An eye disease diagnosis system and method based on lesion spatial structure relationship

Abstract

The application discloses an eye disease diagnosis system and method based on lesion spatial structure relationship, and belongs to the technical field of intelligent eye disease diagnosis. The eye disease diagnosis system comprises a lesion saliency enhancement module and a lesion visual distribution information fusion module. Saliency features related to lesions are extracted from fundus images, and the saliency features are further refined to obtain features of eye lesion pixels, an eye lesion image is constructed, and finally, a global distribution mode of a lesion area is obtained through a pooling operation to complete eye disease diagnosis. The eye disease diagnosis system and method provided by the application can provide effective information for the recognition and diagnosis of different types of eye diseases, obtain accurate retinal disease diagnosis results, comprehensively understand scattered pathological features in fundus images and their global distribution mode, accelerate the calculation speed of the network, and reduce the calculation cost.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent eye disease diagnosis, and particularly relates to an eye disease diagnosis system and method based on lesion spatial structure relationship. BACKGROUND

[0002] Retinal diseases are of various types, have a high incidence rate, and cover a wide age range, and are major public health problems. Common retinal diseases include diabetic retinopathy, glaucoma, age-related macular degeneration, retinal vein occlusion, etc., and 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 crucial role in slowing down the disease progression and reducing the risk of irreversible visual impairment and permanent blindness.

[0003] However, the ophthalmic medical talent resources in China are relatively scarce and extremely uneven in regional distribution, so it is difficult to achieve large-scale ophthalmic disease screening, resulting in many patients missing the best treatment opportunity. At the same time, the diagnosis of retinal diseases highly depends on fundus medical image analysis. However, these pathological regions in fundus images are usually small in size and scattered throughout the retina, greatly increasing the difficulty of diagnosis. The research and development of retinal disease auxiliary diagnosis and treatment tools combined with deep learning can greatly reduce the workload of clinicians, improve the diagnosis efficiency, help achieve early detection and early intervention of retinal diseases, minimize the harm of diseases as much as possible, and realize high-accuracy automated identification of retinal diseases based on fundus images, which has a high application prospect 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 the publication number CN 118380157 A discloses a training method of a model for assisting eye disease diagnosis and related devices thereof, which obtains real eye images of multiple eye disease patients and eye disease annotations corresponding to the real eye images; extracts general features in all real eye images by using an ophthalmic basic model; for each eye disease patient, all general features belonging to the eye disease patient are spliced to obtain multi-modal fusion eye data; and the model for assisting eye disease diagnosis is trained by taking the multi-modal fusion eye data as training samples and the eye disease annotations as labels, thereby solving the problems of single eye disease diagnosis and poor robustness of the original diagnosis model. However, this method leads to relatively poor results in classifying retinal diseases, because it usually ignores some small and subtle pathological regions.

[0005] Later, models that extract features from subtle lesion areas for diagnosis are developed, and the key of these methods is to use significant features about eye lesions to match the distribution of lesions, thereby guiding the attention of the diagnosis model to focus on these subtle lesion areas, and thus improving the diagnostic accuracy. Although these methods have proven effective, 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 certain retinal diseases tend to occur in specific areas of the retina, showing unique patterns of lesion distribution. For example, in diabetic macular edema, hard exudates usually appear in isolated strips or clusters, or around large, ring-distributed microaneurysms. Therefore, a comprehensive understanding of the specific distribution patterns of lesion areas related to retinal diseases is crucial to achieving accurate diagnosis.

[0007] In order to fully exploit the spatial distribution patterns of subtle lesions in fundus images and improve the classification performance of retinal diseases, a graph convolution network (GCN) is used to represent the mutual connection relationship between various lesion areas in retinal lesions, and a comprehensive understanding of the recognition features of retinal diseases is obtained from the distribution patterns of lesion areas. SUMMARY

[0008] The purpose of the present application is to provide an eye disease diagnosis system and method based on lesion spatial structure relationship, which extracts significant features related to lesions from fundus images, further refines the significant features to obtain features of eye lesion pixels, constructs an eye lesion graph, 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 application, an eye disease diagnosis system based on lesion spatial structure relationship is provided, which comprises a lesion saliency enhancement module and a lesion visual distribution information fusion module.

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

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

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

[0013] In one embodiment, identifying image blocks containing ocular lesions based on saliency features includes: generating block embeddings with saliency feature information using a saliency encoder based on saliency features, inputting the block embeddings into a discriminative multilayer perceptron, and obtaining a plurality of block embeddings P containing ocular lesions and image blocks containing ocular lesions.

[0014] In one embodiment, the feature of identifying and obtaining pixels of eye lesions based on image blocks containing eye lesions includes: embedding P based on the i-th block. i And the j-th pixel C in the i-th block embedding ij , will pixel C ij Laplacian features and block embedding P i Addition calculation of the fused features l ij Then, the feature P of the pixels of the eye lesion is generated by weighted summation. ij .

[0015] In one embodiment, generating pixel lesion nodes based on the features of the pixels with eye lesions includes: determining the pixels with eye lesions using a threshold based on the features of the pixels with eye lesions, setting pixels exceeding the threshold as pixel lesion nodes, and obtaining pixel node features based on the pixel lesion nodes.

[0016] In one embodiment, the method of generating an eye lesion map by establishing edges based on pixel lesion nodes includes: a new graph construction algorithm based on the minimum Manhattan distance in four quadrants, which divides the plane associated with each pixel lesion node into four quadrants, calculates the distance between a given pixel lesion node and the nearest pixel lesion node in each quadrant using the Manhattan distance, obtains the edges of the eye lesion map, calculates the weights of the edges based on the Manhattan distance, and generates the eye lesion map.

[0017] In one instance, the introduction of hybrid pooling operations to construct a global distribution pattern of ocular lesion regions includes: introducing edge-cutting pooling and edge-folding pooling, and deriving a global distribution pattern of ocular lesion regions based on the information transmission of edge-cutting pooling and edge-folding pooling.

[0018] In one example, the edge-cut pooling includes: calculating an attention weight of each edge based on the eye lesion map and the pixel node feature, deriving a new pixel node feature by aggregating adjacent pixel nodes based on the attention weight, obtaining a potential eye lesion cluster node by edge-cut pooling based on the eye lesion map and the new pixel node feature, and constructing a new lesion cluster graph.

[0019] The edge folding pooling includes: based on the graph attention mechanism, taking the eye lesion cluster and the node feature of the eye lesion cluster as input, calculating the down-sampled eye lesion cluster and the node feature of the eye lesion cluster by edge folding pooling, and introducing auxiliary loss to regulate the difference between the eye lesion cluster and the down-sampled eye lesion cluster.

[0020] The application also provides an eye disease diagnosis method based on the spatial structure relationship of lesions, which uses the eye disease diagnosis system and includes the following steps:

[0021] A plurality of fundus images are collected, image preprocessing is performed, the resolution of all images is normalized and format conversion is performed, and then data augmentation is performed to obtain a training data set;

[0022] The training data set is input into the eye disease diagnosis system for training, parameter adjustment is performed during the training process, the parameters of the eye disease diagnosis system are optimized, and the training is completed;

[0023] A plurality of fundus images are collected and input into 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 jittering and mirror operation.

[0025] Compared with the prior art, the application has at least the following beneficial effects:

[0026] (1) By constructing the eye disease diagnosis system based on the lesion saliency enhancement module and the lesion visual distribution information fusion module and iteratively training, the dispersed pathological features and global distribution patterns in the fundus image can be comprehensively understood, the shortcomings of the prior model limited to the most significant lesion features are overcome, the pathological semantics is enhanced from the perspective of lesion global distribution pattern, the accuracy of retinal disease classification is improved, and the accuracy of eye disease judgment is not less than the judgment level of experienced doctors.

[0027] (2) The mixed pooling operation is introduced in the lesion visual distribution information fusion module, which can fully utilize the dependency relationship between different lesion areas and analyze the spatial distribution pattern on the image, and significantly reduce the complexity of calculation. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description.

[0029] Figure 1 The structural schematic diagram of the eye disease diagnosis system provided by the present application.

[0030] Figure 2 The structural schematic diagram of the potential eye lesion pixel recognition based on Laplacian operator provided by the present application.

[0031] Figure 3 The structural schematic diagram of the eye lesion map provided by the present application.

[0032] Figure 4 The structural schematic diagram of the mixed pooling operation provided by the present application.

[0033] Figure 5 The flow chart of the eye disease diagnosis method provided by the present application. DETAILED DESCRIPTION

[0034] In order to make the objectives, technical solutions and advantages of the present application clearer, the following will further describe the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and do not limit the protection scope of the present application.

[0035] In order to diagnose and classify various retinal diseases, including age-related macular degeneration, high myopia, glaucoma, venous obstruction image, diabetic retinopathy and normal fundus image, to improve the efficient intelligent diagnosis of doctors, the embodiments provide an eye disease diagnosis system based on lesion spatial structure relationship, as shown in Figure 1 The eye disease diagnosis system includes a lesion saliency 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 a plurality of image blocks, and to mine and extract lesion features of each image block to generate lesion saliency features of the image blocks.

[0037] In the embodiments, a specific fundus image X e R h×w×3 is given, h x w x 3 represents that the input is a three-channel image of h x w, 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 different size convolution kernels (3 x 3, 5 x 5) to sample the image blocks of each stage with a step of 2, to obtain a plurality of image blocks.

[0038] Feature extraction and mining are performed on each image patch. Specifically, the feature extraction layer is a hierarchical structure with four stages, consisting of four progressive stages. The image patch sequentially undergoes feature representation in each of the four stages, and each stage contains several Transformer blocks. The feature information from the final stage is processed by a Multi-layer Perception (MLP) to determine whether the image patch contains classification information for ocular lesions, thus completing the extraction of lesion features. In the final stage, a saliency encoder is used to generate block embeddings with saliency information. These block embeddings are then input into a discriminative MLP to obtain the block embeddings P∈R of n potential lesion patches. 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 ocular lesions based on salient features. It identifies and obtains the features of the pixels of ocular lesions by identifying the image blocks containing ocular lesions, generates pixel lesion nodes based on the features of the pixels of ocular lesions, establishes edges based on the pixel lesion nodes and then generates the ocular lesion map, and introduces hybrid pooling operation to construct the global distribution pattern of the ocular lesion region.

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

[0041]

[0042] Among them, Laplacian (C ij , o) represents the Laplacian operation of the o-th order neighbor. Describe a linear transformation that makes l ij Dimensions and P i Alignment (from 3 to d). Then, pixel features P are obtained through a weighted summation operation. ij :

[0043] P ij =l ij+ s • l ij

[0044] s = Softmax(GELU(g(l ij )))

[0045] where s denotes the saliency value of pixel C ij , GELU denotes the activation function, and g denotes a linear transformation to map the dimension of l ij to 1. By applying these operations to all pixels, the pixel-level saliency features in the image patches of potential eye lesions are obtained.

[0046] In the embodiment, an additional threshold value e is used to determine the pixel saliency value s, and if the pixel saliency value exceeds the threshold value e, each pixel is regarded as an initial pixel lesion node of the eye lesion map. Based on this way, 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, a new graph construction algorithm based on the minimum Manhattan distance in four quadrants is used in the embodiment. As shown in Figure 3 , the connectivity of each pixel lesion node is redefined by dividing 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 map are obtained. In addition, for any edge (v i , v j ) E, the edge weight is calculated using the Manhattan distance in the following way:

[0048]

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

[0050] Subsequently, based on the eye 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 passing 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 shortcomings of a single pooling operation. The components of the hybrid pooling operation are as follows: Figure 4Edge-cut pooling and edge-fold pooling are shown.

[0052] Edge-cut pooling component: In the edge-cut pooling component, a three-step procedure is followed to reduce the pixel lesion nodes to lesion cluster nodes. First, given a lesion map G = (V, E; W) and pixel node features F, the attention weight A i , v j ) e E is computed. The attention weight A ij is computed as follows:

[0053]

[0054] where F i and F j are the pixel node features of v i and v j , respectively, adj(i) denotes the adjacent nodes of v i , and σ denotes a linear transformation that changes the dimension of the pixel node feature F from d to , is a concatenation operation, and LeakyReLu is an activation function.

[0055] After that, a new node feature F' is derived by aggregating the adjacent nodes based on the attention weight A. For any pixel lesion node v i , the node feature F' i is computed as follows:

[0056]

[0057] where δ is a linear transformation that maps F to another space without changing its dimension d. Then we use a cut-pool operation on the boundary of the lesion area to generate the feature of the lesion cluster node by:

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

[0059] where C denotes the feature of the new ocular lesion cluster node, G' = (V', E') is the new lesion cluster map with lesion cluster nodes V' and edges E', and L cut is an auxiliary loss that aims to maximize the probability of each lesion pixel being 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 ocular lesion map to a manageable range.

[0060] Edge-fold pooling component: The edge-fold component includes t pairs of graph attention mechanisms and edge-fold pooling layers. In the i-th pair, the lesion cluster map G' i and the lesion cluster node features As input, and by edge folding pooling operation to calculate down-sampling lesion cluster graph G' i+1 And the corresponding lesion cluster feature Specific 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] Where GAT is a graph attention mechanism, D i is the same as d i dimensional aggregated cluster feature; r is the edge folding ratio, representing the percentage of reduced edges; L diff is an auxiliary loss, used to normalize the difference between 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 layer, d i+1 =d i , d i+1 =2d i .

[0064] After the last edge folding pooling layer, the lesion distribution pattern feature C t is obtained In addition, the average value of C t is calculated 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 features of the lesion graph, thereby avoiding the limitations of a single pooling operation.

[0065] As Figure 5 shown, based on the structure of the above eye disease diagnosis system based on the spatial structure relationship of the lesion, the present application also provides an eye disease diagnosis method based on the spatial structure relationship of the lesion, comprising the steps of:

[0066] Collecting several fundus images, and performing image preprocessing, normalizing the resolution of all pictures and performing format conversion, and then performing data augmentation to obtain a training data set;

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

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

[0069] To better illustrate the effectiveness of the system and method provided by the present application, 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 is used for preliminary adjustment for 20 epochs, followed by 500 epochs of training using a cosine decay learning rate scheduler and AdamW optimizer; the cross-entropy loss function is used to measure the difference between the model prediction value and the true value. The batch size is 128, the initial learning rate is 0.001, the weight decay is 0.05, and the gradient clipping with a maximum norm of 1.0 is 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 mirror operation are used during the training process. The experimental results are shown in Table 1. In the embodiment, the latest technical methods are compared, and the eye disease diagnosis method based on the spatial structure relationship of the lesion (Saliency Image-Graph, SIGraph) in the embodiment significantly outperforms these advanced architectures in performance.

[0071] Table 1

[0072]

[0073] For all the compared models, their original hyperparameter settings are maintained in the embodiment, while the training period and data division are standardized for fair comparison with SIGraph of the present application, especially when the original method lacks performance indicators for specific datasets. 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 application has achieved significant performance improvement, with a Kappa index on the EyePACS dataset at least 3.1% higher.

[0075] ​In comparison with the Transformer-based methods, including Vit-B / 16, CrossFormer-L, Swin-B, PVTv2-B5, SatFormer-B and RETFound, the method of the present application makes significant progress, with an increase of at least 1.5% in Kappa on the EyePACS dataset and an increase of 2.0% in F1 score on the RFMiD. This confirms that the model provided by the present application can capture information that may be overlooked by large-scale transformers, especially the spatial distribution patterns of lesion areas, which indicates that the proposed image-tile can further improve the performance of the model.

[0076] Based on the above eye lesion diagnosis system and method, an accurate retinal disease diagnosis result can be obtained, and the dispersed pathological features and their global distribution patterns in the fundus image can be comprehensively understood. The previous model is limited to the most significant lesion features, and the pathological semantics is enhanced from the perspective of the global distribution pattern of eye lesions, which significantly improves the accuracy of eye lesion recognition. The type of eye lesion displayed in the fundus image can be intelligently diagnosed, and the accuracy can be further improved through continuous training, so as to realize accurate and effective retinal disease classification.

[0077] The specific embodiments described above have described the technical solutions and beneficial effects of the present application in detail. It should be understood that the above description is only the most preferred embodiment of the present application and is not intended to limit the present application. Any modification, supplement and equivalent replacement within the principle range of the present application should be included in the protection scope of the present application.

Claims

1. A diagnostic system for ocular diseases based on the spatial structural relationship of lesions, characterized in that, The eye disease diagnostic system includes a lesion saliency enhancement module and a lesion visual distribution information fusion module; The lesion saliency enhancement module is used to divide the input fundus image into several image blocks, mine and extract lesion features for each image block, and generate lesion saliency features for the image block. The lesion visual distribution information fusion module is used to identify image patches containing ocular lesions based on salient features. It identifies and obtains the features of ocular lesion pixels from these image patches, generates pixel lesion nodes based on these features, establishes edges based on these pixel lesion nodes to regenerate the ocular lesion map, and introduces a hybrid pooling operation to construct a global distribution pattern of the ocular lesion region. Specifically, this includes: based on the first... Block embedding and the The first block embedding in the _th block 1 pixel , will pixels Laplacian features and block embeddings Addition calculation of the fused features Then, the features of the pixels of the eye lesion are generated by weighted summation. Based on the characteristics of pixels with eye lesions, a threshold is used for judgment. Pixels exceeding the threshold are set as pixel lesion nodes. The pixel node characteristics are obtained based on the pixel lesion nodes. A new graph construction algorithm based on the minimum Manhattan distance in the four quadrants is used to divide the plane associated with each pixel lesion node into four quadrants. The distance between a given pixel lesion node and the nearest pixel lesion node in each quadrant is calculated using the Manhattan distance to obtain the edges of the eye lesion graph. The weight of the edges is calculated based on the Manhattan distance to generate the eye lesion graph.

2. The eye disease diagnostic system according to claim 1, characterized in that, The aforementioned mining and extraction of lesion features for each image block includes: each image block undergoes feature representation through four progressive stages, each stage containing several Transformer blocks, and the feature representation of the last stage is processed by a multilayer perceptron to determine whether the image block contains classification information of eye lesions, thus completing the extraction of lesion features.

3. The eye disease diagnostic system according to claim 1, characterized in that, The method for identifying image patches containing ocular lesions based on saliency features includes: generating block embeddings with saliency feature information using a saliency encoder based on saliency features; inputting the block embeddings into a discriminative multilayer perceptron to obtain several block embeddings containing ocular lesions. And image blocks containing eye lesions.

4. The eye disease diagnostic system according to claim 1, characterized in that, The introduction of hybrid pooling operations to construct a global distribution pattern of ocular lesion regions includes: introducing edge-cutting pooling and edge-folding pooling, and deriving and constructing a global distribution pattern of ocular lesion regions based on the information transmission of edge-cutting pooling and edge-folding pooling.

5. The eye disease diagnostic system according to claim 4, characterized in that, The edge-cutting pooling includes: calculating the attention weight of each edge based on the ocular lesion map and pixel node features; deriving new pixel node features by aggregating adjacent pixel nodes based on the attention weights; and obtaining potential ocular lesion cluster nodes through edge-cutting pooling based on the ocular lesion map and the new pixel node features to construct a new lesion cluster map. The edge folding pooling method includes: based on the graph attention mechanism, taking the ocular lesion cluster and its node features as input, calculating the downsampled ocular lesion cluster and its node features through edge folding pooling, and introducing an auxiliary loss to normalize the difference between the ocular lesion cluster and the downsampled ocular lesion cluster.

6. A diagnostic method for ocular diseases based on the spatial structural relationship of lesions, characterized in that, The method for diagnosing eye diseases using the eye disease diagnostic system according to any one of claims 1 to 5 includes the following steps: Several fundus images were acquired and preprocessed. The resolution of all images was normalized and the format was converted. Then, data augmentation was performed to obtain the training dataset. The training dataset is input into the eye disease diagnosis system for training. During the training process, the parameters are adjusted and optimized to complete the training. Several fundus images were collected and fed into a trained eye disease diagnosis system to verify the accuracy of the system in identifying eye diseases.

7. The method for diagnosing eye diseases according to claim 6, characterized in that, The data augmentation includes rotation, scaling, Gaussian blur, color dithering, and mirroring operations.

Images