Focus identification system and method for diabetic retinopathy image

Abstract

The invention discloses a focus recognition system and method for a diabetic retinopathy image, and relates to the technical field of medical image processing, and the system comprises a fundus image collection module, an image feature processing module, a focus feature extraction module, a network model training module and a focus model recognition module. By integrating the fundus image collection module, the image feature processing module, the lesion feature extraction module, the network model training module and the lesion model recognition module, full-process automation from original fundus image acquisition to lesion intelligent recognition is realized, and the efficiency of diabetic retinopathy lesion recognition is remarkably improved; in addition, errors possibly caused by manual intervention are reduced, the accuracy and consistency of recognition results are ensured, specifically, the fundus image collection module can automatically obtain images from ophthalmology equipment or a hospital image storage system, format unification and index construction are carried out, and great convenience is provided for follow-up processing.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a lesion recognition system and method for diabetic retinopathy images. Background Art

[0002] Diabetic retinopathy (DR) is one of the most common microvascular complications of diabetes and the leading cause of blindness in working-age people worldwide. With the increasing incidence of diabetes, early screening and diagnosis of DR are particularly important. Fundus images, as an important basis for diagnosing DR, are crucial for timely detection of the disease and formulation of treatment plans. However, manual interpretation of fundus images is not only time-consuming and labor-intensive, but also easily affected by the doctor's experience and fatigue, resulting in subjective and inconsistent diagnostic results. Therefore, developing a system that can automatically and accurately identify diabetic retinopathy lesions has become a research hotspot in the current field of medical image processing.

[0003] Traditional methods for identifying diabetic retinopathy lesions mainly rely on doctors' naked eye observation and experience. This method has many limitations. First, manual interpretation of fundus images is inefficient and cannot meet the needs of large-scale screening. Second, diagnostic standards may differ between different doctors, resulting in poor consistency in diagnostic results. In addition, manual interpretation is easily affected by doctors' subjective factors, such as fatigue and emotions, which further reduces the accuracy of diagnosis. With the rapid development of computer technology and artificial intelligence, although some image processing-based DR automatic recognition systems have been proposed, these systems often have problems with incomplete feature extraction and weak model generalization capabilities, making them difficult to be widely used in actual clinical practice.

[0004] Therefore, the development of a lesion recognition system and method for diabetic retinopathy images not only solves the problems of low efficiency and high subjectivity of traditional manual interpretation of fundus images, but also provides an efficient and reliable automated solution for the early screening and diagnosis of diabetic retinopathy. Summary of the Invention

[0005] The purpose of the present invention is to make up for the shortcomings of the existing technology and provide a lesion recognition system and method for diabetic retinopathy images. By integrating five major modules: fundus image collection, feature processing, lesion feature extraction, network model training and lesion recognition, the whole process from image acquisition to lesion recognition is automated. The use of multi-scale feature fusion and hierarchical association network model significantly improves the accuracy and efficiency of lesion recognition, effectively solves the problems of subjectivity and inconsistency in traditional manual interpretation, and provides strong technical support for the early screening and diagnosis of diabetic retinopathy.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: On the one hand, a lesion recognition system for diabetic retinopathy images includes: a fundus image collection module, an image feature processing module, a lesion feature extraction module, a network model training module and a lesion model recognition module;

[0007] The fundus image collection module is configured to obtain original fundus image data of diabetic retinopathy, convert the images into grayscale image format, store them in an established database, and create an index for them;

[0008] The image feature processing module extracts image data from a database, performs feature preprocessing and enhancement on the image data, determines the positions of the fovea and optic disc using a particle swarm optimization algorithm, and transforms the image into a standard coordinate system based on the positions of the fovea and optic disc using a transformation matrix;

[0009] The lesion feature extraction module: for the image that has been converted to the standard coordinate system, a multi-scale feature extraction neural network is constructed, and convolution operations are performed by setting convolution kernels with different scale parameters to obtain the detailed texture, edge features and global features of the lesion, and the feature maps of different scales are fused into a complete feature vector;

[0010] The network model training module: constructs a hierarchical association network model, uses a data set to calculate the correlation between feature dimensions, combines a weighted loss function based on the clinical importance of the lesion, uses a weight update method guided by information entropy, and iteratively adjusts the network weights for training, so that the network has the ability to identify diabetic retinopathy lesions;

[0011] The lesion model recognition module inputs the feature vector generated by the lesion feature extraction module into the optimized hierarchical association network model, and the model outputs the recognition result of the diabetic retinopathy lesion after calculation processing.

[0012] Furthermore, the original fundus image in the fundus image collection module is acquired by establishing a data interface with an ophthalmic optical coherence tomography scanner or a fundus camera, or is read from a hospital image storage and transmission system, and the image format is DICOM or JPEG-XR.

[0013] Furthermore, the image data features are preprocessed and enhanced in the image feature processing module. Assume that the image pixel matrix is ​​I(x, y), where x and y are pixel coordinates. A feature mapping function Φ(I) is constructed to preprocess the image features and map the image I to the feature space. The formula is: where w ijIs the weight coefficient, which is used to adjust the contribution of different neighborhood pixels to the current pixel feature. a, b are the local neighborhood sizes, which determine the neighborhood range involved in calculating the current pixel feature. The key structural features are highlighted through the structure-sensitive function Ψ(Φ(I)): Among them, α, β are adjustment parameters, Φ(I) is the image feature after being processed by the feature mapping function, is the second-order derivative of Φ(I) in the x-direction, reflecting the rate of change of image features in the x-direction. It is the second-order derivative of Φ(I) in the y direction, reflecting the change rate of image features in the y direction.

[0014] Furthermore, the specific steps of determining the positions of the fovea and optic disc using the particle swarm optimization algorithm in the image feature processing module are as follows:

[0015] (1) According to the complexity of the image and the computing resources, the number of particles is set. Each particle represents the possible location of the fovea or optic disc in the image. The position and velocity of each particle are randomly determined in the two-dimensional coordinate space of the image.

[0016] (2) Combining the image characteristics after feature preprocessing and enhancement, features related to the fovea and optic disc are selected to construct a fitness function to evaluate the quality of particle positions;

[0017] (3) For each particle, calculate its fitness value according to its current position based on the fitness function, and record the individual optimal position and the global optimal position and the corresponding fitness value;

[0018] (4) Update the speed and position of each particle according to the update formula of the particle swarm optimization algorithm;

[0019] (5) For each particle after the position is updated, recalculate the particle fitness, update the individual and global optimal positions and corresponding fitness values;

[0020] (6) Check whether the maximum number of iterations is met. If not, continue the iteration. If the maximum number of iterations is met, terminate the iteration. The global optimal position is the position of the identified retinal fovea or optic disc.

[0021] Furthermore, the image feature processing module transforms the image into a standard coordinate system through a transformation matrix. Let the transformation matrix be T. The transformation from the current image coordinate system to the standard coordinate system includes rotation R(θ), scaling S(s x ,s y ) and translation T(t x ,t y ), construct the comprehensive transformation matrix, the formula is: T = S (s x ,s y )·R(θ)·T(tx ,t y ), where the rotation matrix R(θ) is: Scaling matrix S(s x ,s y )for: Translation matrix T(t x ,t y )for: Where S(s x ,s y ) is the scaling matrix, s x 、s y are the scaling factors in the x and y directions, respectively, used to adjust the size of the image in two directions, R(θ) is the rotation matrix, θ is the rotation angle, which determines the degree of image rotation T(t x ,t y ) is the translation matrix, t x , t y The x and y directions are the translation amounts respectively, which control the translation position of the image in the coordinate system. The parameters θ and s are determined by calculating the difference between the ideal position and the current position of the fovea or optic disc in the standard coordinate system. x ,s y ,t x ,t y , transform the image.

[0022] Furthermore, the lesion feature extraction module extracts image details and global features by defining convolution kernels at different scales, adopts a scale fusion convolution structure, and defines a scale transformation convolution kernel function K σ (x, y), σ is the scale parameter. Different σ values ​​correspond to convolution kernels of different scales. x and y are the coordinates of the convolution kernel. The formula is: At a small scale of σ1, the texture and edge detail features of the lesion are extracted through convolution operation. The formula is: At a large scale σ2, the global characteristics of the spatial relationship between the lesion and the surrounding tissue are obtained through convolution operation, and the formula is: Where σ1 is a small-scale parameter, which is used to extract the convolution kernel scale of the fine texture and edge detail features of the lesion; σ2 is a large-scale parameter, which is used to obtain the convolution kernel scale of the global features of the spatial relationship between the lesion and the surrounding tissue; C1(I) is the result of the convolution operation at a small scale, which represents the fine texture and edge detail feature map of the lesion; C2(I) is the result of the convolution operation at a large scale, which represents the global feature map of the spatial relationship between the lesion and the surrounding tissue; * is the convolution operation symbol.

[0023] Furthermore, in the lesion feature extraction module, feature fusion function is used to fuse feature maps of different scales to generate a complete feature vector. The formula is: F(C1(I), C2(I)) = γ·C1(I) + (1-γ)·C2(I), where F is the feature fusion function, which realizes the fusion of feature maps of different scales into a complete feature vector, C1(I) is the lesion feature map extracted at a small scale, C2(I) is the lesion feature map extracted at a large scale, and γ is the fusion weight dynamically adjusted according to training, which determines the relative importance of the small-scale feature map C1(I) and the large-scale feature map C2(I) in the fused feature vector.

[0024] Furthermore, the construction of the hierarchical association network in the network model training module adopts a multi-level structure. The input layer receives the image feature vector output by the lesion feature extraction module, and transmits it to the middle layer after data standardization. The middle layer is composed of multiple association layers stacked together. The association layer consists of a feature association unit and a weight adjustment unit. The feature association unit of each association layer uses the formula to calculate the correlation between feature dimensions. The formula is: Where A(X i ,X j ) is the correlation between feature dimensions i and j, X i,k is the kth element of feature dimension i, X j,k is the kth element of feature dimension j, is the mean of feature dimension i, is the mean of feature dimension j, n is the number of samples, the feature correlation matrix and the weighted cross entropy loss function, and the weight adjustment unit is based on the feature correlation matrix and the weighted cross entropy loss function. The formula is: Where L is the weighted cross entropy loss value, which is used to measure the difference between the model prediction results and the actual results, and w m is the weight of sample m, reflecting the importance of different samples in training, combined with the clinical importance of the lesion, y m is the true label value of sample m, is the predicted probability value of sample m, M is the total number of samples, and the network weights are updated using a method guided by information entropy. Nonlinear transformations are performed between layers through the ReLU function, and residual connections are added to avoid gradient disappearance. The output layer uses a fully connected layer combined with the Softmax function to convert the processed feature vector into a probability prediction result for lesion recognition.

[0025] On the other hand, a lesion recognition method for diabetic retinopathy images includes the following steps:

[0026] S100, image acquisition and index construction: Obtain original diabetic retinopathy fundus images from an ophthalmic optical coherence tomography scanner, fundus camera, or hospital image storage and transmission system, convert them into grayscale images, store them in a database, and create an index for each image;

[0027] S200, Feature Enhancement and Coordinate Standardization: Extract image data from the database and perform feature preprocessing and enhancement on the image by constructing feature mapping functions and structure-sensitive functions. Use a particle swarm optimization algorithm to determine the location of the fovea and optic disc. Based on these locations, a transformation matrix including rotation, scaling, and translation is constructed to transform the image into a standard coordinate system.

[0028] S300, Multi-scale Feature Fusion Extraction: A multi-scale feature extraction neural network is constructed for images transformed to a standard coordinate system. Convolution kernels of different scales are defined. Small-scale kernels are used to extract fine texture and edge details of the lesion, while large-scale kernels capture global features of the spatial relationship between the lesion and surrounding tissue. Feature fusion functions are used to fuse feature maps of different scales into a complete feature vector.

[0029] S400, Hierarchical Association Network Training: A hierarchical association network model is constructed. Its input layer receives the generated feature vectors and performs normalization processing. The middle layer is composed of multiple stacked association layers. The association layer calculates the correlation between feature dimensions and the weighted cross-entropy loss function, and updates the network weights using a method guided by information entropy. The ReLU function is used for nonlinear transformation between layers and residual connections are added. The output layer uses a fully connected layer combined with a Softmax function. The model is iteratively trained using the dataset to enable it to identify lesions.

[0030] S500, intelligent lesion identification output: the feature vector generated by the lesion feature extraction module is input into the trained and optimized hierarchical association network model, and the model outputs the identification result of the diabetic retinopathy lesion after calculation and processing.

[0031] Compared with the existing technology, the lesion recognition system and method for diabetic retinopathy images have the following beneficial effects:

[0032] 1. The present invention realizes the automation of the entire process from original fundus image acquisition to intelligent lesion identification by integrating a fundus image collection module, an image feature processing module, a lesion feature extraction module, a network model training module and a lesion model identification module. It not only significantly improves the efficiency of diabetic retinopathy lesion identification, but also reduces the errors that may be caused by manual intervention, ensuring the accuracy and consistency of the identification results. Specifically, the fundus image collection module can automatically acquire images from ophthalmic equipment or hospital image storage systems, and perform format unification and index construction, which greatly facilitates subsequent processing. The image feature processing module accurately locates the retinal fovea and optic disc through the particle swarm optimization algorithm, and uses the transformation matrix to convert the image to the standard coordinate system, laying the foundation for lesion feature extraction.

[0033] 2. The present invention constructs a multi-scale feature extraction neural network, combines convolution kernels of different scales, extracts the detailed texture, edge features and global features of the lesions respectively, and fuses these feature maps into a complete feature vector through a feature fusion function, which effectively solves the problem of feature loss or redundancy that may be caused by single-scale feature extraction, and improves the comprehensiveness and accuracy of feature extraction. In the network model training module, the present invention adopts a hierarchical association network model, calculates the correlation between feature dimensions and combines the weighted loss function of the clinical importance of the lesion, and uses an information entropy-guided weight update method to iteratively train the network. This training method not only enhances the network's ability to recognize lesion features, but also enables the model to better adapt to the complexity and diversity of different cases. Finally, the lesion model recognition module can output high-precision lesion recognition results based on the trained and optimized model, providing doctors with a reliable diagnostic basis, which is conducive to early detection and treatment of diabetic retinopathy and protection of patients' vision health.

[0034] Other advantages, objects and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be learned from the practice of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort.

[0036] Figure 1 This is a framework diagram of a lesion recognition system for diabetic retinopathy images;

[0037] Figure 2 Flowchart of a lesion recognition method for diabetic retinopathy images. DETAILED DESCRIPTION

[0038] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0039] Example 1:

[0040] Hospital ophthalmology clinical diagnosis auxiliary scenario.

[0041] Image acquisition and index construction: In the ophthalmology department of a well-equipped tertiary hospital, doctors perform routine fundus examinations on diabetic patients. Through the data interface established with the fundus camera, they obtain original fundus images of diabetic retinopathy. These images are in DICOM format and are transmitted in real time to the hospital's picture archiving and communication system (PACS). Figure 1 As shown in the figure, the lesion recognition system for diabetic retinopathy images reads the image data from the PACS, converts the images into grayscale image format according to the system settings, and stores them in a database specially established for the system. A unique index is created for each image based on the image acquisition time and patient ID information to facilitate subsequent fast and accurate retrieval, as shown in the figure. Figure 2 shown.

[0042] Feature enhancement and coordinate standardization: When a doctor needs to perform a detailed analysis of a specific patient's fundus image, the system extracts the corresponding image data from the database and first preprocesses and enhances the image data features. Let the image pixel matrix be I(x, y), and construct a feature mapping function Φ(I) to preprocess the image features. The formula is: where w ij is a weight coefficient used to adjust the contribution of different neighborhood pixels to the current pixel feature. a and b are the local neighborhood sizes, which determine the neighborhood range involved in calculating the current pixel feature. Afterwards, the key structural features are highlighted through the structure-sensitive function Ψ(Φ(I)). The formula is: Among them, α and β are adjustment parameters. is the second-order derivative of Φ(I) in the x-direction, reflecting the rate of change of image features in the x-direction. It is the second-order derivative of Φ(I) in the y direction, reflecting the change rate of image features in the y direction.

[0043] The particle swarm optimization algorithm is used to determine the position of the fovea and optic disc. The specific steps are as follows: according to the complexity of the image and the current computing resources of the system, an appropriate number of particles are set, each particle represents a possible position of the fovea or optic disc in the image, the initial position and velocity of each particle are randomly determined in the two-dimensional coordinate space of the image, and the features related to the fovea and optic disc are selected based on the characteristics of the image after feature preprocessing and enhancement. A fitness function is constructed to evaluate the quality of the particle position. For each particle, its fitness value is calculated according to the fitness function based on its current position, and the individual optimal position and the global optimal position and the corresponding fitness value are recorded. The velocity and position of each particle are updated according to the update formula of the particle swarm optimization algorithm. For each particle after the updated position, the particle fitness is recalculated, and the individual and global optimal positions and the corresponding fitness values ​​are updated. The above process is repeated continuously to check whether the maximum number of iterations is met. If not, the iteration is continued. When the maximum number of iterations is met, the iteration is terminated. At this time, the global optimal position is the identified position of the fovea or optic disc.

[0044] According to the position of the fovea and optic disc, the image is transformed into the standard coordinate system through the transformation matrix. Let the transformation matrix be T. The transformation from the current image coordinate system to the standard coordinate system includes rotation R(θ), scaling S(s x ,s y ) and translation T(t x ,t y ), the formula for constructing the comprehensive transformation matrix is: T=S(s x ,s y )·R(θ)·T(t x ,t y ), where the rotation matrix R(θ) is: Scaling matrix S(s x ,s y )for: Translation matrix TT(t x ,t y )for: Determine the parameter θ,s by calculating the difference between the ideal position and the current position of the retinal fovea or optic disc in the standard coordinate system x ,s y ,t x ,t y , transform the image.

[0045] Multi-scale feature fusion extraction: For images that have been converted to the standard coordinate system, a multi-scale feature extraction neural network is constructed and the scale transformation convolution kernel function is defined. σ is the scale parameter. Different σ values ​​correspond to convolution kernels of different scales. x and y are the coordinates of the convolution kernel. The formula is: At a small scale of σ1, the texture and edge detail features of the lesion are extracted through convolution operation. The formula is: At a large scale σ2, the global characteristics of the spatial relationship between the lesion and the surrounding tissue are obtained through convolution operation, and the formula is: Where σ1 is a small-scale parameter, which is used to extract the convolution kernel scale of the fine texture and edge detail features of the lesion; σ2 is a large-scale parameter, which is used to obtain the convolution kernel scale of the global features of the spatial relationship between the lesion and the surrounding tissue; C1(I) is the result of the convolution operation at a small scale, which represents the fine texture and edge detail feature map of the lesion; C2(I) is the result of the convolution operation at a large scale, which represents the global feature map of the spatial relationship between the lesion and the surrounding tissue; * is the convolution operation symbol.

[0046] Hierarchical association network training: A hierarchical association network model is constructed. The input layer receives the image feature vector output by the lesion feature extraction module and passes it to the middle layer after data normalization. The middle layer is composed of multiple association layers stacked together. The association layer consists of feature association units and weight adjustment units. The feature association unit of each association layer uses the formula to calculate the correlation between feature dimensions. The formula is: Where A(X i ,X j ) is the correlation between feature dimensions i and j, X i ,k is the kth element of feature dimension i, X j ,k is the kth element of feature dimension j, is the mean of feature dimension i, is the mean of feature dimension j, n is the number of samples, and the weight adjustment unit is based on the feature correlation matrix and the weighted cross entropy loss function. The formula is: Where L is the weighted cross entropy loss value, which is used to measure the difference between the model prediction results and the actual results, and w m is the weight of sample m, set in combination with the clinical importance of the lesion, y m is the true label value of sample m, is the predicted probability value of sample m, M is the total number of samples, and the network weights are updated using a method guided by information entropy. Nonlinear transformations are performed between layers through the ReLU function, and residual connections are added to avoid gradient vanishing. The output layer uses a fully connected layer combined with a Softmax function to convert the processed feature vector into a probability prediction result for lesion identification. The model is iteratively trained using a large number of diabetic retinopathy image datasets accumulated in the hospital, so that the network has the ability to accurately identify diabetic retinopathy lesions.

[0047] Intelligent lesion identification output: The doctor inputs the feature vector generated by the lesion feature extraction module into the trained and optimized hierarchical association network model. After calculation and processing, the model outputs the identification results of diabetic retinopathy lesions, such as clearly indicating whether there are different types of lesions such as microaneurysms, bleeding points, and exudates in the image, as well as the specific location and approximate range of the lesions in the fundus image. The doctor combines these identification results with the patient's medical history, blood sugar control status and other clinical information to more accurately judge the severity of the patient's diabetic retinopathy and formulate a more scientific and reasonable treatment plan.

[0048] In summary, in the clinical diagnosis assistance scenario of hospital ophthalmology, the present invention constructs a complete diabetic retinopathy image lesion recognition system, starting from image acquisition, unifying the format and establishing indexes for efficient management. Through feature enhancement and coordinate standardization, the key features of the image are highlighted and the image coordinates are standardized. Multi-scale feature fusion extraction comprehensively obtains lesion information. Hierarchical association network training enables the model to accurately identify lesions, and finally realizes intelligent lesion recognition output to assist doctors in diagnosis. This not only improves diagnostic efficiency, but also helps doctors formulate treatment plans based on more accurate results, providing strong support for the clinical diagnosis and treatment of patients with diabetic retinopathy.

[0049] Example 2:

[0050] Telemedicine health management scenario.

[0051] Image acquisition and index construction: In primary medical clinics in remote areas, doctors use portable fundus cameras to perform fundus examinations on local diabetic patients. The original fundus images captured are transmitted to the telemedicine service platform via the network in JPEG-XR format. The lesion recognition system for diabetic retinopathy images on the telemedicine platform obtains these images, converts them into grayscale image format, and stores them in the platform's dedicated database. At the same time, the system creates an index for each image. The index information covers the patient's basic identity information, examination time, and location, facilitating subsequent classification management and quick search of images.

[0052] Feature enhancement and coordinate standardization: After receiving a request for fundus image analysis from a primary care physician, the telemedicine platform's experts extract the corresponding image data from the database. The system first preprocesses and enhances the image data features according to a pre-defined algorithm. The feature mapping function Φ(I) is constructed using the formula to preprocess the image features. The formula is: Then the key structural features are highlighted through the structure sensitivity function, the formula is:

[0053] The particle swarm optimization algorithm is used to determine the position of the fovea and optic disc: according to the complexity of the image and the current computing resource status of the remote server, a certain number of particles are set, each particle represents a possible position of the fovea or optic disc in the image, and each particle is randomly assigned an initial position and velocity in the two-dimensional coordinate space of the image. Combined with the characteristics of the preprocessed and enhanced image, features related to the fovea and optic disc are selected to construct a fitness function. For each particle, the fitness value is calculated based on its current position using the fitness function, and the individual optimal position, global optimal position and their corresponding fitness values ​​are recorded. The particle velocity and position are updated according to the update formula of the particle swarm optimization algorithm. After the position is updated, the particle fitness is recalculated, and the individual and global optimal positions and corresponding fitness values ​​are updated again. This process is repeated until the maximum number of iterations is met. At this time, the global optimal position is the identified position of the fovea or optic disc.

[0054] According to the determined positions of the fovea and optic disc, a transformation matrix is ​​constructed to transform the image into a standard coordinate system. Let the transformation matrix be T, and its construction formula is: T = S (s x ,s y )·R(θ)·T(t x ,t y ), where the rotation matrix R(θ) is: Scaling matrix S(s x ,s y )for: Translation matrix T(t x ,t y )for: Determine the parameter θ,s by calculating the difference between the ideal position and the current position of the retinal fovea or optic disc in the standard coordinate system x ,s y ,t x ,t y , completing the transformation of the image.

[0055] Multi-scale feature fusion extraction: For images converted to the standard coordinate system, a multi-scale feature extraction neural network is constructed and the scale transformation convolution kernel function is defined. The formula is: Different scales of convolution kernels are obtained by using different σ values. At a small scale of σ1, the texture and edge detail features of the lesion are extracted through convolution operation. The formula is: At a large scale σ2, the global characteristics of the spatial relationship between the lesion and the surrounding tissue are obtained through convolution operation, and the formula is: Finally, the feature maps of different scales are fused through the feature fusion function to generate a complete feature vector. The formula is: F(C1(I),C2(I))=γ·C1(I)+(1-γ)·C2(I).

[0056] Hierarchical association network training: A hierarchical association network model is constructed. The input layer receives the image feature vector output by the lesion feature extraction module and performs normalization processing, and then passes it to the middle layer. The middle layer is stacked by multiple association layers. The association layer contains feature association units and weight adjustment units. The feature association unit uses the formula to calculate the correlation between feature dimensions. The formula is: The weight adjustment unit is based on the feature correlation matrix and the weighted cross entropy loss function, the formula is: The network weights are updated using a method guided by information entropy. The ReLU function is used for nonlinear transformation between layers, and residual connections are added to avoid gradient vanishing. The output layer uses a fully connected layer combined with a Softmax function to convert the processed feature vector into a probabilistic prediction result for lesion identification. The model is iteratively trained using a large amount of diabetic retinopathy image data accumulated by the telemedicine platform to improve the model's ability to recognize diabetic retinopathy lesions.

[0057] Intelligent lesion identification output: Experts on the telemedicine platform input the feature vectors generated by the lesion feature extraction module into the trained and optimized hierarchical association network model. After computational processing, the model outputs the identification results of diabetic retinopathy lesions, such as whether there are lesions, the type of lesions, and the approximate severity. Based on these identification results, combined with the patient's basic information and the preliminary diagnosis provided by the primary care physician, the experts provide the primary care physician with detailed diagnostic recommendations and treatment guidance. Based on the experts' advice, the primary care physician develops personalized health management plans for the patients, including dietary adjustments, blood sugar monitoring frequency adjustments, and referral recommendations when necessary, to achieve effective management of diabetic patients' eye health under the telemedicine model.

[0058] To sum up, the present invention plays an important role in the scenario of remote medical health management. It uses the network to realize the collection and transmission of fundus images of patients in remote areas. After system processing, it uses the particle swarm optimization algorithm to determine the key positions and convert the image coordinate system. Multi-scale feature fusion and hierarchical association network training enable the model to effectively identify lesions. Experts give diagnostic suggestions based on the identification results, and grassroots doctors formulate health management plans based on this. This invention breaks the geographical restrictions and enables diabetic patients in remote areas to obtain professional diagnosis and health management guidance, thereby improving the quality and effectiveness of telemedicine services.

[0059] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as above in terms of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can, without departing from the scope of the technical solution of the present invention, make some changes or modifications to equivalent embodiments using the technical contents disclosed above. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A lesion recognition system for diabetic retinopathy images, characterized in that: The system includes: fundus image collection module, image feature processing module, lesion feature extraction module, network model training module and lesion model recognition module; The fundus image collection module is configured to obtain original fundus image data of diabetic retinopathy, convert the images into grayscale image format, store them in an established database, and create an index for them; The image feature processing module extracts image data from a database, performs feature preprocessing and enhancement on the image data, determines the positions of the fovea and optic disc using a particle swarm optimization algorithm, and transforms the image into a standard coordinate system based on the positions of the fovea and optic disc using a transformation matrix; The lesion feature extraction module: for the image that has been converted to the standard coordinate system, a multi-scale feature extraction neural network is constructed, and convolution operations are performed by setting convolution kernels with different scale parameters to obtain the detailed texture, edge features and global features of the lesion, and the feature maps of different scales are fused into a complete feature vector; The network model training module: constructs a hierarchical association network model, uses a data set to calculate the correlation between feature dimensions, combines a weighted loss function based on the clinical importance of the lesion, uses a weight update method guided by information entropy, and iteratively adjusts the network weights for training, so that the network has the ability to identify diabetic retinopathy lesions; The lesion model recognition module inputs the feature vector generated by the lesion feature extraction module into the optimized hierarchical association network model, and the model outputs the recognition result of the diabetic retinopathy lesion after calculation processing.

2. The lesion recognition system for diabetic retinopathy images according to claim 1, characterized in that: The original fundus image in the fundus image collection module is acquired by establishing a data interface with an ophthalmic optical coherence tomography scanner or a fundus camera, or is read from a hospital image storage and transmission system. The image format is DICOM or JPEG-XR.

3. The lesion recognition system for diabetic retinopathy images according to claim 1, characterized in that: The image data feature preprocessing and enhancement in the image feature processing module assumes that the image pixel matrix is ​​I(x, y), where x and y are pixel coordinates, and constructs a feature mapping function Φ(I) to preprocess the image features and map the image I to the feature space. The formula is: where w ij Is the weight coefficient, which is used to adjust the contribution of different neighborhood pixels to the current pixel feature. a, b are the local neighborhood sizes, which determine the neighborhood range involved in calculating the current pixel feature. The key structural features are highlighted through the structure-sensitive function Ψ(Φ(I)): Among them, α, β are adjustment parameters, Φ(I) is the image feature after being processed by the feature mapping function, is the second-order derivative of Φ(I) in the x-direction, reflecting the rate of change of image features in the x-direction. It is the second-order derivative of Φ(I) in the y direction, reflecting the change rate of image features in the y direction.

4. The lesion recognition system for diabetic retinopathy images according to claim 1, characterized in that: The specific steps of using the particle swarm optimization algorithm to determine the positions of the fovea and optic disc in the image feature processing module are as follows: (1) According to the complexity of the image and the computing resources, the number of particles is set. Each particle represents the possible location of the fovea or optic disc in the image. The position and velocity of each particle are randomly determined in the two-dimensional coordinate space of the image. (2) Combining the image characteristics after feature preprocessing and enhancement, features related to the fovea and optic disc are selected to construct a fitness function to evaluate the quality of particle positions; (3) For each particle, calculate its fitness value according to its current position based on the fitness function, and record the individual optimal position and the global optimal position and the corresponding fitness value; (4) Update the speed and position of each particle according to the update formula of the particle swarm optimization algorithm; (5) For each particle after the position is updated, recalculate the particle fitness, update the individual and global optimal positions and corresponding fitness values; (6) Check whether the maximum number of iterations is met. If not, continue the iteration. If the maximum number of iterations is met, terminate the iteration. The global optimal position is the position of the identified retinal fovea or optic disc.

5. The lesion recognition system for diabetic retinopathy images according to claim 1, characterized in that: The image feature processing module transforms the image to the standard coordinate system through the transformation matrix. Let the transformation matrix be T. The transformation from the current image coordinate system to the standard coordinate system includes rotation R(θ), scaling S(s x ,s y ) and translation T(t x ,t y ), construct the comprehensive transformation matrix, the formula is: T = S (s x ,s y )·R(θ)·T(t x ,t y ), where the rotation matrix R(θ) is: Scaling matrix S(s x ,s y )for: Translation matrix T(t x ,t y )for: Where S(s x ,s y ) is the scaling matrix, s x 、s y are the scaling factors in the x and y directions, respectively, used to adjust the size of the image in two directions, R(θ) is the rotation matrix, θ is the rotation angle, which determines the degree of image rotation T(t x ,t y ) is the translation matrix, t x , t y The x and y directions are the translation amounts respectively, which control the translation position of the image in the coordinate system. The parameters θ and s are determined by calculating the difference between the ideal position and the current position of the fovea or optic disc in the standard coordinate system. x ,s y ,t x ,t y , transform the image.

6. The lesion recognition system for diabetic retinopathy images according to claim 1, characterized in that: The lesion feature extraction module extracts image details and global features by defining convolution kernels at different scales, adopts a scale fusion convolution structure, and defines a scale transformation convolution kernel function K σ (x, y), σ is the scale parameter. Different σ values ​​correspond to convolution kernels of different scales. x and y are the coordinates of the convolution kernel. The formula is: At a small scale of σ1, the texture and edge detail features of the lesion are extracted through convolution operation. The formula is: At a large scale σ2, the global characteristics of the spatial relationship between the lesion and the surrounding tissue are obtained through convolution operation, and the formula is: Where σ1 is a small-scale parameter, which is used to extract the convolution kernel scale of the fine texture and edge detail features of the lesion; σ2 is a large-scale parameter, which is used to obtain the convolution kernel scale of the global features of the spatial relationship between the lesion and the surrounding tissue; C1(I) is the result of the convolution operation at a small scale, which represents the fine texture and edge detail feature map of the lesion; C2(I) is the result of the convolution operation at a large scale, which represents the global feature map of the spatial relationship between the lesion and the surrounding tissue; * is the convolution operation symbol.

7. The lesion recognition system for diabetic retinopathy images according to claim 1, characterized in that: In the lesion feature extraction module, feature maps of different scales are fused to generate a complete feature vector through a feature fusion function. The formula is: F(C1(I), C2(I)) = γ·C1(I) + (1-γ)·C2(I), where F is the feature fusion function, which realizes the fusion of feature maps of different scales into a complete feature vector, C1(I) is the lesion feature map extracted at a small scale, C2(I) is the lesion feature map extracted at a large scale, and γ is the fusion weight dynamically adjusted according to training, which determines the relative importance of the small-scale feature map C1(I) and the large-scale feature map C2(I) in the fused feature vector.

8. The lesion recognition system for diabetic retinopathy images according to claim 1, characterized in that: The construction of the hierarchical association network in the network model training module adopts a multi-level structure. The input layer receives the image feature vector output by the lesion feature extraction module and transmits it to the middle layer after data standardization. The middle layer is composed of multiple association layers stacked together. The association layer consists of feature association units and weight adjustment units. The feature association unit of each association layer uses the formula to calculate the correlation between feature dimensions. The formula is: Where A(X i ,X j ) is the correlation between feature dimensions i and j, X i,k is the kth element of feature dimension i, X j,k is the kth element of feature dimension j, is the mean of feature dimension i, is the mean of feature dimension j, n is the number of samples, the feature correlation matrix and the weighted cross entropy loss function, and the weight adjustment unit is based on the feature correlation matrix and the weighted cross entropy loss function. The formula is: Where L is the weighted cross entropy loss value, which is used to measure the difference between the model prediction results and the actual results, and w m is the weight of sample m, reflecting the importance of different samples in training, combined with the clinical importance of the lesion, y m is the true label value of sample m, is the predicted probability value of sample m, M is the total number of samples, and the network weights are updated using a method guided by information entropy. Nonlinear transformations are performed between layers through the ReLU function, and residual connections are added to avoid gradient disappearance. The output layer uses a fully connected layer combined with the Softmax function to convert the processed feature vector into a probability prediction result for lesion recognition.

9. A method for identifying lesions in diabetic retinopathy images, the method being applicable to the system for identifying lesions in diabetic retinopathy images according to any one of claims 1 to 8, characterized in that: The specific steps of this identification method are: S100, image acquisition and index construction: Obtain original diabetic retinopathy fundus images from an ophthalmic optical coherence tomography scanner, fundus camera, or hospital image storage and transmission system, convert them into grayscale images, store them in a database, and create an index for each image; S200, Feature Enhancement and Coordinate Standardization: Extract image data from the database and perform feature preprocessing and enhancement on the image by constructing feature mapping functions and structure-sensitive functions. Use a particle swarm optimization algorithm to determine the location of the fovea and optic disc. Based on these locations, a transformation matrix including rotation, scaling, and translation is constructed to transform the image into a standard coordinate system. S300, Multi-scale Feature Fusion Extraction: A multi-scale feature extraction neural network is constructed for images transformed to a standard coordinate system. Convolution kernels of different scales are defined. Small-scale kernels are used to extract fine texture and edge details of the lesion, while large-scale kernels capture global features of the spatial relationship between the lesion and surrounding tissue. Feature fusion functions are used to fuse feature maps of different scales into a complete feature vector. S400, Hierarchical Association Network Training: A hierarchical association network model is constructed. Its input layer receives the generated feature vectors and performs normalization processing. The middle layer is composed of multiple stacked association layers. The association layer calculates the correlation between feature dimensions and the weighted cross-entropy loss function, and updates the network weights using a method guided by information entropy. The ReLU function is used for nonlinear transformation between layers and residual connections are added. The output layer uses a fully connected layer combined with a Softmax function. The model is iteratively trained using the dataset to enable it to identify lesions. S500, intelligent lesion identification output: the feature vector generated by the lesion feature extraction module is input into the trained and optimized hierarchical association network model, and the model outputs the identification result of the diabetic retinopathy lesion after calculation and processing.

Images