Fundus image analysis method, system, terminal and computer readable storage medium

Abstract

The application discloses an eye fundus image analysis method, system, terminal and computer readable storage medium, and the method comprises the steps of acquiring a target eye fundus image and target non-imaging data, inputting the eye fundus image into an encoder to output visual features; taking the visual features as the features of the nodes of a feature hypergraph, constructing a hyperedge group of the feature hypergraph, and obtaining a first correlation matrix of the feature hypergraph; classifying the non-imaging data to obtain a plurality of groups, obtaining hyperedges of an attribute hypergraph, constructing a hyperedge group of the attribute hypergraph according to the plurality of hyperedges, and obtaining a second correlation matrix of the attribute hypergraph; splicing the first correlation matrix and the second correlation matrix to obtain a target correlation matrix of a target hypergraph; inputting the visual features and the target correlation matrix into a weighted hypergraph convolution module to obtain target node features, obtaining the probability of each category according to the target node features, and generating an eye fundus image analysis report according to the probability. The application shows excellent performance in processing eye fundus image analysis tasks.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, system, terminal, and computer-readable storage medium for fundus image analysis. Background Technology

[0002] With the rapid accumulation of multimodal clinical data in the field of ophthalmology, the demand for learning models that integrate imaging data (such as fundus images and optical coherence tomography images) with non-imaging data (such as age and gender) is growing. Developing deep learning models that can fuse imaging and non-imaging data is particularly important. Such models can utilize the increasingly rich multimodal medical data to achieve more comprehensive feature representation and more accurate analysis of fundus images.

[0003] However, current models fuse features extracted from images with non-radiographic data through methods such as concatenation or addition. While these models can automatically extract visual features from fundus images for analysis, they cannot effectively model interactions and relationships between different users, resulting in less than ideal analysis results. Graph convolutional network (GCNN) models offer a new solution to this problem. In these models, individuals are treated as nodes in a graph, and relationships between them are represented by edges. For example, individual fundus images can be used as nodes, and non-radiographic information can be used to construct edges between these nodes. Although GCNNs can establish pairwise relationships between two objects, traditional GCNNs have limitations when handling complex high-order relationships between users.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a fundus image analysis method, system, terminal, and computer-readable storage medium, aiming to solve the problem that existing technologies for analyzing fundus images cannot effectively model the interactions and relationships between different users, and have limitations in handling complex high-order relationships between users, resulting in inaccurate analysis results of fundus images.

[0006] To achieve the above objectives, the present invention provides a fundus image analysis method, which includes the following steps:

[0007] Acquire target fundus images and target non-imaging data of multiple users to be analyzed, input the fundus images into a pre-trained encoder, and output visual features;

[0008] The visual features are used as the features of the nodes of the feature hypergraph. A preset algorithm is used to construct the hyperedge group of the feature hypergraph based on the nodes. The first correlation matrix of the feature hypergraph is obtained based on the hyperedge group of the feature hypergraph.

[0009] The non-image data is classified into multiple groups, and a hyperedge of an attribute hypergraph is obtained according to each group. The hyperedge group of the attribute hypergraph is constructed according to the multiple hyperedges, and the second correlation matrix of the attribute hypergraph is obtained according to the hyperedge group of the attribute hypergraph.

[0010] The first and second correlation matrices are concatenated to obtain the target correlation matrix of the target hypergraph.

[0011] The visual features and the target association matrix are input into a weighted hypergraph convolution module for feature extraction to obtain target node features. The probability of each category is obtained based on the target node features, and a fundus image analysis report is generated based on the probability.

[0012] Optionally, in the fundus image analysis method, the feature hypergraph, the attribute hypergraph, and the target hypergraph all include a node set V and a hyperedge set E;

[0013] For any hyperedge e∈E in the hyperedge set E, there exists a weight w(e)∈W, where W is the hyperedge weight matrix, defined as:

[0014] W = diag(w(e1), w(e2), ..., w(e...) |E| ));

[0015] Where, diag represents a diagonal matrix, and |E| represents the total number of hyperedges;

[0016] The structures of the feature hypergraph, the attribute hypergraph, and the target hypergraph are all represented by an association matrix H of size |V|×|E|, where each element of the association matrix H is defined as:

[0017]

[0018] Where H(v,e)=1 indicates that any node v∈V in the node set V is on the hyperedge e, and H(v,e)=0 indicates that any node v∈V in the node set V is not on the hyperedge e;

[0019] For any node v∈V, the degree of the node is represented as:

[0020] d(v)=∑ e∈E w(e)H(v,e);

[0021] For any hyperedge e∈E, the degree of the hyperedge is represented as:

[0022]

[0023] Optionally, in the fundus image analysis method, the preset algorithm includes a nearest neighbor method and a clustering method;

[0024] The step of constructing the hyperedge group of the feature hypergraph based on the nodes using a preset algorithm specifically includes:

[0025] Calculate the distance between all the nodes in the feature space, and combine the distance between any node v and the set of its neighbor nodes N. k(v) Aggregates to generate multiple first hyperedges e(v):

[0026] e(v) = N k(v) ∪v;

[0027] Where, N k(v) This represents the set of the k nearest neighbor nodes to any given node v;

[0028] The generated multiple first hyperedges are combined together to obtain the hyperedge group E of the feature hypergraph. KNN :

[0029] E KNN ={e(v)|v∈V};

[0030] Obtain a set N of nodes whose distance from any given node v is less than a preset threshold ∈. ∈ (v) connects the arbitrary node v with the set N of nodes. ∈ (v) Aggregate them separately to generate multiple second hyperedges, and combine the generated multiple second hyperedges together to obtain the hyperedge group E of the feature hypergraph. ∈ :

[0031] E ∈ ={N ∈ (v)∪v|v∈V};

[0032] Using a clustering algorithm, the arbitrary node v is grouped according to its characteristics, and a second hyperedge is generated for each group. Nodes in the same group are connected by the second hyperedge.

[0033] Combining all the generated second hyperedges together yields the hyperedge group E of the feature hypergraph. k-means :

[0034] E k-means ={V k |1≤k≤K};

[0035] Where K represents the number of groups, k represents the ordinal number of the group, and V k This represents the superedge containing the k-th node;

[0036] The super-edge group E is respectively KNN The hyper-edge group E ∈ and the super-edge group E k-means Each hyperedge in the array is assigned a weight:

[0037]

[0038] Where u and v represent two nodes, X(u) and X(v) represent the features of u and v respectively, and σ 2 Let represent the variance of the Gaussian kernel, exp() represent the exponential function, and d(X(u),X(v)) represent the distance between the features of nodes u and v.

[0039] Optionally, in the fundus image analysis method, obtaining the first correlation matrix of the feature hypergraph based on the hyperedge group of the feature hypergraph specifically includes:

[0040] According to the hyperedge group E KNN The hyper-edge group E ∈ and the super-edge group E k-means The first correlation matrix H of the feature hypergraph is represented. f :

[0041]

[0042] Optionally, the fundus image analysis method, wherein classifying the non-imaging data to obtain multiple groups, obtaining a hyperedge of an attribute hypergraph based on each group, constructing a hyperedge group of the attribute hypergraph based on multiple hyperedges, and obtaining a second correlation matrix of the attribute hypergraph based on the hyperedge group of the attribute hypergraph, specifically includes:

[0043] The non-image data of multiple users to be analyzed are grouped according to attributes, including discrete attribute data and continuous attribute data;

[0044] Users with the same discrete attribute data are grouped onto the same hyperedge. Attribute intervals are divided using a segmented approach. Users with continuous attribute data falling within the same attribute interval are grouped onto the same hyperedge, thus constructing the hyperedge group E of the attribute hypergraph. attribute :

[0045] E attribute ={N attribute (a)|a∈A};

[0046] Where, N attribute (a) is the set of nodes that have the same attribute a, and A represents the set of all attributes;

[0047] Through the hyperedge group Eattribute Obtain the second association matrix H of the attribute hypergraph attribute and for the hyperedge group E attribute Each hyperedge in the array is assigned a weight:

[0048]

[0049] Where δ(e) represents the degree of the hyperedge e, and ||2| represents the Euclidean distance.

[0050] Optionally, in the fundus image analysis method, the step of concatenating the first correlation matrix and the second correlation matrix to obtain the target correlation matrix of the target hypermap specifically includes:

[0051] The first association matrix of the feature hypergraph and the second association matrix of the attribute hypergraph are merged using a matrix concatenation method to obtain the association matrix H of the initial hypergraph:

[0052] H = H f ||H attribute ;

[0053] As the network depth increases, the correlation matrix of the initial hypergraph is dynamically adjusted using dynamic graph construction techniques to obtain the target correlation matrix of the target hypergraph.

[0054] Optionally, in the aforementioned fundus image analysis method, the step of inputting the visual features and the target correlation matrix into a weighted hypergraph convolution module for feature extraction to obtain target node features, and obtaining the probability of each category based on the target node features, specifically includes:

[0055] The visual features in the l-th layer are X (l) The transpose of the target association matrix is ​​input into a weighted hypergraph convolution module for matrix multiplication. This module aggregates the features of the nodes connected to each hyperedge in the target association matrix onto the hyperedge, resulting in a hyperedge feature matrix H. (l) ;

[0056] Set a matrix that can learn node weights. The matrix Perform a Hadamard product operation with the target correlation matrix, assign weights to each node on the hyperedge, and obtain the target hyperedge feature matrix E. (l) :

[0057]

[0058] Where ⊙ represents the Hadamard product operation, and T represents the matrix transpose;

[0059] Set a matrix that can learn node weights. The matrix A Hadamard product is performed with the target association matrix to assign weights to each hyperedge containing a node, resulting in a weighted node feature matrix. This weighted node feature matrix is ​​then processed by a learnable parameter Θ. (l) A feature space transformation is performed, and a residual operation is introduced to obtain the target node feature X. (l+1) :

[0060]

[0061] Where ρ represents the activation function, and EHGCN() represents the output of the weighted hypergraph convolution module;

[0062] The target node feature X (l+1) The probability of each user to be analyzed in each category is obtained by outputting the softmax layer.

[0063] Furthermore, to achieve the above objectives, the present invention also provides a fundus image analysis system, wherein the fundus image analysis system comprises:

[0064] The user data acquisition module is used to acquire target fundus images and target non-imaging data of multiple users to be analyzed, input the fundus images into a pre-trained encoder, and output visual features.

[0065] The first correlation matrix construction module is used to take the visual features as the features of the nodes of the feature hypergraph, construct the hyperedge group of the feature hypergraph according to the nodes using a preset algorithm, and obtain the first correlation matrix of the feature hypergraph according to the hyperedge group of the feature hypergraph.

[0066] The second correlation matrix construction module is used to classify the non-image data to obtain multiple groups, obtain a hyperedge of the attribute hypergraph according to each group, construct a hyperedge group of the attribute hypergraph according to multiple hyperedges, and obtain the second correlation matrix of the attribute hypergraph according to the hyperedge group of the attribute hypergraph.

[0067] The target association matrix generation module is used to concatenate the first association matrix and the second association matrix to obtain the target association matrix of the target hypergraph.

[0068] The feature extraction module is used to input the visual features and the target association matrix into the weighted hypergraph convolution module for feature extraction, obtain target node features, obtain the probability of each category based on the target node features, and generate a fundus image analysis report based on the probability.

[0069] In addition, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a fundus image analysis program stored in the memory and executable on the processor, wherein when the fundus image analysis program is executed by the processor, it implements the steps of the fundus image analysis method as described above.

[0070] In addition, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a fundus image analysis program, which, when executed by a processor, implements the steps of the fundus image analysis method as described above.

[0071] In this invention, a target fundus image and target non-image data are acquired. The fundus image is input into an encoder to output visual features. The visual features are used as the features of the nodes in a feature hypergraph to construct a hyperedge group of the feature hypergraph, resulting in a first correlation matrix of the feature hypergraph. The non-image data is classified to obtain multiple groups, resulting in hyperedges of an attribute hypergraph. Based on these hyperedges, a hyperedge group of the attribute hypergraph is constructed, resulting in a second correlation matrix of the attribute hypergraph. The first and second correlation matrices are concatenated to obtain the target correlation matrix of the target hypergraph. The nodes of the feature hypergraph and the target correlation matrix are input into a weighted hypergraph convolution module to obtain target node features. The probability of each category is obtained based on the target node features, and a fundus image analysis report is generated based on the probabilities. This invention demonstrates excellent performance in processing fundus image analysis tasks, providing a new perspective and powerful tool for high-order relation modeling. Attached Figure Description

[0072] Figure 1 This is a flowchart of a preferred embodiment of the fundus image analysis method of the present invention;

[0073] Figure 2 This is a diagram of the Multi-Hypergraph learning framework in the fundus image analysis method of this invention;

[0074] Figure 3 This is a comparison diagram of the simple graph and the supergraph in the fundus image analysis method of the present invention;

[0075] Figure 4 This is a schematic diagram of the feature hypermap construction in the fundus image analysis method of the present invention;

[0076] Figure 5 This is a schematic diagram of the attribute hypergraph construction in the fundus image analysis method of the present invention;

[0077] Figure 6 This is a flowchart of the weighted hypergraph convolution calculation in the fundus image analysis method of the present invention;

[0078] Figure 7This is a schematic diagram of a preferred embodiment of the fundus image analysis system of the present invention;

[0079] Figure 8 This is a schematic diagram of the operating environment of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0080] This application provides a method and related equipment for fundus image analysis. To make the purpose, technical solution, and effects of this application clearer and more explicit, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit this application.

[0081] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0082] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0083] The preferred embodiment of the fundus image analysis method of the present invention, such as... Figure 1 and Figure 2 As shown, the fundus image analysis method includes the following steps:

[0084] Step S100: Obtain target fundus images and target non-imaging data of multiple users to be analyzed, input the fundus images into a pre-trained encoder, and output visual features.

[0085] Understandably, firstly, multiple target fundus images (fundus image data) of the users to be analyzed and multiple target non-image data (such as gender, age, eye type, BMI, data source, etc.) of the users to be analyzed are acquired as input to the graph neural network; then, the target fundus images are input into a pre-trained encoder structure to obtain visual features, and then the hypergraph initialization stage is entered.

[0086] Step S200: Use the visual features as the features of the nodes of the feature hypergraph, construct the hyperedge group of the feature hypergraph based on the nodes using a preset algorithm, and obtain the first correlation matrix of the feature hypergraph based on the hyperedge group of the feature hypergraph.

[0087] like Figure 3 As shown in (a), traditional Graph Convolutional Networks (GCNs) are primarily based on simple graph structures and can only handle pairwise relationships between two objects. This limitation makes GCNs ineffective at handling more complex higher-order relationships, which is particularly important in ophthalmology, where complex higher-order relationships often exist between image data and non-image data. For example, as... Figure 3 As shown in (b), fundus images can be connected by sharing similar attributes or feature similarity.

[0088] To overcome the limitations of traditional Graph Convolution Networks (GCNs) in handling complex high-order relationships between individuals, this invention introduces a hypergraph structure to more accurately model these relationships. By introducing the concept of hyperedges, hypergraphs can group multiple nodes with the same relationship onto a single hyperedge, thereby effectively capturing high-order relationships between individuals.

[0089] Specifically, the definition of a hypergraph and related concepts are explained. A hypergraph (i.e., the feature hypergraph and the subsequent attribute hypergraph and target hypergraph) includes a set of nodes V and a set of hyperedges E.

[0090] For any hyperedge e∈E in the hyperedge set E, there exists a weight w(e)∈W, where W is the hyperedge weight matrix, defined as:

[0091] W = diag(w(e1), w(e2), ..., w(e...) |E| ));

[0092] Where diag represents a diagonal matrix, and |E| represents the total number of hyperedges.

[0093] The structures of the feature hypergraph, the attribute hypergraph, and the target hypergraph are all represented by an association matrix H of size |V|×|E|, where each element of the association matrix H is defined as:

[0094]

[0095] Where H(v,e)=1 indicates that any node v∈V in the node set V is on the hyperedge e, and H(v,e)=0 indicates that any node v∈V in the node set V is not on the hyperedge e.

[0096] For any node v∈V, the degree of the node (i.e. how many hyperedges the node lies on) is expressed as:

[0097] d(v)=∑ e∈E w(e)H(v,e);

[0098] For any hyperedge e∈E, the degree of the hyperedge (i.e., how many nodes the hyperedge contains) is expressed as:

[0099]

[0100] Therefore, the degree matrix D of the hypergraph G is... v and the boundary degree matrix D e They are respectively:

[0101] diag(D v )=[d(v1),d(v2),…,d(v |V| ];

[0102] diag(D e )=[δ(e1),δ(e2),…,δ(e |E| )];

[0103] Where |V| and |E| are the number of nodes and superedges contained in the node set V and the superedge set E, respectively.

[0104] Furthermore, considering the simultaneous presence of image data and non-image data, where image data represents the semantic information of nodes and non-image data reflects the structural and topological information of the graph, this invention introduces the concept of hyperedge groups to unify and integrate these two types of information in the hypergraph. These hyperedge groups are used to construct feature hypergraphs reflecting semantic information and attribute hypergraphs reflecting structural and topological information. Hyperedge groups can be constructed based on different types of data to highlight the different characteristics of various types of information.

[0105] This invention uses a distance-based method to construct a feature hypergraph. In this method, the direct relationships between objects are not directly given, but rather constructed using the semantic information or features of all objects, thus forming corresponding hyperedges. In this method, a hyperedge typically consists of a central node and its neighboring nodes. Nearest neighbor methods and clustering-based methods are commonly used to select neighboring nodes. The nearest neighbor method focuses on the node closest to the central node, reflecting the local structure of the graph, while clustering algorithms operate on all nodes in each iteration, reflecting the global structure of the graph.

[0106] Specifically, such as Figure 4 As shown in (a), (b) and (c) in the figure, the construction of the hyperedge group of the feature hypergraph based on the nodes by the preset algorithm, wherein the preset algorithm includes the nearest neighbor method and the clustering method.

[0107] Calculate the distance between all the nodes in the feature space, and combine the distance between any node v and the set of its neighbor nodes N. k(v) Aggregates to generate multiple first hyperedges e(v):

[0108] e(v) = N k(v) ∪v;

[0109] Where, N k(v) This represents the set of the k nearest neighbor nodes to any given node v;

[0110] The generated multiple first hyperedges are combined together to obtain the hyperedge group E of the feature hypergraph. KNN :

[0111] E KNN ={e(v)|v∈V}.

[0112] It is understandable that the hyperedge group E of the generated feature hypergraph is... KNN The process is based on the nearest neighbor method. When generating a feature hypergraph based on the existing nearest neighbor hyperedge generation strategy, the distance between all vertex pairs in the feature space is first calculated. For any node v in the hypergraph, its feature is represented by X(v). Then, the distance function dis(v1,v2) is used to measure the similarity between nodes v1 and v2. Usually, Euclidean distance or cosine similarity can be chosen as the distance function. However, the hypergraph constructed by measuring the similarity between two points using this distance method is a fully connected dense graph. Such a dense graph will introduce noise due to too many edges, which is not conducive to the aggregation of features on the same hyperedge. Therefore, this invention adopts a KNN-based method to sparsify the graph.

[0113] Furthermore, when constructing the hyperedge group of the feature hypergraph using the nearest neighbor method, in addition to using a KNN-based method to sparsify the graph, a threshold-based method can also be used to sparsify the graph:

[0114] Obtain a set N of nodes whose distance from any given node v is less than a preset threshold ∈. ∈ (v) connects the arbitrary node v with the set N of nodes. ∈ (v) Aggregate them separately to generate multiple second hyperedges, and combine the generated multiple second hyperedges together to obtain the hyperedge group E of the feature hypergraph. ∈ :

[0115] E ∈ ={N ∈ (v)∪v|v∈V}.

[0116] Furthermore, a hyperedge group of the feature hypergraph is constructed by a clustering algorithm: the arbitrary node v is grouped according to its features, a second hyperedge is generated for each group, and the nodes in the same group are connected by a second hyperedge.

[0117] Combining all the generated second hyperedges together yields the hyperedge group E of the feature hypergraph. k-means :

[0118] E k-means ={V k |1≤k≤K};

[0119] Where K represents the number of groups, k represents the ordinal number of the group, and V k This represents the superedge containing the k-th group of nodes.

[0120] As can be seen, the flexible selection strategy of the present invention for constructing the hyperedge group of the feature hypergraph can effectively capture the potential relationships between objects, providing a strong information foundation for hypergraph construction.

[0121] It should be noted that since each hyperedge has a different importance, different weights need to be assigned to them. In the distance-based hyperedge construction method, a Gaussian kernel-based method is used to calculate the weight of each hyperedge. The Gaussian kernel method determines the weight of a hyperedge by calculating the average distance between the features of any two nodes on the hyperedge.

[0122] Specifically, these are the hyperedge groups E KNN The hyper-edge group E ∈ and the super-edge group E k-means Each hyperedge in the array is assigned a weight:

[0123]

[0124] Where u and v represent two nodes, X(u) and X(v) represent the features of u and v respectively, and σ 2 Let represent the variance of the Gaussian kernel, exp() represent the exponential function, and d(X(u),X(v)) represent the distance between the features of nodes u and v.

[0125] Furthermore, according to the hyperedge group E KNN The hyper-edge group E ∈ and the super-edge group E k-means The first correlation matrix H of the feature hypergraph is represented. f :

[0126]

[0127] Step 300: Classify the non-image data to obtain multiple groups, obtain a hyperedge of the attribute hypergraph according to each group, construct a hyperedge group of the attribute hypergraph according to the multiple hyperedges, and obtain the second correlation matrix of the attribute hypergraph according to the hyperedge group of the attribute hypergraph.

[0128] like Figure 5 As shown, it can be understood that each individual possesses a series of inherent attributes, including discrete attribute data such as gender, eye type, and data source, as well as continuous attribute data such as age and BMI (Body Mass Index). These attributes allow for a direct connection between individuals. This invention treats each attribute as a hyperedge, aggregating individuals with the same attribute value onto the same hyperedge.

[0129] Specifically, the non-image data of multiple users to be analyzed are grouped according to attributes, and the non-image data includes discrete attribute data and continuous attribute data;

[0130] Users with the same discrete attribute data are grouped onto the same hyperedge. Attribute intervals are divided using a segmented approach. Users with continuous attribute data falling within the same attribute interval are grouped onto the same hyperedge, thus constructing the hyperedge group E of the attribute hypergraph. attribute :

[0131] E attribute ={N attribute (a)|a∈A};

[0132] Where, N attribute (a) is the set of nodes that have the same attribute a, and A represents the set of all attributes.

[0133] It is understandable that in the process of constructing the hyperedge group of the attribute hypergraph, for discrete attributes such as gender, eye type, and data source, this invention forms a hyperedge by connecting nodes with the same attribute value. Using mathematical notation, let A denote the set of all attributes. For any discrete attribute a∈A, a discrete value a i All individuals with the same ai value form a hyperedge; for continuous attribute data such as age and BMI, this invention adopts a segmented partitioning method. For each continuous attribute value, it is divided into several intervals to ensure that individuals whose attribute values ​​fall into the same interval are aggregated on the same hyperedge, thereby accurately constructing the hyperedge group of the attribute hypergraph.

[0134] Furthermore, for attribute-based constructed hyperedges, the number of shared attributes between samples connected by the hyperedge can quantitatively reflect the relative correlation strength. The more attributes the samples share, the stronger the connection between these corresponding vertices, and the greater the weight assigned to the hyperedge. Therefore, the heatkerne method is used to calculate the weight of the hyperedge.

[0135] Specifically, through the hyperedge group E attribute Obtain the second association matrix H of the attribute hypergraph attribute and for the hyperedge group E attribute Each hyperedge in the array is assigned a weight:

[0136]

[0137] Where δ(e) represents the degree of the hyperedge e, and ||2| represents the Euclidean distance.

[0138] As can be seen, this invention uses a distance-based method to construct a feature hypergraph and an attribute-based method to construct an attribute hypergraph, effectively integrating image data and non-image data, and providing a multi-dimensional information foundation for hypergraph modeling. This fusion strategy aims to fully explore the correlations between different data types, providing more comprehensive information support for integrated analysis and modeling.

[0139] Step 400: Concatenate the first association matrix and the second association matrix to obtain the target association matrix of the target hypergraph.

[0140] This invention constructs feature hypergraphs and attribute hypergraphs respectively, and obtains their correlation matrices (the first correlation matrix H). f Second correlation matrix and H attribute To effectively fuse image and non-image data, this invention employs a matrix stitching method, merging two initialized correlation matrices to form the correlation matrix H of the initial hypergraph, i.e., Multi-Hypergraph.

[0141] Specifically, the first correlation matrix of the feature hypergraph and the second correlation matrix of the attribute hypergraph are merged using a matrix concatenation method to obtain the correlation matrix H of the initial hypergraph:

[0142] H = H f ||H attribute ;

[0143] As the network depth increases, the correlation matrix of the initial hypergraph is dynamically adjusted using dynamic graph construction techniques to obtain the target correlation matrix of the target hypergraph.

[0144] Understandably, the initially constructed graph structure may not be perfectly suited to representing the data, as feature embeddings dynamically adjust with increasing network depth. Therefore, this invention introduces a dynamic graph construction process. As the network layers deepen, the structure of the feature hypergraph adjusts in real-time according to changes in feature embeddings, ensuring the graph structure more closely reflects the true characteristics of the data. This dynamically adaptable strategy provides a more accurate and effective method for representing and processing complex datasets.

[0145] Step 500: Input the visual features and the target association matrix into the weighted hypergraph convolution module for feature extraction to obtain target node features. Obtain the probability of each category based on the target node features and generate a fundus image analysis report based on the probability.

[0146] like Figure 6 As shown, the weighted hypergraph convolution module consists of two sub-modules: a node convolution sub-module and a hyperedge convolution sub-module. First, the node convolution sub-module aggregates node features onto the hyperedge, and then the hyperedge convolution sub-module aggregates adjacent hyperedge features onto the centroid vertex through hyperedge convolution.

[0147] Specifically, the visual features in the l-th layer are the features X. (l) The transpose of the target association matrix is ​​input into a weighted hypergraph convolution module for matrix multiplication. This module aggregates the features of the nodes connected to each hyperedge in the target association matrix onto the hyperedge, resulting in a hyperedge feature matrix H. (l) .

[0148] However, the aggregation operation described above is actually a form of average pooling, which is difficult to capture well the differences between node features. To better model the differences between node features, we introduce a matrix of learnable node weights.

[0149] Set a matrix that can learn node weights. The matrix Perform a Hadamard product operation with the target correlation matrix, assign weights to each node on the hyperedge, and obtain the target hyperedge feature matrix E. (l) :

[0150]

[0151] Where ⊙ represents the Hadamard product operation, and T represents the matrix transpose.

[0152] Furthermore, considering the varying importance of node features, a matrix of learnable node weights is set up. The matrix A Hadamard product is performed with the target association matrix to assign weights to each hyperedge containing a node, resulting in a weighted node feature matrix. This weighted node feature matrix is ​​then processed by a learnable parameter Θ. (l) A feature space transformation is performed, and a residual operation is introduced to obtain the target node feature X. (l+1) :

[0153]

[0154] Where ρ represents the activation function, and WHGCN() represents the output of the weighted hypergraph convolution module.

[0155] Understandably, the residual operation is introduced to mitigate the smoothing problem that may result from an excessive number of weighted hypergraph convolutional layers. By dynamically adjusting the graph structure, the relationship between node features is effectively captured, thereby improving the model's performance and expressive power.

[0156] Furthermore, the target node feature X (l+1) The probability of each user under each category is obtained by outputting the softmax (normalized exponential function) layer.

[0157] As can be seen, this invention addresses the limitations of existing technologies in analyzing fundus images, which cannot effectively model interactions and relationships between different users and handle complex high-order relationships between users, leading to inaccurate analysis results. To address these limitations, this invention proposes a hypergraph model for fundus image analysis. This model combines image data and non-image data to construct a multi-hypergraph structure to model high-order relationships between individuals. Hypergraph convolution is used to efficiently transfer information within the multi-hypergraph structure. During model training, feature hypergraphs are dynamically constructed based on changes in node embeddings, better adapting to the true characteristics of the data and improving the model's ability to handle complex data structures. The method of this invention can dynamically capture the relationships between features while considering both fundus image and non-image data, and fully utilize the hypergraph structure for comprehensive learning to improve the performance and accuracy of fundus image classification.

[0158] Furthermore, such as Figure 7 As shown, based on the above-described fundus image analysis method, the present invention also provides a fundus image analysis system, wherein the fundus image analysis system includes:

[0159] User data acquisition module 51 is used to acquire target fundus images and target non-imaging data of multiple users to be analyzed, input the fundus images into a pre-trained encoder, and output visual features.

[0160] The first association matrix construction module 52 is used to take the visual features as the features of the nodes of the feature hypergraph, construct the hyperedge group of the feature hypergraph according to the nodes through a preset algorithm, and obtain the first association matrix of the feature hypergraph according to the hyperedge group of the feature hypergraph.

[0161] The second association matrix construction module 53 is used to classify the non-image data to obtain multiple groups, obtain a hyperedge of the attribute hypergraph according to each group, construct a hyperedge group of the attribute hypergraph according to multiple hyperedges, and obtain the second association matrix of the attribute hypergraph according to the hyperedge group of the attribute hypergraph.

[0162] The target association matrix generation module 54 is used to concatenate the first association matrix and the second association matrix to obtain the target association matrix of the target hypergraph.

[0163] The feature extraction module 55 is used to input the visual features and the target association matrix into the weighted hypergraph convolution module for feature extraction, obtain target node features, obtain the probability of each category based on the target node features, and generate a fundus image analysis report based on the probability.

[0164] Furthermore, such as Figure 8 As shown, based on the above-mentioned fundus image analysis method and system, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 8 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0165] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a fundus image analysis program 40, which can be executed by the processor 10 to implement the fundus image analysis method of this application.

[0166] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the fundus image analysis method.

[0167] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The components 10-30 of the terminal communicate with each other via a system bus.

[0168] In one embodiment, when the processor 10 executes the fundus image analysis program 40 in the memory 20, the following steps are performed:

[0169] Acquire target fundus images and target non-imaging data of multiple users to be analyzed, input the fundus images into a pre-trained encoder, and output visual features;

[0170] The visual features are used as the features of the nodes of the feature hypergraph. A preset algorithm is used to construct the hyperedge group of the feature hypergraph based on the nodes. The first correlation matrix of the feature hypergraph is obtained based on the hyperedge group of the feature hypergraph.

[0171] The non-image data is classified into multiple groups, and a hyperedge of an attribute hypergraph is obtained according to each group. The hyperedge group of the attribute hypergraph is constructed according to the multiple hyperedges, and the second correlation matrix of the attribute hypergraph is obtained according to the hyperedge group of the attribute hypergraph.

[0172] The first and second correlation matrices are concatenated to obtain the target correlation matrix of the target hypergraph.

[0173] The visual features and the target association matrix are input into a weighted hypergraph convolution module for feature extraction to obtain target node features. The probability of each category is obtained based on the target node features, and a fundus image analysis report is generated based on the probability.

[0174] The feature hypergraph, the attribute hypergraph, and the target hypergraph all include a node set V and a hyperedge set E.

[0175] For any hyperedge e∈E in the hyperedge set E, there exists a weight w(e)∈W, where W is the hyperedge weight matrix, defined as:

[0176] W = diag(w(e1), w(e2), ..., w(e...)|E| ));

[0177] Where, diag represents a diagonal matrix, and |E| represents the total number of hyperedges;

[0178] The structures of the feature hypergraph, the attribute hypergraph, and the target hypergraph are all represented by an association matrix H of size |V|×|E|, where each element of the association matrix H is defined as:

[0179]

[0180] Where H(v,e)=1 indicates that any node v∈V in the node set V is on the hyperedge e, and H(v,e)=0 indicates that any node v∈V in the node set V is not on the hyperedge e;

[0181] For any node v∈V, the degree of the node is represented as:

[0182] d(v)=∑ e∈E w(e)H(v,e);

[0183] For any hyperedge e∈E, the degree of the hyperedge is represented as:

[0184]

[0185] The preset algorithm includes the nearest neighbor method and the clustering method;

[0186] The step of constructing the hyperedge group of the feature hypergraph based on the nodes using a preset algorithm specifically includes:

[0187] Calculate the distance between all the nodes in the feature space, and combine the distance between any node v and the set of its neighbor nodes N. k(v) Aggregates to generate multiple first hyperedges e(v):

[0188] e(v) = N k(v) ∪v;

[0189] Where, N k(v) This represents the set of the k nearest neighbor nodes to any given node v;

[0190] The generated multiple first hyperedges are combined together to obtain the hyperedge group E of the feature hypergraph. KNN :

[0191] E KNN ={e(v)|v∈V};

[0192] Obtain a set N of nodes whose distance from any given node v is less than a preset threshold ∈. ∈(v) connects the arbitrary node v with the set N of nodes. ∈ (v) Aggregate them separately to generate multiple second hyperedges, and combine the generated multiple second hyperedges together to obtain the hyperedge group E of the feature hypergraph. ∈ :

[0193] E ∈ ={N ∈ (v)∪v|v∈V};

[0194] Using a clustering algorithm, the arbitrary node v is grouped according to its characteristics, and a second hyperedge is generated for each group. Nodes in the same group are connected by the second hyperedge.

[0195] Combining all the generated second hyperedges together yields the hyperedge group E of the feature hypergraph. k-means :

[0196] K k-means ={V k |1≤k≤K};

[0197] Where K represents the number of groups, k represents the ordinal number of the group, and V k This represents the superedge containing the k-th node;

[0198] The super-edge group E is respectively KNN The hyper-edge group E ∈ and the super-edge group E k-means Each hyperedge in the array is assigned a weight:

[0199]

[0200] Where u and v represent two nodes, X(u) and X(v) represent the features of u and v respectively, and σ 2 Let represent the variance of the Gaussian kernel, exp() represent the exponential function, and d(X(u),X(v)) represent the distance between the features of nodes u and v.

[0201] Specifically, obtaining the first correlation matrix of the feature hypergraph based on the hyperedge group of the feature hypergraph includes:

[0202] According to the hyperedge group E KNN The hyper-edge group E ∈ and the super-edge group E k-means The first correlation matrix H of the feature hypergraph is represented. f :

[0203]

[0204] Specifically, the process of classifying the non-imagery data to obtain multiple groups, obtaining a hyperedge of an attribute hypergraph based on each group, constructing a hyperedge group of the attribute hypergraph based on multiple hyperedges, and obtaining a second correlation matrix of the attribute hypergraph based on the hyperedge group of the attribute hypergraph includes:

[0205] The non-image data of multiple users to be analyzed are grouped according to attributes, including discrete attribute data and continuous attribute data;

[0206] Users with the same discrete attribute data are grouped onto the same hyperedge. Attribute intervals are divided using a segmented approach. Users with continuous attribute data falling within the same attribute interval are grouped onto the same hyperedge, thus constructing the hyperedge group E of the attribute hypergraph. attribute :

[0207] E attribute ={N attribute (a)|a∈A};

[0208] Where, N attribute (a) is the set of nodes that have the same attribute a, and A represents the set of all attributes;

[0209] Through the hyperedge group E attribute Obtain the second association matrix H of the attribute hypergraph attribute and for the hyperedge group E attribute Each hyperedge in the array is assigned a weight:

[0210]

[0211] Where δ(e) represents the degree of the hyperedge e, and ||2| represents the Euclidean distance.

[0212] Specifically, the step of concatenating the first association matrix and the second association matrix to obtain the target association matrix of the target hypergraph includes:

[0213] The first association matrix of the feature hypergraph and the second association matrix of the attribute hypergraph are merged using a matrix concatenation method to obtain the association matrix H of the initial hypergraph:

[0214] H = H f ||H attribute ;

[0215] As the network depth increases, the correlation matrix of the initial hypergraph is dynamically adjusted using dynamic graph construction techniques to obtain the target correlation matrix of the target hypergraph.

[0216] Specifically, the step of inputting the visual features and the target association matrix into a weighted hypergraph convolution module for feature extraction to obtain target node features, and obtaining the probability of each category based on the target node features, includes:

[0217] The visual features in the l-th layer are X (l) The transpose of the target association matrix is ​​input into a weighted hypergraph convolution module for matrix multiplication. This module aggregates the features of the nodes connected to each hyperedge in the target association matrix onto the hyperedge, resulting in a hyperedge feature matrix H. (l) ;

[0218] Set a matrix that can learn node weights. The matrix Perform a Hadamard product operation with the target correlation matrix, assign weights to each node on the hyperedge, and obtain the target hyperedge feature matrix E. (l) :

[0219]

[0220] Where ⊙ represents the Hadamard product operation, and T represents the matrix transpose;

[0221] Set a matrix that can learn node weights. The matrix A Hadamard product is performed with the target association matrix to assign weights to each hyperedge containing a node, resulting in a weighted node feature matrix. This weighted node feature matrix is ​​then processed by a learnable parameter Θ. (l) A feature space transformation is performed, and a residual operation is introduced to obtain the target node feature X. (l+1) :

[0222]

[0223] Where ρ represents the activation function, and WHGCN() represents the output of the weighted hypergraph convolution module;

[0224] The target node feature X (l+1) The probability of each user to be analyzed in each category is obtained by outputting the softmax layer.

[0225] In summary, this invention provides a method and related equipment for fundus image analysis. The method includes: acquiring a target fundus image and target non-imaging data; inputting the fundus image into an encoder to output visual features; constructing a hyperedge group of the feature hypergraph using the visual features as features of the nodes, obtaining a first correlation matrix of the feature hypergraph; classifying the non-imaging data to obtain multiple groups to obtain hyperedges of the attribute hypergraph; constructing a hyperedge group of the attribute hypergraph based on the multiple hyperedges, obtaining a second correlation matrix of the attribute hypergraph; concatenating the first and second correlation matrices to obtain a target correlation matrix of the target hypergraph; inputting the nodes of the feature hypergraph and the target correlation matrix into a weighted hypergraph convolution module to obtain target node features; obtaining the probability of each category based on the target node features; and generating a fundus image analysis report based on the probabilities. This invention demonstrates superior performance in processing fundus image analysis tasks; it can dynamically capture the correlation between features while considering fundus images and non-imaging data, and fully utilize the hypergraph structure for comprehensive learning to improve the performance and accuracy of fundus image classification.

[0226] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0227] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0228] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for analyzing fundus images, characterized in that, The fundus image analysis method includes: Acquire target fundus images and target non-imaging data of multiple users to be analyzed, input the fundus images into a pre-trained encoder, and output visual features; The visual features are used as the features of the nodes of the feature hypergraph. A preset algorithm is used to construct the hyperedge group of the feature hypergraph based on the nodes. The first correlation matrix of the feature hypergraph is obtained based on the hyperedge group of the feature hypergraph. The non-image data is classified into multiple groups, and a hyperedge of an attribute hypergraph is obtained according to each group. The hyperedge group of the attribute hypergraph is constructed according to the multiple hyperedges, and the second correlation matrix of the attribute hypergraph is obtained according to the hyperedge group of the attribute hypergraph. The first and second correlation matrices are concatenated to obtain the target correlation matrix of the target hypergraph. The visual features and the target association matrix are input into a weighted hypergraph convolution module for feature extraction to obtain target node features. The probability of each category is obtained based on the target node features, and a fundus image analysis report is generated based on the probability. The process of classifying the non-imagery data to obtain multiple groups, obtaining a hyperedge of an attribute hypergraph based on each group, constructing a hyperedge group of the attribute hypergraph based on multiple hyperedges, and obtaining a second correlation matrix of the attribute hypergraph based on the hyperedge group of the attribute hypergraph specifically includes: The non-image data of multiple users to be analyzed are grouped according to attributes, including discrete attribute data and continuous attribute data; Users with the same discrete attribute data are grouped onto the same hyperedge. Attribute intervals are divided using a segmented approach. Users with continuous attribute data falling within the same attribute interval are grouped onto the same hyperedge to construct the hyperedge group of the attribute hypergraph. : ； in, They have the same attributes The set of nodes, Represents a collection of all attributes; Through the hyperedge group Obtain the second association matrix of the attribute hypergraph and for the hyper-edge group Each hyperedge in the array is assigned a weight: ； in, Indicate the degree of the hyperedge e. Represents Euclidean distance; The step of inputting the visual features and the target association matrix into a weighted hypergraph convolution module for feature extraction to obtain target node features, and obtaining the probability of each category based on the target node features, specifically includes: The visual features in the first Layer characteristics The transpose of the target correlation matrix is ​​input into a weighted hypergraph convolution module for matrix multiplication. This module aggregates the features of the nodes connected to each hyperedge in the target correlation matrix onto the hyperedge to obtain the hyperedge feature matrix. ; Set a matrix that can learn node weights. , the matrix Perform a Hadamard product operation with the target correlation matrix, assign weights to each node on the hyperedge, and obtain the target hyperedge feature matrix. : ； in, This represents the Hadamard product operation. Indicates matrix transpose; Set a matrix that can learn node weights. , the matrix Perform a Hadamard product operation with the target association matrix, assign weights to each hyperedge containing a node, and obtain a weighted node feature matrix. Then, pass learnable parameters to the weighted node feature matrix. A feature space transformation is performed, and a residual operation is introduced to obtain the target node features. : ； in, This represents the activation function. This represents the output of the weighted hypergraph convolution module; The target node features The probability of each user to be analyzed in each category is obtained by outputting the softmax layer.

2. The fundus image analysis method according to claim 1, characterized in that, The feature hypergraph, the attribute hypergraph, and the target hypergraph all include a node set. and super edge ; The hyperedge set any superedge There is a weight for each. , Let be the hyperedge weight matrix, defined as: ； in, Represents a diagonal matrix. This represents the total number of superedges; The structure of the feature hypergraph, the attribute hypergraph, and the target hypergraph all consists of a structure of size [size missing]. correlation matrix The association matrix indicates that... Each element in is defined as: ； in, Represents the node set any node on In the super-edge superior, Represents the node set any node on Not beyond the edge superior; For any node The degree of a node is represented as: ； For any hyperedge The degree of a hyperedge is represented as: 。 3. The fundus image analysis method according to claim 2, characterized in that, The preset algorithm includes the nearest neighbor method and the clustering method; The step of constructing the hyperedge group of the feature hypergraph based on the nodes using a preset algorithm specifically includes: Calculate the distance between all the nodes in the feature space, and then set the arbitrary node... and any node The set of neighboring nodes Multiple first super edges are generated by aggregation. : ； in, Indicates the relationship with any of the nodes. The set of the k nearest neighbor nodes; The generated multiple first hyperedges are combined together to obtain the hyperedge group of the feature hypergraph. : ； Get the connection with any node The distance is less than a preset threshold Multiple node sets , to the arbitrary node With multiple sets of nodes Multiple second hyperedges are generated by aggregating them separately, and then the generated multiple second hyperedges are combined together to obtain the hyperedge group of the feature hypergraph. : ； Using a clustering algorithm, based on any node Features for any node Group the nodes and generate a second hyperedge for each group. Connect the nodes in the same group through the second hyperedge. Combining all the generated second hyperedges together yields the hyperedge group of the feature hypergraph. : ； in, Indicates the number of groups. Indicates the ordinal number of the group. Indicates the first The superedge containing the group node; These are the hyper-edge groups The hyper-edge group and the hyper-edge group Each hyperedge in the array is assigned a weight: ； in, and These represent two nodes respectively. and They represent and Features The variance of the Gaussian kernel is represented. Represents an exponential function. Represents a node and The distance between features.

4. The fundus image analysis method according to claim 3, characterized in that, The step of obtaining the first correlation matrix of the feature hypergraph based on the hyperedge group of the feature hypergraph specifically includes: According to the hyperedge group The hyper-edge group and the hyper-edge group Represent the first correlation matrix of the feature hypergraph. : 。 5. The fundus image analysis method according to claim 1, characterized in that, The step of concatenating the first and second correlation matrices to obtain the target correlation matrix of the target hypergraph specifically includes: The first association matrix of the feature hypergraph and the second association matrix of the attribute hypergraph are merged using a matrix concatenation method to obtain the association matrix of the initial hypergraph. : ； As the network depth increases, the correlation matrix of the initial hypergraph is dynamically adjusted using dynamic graph construction techniques to obtain the target correlation matrix of the target hypergraph.

6. A fundus image analysis system, characterized in that, The fundus image analysis system is used to implement the fundus image analysis method according to any one of claims 1-5, and the fundus image analysis system includes: The user data acquisition module is used to acquire target fundus images and target non-imaging data of multiple users to be analyzed, input the fundus images into a pre-trained encoder, and output visual features. The first correlation matrix construction module is used to take the visual features as the features of the nodes of the feature hypergraph, construct the hyperedge group of the feature hypergraph according to the nodes using a preset algorithm, and obtain the first correlation matrix of the feature hypergraph according to the hyperedge group of the feature hypergraph. The second correlation matrix construction module is used to classify the non-image data to obtain multiple groups, obtain a hyperedge of the attribute hypergraph according to each group, construct a hyperedge group of the attribute hypergraph according to multiple hyperedges, and obtain the second correlation matrix of the attribute hypergraph according to the hyperedge group of the attribute hypergraph. The target association matrix generation module is used to concatenate the first association matrix and the second association matrix to obtain the target association matrix of the target hypergraph. The feature extraction module is used to input the visual features and the target association matrix into the weighted hypergraph convolution module for feature extraction, obtain target node features, obtain the probability of each category based on the target node features, and generate a fundus image analysis report based on the probability.

7. A terminal, characterized in that, The terminal includes: a memory, a processor, and a fundus image analysis program stored in the memory and executable on the processor. When the fundus image analysis program is executed by the processor, it implements the steps of the fundus image analysis method as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a fundus image analysis program, which, when executed by a processor, implements the steps of the fundus image analysis method as described in any one of claims 1-5.

Images