A method and system for optimizing orthokeratology lenses based on the 3D retinal myopia defocus plane
By optimizing the parameters of orthokeratology lenses through multimodal feature fusion deep learning algorithms, the problem of personalized orthokeratology lens design and deformation prediction in existing technologies has been solved, achieving accurate prediction of corneal and retinal morphology and efficient optimization of orthokeratology lens parameters.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI JIAOTONG UNIV
- Filing Date
- 2023-04-06
- Publication Date
- 2026-07-03
AI Technical Summary
Existing orthokeratology lens design technology is difficult to meet the personalized parameter needs of wearers, especially the changes in shape parameters caused by deformation after stress are difficult to predict accurately.
An optimization method for orthokeratology lenses based on the 3D retinal myopia defocus surface was adopted. Through multimodal feature extraction and selection fusion deep learning algorithm, combined with corneal topography, optical coherence tomography and physiological parameters, the parameters of orthokeratology lenses were optimized.
Precise optimization of orthokeratology lens parameters was achieved, improving the accuracy of corneal and retinal morphology prediction after wearing. A correlation function between the orthokeratology lens and the defocus surface was established, improving design efficiency and accuracy.
Smart Images
Figure CN116597072B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a technology in the field of medical device manufacturing, specifically a method and system for optimizing orthokeratology lenses based on a 3D retinal myopia defocus surface. Background Technology
[0002] Current orthokeratology lens design technologies struggle to meet the reliability and accuracy requirements of wearers with varying parameters, especially the shape parameter changes caused by deformation under stress. Improved schemes based on corneal topography can only predict the treatment effect of orthokeratology lenses using a few simple parameters, and cannot predict the optimal matching parameters using multimodal fine-grained medical image data. Summary of the Invention
[0003] This invention addresses the problem that the structural parameters of existing orthokeratology lenses are usually generated based on human experience, resulting in low accuracy. It proposes an optimization method and system for orthokeratology lenses based on the 3D retinal myopia defocus surface. The optimal parameters are predicted by a deep learning algorithm that extracts and selects multimodal features from corneal topography, optical coherence tomography, and other data.
[0004] This invention is achieved through the following technical solution:
[0005] This invention relates to an optimization method for orthokeratology lenses based on a 3D retinal defocus surface. The method utilizes anterior and posterior segment optical coherence tomography (OCT) images, corneal detection data, physiological parameters, and orthokeratology lens parameters. A multimodal feature extraction and fusion algorithm is used to obtain multimodal data fusion features. A multi-scale image prediction network based on feature interaction function operators is then used to obtain a corneal topography map and an posterior segment retinal OCT image after orthokeratology lens wear. Furthermore, ray tracing and sampling using geometric optics are used to obtain the path of light rays passing through the cornea. This path, combined with the posterior segment retinal OCT image, is then used for three-dimensional reconstruction to obtain a 3D defocus surface. By adjusting the parameters of the 3D defocus surface, optimized orthokeratology lens manufacturing parameters are obtained.
[0006] This invention relates to a corneal reshaping lens optimization system based on a 3D retinal myopia defocus surface for implementing the above-mentioned method, comprising: an information input and preprocessing module, a feature extraction and fusion module, a corneal and retinal prediction module, a predicted optical path geometry and 3D defocus surface reconstruction module, and a parameter adjustment correction module. The information input and preprocessing module collects multimodal data such as corneal detection data, axial length parameter detection data, ocular imaging detection data, genomics test data, and physiological parameters, converts the data in a different format, adjusts its size, and outputs it to the feature extraction and fusion module. The feature extraction and fusion module performs feature extraction, feature selection, and fusion to obtain the user's multimodal fusion. Features: The corneal and retinal prediction module predicts changes in corneal and retinal morphology based on the orthokeratology lens and the user's multimodal fusion features. The optical path geometry and 3D defocus reconstruction module obtains a set of corneal incident light rays and corresponding corneal outgoing light ray paths through geometric optics light sampling technology, reconstructs the user's 3D defocus surface, and obtains the distance between the defocus surface and the retina. Then, it establishes and fits a physical correlation function between the parameters and the distance by combining the parameters of the orthokeratology lens. The parameter adjustment correction module adjusts the parameters of the orthokeratology lens according to the physical correlation function to obtain optimized orthokeratology lens manufacturing parameters.
[0007] The information input and preprocessing module includes a multimodal data acquisition unit and an information preprocessing unit. The multimodal data acquisition unit acquires corneal detection data, axial length parameter detection data, corneal OCT images, retinal OCT images, genomics test data, and physiological parameter data. The preprocessing unit converts the format of the eye information and converts some discrete data into image format, which is then output to the feature extraction and fusion module.
[0008] The feature extraction and fusion module includes a feature extraction unit and a feature fusion unit. The feature extraction unit uses a neural network to extract higher-dimensional image features from corneal topography and corneal and retinal OCT images in multimodal data. It uses embedding to encode axial length parameter detection data, genomic test data, and physiological parameters in multimodal data, converting discrete parameters into continuous parametric features. The image features and parametric features are then output to the feature fusion unit. The feature fusion unit assigns weights to the extracted image features and parametric features and dynamically adjusts the weights of each modality of data through training.
[0009] The corneal and retinal prediction module includes an external force encoding unit and an orthokeratology lens-induced corneal and retinal prediction unit. The external force encoding unit encodes the pressure applied to the cornea based on the preset geometric parameters and wearing parameters of the orthokeratology lens. The orthokeratology lens-induced corneal and retinal prediction unit predicts the shape of the cornea and retina after wearing the orthokeratology lens by using a Transformer network based on a U-shaped network and a mechanical feature interaction modeling (PhyInt) component, based on the input multimodal data and the mechanical features encoded by the external force encoding unit.
[0010] The Transformer network includes an encoder and a decoder, with a total of 4 layers. The encoder path extracts spatial downsampling, and the decoder path performs feature reconstruction and spatial upsampling. Feature maps are stacked from the encoder path of the same resolution through jump connections to gradually recover the input of the same size.
[0011] Each PhyInt component in the Transformer network is further preceded by a CNN structure.
[0012] The encoding of the pressure applied to the cornea specifically includes:
[0013] The first step is to synthesize the curvature map of the orthokeratology lens: The orthokeratology lens is used to form a curvature map that is aligned with each point of the corneal region according to the preset size parameters in the base curve region, reverse curve region, positioning curve region and peripheral curve region. At each point in the curvature map, the corresponding curvature value of the lens is filled in to represent the pressure of the orthokeratology lens on the cornea. This makes it convenient to use a neural network to simulate the continuous force generation and application process with an implicit feature encoding.
[0014] The second step involves extracting orthokeratology lens pressure features based on time encoding: The orthokeratology lens curvature map synthesized in the first step is input into a funnel-shaped network structure containing 5 convolutional layers of the external force encoding unit for feature extraction and dimensionality reduction, obtaining the feature encoding information of the pressure applied by the orthokeratology lens; the time information from the duration of orthokeratology lens wear is encoded, and a decay function e is applied. -at The duration of lens wear indicates the degree of impact on the cornea, and is multiplied with the original pressure feature encoding information to obtain the time-decayed pressure feature encoding information; the feature encoding information of the pressure applied by the orthokeratology lens and the time-decayed pressure features are stacked together, and after several different cascaded convolution and reshaping operations, the external force feature encoding is obtained.
[0015] The prediction of the corneal and retinal shape after wearing orthokeratology lenses specifically includes:
[0016] Step 1, Injecting multimodal data features: Encoding multimodal data features using a combination of U-shaped structure and Transformer network to match the input of the prediction model;
[0017] Step 2, Injecting external mechanical feature information: Mechanical features are encoded and injected into the PhyInt component through the CNN structure before the PhyInt component to achieve image prediction guided by external mechanical information;
[0018] Step 3, predicting changes in the cornea and retina based on mechanical feature interaction components: the cornea and retina are discretized into a pentahedral discrete mechanical system, and self-attention operations are calculated. Wherein: Q and K encode the location information, and their difference represents the spatial distance information between two points; then the shape of the cornea and retina after wearing orthokeratology lenses is predicted by the Transformer network with mechanical feature interaction modeling components.
[0019] The aforementioned pentahedral discrete mechanical system refers to a pentahedral discrete mechanical system in which any corneal location O corresponds to four nearby points A, B, C, and D, forming an abstract and approximately rectangular structure. OA =l OB =l OC =l OD =l, the length of the pentahedral discrete mechanical system width The curvature of point O is That is, the curvature of a vertex in a pentahedral discrete mechanical system is inversely proportional to the second square of the distance between that point and the surrounding points.
[0020] The aforementioned predictive optical path geometry and 3D defocus surface reconstruction module includes an optical path prediction unit and a 3D defocus surface reconstruction unit. The optical path prediction unit predicts the exit angle of light after passing through the cornea based on the corneal morphology predicted by the cornea and retina prediction module, i.e., the radius of curvature of the anterior corneal surface, using geometric optics light sampling technology. The 3D defocus surface reconstruction unit uses Monte Carlo sampling technology to sample the point light source located in front of the cornea, obtains the point cloud data of the defocus surface through the incident light, and obtains the expression of 10 sets of function surfaces that make up the defocus surface through the built-in surface parameter prediction network, fits the reconstructed defocus surface, and calculates the distance between each point on the reconstructed defocus surface and the retinal surface.
[0021] The parameter adjustment correction module includes a correlation function establishment unit and an orthokeratology lens parameter correction unit. The correlation function establishment unit dynamically adjusts the radius of the base arc region, the width of the inversion arc region, the width of the positioning arc region, and the width of the peripheral arc region of the orthokeratology lens and outputs the results to the prediction optical path geometry and 3D defocus surface reconstruction module. This allows the unit to obtain the changes in the defocus surface under different orthokeratology lens parameters and the distance between the reconstructed defocus surface and the retina. It also generates a correlation function between the orthokeratology lens parameters and the distance to the defocus surface. The orthokeratology lens parameter correction unit obtains the optimal parameters of the orthokeratology lens under the optimal defocus surface condition through the correlation function. Attached Figure Description
[0022] Figure 1 This is a schematic diagram of the end-to-end method for generating optimal fitting parameters for orthokeratology lenses according to the present invention.
[0023] Figure 2 This is a schematic diagram of the 3D image-based geometric reconstruction method for myopic defocus surfaces in the embodiment.
[0024] Figure 3 This is a schematic diagram of the corneal particle interactive deformation prediction algorithm based on physical and mechanical rules in the embodiment.
[0025] Figure 4 This is a schematic diagram of the algorithm for predicting the defocus surface function based on MLP in the embodiment. Detailed Implementation
[0026] like Figure 1 As shown in the figure, this embodiment relates to a corneal reshaping lens optimization system based on a 3D retinal myopia defocus surface, which includes: an information input and preprocessing module, a feature extraction and fusion module, a predictive optical path geometry and 3D defocus surface reconstruction module, and a parameter adjustment correction module.
[0027] The information input and preprocessing module includes a multimodal data acquisition unit and an information preprocessing unit. The multimodal data acquisition unit uses corneal detection data and axial length parameter detection data obtained through Pentacam scanning, corneal OCT images and retinal OCT images obtained through the Vision OCT scanner, genomics test data, and physiological parameter data. In the information preprocessing unit, for the CSV file storing corneal detection data, the curvature radius of each point on the cornea is exported, and the data is converted into a corneal topography map in image form. The center of the corneal topography map is aligned with the pupil center, and the central area is cropped to obtain a 128×128 grayscale image, corresponding to a 6.00mm×6.00mm corneal region. For the corneal OCT images and retinal OCT images, a noise-free image is restored using a denoising algorithm, and the image is reshaped to a size of 160×100.
[0028] The feature extraction and fusion module includes a feature extraction unit and a feature fusion unit. The feature extraction unit uses a ResNet50 network to extract corneal topography and corneal and retinal OCT images from multimodal data to obtain image features. It also uses embedding technology to encode axial length parameter detection data, genomics test data, and physiological parameter data, converting discrete parameters into continuous vector representations. The feature fusion unit assigns weights to the extracted image features and vector representations and can dynamically adjust the weights of each modality of data.
[0029] In this embodiment, the pressure applied to the cornea is encoded through the following operations, specifically including:
[0030] The first step is to synthesize the curvature map of the orthokeratology lens: a pre-defined orthokeratology lens with a base curve radius of 6mm, a reverse curve width of 0.68mm, a positioning curve width of 1.23mm, and a peripheral curve width of 0.25mm is used to form a curvature map of the orthokeratology lens that is aligned with each point of the corneal region. At each point in the curvature map of the orthokeratology lens, the corresponding curvature value of the lens is filled in to represent the pressure of the orthokeratology lens on the cornea. This allows the use of neural networks to simulate the generation and application of continuous forces in an implicit feature encoding manner.
[0031] The second step involves extracting latent pressure features based on time-encoded data: the curvature map of the orthokeratology lenses is input into a funnel-shaped convolutional network with five 3×3 kernels to extract latent mechanical features. Since the wearing time of the orthokeratology lenses is also a significant factor affecting the medical process, a decay function e is applied to encode the time information. -at The degree of impact of the lens on the cornea is represented, and multiplied with the original pressure feature encoding information to obtain the time-decayed pressure feature encoding information. The feature encoding information of the pressure applied by the orthokeratology lens and the time-decayed pressure features are stacked together, and after a series of different cascaded convolution and reshaping operations, the external force feature encoding is obtained and input into the corneal and retinal prediction unit induced by the orthokeratology lens.
[0032] The input to the optical path prediction unit is the corneal data predicted by the corneal and retinal prediction module, which includes the radius of curvature of the anterior corneal surface. Due to the difference in refractive indices between the anterior and posterior media of the cornea, light refracts after passing through the cornea. Clinically, the cornea is generally considered as a whole, with a refractive index of 1.3375. For a point p on the cornea, let the incident angle θ1 and exit angle θ2 of a ray incident on point p satisfy: n1 sinθ1 = n2 sinθ2, where n1 is the air refractive index (1), and n2 is the corneal equivalent refractive index (1.3375). Therefore, at this time... The angle of the emitted light can be determined from this formula. Based on this refraction formula, the propagation path of light after passing through the cornea can be obtained using the ray tracing software Zemax.
[0033] The 3D defocus surface reconstruction unit includes a dense ray sampling subunit and a defocus surface function reconstruction subunit. The dense ray sampling subunit uses Monte Carlo ray sampling technology to obtain ray information by densely and efficiently sampling a light source placed in front of the cornea. For each ray, based on the refraction formula, its propagation path through the cornea is calculated, and point cloud data of the defocus surface is formed using an optical path array. The 3D defocus surface reconstruction subunit inputs the defocus surface point cloud data into a surface parameter prediction network based on the point cloud model of the defocus surface and the retinal OCT image to obtain the expression of the defocus surface, and then calculates the distance between each point on the defocus surface and the retinal surface.
[0034] The sampling process specifically involves: for a point light source placed 10 cm in front of the cornea, a random number generator is used to generate a set of random samples to determine the point and direction of the light ray. These random samples are typically points uniformly distributed on the surface of the light source or vectors randomly generated within a certain range. For each random sample, a light ray is generated. The point of the light ray is the sampling point on the surface of the light source, and the direction is the direction from the sampling point to any point on the cornea.
[0035] The surface parameter prediction network is a 5-layer MLP network. After obtaining the coordinate information of each point on the point cloud, the network predicts the expression of 10 sets of function surfaces that make up the defocus surface. The reconstructed defocus surface expression is then fitted using these 10 sets of surface functions.
[0036] Through specific practical experiments, using Python and PyTorch as the deep learning framework, the accuracy of corneal prediction after wearing orthokeratology lenses was measured by two indicators: PSNR and SSIM. Table 1 shows the data obtained by comparing with existing commonly used neural network models.
[0037] Table 1
[0038] U-Net SHM FCN U-former Ours PSNR 27.57 26.94 23.42 27.98 28.37 SSIM 0.7386 0.6977 0.6342 0.7591 0.7642
[0039] For the prediction of retinal morphology after wearing orthokeratology lenses, the accuracy of the prediction is measured by two indicators: PSNR and SSIM. Table 2 shows the data obtained by comparing with existing commonly used neural network models.
[0040] Table 2
[0041] U-Net SHM FCN U-former Ours PSNR 28.87 27.23 24.65 29.56 30.86 SSIM 0.7579 0.7235 0.6853 0.7734 0.8153
[0042] The expression for the reconstructed retinal defocus plane in this embodiment is: After adjusting the parameters, the optimized shaped mirror base arc area radius is 6.1mm, the reverse arc area width is 0.64mm, the positioning arc area width is 1.28mm, and the peripheral arc area width is 0.45mm.
[0043] This invention enables rapid and accurate optimization of orthokeratology lens parameters. The orthokeratology lenses automatically designed by this system exhibit a high degree of consistency with those designed by doctors, assisting them in designing orthokeratology lens parameters that better suit the eye and making the process more efficient. To more accurately predict changes in corneal and retinal morphology after wearing orthokeratology lenses, this invention incorporates biomechanical feature interaction modeling inspired by physical laws into the prediction network. Furthermore, to more accurately express the defocus surface function, it uses MLP to predict the multi-surface formulas that constitute the defocus surface. These are the key features of this invention.
[0044] In summary, compared to existing technologies, this invention can more accurately predict the morphology of the cornea and retina after wearing orthokeratology lenses through a mechanical feature interaction modeling unit, thereby obtaining a more accurate retinal defocus surface. This invention establishes a correlation function between the orthokeratology lens and the defocus surface, making the orthokeratology lens more precise and efficient.
[0045] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.
Claims
1. A corneal reshaping lens optimization system based on a 3D retinal myopia defocus surface, characterized in that, include: The system comprises an information input and preprocessing module, a feature extraction and fusion module, a corneal and retinal prediction module, a predicted optical path geometry and 3D defocus plane reconstruction module, and a parameter adjustment correction module. Specifically: the information input and preprocessing module collects corneal detection data, axial length parameter detection data, ocular imaging detection data, genomics test data, and multimodal physiological parameter data, converts the format, adjusts the size, and outputs it to the feature extraction and fusion module; the feature extraction and fusion module performs feature extraction, feature selection, and fusion to obtain the user's multimodal fused features; and the corneal and retinal prediction module, based on orthokeratology lenses... In addition to user multimodal fusion features, the system predicts changes in corneal and retinal morphology when using orthokeratology lenses; the optical path geometry prediction and 3D defocus reconstruction module obtains a set of corneal incident light rays and corresponding corneal outgoing light ray paths through geometric optics light sampling technology, reconstructs the user's 3D defocus surface and obtains the distance between the defocus surface and the retina, and then establishes and fits a physical correlation function between the parameters and the distance by combining the parameters of the orthokeratology lens; the parameter adjustment correction module adjusts the parameters of the orthokeratology lens according to the physical correlation function to obtain optimized orthokeratology lens manufacturing parameters; The corneal and retinal prediction module includes an external force encoding unit and an orthokeratology lens-induced corneal and retinal prediction unit. The external force encoding unit encodes the pressure applied to the cornea based on the preset geometric parameters and wearing parameters of the orthokeratology lens. The orthokeratology lens-induced corneal and retinal prediction unit predicts the shape of the cornea and retina after wearing the orthokeratology lens through a Transformer network based on a U-shaped network and a mechanical feature interaction modeling component, based on the input multimodal data and the mechanical features encoded by the external force encoding unit. The encoding of the pressure applied to the cornea specifically includes: The first step is to synthesize the curvature map of the orthokeratology lens: The orthokeratology lens is used to form a curvature map that is aligned with each point of the corneal region according to the preset size parameters in the base curve region, reverse curve region, positioning curve region and peripheral curve region. At each point in the curvature map, the corresponding curvature value of the lens is filled in to represent the pressure of the orthokeratology lens on the cornea. This makes it convenient to use a neural network to simulate the continuous force generation and application process with an implicit feature encoding. The second step involves extracting orthokeratology lens pressure features based on time encoding: The orthokeratology lens curvature map synthesized in the first step is input into a funnel-shaped network structure containing five convolutional layers of the external force encoding unit for feature extraction and dimensionality reduction, obtaining the feature encoding information of the pressure applied by the orthokeratology lens; the time information from the duration of orthokeratology lens wear is encoded, and a decay function is applied. The duration of lens wear indicates the degree of impact on the cornea, and is multiplied with the original pressure feature encoding information to obtain the time-decayed pressure feature encoding information; the feature encoding information of the pressure applied by the orthokeratology lens and the time-decayed pressure features are stacked together, and after several different cascaded convolution and reshaping operations, the external force feature encoding is obtained; The prediction of the corneal and retinal shape after wearing orthokeratology lenses specifically includes: Step 1, Injecting multimodal data features: Encoding multimodal data features using a combination of U-shaped structure and Transformer network to match the input of the prediction model; Step 2, Injecting external mechanical feature information: Mechanical features are encoded and injected into the PhyInt component through the CNN structure before the PhyInt component to achieve image prediction guided by external mechanical information; Step 3, predicting changes in the cornea and retina based on mechanical feature interaction components: the cornea and retina are discretized into a pentahedral discrete mechanical system, and self-attention operations are calculated. Where Q and K encode the location information, It represents the spatial distance information between two points; then, the shape of the cornea and retina after wearing orthokeratology lenses is predicted by the Transformer network with mechanical feature interactive modeling components.
2. The corneal reshaping lens optimization system based on the 3D retinal myopia defocus surface according to claim 1, characterized in that, The information input and preprocessing module includes a multimodal data acquisition unit and an information preprocessing unit. The multimodal data acquisition unit acquires corneal detection data, axial length parameter detection data, corneal OCT images, retinal OCT images, genomics test data, and physiological parameter data. The preprocessing unit converts the format of the eye information and converts some discrete data into image format, which is then output to the feature extraction and fusion module.
3. The corneal reshaping lens optimization system based on the 3D retinal myopia defocus surface according to claim 1, characterized in that, The feature extraction and fusion module includes a feature extraction unit and a feature fusion unit. The feature extraction unit uses a neural network to extract higher-dimensional image features from corneal topography and corneal and retinal OCT images in multimodal data. It uses embedding operations to encode axial length parameter detection data, genomic test data, and physiological parameters in multimodal data, converting discrete parameters into continuous parametric features. The image features and parametric features are then output to the feature fusion unit. The feature fusion unit assigns weights to the extracted image features and parametric features and dynamically adjusts the weights of each modality of data through training.
4. The corneal reshaping lens optimization system based on the 3D retinal myopia defocus surface according to claim 1, characterized in that, The Transformer network includes an encoder and a decoder, with a total of 4 layers. The encoder path extracts spatial downsampling, and the decoder path performs feature reconstruction and spatial upsampling. The input of the same size is gradually recovered by hopping and stacking feature maps from the encoder path of the same resolution. Each mechanical feature interaction modeling component in the Transformer network is preceded by a CNN structure.
5. The corneal reshaping lens optimization system based on the 3D retinal myopia defocus surface according to claim 1, characterized in that, The aforementioned predictive optical path geometry and 3D defocus surface reconstruction module includes an optical path prediction unit and a 3D defocus surface reconstruction unit. The optical path prediction unit predicts the exit angle of light after passing through the cornea based on the corneal morphology predicted by the cornea and retina prediction module, i.e., the radius of curvature of the anterior corneal surface, using geometric optics light sampling technology. The 3D defocus surface reconstruction unit uses Monte Carlo sampling technology to sample the point light source located in front of the cornea, obtains the point cloud data of the defocus surface through the incident light, and obtains the expression of 10 sets of function surfaces that make up the defocus surface through the built-in surface parameter prediction network, fits the reconstructed defocus surface, and calculates the distance between each point on the reconstructed defocus surface and the retinal surface.
6. The corneal reshaping lens optimization system based on the 3D retinal myopia defocus surface according to claim 1, characterized in that, The parameter adjustment correction module includes a correlation function establishment unit and an orthokeratology lens parameter correction unit. The correlation function establishment unit dynamically adjusts the radius of the base arc region, the width of the inversion arc region, the width of the positioning arc region, and the width of the peripheral arc region of the orthokeratology lens and outputs the results to the prediction optical path geometry and 3D defocus surface reconstruction module. This allows the unit to obtain the changes in the defocus surface under different orthokeratology lens parameters and the distance between the reconstructed defocus surface and the retina. It also generates a correlation function between the orthokeratology lens parameters and the distance to the defocus surface. The orthokeratology lens parameter correction unit obtains the optimal parameters of the orthokeratology lens under the optimal defocus surface condition through the correlation function.
7. A method for optimizing orthokeratology lenses, characterized in that, According to any one of claims 1-6, the corneal reshaping lens optimization system based on a 3D retinal myopia defocus surface obtains multimodal data fusion features by using multimodal feature extraction and selection fusion algorithms based on anterior and posterior segment optical coherence tomography images, corneal detection data, physiological parameters, and corneal reshaping lens parameters. Then, a multi-scale image prediction network based on feature interaction function operators is used to obtain corneal topography and posterior segment retinal OCT images after wearing the corneal reshaping lens. Finally, ray tracing and sampling using geometric optics are used to obtain the path of light rays passing through the cornea. This path, combined with the posterior segment retinal OCT image, is used for three-dimensional reconstruction to obtain a 3D defocus surface. Finally, by adjusting the parameters of the 3D defocus surface, optimized corneal reshaping lens manufacturing parameters are obtained.