Adaptation method and system for fluorescein staining of orthokeratology lenses

By using fluorescein stained images in corneal resizing mirror adaptation technology to construct differential models, including feature resolvers, self-attention reconstructors and differential predictors, the problem of complex adaptation and low accuracy in the prior art is solved. , achieving more efficient and accurate corneal resizing lens adaptation.

CN118605039BActive Publication Date: 2024-11-15TIANJIN NORMAL UNIVERSITY +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411075703.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-07
Publication Date
2024-11-15
Estimated Expiration
2044-08-07

AI Technical Summary

Technical Problem

The existing corneal resizing lens adaptation technology is complex and has low accuracy, resulting in inaccurate adaptation results.

Method used

The differential model was constructed using fluorescein stained images, including feature parser, self-attention reconstructor and differential predictor. The parameters of the trial-mounted film were adjusted by training the model to improve the matching accuracy and speed.

Benefits of technology

Through the use of the differential model, the adaptation accuracy and speed of the corneal resizing mirror are improved, the accurate calculation of deviation values ​​is ensured, and the reliability of the adaptation results is improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118605039B_ABST
    Figure CN118605039B_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of orthokeratology lenses, and in particular to an orthokeratology lens fluorescein staining fitting method and fitting system, the fitting method comprising: defining a fluorescein staining image before adjusting a deviation value as a non-standard fitting image, and defining a fluorescein staining image after adjusting the deviation value as a standard fitting image; constructing a differential model for determining the deviation value, the differential model comprising a feature parser, a self-attention reconstructor, and a differential predictor; training the differential model using the non-standard fitting image and the standard fitting image; obtaining a fluorescein staining image of a trial lens, and obtaining original parameters of several variables of the trial lens; importing the fluorescein staining image into the trained differential model, the differential model deriving a deviation value for adjusting the original parameter, and then adjusting the original parameter to the final parameter according to the deviation value, so as to adjust the trial lens to the orthokeratology lens. The present invention improves the fitting rate and fitting accuracy of orthokeratology lenses.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of orthokeratology lenses, and in particular to a method and system for fitting orthokeratology lenses using fluorescein staining. Background Art

[0002] In the field of intelligent fitting of orthokeratology lenses, some studies have proposed different methods to optimize the fitting process of orthokeratology lenses.

[0003] One is a calculation method that uses corneal height data. By applying a fitting algorithm, the AC value of orthokeratology lenses can be estimated more accurately. However, this method requires the preparation of the patient's corneal data file and requires complex cross-platform model building for fitting, resulting in longer data preparation and program running time.

[0004] The second is to establish an algorithm model for trial-wearing parameter selection of orthokeratology lenses based on machine learning methods, which is used to reduce the number of trial-wearing times during the orthokeratology lens fitting process. However, this method still fits the model through numerical data generated by corneal topography, and cannot provide predictive calibration through corneal fluorescence staining images.

[0005] Therefore, the existing orthokeratology lens fitting technology is relatively complex and the accuracy of the fitting results is low. Summary of the invention

[0006] The present invention aims to solve at least one of the technical problems existing in the related art. To this end, the present invention provides a method and system for fitting fluorescein staining of orthokeratology lenses, which solves the technical problems of relatively complex orthokeratology lens fitting technology and low accuracy of fitting results in traditional technologies, and improves the fitting rate and accuracy of orthokeratology lenses.

[0007] The present invention provides a method for fitting orthokeratology lenses with fluorescein staining, comprising the following steps:

[0008] The fluorescein staining image before adjusting the deviation value is defined as a non-standard fitting image, and the fluorescein staining image after adjusting the deviation value is defined as a standard fitting image;

[0009] Constructing a differential model for determining the deviation value, the differential model comprising a feature parser, a self-attention reconstructor, and a differential predictor;

[0010] Providing several groups of non-standard fitting images and standard fitting images, and using the non-standard fitting images and the standard fitting images to train the difference model;

[0011] Obtaining a fluorescein-stained image of the trial lens and obtaining original parameters of several variables of the trial lens;

[0012] The fluorescein-stained image is imported into the trained differential model, and the differential model obtains the deviation value for adjusting the original parameters, and then the original parameters are adjusted to the final parameters according to the deviation value, so as to adjust the trial lens to the orthokeratology lens.

[0013] A further improvement of the fluorescein staining adaptation method for orthokeratology lenses of the present invention is that the fluorescein staining image is imported into the trained differential model, and when the differential model obtains the deviation value for adjusting the original parameter, it includes:

[0014] The feature analyzer obtains the fluorescein-stained image, and then extracts non-standard features from the fluorescein-stained image, wherein the non-standard features include original parameters of the fluorescein-stained image, and the self-attention reconstructor generates reconstructed features based on the non-standard features. The differential predictor differentially fuses the non-standard features and the reconstructed features, and then the differential predictor obtains the deviation value.

[0015] A further improvement of the orthokeratology lens fluorescein staining fitting method of the present invention is that when the non-standard fitting image and the standard fitting image are used to train the differential model, the method comprises:

[0016] Setting a segmentation enhancement trainer, and connecting the segmentation enhancement trainer to the feature parser;

[0017] Training the feature parser;

[0018] Freeze the trained feature parser and train the self-attention reconstructor;

[0019] The trained feature parser and the self-attention reconstructor are frozen, and the differential predictor is trained to complete the training of the differential model.

[0020] A further improvement of the fluorescein staining matching method for orthokeratology lenses of the present invention is that the arc region of the fluorescein staining image is defined as the foreground, and other regions of the fluorescein staining image except the arc region are defined as the background, and when training the feature parser, it includes:

[0021] Determine the loss function of the feature parser ,

[0022]

[0023] in, represents the width of the fluorescein staining image, represents the height of the fluorescein staining image, each of which includes a number of pixel values , Indicates The true value of the foreground and background categories corresponding to the pixel values, Indicates The predicted probability corresponding to the pixel value.

[0024] A further improvement of the method for adapting fluorescein staining of orthokeratology lenses of the present invention is that the feature parser after freezing the training is trained, and when training the self-attention reconstructor, it includes:

[0025] Provides several sets of non-standard fitting images and standard fitting images ,

[0026] The non-standard fitting image The trained feature parser is processed to obtain non-standard features , and then the self-attention reconstructor reconstructs the non-standard features into reconstructed features ;

[0027] The standard fitting image The trained feature parser is processed to obtain contrast features ,

[0028] Construct the loss function of the self-attention reconstructor ,

[0029]

[0030] in, represents the number of groups of non-standard fitting images and standard fitting images used for training, Indicates The non-standard fitting images and the standard fitting images are grouped together. Indicates except The corresponding group Other than The contrast features of the standard fitting images of the group, express and The cosine similarity of express and The cosine similarity of represents the temperature coefficient;

[0031] Through the loss function Calculate and The contrast loss is then used to adjust the weight of the self-attention reconstructor according to the contrast loss.

[0032] A further improvement of the adaptation method of orthokeratology lens fluorescein staining of the present invention is that the feature parser and the self-attention reconstructor after freezing the training are used to train the differential predictor, including:

[0033] The differential predictor includes a fusion module and a prediction module;

[0034] Providing original parameters of several variables in the non-standard fitting image, providing final parameters of several variables in the standard fitting image, and then obtaining the true values ​​of the original parameters adjusted to the final parameters ,

[0035] Through the fusion module and Fusion to obtain differential features , and then the difference feature Input the prediction module to make predictions and get the predicted value ,

[0036] The loss function of the differential predictor is ,

[0037]

[0038] in, represents the number of groups of non-standard fitting images and standard fitting images used for training, Indicates The non-standard fitting images and the standard fitting images are grouped together. Indicates The predicted values ​​corresponding to the non-standard fitting images and the standard fitting images, Indicates The true values ​​corresponding to the non-standard fitting images and the standard fitting images;

[0039] The weights of the fusion module and the prediction module are adjusted according to the loss function.

[0040] A further improvement of the method for adapting fluorescein staining of orthokeratology lenses of the present invention is that a reconstruction network is set in the self-attention reconstructor, and the reconstruction network includes a sampling unit, a convolution unit and a linear attention unit;

[0041] The sampling unit is used to adjust the input size of the non-standard feature, and the convolution unit is used to encode and decode the non-standard feature, and the linear attention unit is used to provide a connection between multiple non-standard features and obtain a reconstructed feature.

[0042] A further improvement in the method for fitting fluorescein staining for orthokeratology lenses of the present invention is that the variables include amplitude reduction, AC, CP and DIA parameters.

[0043] The present invention also provides a corneal reshaping lens fluorescein staining fitting system, the fitting system is used to perform the above-mentioned fitting method, the fitting system includes a differential model, the differential model includes a feature parser, a self-attention reconstructor and a differential predictor;

[0044] The feature analyzer is used to obtain a fluorescein-stained image, and then extract non-standard features from the fluorescein-stained image;

[0045] The self-attention reconstructor is used to generate a reconstructed feature according to the non-standard feature;

[0046] The differential predictor is used to perform differential fusion on the non-standard features and the reconstructed features, and then the differential predictor obtains the deviation value.

[0047] The fitting method of fluorescein staining for orthokeratology lenses of the present invention constructs a differential model, and the differential model is specifically used for adjusting the parameters of variables in the orthokeratology lens fitting process, thereby improving the fitting speed and accuracy of the orthokeratology lenses. By providing paired standard fitting images and non-standard fitting images, the differential model is correspondingly trained to ensure that the feature parser, self-attention reconstructor and differential predictor in the differential model are adjusted to accurate weights and parameters, ensuring that the deviation value can be accurately obtained after the differential model is put into use, thereby improving the fitting accuracy of the orthokeratology lenses.

[0048] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 It is a flow chart of the adaptation method of membrane orthophotofluorescein staining provided by the present invention.

[0051] Figure 2 It is a block diagram of the adaptation system for membrane shaping mirror fluorescein staining provided by the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention. The following embodiments are used to illustrate the present invention, but cannot be used to limit the scope of the present invention.

[0053] Combine the following Figure 1 and Figure 2 A method for matching orthokeratology lenses with fluorescein staining according to the present invention is described, comprising the following steps:

[0054] The fluorescein staining image before adjusting the deviation value is defined as a non-standard fitting image, and the fluorescein staining image after adjusting the deviation value is defined as a standard fitting image;

[0055] Constructing a differential model for determining the deviation value, the differential model includes a feature parser, a self-attention reconstructor, and a differential predictor;

[0056] Providing several groups of non-standard fitting images and standard fitting images, and using the non-standard fitting images and the standard fitting images to train the difference model;

[0057] Obtaining a fluorescein-stained image of the trial lens and obtaining original parameters of several variables of the trial lens;

[0058] The fluorescein-stained image is imported into the trained differential model, and the differential model obtains the deviation value for adjusting the original parameters, and then the original parameters are adjusted to the final parameters according to the deviation value, so as to adjust the trial lens to the orthokeratology lens.

[0059] Preferably, the trial wear lens is a corneal reshaping lens with preset parameters, and the trial wear lens becomes the corneal reshaping lens after the deviation value is adjusted.

[0060] Furthermore, the variables include the drop, AC, CP and DIA parameters, where the drop indicates the expected visual acuity correction of the trial lens; AC indicates the curvature of the lens, which corresponds to the cornea. The smaller the eyeball, the steeper the cornea, the higher the AC value; the larger the eyeball, the flatter the cornea, the lower the AC value, the AC value is the positioning arc, and the positioning arc is the curvature; CP indicates the difference between the higher arc and the lower arc in the lens positioning arc, which is called CP "astigmatism"; DIA indicates the "diameter" of the lens.

[0061] In a preferred implementation case of the fluorescein staining adaptation method for orthokeratology lenses of the present invention, the fluorescein staining image is imported into the trained differential model, and when the differential model obtains the deviation value for adjusting the original parameter, it includes:

[0062] The feature parser obtains the fluorescein-stained image and then extracts non-standard features from the fluorescein-stained image. The non-standard features include the original parameters of the fluorescein-stained image. The self-attention reconstructor generates reconstructed features based on the non-standard features. The differential predictor differentially fuses the non-standard features and the reconstructed features, and then the differential predictor obtains the deviation value.

[0063] Furthermore, when the difference model is trained using the non-standard fitting images and the standard fitting images, it includes:

[0064] Set up a segmentation enhancement trainer and connect the segmentation enhancement trainer to the feature parser;

[0065] Train the feature parser;

[0066] Freeze the trained feature parser and train the self-attention reconstructor;

[0067] Freeze the trained feature parser and self-attention reconstructor, train the differential predictor, and complete the training of the differential model.

[0068] Furthermore, the arc region of the fluorescein-stained image is defined as the foreground, and the other regions of the fluorescein-stained image except the arc region are defined as the background. When training the feature parser, it includes:

[0069] Determine the loss function of the feature parser ,

[0070]

[0071] in, represents the width of the fluorescein staining image, represents the height of the fluorescein staining image, each of which includes a number of pixel values , Indicates The true value of the foreground and background categories corresponding to the pixel values, Indicates The predicted probability corresponding to the pixel value.

[0072] Furthermore, the trained feature parser is frozen and the self-attention reconstructor is trained, including:

[0073] Provides several sets of non-standard fitting images and standard fitting images ,

[0074] Non-standard fitting images The trained feature parser is processed to obtain non-standard features , and then the self-attention reconstructor reconstructs the non-standard features into reconstructed features ;

[0075] Standard fitting image The trained feature parser is processed to obtain contrast features ,

[0076] Constructing the loss function of the self-attention reconstructor ,

[0077]

[0078] in, represents the number of groups of non-standard fitting images and standard fitting images used for training, Indicates The non-standard fitting images and the standard fitting images are grouped together. Indicates except The corresponding group Other than The contrast features of the standard fitting images of the group, express and The cosine similarity of express and The cosine similarity of represents the temperature coefficient;

[0079] Through the loss function Calculate and The contrast loss is then used to adjust the weight of the self-attention reconstructor according to the contrast loss.

[0080] Furthermore, the trained feature parser and self-attention reconstructor are frozen, and the differential predictor is trained including:

[0081] The differential predictor includes a fusion module and a prediction module;

[0082] Provide the original parameters of several variables in non-standard fitting images, provide the final parameters of several variables in standard fitting images, and then obtain the true value of the original parameters adjusted to the final parameters ,

[0083] Through the fusion module and Fusion to obtain differential features , and then the difference feature Input the prediction module to make predictions and get the predicted value ,

[0084] Loss function for the differential predictor ,

[0085]

[0086] in, represents the number of groups of non-standard fitting images and standard fitting images used for training, Indicates The non-standard fitting images and the standard fitting images are grouped together. Indicates The predicted values ​​corresponding to the non-standard fitting images and the standard fitting images, Indicates The true values ​​corresponding to the non-standard fitting images and the standard fitting images;

[0087] The weights of the fusion module and the prediction module are adjusted according to the loss function.

[0088] Furthermore, a reconstruction network is set in the self-attention reconstructor, and the reconstruction network includes a sampling unit, a convolution unit, and a linear attention unit;

[0089] The sampling unit is used to adjust the input size of the non-standard features, and the convolution unit is used to encode and decode the non-standard features. The linear attention unit is used to provide connections between multiple non-standard features and obtain reconstructed features.

[0090] Specifically, in the inference stage of the differential model, based on the difference between the standard features of the standard fitting image generated by the self-attention reconstructor and the non-standard features of the non-standard fitting image extracted by the feature parser, the differential method is used to extract the differential features of the drop, AC, CP, and DIA parameters. This differential analysis can better capture the feature changes between standard features and non-standard features, and use the nonlinear ability of the differential predictor to extract relevant information of the differential features to obtain the adjustment values ​​of each parameter.

[0091] The first stage of reasoning of the differential model: a feature parser is used to extract representation features from the fluorescein-stained image (non-standard features correspond to non-standard fitting images, and standard features correspond to standard fitting images). This representation feature contains semantic information within the peripheral arc of the fluorescein-stained image. This information represents the global structure of the fluorescein-stained image, including important visual features that can represent corneal shape, texture, etc.; the choice of the backbone network of the feature parser can select network models with different parameter amounts according to the actual application situation. After the backbone network, a 1×1 convolutional network is used to adjust the number of output channels, and instance normalization and Relu activation function are used to further enhance the nonlinear expression ability of the network model; the feature parser and the segmentation enhancement trainer together constitute a U-shaped encoder residual structure, in which the feature parser corresponds to the encoding part, and the segmentation enhancement trainer corresponds to the decoding part; the main The main function is to generate feature images during the model training stage, and calculate the loss value of the feature parser in combination with the image segmentation task within the peripheral arc to guide the optimization direction of the feature parser; in the actual prediction and reasoning task, the segmentation enhancement trainer is abandoned, and the output of the feature parser is sent to the self-attention reconstructor to generate reconstructed features to provide support for subsequent prediction tasks; the U-shaped structure design allows the model to directly pass lower-level features to higher layers, more comprehensively capture the local and overall information of the fluorescein-stained image, help improve the network's understanding of the target semantics, and improve the modeling capabilities of fluorescein-stained images of different scales and structures; in addition, this structure helps the model focus on the image features within the peripheral arc of the fluorescein-stained image, reduce learning of other areas, thereby helping the backbone network to better learn the semantic information of the fluorescein-stained image and provide more targeted features for subsequent tasks.

[0092] The second stage of the difference model’s reasoning: using the self-attention reconstructor to transform non-standard features Refactoring to Refactoring Features , non-standard features It is a non-standard image feature extracted by the feature parser. This feature contains the visual information within the peripheral arc of the non-standard fitting image and plays a decisive role in the prediction of the subsequent adjustment value. The non-standard feature extracted by the feature parser is input into the reconstruction network. , the output is the reconstructed features after reconstruction , which can contain the visual local and global features of the standard fitting image; the up-sampling and down-sampling units are used to adjust the input size of the reconstruction network to ensure that the features can be effectively extracted and reconstructed at different scales; the convolution unit is responsible for non-standard features Encoding and decoding are performed to retain the key information of the image; while the linear attention unit is used to enhance non-standard features The relationship between them can more accurately obtain the reconstructed features .

[0093] The CrossChannelKans transformer module is a key component of the reconstruction network, which is designed to enhance the reconstruction network's tolerance to non-standard features. The CrossChannelKanstransformer module receives non-standard features As feature X; first, perform global average pooling and global maximum pooling operations on feature X. This step helps capture the overall statistical information of feature X and can better understand the global structure and feature distribution of feature X; then, calculate the cross-channel attention weight , achieving global attention to feature X, which can more comprehensively grasp the overall characteristics of the fluorescein-stained image and fully learn the expression of the parameters of each variable on the image. This process can guide the model to focus on the most representative and important feature parts of the image. If the globality is not considered, the important global information in the image will be ignored, resulting in the loss of some features; then, the linear attention unit is used to weight the features to obtain feature V, which can make the attention weight calculation more accurate and ensure that every detail of feature V is included in the global information; the linear attention unit enables the reconstruction network to more finely adjust the representation of feature V, so as to better capture the subtle changes and important features in feature V; then, the overall attention weight is calculated to achieve feature aggregation. Finally, the KAN model is used to automatically extract a smaller and more concise feature V architecture.

[0094] The design of the reconstruction network can pay more attention to the details of non-standard features, further improve the abstraction and representation capabilities of non-standard features, and by comprehensively considering global and local information, it can better understand the semantic content of non-standard features, providing a more reliable and comprehensive foundation for subsequent prediction tasks.

[0095] The third stage of reasoning of the differential model: the differential predictor is used to integrate the non-standard features of the feature parser and the reconstructed features of the self-attention reconstructor, and used to predict the final result; the differential predictor consists of three consecutive convolutional layers, and finally the feature map is unfolded through a flattening operation, and a linear layer is used to predict the adjustment values ​​of each parameter of the trial lens; the difference features between the reconstructed features and the non-standard features are represented by the fusion module; then, the offsets of these difference features are learned through the prediction module.

[0096] The present invention also provides a corneal reshaping lens fluorescein staining fitting system, the fitting system is used to perform the above fitting method, the fitting system includes a differential model, the differential model includes a feature parser, a self-attention reconstructor and a differential predictor;

[0097] The feature parser is used to obtain the fluorescein staining image, and then extract non-standard features from the fluorescein staining image;

[0098] The self-attention reconstructor is used to generate reconstructed features based on non-standard features;

[0099] The differential predictor is used to differentially fuse the non-standard features and the reconstructed features, and then the differential predictor obtains the deviation value.

[0100] The fitting method of fluorescein staining for orthokeratology lenses of the present invention constructs a differential model, and the differential model is specifically used for adjusting the parameters of variables in the orthokeratology lens fitting process, thereby improving the fitting speed and accuracy of the orthokeratology lenses. By providing paired standard fitting images and non-standard fitting images, the differential model is correspondingly trained to ensure that the feature parser, self-attention reconstructor and differential predictor in the differential model are adjusted to accurate weights and parameters, ensuring that the deviation value can be accurately obtained after the differential model is put into use, thereby improving the fitting accuracy of the orthokeratology lenses.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for fitting orthokeratology lenses with fluorescein staining, characterized in that: The steps include: The fluorescein staining image before adjusting the deviation value is defined as a non-standard fitting image, and the fluorescein staining image after adjusting the deviation value is defined as a standard fitting image; Constructing a differential model for determining the deviation value, the differential model comprising a feature parser, a self-attention reconstructor, and a differential predictor; Providing several groups of non-standard fitting images and standard fitting images, and using the non-standard fitting images and the standard fitting images to train the difference model; Acquire a fluorescein-stained image of the trial lens, and obtain original parameters of several variables of the trial lens, wherein the variables include drop amplitude, AC, CP, and DIA parameters; The fluorescein-stained image is imported into the trained differential model, and the differential model obtains a deviation value for adjusting the original parameter, and then the original parameter is adjusted to the final parameter according to the deviation value, so as to adjust the trial lens to the orthokeratology lens; The fluorescein-stained image is imported into the trained differential model, and the differential model obtains the deviation value for adjusting the original parameter, comprising: The feature analyzer obtains the fluorescein-stained image, and then extracts non-standard features from the fluorescein-stained image, wherein the non-standard features include original parameters of the fluorescein-stained image, the self-attention reconstructor generates a reconstructed feature according to the non-standard features, the differential predictor performs differential fusion on the non-standard features and the reconstructed features, and then the differential predictor obtains the deviation value; When the non-standard fitting image and the standard fitting image are used to train the difference model, the method includes: Setting a segmentation enhancement trainer, and connecting the segmentation enhancement trainer to the feature parser; Training the feature parser; Freeze the trained feature parser and train the self-attention reconstructor; Freezing the trained feature parser and the self-attention reconstructor, and training the differential predictor, thereby completing the training of the differential model; A reconstruction network is set in the self-attention reconstructor, wherein the reconstruction network includes a sampling unit, a convolution unit and a linear attention unit; The sampling unit is used to adjust the input size of the non-standard feature, and the convolution unit is used to encode and decode the non-standard feature, and the linear attention unit is used to provide a connection between multiple non-standard features and obtain a reconstructed feature.

2. The method for fitting orthokeratology lenses with fluorescein staining according to claim 1, characterized in that: The arc region of the fluorescein-stained image is defined as the foreground, and other regions of the fluorescein-stained image except the arc region are defined as the background. When training the feature parser, the training includes: Determine the loss function of the feature parser , in, represents the width of the fluorescein staining image, represents the height of the fluorescein staining image, each of which includes a number of pixel values , Indicates The true value of the foreground and background categories corresponding to the pixel values, Indicates The predicted probability corresponding to the pixel value.

3. The method for fitting orthokeratology lenses with fluorescein staining according to claim 1, characterized in that: Freezing the trained feature parser, and training the self-attention reconstructor, includes: Provides several sets of non-standard fitting images and standard fitting images , The non-standard fitting image The trained feature parser is processed to obtain non-standard features , and then the self-attention reconstructor reconstructs the non-standard features into reconstructed features ; The standard fitting image The trained feature parser is processed to obtain contrast features , Construct the loss function of the self-attention reconstructor , in, represents the number of groups of the non-standard fitting images and standard fitting images used for training, Indicates The non-standard fitting images and the standard fitting images are grouped together. Indicates except The corresponding group Other than The contrast features of the standard fitting images of the group, express and The cosine similarity of express and The cosine similarity of represents the temperature coefficient; Through the loss function Calculate and The contrast loss is then used to adjust the weight of the self-attention reconstructor according to the contrast loss.

4. The method for fitting orthokeratology lenses with fluorescein staining according to claim 1, characterized in that: Freezing the trained feature parser and the self-attention reconstructor, when training the differential predictor, includes: The differential predictor includes a fusion module and a prediction module; Providing original parameters of several variables in the non-standard fitting image, providing final parameters of several variables in the standard fitting image, and then obtaining the true values ​​of the original parameters adjusted to the final parameters , Through the fusion module and Fusion to obtain differential features , and then the difference feature Input the prediction module to make predictions and get the predicted value , The loss function of the differential predictor is , in, represents the number of groups of the non-standard fitting images and standard fitting images used for training, Indicates The non-standard fitting images and the standard fitting images are grouped together. Indicates The predicted values ​​corresponding to the non-standard fitting images and the standard fitting images, Indicates The true values ​​corresponding to the non-standard fitting images and the standard fitting images; The weights of the fusion module and the prediction module are adjusted according to the loss function.

5. A corneal reshaping lens fluorescein staining matching system, characterized in that: The adaptation system is used to perform the adaptation method according to any one of claims 1 to 4, wherein the adaptation system comprises a differential model, and the differential model comprises a feature parser, a self-attention reconstructor, and a differential predictor; The feature analyzer is used to obtain a fluorescein-stained image, and then extract non-standard features from the fluorescein-stained image; The self-attention reconstructor is used to generate a reconstructed feature according to the non-standard feature; The differential predictor is used to perform differential fusion on the non-standard features and the reconstructed features, and then the differential predictor obtains the deviation value.