender tubular structures. As mentioned earlier, existing methods utilize the slender and continuous tubular structure in retinal images to simultaneously enhance perception in the following three stages of the neural network model: feature extraction, feature fusion, and loss constraint. However, the convolutional kernel used in feature extraction only considers first-order continuity (i.e., continuity between preceding and following positions). This makes it difficult for the model to capture more complex structural features, especially when the target structure exhibits high-frequency changes, resulting in poor model accuracy.
[0034] Based on this, this application provides a scheme for training a classification model for metabolic diseases. By introducing higher-order continuous constraints, morphological loss functions, adaptive weight layers, and confidence functions in the feature extraction, feature fusion, and classification processes, the accuracy of the classification model can be improved, and more accurate classification results for metabolic diseases can be obtained.
[0035] The specific embodiments of this application will now be described in detail with reference to the accompanying drawings.
[0036] Figure 2 This is an exemplary flowchart illustrating a method 200 for training a classification model for metabolic diseases according to an embodiment of this application. In one embodiment, the classification model may include a serpentine convolution module, a fusion module, and a classifier module. Figure 2 As shown, in step S201, the original retinal image and the retinal grayscale image are input into the classification model, respectively. The serpentine convolution module, the fusion module, and the classifier module sequentially perform serpentine convolution operations, multi-view fusion operations, and classification operations to obtain their respective first and second classification results. That is, the original retinal image and the retinal grayscale image are input into the classification model. First, the serpentine convolution module performs a serpentine convolution operation to obtain features. Then, the fusion module performs a multi-view fusion operation on the extracted features to obtain multiple viewpoint features. These multiple viewpoint features are then input into the classifier module to perform a classification operation, obtaining the first and second classification results based on the original retinal image and the retinal grayscale image, respectively. In some embodiments, the aforementioned first and second classification results may include probability values corresponding to multiple disease categories. These multiple disease categories may include, for example, various endocrine and nutritional metabolic diseases such as dyslipidemia, obesity, metabolic syndrome (MS), and hyperuricemia.
[0037] Next, in step S202, the final classification result of the metabolic disease is determined based on the first classification result and the second classification result, and the initial loss function of the classification model is calculated. In one embodiment, the final classification result of the metabolic disease can be determined based on the first classification result and the second classification result after setting corresponding confidence levels for the first classification result and the second classification result with corresponding confidence levels. It is understood that a simple ensemble method of averaging or voting on the outputs of different models may not fully utilize the advantages of each model, especially when a certain model has high confidence in predicting a certain class. Therefore, this embodiment of the application sets confidence levels for the first and second classification results, for example, by introducing a confidence-based weight function to realize a confidence-based dynamic weight allocation mechanism, which allows the classification model to better utilize the advantages of each classification result in determining the final classification result (i.e., ensemble learning) and obtain a better classification result. As an example, the confidence level can be set by the following formula:
[0038] P = ∑(w_i*P_i) / ∑w_i (1)
[0039] Here, P represents the final classification result, P_i is the i-th classification result, w_i is the weight of the corresponding classification result, and w_i = confidence(P_i), i.e., the set confidence level. Therefore, when a corresponding classification result has a high confidence level, its weight in the final classification result will increase. It can be understood that the final classification result also includes various endocrine and nutritional metabolic diseases such as dyslipidemia, obesity, metabolic syndrome (MS), and hyperuricemia, and the disease category corresponding to the largest P-value is taken as the final identified disease.
[0040] In one implementation scenario, the initial loss function can be obtained based on the difference between the output of the classification model (i.e., the final classification result) and the true classification result, and it is used to measure the matching degree between the output of the classification model and the classification task. As an example, this initial loss function can be denoted as L_origin. Based on the initial loss function obtained above, in step S203, higher-order continuous constraint terms from the serpentine convolution operation are introduced into the initial loss function to form a first loss function. In one embodiment, the aforementioned serpentine convolution operation includes serpentine convolution in the horizontal direction (e.g., the x-direction) and serpentine convolution in the vertical direction (e.g., the y-direction), and the higher-order continuous constraint terms include a first higher-order continuous constraint term for the serpentine convolution in the horizontal direction and a second higher-order continuous constraint term for the serpentine convolution in the vertical direction.
[0041] In the above serpentine convolution operation, a serpentine convolution kernel is used (e.g., Figure 4 As shown in Figure (b), it can be understood as converting standard convolutional kernels (e.g., Figure 4As shown in (a), it is straightened in both the x and y directions. In an exemplary scenario, assuming a convolution kernel of size 9 is used, and taking the x direction as an example, the specific position of each cell can be represented as K. i±c =(x i±c ,y i±c ), where c = {0, 1, 2, 3, 4}, which represents the horizontal distance from the center grid; K i±c This represents a position of the convolution kernel K, and this position is obtained through an accumulation process. Specifically, starting from the center position K... i Initially, the position of a cell furthest from the center depends on the position K of the previous cell. i±1 Therefore, relative to the center position K i The position of the cell preceding it, K i±1 An offset Δ = {δ|δ∈[-1,1]} is added. In this scenario, a serpentine convolution in the x-direction can be obtained:
[0042]
[0043] Similarly, a serpentine convolution in the y-direction can be obtained:
[0044]
[0045] Based on the aforementioned horizontal and vertical serpentine convolutions, the contour points of the convolution kernel can be obtained, which can be described by, for example, some control points on a Bezier curve. Furthermore, a first higher-order continuity constraint term for the horizontal serpentine convolution and a second higher-order continuity constraint term for the vertical serpentine convolution can be obtained. In one implementation scenario, the first and second higher-order continuity constraint terms can be obtained by calculating the higher-order derivatives of the positional changes of the contour points of the serpentine convolution kernel in the horizontal and vertical directions, respectively. Taking second-order as an example, the aforementioned first higher-order continuity constraint term can be expressed as d 2 offset_x / dx 2 , representing the second derivative of the horizontal position change of the contour points of the serpentine convolution kernel, i.e., the curvature in the x-direction. The aforementioned second higher-order continuity constraint term can be expressed as d 2 offset_y / dy 2 , which represents the second derivative of the vertical position change of the contour points of the shaped convolution kernel, that is, the curvature in the y-direction.
[0046] In one implementation scenario, assuming the first loss function is denoted as L_task, it can be formed using the following formula (e.g., cross-entropy or mean squared error):
[0047] L_task=L_origin+λ_x*∑|d 2offset_x / dx 2 |+λ_y*∑|d 2 offset_y / dy 2 |(5)
[0048] Here, L_origin represents the initial loss function, and λ_x and λ_y represent hyperparameters used to balance the importance of second-order continuity in the x and y directions, respectively. Therefore, by subsequently optimizing this first loss function, we can maintain first-order continuity while considering higher-order continuity, thereby capturing more complex features in retinal images and improving the accuracy of classification results.
[0049] After obtaining the first loss function, in step S204, the classification model for metabolic diseases is trained based on the first loss function. For example, the parameters in the classification model are updated backward using gradient descent to complete the training of the classification model for metabolic diseases.
[0050] As described above, this embodiment of the application obtains a first loss function by introducing a high-order continuous constraint term from the serpentine convolution operation into the initial loss function of the classification model, thereby training the classification model. This allows the classification model to better adapt to and capture complex features in retinal images, thus facilitating more accurate classification results and improving the accuracy of the classification model. Furthermore, this embodiment of the application obtains even better classification results by combining the first and second classification results based on the original retinal image and the corresponding retinal grayscale image, and by setting dynamically assigned weights based on confidence levels.
[0051] In one embodiment, this application further proposes adding a morphological loss function from the serpentine convolution operation to the first loss function to form a second loss function, and then training a classification model for metabolic diseases based on the second loss function. Specifically, in one implementation scenario, a morphological generator can be connected after the serpentine convolution module so that the morphological generator can predict the morphology of the serpentine convolution kernel based on the features extracted by the serpentine convolution module, obtain the morphological prediction result, and then construct a morphological loss function based on the difference between the morphological prediction result and the target morphology.
[0052] It's important to understand that the shape generator described above is an auxiliary network responsible for generating a set of parameters describing the shape of the convolutional kernel. During training, it receives feature maps from the previous convolutional layer in the serpentine convolutional module as input and outputs a series of control points or transformation parameters that define the spatial shape changes of the convolutional kernel. As mentioned earlier, the control points are points on the Bézier curve used to describe the contour of the serpentine convolutional kernel, and the coordinates of each control point can be used to construct the final dynamic serpentine convolutional kernel. The aforementioned transformation parameters are mathematical parameters used to describe the transformation of the serpentine convolutional kernel from its initial shape to a new shape. In some embodiments, the aforementioned transformation parameters may include six parameters: two for translation, two for scaling, one for rotation, and one for shearing. Based on the aforementioned parameters output by the shape generator, the shape of the serpentine convolutional kernel can be adjusted during each forward propagation. Furthermore, the shape generator learns the complexity of the input feature maps and predicts the serpentine convolutional kernel shape best suited to the current features, constructing a shape loss function based on the difference between the shape generator and the target shape.
[0053] In an exemplary scenario, assuming the above morphological loss function is denoted as L_shape, the second loss function (i.e., the total loss function) L_total can be obtained by the following formula:
[0054] L_total=L_task+α*L_shape(6)
[0055] Based on the obtained second loss function, the parameters in the classification model are updated using methods such as gradient descent to complete the training of the classification model and further improve the accuracy of the classification model.
[0056] In one embodiment, this application further proposes setting a weight layer before the multi-view fusion operation to adjust the weight of each view feature. This is because different view features may have different importance. When all view features are treated equally, the model may not fully utilize information from important viewpoints and may over-rely on unimportant or redundant view features. Therefore, this application adds an adaptive weight layer before multi-view feature fusion to provide an adaptive feature selection mechanism, enabling the classification model to automatically learn and select more important view features. Specifically:
[0057] F = W * F_origin(7)
[0058] Here, F_origin represents the original multi-view features, and W is a weight matrix. This weight matrix allows the classification model to assign weights based on the importance of each view feature. It can be understood that the aforementioned importance represents the contribution of each view feature to the final classification task, which is automatically learned through gradient descent during training. The aforementioned weight matrix W is not a fixed value but is optimized as part of the classification model parameters. Specifically, each element in W corresponds to a view feature in F_origin. After model training, a set of weight values can be found such that the weighted feature F maximizes the model's performance metrics (such as accuracy, IOU, etc.).
[0059] Figure 3 This is an exemplary schematic diagram illustrating the overall training of a classification model for metabolic diseases according to an embodiment of this application. Figure 3 As shown, the classification model may include a serpentine convolution module 301, a fusion module 302, and a classifier module 303. The original retinal image 304 and the retinal grayscale image 305 are input into the classification model, and the serpentine convolution operation, multi-view fusion operation, and classification operation are sequentially performed by the serpentine convolution module 301, the fusion module 302, and the classifier module 303 to obtain the corresponding first classification result 306 and second classification result 307. In some embodiments, the aforementioned first and second classification results may include probability values corresponding to various endocrine and nutritional metabolic diseases such as dyslipidemia, obesity, metabolic syndrome (MS), and hyperuricemia. In one embodiment, based on the aforementioned first and second classification results, a confidence function can be set based on the above formula (1), that is, through an integration module 308, the final classification result 309 is output. The final classification result 309 is the disease category corresponding to the highest final probability value among various endocrine and nutritional metabolic diseases such as dyslipidemia, obesity, metabolic syndrome (MS), and hyperuricemia.
[0060] Based on the difference between the output of the classification model (i.e., the final classification result) and the true classification result, an initial loss function L_origin can be constructed. By introducing a higher-order continuous constraint term in the snake convolution operation process into the initial loss function based on the above formula (5), a first loss function L_task can be formed to train the classification model. Based on this, more complex features in the retinal image can be captured to improve the accuracy of the classification result. In one implementation scenario, a morphological loss function L_shape can also be introduced into the first loss function L_task to obtain a second loss function L_total (see the above formula (6)) to train the classification model. This enables the dynamic adjustment of the shape of the snake convolution kernel to ensure the performance and accuracy of the classification model. Specifically, a shape generator (not shown in the figure) can be connected after the snake convolution module 301 so that the shape generator can predict the shape of the snake convolution kernel based on the features extracted by the snake convolution module to obtain the shape prediction result, and then construct the shape loss function based on the difference between the shape prediction result and the target shape.
[0061] As described above, this embodiment adds an adaptive weight layer (e.g., formula (7) above) before the multi-view feature fusion is performed in the fusion module 302. This adaptive feature selection mechanism enables the classification model to automatically learn and select more important viewpoint features, thereby preventing the model from overfitting and improving model performance. For more details on the training of the classification model, please refer to the above... Figure 2 The content described herein will not be repeated here.
[0062] Figure 4 This is an exemplary schematic diagram illustrating a serpentine convolution kernel and a standard convolution kernel according to embodiments of this application. Figure 4 Figure (a) shows a standard convolution kernel, which has a square outline, as shown in c1, c2, c3, c4, c5, c6, c7, c8, and c9 in the figure. The standard convolution is calculated by weighting the standard convolution kernel with the pixels at its corresponding positions. Figure 4 Figure (b) shows a serpentine convolution kernel, which has an irregular outline, as shown in P1, P2, P3, P4, P5, P6, P7, P8, and P9 in the figure. Referring to the previous description, the calculation of the serpentine convolution is performed by starting from the center position (e.g., P5) K... i Initially, the position of a cell furthest from the center depends on the position K of the previous cell. i±1 The results are accumulated and obtained, as detailed in formulas (3) and (4) above. In this embodiment, by adding a high-order continuity constraint term to the serpentine convolution operation, each convolution position is based on its previous position and can freely choose the swing direction, thereby ensuring the continuity of extracted features while allowing for free selection.
[0063] Figure 5 This is an exemplary structural block diagram illustrating an apparatus 500 for classifying metabolic diseases according to an embodiment of this application. Figure 5 As shown, the device 500 may include a processor 501 and a memory 502, wherein the processor 501 and the memory 502 communicate via a bus. The memory 502 stores program instructions for classifying metabolic diseases. When these program instructions are executed by the processor 501, the following operations are performed: acquiring a raw retinal image and a retinal grayscale image; then inputting the acquired raw retinal image and retinal grayscale image into a classification model trained according to this embodiment of the application for metabolic disease classification, to obtain the final classification result of the metabolic disease. That is, based on the above-mentioned trained classification model, by inputting the raw retinal image and the corresponding retinal grayscale image into the classification model, the final classification result of the metabolic disease can be directly obtained.
[0064] Figure 6 This is an exemplary structural block diagram illustrating a device 600 for training a classification model for metabolic diseases according to an embodiment of this application.
[0065] like Figure 6 As shown, the device of this application may include a central processing unit (“CPU”) 611, which may be a general-purpose CPU, a special-purpose CPU, or other information processing and program execution unit. Furthermore, the device 600 may also include a mass storage 612 and a read-only memory (“ROM”) 613. The mass storage 612 may be configured to store various types of data, including various raw retinal images, retinal grayscale images, first and second classification results, final classification results, algorithm data, intermediate results, and various programs required to run the device 600. The ROM 613 may be configured to store power-on self-test (POST) data for the device 600, initialization of various functional modules in the system, drivers for the system's basic input / output, and data and instructions required to boot the operating system.
[0066] Optionally, device 600 may also include other hardware platforms or components, such as the tensor processing unit (“TPU”) 614, graphics processing unit (“GPU”) 615, field-programmable gate array (“FPGA”) 616, and machine learning unit (“MLU”) 617 shown. It is understood that although various hardware platforms or components are shown in device 600, they are merely exemplary and not limiting, and those skilled in the art can add or remove appropriate hardware as needed. For example, device 600 may include only a CPU, associated storage devices, and interface devices to implement the method of training a classification model for metabolic diseases as described in this application.
[0067] In some embodiments, to facilitate data transmission and interaction with external networks, the device 600 of this application further includes a communication interface 618, through which it can connect to a local area network / wireless local area network (“LAN / WLAN”) 605, and further through the LAN / WLAN to connect to a local server 606 or to the Internet (“Internet”) 607. Alternatively or additionally, the device 600 of this application can also directly connect to the Internet or cellular network via the communication interface 618 based on wireless communication technology, such as wireless communication technology based on 3G (“3G”), 4G (“4G”), or 5G (“5G”). In some application scenarios, the device 600 of this application can also access the server 608 and database 609 of an external network as needed to obtain various known algorithms, data, and modules, and can remotely store various data, such as various data or instructions used to present raw retinal images, retinal grayscale images, first and second classification results, final classification results, etc.
[0068] Peripherals of device 600 may include a display device 602, an input device 603, and a data transmission interface 604. In one embodiment, the display device 602 may include, for example, one or more speakers and / or one or more visual displays, configured to provide voice prompts and / or display images and videos for training classification models for metabolic diseases according to this application. The input device 603 may include, for example, a keyboard, mouse, microphone, posture capture camera, and other input buttons or controls, configured to receive audio data input and / or user commands. The data transmission interface 604 may include, for example, a serial interface, parallel interface, or Universal Serial Bus interface (“USB”), Small Computer System Interface (“SCSI”), Serial ATA, FireWire (“FireWire”), PCI Express, and High Definition Multimedia Interface (“HDMI”), configured for data transmission and interaction with other devices or systems. According to the scheme of this application, the data transmission interface 604 can receive raw retinal images acquired from a fundus acquisition device and transmit raw retinal images or various other types of data or results to device 600.
[0069] The CPU 611, mass storage 612, ROM 613, TPU 614, GPU 615, FPGA 616, MLU 617, and communication interface 618 of the device 600 of this application can be interconnected via bus 619, and can interact with peripheral devices through this bus. In one embodiment, the CPU 611 can control other hardware components in the device 600 and its peripheral devices through bus 619.
[0070] The above combination Figure 6 This paper describes a device that can be used to train a classification model for metabolic diseases, as described in this application. It should be understood that the device structure or architecture described herein is merely exemplary, and the implementation methods and entities described herein are not limited thereto, but can be modified without departing from the spirit of this application.
[0071] Based on the foregoing description in conjunction with the accompanying drawings, those skilled in the art will understand that the embodiments of this application can also be implemented by software programs. Therefore, this application also provides a computer-readable storage medium storing computer-readable instructions for training a classification model for metabolic diseases or for classifying metabolic diseases. When executed by one or more processors, these computer-readable instructions implement the embodiments of this application in conjunction with the accompanying drawings. Figure 2 The described method for training or combining classification models for metabolic diseases. Figure 5 The operation performed by the device described for classifying metabolic diseases.
[0072] It should be noted that although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the steps depicted in the flowchart can be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.
[0073] It should be understood that when the terms "first," "second," "third," and "fourth," etc., are used in the claims, specification, and drawings of this application, they are used only to distinguish different objects and not to describe a specific order. The terms "comprising" and "including" as used in the specification and claims of this application indicate the presence of the described features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof.
[0074] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this specification and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this specification and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.
[0075] Although the embodiments of this application are described above, the content is merely an example adopted for the purpose of facilitating understanding of this application and is not intended to limit the scope and application scenarios of this application. Any person skilled in the art described in this application may make any modifications and changes in the form and details of the implementation without departing from the spirit and scope disclosed in this application, but the scope of patent protection of this application shall still be determined by the scope defined in the appended claims.
Claims
1. A method for training a classification model for metabolic diseases, wherein the classification model includes a serpentine convolution module, a fusion module, and a classifier module, and the method includes: The original retinal image and the grayscale retinal image are respectively input into the classification model. The snake convolution module, the fusion module and the classifier module sequentially perform snake convolution operation, multi-view fusion operation and classification operation to obtain the corresponding first classification result and second classification result. The final classification result of metabolic diseases is determined based on the first classification result and the second classification result, and the initial loss function of the classification model is calculated. A higher-order continuous constraint term from the serpentine convolution operation is introduced into the initial loss function to form the first loss function; The morphological loss function from the serpentine convolution operation is added to the first loss function to form a second loss function; and The classification model for metabolic diseases is trained based on the second loss function. The serpentine convolution module is followed by a morphology generator, and the morphological loss function is constructed through the following operations: Based on the features extracted by the snake-shaped convolution module, the shape generator is used to predict the shape of the snake-shaped convolution kernel to obtain the shape prediction result; and The morphological loss function is constructed based on the difference between the morphological prediction result and the target morphology; The first loss function is expressed as L_task=L_origin+λ_x ∑|d²offset_x / dx²|+λ_y ∑|d²offset_y / dy²|, where L_origin represents the initial loss function, λ_x and λ_y represent hyperparameters used to balance the importance of second-order continuity in the x and y directions, respectively, and d 2 offset_x / dx 2 and d 2 offset_y / dy 2 These represent the second derivatives of the contour points of the serpentine convolution kernel in the horizontal and vertical directions, respectively.
2. The method according to claim 1, wherein the serpentine convolution operation includes horizontal serpentine convolution and vertical serpentine convolution.
3. The method according to claim 1, further comprising: A weighting layer is set before the multi-view fusion operation to adjust the weights of each view feature.
4. The method according to claim 1, wherein determining the final classification result of the metabolic disease based on the first classification result and the second classification result includes: Set corresponding confidence levels for the first classification result and the second classification result respectively; as well as The final classification result of metabolic diseases is determined based on the first and second classification results after corresponding confidence levels.
5. An apparatus for training a classification model for metabolic diseases, comprising: processor; as well as A memory storing program instructions for training a classification model for metabolic diseases, which, when executed by the processor, cause the device to implement the method according to any one of claims 1-4.
6. A device for classifying metabolic diseases, comprising: processor; as well as A memory storing program instructions for classifying metabolic diseases, which, when executed by the processor, cause the device to perform the following operations: Acquire the raw retinal image and the grayscale retinal image; and The original retinal image and the grayscale retinal image are input into a classification model trained according to any one of the methods described in requirements 1-4 to classify metabolic diseases, so as to obtain the final classification result of metabolic diseases.
7. A computer-readable storage medium storing computer-readable instructions for training a classification model for metabolic diseases, wherein the computer-readable instructions, when executed by one or more processors, implement the method as described in any one of claims 1-4; or storing computer-readable instructions for classifying metabolic diseases, wherein the computer-readable instructions, when executed by one or more processors, implement the operation performed by the apparatus as described in claim 6.
