Eye disease diagnosis system and method based on lesion space structure relationship

Abstract

The invention discloses an eye disease diagnosis system and method based on a lesion space structure relation, and belongs to the technical field of intelligent eye disease diagnos.The eye disease diagnosis system comprises a lesion significance enhancement module and a lesion visual distribution information fusion module; the method comprises the following steps of: firstly, obtaining the characteristics of an eye lesion pixel, further refining the saliency characteristics to obtain the characteristics of the eye lesion pixel, constructing an eye lesion graph, and finally obtaining a global distribution mode of a lesion area through pooling operation to complete the diagnosis of the eye disease. According to the eye disease diagnosis system and method provided by the invention, effective information can be provided for identification and diagnosis of different types of eye diseases, an accurate retinal disease diagnosis result can be obtained, dispersed pathological features in the fundus image and a global distribution mode of the dispersed pathological features can be comprehensively understood, the calculation speed of a network is accelerated, and the calculation cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent eye disease diagnosis, and in particular to an eye disease diagnosis system and method based on the spatial structural relationship of lesions. Background Art

[0002] Retinal diseases are a major public health issue with a wide variety of types, high incidence, and a wide age range. Common retinal diseases include diabetic retinopathy, glaucoma, age-related macular degeneration, retinal vein occlusion, etc., which are the main causes of irreversible visual impairment and permanent blindness among the working-age population worldwide. Early screening and timely treatment of retinal diseases play a vital role in slowing the progression of the disease and reducing the risk of irreversible visual impairment and even permanent blindness.

[0003] However, my country's ophthalmic medical talent resources are relatively scarce and extremely unbalanced in geographical distribution, making it difficult to achieve large-scale screening of ophthalmic diseases, causing many patients to miss the best time for treatment. At the same time, the diagnosis of retinal diseases is highly dependent on fundus medical image analysis. However, these pathological areas in fundus images are usually small in size and scattered throughout the retina, greatly increasing the difficulty of diagnosis. The research and development of auxiliary diagnosis and treatment tools for retinal diseases combined with deep learning can greatly reduce the workload of clinicians, improve diagnostic efficiency, help achieve early detection and early intervention of retinal diseases, minimize the harm of the disease, and achieve high-accuracy automated identification of retinal diseases based on fundus images, which has high application prospects and social value.

[0004] With the development of artificial intelligence, such as convolutional neural networks (CNN) and Transformers, great success has been achieved in retinal disease diagnosis tasks. For example, the invention application with publication number CN 118380157 A discloses a training method for a model for assisting eye disease diagnosis and its related device, by obtaining real eye images of multiple eye disease patients and eye disease annotations corresponding to the real eye images; using the basic ophthalmology model to extract common features in all real eye images; for each eye disease patient, all common features belonging to the eye disease patient are spliced ​​to obtain multimodal fused eye data; using the multimodal fused eye data as training samples and the eye disease annotations as labels, the model for assisting eye disease diagnosis is trained, solving the problem of single eye disease diagnosis and poor robustness of the original diagnostic model. However, this method leads to relatively poor results in classifying retinal diseases because they usually ignore some small and subtle lesion areas.

[0005] Later, models were developed to extract features from subtle lesion areas for diagnosis. The key to these methods is to use significant features about ocular lesions to match the distribution of lesions, thereby guiding the attention of the diagnostic model to focus on these subtle lesion areas and thus improve diagnostic accuracy. Although these methods have proven their effectiveness, they ignore a key aspect of retinal diseases: the spatial distribution pattern of lesion areas, that is, the relative topological relationship of these lesions.

[0006] Studies have shown that specific retinal diseases tend to appear in specific areas of the retina, showing unique patterns of lesion distribution. For example, in diabetic macular edema, hard exudates often appear as isolated streaks or clusters, or around large, ring-shaped microaneurysms. Therefore, a comprehensive understanding of the specific distribution patterns of lesions associated with retinal diseases is essential to achieve accurate diagnosis.

[0007] In order to fully explore the spatial distribution patterns of subtle lesions in fundus images and improve the classification performance of retinal diseases, it is necessary to use graph convolutional networks (GCNs) to represent the interconnected relationships between various lesion areas in retinal lesions and obtain a comprehensive understanding of the identification characteristics of retinal diseases from the distribution patterns of lesion areas. Summary of the invention

[0008] The purpose of the present invention is to provide an eye disease diagnosis system and method based on the spatial structural relationship of lesions. The system extracts significant features related to lesions from fundus images, further refines the significant features to obtain the characteristics of eye lesion pixels, constructs an eye lesion map, and finally obtains the global distribution pattern of the lesion area through pooling operation to complete the diagnosis of eye diseases.

[0009] To achieve the above-mentioned purpose of the invention, an embodiment provides an eye disease diagnosis system based on the spatial structural relationship of lesions, wherein the eye disease diagnosis system includes a lesion significance enhancement module and a lesion visual distribution information fusion module;

[0010] The lesion significance enhancement module is used to divide the input fundus image into a number of image blocks, mine and extract lesion features for each image block, and generate lesion significance features of the image block;

[0011] The lesion visual distribution information fusion module is used to identify image blocks containing eye lesions based on significant features, identify and obtain the features of eye lesion pixels through image blocks containing eye lesions, generate pixel lesion nodes according to the features of eye lesion pixels, establish edges based on pixel lesion nodes to regenerate an eye lesion map, and introduce a hybrid pooling operation to construct a global distribution pattern of the eye lesion area.

[0012] In one embodiment, the mining and extraction of lesion features for each image block includes: each image block undergoes four progressive stages of feature representation, each stage includes a number of Transformer blocks, and the feature representation of the last stage is processed by a multi-layer perceptron to determine whether the image block contains classification information of eye lesions, thereby completing the extraction of lesion features.

[0013] In one embodiment, the identifying of image blocks containing eye lesions based on significant features includes: based on the significant features, using a significant encoder to generate block embeddings with significant feature information, inputting the block embeddings into a discriminative multi-layer perceptron, and obtaining a plurality of block embeddings P containing eye lesions and image blocks containing eye lesions.

[0014] In one embodiment, the method of identifying and obtaining the features of eye lesion pixels based on the image block containing the eye lesion comprises: embedding P based on the i-th block i and the jth pixel C in the i-th block embedding ij , pixel C ij Laplace features and block embedding P i Add and calculate the fused features l ij , and then generate the feature P of the eye lesion pixel through weighted summation ij .

[0015] In one embodiment, generating a pixel lesion node based on the characteristics of eye lesion pixels includes: based on the characteristics of the eye lesion pixels, using a threshold to make a determination, setting pixels exceeding the threshold as pixel lesion nodes, and obtaining pixel node characteristics based on the pixel lesion nodes.

[0016] In one embodiment, the method of establishing edges based on pixel lesion nodes to regenerate an eye lesion map includes: a new graph construction algorithm based on the minimum Manhattan distance in four quadrants, by dividing the plane associated with each pixel lesion node into four quadrants, calculating the distance between a given pixel lesion node and the nearest pixel lesion node in each quadrant through Manhattan distance, obtaining the edges of the eye lesion map, calculating the weights of the edges according to the Manhattan distance, and generating the eye lesion map.

[0017] In one example, the introduction of a hybrid pooling operation to construct a global distribution pattern of the eye lesion area includes: introducing edge cutting pooling and edge folding pooling, and deriving and constructing a global distribution pattern of the eye lesion area based on information transfer between edge cutting pooling and edge folding pooling.

[0018] In one example, the edge cutting and pooling includes: calculating the attention weight of each edge based on the eye lesion map and the pixel node features, deriving new pixel node features by aggregating adjacent pixel nodes based on the attention weights, obtaining potential eye lesion cluster nodes through edge cutting and pooling based on the eye lesion map and the new pixel node features, and constructing a new lesion cluster map;

[0019] The edge folding pooling includes: based on the graph attention mechanism, taking the eye lesion cluster and the node features of the eye lesion cluster as input, calculating the downsampled eye lesion cluster and the node features of the eye lesion cluster through edge folding pooling, and introducing auxiliary loss to standardize the difference between the eye lesion cluster and the downsampled eye lesion cluster.

[0020] The present invention also provides an eye disease diagnosis method based on the spatial structural relationship of lesions, wherein the eye disease diagnosis method uses the eye disease diagnosis system and comprises the following steps:

[0021] Collect several fundus images and perform image preprocessing, normalize the resolution of all images and convert the format, and then perform data augmentation to obtain the training data set;

[0022] Inputting the training data set into the eye disease diagnosis system for training, adjusting parameters during the training process, optimizing the parameters of the eye disease diagnosis system, and completing the training;

[0023] Collect several fundus images and send them to the trained eye disease diagnosis system to verify the eye disease recognition accuracy of the eye disease diagnosis system.

[0024] In one embodiment, the data augmentation includes: rotation, scaling, Gaussian blur, color jitter and mirroring operations.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] (1) By constructing an eye disease diagnosis system based on a lesion saliency enhancement module and a lesion visual distribution information fusion module and iteratively training it, it is able to fully understand the scattered pathological features and global distribution patterns in fundus images, overcoming the shortcomings of previous models that are limited to the most significant lesion features, enhancing pathological semantics from the perspective of the global distribution pattern of lesions, and improving the accuracy of retinal disease classification, so that the accuracy of eye disease judgment is not lower than that of experienced doctors;

[0027] (2) The hybrid pooling operation is introduced into the lesion visual distribution information fusion module, which can fully utilize the dependencies between different lesion areas and analyze the spatial distribution patterns on the image, significantly reducing the complexity of the calculation. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below.

[0029] Figure 1 This is a schematic diagram of the structure of the eye disease diagnosis system provided by the present invention.

[0030] Figure 2 Schematic diagram of the potential eye lesion pixel recognition structure based on the Laplace operator provided by the present invention.

[0031] Figure 3 This is a schematic diagram of constructing an eye lesion map provided by the present invention.

[0032] Figure 4 This is a schematic diagram of the mixing pooling operation provided by the present invention.

[0033] Figure 5 This is a flow chart of the eye disease diagnosis method provided by the present invention. DETAILED DESCRIPTION

[0034] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0035] In order to diagnose and classify various retinal diseases, including age-related macular degeneration, high myopia, glaucoma, venous occlusion images, diabetic retinopathy and normal fundus images, so as to enhance the ability of doctors to perform efficient intelligent diagnosis, an embodiment provides an eye disease diagnosis system based on the spatial structure relationship of lesions, such as Figure 1 As shown, the eye disease diagnosis system includes a lesion significance enhancement module and a lesion visual distribution information fusion module (Image-Graph).

[0036] The lesion saliency enhancement module is used to divide the input fundus image into several image blocks, mine and extract the lesion features of each image block, and generate the lesion saliency features of the image block.

[0037] In the embodiment, given a specific fundus image X∈R h×w×3 , h×w×3 represents a three-channel image with an input of h×w, where h and w represent the height and width of the input fundus image, respectively. Based on the input fundus image, the lesion saliency enhancement module uses two convolution kernels of different sizes (3×3, 5×5) to sample the image blocks of each stage with a step size of 2 to obtain several image blocks.

[0038] Feature extraction and mining are performed on each image block. Specifically, the feature extraction layer is a four-stage hierarchical structure consisting of four progressive stages. The image block undergoes feature representation in four stages in turn, and each stage contains several Transformer blocks. The feature information of the last stage is processed by a multi-layer perception (MLP) to determine whether the image block contains classification information of eye lesions and complete the extraction of lesion features. In the last stage, a saliency encoder is used to generate block embeddings with saliency information. By inputting these block embeddings into a discriminative multi-layer perception machine, the block embeddings P∈R of n potential lesion blocks are obtained. n×d , where d represents the embedding dimension. Through this process, n potential lesion blocks and their corresponding block embeddings P are identified.

[0039] The lesion visual distribution information fusion module is used to identify image blocks containing eye lesions based on significant features, identify and obtain the features of eye lesion pixels through image blocks containing eye lesions, generate pixel lesion nodes according to the features of eye lesion pixels, establish edges based on pixel lesion nodes to regenerate the eye lesion map, and introduce hybrid pooling operations to construct a global distribution pattern of the eye lesion area.

[0040] To obtain more fine-grained information, such as Figure 2 As shown in Fig. 1, the features of the potential lesion pixels in the potential lesion block are further identified in the lesion map node generation module of the potential pixel. In this process, given the i-th block embedding P i ∈R 1 ×d and the jth pixel C in the i-th block ij ∈R 1×3 , generate a new pixel feature P ij ∈R 1×d Specifically, the pixel C is calculated by adding the pixel Laplacian operator and the block embedding. ij Features:

[0041]

[0042] Among them, Laplacian (C ij , o) represents the Laplacian operation of the o-order neighbor, Denotes a linear transformation so that l ij The dimension and P i (from 3 to d) are aligned. Then, the pixel feature P is obtained by weighted summation operation ij :

[0043] P ij = l ij+s·l ij

[0044] s=Softmax(GELU(γ(l ij )))

[0045] Where s represents pixel C ij The significance value of GELU is the activation function, and γ is a linear transformation to convert l ij The dimension of is mapped to 1. By applying these operations to all pixels, pixel-level saliency features in image patches of potential eye lesions are obtained.

[0046] In the embodiment, an additional threshold ε is used to determine the pixel significance value s. If the pixel significance value exceeds the threshold ε, each pixel is regarded as an initial pixel lesion node of the eye lesion map. Based on this method, the pixel lesion node V and the pixel node feature F are obtained.

[0047] After obtaining the potential lesion pixels, in order to construct a connected eye lesion map, the embodiment uses a new map construction algorithm based on the minimum Manhattan distance in four quadrants. Figure 3 As shown in Figure 1, the connectivity of each pixel lesion node is redefined by splitting the plane associated with each pixel lesion node into four quadrants. Specifically, an edge is established between a given pixel lesion node and the nearest pixel lesion node in each quadrant. The distance between any two pixel lesion nodes can be calculated using the Manhattan distance. By applying this operation to all pixel lesion nodes, the edges of the eye lesion graph are obtained. In addition, for any edge (v i , v j )∈E, the edge weight is calculated using Manhattan distance as follows:

[0048]

[0049] In this way, an eye lesion graph G = (V, E; W) is constructed. This process not only accurately defines the spatial relationship between nodes in the eye lesion graph, but also reflects the relative distance between nodes through weights, which is helpful for the subsequent analysis of the eye lesion graph and the lesion detection process.

[0050] Subsequently, based on the ocular lesion graph G, a variant of the node cluster pooling operation is used to extract the global distribution pattern of the lesion area while limiting the use of computing resources.

[0051] In the embodiment, by introducing a hybrid pooling operation, several information transfer operations are used to derive the global distribution of the eye lesion area and avoid the loss of hierarchical and topological information. Therefore, the introduced hybrid pooling operation can avoid the disadvantages of a single pooling operation. The components of the hybrid pooling operation are as follows: Figure 4As shown, it includes edge cutting pooling and edge folding pooling.

[0052] Edge Cut Pooling Component: In the edge cut pooling component, a three-step process is followed to reduce pixel lesion nodes to lesion cluster nodes. First, given the lesion graph G = (V, E; W) and the pixel node features F, the edge (v i , v j )∈E’s attention weight A ij As shown below:

[0053]

[0054] where F i and F j They are v i and v j The pixel node features of adj(i) represent v i The adjacent nodes of , σ represents the dimension of the pixel node feature F from d to The linear transformation of is the connection operation, and LeakyReLu is the activation function.

[0055] Afterwards, a new node feature F′ is derived by aggregating neighboring nodes based on the attention weight A. For any pixel-lesion node v i , node feature F′ i It is calculated as follows:

[0056]

[0057] where δ is a linear transformation that maps F to another space without changing its dimension d. We then use a cut-pooling operation on the boundary of the lesion region to generate features for the lesion cluster nodes in the following way:

[0058] (C, G′, L cut ) = CutPool(F′, G′)

[0059] Where C represents the characteristics of the new eye lesion cluster node, G′=(V′, E′) is the new lesion cluster graph with lesion cluster node V′ and edge E′, and L cut is an auxiliary loss that aims to maximize the probability that each lesion pixel is associated with a single lesion cluster. In this way, a latent lesion cluster map is obtained, thereby reducing the number of nodes in the original eye lesion map to a manageable size.

[0060] Edge Collapse Pooling Component: The edge collapse component includes t-pair graph attention mechanism and edge collapse pooling layer. In the i-th pair, the lesion cluster graph G′ is used i and lesion cluster node characteristics As input, the downsampled lesion cluster graph G′ is calculated by edge-folding pooling operation i+1 And the corresponding lesion cluster characteristics The details are as follows:

[0061] D i =GAT(C i , G i )

[0062] (C′ i+1 , G′ i+1 , L diff )=CLusterPool(D i , G′ i , r)

[0063] Among them, GAT is the graph attention mechanism, D i is with d i The cluster features of the same dimension; r is the edge collapse ratio, which represents the percentage of reduced edges; L diff is an auxiliary loss used to normalize G′ i and G′ i+1 In particular, when i = 1, G′ 1 , C′ 1 represent G′ and C′ respectively, and except for the first and last edge-folding pooling layers, d i+1 =d i When i+1 =2d i .

[0064] After the last edge-folding pooling layer, the lesion distribution pattern characteristics are obtained. In addition, by calculating C t The average value of is taken as the representation of the global distribution pattern of eye lesions. By integrating the above two pooling components, the pooling operation is customized according to the specific characteristics of the lesion map, thus circumventing the limitations of a single pooling operation.

[0065] like Figure 5 As shown, based on the structure of the above-mentioned eye disease diagnosis system based on the spatial structural relationship of lesions, the present invention also provides an eye disease diagnosis method based on the spatial structural relationship of lesions, comprising the steps of:

[0066] Collect several fundus images and perform image preprocessing, normalize the resolution of all images and convert the format, and then perform data augmentation to obtain the training data set;

[0067] Inputting the training data set into the eye disease diagnosis system for training, adjusting parameters during the training process, optimizing the parameters of the eye disease diagnosis system, and completing the training;

[0068] Collect several fundus images and send them to the trained eye disease diagnosis system to verify the eye disease recognition accuracy of the eye disease diagnosis system.

[0069] In order to better illustrate the effectiveness of the system and method provided by the present invention, the collected fundus image dataset, EyePACS dataset and RFMiD dataset are used as sample data, and the accuracy, AUC value, F1-score and Kappa coefficient are used as indicators for comparison.

[0070] During the training process, a linear warm-up learning rate was used for preliminary adjustment for 20 epochs, followed by a cosine decay learning rate scheduler and an AdamW optimizer for 500 epochs of training; the cross entropy loss function was used to measure the difference between the model prediction value and the true value. The batch size was 128, the initial learning rate was 0.001, the weight decay was 0.05, and the gradient clipping with a maximum norm of 1.0 was used to stabilize the training process. In order to improve the generalization ability and robustness of the model, data augmentation strategies such as rotation, scaling, Gaussian blur, color jitter and mirroring operations were used during the training process, and the experimental results are shown in Table 1. In the embodiment, a comparison was made with the latest technical methods, in which the eye disease diagnosis method based on the spatial structural relationship of lesions (Saliency Image-Graph, SIGraph) of this embodiment significantly surpassed these advanced architectures in performance.

[0071] Table 1

[0072]

[0073] For all compared models, the embodiments maintain their original hyperparameter settings, while standardizing the training cycle and data partitioning to facilitate fair comparison with the SIGraph of the present invention, especially when the original method lacks performance indicators for specific data sets. Among them, model 1 is Zoom-in-Net, model 2 is GREEN, model 3 is GRADING, model 4 is Vit-B / 16, model 5 is CrossFormer-L, model 6 is Swin-B, model 7 is PVTv2-B5, model 8 is SatFormer-B, and model 9 is RETFound.

[0074] The results in Table 1 show that compared with the latest CNN- and GCN-based methods (Zoom-in-Net, GREEN, and GRADING), the SIGraph method provided by the present invention has achieved significant improvement in performance, and the Kappa index on the EyePACS dataset has been improved by at least 3.1%.

[0075] In comparison with Transformer-based methods, including the latest methods such as Vit-B / 16, CrossFormer-L, Swin-B, PVTv2-B5, SatFormer-B and RETFound, the proposed method has achieved significant improvements, at least 1.5% increase in Kappa on the EyePACS dataset and 2.0% increase in F1 score on RFMiD. This confirms that the model provided by the present invention can capture information that large-scale transformers may ignore, especially the spatial distribution pattern of lesion areas, which indicates that the proposed image-patches can further improve model performance.

[0076] Based on the above-mentioned eye lesion diagnosis system and method, accurate retinal disease diagnosis results can be obtained, and the scattered pathological features in fundus images and their global distribution patterns can be fully understood, overcoming the shortcomings of previous models that are limited to the most significant lesion features, enhancing pathological semantics from the perspective of the global distribution pattern of eye lesions, and significantly improving the accuracy of eye lesion recognition. It is possible to perform intelligent diagnosis on the types of eye lesions shown in fundus images, and at the same time achieve higher accuracy through continuous training to achieve accurate and effective retinal disease classification.

[0077] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An eye disease diagnosis system based on the spatial structural relationship of lesions, characterized in that: The eye disease diagnosis system includes a lesion significance enhancement module and a lesion visual distribution information fusion module; The lesion significance enhancement module is used to divide the input fundus image into a number of image blocks, mine and extract lesion features for each image block, and generate lesion significance features of the image block; The lesion visual distribution information fusion module is used to identify image blocks containing eye lesions based on significant features, identify and obtain the features of eye lesion pixels through image blocks containing eye lesions, generate pixel lesion nodes according to the features of eye lesion pixels, establish edges based on pixel lesion nodes to regenerate an eye lesion map, and introduce a hybrid pooling operation to construct a global distribution pattern of the eye lesion area.

2. The eye disease diagnosis system according to claim 1, characterized in that: The mining and extraction of lesion features for each image block includes: each image block undergoes four progressive stages of feature representation, each stage includes a number of Transformer blocks, and the feature representation of the last stage is processed by a multi-layer perceptron to determine whether the image block contains classification information of eye lesions, thereby completing the extraction of lesion features.

3. The eye disease diagnosis system according to claim 1, characterized in that: The method of identifying image blocks containing eye lesions based on significant features includes: based on the significant features, using a significant encoder to generate block embeddings with significant feature information, inputting the block embeddings into a discriminative multi-layer perceptron, and obtaining a plurality of block embeddings P containing eye lesions and image blocks containing eye lesions.

4. The eye disease diagnosis system according to claim 1, characterized in that: The method of identifying and obtaining the features of eye lesion pixels based on the image block containing the eye lesion includes: embedding P based on the i-th block i and the jth pixel C in the i-th block embedding ij , pixel C ij Laplace features and block embedding P i Add and calculate the fused features l ij , and then generate the feature P of the eye lesion pixel through weighted summation ij .

5. The eye disease diagnosis system according to claim 1, characterized in that: The generating of pixel lesion nodes according to the characteristics of eye lesion pixels includes: based on the characteristics of eye lesion pixels, using a threshold to make a judgment, setting pixels exceeding the threshold as pixel lesion nodes, and obtaining pixel node characteristics according to the pixel lesion nodes.

6. The eye disease diagnosis system according to claim 1, characterized in that: The method of establishing edges based on pixel lesion nodes and regenerating an eye lesion map includes: a new graph construction algorithm based on the minimum Manhattan distance in four quadrants, by dividing the plane associated with each pixel lesion node into four quadrants, calculating the distance between a given pixel lesion node and the nearest pixel lesion node in each quadrant through Manhattan distance, obtaining the edges of the eye lesion map, calculating the weights of the edges according to the Manhattan distance, and generating the eye lesion map.

7. The eye disease diagnosis system according to claim 1, characterized in that: The introduction of the hybrid pooling operation to construct the global distribution pattern of the eye lesion area includes: introducing edge cutting pooling and edge folding pooling, and deriving and constructing the global distribution pattern of the eye lesion area based on the information transmission of edge cutting pooling and edge folding pooling.

8. The eye disease diagnosis system according to claim 7, characterized in that: The edge cutting pooling includes: calculating the attention weight of each edge based on the eye lesion map and the pixel node features, deriving new pixel node features by aggregating adjacent pixel nodes based on the attention weights, obtaining potential eye lesion cluster nodes through edge cutting pooling based on the eye lesion map and the new pixel node features, and constructing a new lesion cluster map; The edge folding pooling includes: based on the graph attention mechanism, taking the eye lesion cluster and the node features of the eye lesion cluster as input, calculating the downsampled eye lesion cluster and the node features of the eye lesion cluster through edge folding pooling, and introducing auxiliary loss to standardize the difference between the eye lesion cluster and the downsampled eye lesion cluster.

9. A method for diagnosing eye diseases based on the spatial structural relationship of lesions, characterized in that: The eye disease diagnosis method uses the eye disease diagnosis system according to any one of claims 1 to 8, comprising the following steps: Collect several fundus images and perform image preprocessing, normalize the resolution of all images and convert the format, and then perform data augmentation to obtain the training data set; Inputting the training data set into the eye disease diagnosis system for training, adjusting parameters during the training process, optimizing the parameters of the eye disease diagnosis system, and completing the training; Collect several fundus images and send them to the trained eye disease diagnosis system to verify the eye disease recognition accuracy of the eye disease diagnosis system.

10. The method for diagnosing eye diseases according to claim 9, characterized in that: The data augmentation includes: rotation, scaling, Gaussian blur, color jitter and mirroring operations.

Images