Fundus OCT image data classification method with key area identification

Abstract

The invention relates to a fundus OCT (Optical Coherence Tomography) image data classification method with key area identification. According to the method, on the basis of a deep convolutional neural network model VGG, a retina image obtained by optical coherence tomography (OCT) is refined through an improved bilinear attention model, and rough classification and class activation regions of the image are predicted. Besides, the invention provides a new auxiliary network model which is used for predicting the region of interest of the fundus OCT image, so that global coarse classification and local independent regions are combined to obtain a more accurate final classification result.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, and in particular relates to a fundus OCT image data classification method with key area recognition. Background Art

[0002] Ophthalmic diseases are complex and diverse, and their clinical diagnosis often requires ophthalmologists to make judgments based on a large number of imaging examinations. In the process of clinical diagnosis and treatment, low image quality and relatively insufficient physician experience can lead to misjudgment of imaging data, failure to diagnose and treat the disease early, worsening the disease, and even causing blindness. At present, the diagnosis of retinal fundus disease images is mainly done manually, but manual analysis by doctors alone is inefficient and is affected by subjective judgment.

[0003] With the development of computer technology, many scholars and institutions have used computer vision technology to process retinal fundus images to achieve the purpose of classification and recognition. Some existing technical solutions use classic machine learning algorithms, such as SVM algorithm and K-nearest neighbor algorithm to classify fundus images, while others use convolutional neural networks to classify them, but all of them have the problem of low classification accuracy.

[0004] The disadvantage of the existing technology is that it only classifies and identifies the retinal fundus image as a whole, and cannot accurately extract the features of the retinal fundus image; at the same time, the grading method of fundus nerve tortuosity relies on artificially designed filters, which is not suitable for a large number of nerve tortuosity with different morphologies, resulting in low classification accuracy. Summary of the invention

[0005] The purpose of the present invention is to solve the above problems and provide a method for classifying fundus OCT image data with key area recognition. Based on the deep convolutional neural network model VGG, the method refines the retinal image obtained by optical coherence tomography (OCT) through an improved bilinear attention model to predict the coarse classification and class activation area of ​​the image. In addition, the present invention proposes a new auxiliary network model for predicting the region of interest of fundus OCT images, so as to combine the global coarse classification and the local independent area to obtain a more accurate final classification result.

[0006] To achieve the above purpose, the technical solution of the present invention is: a fundus OCT image data classification method with key area recognition, a fundus retinal OCT image classification and recognition network is proposed, which includes two parts:

[0007] The upper part is the key area recognition network, which aims to identify possible lesion areas to achieve auxiliary classification and recognition of fundus retinal OCT images;

[0008] The lower part is the subject recognition network, which uses the pre-trained VGG16 deep convolutional neural network model as the backbone and replaces the fully connected layer with a bilinear attention model;

[0009] The classification and recognition results of the upper and lower parts are then passed through a fully connected layer to obtain accurate fundus disease classification information.

[0010] In one embodiment of the present invention, the bilinear attention model is a bilinear attention model that integrates key areas, which achieves the purpose of weakening irrelevant characteristics and enhancing the relevance of key area features, and is specifically implemented as follows:

[0011] The output characteristics of the convolutional block C5 of the VGG16 deep convolutional neural network model are denoted as M, and the output characteristics of the convolutional block V3 of the key area recognition network are denoted as N, where M, N∈R C×W×H , C, W, and H represent the number of channels, width, and height, respectively;

[0012] (1) Channel conversion

[0013] First, reconstruct M and N as M, N∈R C×K , where K = W × H; then, transpose N to obtain N∈R K×C , find the matrix outer product of M and N to obtain the bilinear features of the fundus retinal OCT image:

[0014] T=M×N

[0015] Where T∈R C×C Represents the correlated bilinear features of the channels of fundus retinal OCT images;

[0016] (2) Numerical processing

[0017] The obtained related bilinear features are reshaped and normalized using the following function:

[0018]

[0019] Among them, T θ ∈R C×C It is the enhanced bilinear feature;

[0020] T θ The data is input into the fully connected layer, and after the operation of three fully connected layers, accurate classification information of fundus diseases is obtained.

[0021] In one embodiment of the present invention, the key area identification network is specifically implemented as follows:

[0022] (1) Generate key areas

[0023] The class activation map in the convolution block C5 of the VGG16 deep convolutional neural network model is used to identify the most likely lesion area, i.e., the key area, of different diseases in each fundus retinal OCT image;

[0024] Threshold processing is performed on the heat map of the key area to obtain a binary class activation mask, and a segmentation map of the key area is obtained, and scale transformation is performed to make it consistent with the input fundus retinal OCT image;

[0025] (2) Identification of key areas

[0026] The generated key area segmentation map is element-wise multiplied with the input fundus retinal OCT image and then scaled. The output feature is 56*56*64, and then input into the convolution blocks V1, V2, and V3 in the key area recognition network for processing. The output feature is 7*7*512, and then used as one of the inputs of the bilinear attention model.

[0027] In one embodiment of the present invention, in (1), the threshold is set to S=0.62.

[0028] In one embodiment of the present invention, the key area recognition network adopts the structure of a VGG16 deep convolutional neural network model, and the VGG16 deep convolutional neural network model adopted by the key area recognition network is independent of the VGG16 deep convolutional neural network model adopted by the subject recognition network.

[0029] In one embodiment of the present invention, the method further proposes an expert system to further identify the accurate fundus disease classification information output by the fundus retinal OCT image classification recognition network.

[0030] In one embodiment of the present invention, the expert system is constructed using a neutronomic set.

[0031] In one embodiment of the present invention, the expert system construction process is specifically as follows:

[0032] (1) Identify possible fundus diseases, including age-related macular degeneration (AMD) 1 ), choroidal neovascularization CNV (α 2 ), diabetic macular edema DME (α 3 ), fundus normal without disease NOR (α 4 );

[0033] (2) Determine the attributes of fundus diseases: To evaluate the possible types of fundus diseases, select the common attributes of fundus diseases, including vascular configuration (ψ 1 ), vascular tortuosity (ψ 2 ), reflection strips (ψ 3 ), exudate (ψ4 ), reflection structure (ψ 5 ), blood flow signal (ψ 6 ) and retinal thickness (ψ 7 );

[0034] (3) Determine the weights of fundus disease attributes;

[0035] (4) Experts evaluate fundus diseases based on their attributes and provide evaluation indicators including very good, good, poor, and very poor;

[0036] (5) Generate evaluation matrix: Use the single-valued neutral intelligence number SVNS to quantify the evaluation indicators given by experts. The formula of the single-valued neutral intelligence number is as follows:

[0037] δ={ <x,p δ (x),l δ (x),q δ (x)>:x∈X}

[0038] Among them, p δ (x), l δ (x) and q δ (x) represent true membership, uncertain membership and false membership respectively, and satisfy p δ (x),l δ (x),q δ (x)∈[0,1] and 0≤p δ (x)+l δ (x)+q δ (x)≤3;

[0039] (6) Evaluation information standardization: Assume X = [δ] m×n =<T,I,F> m×n Is an original evaluation index matrix, and the evaluation information standardization is processed as follows:

[0040]

[0041] Among them, T, I, F represent the true membership, uncertain membership and false membership in the original evaluation index matrix, p, l, q represent the true membership, uncertain membership and false membership after standardization;

[0042] (7) Calculate the aggregate value of each fundus disease attribute: A Maclaurin symmetric mean operator (STTMSM) based on the single-valued neutron set Schweizer–Sklar algorithm is proposed to calculate the aggregate value of each alternative solution α i The evaluation vector The aggregate value of is calculated as follows:

[0043]

[0044] in i 1 ,i 2 ,…,i k are k integers in the set of integers {1,2,…,n}, where 1≤i 1 <i 2 <…<i k ≤n, r is the parameter of the operator, yes The weight of is the binomial coefficient;

[0045] (8) Sorting alternative solutions: Compare the aggregate values ​​obtained in (7) and sort the results from large to small. The higher the aggregate value, the greater the possibility that the patient has the corresponding fundus disease. The aggregate value comparison function is implemented by the following formula:

[0046]

[0047] The present invention also provides a fundus OCT image data classification system with key area recognition, including a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0048] The present invention also provides a computer-readable storage medium, on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] 1. The present invention proposes a new network architecture to automatically learn tissue features in fundus retinal OCT images. It does not need to rely on manually designed filters like the existing technical solutions, and can adaptively learn pathological tissue-related features.

[0051] 2. Compared with the existing solutions, the present invention can better characterize the pathological characteristics of relevant tissues in fundus retinal OCT images and obtain the second-order statistics of the pathological characteristics.

[0052] 3. Compared with the existing solutions, the present invention obtains the local tissue images with important significance in the fundus retinal OCT images, greatly expands the scope of mining pathological features, and improves the classification performance from the perspective of interpretability. It enhances the interpretability of the classification model and improves the classification performance.

[0053] 4. The same data set is classified using different schemes. The results are shown in Table 1. The scheme proposed in the present invention has obvious advantages in all aspects.

[0054] Table 1

[0055] method Precision Sensitivity F1-score Accuracy KNN 0.725 0.725 0.725 0.730 SVN 0.795 0.800 0.790 0.790 3D-CNN 0.885 0.885 0.890 0.890 The present invention 0.925 0.920 0.915 0.920 BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a system block diagram of the present invention.

[0057] Figure 2 This is a system architecture diagram of the present invention.

[0058] Figure 3 This is an OCT image of DME.

[0059] Figure 4 It is the key area of ​​the fundus retinal OCT image.

[0060] Figure 5 Heat map of key areas.

[0061] Figure 6 Identify network architectures for key areas.

[0062] Figure 7 This is the flow chart of the expert system. DETAILED DESCRIPTION

[0063] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0064] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0065] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0066] The present invention provides a fundus OCT image data classification method with key area recognition, and proposes a fundus retinal OCT image classification and recognition network, which includes two parts:

[0067] The upper part is the key area recognition network, which aims to identify possible lesion areas to achieve auxiliary classification and recognition of fundus retinal OCT images;

[0068] The lower part is the subject recognition network, which uses the pre-trained VGG16 deep convolutional neural network model as the backbone and replaces the fully connected layer with a bilinear attention model;

[0069] The classification and recognition results of the upper and lower parts are then passed through a fully connected layer to obtain accurate fundus disease classification information.

[0070] The following is the specific implementation process of the present invention.

[0071] With the substantial growth of productivity in my country, daily necessities have been greatly enriched, and the number of diabetic patients has also increased. Diabetes brings great pain and inconvenience to patients, and its complications also bring great harm. Among them, age-related macular degeneration (AMD), choroidal neovascularization (CNV) and diabetic macular edema (DME) can damage patients' vision, and even cause blindness in severe cases. At present, the diagnosis of retinal fundus diseases still mainly relies on the doctor's manual judgment. For hospitals with many patients, it often causes great work pressure for doctors, and there may also be misdiagnosis and missed diagnosis.

[0072] The present invention discloses a method for classifying fundus OCT image data with key area recognition, which applies convolutional neural networks to the classification and recognition of retinal fundus OCT images, assists doctors in diagnosis, reduces workload, and reduces the number of misdiagnoses and missed diagnoses. Existing classification methods based on convolutional neural network models have achieved good results in various industries, but require a large amount of data for training to achieve better results, otherwise there will be a risk of overfitting. However, due to medical ethics and patient privacy factors, it is impossible to obtain a large amount of medical data. Therefore, the present invention uses a pre-trained VGG16 as the backbone to speed up the training convergence and improve learning efficiency; then, a bilinear attention module and an auxiliary network are used to improve the accuracy of classification. Its architecture is as follows: Figure 1 shown.

[0073] The fundus retinal OCT image classification and recognition network proposed in the present invention consists of two parts. The upper part is the key area recognition network, and its main purpose is to identify possible lesion areas to achieve auxiliary classification and recognition of OCT images; the lower part is the main recognition network, which uses the pre-trained VGG16 deep convolutional neural network model as the backbone, and its fully connected layer is transformed and replaced with a bilinear attention model; after the classification and recognition results of the upper and lower networks pass through a fully connected layer, more accurate fundus disease classification information can be obtained.

[0074] 1. Bilinear Attention Network Architecture

[0075] The subject recognition network uses the VGG16 deep convolutional neural network model as the backbone. In order to better extract tissue features in OCT images, the present invention proposes a bilinear attention mechanism, whose architecture is as follows: Figure 2 As shown. We trimmed the classic VGG16 convolutional neural network and replaced the fully connected layer in the VGG16 network with bilinear attention. The final classification recognition in the convolutional neural network is determined by the high-level features extracted from the intermediate hidden layer represented by the number of channels. Therefore, in order to improve the accuracy of fundus disease classification, the present invention designs a bilinear attention model that integrates the focus area to achieve the purpose of weakening irrelevant characteristics and enhancing the correlation of key area features. In order to facilitate the description of the bilinear attention network, the output characteristics of C5 are represented as M, and the output characteristics of V3 are represented as N, where M, N∈R C×W×H , C, W, and H represent the number of channels, width, and height, respectively.

[0076] (1) Channel conversion

[0077] First, reconstruct M and N as M, N∈R C×K , where K = W × H. Then, transpose N to get N∈R K×C , so we have M∈R C×K and N∈R K×C , it can be seen that the transformation methods of M and N channels are inconsistent. Finally, the outer product of the matrix of M and N is calculated to obtain the bilinear features of the fundus retinal image.

[0078] T=M×N

[0079] Where T∈R C×C Correlated bilinear features of channels representing retinal tissue.

[0080] (2) Numerical processing

[0081] In order to enable the fully connected layer to use it for classification recognition, we reshape it and normalize it using the following function.

[0082]

[0083] Among them, T θ ∈R C×C is an enhanced bilinear vector. Then T θ After inputting the fully connected layer and going through three fully connected layers, more accurate classification and recognition data can be output.

[0084] 2. Key Area Identification Network

[0085] AMD, CNV, and DME have high similarities with normal retinal OCT images, which makes them difficult to classify and identify. Therefore, it is necessary to mine the decisive tissue features in OCT images to improve the accuracy of classification and identification. For example, in the OCT images of DME, the reflection signals of the ellipsoidal band and RPE (retinal pigment epithelium) layer will be reduced and discontinuous, and a dot-like high reflection signal can be seen under the fovea, such as Figure 3 By analyzing the pathological characteristics of AMD, CNV, and DME, we decided to mine the decisive tissue features in fundus OCT images as key areas to assist the recognition task of the main convolutional neural network, thereby improving the accuracy of the overall network classification and recognition.

[0086] (1) Generate key areas

[0087] The key areas of the fundus retinal OCT image are as follows: Figure 4 As shown in the figure, from left to right are OCT images of DME, CNV and AMD. Figure 4 It can be seen that the relevant fundus diseases can be more accurately identified through the key areas of the OCT image. The present invention uses the class activation map in the convolution block C5 in the VGG16 convolutional neural network to identify the most likely key areas of different diseases in each OCT image. The obtained image is as follows Figure 5 shown.

[0088] In order to more accurately characterize the tissue features of the fundus retinal OCT image, the obtained heat map is then thresholded to obtain a binary class activation mask. Here, we experimentally found that setting the threshold to S = 0.62 can achieve better classification results. The purpose of thresholding is to obtain a segmentation map of the key area. Then we scale it to make it consistent with the input fundus OCT image.

[0089] (2) Key Area Identification Network

[0090] This network also uses the structure of convolutional neural network, but it is independent of the main VGG16 network. The key area segmentation map generated in the previous step is used as the input of the network after element-wise product calculation with the original retinal OCT image, and is scaled again. Its output feature is 56*56*64. Then it is input into the convolution blocks V1, V2, and V3 for processing. The network structure of the convolution block is as follows: Figure 6 shown.

[0091] After the input image passes through 12 consecutive convolutional layers with a kernel size of 3, its output feature is 7*7*512, and it is used as one of the inputs of the bilinear network.

[0092] 3. Expert System

[0093] Although the recognition of fundus retinal OCT images by the bilinear attention network has achieved a high accuracy, there are still some relatively complex images that require experts to identify. Since experts may have ambiguity and uncertainty when giving opinions on certain aspects, in order to better describe this decision model, the present invention uses the Neutrino set to construct an expert system, and its process is as follows: Figure 7 shown.

[0094] (1) Identify possible fundus diseases, including age-related macular degeneration (AMD) 1 ), choroidal neovascularization CNV (α 2 ), diabetic macular edema DME (α 3 ), fundus normal without disease NOR (α 4 );

[0095] (2) Determine the attributes of fundus diseases: To evaluate the possible types of fundus diseases, select the common attributes of fundus diseases, including vascular configuration (ψ 1 ), vascular tortuosity (ψ 2 ), reflection strips (ψ 3 ), exudate (ψ 4 ), reflection structure (ψ 5 ), blood flow signal (ψ 6 ) and retinal thickness (ψ 7 );

[0096] (3) Determine the weights of fundus disease attributes;

[0097] (4) Experts evaluate fundus diseases based on their attributes and provide evaluation indicators including very good, good, poor, and very poor;

[0098] (5) Generate evaluation matrix: Use the single-valued neutral intelligence number SVNS to quantify the evaluation indicators given by experts. The formula of the single-valued neutral intelligence number is as follows:

[0099] δ={ <x,p δ (x),l δ (x),q δ (x)>:x∈X}

[0100] Among them, p δ (x), l δ (x) and q δ (x) represent true membership, uncertain membership and false membership respectively, and satisfy p δ (x),l δ (x),q δ (x)∈[0,1] and 0≤p δ (x)+l δ (x)+qδ (x)≤3;

[0101] (6) Evaluation information standardization: Assume X = [δ] m×n =<T,I,F> m×n Is an original evaluation index matrix, and the evaluation information standardization is processed as follows:

[0102]

[0103] Among them, T, I, F represent the true membership, uncertain membership and false membership in the original evaluation index matrix, p, l, q represent the true membership, uncertain membership and false membership after standardization;

[0104] (7) Calculate the aggregate value of each fundus disease attribute: A Maclaurin symmetric mean operator (STTMSM) based on the single-valued neutron set Schweizer–Sklar algorithm is proposed to calculate the aggregate value of each alternative solution α i The evaluation vector The aggregate value of is calculated as follows:

[0105]

[0106] in i 1 ,i 2 ,…,i k are k integers in the set of integers {1,2,…,n}, where 1≤i 1 <i 2 <…<i k ≤n, r is the parameter of the operator, yes The weight of is the binomial coefficient;

[0107] (8) Sorting alternative solutions: Compare the aggregate values ​​obtained in (7) and sort the results from large to small. The higher the aggregate value, the greater the possibility that the patient has the corresponding fundus disease. The aggregate value comparison function is implemented by the following formula:

[0108]

[0109] The present invention also provides a fundus OCT image data classification system with key area recognition, including a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0110] The present invention also provides a computer-readable storage medium, on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0111] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0112] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0113] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0114] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0115] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A method for classifying fundus OCT image data with key area recognition, characterized in that: A fundus retinal OCT image classification and recognition network is proposed, which consists of two parts: The upper part is the key area recognition network, which aims to identify possible lesion areas to achieve auxiliary classification and recognition of fundus retinal OCT images; The lower part is the subject recognition network, which uses the pre-trained VGG16 deep convolutional neural network model as the backbone and replaces the fully connected layer with a bilinear attention model; The classification and recognition results of the upper and lower parts are then passed through a fully connected layer to obtain accurate fundus disease classification information.

2. The method for classifying fundus OCT image data with key area recognition according to claim 1, characterized in that: The bilinear attention model is a bilinear attention model that integrates key areas, which can weaken irrelevant characteristics and enhance the relevance of key area features. The specific implementation is as follows: The output characteristics of the convolutional block C5 of the VGG16 deep convolutional neural network model are denoted as M, and the output characteristics of the convolutional block V3 of the key area recognition network are denoted as N, where M, N∈R C×W×H , C, W, and H represent the number of channels, width, and height, respectively; (1) Channel conversion First, reconstruct M and N as M, N∈R C×K , where K = W × H; then, transpose N to obtain N∈R K×C , find the matrix outer product of M and N to obtain the bilinear features of the fundus retinal OCT image: T=M×N Where T∈R C×C Represents the correlated bilinear features of the channels of fundus retinal OCT images; (2) Numerical processing The obtained related bilinear features are reshaped and normalized using the following function: Among them, T θ ∈R C×C It is the enhanced bilinear feature; T θ The data is input into the fully connected layer, and after the operation of three fully connected layers, accurate classification information of fundus diseases is obtained.

3. The method for classifying fundus OCT image data with key area recognition according to claim 1, characterized in that: The key area identification network is specifically implemented as follows: (1) Generate key areas The class activation map in the convolution block C5 of the VGG16 deep convolutional neural network model is used to identify the most likely lesion area, i.e., the key area, of different diseases in each fundus retinal OCT image; Threshold processing is performed on the heat map of the key area to obtain a binary class activation mask, and a segmentation map of the key area is obtained, and scale transformation is performed to make it consistent with the input fundus retinal OCT image; (2) Identification of key areas The generated key area segmentation map is element-wise multiplied with the input fundus retinal OCT image and then scaled. The output feature is 56*56*64, and then input into the convolution blocks V1, V2, and V3 in the key area recognition network for processing. The output feature is 7*7*512, and then used as one of the inputs of the bilinear attention model.

4. The method for classifying fundus OCT image data with key area identification according to claim 3, characterized in that: In (1), the threshold is set to S = 0.

62.

5. A method for classifying fundus OCT image data with key area recognition according to claim 1 or 3, characterized in that: The key area recognition network adopts the structure of the VGG16 deep convolutional neural network model, and the VGG16 deep convolutional neural network model adopted by the key area recognition network is independent of the VGG16 deep convolutional neural network model adopted by the subject recognition network.

6. The method for classifying fundus OCT image data with key area identification according to claim 1, characterized in that: An expert system is also proposed to further identify the accurate fundus disease classification information output by the fundus retinal OCT image classification and recognition network.

7. The method for classifying fundus OCT image data with key area identification according to claim 6, characterized in that: The expert system is constructed using a neutrino set.

8. The method for classifying fundus OCT image data with key area identification according to claim 7, characterized in that: The expert system construction process is as follows: (1) Determine the possible types of fundus diseases, including age-related macular degeneration (AMD) (α1), choroidal neovascularization (CNV) (α2), diabetic macular edema (DME) (α3), and normal fundus (NOR) (α4); (2) Determine the attributes of fundus diseases: To evaluate the possible types of fundus diseases, the common attributes of fundus diseases are selected, including vascular configuration (ψ1), vascular tortuosity (ψ2), reflection stripes (ψ3), exudates (ψ4), reflection structure (ψ5), blood flow signal (ψ6) and retinal thickness (ψ7); (3) Determine the weights of fundus disease attributes; (4) Experts evaluate fundus diseases based on their attributes and provide evaluation indicators including very good, good, poor, and very poor; (5) Generate evaluation matrix: Use the single-valued neutral intelligence number SVNS to quantify the evaluation indicators given by experts. The formula of the single-valued neutral intelligence number is as follows: δ={<x,p δ (x),l δ (x),q δ (x)>:x∈X} Among them, p δ (x), l δ (x) and q δ (x) represent true membership, uncertain membership and false membership respectively, and satisfy p δ (x),l δ (x),q δ (x)∈[0,1] and 0≤p δ (x)+l δ (x)+q δ (x)≤3; (6) Evaluation information standardization: Assume X = [δ] m×n =<T,I,F> m×n Is an original evaluation index matrix, and the evaluation information standardization is processed as follows: Among them, T, I, F represent the true membership, uncertain membership and false membership in the original evaluation index matrix, p, l, q represent the true membership, uncertain membership and false membership after standardization; (7) Calculate the aggregate value of each fundus disease attribute: A Maclaurin symmetric mean operator (STTMSM) based on the single-valued neutron set Schweizer–Sklar algorithm is proposed to calculate the aggregate value of each alternative solution α i The evaluation vector The aggregate value of is calculated as follows: in i1,i2,…,i k are k integers in the set of integers {1,2,…,n}, where 1≤i1<i2<…<i k ≤n, r is the parameter of the operator, yes The weight of is the binomial coefficient; (8) Sorting alternative solutions: Compare the aggregate values ​​obtained in (7) and sort the results from large to small. The higher the aggregate value, the greater the possibility that the patient has the corresponding fundus disease. The aggregate value comparison function is implemented by the following formula:

9. A fundus OCT image data classification system with key area recognition, characterized in that: The method comprises a memory, a processor and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps as claimed in any one of claims 1 to 8 can be implemented.

10. A computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor, and when the processor executes the computer program instructions, the method steps according to any one of claims 1 to 8 can be implemented.

Images