A method and system for assessing the effectiveness of vision training

Abstract

The application provides a vision training effect evaluation method and system, the method comprising: obtaining sample vision training effect data; inputting the sample vision training effect data into a preset data expansion model to obtain expanded data; constructing a classifier based on a Monte Carlo algorithm and a preset decision tree model, and constructing an initial evaluation model based on the classifier; training the initial evaluation model according to the expanded data and the sample vision training effect data to obtain a target evaluation model; obtaining to-be-tested vision training effect data, and inputting the vision training effect data into the target evaluation model to enable the target evaluation model to perform feature processing on the to-be-tested vision, obtain to-be-tested feature data, and then enable the target evaluation model to classify the to-be-tested feature data based on the classifier to obtain a vision training effect evaluation result, thereby improving the accuracy of vision training effect evaluation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vision assessment, in particular to a vision training effect evaluation method and system. BACKGROUND

[0002] Currently, the evaluation of vision training effect is increasingly valued, which has a wide range of applications, for example: for patients with myopia, hyperopia, astigmatism, amblyopia, etc., the evaluation of vision training effect can well judge the improvement degree of a series of visual activities and exercises on their vision. However, the traditional vision training effect evaluation method mainly relies on subjective judgment or simple vision test, and these methods cannot fully reflect the training effect, and lack of precision and objectivity. With the development of artificial intelligence technology, especially the application of machine learning and deep learning, it provides a new way to deal with the complex data analysis problem of vision training effect evaluation. However, there are still limitations in the model construction process of traditional artificial intelligence, for example: it may not be able to fully simulate the real data distribution when data augmentation, resulting in large training sample bias; it is too conservative in structure selection, etc. All of these will lead to the final evaluation result not accurate enough. SUMMARY

[0003] The present application aims to provide a vision training effect evaluation method and system to solve the above technical problems and improve the accuracy of vision training effect evaluation.

[0004] In order to solve the above technical problems, the present application provides a vision training effect evaluation method, comprising the following steps:

[0005] Obtain sample vision training effect data;

[0006] Input the sample vision training effect data into a preset data augmentation model to obtain augmented data;

[0007] Based on the Monte Carlo algorithm and a preset decision tree model, a classifier is constructed, and an initial evaluation model is constructed based on the classifier;

[0008] According to the augmented data and the sample vision training effect data, the initial evaluation model is trained to obtain a target evaluation model;

[0009] Obtain the vision training effect data to be tested, and input the vision training effect data into the target evaluation model, so that the target evaluation model processes the features of the vision training effect data to be tested to obtain feature data to be tested, and then the target evaluation model classifies the feature data to be tested based on the classifier to obtain the vision training effect evaluation result.

[0010] In the scheme, the sample vision training effect data is expanded by the data expansion model to improve the quantity and authenticity of the data used for training the initial evaluation model; based on the Monte Carlo algorithm and the preset decision tree model, a classifier is constructed, and the initial evaluation model is constructed based on the classifier, so as to explore more possible tree structures, so that the model structure of the constructed classifier has more accurate vision training effect prediction ability, and the prediction accuracy of the initial evaluation model is further improved; the initial evaluation model is trained to obtain a target evaluation model, and the to-be-tested vision training effect data is input into the target evaluation model to obtain a more accurate vision training effect evaluation result.

[0011] In the scheme, the sample vision training effect data is expanded by the data expansion model to expand the quantity and authenticity of the sample vision training effect data; the structure of the decision tree is selected from the preset extreme gradient boosting decision tree model by the Monte Carlo algorithm, which can explore a wider model structure compared with the traditional decision tree structure selection method relying on the greedy algorithm, and improves the generalization ability of the model. The above all help to improve the accuracy of vision training effect evaluation.

[0012] Further, the data expansion model includes a generative adversarial network and a variational autoencoder, and the input of the sample vision training effect data into the preset data expansion model to obtain expanded data includes: performing data generation processing on the sample vision training effect data by the generative adversarial network to generate synthetic data; performing automatic differentiation and variational inference on the synthetic data by the variational autoencoder to obtain the expanded data.

[0013] In the scheme, the sample vision training effect data is expanded by the data expansion model to expand the quantity and authenticity of the sample vision training effect data; the structure of the decision tree is selected from the preset extreme gradient boosting decision tree model by the Monte Carlo algorithm, which can explore a wider model structure compared with the traditional decision tree structure selection method relying on the greedy algorithm, and improves the generalization ability of the model. The above all help to improve the accuracy of vision training effect evaluation.

[0014] Further, the input of the sample vision training effect data into the preset data expansion model to obtain expanded data includes: the training process of the generative adversarial network is that the preset initial generative adversarial network is trained based on the sample vision training effect data to obtain the generative adversarial network, and then the intermediate training data is obtained based on the generative adversarial network; the training process of the variational autoencoder is that the preset initial variational autoencoder is trained based on the intermediate training data to obtain the variational autoencoder.

[0015] In the scheme, because the purpose of the variational autoencoder is to perform automatic differentiation and variational inference on the synthetic data generated by the generative adversarial network, in the training process, the preset initial variational autoencoder is used to train the intermediate training data output by the generative adversarial network, avoiding separate training of different data sources, ensuring the unity of the training process, and improving the reliability of the finally obtained generative adversarial network and variational autoencoder.

[0016] Further, the preset initial generative adversarial network includes a generator and a discriminator, and the preset initial generative adversarial network is trained based on the sample visual training effect data to obtain the generative adversarial network, and then the intermediate training data is obtained based on the generative adversarial network, including: initializing parameters of the generator and parameters of the discriminator; inputting a preset random noise vector into the generator to generate fake data; constructing a loss function of the discriminator based on the fake data, the sample visual training effect data and the parameters of the discriminator; constructing a loss function of the generator based on the random noise vector, the parameters of the generator and the parameters of the discriminator; updating the parameters of the discriminator based on the gradient of the loss function of the discriminator, so that the preset initial generative adversarial network updates the parameters of the generator based on the gradient of the loss function of the generator, and then the loss function of the generator and the loss function of the discriminator converge, and the updating of the parameters of the generator and the parameters of the discriminator is stopped when the loss function of the generator and the loss function of the discriminator converge simultaneously, and the generative adversarial network is obtained.

[0017] In the scheme, the generator and the discriminator are trained in an adversarial manner, and they influence each other, so that the finally trained generator and discriminator have high precision, and the generated intermediate training data is closer to the real data.

[0018] Further, when the loss function of the generator and the loss function of the discriminator converge simultaneously, the updating of the parameters of the generator and the parameters of the discriminator is stopped, and the generative adversarial network is obtained, including: constructing a dual loss function based on the loss function of the generator and the loss function of the discriminator; updating the parameters of the discriminator and the parameters of the generator in the dual loss function based on the gradient of the dual loss function, and stopping the updating of the parameters of the generator and the parameters of the discriminator when the dual loss function converges, and obtaining the generative adversarial network.

[0019] In the scheme, a dual loss function is constructed based on the loss function of the generator and the loss function of the discriminator, that is, the training intensity of the generator and the discriminator is balanced by the dual loss function to prevent one party from being too strong to cause training imbalance, so as to obtain a more accurate generative adversarial network, so that the intermediate training data output by the generative adversarial network is more diverse and similar to the real sample visual training effect data in the overall data distribution.

[0020] Further, the preset initial variational autoencoder includes an encoder and a decoder, and the training of the preset initial variational autoencoder based on the intermediate training data includes: constructing a loss function of the preset initial variational autoencoder based on the intermediate training data, parameters of the encoder, parameters of the decoder and KL divergence; updating the parameters of the encoder and the parameters of the decoder based on the gradient of the loss function of the preset initial variational autoencoder, respectively, and stopping the update of the parameters of the encoder and the parameters of the decoder when the loss function of the variational autoencoder converges, to obtain the variational autoencoder.

[0021] In the scheme, during the training of the preset initial variational autoencoder, the KL divergence constraint is used to make the variational autoencoder finally trained to generate expansion data with a distribution closer to the real data distribution and more uniformly covering the entire data space.

[0022] Further, the construction of the classifier based on the Monte Carlo algorithm and the preset decision tree model and the construction of the initial evaluation model based on the classifier include: initializing the preset decision tree model; performing upper confidence bound exploration on the visual training effect features in the sample visual training effect data and the expansion data by the Monte Carlo algorithm to obtain a target feature combination; performing segmentation point evaluation on the visual training effect features in the target feature combination by the Monte Carlo algorithm to obtain a target segmentation point; updating the preset decision tree model based on the target feature combination and the target segmentation point to obtain the classifier.

[0023] In the scheme, because the traditional decision tree model usually relies on a greedy algorithm in structure selection, it ignores the exploration of a better model structure and limits the generalization ability of the model. Therefore, the scheme explores more possible tree structures by the Monte Carlo algorithm, thereby improving the generalization ability of the model.

[0024] Further, the updating of the preset decision tree model based on the target feature combination and the target segmentation point comprises: obtaining a pruning value of each decision tree node in the preset decision tree model based on a preset pruning algorithm; and when the pruning value is less than a preset pruning threshold, pruning the decision tree node to obtain the classifier.

[0025] In the above scheme, in order to prevent overfitting, each decision tree node in the preset decision tree model is pruned to remove branches that do not greatly improve the performance of the preset decision tree model, thereby deleting tree branches that do not significantly improve the performance of the model, thereby improving the stability of the preset decision tree model.

[0026] Further, the target evaluation model comprises a feature extraction sub-model and a feature dimension reduction sub-model, and the target evaluation model is used to perform feature processing on the to-be-tested vision to obtain to-be-tested feature data, which comprises: performing feature extraction on the to-be-tested vision training effect data by the feature extraction sub-model to obtain initial vision features; and performing feature dimension reduction on the initial vision features by the feature dimension reduction sub-model to obtain the to-be-tested feature data.

[0027] In the above scheme, the feature extraction sub-model performs feature extraction on the to-be-tested vision training effect data to obtain initial vision features, and the feature dimension reduction sub-model performs feature dimension reduction on the initial vision features to obtain the to-be-tested feature data, thereby improving the quality and efficiency of feature representation, and improving the accuracy of subsequent evaluation and classification of the to-be-tested feature data by the classifier.

[0028] The application further provides a vision training effect evaluation system, which comprises: a first acquisition module configured to acquire sample vision training effect data; an input module configured to input the sample vision training effect data into a preset data expansion model to obtain expanded data; a construction module configured to construct a classifier based on a Monte Carlo algorithm and a preset decision tree model, and construct an initial evaluation model based on the classifier; a training module configured to train the initial evaluation model based on the expanded data and the sample vision training effect data to obtain a target evaluation model; and an evaluation module configured to acquire to-be-tested vision training effect data, input the vision training effect data into the target evaluation model, so that the target evaluation model performs feature processing on the to-be-tested vision to obtain to-be-tested feature data, and then makes the target evaluation model classify the to-be-tested feature data based on the classifier to obtain a vision training effect evaluation result.

[0029] The system provided by the scheme has the data expansion model expanding the sample vision training effect data, and the quantity and authenticity of the sample vision training effect data are expanded; the structure of the decision tree is selected from the preset extreme gradient boosting decision tree model through the Monte Carlo algorithm, compared with the traditional decision tree structure selection method which depends on the greedy algorithm, a wider model structure can be explored, and the generalization ability of the model is improved. The above all help to improve the accuracy of the vision training effect evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 A flowchart of a vision training effect evaluation method is provided for an embodiment of the present application.

[0031] Figure 2 A system architecture diagram of a vision training effect evaluation system is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0033] Please refer to Figure 1 The present embodiment provides a vision training effect evaluation method, comprising the following steps:

[0034] Step S1: obtaining sample vision training effect data;

[0035] Step S2: inputting the sample vision training effect data into a preset data expansion model to obtain expanded data;

[0036] Step S3: constructing a classifier based on the Monte Carlo algorithm and a preset decision tree model, and constructing an initial evaluation model based on the classifier;

[0037] Step S4: training the initial evaluation model according to the expanded data and the sample vision training effect data to obtain a target evaluation model;

[0038] Step S5: obtaining to-be-tested vision training effect data, and inputting the vision training effect data into the target evaluation model to enable the target evaluation model to perform feature processing on the to-be-tested vision, to obtain to-be-tested feature data, and then enabling the target evaluation model to classify the to-be-tested feature data based on the classifier to obtain a vision training effect evaluation result.

[0039] In this embodiment, the sample vision training effect data is expanded by a data augmentation model to increase the quantity and authenticity of the data used to train the initial evaluation model. Based on the Monte Carlo algorithm and a preset decision tree model, a classifier is constructed, and the initial evaluation model is built based on the classifier to explore more possible tree structures. This makes the constructed classifier model structure have a more accurate vision training effect prediction capability, further improving the prediction accuracy of the initial evaluation model. The initial evaluation model is trained to obtain the target evaluation model, and then the vision training effect data to be tested is input into the target evaluation model to obtain a more accurate vision training effect evaluation result.

[0040] In this embodiment, the data augmentation model expands the sample vision training effect data, increasing the quantity and accuracy of the data. The Monte Carlo algorithm selects the decision tree structure from a pre-defined extreme gradient boosting decision tree model. Compared to traditional decision tree structure selection methods that rely on greedy algorithms, this approach explores a wider range of model structures, improving the model's generalization ability. All of these contribute to improving the accuracy of vision training effect evaluation.

[0041] It should be noted that the sample visual acuity training effect data can be derived from actual records of clinical amblyopia training and training data simulated under laboratory conditions, and the acquired data is stored in vector form. The features in the sample visual acuity training effect data can be divided into multiple categories. For example, in the evaluation of visual acuity training effect for amblyopia, a1 represents the visual acuity level before training; a2 represents the visual acuity level after training; a3 represents the training duration; a4 represents the training frequency; a5 represents the training intensity; a6 represents the training method; a7 represents the patient's age; a8 represents the patient's gender; a9 represents the patient's initial amblyopia degree; a 10 This indicates other auxiliary treatment methods, etc. The visual acuity training effect data of the samples are manually labeled; for example, three evaluation categories are labeled: significant training effect, moderate training effect, and insignificant training effect.

[0042] Furthermore, the data augmentation model includes a generative adversarial network and a variational autoencoder. Sample vision training data is input into the pre-defined data augmentation model to obtain augmented data, including:

[0043] Generative adversarial networks are used to process sample vision training effect data to generate synthetic data.

[0044] Augmented data is obtained by performing automatic differentiation and variational inference on the synthetic data using a variational autoencoder.

[0045] It should be noted that the sample vision training effect data is used to train the initial evaluation model, but the amount of sample vision training effect data obtained is insufficient for model training, which easily leads to poor generalization ability of the target evaluation model obtained by training, and affects the evaluation accuracy of the target evaluation model. Therefore, the vision training effect data needs to be expanded in data amount. The application expands the vision training effect data through a data expansion model: inputting the sample vision training effect data into a preset data expansion model, generating synthetic data close to the sample vision training effect data through a generative adversarial network for data generation processing, and using the synthetic data together with the sample vision training effect data to train the initial evaluation model, which not only ensures the sufficiency of the data amount, but also ensures the authenticity of the data. However, the data distribution of the synthetic data generated by the generative adversarial network is uncertain, which may be limited to a certain category, which will lead to inaccurate evaluation of the target evaluation model finally constructed, therefore, it is still necessary to automatically differentiate and variational inference the synthetic data through the variational autoencoder, so that the expansion data generated is closer to the distribution of the real data, further improving the authenticity of the expansion data. The generative adversarial network optimizes the authenticity of the synthetic data generated, and the variational autoencoder optimizes the data distribution matching of the expansion data generated, both of which improve the quality of the expansion data.

[0046] Further, inputting the sample vision training effect data into the preset data expansion model obtains expansion data, including:

[0047] The training process of the generative adversarial network is:

[0048] Based on the sample vision training effect data, the preset initial generative adversarial network is trained to obtain the generative adversarial network, and then the intermediate training data is obtained based on the generative adversarial network;

[0049] The training process of the variational autoencoder is:

[0050] Based on the intermediate training data, the preset initial variational autoencoder is trained to obtain the variational autoencoder.

[0051] It should be noted that the processing object of the variational autoencoder is the output result of the generative adversarial network, therefore, in the training process, the preset initial variational autoencoder needs to obtain the intermediate training data based on the generative adversarial network, and then train the preset initial variational autoencoder based on the intermediate training data. The training of the preset initial generative adversarial network and the preset initial variational autoencoder is repeated and iterated until the preset stop iteration condition is met, that is, the data expansion model training is completed.

[0052] Further, the preset initial generative adversarial network includes a generator and a discriminator, the preset initial generative adversarial network is trained based on the sample vision training effect data, a generative adversarial network is obtained, and then intermediate training data is obtained based on the generative adversarial network, including:

[0053] parameters of the generator and parameters of the discriminator are initialized;

[0054] preset random noise vectors are input into the generator to generate fake data;

[0055] a loss function of the discriminator is constructed based on the fake data, the sample vision training effect data and the parameters of the discriminator;

[0056] a loss function of the generator is constructed based on the random noise vectors, the parameters of the generator and the parameters of the discriminator;

[0057] the parameters of the discriminator are updated based on the gradient of the loss function of the discriminator, so that the preset initial generative adversarial network updates the parameters of the generator based on the gradient of the loss function of the generator, and then the loss function of the generator and the loss function of the discriminator are converged, the update of the parameters of the generator and the parameters of the discriminator is stopped when the loss function of the generator and the loss function of the discriminator are simultaneously converged, and a generative adversarial network is obtained.

[0058] It should be noted that the preset initial generative adversarial network includes a generator (G) and a discriminator (D), wherein the generator is responsible for generating fake data, and the discriminator is responsible for distinguishing real data and fake data. In the training process, the generator is continuously optimized to make the generated fake data deceive the discriminator, and the discriminator continuously improves the discrimination ability to distinguish whether the fake data output by the generator is real data, that is, the generator and the discriminator are in an adversarial training relationship, and they influence each other, so that the final generated data is closer to the real data. Specifically: the parameters of the generator and the discriminator can be initialized by random weights, and then preset random noise vectors are input into the generator to generate fake data, wherein the random noise vectors can be sampled from a Gaussian distribution, and the random noise vectors can be expressed as:

[0059] bz~N(0,I);

[0060] In the formula, bz is a random noise vector, N represents a normal distribution, 0 is a mean vector, and I is an identity matrix.

[0061] The fake data can be expressed as:

[0062] bx=G(bz;b G );

[0063] wherein, bx is the generated fake data, G() is an existing generation function, bz is a random noise vector, b G is a parameter of the generator. Exemplarily, in a sample vision training effect data augmentation task for amblyopia, the generator can be used to simulate the change of the vision level of different patients after training, and generate fake data similar to the sample vision training effect data with different training time, training method, and auxiliary treatment, etc.

[0064] It should be noted that based on the parameters of the fake data, the sample vision training effect data and the discriminator, a loss function of the discriminator is constructed, and exemplarily, the loss function of the discriminator can be represented as:

[0065]

[0066] wherein, L D is the loss function of the discriminator, m is the batch size, D() is an existing discrimination function, bx real is the sample vision training effect data, bx is the generated fake data, b D is a parameter of the discriminator. In the training process, the data set corresponding to the sample vision training effect data is divided into multiple small subsets, and the batch represents the subset sequence number.

[0067] It should be noted that based on the random noise vector, the parameters of the generator and the discriminator, a loss function of the generator is constructed, and exemplarily, the loss function of the generator can be represented as:

[0068]

[0069] wherein, L G is the loss function of the generator, m is the batch size, D() is the discrimination function, G() is the generation function, bz is the random noise vector, b G is a parameter of the generator, b D is a parameter of the discriminator.

[0070] It should be noted that based on the gradient of the loss function of the discriminator, the parameters of the discriminator are updated, and exemplarily, the parameters of the discriminator can be updated by back propagation, represented as:

[0071]

[0072] wherein, η is a learning rate, is the gradient of the loss function of the discriminator. The network parameters are adjusted in the opposite direction by the learning rate, so that the discrimination result is more accurate.

[0073] It should be noted that based on the gradient of the loss function of the generator, the parameters of the generator are updated, and exemplarily, the parameters of the generator can be updated by back propagation, represented as:

[0074]

[0075] wherein η is a learning rate, is the gradient of the generator loss function. The generator adjusts its parameters in reverse according to the error gradient fed back by the discriminator, so that the generated fake data is closer to the real sample vision training effect data considered by the discriminator. Then the parameters of the discriminator and the generator are continuously updated, so that the loss function of the generator and the loss function of the discriminator converge, and the update of the parameters of the generator and the parameters of the discriminator is stopped when the loss function of the generator and the loss function of the discriminator converge, and the generative adversarial network is obtained.

[0076] Further, the parameters of the generator and the parameters of the discriminator are stopped when the loss function of the generator and the loss function of the discriminator converge, and the generative adversarial network is obtained, comprising:

[0077] Based on the loss function of the generator and the loss function of the discriminator, a dual loss function is constructed;

[0078] Based on the gradient of the dual loss function, the parameters of the discriminator and the parameters of the generator in the dual loss function are updated, and the update of the parameters of the generator and the parameters of the discriminator is stopped when the dual loss function converges, and the generative adversarial network is obtained.

[0079] It should be noted that the dual loss function obtained based on the loss function of the generator and the loss function of the discriminator can be expressed as:

[0080] L GD =L G +λ vg L D ;

[0081] wherein L GD is a joint loss function, L G is a loss function of the generator, L D is a loss function of the discriminator, and λ vg is a balance coefficient. Exemplarily, λ vg is set to 0.3. The dual loss function L GD is used when the parameters of the generator and the discriminator are updated jointly, and balances the training strength of the two, preventing one side from being too strong and causing training imbalance. By balancing the training of the generator and the discriminator, a more accurate generative adversarial network is obtained, so that the output intermediate training data has more diversity, not only similar to the real sample vision training effect data in overall distribution, but also consistent with different vision training methods of different types of patients.

[0082] Further, the preset initial variational autoencoder comprises an encoder and a decoder, the preset initial variational autoencoder is trained based on the intermediate training data, and a variational autoencoder is obtained, comprising:

[0083] Based on the intermediate training data, the parameters of the encoder, the parameters of the decoder and the KL divergence, a loss function of the preset initial variational autoencoder is constructed;

[0084] Based on the gradient of the loss function of the preset initial variational autoencoder, the parameters of the encoder and the parameters of the decoder are updated respectively, and the updating of the parameters of the encoder and the parameters of the decoder is stopped when the loss function of the variational autoencoder converges, and the variational autoencoder is obtained.

[0085] It should be noted that the preset initial variational autoencoder is used to perform automatic differentiation and variational inference on the intermediate training data output by the generative adversarial network, and the KL divergence constraint of the variational autoencoder is used to make the variational autoencoder finally trained to generate expansion data with a distribution closer to the real data distribution and more uniformly covering the entire data space. Specifically, the loss function of the preset initial variational autoencoder can be expressed as:

[0086]

[0087] In the formula, L VAE is the loss function of the preset initial variational autoencoder, represents an existing expectation function, b enc is the parameter of the encoder, b dec is the parameter of the decoder, bx is the intermediate training data generated by the generator, q(b enc |bx) is the posterior distribution of the encoder, p(bx|b dec ) is the generation distribution of the decoder, D KL is the KL divergence, wherein the KL divergence is used to represent the difference between the posterior distribution of the encoder and the generation distribution of the decoder.

[0088] It should be noted that based on the gradient of the loss function of the preset initial variational autoencoder, the parameters of the encoder and the parameters of the decoder are updated respectively, and the updating of the parameters of the encoder and the parameters of the decoder is stopped when the loss function of the variational autoencoder converges, and the variational autoencoder is obtained. Illustratively, the parameters of the encoder and the decoder are updated by back propagation, which can be expressed as:

[0089]

[0090] In the formula, η bf is the learning rate, is the gradient of the loss function of the variational autoencoder with respect to the parameters of the encoder, is the gradient of the variational autoencoder loss function with respect to the decoder parameters. Exemplarily, η bf is set to 0.02. The encoder and the decoder update the parameters according to the variational loss gradient with a fixed learning rate, so that the finally generated data conforms to the original distribution and retains key features.

[0091] Further, based on the Monte Carlo algorithm and the preset decision tree model, a classifier is constructed, and an initial evaluation model is constructed based on the classifier, including:

[0092] The preset decision tree model is initialized;

[0093] The vision training effect features in the sample vision training effect data and the augmented data are explored by the Monte Carlo algorithm, and a target feature combination is obtained;

[0094] The vision training effect features in the target feature combination are evaluated by the Monte Carlo algorithm, and a target split point is obtained;

[0095] Based on the target feature combination and the target split point, the preset decision tree model is updated to obtain the classifier.

[0096] It should be noted that the traditional decision tree model usually relies on the greedy algorithm in structure selection, so it ignores the exploration of better model structures and limits the generalization ability of the model. In this scheme, the Monte Carlo algorithm is used to explore more possible tree structures, thereby improving the generalization ability of the model. Specifically, the parameters of the preset decision tree model are initialized, and exemplarily, the preset decision tree model can adopt an extreme gradient boosting decision tree model. The parameters of the preset decision tree model include a learning rate, a maximum depth of a tree, and a maximum number of trees, and the preset decision tree model can be represented as:

[0097] θ (0) =(λ q ,Y q ,α q ,σ q );

[0098] In the formula, θ (0) is the preset decision tree model, λ q represents the learning rate; γ q represents the maximum depth of the tree; α q represents the sum of the minimum leaf node sample weights; and σ q represents the maximum number of trees. Preferably, the learning rate λ q of the extreme gradient boosting decision tree model is set to 0.1, the maximum depth γ q of the tree is set to 6, and the maximum number σ q of trees is set to 100.

[0099] It should be noted that, since the sample vision training effect data and the augmented data have many vision training effect features, exemplarily including “vision level before and after training, training duration, training frequency, training method, patient age” and the like, it is necessary to screen the most valuable vision training effect features for evaluation classification, so as to improve the evaluation accuracy of the initial evaluation model. In the present application, the upper confidence bound of different vision training effect features in the sample vision training effect data and the augmented data is explored by the Monte Carlo algorithm, so as to select the optimal target feature combination. Specifically, for each vision training effect feature, the contribution of the feature to the model prediction ability is evaluated, and the feature combination with the largest model performance improvement is selected. The evaluation function of the process can be expressed as:

[0100]

[0101] In the formula, U(bm) is the upper confidence bound of the vision training effect feature bm; represents the average improvement effect of the vision training effect feature bm in the historical simulation; c is a preset exploration parameter for balancing exploration and utilization. In this task, since some vision training effect features, such as patient age, may have a greater impact on the training effect, the adjustment of the exploration parameter ensures that the preset decision tree model can not only utilize known effective vision training effect features such as training time, but also try new vision training effect features such as training intensity; N is the total number of simulations; n(bm) is the number of times the vision training effect feature bm is simulated.

[0102] In the formula, the historical simulation utilizes the vision training effect feature contribution in the historical data for evaluation. For example, if “training frequency” always has a greater impact on the result in the past evaluation tasks, the probability of being selected will be higher, and therefore, The calculation method of can be expressed as:

[0103]

[0104] In the formula, ΔL j (bm) represents the loss reduction amount brought by using the vision training effect feature bm as the segmentation feature in the jth simulation.

[0105] It should be noted that the evaluation classification of the visual training effect involves multiple continuous variables such as visual changes, training time, etc., and therefore how to find the best segmentation point is the key to determining the effect of evaluation classification. The Monte Carlo method calculates the expected total loss reduction by simulating different segmentation points, and selects the optimal segmentation scheme. For each decision tree in the preset decision tree model, the Monte Carlo method is used to determine the best segmentation point, and the selection of the segmentation point directly affects the evaluation classification result in the amblyopia evaluation data. For example: if the likelihood of visual improvement significantly increases after the "training time" exceeds a certain threshold, a segmentation point can be set at that point; by calculating the loss reduction, the boundary point of significant visual improvement is found to better predict the training effect. Therefore, the target segmentation point is obtained by evaluating the segmentation point of the visual training effect feature in the target feature combination through the Monte Carlo algorithm. Specifically, for each candidate segmentation point s, the expected total loss reduction can be represented as:

[0106]

[0107] In the formula, ΔL(s) is the expected total loss of the segmentation point s, G L and G R represent the gradients of the left and right child nodes respectively; H L and H R represent the second derivatives of the left and right child nodes respectively; λ q is a regularization parameter. The child nodes in the decision tree are the visual training effect features in the target feature combination. According to the comparison between the expected total loss and the preset threshold, the target segmentation point is obtained.

[0108] It should be noted that based on the target feature combination and the target segmentation point, each decision tree in the preset decision tree model is updated. In each iteration, the loss of the decision tree in the current preset decision tree model is calculated, including the classification error and the consideration of model complexity. Further, the parameters of the decision tree are adjusted to minimize this loss, so as to find a balance between the accuracy and complexity of the decision tree. The calculation method of the loss function can be represented as:

[0109]

[0110] In the formula, is the loss between the predicted value and the true value output by the decision tree in the preset decision tree model; Ω(θ) is the complexity of the current preset decision tree model; n is the number of samples input in batches; λ qis a preset regularization coefficient, used to adjust the balance between model complexity and fitting degree. In the sample vision training effect data and the expanded data, some vision training effect features may cause the model to overfit, for example, age may perform as a very strong feature in the training set, but it is not stable in the test set. Therefore, the complexity is controlled by the regularization parameter to prevent overfitting.

[0111] In one embodiment, The calculation method of the logical loss can be represented as:

[0112]

[0113] In the formula, y i is the actual label of the i-th data point; is the probability predicted by the decision tree, which is calculated by a logical function.

[0114] In this embodiment, the model complexity Ω(θ) includes the number of trees and the depth of the trees in the preset decision tree model, and the calculation method can be represented as:

[0115]

[0116] In the formula, T is the total number of trees in the preset decision tree model; ω t is the vector of leaf node values of the t-th tree; γ q and λ q are parameters for controlling the complexity of the tree, respectively.

[0117] It should be noted that after training each decision tree in the preset decision tree model, all the constructed decision trees are fused, and the outputs of all the decision trees are fused by weighting, and the weight w t is calculated as:

[0118] w t = exp(-α q ·R(t));

[0119] In the formula, exp() is an exponential function with a natural constant as the base; R(t) is the loss of decision tree t on the validation set; α q is a weight decay factor. Preferably, the weight decay factor α q is set to 0.9. The calculation method of R(t) can be represented as:

[0120]

[0121] In the formula, G t is the sum of all sample gradients of the node; H t is the sum of all second-order derivatives of the node, and λ qa preset regularization parameter.

[0122] Further, based on the target feature combination and the target segmentation point, the preset decision tree model is updated to obtain a classifier, including:

[0123] Based on a preset pruning algorithm, a pruning value of each decision tree node in the preset decision tree model is obtained.

[0124] When the pruning value is less than a preset pruning threshold, the decision tree node is pruned to obtain the classifier.

[0125] It should be noted that after updating the parameters of each decision tree in the preset decision tree model and then training, the preset decision tree model can be pruned, and all constructed decision trees can be fused. Because in the vision training effect evaluation task, the sample vision training effect evaluation data and the expanded data used for training usually contain noise, for example, the special treatment method of individual patients may interfere with model learning. Therefore, it is necessary to remove the branches that affect the stability of the decision tree model through post-pruning. In order to prevent overfitting, the generated decision tree is post-pruned, and according to the performance of the verification data, branches that do not greatly improve the performance of the preset decision tree model are removed to delete tree branches that do not significantly improve the performance of the model. The preset pruning algorithm can be used to obtain the pruning value of each decision tree node in the preset decision tree model, wherein the pruning algorithm is the above-mentioned R(t) obtaining formula, R(t) obtained is taken as the pruning value, and δ is set as a preset pruning threshold, for example, the preset pruning threshold δ is set to 0.01, and if R(t) is less than the set threshold δ, the node is pruned.

[0126] Further, the target evaluation model includes a feature extraction sub-model and a feature dimension reduction sub-model, and the target evaluation model is used for feature processing on the to-be-tested vision to obtain to-be-tested feature data, including:

[0127] The feature extraction sub-model is used for feature extraction on the to-be-tested vision training effect data to obtain initial vision features.

[0128] The feature dimension reduction sub-model is used for feature dimension reduction on the initial vision features to obtain the to-be-tested feature data.

[0129] It should be noted that the feature extraction sub-model can adopt a method of dynamically adjusting the network topology structure, allowing the weights and biases to be adjusted during the learning process, thereby improving the adaptability and training stability of the feature extraction sub-model. The initial visual feature is obtained by performing feature extraction on the to-be-measured visual training effect data through the trained feature extraction sub-model. The feature dimension reduction sub-model can adopt a self-encoding network with a differentiable adjustment strategy, which not only reconstructs the input data but also optimizes the matching degree between the encoded features and the preset low-dimensional target features, thereby improving the quality and efficiency of feature representation. The to-be-measured feature data is obtained by performing feature dimension reduction on the initial visual feature through the trained feature dimension reduction sub-model. The trained target evaluation model is used to evaluate the visual training effect. In one embodiment, for a newly collected to-be-measured visual training effect data, the initial visual feature is first extracted through the trained feature extraction sub-model. Further, the extracted initial visual feature is input into the feature dimension reduction model for feature dimension reduction. Further, the dimension-reduced to-be-measured feature data is input into the classifier for classification. Exemplarily, the classification categories include: obvious training effect, general training effect, and non-obvious training effect, a total of 3 categories.

[0130] See Figure 2 The embodiment also provides a visual training effect evaluation system, including: a first acquisition module configured to acquire sample visual training effect data; an input module configured to input the sample visual training effect data into a preset data expansion model to obtain expanded data; a construction module configured to construct a classifier based on a Monte Carlo algorithm and a preset decision tree model, and construct an initial evaluation model based on the classifier; a training module configured to train the initial evaluation model based on the expanded data and the sample visual training effect data to obtain a target evaluation model; and an evaluation module configured to acquire to-be-measured visual training effect data, and input the visual training effect data into the target evaluation model, so that the target evaluation model performs feature processing on the to-be-measured visual training effect data to obtain to-be-measured feature data, and then the target evaluation model classifies the to-be-measured feature data based on the classifier to obtain a visual training effect evaluation result.

[0131] The system provided in the embodiment can well implement the above-mentioned visual training effect evaluation method. The sample visual training effect data is expanded by the data expansion model, which expands the quantity and authenticity of the sample visual training effect data. The structure of the decision tree is selected from the preset extreme gradient boosting decision tree model by the Monte Carlo algorithm, which can explore a wider model structure compared with the traditional decision tree structure selection method relying on the greedy algorithm, thereby improving the generalization ability of the model. The above all helps to improve the accuracy of visual training effect evaluation.

[0132] The above is the preferred embodiment of the present application, it should be pointed out that, for those skilled in the art, without departing from the principles of the present application, can also make a number of improvements and refinements, these improvements and refinements are also considered to be within the scope of the present application.

Claims

1. A method for evaluating the effectiveness of vision training, characterized in that, include: Obtain data on the effectiveness of vision training in samples; The sample vision training effect data is input into a preset data augmentation model to obtain augmented data; A classifier is constructed based on the Monte Carlo algorithm and a pre-defined decision tree model, and an initial evaluation model is built based on the classifier. The initial evaluation model is trained based on the expanded data and the sample vision training effect data to obtain the target evaluation model; Acquire the vision training effect data to be tested, and input the vision training effect data into the target evaluation model so that the target evaluation model performs feature processing on the vision to be tested to obtain the feature data to be tested. Then, the target evaluation model classifies the feature data to be tested based on the classifier to obtain the vision training effect evaluation result. The data augmentation model includes a generative adversarial network and a variational autoencoder. The step of inputting the sample vision training effect data into the preset data augmentation model to obtain augmented data includes: The generative adversarial network is used to process the sample vision training effect data to generate synthetic data. The augmented data is obtained by performing automatic differentiation and variational inference on the synthetic data using the variational autoencoder. The step of constructing a classifier based on the Monte Carlo algorithm and a pre-defined decision tree model, and then constructing an initial evaluation model based on the classifier, includes: The preset decision tree model is initialized; The upper confidence bound of the vision training effect features in the sample vision training effect data and the expanded data is explored using the Monte Carlo algorithm to obtain the target feature combination; The vision training effect features in the target feature combination are evaluated for segmentation points using the Monte Carlo algorithm to obtain the target segmentation points. Based on the target feature combination and the target segmentation point, the preset decision tree model is updated to obtain the classifier.

2. The method for evaluating the effectiveness of vision training according to claim 1, characterized in that, The step of inputting the sample vision training effect data into a preset data augmentation model to obtain augmented data includes: The training process of the generative adversarial network is as follows: The preset initial generative adversarial network is trained based on the sample vision training effect data to obtain the generative adversarial network, and then intermediate training data is obtained based on the generative adversarial network. The training process of the variational autoencoder is as follows: The preset initial variational autoencoder is trained based on the intermediate training data to obtain the variational autoencoder.

3. The method for evaluating the effectiveness of vision training according to claim 2, characterized in that, The preset initial generative adversarial network includes a generator and a discriminator. The process of training the preset initial generative adversarial network based on the sample vision training effect data to obtain the generative adversarial network, and then obtaining intermediate training data based on the generative adversarial network, includes: Initialize the parameters of the generator and the parameters of the discriminator; A preset random noise vector is input into the generator to generate fake data; Based on the forged data, the sample vision training effect data, and the parameters of the discriminator, the loss function of the discriminator is constructed; Based on the random noise vector, the parameters of the generator, and the parameters of the discriminator, the loss function of the generator is constructed; Based on the gradient of the discriminator's loss function, the parameters of the discriminator are updated so that the preset initial generative adversarial network updates the parameters of the generator based on the gradient of the generator's loss function, thereby causing the loss functions of the generator and the discriminator to converge. When the loss functions of the generator and the discriminator converge simultaneously, the updating of the parameters of the generator and the discriminator is stopped, and the generative adversarial network is obtained.

4. The method for evaluating the effectiveness of vision training according to claim 3, characterized in that, When the loss functions of the generator and the discriminator converge simultaneously, the updating of the parameters of the generator and the discriminator is stopped, resulting in the generative adversarial network, which includes: Based on the loss function of the generator and the loss function of the discriminator, a dual loss function is constructed. Based on the gradient of the dual loss function, the parameters of the discriminator and the generator in the dual loss function are updated. When the dual loss function converges, the updating of the parameters of the generator and the discriminator is stopped, thus obtaining the generative adversarial network.

5. The method for evaluating the effectiveness of vision training according to claim 2, characterized in that, The preset initial variational autoencoder includes an encoder and a decoder. The step of training the preset initial variational autoencoder based on the intermediate training data to obtain the variational autoencoder includes: Based on the intermediate training data, the parameters of the encoder, the parameters of the decoder, and the KL divergence, the loss function of the preset initial variational autoencoder is constructed. Based on the gradient of the loss function of the preset initial variational autoencoder, the parameters of the encoder and the decoder are updated respectively. When the loss function of the variational autoencoder converges, the updating of the parameters of the encoder and the decoder is stopped, thus obtaining the variational autoencoder.

6. The method for evaluating the effectiveness of vision training according to claim 1, characterized in that, The step of updating the preset decision tree model based on the target feature combination and the target segmentation point to obtain the classifier includes: Based on a preset pruning algorithm, the pruning values ​​of each decision tree node in the preset decision tree model are obtained; When the pruning value is less than a preset pruning threshold, the decision tree node is pruned to obtain the classifier.

7. The method for evaluating the effectiveness of vision training according to claim 1, characterized in that, The target evaluation model includes a feature extraction sub-model and a feature dimensionality reduction sub-model. The target evaluation model performs feature processing on the visual acuity to be tested to obtain the feature data to be tested, including: The initial vision features are obtained by extracting features from the vision training effect data to be tested using the feature extraction sub-model. The initial visual features are reduced in dimensionality using the feature reduction sub-model to obtain the feature data to be tested.

8. A vision training effectiveness evaluation system, characterized in that, include: The first acquisition module is used to acquire sample vision training effect data; The input module is used to input the sample vision training effect data into a preset data augmentation model to obtain augmented data; The module is used to build a classifier based on the Monte Carlo algorithm and a preset decision tree model, and to build an initial evaluation model based on the classifier. The training module is used to train the initial evaluation model based on the expanded data and the sample vision training effect data to obtain the target evaluation model; The evaluation module is used to acquire the vision training effect data to be tested, and input the vision training effect data into the target evaluation model, so that the target evaluation model performs feature processing on the vision to be tested to obtain the feature data to be tested, and then the target evaluation model classifies the feature data to be tested based on the classifier to obtain the vision training effect evaluation result. The data augmentation model includes a generative adversarial network and a variational autoencoder. The step of inputting the sample vision training effect data into the preset data augmentation model to obtain augmented data includes: The generative adversarial network is used to process the sample vision training effect data to generate synthetic data. The augmented data is obtained by performing automatic differentiation and variational inference on the synthetic data using the variational autoencoder. The step of constructing a classifier based on the Monte Carlo algorithm and a pre-defined decision tree model, and then constructing an initial evaluation model based on the classifier, includes: The preset decision tree model is initialized; The upper confidence bound of the vision training effect features in the sample vision training effect data and the expanded data is explored using the Monte Carlo algorithm to obtain the target feature combination; The vision training effect features in the target feature combination are evaluated for segmentation points using the Monte Carlo algorithm to obtain the target segmentation points. Based on the target feature combination and the target segmentation point, the preset decision tree model is updated to obtain the classifier.

Images