Training method of ophthalmic disease prediction model based on artificial intelligence

Abstract

The present application relates to the field of medical care informatics, and specifically to an artificial intelligence-based ophthalmic disease prediction model training method; comprising the following steps performed in sequence: S1: obtaining an ophthalmic disease data set, and annotating each image in the data set with a disease; S2: performing data expansion on the annotated data set to obtain an ophthalmic disease training data set; S3: using a neural network algorithm based on dynamic weight redistribution as a feature extraction model to extract features from the ophthalmic disease training data set to obtain extracted features of the ophthalmic disease; S4: using an autoencoder neural network based on feature refinement as a feature dimensionality reduction model to reduce the dimensionality of the extracted features to obtain the initial low-dimensional features of the ophthalmic disease after dimensionality reduction; S5: inputting the initial low-dimensional features of the ophthalmic disease after dimensionality reduction into a classifier model for training to obtain classification results; S6: classifying the data set using the trained model; effectively capturing the features of the ophthalmic disease data.

Description

Technical Field

[0001] The present application relates to the field of healthcare informatics, and specifically to an artificial intelligence-based ophthalmic disease prediction model training method. Background Art

[0002] With the increasing incidence of ophthalmic diseases around the world, early diagnosis has become an important need in the medical field. There are many types of ophthalmic diseases, including glaucoma, diabetic retinopathy and macular degeneration. Traditional diagnostic methods rely on the doctor's experience and fewer diagnostic tools, which often leads to misdiagnosis or missed diagnosis.

[0003] In order to improve the accuracy and efficiency of diagnosis, artificial intelligence-based technologies have gradually attracted attention. In this context, artificial intelligence technologies such as machine learning and deep learning have been introduced into the field of ophthalmic disease prediction. These technologies are able to process large amounts of clinical data and extract potential disease features from them. However, existing methods still face some challenges in feature extraction. Summary of the invention

[0004] The purpose of the present invention is to provide an artificial intelligence-based ophthalmic disease prediction model training method that improves feature extraction capabilities.

[0005] In order to achieve the above object, the present invention adopts such technical solution:

[0006] The artificial intelligence-based ophthalmic disease prediction model training method comprises the following steps performed in sequence:

[0007] S1: Obtain a dataset of ophthalmic diseases, and annotate each image in the dataset with the disease;

[0008] S2: using a generative adversarial network algorithm based on multi-sample resonance optimization to perform data expansion on the labeled data set to obtain a training data set for ophthalmic diseases;

[0009] S3: Using a neural network algorithm based on dynamic weight redistribution as a feature extraction model to extract features from the training data set of ophthalmic diseases, and obtaining the extracted features of ophthalmic diseases;

[0010] S4: Using the autoencoder neural network based on feature refinement as the feature dimension reduction model to reduce the dimension of the extracted features and obtain the initial low-dimensional features of the ophthalmic disease after dimension reduction;

[0011] S5: Input the initial low-dimensional features of the ophthalmic disease after dimensionality reduction into the classifier model for training, and output the ophthalmic disease classification results;

[0012] S6: Inputting the data set into the trained feature extraction model and the feature dimension reduction model for feature processing, and inputting the processed features into the trained classifier model for classification to obtain ophthalmic disease prediction results.

[0013] Preferably, in step S2, a generative adversarial network algorithm based on multi-sample resonance optimization is used for data expansion, and the specific steps are as follows:

[0014] S201: Initialize the parameters of the generator and discriminator: The generator is responsible for generating nearly real ophthalmic disease data, and the discriminator tries to distinguish the generated samples from the real images. Let the generator be , the discriminator is , are the parameters of the generator, are the parameters of the discriminator;

[0015] S202: In the adversarial training stage, a positive and negative constraint mechanism is used to optimize the training process by dynamically adjusting the weights in the loss function. The constraints are:

[0016] ;

[0017] in, Represents the value function of the generative adversarial network algorithm, which is the constraint condition in the adversarial training process. Represents real data, represents the noise sampled from the prior distribution, It means that it obeys a specific distribution. represents the distribution of real data, is the discriminator function, is the noise distribution, is the generator function, is the regularization weight coefficient of the generative adversarial network, is the regularization term of the generative adversarial network, represents the discriminator, represents a generator, express expectations;

[0018] The regularization term of the generative adversarial network algorithm improves the generalization ability of the generative adversarial network algorithm by penalizing the sensitivity of the discriminator output to the input. The calculation method is expressed as:

[0019] ;

[0020] in, is the smoothing constant, is the number of samples input to the generative adversarial network algorithm in the current batch, It is Real data, represents the gradient of the real data, is the L2 norm, is the sparse loss term, is the sample sparsification weight, Represents the weight of the sparsification loss term, which controls the influence of the sparsification loss term on the total loss;

[0021] An adaptive generated sample sparsification strategy is used to optimize the distribution of generated samples. The calculation method of the sparsification loss term is expressed as:

[0022] ;

[0023] in, It is The sparse weight of each data point, is the mean of the currently generated samples, is a preset constant added to prevent division by zero. It is The noise input to the generator;

[0024] The calculation of the sparsification weight depends on the density estimation of the sample in the generated space, and its dynamic adjustment is expressed as:

[0025] ;

[0026] In the formula, is the sparsification weight adjustment factor, Is the measurement generation sample The function of the density between the currently generated sample sets, Indicates the currently generated sample set;

[0027] S203: When the generator and the discriminator reach a certain confrontational balance, the multi-sample resonance optimization strategy is set based on the resonance index Select samples for key training, and the resonance optimization function is expressed as:

[0028] ;

[0029] In the formula, represents the learning rate of the generator, represents the weight of resonance optimization, represents the resonance optimization function, Indicates a parameter update operation. represents the gradient with respect to the generator parameters, is the resonance standard sample;

[0030] The resonance optimization function depends on the relationship between the generated samples of the generator and the resonance index. Specifically, the similarity between the generated samples and the resonance samples is enhanced through the exponentially decaying sum of square errors to increase the responsiveness of the model to specific data features. The calculation method of the exponentially decaying sum of square errors is expressed as:

[0031] ;

[0032] In the formula, is the adjustment factor that controls the resonance sensitivity, is the number of generated samples, It is Resonance standard samples, It is The noise input to the generator is represents the natural exponential function, represents the square of L2 norm;

[0033] S204: Repeat the above steps S202-S203 until the preset stop iteration condition is met, and output the training data set of ophthalmic diseases.

[0034] Preferably, in step S3, the feature extraction model is a fully connected neural network, and the specific training steps are as follows:

[0035] S301: Initialize the parameters of the fully connected neural network: Use the following formula (1) to calculate the initial weight matrix of the neural network: , the following formula (2) is used to calculate the initial bias vector of the neural network :

[0036] (1);

[0037] (2);

[0038] in, is the dimension of the feature space of the training dataset of ophthalmic diseases, is the dimension of the number of samples in the training data set for ophthalmic diseases. The randn() function is used to generate random numbers with a standard normal distribution.

[0039] S302: In each iteration of the fully connected neural network, the training data set of ophthalmic diseases is forward propagated. The input of each layer is determined by the output of the previous layer and the parameters of the current layer. The data of each ophthalmic disease training data set is mapped to a new feature space defined by a hyperplane. The feature vector after feature extraction is expressed by the following formula :

[0040] ;

[0041] in, The feature vector representing the training dataset of ophthalmic diseases input to the fully connected neural network, is the Sigmoid activation function, and are the weights and biases of the fully connected neural network after training the tth iteration;

[0042] S303: Assumption For the output of layer I in the fully connected neural network, a dynamic weight redistribution network is set, and the evaluation characteristics of the dynamic weight redistribution network of layer I of the fully connected neural network are calculated using the following formula: :

[0043] ;

[0044] in, It is a normalization function used to evaluate the activity or information content of the layer output;

[0045] The weight adjustment coefficient of the first layer of the fully connected neural network is calculated using the following formula: :

[0046] ;

[0047] in, and are the weights and biases of the dynamic weight redistribution network, which are used to generate the weight adjustment coefficients of each layer from the evaluation indicators. is the preset Softmax function, is the evaluation characteristic of the dynamic weight redistribution network of layer I of the fully connected neural network;

[0048] After the above weight adjustment, the output of the Ith layer after weight adjustment is calculated using the following formula: :

[0049] ;

[0050] in, To dynamically adjust the factor, the output of the fully connected neural network is adaptively amplified or suppressed according to the strength of the gradient, enhancing the sensitivity of the feature extraction model to important features;

[0051] Adopting the gradient-guided adaptive feature enhancement strategy, using the gradient information in the back-propagation process, adaptively strengthen the important feature representation in the feature extraction model. The gradient of the loss function is , through gradient-guided adaptive feature enhancement, dynamically adjust the factor The calculation method is expressed as:

[0052] ;

[0053] in, yes The modulus of is used to evaluate the influence of this layer in the training process;

[0054] S304: In the feature conversion process, the feature is sparsely processed by regularization, and the L1 norm is used for sparse processing in the loss function of the fully connected neural network. The loss function of the fully connected neural network is As shown below:

[0055] ;

[0056] in, is the target output, that is, the input sample and the corresponding true label vector; is the square of the L2 norm, is the L1 norm, which is used to achieve the sparsification of weights. is the regularization coefficient of the fully connected neural network, which regulates the degree of sparsity. is the mapping function from the layer output of the fully connected neural network to the predicted output, After weight adjustment, Layer output, is the sparse regularization term;

[0057] Sparse regularization term Used to enhance the generalization ability of feature extraction models, The calculation method is expressed as:

[0058] ;

[0059] in, is the weight matrix of the neural network. Row and Elements of a column;

[0060] Adopting an adaptive adjustment mechanism based on training error, the regularization parameter decays exponentially with the progress of iterations. The regularization coefficient of the tth iteration of the fully connected neural network is The calculation method is expressed as:

[0061] ;

[0062] in, is the regularization strength of the initial fully connected neural network, is the decay rate of the fully connected neural network, is the current iteration number, is the total number of iterations;

[0063] S305: According to the loss function The calculated gradient is used to adjust the parameters of the fully connected neural network through the back propagation algorithm to optimize the hyperplane position and the weights of the fully connected neural network. The parameters of the fully connected neural network are updated as follows:

[0064] ;

[0065] in, is the parameter update learning rate of the fully connected neural network, which determines the step size of the parameter update. and are the weights and biases of the fully connected neural network after training the tth iteration, and are the gradients of the weights and biases of the fully connected neural network, For the The weights of the fully connected neural network at iterations, For the The weights of the fully connected neural network at iterations, For the The bias of the fully connected neural network for the iteration, For the The bias of the fully connected neural network for the iteration, is the symbol of partial derivative;

[0066] S306: Repeat the above steps S302-S305 until the preset stop iteration condition is met, and output the extracted features ;

[0067] Preferably, the extracted features The input is put into the feature dimensionality reduction model for training. The feature dimensionality reduction model includes an encoder, a decoder and a feature adjustment module. The specific training steps of the feature dimensionality reduction model are as follows:

[0068] S401: The encoder uses a multi-layer nonlinear mapping structure to extract features Mapping from high-dimensional data to the initial low-dimensional feature space, the initial low-dimensional features It is expressed as:

[0069] ;

[0070] in, is the weight matrix of the encoder, is the encoder bias vector, is the multi-layer Sigmoid activation function of the encoder;

[0071] S402: After the initial low-dimensional features are generated, the feature adjustment module automatically generates feature weights according to the feature importance in the current feature space. The feature adjustment module assigns the same initial weights to all features during initialization so that they can be gradually adjusted according to feature contributions in subsequent steps. The initial weight matrix The calculation method is expressed as:

[0072] ;

[0073] in, is the feature weight vector, Each element in Initialized to the same value, indicating that all initial low-dimensional features have the same importance in the initial stage. is a function for extracting the diagonal elements of a matrix;

[0074] Initial low-dimensional features of ophthalmic diseases after dimensionality reduction It is expressed as:

[0075] ;

[0076] S403: The feature adjustment module adjusts the initial low-dimensional features of the ophthalmic disease after dimensionality reduction Recursive optimization is performed. In each round of iteration, the feature adjustment module adjusts the weight of each feature according to the performance of the previous round of features, gradually enhances those features with important influence, and gradually weakens redundant or noise features. In the round of iteration, The feature weight of the round iteration is :

[0077] ;

[0078] in, Indicates The feature weights of the round iteration, is the learning rate of the feature dimensionality reduction model, is the loss function of the feature dimensionality reduction model, is the label data, Represents the gradient of the loss function with respect to the feature weight;

[0079] S404: The decoder remaps the low-dimensional features back to high-dimensional features to ensure that no important information is lost during the dimensionality reduction process. The decoder reconstructs the high-dimensional data for:

[0080] ;

[0081] in, is the weight matrix of the decoder, is the bias vector of the decoder, is the multi-layer Sigmoid activation function of the decoder;

[0082] S405: Repeat the iteration steps S402-S404 until the preset stop iteration condition is met, and output the initial low-dimensional features of the ophthalmic disease after dimensionality reduction ;

[0083] Preferably, in step S5, the classifier model includes an input layer, a hidden layer and an output layer, and the specific training steps of the classifier model are as follows:

[0084] S501: Initialize the hidden layer weights of the classifier model, randomly generate the connection weights of the hidden layer, and initialize the weights of the hidden layer It is expressed by the following formula:

[0085] ;

[0086] in, is the initialization scalar that controls the range of weights, Indicates the generation of a A random matrix is ​​used to provide initial weights for each hidden node;

[0087] S502: Initial low-dimensional features of ophthalmic diseases after dimensionality reduction Input into the classifier model to form a feature matrix, which is represented as a feature input vector :

[0088] ;

[0089] in, For the The characteristic value of the sample, is the feature dimension;

[0090] S503: Use activation function to input feature vector Processing to get the output of the hidden layer :

[0091] ;

[0092] in, is the activation function of the classifier model, is the bias term of the hidden layer;

[0093] S504: Output to hidden layer Perform inverse correction by calculating the output of the output layer The difference between the actual label and the hidden layer weights are adjusted inversely to reduce the prediction error. The prediction error of the classifier model during training As shown below:

[0094] ;

[0095] in, is the true label of the sample, ) is the Softmax function, Indicates the use of a function to perform category calculation on the hidden layer output of the classifier model;

[0096] The weight of the hidden layer is updated by the inverse correction strategy. It is expressed by the following formula:

[0097] ;

[0098] in, is the first input to the classifier model The dynamic importance of a feature, indicating the contribution of the feature in the current sample, is the first input to the classifier model eigenvalues, is the learning rate of the classifier model, The number of samples input to the classifier model for the current batch, is the first-order derivative of the activation function of the classifier model;

[0099] Dynamically adjust feature weights using fuzzy logic: The dynamic importance of features input to the separator model adjustment is calculated as:

[0100] ;

[0101] in, is the fuzzy logic evaluation function, Features The fuzzy factor reflects the sensitivity of the feature to the prediction result;

[0102] The calculation method of the fuzzy logic evaluation function is expressed as:

[0103] ;

[0104] in, Features The fuzzy factor, is the first input to the classifier model eigenvalues;

[0105] S505: According to the uncertainty of the prediction result of each sample by the classifier model, the weight of the feature is dynamically adjusted. For the features with uncertain classification results, a higher weight is given, so that the classifier model pays more attention to these challenging features during training. The method of dynamically adjusting the feature weight is expressed as follows:

[0106] ;

[0107] in, is the first input to the classifier model The weight of the samples, To adjust the uncertainty sensitivity parameter, is the threshold used to determine whether a sample is considered uncertain, For the classifier model The prediction error during the training process for samples;

[0108] S506: Calculate the final classification result by weighted calculation of the hidden layer output as the predicted output of the classifier model output layer :

[0109] ;

[0110] in, is the output layer weight of the extreme learning machine, is the output layer bias of the extreme learning machine;

[0111] S507: Output the prediction With the true label Compare, calculate the error, and feed the error back to the classifier model. Further iterate the training process, update the hidden layer weights and feature weights in each iteration to gradually improve the performance of the separator model. The final error is calculated and the feedback adjustment is expressed as: ;

[0112] in, is the adjustment amount of the output layer weight of the classifier model, is the activation function of the classifier model The second-order derivative of is the true label of the sample;

[0113] S508: Repeat the above steps S502-S507 until the preset stop condition is met, and output the ophthalmic disease classification result.

[0114] Preferably, in step S2, data expansion is performed on the labeled data set by adding noise, vector concatenation or vector translation.

[0115] Preferably, in step S5, a support vector machine, a random forest, a decision tree or an extreme learning machine is used as a classifier model.

[0116] By adopting the aforementioned design scheme, the beneficial effects of the present invention are as follows: the present application adopts a neural network algorithm with dynamic weight redistribution for feature extraction, effectively capturing the important features of ophthalmic disease data, solving the problem of insufficient adaptability of existing neural networks to dynamic environments in feature extraction, and improving the feature extraction capability by dynamically adjusting the weights of each layer; adopting an autoencoder neural network based on feature refinement for feature dimensionality reduction, solving the problem of information loss in the process of high-dimensional data processing, ensuring that the dimensionality reduction process is reversible, and gradually adjusting the importance of features through recursive feature adaptive optimization to strengthen key features. BRIEF DESCRIPTION OF THE DRAWINGS

[0117] Figure 1 It is a data flow diagram of the present invention;

[0118] Figure 2 This is the training process of the feature dimensionality reduction model of the present invention. DETAILED DESCRIPTION

[0119] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only partial embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0120] Artificial intelligence-based ophthalmic disease prediction model training method, such as Figure 1 As shown, the steps include the following steps performed in sequence:

[0121] S1: Obtain a dataset of ophthalmic diseases and annotate each image in the dataset with the disease;

[0122] In this embodiment, the data set of ophthalmic diseases comes from, including but not limited to, hospitals or professional ophthalmology clinics.

[0123] The data were collected using non-invasive fundus photography and optical coherence tomography to obtain clinical diagnostic data about the patient's fundus.

[0124] The data is stored in the form of discrete vectors. In one embodiment, the attributes of the data include: age R_a, gender S_a, intraocular pressure I_a, retinal thickness T_a, vascular density V_a, foveal thickness M_a, diabetes history D_a, myopia degree N_a, astigmatism degree A_a, and genetic tendency G_a. It should be noted that this embodiment is only used to illustrate a data format and type of the present invention. In actual applications, the attributes of the data are usually more than 10 attributes, and the number of attributes of the data may reach dozens or even hundreds.

[0125] In this embodiment, disease labeling is performed by manual labeling, and the labeled categories include various ophthalmic diseases, such as glaucoma, diabetic retinopathy, macular degeneration, etc.

[0126] S2: The labeled data set is expanded using a generative adversarial network algorithm based on multi-sample resonance optimization to obtain a training data set for ophthalmic diseases; the adversarial process between the generator and the discriminator of the generative adversarial network is used to continuously optimize the quality of the generated samples. On the basis of the traditional generative adversarial network architecture, a positive and negative constraint mechanism is used to adjust the information entropy between the generated template and the real sample, so as to enhance the data expansion model's ability to capture the characteristics of ophthalmic diseases. The specific steps are as follows:

[0127] S201: Initialize the parameters of the generator and discriminator: The generator is responsible for generating nearly real ophthalmic disease data, and the discriminator tries to distinguish the generated samples from the real images. Let the generator be , the discriminator is , are the parameters of the generator, is the parameter of the discriminator; in this embodiment, the parameter initialization method of the generative adversarial network is random initialization, and the random value obeys a normal distribution with a mean of 0 and a variance of the unit matrix.

[0128] S202: In the adversarial training stage, the generator and the discriminator perform adversarial training. The generator attempts to generate more and more realistic samples to deceive the discriminator, and the discriminator continuously learns how to distinguish between real data and generated samples.

[0129] In the adversarial training stage, the positive and negative constraint mechanism is used to optimize the training process by dynamically adjusting the weights in the loss function. The constraints are:

[0130] ;

[0131] in, Represents the value function of the generative adversarial network algorithm, which is the constraint condition in the adversarial training process. Represents real data, represents the noise sampled from the prior distribution, It means that it obeys a specific distribution. represents the distribution of real data, is the discriminator function, is the noise distribution, is the generator function, is the regularization weight coefficient of the generative adversarial network, is the regularization term of the generative adversarial network, represents the discriminator, represents a generator, express expectations;

[0132] The regularization term of the generative adversarial network algorithm improves the generalization ability of the generative adversarial network algorithm by penalizing the sensitivity of the discriminator output to the input. The calculation method is expressed as:

[0133] ;

[0134] in, is the smoothing constant, is the number of samples input to the generative adversarial network algorithm in the current batch, It is Real data, represents the gradient of the real data, is the L2 norm, is the sparse loss term, is the sample sparsification weight, Represents the weight of the sparsification loss term, which controls the influence of the sparsification loss term on the total loss;

[0135] An adaptive sample sparsification strategy is used to optimize the distribution of generated samples. By reducing the generation frequency of samples in high-density areas and increasing exploration in low-density areas, the ability of the generative adversarial network algorithm to identify rare diseases is improved. The distribution of generated samples is adjusted in an adaptive manner to make it more consistent with the sparsity of the actual disease distribution. The calculation method of the sparseness loss term is expressed as:

[0136] ;

[0137] in, It is The sparse weight of each data point, is the mean of the currently generated samples, is a preset constant added to prevent division by zero. It is The noise input to the generator;

[0138] The calculation of the sparsification weight depends on the density estimation of the sample in the generated space, and its dynamic adjustment is expressed as:

[0139] ;

[0140] In the formula, is the sparsification weight adjustment factor, Is the measurement generation sample The function of the density between the currently generated sample sets, Indicates the currently generated sample set;

[0141] S203: When the generator and the discriminator reach a certain adversarial balance, a multi-sample resonance optimization strategy is adopted. By selecting samples that can cause resonance in the adversarial network algorithm for key training, it helps the adversarial network algorithm find key variation features in complex ophthalmic disease data sets. The multi-sample resonance optimization strategy is based on the resonance index Select samples for key training, and the resonance optimization function is expressed as:

[0142] ;

[0143] In the formula, represents the learning rate of the generator, represents the weight of resonance optimization, represents the resonance optimization function, Indicates a parameter update operation. represents the gradient with respect to the generator parameters, is a resonance standard sample; in this embodiment, Set to 0.01.

[0144] The resonance optimization function depends on the relationship between the generated samples of the generator and the resonance index. Specifically, the similarity between the generated samples and the resonance samples is enhanced through the exponentially decaying sum of square errors to increase the responsiveness of the model to specific data features. The calculation method of the exponentially decaying sum of square errors is expressed as:

[0145] ;

[0146] In the formula, is the adjustment factor that controls the resonance sensitivity, is the number of generated samples, It is Resonance standard samples, It is The noise input to the generator is represents the natural exponential function, represents the square of L2 norm;

[0147] S204: Repeat the above steps S202-S203 until the preset stop iteration condition is met, which means that the data expansion model training is completed and the training data set of ophthalmic diseases is output.

[0148] In this embodiment, assuming that the original collected samples are 800, and the data expansion model generates 200 samples through expansion, the expanded data set contains 1000 samples.

[0149] In step S2, the labeled data set may also be augmented by adding noise, vector concatenation, or vector translation.

[0150] S3: Using a neural network algorithm based on dynamic weight redistribution as a feature extraction model to extract features from the training data set of ophthalmic diseases can maintain the important features and structural information of ophthalmic disease data, more flexibly capture and express the distinctive features of ophthalmic disease data, and obtain the extracted features of ophthalmic diseases;

[0151] In step S3, the feature extraction model is a 6-layer fully connected neural network, and the specific training steps are as follows:

[0152] S301: Initialize the parameters of the fully connected neural network: Use the following formula (1) to calculate the initial weight matrix of the neural network: , the following formula (2) is used to calculate the initial bias vector of the neural network :

[0153] (1);

[0154] (2);

[0155] in, is the dimension of the feature space of the training dataset of ophthalmic diseases, is the dimension of the number of samples of the ophthalmic disease training data set, and the randn() function is used to generate random numbers with a standard normal distribution to ensure the diversity of initial weights and biases.

[0156] S302: In each iteration of the fully connected neural network, the training data set of ophthalmic diseases is forward propagated. The input of each layer is determined by the output of the previous layer and the parameters of the current layer. The data of each ophthalmic disease training data set is mapped to a new feature space defined by a hyperplane. The feature vector after feature extraction is expressed by the following formula :

[0157] ;

[0158] in, The feature vector representing the training dataset of ophthalmic diseases input to the fully connected neural network, is the Sigmoid activation function, and are the weights and biases of the fully connected neural network after training the tth iteration;

[0159] S303: Traditional feature extraction methods mainly focus on static feature processing and lack the adaptability to the dynamic ophthalmic disease data environment changes. The present invention uses a dynamic weight redistribution network to adjust the importance of each layer of weights in the neural network, and dynamically optimizes the network structure according to the changes in the ophthalmic disease data flow, thereby improving the model's feature extraction ability for the changing ophthalmic disease data characteristics. For the output of layer I in the fully connected neural network, a dynamic weight redistribution network is set, and the evaluation characteristics of the dynamic weight redistribution network of layer I of the fully connected neural network are calculated using the following formula: :

[0160] ;

[0161] in, It is a normalization function used to evaluate the activity or information content of the layer output;

[0162] The weight adjustment coefficient of the first layer of the fully connected neural network is calculated using the following formula: :

[0163] ;

[0164] in, and are the weights and biases of the dynamic weight redistribution network, which are used to generate the weight adjustment coefficients of each layer from the evaluation indicators. is the preset Softmax function, is the evaluation characteristic of the dynamic weight redistribution network of layer I of the fully connected neural network;

[0165] After the above weight adjustment, the output of the Ith layer after weight adjustment is calculated using the following formula: :

[0166] ;

[0167] in, To dynamically adjust the factor, the output of the fully connected neural network is adaptively amplified or suppressed according to the strength of the gradient, enhancing the sensitivity of the feature extraction model to important features;

[0168] A gradient-guided adaptive feature enhancement strategy is adopted to use the gradient information in the back-propagation process to adaptively enhance the important feature representation in the feature extraction model, thereby further improving the expression ability of the training data set for ophthalmic diseases. The gradient of the loss function is , through gradient-guided adaptive feature enhancement, dynamically adjust the factor The calculation method is expressed as:

[0169] ;

[0170] in, yes The modulus of is used to evaluate the influence of this layer in the training process;

[0171] S304: In the feature conversion process, regularization is used to achieve feature sparsification, which helps to remove redundant and unimportant features and retain the features that are most beneficial to subsequent tasks. The L1 norm is used for sparsification in the loss function of the fully connected neural network. The loss function of the fully connected neural network As shown below:

[0172] ;

[0173] in, is the target output, that is, the input sample and the corresponding true label vector; is the square of the L2 norm, is the L1 norm, which is used to achieve the sparsification of weights. is the regularization coefficient of the fully connected neural network, which regulates the degree of sparsity. is the mapping function from the layer output of the fully connected neural network to the predicted output, After weight adjustment, Layer output, is the sparse regularization term;

[0174] Sparse regularization term It is used to enhance the generalization ability of the feature extraction model. By forcing many elements of the weight vector to approach zero and summing the absolute values ​​of each weight element, the feature extraction model is driven to reduce unnecessary parameter complexity while maintaining the expression of key features. The calculation method is expressed as:

[0175] ;

[0176] in, is the weight matrix of the neural network. Row and Elements of a column;

[0177] In order to adapt to the needs of different training stages, an adaptive adjustment mechanism based on training error is adopted. The regularization parameter decays exponentially with the progress of iterations, aiming to provide strong regularization in the early stage of training to prevent overfitting, and gradually relax it in the later stage of training, allowing the feature extraction model to adjust the weights more freely to fine-tune the training data set of ophthalmic diseases. The regularization coefficient of the tth iteration of the fully connected neural network is The calculation method is expressed as:

[0178] ;

[0179] in, is the regularization strength of the initial fully connected neural network, Set to 0.1; is the decay rate of the fully connected neural network, is the current iteration number, is the total number of iterations;

[0180] S305: According to the loss function The calculated gradient is used to adjust the parameters of the fully connected neural network through the back propagation algorithm to optimize the hyperplane position and the weights of the fully connected neural network to reduce the prediction error and improve the generalization ability of the model. The parameter update of the fully connected neural network is expressed as follows: ;

[0181] in, is the parameter update learning rate of the fully connected neural network, which determines the step size of the parameter update. Set to 0.01; and are the weights and biases of the fully connected neural network after training the tth iteration, and are the gradients of the weights and biases of the fully connected neural network, For the The weights of the fully connected neural network at iterations, For the The weights of the fully connected neural network at iterations, For the The bias of the fully connected neural network for the iteration, For the The bias of the fully connected neural network for the iteration, is the symbol of partial derivative;

[0182] S306: Repeat the above steps S302-S305 until the preset stop iteration condition is met, and output the extracted features In this embodiment, the preset condition for stopping iteration is reaching a preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0183] S4: Using the autoencoder neural network based on feature refinement as the feature dimension reduction model to reduce the dimension of the extracted features and obtain the initial low-dimensional features of the ophthalmic disease after dimension reduction;

[0184] Features will be extracted The input is trained in the feature dimensionality reduction model, which includes an encoder, a decoder, and a feature adjustment module. The encoder is responsible for mapping high-dimensional input data to a low-dimensional feature space, and the decoder is used to reconstruct the reduced features back to the original space to ensure the reversibility of the dimensionality reduction process. The feature adjustment module dynamically adjusts the reduced feature space through recursive feature adaptive optimization, so that important features can be strengthened and minor features are gradually weakened, so that the reduced feature representation has better simplicity while retaining data information.

[0185] like Figure 2 As shown in Figure 2, the specific training steps of the feature dimensionality reduction model are as follows:

[0186] S401: The encoder uses a multi-layer nonlinear mapping structure to extract features Mapping from high-dimensional data to the initial low-dimensional feature space, the initial low-dimensional features It is expressed as:

[0187] ;

[0188] in, is the weight matrix of the encoder, is the encoder bias vector, is the multi-layer Sigmoid activation function of the encoder;

[0189] S402: After the initial low-dimensional features are generated, the feature adjustment module automatically generates feature weights according to the feature importance in the current feature space. When the feature adjustment module is initialized, the same initial weights are assigned to all features so that they can be gradually adjusted according to feature contributions in subsequent steps. The initial weight matrix The calculation method is expressed as:

[0190] ;

[0191] in, is the feature weight vector, Each element in Initialized to the same value, indicating that all initial low-dimensional features have the same importance in the initial stage. is a function for extracting the diagonal elements of a matrix;

[0192] Initial low-dimensional features of ophthalmic diseases after dimensionality reduction It is expressed as:

[0193] ;

[0194] S403: The feature adjustment module adjusts the initial low-dimensional features of ophthalmic diseases after dimensionality reduction Recursive optimization is performed. In each round of iteration, the feature adjustment module adjusts the weight of each feature according to the performance of the previous round of features, gradually enhances those features with important influence, and gradually weakens redundant or noise features. In the round of iteration, The feature weight of the round iteration is :

[0195] ;

[0196] in, Indicates The feature weights of the round iteration, is the learning rate of the feature dimensionality reduction model, is the loss function of the feature dimensionality reduction model, is the label data, Represents the gradient of the loss function with respect to the feature weight;

[0197] S404: To ensure the effectiveness of the dimensionality reduction process, the decoder remaps the low-dimensional features back to high-dimensional features to ensure that no important information is lost during the dimensionality reduction process. The decoder reconstructs the high-dimensional data for:

[0198] ;

[0199] in, is the weight matrix of the decoder, is the bias vector of the decoder, is the multi-layer Sigmoid activation function of the decoder;

[0200] S405: Repeat the iteration steps S402-S404 until the preset stop iteration condition is met, and output the initial low-dimensional features of the ophthalmic disease after dimensionality reduction ; In this embodiment, the preset condition for stopping iteration is reaching a preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0201] S5: Input the initial low-dimensional features of the ophthalmic disease after dimensionality reduction into the classifier model for training, and output the ophthalmic disease classification results; the present invention adopts the extreme learning machine algorithm based on inverse correction as the classifier model, and can also use support vector machine, random forest, decision tree or extreme learning machine as the classifier model.

[0202] In step S5, the classifier model includes an input layer, a hidden layer and an output layer. The input layer is used to receive feature vectors, the hidden layer is connected to the output layer through randomly generated weights, and the output layer performs final classification.

[0203] In this embodiment, the specific training steps of the classifier model are as follows:

[0204] S501: Initialize the hidden layer weights of the classifier model, randomly generate the connection weights of the hidden layer, and initialize the weights of the hidden layer It is expressed by the following formula:

[0205] ;

[0206] in, is the initialization scalar that controls the range of weights, Indicates the generation of a A random matrix is ​​used to provide initial weights for each hidden node;

[0207] S502: Initial low-dimensional features of ophthalmic diseases after dimensionality reduction Input into the classifier model to form a feature matrix, which is represented as a feature input vector :

[0208] ;

[0209] in, For the The characteristic value of the sample, is the feature dimension;

[0210] S503: Use activation function to input feature vector The present invention uses a nonlinear activation function to improve the expression ability of the classification model and obtain the output of the hidden layer. :

[0211] ;

[0212] in, is the activation function of the classifier model, is the bias term of the hidden layer;

[0213] S504: Output to hidden layer Perform inverse correction by calculating the output of the output layer The difference between the actual label and the hidden layer weights are adjusted inversely to reduce the prediction error. The prediction error of the classifier model during training As shown below:

[0214] ;

[0215] in, is the true label of the sample, ) is the Softmax function, Indicates the use of a function to perform category calculation on the hidden layer output of the classifier model;

[0216] The weight of the hidden layer is updated by the inverse correction strategy. It is expressed by the following formula:

[0217] ;

[0218] in, is the first input to the classifier model The dynamic importance of a feature, indicating the contribution of the feature in the current sample, is the first input to the classifier model eigenvalues, is the learning rate of the classifier model, Set to 0.001. The number of samples input to the classifier model for the current batch, is the first-order derivative of the activation function of the classifier model;

[0219] Use fuzzy logic to dynamically adjust feature weights: Not only consider the prediction error, but also dynamically adjust the feature weights according to the contribution of the feature to the prediction result, so as to achieve more refined separator model adjustment. The dynamic importance calculation method of the feature input to the separator model adjustment is expressed as:

[0220] ;

[0221] in, is the fuzzy logic evaluation function, Features The fuzzy factor reflects the sensitivity of the feature to the prediction result;

[0222] The calculation method of the fuzzy logic evaluation function is expressed as:

[0223] ;

[0224] in, Features The fuzzy factor, is the first input to the classifier model eigenvalues;

[0225] S505: According to the uncertainty of the prediction result of each sample by the classifier model, the weight of the feature is dynamically adjusted. For the features with uncertain classification results, a higher weight is given, so that the classifier model pays more attention to these challenging features during training. The method of dynamically adjusting the feature weight is expressed as follows:

[0226] ;

[0227] in, is the first input to the classifier model The weight of the samples, To adjust the uncertainty sensitivity parameter, is the threshold used to determine whether a sample is considered uncertain, For the classifier model The prediction error during the training process for samples; Set to 0.3, Set to 0.5.

[0228] S506: Calculate the final classification result by weighted calculation of the hidden layer output as the predicted output of the classifier model output layer :

[0229] ;

[0230] in, is the output layer weight of the extreme learning machine, is the output layer bias of the extreme learning machine;

[0231] S507: Output the prediction With the true label Compare, calculate the error, and feed the error back to the classifier model. Further iterate the training process, update the hidden layer weights and feature weights in each iteration to gradually improve the performance of the separator model. The final error is calculated and the feedback adjustment is expressed as: ;

[0232] in, is the adjustment amount of the output layer weight of the classifier model, is the activation function of the classifier model The second-order derivative of is used to provide more refined adjustment information. The updating process of the model can more comprehensively consider the shape and change trend of the loss function, thereby achieving more efficient convergence. is the true label of the sample;

[0233] S508: Repeat the above steps S502-S507 until the preset stop condition is met, and output the ophthalmic disease classification result. In this embodiment, the preset stop condition is reaching the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0234] S6: Input the data set into the trained feature extraction model and feature dimensionality reduction model for feature processing, and input the processed features into the trained classifier model for classification to obtain the prediction results of ophthalmic diseases.

[0235] In summary, this application solves the problem of poor model generalization ability caused by insufficient training samples by performing data expansion based on a generative adversarial network algorithm based on multi-sample resonance optimization, and adopts a positive and negative constraint mechanism to dynamically adjust the information entropy between generated samples and real samples to improve the quality of generated samples;

[0236] The neural network algorithm with dynamic weight redistribution is used for feature extraction to effectively capture the important features of ophthalmic disease data, solving the problem of insufficient adaptability of existing neural networks to dynamic environments in feature extraction. The feature extraction capability is improved by dynamically adjusting the weights of each layer.

[0237] The feature dimensionality reduction is performed by using an autoencoder neural network based on feature refinement, which solves the problem of information loss in the process of high-dimensional data processing, ensures that the dimensionality reduction process is reversible, and gradually adjusts the importance of features through recursive feature adaptive optimization to strengthen key features.

[0238] The extreme learning machine algorithm based on inverse correction is used as the classifier. The uncertainty problem in the classification process is solved by dynamically adjusting the sample weights, focusing on samples with higher uncertainty, and optimizing the learning efficiency and generalization ability of the model.

[0239] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An artificial intelligence-based ophthalmic disease prediction model training method, characterized by: The process includes the following steps: S1: Obtain a dataset of ophthalmic diseases, and annotate each image in the dataset with the disease; S2: using a generative adversarial network algorithm based on multi-sample resonance optimization to perform data expansion on the labeled data set to obtain a training data set for ophthalmic diseases; In step S2, a generative adversarial network algorithm based on multi-sample resonance optimization is used for data expansion. The specific steps are as follows: S201: Initialize the parameters of the generator and discriminator: The generator is responsible for generating nearly real ophthalmic disease data, and the discriminator tries to distinguish the generated samples from the real images. Let the generator be , the discriminator is , are the parameters of the generator, are the parameters of the discriminator; S202: In the adversarial training stage, a positive and negative constraint mechanism is used to optimize the training process by dynamically adjusting the weights in the loss function. The constraints are: ； in, Represents the value function of the generative adversarial network algorithm, which is the constraint condition in the adversarial training process. Represents real data, represents the noise sampled from the prior distribution, Indicates that it obeys the preset distribution. represents the distribution of real data, is the discriminator function, is the noise distribution, is the generator function, is the regularization weight coefficient of the generative adversarial network, is the regularization term of the generative adversarial network, represents the discriminator, represents a generator, express expectations; The regularization term of the generative adversarial network algorithm improves the generalization ability of the generative adversarial network algorithm by penalizing the sensitivity of the discriminator output to the input. The calculation method is expressed as: ； in, is the smoothing constant, is the number of samples input to the generative adversarial network algorithm in the current batch, It is Real data, represents the gradient of the real data, is the L2 norm, is the sparse loss term, is the sample sparsification weight, Represents the weight of the sparsification loss term, which controls the influence of the sparsification loss term on the total loss; An adaptive generated sample sparsification strategy is used to optimize the distribution of generated samples. The calculation method of the sparsification loss term is expressed as: ； in, It is The sparse weight of each data point, is the mean of the currently generated samples, is a preset constant added to prevent division by zero. It is The noise input to the generator; The calculation of the sparsification weight depends on the density estimation of the sample in the generated space, and its dynamic adjustment is expressed as: ； In the formula, is the sparsification weight adjustment factor, Is the measurement generated sample The function of the density between the currently generated sample sets, Indicates the currently generated sample set; S203: When the generator and the discriminator reach a confrontational balance, a multi-sample resonance optimization strategy is adopted, and the multi-sample resonance optimization strategy is set based on the resonance index Select samples for key training, and the resonance optimization function is expressed as: ； In the formula, represents the learning rate of the generator, represents the weight of resonance optimization, represents the resonance optimization function, Indicates a parameter update operation. represents the gradient with respect to the generator parameters, is the resonance standard sample; The resonance optimization function depends on the relationship between the generated samples of the generator and the resonance index. Specifically, the similarity between the generated samples and the resonance samples is enhanced through the exponentially decaying sum of square errors to increase the responsiveness of the model to data features. The calculation method of the exponentially decaying sum of square errors is expressed as: ； In the formula, is the adjustment factor that controls the resonance sensitivity, is the number of generated samples, It is Resonance standard samples, It is The noise input to the generator is represents the natural exponential function, represents the square of L2 norm; S204: repeating the above steps S202-S203 until a preset stop iteration condition is met, and outputting a training data set of ophthalmic diseases; S3: Using a neural network algorithm based on dynamic weight redistribution as a feature extraction model to extract features from the training data set of ophthalmic diseases, and obtaining the extracted features of ophthalmic diseases; In step S3, the feature extraction model is a fully connected neural network, and the specific training steps are as follows: S301: Initialize the parameters of the fully connected neural network: Use the following formula (1) to calculate the initial weight matrix of the neural network: , the following formula (2) is used to calculate the initial bias vector of the neural network : （1）； （2）； in, is the dimension of the feature space of the training dataset of ophthalmic diseases, is the dimension of the number of samples in the training data set for ophthalmic diseases. The randn() function is used to generate random numbers with a standard normal distribution. S302: In each iteration of the fully connected neural network, the training data set of ophthalmic diseases is forward propagated. The input of each layer is determined by the output of the previous layer and the parameters of the current layer. The data of each ophthalmic disease training data set is mapped to a new feature space defined by a hyperplane. The feature vector after feature extraction is expressed by the following formula ; ； in, The feature vector representing the training dataset of ophthalmic diseases input to the fully connected neural network, is the Sigmoid activation function, and are the weights and biases of the fully connected neural network after training the tth iteration; S303: Assumption For the output of layer I in the fully connected neural network, a dynamic weight redistribution network is set, and the evaluation characteristics of the dynamic weight redistribution network of layer I of the fully connected neural network are calculated using the following formula: : ； in, It is a normalization function used to evaluate the activity or information content of the layer output; The weight adjustment coefficient of the first layer of the fully connected neural network is calculated using the following formula: : ； in, and are the weights and biases of the dynamic weight redistribution network, used to generate the weight adjustment coefficients of each layer from the evaluation indicators, is the preset softmax function, is the evaluation characteristic of the dynamic weight redistribution network of layer I of the fully connected neural network; After the above weight adjustment, the output of the Ith layer after weight adjustment is calculated using the following formula: : ； in, To dynamically adjust the factor, the output of the fully connected neural network is adaptively amplified or suppressed according to the strength of the gradient, enhancing the sensitivity of the feature extraction model to important features; Adopting the gradient-guided adaptive feature enhancement strategy, using the gradient information in the back-propagation process, adaptively strengthen the important feature representation in the feature extraction model. The gradient of the loss function is , through gradient-guided adaptive feature enhancement, dynamically adjust the factor The calculation method is expressed as: ； in, yes The modulus of is used to evaluate the influence of this layer in the training process; S304: In the feature conversion process, the feature is sparsely processed by regularization, and the L1 norm is used for sparse processing in the loss function of the fully connected neural network. The loss function of the fully connected neural network is As shown below: ； in, is the target output, that is, the input sample and the corresponding true label vector; is the square of the L2 norm, is the L1 norm, which is used to achieve the sparsification of weights. is the regularization coefficient of the fully connected neural network, which regulates the degree of sparsity. is the mapping function from the layer output of the fully connected neural network to the predicted output, After weight adjustment, Layer output, is the sparse regularization term; Sparse regularization term Used to enhance the generalization ability of feature extraction models, The calculation method is expressed as: ； in, is the weight matrix of the neural network. Row and Elements of a column; Adopting an adaptive adjustment mechanism based on training error, the regularization parameter decays exponentially with the progress of iterations. The regularization coefficient of the tth iteration of the fully connected neural network is The calculation method is expressed as: ； in, is the regularization strength of the initial fully connected neural network, is the decay rate of the fully connected neural network, is the current iteration number, is the total number of iterations; S305: According to the loss function The calculated gradient is used to adjust the parameters of the fully connected neural network through the back propagation algorithm to optimize the hyperplane position and the weights of the fully connected neural network. The parameter update of the fully connected neural network is expressed as follows: ； in, is the parameter update learning rate of the fully connected neural network, which determines the step size of the parameter update. and are the weights and biases of the fully connected neural network after training the tth iteration, and are the gradients of the weights and biases of the fully connected neural network, For the The weights of the fully connected neural network at the iteration, For the The weights of the fully connected neural network at the iteration, For the The bias of the fully connected neural network for the iteration, For the The bias of the fully connected neural network for the iteration, is the symbol of partial derivative; S306: Repeat the above steps S302-S305 until the preset stop iteration condition is met, and output the extracted features ; S4: Using the autoencoder neural network based on feature refinement as the feature dimension reduction model to reduce the dimension of the extracted features and obtain the initial low-dimensional features of the ophthalmic disease after dimension reduction; The extracted features The input is put into the feature dimensionality reduction model for training. The feature dimensionality reduction model includes an encoder, a decoder and a feature adjustment module. The specific training steps of the feature dimensionality reduction model are as follows: S401: The encoder uses a multi-layer nonlinear mapping structure to extract features Mapping from high-dimensional data to the initial low-dimensional feature space, the initial low-dimensional features It is expressed as: ； in, is the weight matrix of the encoder, is the encoder bias vector, is the multi-layer Sigmoid activation function of the encoder; S402: After the initial low-dimensional features are generated, the feature adjustment module automatically generates feature weights according to the feature importance in the current feature space. The feature adjustment module assigns the same initial weights to all features during initialization so that they can be gradually adjusted according to feature contributions in subsequent steps. The initial weight matrix The calculation method is expressed as: ； in, is the feature weight vector, Each element in Initialized to the same value, indicating that all initial low-dimensional features have the same importance in the initial stage. is a function for extracting the diagonal elements of a matrix; Initial low-dimensional features of ophthalmic diseases after dimensionality reduction It is expressed as: ； S403: The feature adjustment module adjusts the initial low-dimensional features of the ophthalmic disease after dimensionality reduction Recursive optimization is performed. In each round of iteration, the feature adjustment module adjusts the weight of each feature according to the performance of the previous round of features, gradually enhances those features with important influence, and gradually weakens redundant or noise features. In the round of iteration, The feature weight of the round iteration is : ； in, Indicates The feature weights of the round iteration, is the learning rate of the feature dimensionality reduction model, is the loss function of the feature dimensionality reduction model, is the label data, Represents the gradient of the loss function with respect to the feature weight; S404: The decoder remaps the low-dimensional features back to high-dimensional features to ensure that no important information is lost during the dimensionality reduction process. The decoder reconstructs the high-dimensional data for: ； in, is the weight matrix of the decoder, is the bias vector of the decoder, is the multi-layer Sigmoid activation function of the decoder; S405: Repeat the iteration steps S402-S404 until the preset stop iteration condition is met, and output the initial low-dimensional features of the ophthalmic disease after dimensionality reduction ; S5: Input the initial low-dimensional features of the ophthalmic disease after dimensionality reduction into the classifier model for training, and output the ophthalmic disease classification results; S6: Inputting the data set into the trained feature extraction model and the feature dimension reduction model for feature processing, and inputting the processed features into the trained classifier model for classification to obtain ophthalmic disease prediction results.

2. The artificial intelligence-based ophthalmic disease prediction model training method according to claim 1, characterized in that: In step S5, the classifier model includes an input layer, a hidden layer and an output layer. The specific training steps of the classifier model are as follows: S501: Initialize the hidden layer weights of the classifier model, randomly generate the connection weights of the hidden layer, and initialize the weights of the hidden layer It is expressed by the following formula: ； in, is the initialization scalar that controls the range of weights, Indicates the generation of a A random matrix is ​​used to provide initial weights for each hidden node; S502: Initial low-dimensional features of ophthalmic diseases after dimensionality reduction Input into the classifier model to form a feature matrix, which is represented as a feature input vector : ； in, For the The characteristic value of the sample, is the feature dimension; S503: Use activation function to input feature vector Processing to get the output of the hidden layer : ； in, is the activation function of the classifier model, is the bias term of the hidden layer; S504: Output to hidden layer Perform inverse correction by calculating the output of the output layer The difference between the actual label and the hidden layer weights are adjusted inversely to reduce the prediction error. The prediction error of the classifier model during training As shown below: ； in, is the true label of the sample, ) is the Softmax function, Indicates the use of functions to perform category calculations on the hidden layer outputs of the classifier model; The weight of the hidden layer is updated by the inverse correction strategy. It is expressed by the following formula: ； in, is the first input to the classifier model The dynamic importance of a feature indicates the contribution of the feature in the current sample. is the first input to the classifier model eigenvalues, is the learning rate of the classifier model, The number of samples input to the classifier model for the current batch, is the first-order derivative of the activation function of the classifier model; Dynamically adjust feature weights using fuzzy logic: The dynamic importance of features input to the separator model adjustment is calculated as: ； in, is the fuzzy logic evaluation function, Features The fuzzy factor reflects the sensitivity of the feature to the prediction result; The calculation method of the fuzzy logic evaluation function is expressed as: ； in, Features The fuzzy factor, is the first input to the classifier model eigenvalues; S505: According to the uncertainty of the prediction result of each sample by the classifier model, the weight of the feature is dynamically adjusted. For the features with uncertain classification results, a higher weight is given, so that the classifier model pays more attention to these challenging features during training. The method of dynamically adjusting the feature weight is expressed as follows: ； in, is the first input to the classifier model The weight of the samples, To adjust the uncertainty sensitivity parameter, is the threshold used to determine whether a sample is considered uncertain, For the classifier model The prediction error during the training process for samples; S506: Calculate the final classification result by weighted calculation of the hidden layer output as the predicted output of the classifier model output layer : ； in, is the output layer weight of the extreme learning machine, is the output layer bias of the extreme learning machine; S507: Output the prediction With the true label Compare, calculate the error, and feed the error back to the classifier model. Further iterate the training process, update the hidden layer weights and feature weights in each iteration to gradually improve the performance of the separator model. The final error is calculated and the feedback adjustment is expressed as: ; in, is the adjustment amount of the output layer weight of the classifier model, is the activation function of the classifier model The second-order derivative of is the true label of the sample; S508: Repeat the above steps S502-S507 until the preset stop condition is met, and output the ophthalmic disease classification result.

3. The artificial intelligence-based ophthalmic disease prediction model training method according to claim 1, characterized in that: In step S2, the labeled data set is expanded by adding noise, vector concatenation or vector translation.

4. The artificial intelligence-based ophthalmic disease prediction model training method according to claim 1, characterized in that: In step S5, a support vector machine, a random forest, a decision tree or an extreme learning machine is used as a classifier model.

Images