Eye fundus disease diagnosis method and system based on disease domain

Abstract

The invention discloses a disease domain-based fundus disease diagnosis method and system, and the method comprises the steps: obtaining a plurality of fundus images, constructing an image set, and dividing the images in the image set into N disease domains according to a clinical medical diagnosis result; constructing a visual geometric neural network model and a classifier thereof, carrying out multiple rounds of training, and correcting the current visual geometric neural network model and the classifier thereof; and according to the vision geometric neural network model corrected after multiple rounds of training and the classifier thereof, diagnosing a to-be-diagnosed eye fundus image. According to the method, the to-be-diagnosed OCT image is automatically diagnosed by adopting the vision geometric neural network and the classifier which are corrected for multiple times, the diagnosis analysis and treatment efficiency is relatively high, and the diagnosis accuracy of the OCT image is improved by adopting a plurality of groups.

Description

Technical Field

[0001] The present invention relates to the technical field of fundus image diagnosis, and particularly relates to a method and system for diagnosing fundus diseases based on disease domains. Background Art

[0002] Optical coherence tomography (OCT) images play a crucial role in the detection and analysis of retinal diseases. This non-invasive imaging technique provides a high-resolution cross-sectional view of the retina and can be used to evaluate structural changes and abnormalities. In clinical practice, a large number of OCT images are analyzed every day for diagnosing retinal diseases. Manual diagnosis is labor-intensive and inefficient, and the existing automated diagnosis methods for OCT images cannot classify according to the types of fundus diseases, resulting in low treatment efficiency and accuracy of automated diagnosis. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for diagnosing fundus diseases based on disease domains, aiming to solve the problems in the prior art that manual diagnosis is labor-intensive and inefficient, and the existing automated diagnosis methods for fundus images cannot classify according to the types of fundus diseases, resulting in low treatment efficiency and accuracy of automated diagnosis.

[0004] To achieve the above purpose, the present invention is realized through the following technical solutions:

[0005] On the one hand, the present invention provides a method for diagnosing fundus diseases based on disease domains, including:

[0006] Step S1: Obtain a number of fundus images, construct an image set, and divide the images in the image set into N disease domains according to clinical medical diagnosis results;

[0007] Step S2: Construct a visual geometry neural network model and its classifier, and conduct multiple rounds of training:

[0008] In each round of training, dynamically and randomly assign a main domain and N - 1 auxiliary domains from the N disease domains, randomly select two images from the main domain, and randomly select one image from each auxiliary domain. Input all the selected images into the visual geometry neural network model and its classifier respectively, calculate the co-domain value of the two images selected from the main domain, and the foreign-domain value of any one of the two images selected from the main domain for each one image selected from the auxiliary domains. According to the co-domain value and the foreign-domain value, correct the current visual geometry neural network model and its classifier;

[0009] Step S3: According to the visual geometry neural network model and its classifier corrected after multiple rounds of training, diagnose the fundus image to be diagnosed.

[0010] Preferably, in the step S2, the construction of the visual geometry neural network model and its classifier includes: removing the last layer from the original visual geometry neural network model, and inserting two fully connected layers as the classifier into the last layer of the original visual geometry neural network model to obtain the visual geometry neural network model and its classifier.

[0011] Preferably, in the step S2, before all the selected images are respectively input into the visual geometry neural network model and its classifier, preprocess all the selected images.

[0012] Preferably, in the step S2, the operation of respectively inputting all the selected images into the visual geometry neural network model and its classifier specifically includes: recording the input picture of the visual geometry neural network model as I, then the output of the first fully connected layer of the classifier is the general feature F I , and input the general feature F I into the second fully connected layer of the classifier to obtain classification indicators in N dimensions, and perform normalization processing on the classification indicators in the N dimensions so that the sum of the classification indicators in the N dimensions is equal to 1.

[0013] Preferably, the expression for performing normalization processing on the classification indicators in the N dimensions is:

[0014]

[0015] where p k represents the k-th item among the classification indicators in N dimensions, k = 1,…,N, and k is a positive integer.

[0016] Preferably, perform cross-entropy calculation on the normalized classification indicators to obtain the classification residuals corresponding to the classification indicators in the N dimensions, and the calculation expression of the classification residual Loss cls is:

[0017]

[0018] where Loss cls represents the classification residual, y k represents the clinical medical diagnosis result for the k-th fundus disease, y k = 0 represents negative, y k = 1 represents positive.

[0019] Preferably, in the step S2, according to the co-domain value and the foreign-domain value, correct the current visual geometry neural network model and its classifier, and the specific operations include:

[0020] Step S2.1: Calculate the co-domain residual and the foreign-domain residual respectively according to the co-domain value and the foreign-domain value;

[0021] Step S2.2: Calculate the total residual according to the co-domain residual, the foreign-domain residual and the classification residual;

[0022] Step S2.3: Modify the visual geometry neural network model and its classifier according to the total residual by using the backpropagation algorithm to obtain the modified visual geometry neural network model and its classifier.

[0023] Preferably, the input dimension of the first fully connected layer among the two fully connected layers is 2048, and the output dimension is 1024; the input dimension of the second fully connected layer among the two fully connected layers is 1024, and the output dimension is 8.

[0024] Preferably, the calculation of the co-domain value specifically includes: respectively record two images randomly selected from the main domain as picture A and picture B, and input picture A and picture B into the visual geometry neural network model and its classifier to obtain ordinary features F A and F B , respectively cut the two ordinary features in the corresponding dimensions to obtain the corresponding matrix groups Ω(A) and Ω(B) as:

[0025]

[0026] where, means taking the data from the 1st to 511th segments in F A and recording it as u; means taking the 512th data in F A and recording it as α; means taking the data from the 513th to 1023rd segments in F A and recording it as v; means taking the 1024th data in F A and recording it as β; means taking the data from the 1st to 511th segments in F B and recording it as χ; means taking the 512th data in F B and recording it as γ; means taking the data from the 513th to 1023rd segments in F B and recording it as y; means taking the 1024th data in F B and recording it as δ;

[0027] Then the expression of the co-domain value H(A|B) of picture A with respect to picture B is:

[0028] H(A|B) = uT u + χ T χ - 2u T χ + α - γ + u T v - u T y + β - δ

[0029] The calculation of the outlier value specifically includes: Selecting an image in any auxiliary domain as Picture C, and inputting Picture C into the visual geometric neural network model and its classifier to obtain the general feature F C , for F C Cutting on the corresponding dimension to obtain F C The corresponding matrix group ΩC) is:

[0030]

[0031] Among them, means taking the data in the 1st - 511 segments of F C and recording it as means taking the 512th data in F C and recording it as ε; means taking the data in the 513th - 1023rd segments of F C and recording it as t; means taking the 1024th data in F C and recording it as σ;

[0032] Then the expression of the outlier value G(A|C) of Picture A with respect to Picture C is:

[0033]

[0034] The co - domain residual Loss H The calculation expression is:

[0035]

[0036] The outlier residual Loss G The calculation expression is:

[0037]

[0038] On the other hand, the present invention also provides a fundus disease diagnosis system based on the disease domain, including:

[0039] An image acquisition module, which acquires a plurality of fundus images, constructs an image set, and divides the images in the image set into N disease domains according to the clinical medical diagnosis results;

[0040] A model construction and training module, which is connected to the image acquisition module, is used to construct a visual geometric neural network model and its classifier, and perform multiple rounds of training: in each round of training, a main domain and N - 1 auxiliary domains are dynamically and randomly assigned from the N disease domains, two images are randomly selected from the main domain, and one image is randomly selected from each auxiliary domain. All the selected images are respectively input into the visual geometric neural network model and its classifier, and the co-domain value of the two images selected from the main domain and the foreign-domain value of any one of the two images selected from the main domain with respect to one image selected from each auxiliary domain are calculated. According to the co-domain value and the foreign-domain value, the current visual geometric neural network model and its classifier are corrected;

[0041] A diagnosis module, which is connected to the model construction and training module, diagnoses the fundus image to be diagnosed according to the visual geometric neural network model and its classifier corrected after multiple rounds of training.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] The present invention proposes a method and system for diagnosing fundus diseases based on disease domains. The optical coherence tomography images are classified into several disease domains based on disease domains, a visual geometric neural network and its classifier are constructed, and feature mining and classification of pictures are performed; the co-domain value and the foreign-domain value are calculated to obtain the corresponding co-domain residual and foreign-domain residual, and the visual geometric neural network and the classifier are corrected. The corrected visual geometric neural network and classifier are used to automatically diagnose the OCT image to be diagnosed, and the efficiency of diagnosis analysis and treatment is relatively high. Moreover, several groups are adopted to improve the diagnosis accuracy of the OCT image. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the present invention, the drawings required for description will be briefly introduced below. Obviously, the drawings in the following description are an embodiment of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts:

[0045] Figure 1 It is a flowchart of a method for diagnosing fundus diseases based on disease domains provided by an embodiment of the present invention;

[0046] Figure 2 It is a schematic diagram of the cutting of the ordinary feature F of Picture A provided by an embodiment of the present invention A on the corresponding dimension. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] The following is combined with the attached Figure 1-2The following detailed description and specific embodiments further elaborate on the method and system for diagnosing fundus diseases based on disease domains proposed by the present invention. According to the following description, the advantages and features of the present invention will become clearer. It should be noted that the attached drawings are in a very simplified form and use non-precise scales, solely for the purpose of conveniently and clearly assisting in explaining the embodiments of the present invention. To make the objectives, features, and advantages of the present invention more obvious and understandable, please refer to the attached drawings. It should be known that the structures, proportions, sizes, etc. shown in the drawings of this specification are only used to cooperate with the content disclosed in the specification for those skilled in this technology to understand and read, and are not used to limit the limiting conditions for the implementation of the present invention. Therefore, they do not have any substantial technical meaning. Any modification of the structure, change in the proportional relationship, or adjustment of the size, without affecting the efficacy that the present invention can produce and the objectives that can be achieved, should still fall within the scope covered by the technical content disclosed by the present invention.

[0048] In view of the large workload and low efficiency of manual diagnosis in the prior art, and the existing automated diagnosis methods for OCT images have problems such as being unable to classify according to fundus disease types, with both the treatment efficiency and the accuracy of automated diagnosis being relatively low.

[0049] On the one hand, this embodiment provides a method for diagnosing fundus diseases based on disease domains, including:

[0050] Step S1: Obtain a number of fundus images (fundus images of optical coherence tomography in this embodiment), construct an image set, and divide the images in the image set into N disease domains according to the clinical medical diagnosis results.

[0051] In this embodiment, the pictures in the image set are divided into 8 disease domains, namely: choroidal neovascularization (CNV) domain, polypoidal choroidal vasculopathy (PCV) domain, dry age-related macular degeneration (DRY) domain, pathologic myopia (PM) domain, diabetic macular edema (DME) domain, retinal vein occlusion (RVO) domain, central serous chorioretinopathy (CSC) domain, and normal retina (NM) domain.

[0052] The optical coherence tomography images of each patient are stratified to obtain multiple two-dimensional pictures. Each two-dimensional picture is reviewed by three experts, and in the form of majority vote, all pictures containing fundus lesions are screened out. According to the clinical medical diagnosis results of the patients, all the pictures are grouped. If a patient has no fundus disease, all the pictures of this patient are classified into the NM domain.

[0053] Step S2: Construct a visual geometry neural network model and its classifier, and conduct multiple rounds of training:

[0054] In each round of training, a main domain and N - 1 auxiliary domains are dynamically and randomly assigned from the N disease domains. Two images are randomly selected from the main domain, and one image is randomly selected from each auxiliary domain. All the selected images are respectively input into the visual geometry neural network model and its classifier, and the co - domain value of the two images selected from the main domain and the hetero - domain value of any one of the two images selected from the main domain with respect to one image selected from each auxiliary domain are calculated. Based on the co - domain value and the hetero - domain value, the current visual geometry neural network model and its classifier are corrected.

[0055] Before all the selected images are respectively input into the visual geometry neural network model and its classifier, pre - processing is performed on all the selected images. In this embodiment, the way to perform pre - processing on all the selected images is: normalizing each pixel point x of the image.

[0056] The construction of the visual geometry neural network model and its classifier includes: removing the last layer in the original visual geometry neural network model, and inserting two fully - connected layers as the classifier in the last layer of the original visual geometry neural network model to obtain the visual geometry neural network model and its classifier. The input dimension of the first fully - connected layer among the two fully - connected layers is 2048, and the output dimension is 1024; the input dimension of the second fully - connected layer among the two fully - connected layers is 1024, and the output dimension is 8.

[0057] Record the input picture of the visual geometry neural network model as I, then the output of the first fully - connected layer of the classifier is the ordinary feature F I and input the ordinary feature F I into the second fully - connected layer of the classifier to obtain classification metrics of N dimensions, and perform normalization processing on the classification metrics of N dimensions so that the sum of the classification metrics of N dimensions is equal to 1.

[0058] Set classification metrics of N dimensions and perform Softmax normalization processing on the classification metrics of N dimensions. The expression of the normalization processing is:

[0059]

[0060] where p k represents the k - th item among the classification metrics of N dimensions, k = 1, …, N, and k is a positive integer.

[0061] In this embodiment, classification metrics of 8 dimensions are set, then the expression of the normalization processing is:

[0062]

[0063] Calculate the cross-entropy of the normalized classification metrics to obtain the classification residuals corresponding to the classification metrics in the N dimensions, where the classification residual is Loss cls The calculation expression is:

[0064]

[0065] where p k represents one of the classification metrics, and Loss cls represents the classification residual, and y k represents the clinical medical diagnosis result for the k-th fundus disease. y k =0 indicates negative, and y k =1 indicates positive.

[0066] The calculation of the co-domain value specifically includes: respectively record two images randomly selected from the main domain as Picture A and Picture B, input Picture A and Picture B into the visual geometry neural network model and its classifier, and respectively obtain the general features F A and F B . Cut the two general features in the corresponding dimensions respectively to obtain the corresponding matrix groups Ω(A) and Ω(B) as follows:

[0067]

[0068] where means to take the data from the 1st to 511th segments in F A and record it as u; means to take the 512th data in F A and record it as α; means to take the data from the 513th to 1023rd segments in F A and record it as v; means to take the 1024th data in F A and record it as β; means to take the data from the 1st to 511th segments in F B and record it as χ; means to take the 512th data in F B and record it as γ; means to take the data from the 513th to 1023rd segments in F B and record it as y; means to take the 1024th data in F B and record it as δ;

[0069] Then the expression of the co-domain value H(A|B) of Picture A with respect to Picture B is:

[0070] H(A|B)=u T u + χ T χ - 2uT χ + α - γ + u T v - u T y + β - δ (6)

[0071] The calculation of the outlier value specifically includes: Select an image in any auxiliary domain as Picture C, and input Picture C into the visual geometric neural network model and its classifier to obtain the general feature F C , for F C Perform cutting on the corresponding dimension to obtain F C The corresponding matrix group Ω(C) is:

[0072]

[0073] Among them, It means to take the data from the 1st to 511th segments in F C and denote it as It means to take the 512th data in F C and denote it as ε; It means to take the data from the 513th to 1023rd segments in F C and denote it as t; It means to take the 1024th data in F C and denote it as σ.

[0074] Then the expression of the outlier value G(A|C) of Picture A with respect to Picture C is:

[0075]

[0076] According to the co-domain value and the outlier value, correct the current visual geometric neural network model and its classifier, which specifically includes:

[0077] Step S2.1: Calculate the co-domain residual and the outlier residual respectively according to the co-domain value and the outlier value.

[0078] The calculation expression of the co-domain residual Loss H is:

[0079]

[0080] The calculation expression of the outlier residual Loss G is:

[0081]

[0082] In each round of training, a main domain and N - 1 auxiliary domains are dynamically and randomly assigned from the N disease domains. Two images are randomly selected from the main domain, and one image is randomly selected from each auxiliary domain. All the selected images are respectively input into the visual geometry neural network model and its classifier, and the co-domain value of the two images selected from the main domain and the foreign-domain value of any one of the two images selected from the main domain with respect to one image selected from each auxiliary domain are calculated respectively. In this embodiment, there are 8 disease domains. According to the above formula for calculating the foreign-domain residual, the foreign-domain value and the foreign-domain residual of picture A with respect to one image selected from each of the remaining 7 auxiliary domains are calculated, and the average value of the 8 obtained foreign-domain residuals is calculated to obtain the average foreign-domain residual

[0083] Step S2.2: Calculate the total residual according to the co-domain residual, the foreign-domain residual and the classification residual, that is, perform weighted summation on the co-domain residual, the average foreign-domain residual and the classification residual to obtain the total residual. The expression of the total residual Loss is:

[0084]

[0085] where, w 1 , w 2 and w 3 respectively represent the weights corresponding to the co-domain residual, the foreign-domain residual and the classification residual. In this embodiment, preferably, w 1 = 0.3, w 2 = 0.3, w 3 = 0.4.

[0086] Step S2.3: According to the total residual, use the backpropagation algorithm to correct the visual geometry neural network model and its classifier to obtain the corrected visual geometry neural network model and its classifier.

[0087] Repeat the training process in Step S2 to perform multiple rounds of correction on the visual geometry neural network model and its classifier until the preset conditions are met and then terminate. The termination conditions in this embodiment include: 1) After all pictures participate in the training, calculate the average value of the total residual Loss. If the average value of the total residual is less than a preset average value, that is, terminate the training. The preset average value in this embodiment is 0.0005; 2) Record all pictures participating in the training as one round. When the total number of training rounds reaches a preset number of rounds, terminate the training. The preset number of rounds in this embodiment is 300 rounds. Terminate the training if any one of the above two conditions is met.

[0088] Step S3: Diagnose the fundus images to be diagnosed according to the visually geometric neural network model and its classifier corrected after multiple rounds of training. Specifically, it includes: inputting the fundus images to be diagnosed into the visually network model and its classifier corrected after multiple rounds of training, and the classifier obtains the diagnostic result, that is, determines whether the fundus images to be diagnosed belong to CVN disease, or PCV disease, or DRY disease, or PM disease, or DME disease, or RVO disease, or CSC disease, or normal disease-free pictures.

[0089] The following further introduces the above method with specific embodiments:

[0090] As Figure 1 shown, 5000 OCT fundus images are collected to form an image set. The image set is grouped according to the clinical medical diagnosis results. There are a total of 8 groups including 7 disease groups and 1 healthy group. Each group forms a disease domain by itself, namely CVN domain, PCV domain, DRY domain, PM domain, DME domain, RVO domain, CSC domain and NM domain. Randomly select one of the 8 groups as the main domain, and the other 7 groups are used as auxiliary domains. Randomly select 2 pictures from the main domain and record them as Im_1 and Im_2 respectively; randomly select 1 picture from each of the 7 auxiliary domains and record them as Im_3, Im_4, Im_5, Im_6, Im_7, Im_8, Im_9 respectively.

[0091] Preprocess each of the above pictures. The specific processing method is: adjust the length and width of the pictures to 512, and then perform normalization processing on each pixel point x of the pictures. The normalization processing method is:

[0092]

[0093] Successively input the preprocessed pictures into the VGG neural network, and the general features of the corresponding 2 main domain pictures (i.e., Im_1 and Im_2) and the general features of the 7 auxiliary domain pictures (i.e., Im_3, Im_4, Im_5, Im_6, Im_7, Im_8, Im_9) can be obtained. Input the general features of the Im_1 picture into the classifier and calculate the corresponding classification residuals. Calculate the co-domain value of the Im_1 picture for the Im_2 picture. Calculate the cross-domain values of the Im_1 picture for the Im_3, Im_4, Im_5, Im_6, Im_7, Im_8, Im_9 pictures respectively, and 7 cross-domain values can be obtained.

[0094] Determine whether the co-domain value is less than 0. If the co-domain value is not less than 0, calculate the co-domain residual; if the co-domain value is less than 0, determine whether the number of out-of-domain values greater than or equal to 0 among the above 7 out-of-domain values is greater than 4. If the number of out-of-domain values greater than or equal to 0 among the 7 out-of-domain values is not greater than 4, calculate the average out-of-domain residual; if the number of out-of-domain values greater than or equal to 0 among the 7 out-of-domain values is greater than 4, select another fundus OCT image to be diagnosed for testing. Calculate the total residual according to the classification residual, co-domain residual, and out-of-domain residual. It should be noted that if the number of out-of-domain values greater than or equal to 0 among the 7 out-of-domain values is greater than 4, it means that the VGG neural network and the classifier have been fully corrected, and the VGG neural network and the classifier at this time have been trained and can be used to diagnose fundus OCT images.

[0095] Use the backpropagation algorithm to correct the VGG neural network and the classifier. Repeat the above process until it terminates after meeting the preset conditions, and obtain the VGG neural network and the classifier corrected after multiple rounds of training.

[0096] Input the fundus OCT image to be diagnosed into the VGG neural network and the classifier corrected after multiple rounds of training for diagnosis. Obtain the diagnosis result through the classifier, that is, determine whether the image belongs to CVN disease, or PCV disease, or DRY disease, or PM disease, or DME disease, or RVO disease, or CSC disease, or a normal disease-free image.

[0097] On the other hand, this embodiment also provides a fundus disease diagnosis system based on the disease domain, including: an image acquisition module, which acquires a plurality of fundus images, constructs an image set, and divides the images in the image set into N disease domains according to the clinical medical diagnosis results; a model construction and training module, which is connected to the image acquisition module and is used to construct a visual geometry neural network model and its classifier, and perform multiple rounds of training: in each round of training, dynamically and randomly allocate a main domain and N - 1 auxiliary domains from the N disease domains, randomly select two images from the main domain, randomly select one image from each auxiliary domain, input all the selected images into the visual geometry neural network model and its classifier respectively, calculate the co-domain values of the two images selected from the main domain, and the out-of-domain values of any one of the two images selected from the main domain for each one image selected from each auxiliary domain, and correct the current visual geometry neural network model and its classifier according to the co-domain values and the out-of-domain values; a diagnosis module, which is connected to the model construction and training module and diagnoses the fundus image to be diagnosed according to the visual geometry neural network model and its classifier corrected after multiple rounds of training.

[0098] In summary, for the fundus disease diagnosis method and system provided in this embodiment, the optical coherence tomography images are classified into several disease domains based on the disease domain, a visual geometry neural network and its classifier are constructed to mine features and classify the pictures; the co-domain value and the foreign-domain value are calculated to obtain the corresponding co-domain residual and foreign-domain residual, and the visual geometry neural network and the classifier are corrected. The OCT images to be diagnosed are automatically diagnosed using the visually geometric neural network and classifier after multiple corrections, and the efficiency of diagnosis analysis and treatment is relatively high. Moreover, several groups are used to improve the diagnostic accuracy of the OCT images.

[0099] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "including a..." does not exclude the presence of additional identical elements in the process, method, article or device including the element.

[0100] It should be noted that the devices and methods disclosed in the embodiments of this article can also be implemented in other ways. The device embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions and operations of devices, methods and computer program products according to multiple embodiments of this article. In this regard, each block in the flowchart or block diagram may represent a module, program or part of the code, and the module, program segment or part of the code includes one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than marked in the accompanying drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and the combination of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system for performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0101] In addition, each functional module in the various embodiments of this text may be integrated together to form an independent part, or each module may exist alone, or two or more modules may be integrated to form an independent part.

[0102] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be recognized that the above description should not be considered as a limitation of the present invention. After those skilled in the art have read the above content, various modifications and alternatives to the present invention will be obvious. Therefore, the protection scope of the present invention should be defined by the appended claims.

Claims

1. A method for diagnosing fundus diseases based on disease domain, characterized in that: include: Step S1: acquiring a number of fundus images, constructing an image set, and dividing the images in the image set into N disease domains according to clinical medical diagnosis results; Step S2: Construct a visual geometric neural network model and its classifier, and perform multiple rounds of training: In each round of training, a main domain and N-1 auxiliary domains are dynamically and randomly allocated from the N disease domains, two images are randomly selected from the main domain, and one image is randomly selected from each auxiliary domain, and all the selected images are respectively input into the visual geometric neural network model and its classifier, and the common domain value of the two images selected from the main domain and the heterodomain value of any image of the two images selected from the main domain with respect to an image selected from each auxiliary domain are respectively calculated, and according to the common domain value and the heterodomain value, the current visual geometric neural network model and its classifier are modified; Step S3: diagnose the fundus image to be diagnosed according to the visual geometric neural network model and its classifier modified after multiple rounds of training.

2. The method for diagnosing fundus diseases based on disease domain according to claim 1, characterized in that: In step S2, constructing the visual geometric neural network model and its classifier includes: removing the last layer in the original visual geometric neural network model, and inserting two fully connected layers as classifiers in the last layer of the original visual geometric neural network model to obtain the visual geometric neural network model and its classifier.

3. The method for diagnosing fundus diseases based on disease domain according to claim 2, characterized in that: In the step S2, all the selected images are preprocessed before being respectively input into the visual geometric neural network model and its classifier.

4. The method for diagnosing fundus diseases based on disease domain according to claim 3, characterized in that: In the step S2, all selected images are input to the visual geometric neural network model and its classifier respectively, which specifically includes: recording the input image of the visual geometric neural network model as I, then the output of the first fully connected layer of the classifier is the common feature F I , the common feature F I The second fully connected layer of the classifier is input to obtain classification indicators of N dimensions, and the classification indicators of the N dimensions are normalized so that the sum of the classification indicators of the N dimensions is equal to 1.

5. The method for diagnosing fundus diseases based on disease domain according to claim 4, characterized in that: The expression for normalizing the classification indicators of the N dimensions is: Among them, p k Represents the kth item in the classification index of N dimensions, k = 1,…,N, k is a positive integer.

6. The method for diagnosing fundus diseases based on disease domain according to claim 5, characterized in that: The cross entropy calculation is performed on the normalized classification index to obtain the classification residual corresponding to the classification index of the N dimensions. The classification residual Loss cls The calculation expression is: Among them, Loss cls represents the classification residual, y k represents the clinical diagnosis result of the kth fundus disease, y k =0 means negative, y k =1 indicates positive.

7. The method for diagnosing fundus diseases based on disease domain according to claim 6, characterized in that: In the step S2, the current visual geometric neural network model and its classifier are modified according to the common domain value and the foreign domain value, which specifically includes: Step S2.1: Calculate the common domain residual and the foreign domain residual respectively according to the common domain value and the foreign domain value; Step S2.2: Calculate the total residual according to the common domain residual, the heterogeneous domain residual and the classification residual; Step S2.3: Based on the total residual, the visual geometric neural network model and its classifier are corrected using a back propagation algorithm to obtain a corrected visual geometric neural network model and its classifier.

8. The method for diagnosing fundus diseases based on disease domain according to claim 7, characterized in that: The input dimension of the first of the two fully connected layers is 2048, and the output dimension is 1024; the input dimension of the second of the two fully connected layers is 1024, and the output dimension is 8.

9. The method for diagnosing fundus diseases based on disease domain according to claim 8, characterized in that: The calculation of the co-domain value specifically includes: recording two images randomly selected from the main domain as picture A and picture B respectively, inputting picture A and picture B into the visual geometric neural network model and its classifier to obtain common features F respectively. A and F B , the two common features are cut in the corresponding dimensions, and the corresponding matrix groups Ω(A) and Ω(B) are obtained as follows: in, Indicates taking F A The 1st to 511th segments of data are recorded as u; Indicates taking F A The 512th data in is recorded as α; Indicates taking F A The data of the 513th to 1023rd paragraph in are recorded as ν; Indicates taking F A The 1024th data in is recorded as β; Indicates taking F B The data of the 1st to 511th paragraphs are recorded as χ; Indicates taking F B The 512th data in is recorded as γ; Indicates taking F B The data from segment 513 to 1023 in the dataset are recorded as y; Indicates taking F B The 1024th data in is recorded as δ; Then the expression of the co-domain value H(A|B) of the picture A for the picture B is: H(A|B)=u T u+x T x-2u T x+a-g+u T wow T y+β-δ The calculation of the foreign domain value specifically includes: selecting an image from any auxiliary domain as picture C, inputting picture C into the visual geometric neural network model and its classifier to obtain the common feature F C , for F C Cut on the corresponding dimension to get F C The corresponding matrix group Ω(C) is: in, Indicates taking F C The data of the 1st to 511th paragraphs in Indicates taking F C The 512th data in is recorded as ε; Indicates taking F C The data of the 513th to 1023rd segment in are recorded as τ; Indicates taking F C The 1024th data in is recorded as σ; Then the expression of the foreign value G(A|C) of the picture A to the picture C is: The co-domain residual Loss H The calculation expression is: The foreign residual Loss G The calculation expression is:

10. A fundus disease diagnosis system based on disease domain, characterized in that: include: An image acquisition module acquires a number of fundus images, constructs an image set, and divides the images in the image set into N disease domains according to clinical medical diagnosis results; A model building and training module, which is connected to the image acquisition module, is used to build a visual geometric neural network model and its classifier, and perform multiple rounds of training: in each round of training, a main domain and N-1 auxiliary domains are dynamically and randomly allocated from the N disease domains, two images are randomly selected from the main domain, and one image is randomly selected from each auxiliary domain, and all the selected images are respectively input into the visual geometric neural network model and its classifier, and the common domain value of the two images selected from the main domain and the heterodomain value of any image of the two images selected from the main domain with respect to an image selected from each auxiliary domain are respectively calculated, and the current visual geometric neural network model and its classifier are corrected according to the common domain value and the heterodomain value; The diagnosis module is connected to the model building and training module, and diagnoses the fundus image to be diagnosed according to the visual geometric neural network model and its classifier modified after multiple rounds of training.

Images