Prediction method and model training method for efficacy of orthokeratology in controlling progression of myopia, and computer program product

Abstract

This patent relates to the field of medical artificial intelligence technology, specifically to a method for predicting the efficacy of orthokeratology in controlling myopia progression based on multimodal data, a model training method for implementing this prediction method, and a computer program product. The prediction method includes: a pre-treatment screening step to obtain a one-year axial elongation prediction value and initial risk recommendations; and an early reassessment step to obtain an updated axial elongation prediction value and intervention suggestions. Through this two-stage prediction framework, the decision-making node for efficacy prediction is decomposed into two steps: a static screening before treatment and an early dynamic reassessment after wearing the device. This significantly improves the accuracy and timeliness of predicting the final one-year efficacy outcome, providing sufficient technical basis for timely adjustments to the intervention plan.

Description

Technical Field

[0001] This patent relates to the fields of medical artificial intelligence, optometry and computer-aided diagnostic technology. Specifically, it relates to a method for predicting the efficacy of orthokeratology in controlling myopia progression based on multimodal data, a model training method for implementing the prediction method, and a computer program product. Background Technology

[0002] Myopia has become a significant public health problem worldwide, particularly among adolescents. Orthokeratology (Ortho-K) is a non-surgical intervention that temporarily reshapes the anterior surface of the cornea by wearing a rigid gas-permeable contact lens with reverse geometry overnight, thereby achieving clear uncorrected vision during the day and slowing axial elongation. However, clinical studies have shown significant differences in the one-year axial length (AL) growth control after orthokeratology among different individuals. Accurately identifying individuals who may not fully benefit from conventional orthokeratology before or very early in treatment is a key challenge in clinical management.

[0003] In existing technologies, there exists a class of technical solutions based on machine learning or empirical models to assist in clinical decision-making for orthokeratology lenses. These solutions primarily focus on recommending lens parameters, optimizing the fitting process, or assessing corneal fit. Specifically, these existing solutions typically employ the following technical path: acquiring low-dimensional structured data of the patient's ocular biometric parameters before fitting, such as refractive error, corneal curvature, and axial length; and using a pre-defined mathematical model or machine learning algorithm (such as a random forest model) to predict the patient's future axial length growth. Based on this, the core objective of these solutions is to deduce and output lens parameters deemed most suitable for the patient based on the prediction results. Alternatively, after the patient wears the lens, by acquiring corneal topography before and after fitting, and using image segmentation and localization networks to process the difference maps, the overlap between the actual treatment area and the ideal simulated area on the cornea is assessed. This results in an evaluation of the lens fit, guiding subsequent adjustments to the lens parameters.

[0004] However, in actual clinical applications, the above-mentioned technical solutions generally have the following technical problems: the core technical problem solved by this type of solution is "what kind of lens to use" or "whether the lens is well-fitted", and it cannot respond to the clinical decision-making needs of "whether the patient is suitable for initiating orthokeratology treatment" and "how to dynamically judge the trend of its efficacy development in the very early window after the treatment is initiated". Summary of the Invention

[0005] To solve, or at least partially solve, the above-mentioned technical problems, the present invention provides a method for predicting the efficacy of orthokeratology in controlling myopia progression based on multimodal data, comprising: Pre-treatment screening steps: Obtain the first corneal topography image and baseline clinical variables of the subject to be evaluated before treatment, and input them into the first prediction model to obtain the one-year axial elongation prediction value and initial risk recommendations; Early reassessment steps: After the subject to be assessed has worn orthokeratology lenses for a preset follow-up period, a second corneal topography image is acquired; based on the first and second corneal topography images, differential features characterizing corneal remodeling are constructed, and the first, second, and differential features are input into a second prediction model to obtain updated axial elongation prediction values ​​and intervention prompts; The first prediction model is configured to process static baseline features, while the second prediction model is configured to process dynamic features that include time-series change information; The difference features include at least a pixel-level difference map or a high-dimensional feature vector extracted based on a convolutional neural network. The pixel-level difference map is a residual matrix obtained by subtracting the refractive power matrix of the first corneal topographic image from the refractive power matrix of the second corneal topographic image.

[0006] The two-stage prediction framework described above decomposes the decision-making nodes for efficacy prediction into two steps: static screening before treatment and early dynamic reassessment after wearing the lenses. The first prediction model can output the predicted value of axial length growth for one year using only pre-treatment static data. This allows clinicians to identify potential patients with insufficient efficacy before the actual cost of fitting is incurred, thus avoiding the irreversible cost investment and intervention window delays caused by the current technology that requires fitting, lens manufacturing, and wearing before assessment. The second prediction model introduces dynamic differential features based on corneal topography. These features can accurately quantify the local and non-uniform morphological remodeling of the cornea after wearing. By fusing this time-series dynamic information with static baseline information, the model can significantly improve the accuracy and timeliness of predicting the final one-year efficacy outcome at a very early follow-up point, providing sufficient technical basis for timely adjustment of intervention plans.

[0007] Optionally, in the step of constructing differential features characterizing corneal reshaping, the construction of the pixel-level difference map is restricted to a preset annular region in the center of the cornea; the preset annular region is an annular band with a radius of 3 mm to 5 mm, and the annular band corresponds to the mid-peripheral defocus area of ​​the orthokeratology lens.

[0008] Since the corneal reshaping effect of orthokeratology lenses is mainly manifested in the defocus ring region of the mid-periphery, changes in refractive power in this region are closely related to myopia control. Limiting the extraction range of the difference map to this specific ring region can filter out irrelevant or noisy changes in the central optical zone and the periphery of the cornea, ensuring that the constructed difference features can more concentratedly and accurately represent corneal remodeling signals related to treatment efficacy, thereby improving the predictive signal-to-noise ratio of the second prediction model.

[0009] Optionally, the logic for generating the intervention prompt includes: obtaining the probability of poor efficacy risk output by the second prediction model; when the probability of poor efficacy risk exceeds a dynamic threshold... At that time, suggestions for combined drug-assisted intervention are generated; the dynamic threshold The optimal threshold is determined based on a validation set, which contains several historical samples with known therapeutic outcomes. The Youden index of the validation set is defined as the sum of sensitivity and specificity minus 1. The dynamic threshold... When iterating through several candidate probability thresholds, the candidate probability threshold that corresponds to the maximum value of the Youden index is selected.

[0010] By comparing the abstract risk probability output by the model with a specific clinical threshold that is dynamically optimized based on statistical principles, clear and actionable intervention prompts can be generated. This threshold is not set arbitrarily, but rather determined by maximizing the Youden exponent of the validation set. This achieves an optimal balance between sensitivity and specificity, and allows the entire early warning mechanism to self-calibrate as data accumulates, improving the system's decision robustness across different populations and scenarios.

[0011] Optionally, prior to the pretreatment screening step and the early reassessment step, the method further includes a step of periodically synchronizing the color mark configuration file released by the device manufacturer via an encrypted API interface, and restoring the pixel color values ​​in the relative color mark mode of the original image to absolute refractive power values ​​based on the mapping function in the color mark configuration file, so as to eliminate non-physiological differences introduced by the device display settings.

[0012] Because corneal topography maps under different devices or settings may use relative color scales, the absolute refractive power values ​​corresponding to the same color may be inconsistent. This non-physiological display difference can become noise that interferes with model learning. This application's technical solution establishes an information update channel with device manufacturers, enabling automatic acquisition of the latest color scale mapping relationships. By restoring the image from a relative color scale mode to a unified absolute refractive power value, the physical consistency of the input data is ensured, eliminating the differences introduced by different display settings, and creating conditions for high-precision prediction across centers and devices.

[0013] Optionally, it also includes a cloud-based dynamic template registration step for unknown devices, including: when the device model of the acquired corneal topography image does not match in the local database, uploading the corneal topography image to the backend server, and receiving a new color mark configuration file generated by the backend server and digitally signed using its private key, the new color mark configuration file containing a digital signature is used to verify the integrity and security of the new color mark configuration file on the local terminal.

[0014] In practical applications, end users constantly encounter new models or versions of corneal topography devices, posing a risk of local template inventory lag. This technical solution constructs a dynamic template registration mechanism with secure verification. When encountering an unknown device, it can instantly trigger a cloud-based collaborative process to generate a trusted mapping file with a digital signature. This ensures the data processing workflow's rapid adaptability to new devices, guaranteeing continuous system availability. Furthermore, the digital signature technology prevents the injection of malicious configuration files, safeguarding the integrity and security of the absolute refractive power values ​​used as the basis for prediction, and providing a solid and reliable data traceability foundation for the entire prediction process.

[0015] The present invention also provides a model training method for training a first prediction model and / or a second prediction model in the aforementioned prediction method, the model training method comprising: Dataset construction steps: Collect multicenter follow-up data of orthokeratology surgery to construct a first training set and a second training set; the first training set includes the first corneal topography image and baseline clinical variables before treatment, and is labeled with the true value of axial growth over one year; the second training set includes the first corneal topography image, baseline clinical variables, the second corneal topography image after wearing for a preset duration, and the differential features characterizing the morphological changes between the first corneal topography image and the second corneal topography image, and is labeled with the true value of axial growth over one year corresponding to the follow-up duration; Loss function definition steps: Define a hybrid loss function, which includes at least a data loss term and a surrogate physical loss term. The data loss term is used to characterize the deviation between the predicted axial growth value output by the prediction model and the true value of the one-year axial growth. The surrogate physical loss term is constructed based on the biomechanical creep characteristics of the cornea and is used to calculate the residual between the predicted topographic map and the theoretical creep topographic map. Parameter optimization steps: Through the backpropagation algorithm, jointly optimize the network weights of the prediction model and a set of globally learnable physical parameters in the proxy physical loss term; For the second training set, the morphological irregularity index of the first corneal topographic image within a preset region of interest is calculated, wherein the morphological irregularity index characterizes the degree of dispersion of the curvature distribution; A univariate learning weight adapter is introduced, which can map the morphological irregularity index into dynamic weight coefficients through nonlinear transformation, so that the proportion of the surrogate physical loss term in the hybrid loss function dynamically fluctuates with the degree of corneal morphological irregularity.

[0016] Compared to black-box training methods that rely solely on data-driven approaches, this patented technology constructs a training architecture jointly driven by physical meaning and data statistics. By introducing a surrogate physical loss term based on corneal creep characteristics into the loss function, and using creep formula parameters as learnable variables for backpropagation updates during optimization, the network can be constrained at the level of physical laws. This allows it to learn feature representations that conform to basic biomechanical principles and have stronger generalization capabilities, solving the problem of overfitting to spurious features in purely data-driven models when training samples are insufficient or unevenly distributed. Simultaneously, through a morphological irregularity index and a meta-learning weight adapter, the strength of the physical constraints on each sample can be dynamically adjusted. For samples with regular corneal morphology that conform to general patterns, physical constraints play a stronger role in training; for samples with highly irregular morphology and large individual differences, physical constraints are automatically weakened, and the model shifts to primarily relying on the data itself for learning. This achieves an adaptive balance of physical constraints, ultimately improving the model's predictive robustness in various complex and extreme cases.

[0017] Optionally, the set of globally learnable physical parameters includes the creep coefficient. Stress index and time index The surrogate physical loss term is calculated based on a modified Norton-Bailey creep formula, which is expressed as: To predict deformation variables, For lens pressure distribution, The creep coefficient is the duration of wear. The stress index and the time index These are globally learnable parameters that are jointly updated via backpropagation during training, rather than pre-set fixed empirical values. It can be constrained within the range of 0.5 to 2.5. It can be constrained to the range of 0.1 to 1.2.

[0018] The deformation of corneal tissue under continuous lens pressure is a typical viscoelastic creep process. Traditional fixed-parameter physical models cannot accurately fit the corneal biomechanical characteristics of different individuals. This invention explicitly employs a modified Norton-Bailey creep formula to construct the physical loss term, defining its core parameters as variables that can participate in global backpropagation optimization. This makes the training process itself an automatic fitting of the "equivalent corneal physical parameters" best suited to the current multi-center data distribution, overcoming the mismatch problem of preset fixed physical experience values ​​in organisms, and greatly enhancing the accuracy and adaptability of physical constraints.

[0019] Optionally, in the parameter optimization step, a physical constraint is imposed on the globally learnable physical parameter: the stress exponent By clipping its gradient during backpropagation, or by adding a penalty term to its corresponding loss term, a first constraint is applied to keep it within the range of 0.5 to 2.5, in order to match the nonlinear viscoelastic response of corneal tissue under lens pressure; for the time exponent A second constraint is applied to keep it within the range of 0.1 to 1.2 to match the creep accumulation effect during nighttime wear; and the creep coefficient is adjusted accordingly. Before performing the backpropagation, a logarithmic transformation is performed to prevent gradient explosion and ensure its positive value.

[0020] Completely unrestricted parameter learning can lead to parameter values ​​escaping their reasonable physical domain, generating singular solutions that cannot be biomechanically interpreted, and causing training instability. By imposing interval constraints based on physiological characteristics on each physical parameter and performing a logarithmic transformation on the creep coefficient, we ensure that the learned parameters always remain within a range with clear biomechanical significance, guaranteeing the interpretability of the model and significantly enhancing the numerical stability and convergence of the training process after introducing physical constraints.

[0021] Optionally, the meta-learning weight adapter is constructed as a sample-specific mapping network, specifically a single-layer fully connected network. Its input vector includes at least the morphological irregularity index, the age of the object to be evaluated, and the baseline axial length of the eye. Its output is processed by a sigmoid activation function to map the output value to a normalized range of 0 to 1 and then multiplied by a preset maximum weight limit, thereby enabling the meta-learning weight adapter to capture the comprehensive impact of individual differences on the credibility of physical constraints.

[0022] Physical constraints are not equally reliable in all patients. For individuals with specific developmental stages or extreme baseline ocular parameters, the applicability of general physical laws decreases. Therefore, this invention constructs a lightweight, single-layer mapping network that not only receives an index representing morphological complexity but also integrates variables reflecting an individual's physiological stage, such as age and baseline axial length. Through a nonlinear transformation of a sigmoid function, it generates a dynamic weight that comprehensively assesses the applicability of physical constraints to the current sample. This elevates the weight decision from a single morphological regularity judgment to a multivariate-driven, individualized confidence assessment, further enhancing the precision and rationality of the adaptive mechanism.

[0023] Optionally, before the dataset construction step, the method further includes a step of numerically reconstructing the original image using a cloud-based dynamic template library. The cloud-based dynamic template library synchronizes color mark configuration files for different device models through an over-the-top update mechanism, mapping the pixel color values ​​of the original image to absolute refractive power values. The method also includes a step of calculating data quality confidence based on the numerical reconstruction. The data quality confidence is obtained by comparing the reconstruction error between the reconstructed refractive power value and the standard physical model; the smaller the reconstruction error, the higher the data quality confidence. The data quality confidence is incorporated as a sample weighting factor into the data loss term. When calculating the total loss function, the data quality confidence is configured to adjust the weight coefficient of the surrogate physical loss term, such that the final physical weight is equal to the product of the sample-specific weight and the data quality confidence, thereby achieving an adaptive degradation mechanism where the lower the physical constraint weight, the worse the data quality.

[0024] Raw images may suffer from degradation and color mark deviation during acquisition and transmission. If low-quality data is indiscriminately input into training, its inherent noise can conflict with physical constraints, thus compromising the stability of model learning. This technical solution introduces a closed-loop data sourcing and quality transfer mechanism at the front end of the training process. A cloud-based template library ensures the accuracy of numerical reconstruction, and a continuous, objective data quality confidence score is calculated based on the reconstruction results. This confidence score serves as a key adjustment factor, simultaneously injected into the weight calculation of both the data loss term and the surrogate physical loss term. In particular, its adjustment of the product of physical constraint weights implements an "adaptive degradation" logic: when the input sample quality is low, the weight of the physical constraints automatically decreases, and the model degenerates to primarily rely on data-driven approaches, preventing unreliable data from being reinforced by incorrect physical laws; when the data quality is high, the physical constraints are fully effective, guiding model learning. This mechanism, for the first time, seamlessly and mathematically transmits and couples the quality information from the data preprocessing stage to the physical constraint strength control of the core training, forming a highly robust training system with uncertainty perception capabilities.

[0025] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the aforementioned prediction methods or model training methods. Thus, the two-stage prediction logic based on surrogate physical constraints and dynamic weight adaptation, along with its training framework, can be packaged as a software product and deployed on a large scale.

[0026] In summary, this invention fundamentally and systematically solves the core technical problems of existing technologies, such as prediction node lag, unavoidable costs of ineffective treatment initiation, and insufficient generalization ability of purely data-driven models in complex biological systems, by decoupling pre-treatment static screening and post-wearing dynamic re-evaluation into two independent prediction models driven by different feature spaces. It also deeply integrates a data-driven deep learning architecture with a surrogate physical constraint training mechanism with adaptive weight adjustment capabilities. The entire technical system forms a complete closed loop from data quality traceability and biomechanical feature constraints to individualized risk decision-making, achieving significant comprehensive technical advantages in terms of the timeliness, accuracy, and robustness to extreme cases in clinical prediction. Attached Figure Description

[0027] To more clearly illustrate the embodiments of this patent, the relevant drawings will be briefly described below. It should be understood that the drawings described below are only for illustrating some embodiments of this patent, and those skilled in the art can obtain many other technical features and connections not mentioned herein based on these drawings.

[0028] Figure 1 A flowchart illustrating the embedding of the prediction method provided in this invention into a clinical workflow; Figure 2 This is a flowchart illustrating the method for predicting the efficacy of orthokeratology in controlling myopia progression based on multimodal data, provided by the embodiments of the invention. Figure 3 Examples of different image modalities and cropping of corneal topography provided in the embodiments of the present invention; Figure 4 Example diagrams of corneal topography provided in embodiments of the present invention in relative and absolute color scale modes; Figure 5 A comparative example diagram of corneal topography provided in the embodiments of the present invention under the states of automatic labeling information retention and removal; Figure 6 This is a flowchart illustrating the adaptive proxy physical constraint training architecture provided in an embodiment of the present invention. Figure 7 This is a schematic diagram of the architecture of the data tracing and confidence soft transfer engine provided in the embodiments of the present invention and its coupling with the training loss function; Figure 8 This is a schematic diagram of the cloud-based dynamic template library crowdsourcing verification and secure OTA update process provided in the embodiments of the present invention; Figure 9 This is a schematic diagram of the first-stage training strategy and parameter cascading process provided in the embodiments of the present invention; Figure 10 This is a schematic diagram of the second-stage training strategy and parameter cascading process provided in the embodiments of the present invention. Detailed Implementation

[0029] The patent will now be described in detail with reference to the accompanying drawings.

[0030] In clinical practice, orthokeratology is widely used to slow the progression of myopia in children and adolescents. Its basic principle is to reshape the anterior surface of the cornea by wearing rigid gas-permeable contact lenses with reverse geometry at night, thereby improving uncorrected visual acuity during the day and slowing the increase in axial length. The increase in axial length is a core indicator for assessing myopia progression; therefore, accurately predicting the amount of axial length growth in a patient over the next year before or in the early stages of treatment is crucial for developing personalized intervention plans.

[0031] In existing technologies, some solutions have attempted to utilize machine learning techniques to process data related to orthokeratology. For example, a typical existing solution constructs a fitting system based on axial growth prediction and corneal information. This solution uses a corneal information acquisition system to obtain the patient's corneal topography and a series of structured clinical information, such as age, refractive power, and axial length. Subsequently, optical character recognition technology in a data intelligent cloud processing system is used to analyze and extract the morphological parameters of the text presented on the corneal topography, such as the corneal eccentricity e-value and central mean curvature. These analyzed low-dimensional parameters, along with the patient's baseline information from refraction examinations, are input into a mathematical model based on ensemble learning methods such as random forests. The core task of this model is to first predict the patient's axial growth one year after wearing orthokeratology lenses, and then use this predicted value as one of the key input features, combined with other clinical information, to ultimately calculate the most suitable orthokeratology lens parameters for the patient, such as lens brand, radius of curvature, and overcorrection power.

[0032] The aforementioned method, which relies on machine learning to predict axial elongation and ultimately recommend lens parameters, can provide a more quantitative reference during fitting than manual trial fitting, thus reducing reliance on the doctor's subjective experience to some extent. Its core approach is to use the prediction of axial elongation as an intermediate variable, with all technical steps ultimately serving the goal of "deriving the optimal lens parameters."

[0033] However, in practical applications, it has been found that the core problem addressed by this technology is how to improve the accuracy of lens parameter recommendations. Its workflow is predicated on the patient already being fitted, and the prediction logic either relies on data only available after lens wearing or fails to move the prediction point forward to the screening stage before treatment begins. This leads to a common negative effect: if a patient's myopia control after orthokeratology lens treatment is already suboptimal (i.e., they belong to the low-benefit group), then the axial elongation prediction and optimal lens parameter recommendation work performed for that patient cannot fundamentally avoid the already incurred and subsequently proven ineffective fitting costs, including but not limited to lens customization fees, discomfort during wear, and follow-up manpower costs. In this case, the computational value of the prediction and recommendation model lies only in improving the "efficiency" of a potentially ineffective fitting process, rather than avoiding costs before they occur.

[0034] From a deeper technical perspective, these solutions suffer from two irreconcilable internal contradictions. First, the basic data used to predict axial elongation and recommend lens parameters is entirely derived from a single examination at the moment of fitting. This means the model can only provide static, one-time predictions at the start of treatment, lacking the ability to dynamically interact with and update predictions through early post-treatment follow-ups. Considering that orthokeratology is a long-term biomechanical remodeling process, the initial corneal topography changes contain extremely rich individual response information. Existing solutions, failing to utilize this temporal information, cannot capture early deviations from the treatment trajectory during prediction. Second, these solutions' utilization of image data remains limited to OCR extraction of low-dimensional parameters or simple grayscale / size transformation. When different manufacturers or different models of corneal topography devices from the same manufacturer use different color scale schemes in their report outputs—such as relative or absolute color scales—or overlay different automatic labeling information on the image, such as system brand logos, auxiliary lines, color bars, and parameter text, artifacts unrelated to the corneal morphology itself are introduced. These artifacts vary greatly across different centers and devices, and are easily mislearned by deep models as characteristics that distinguish samples, leading to a sharp decline in the model's generality across different device environments.

[0035] If one attempts to directly apply another type of post-treatment assessment scheme, such as a system for evaluating corneal topography differences before and after orthokeratology lens fitting, to a pre-treatment prediction scenario, it will encounter obstacles at the technical level. The system's operation relies entirely on the difference analysis between the "pre-fitting topography" and the "post-fitting topography," using convolutional neural networks for region segmentation and center localization to assess the overlap and deviation between the simulated and actual treatment areas. Its output, IOU and Distance indices, are quantitative assessments of the actual positional fit of the lens on the cornea. Since this system fundamentally requires post-fitting topography data to generate the difference map and perform the assessment, this technique cannot be used before treatment. Furthermore, the system's evaluation logic focuses on the static geometric fit between the lens and the cornea, without addressing the prediction and dynamic reassessment of long-term axial elongation, thus failing to provide doctors with data support on when to initiate adjunctive medical interventions. Therefore, neither the scheme for predicting lens parameters nor the scheme for assessing lens fit can constitute a closed-loop decision chain capable of pre-treatment screening and dynamically triggering joint intervention prompts in the early follow-up window.

[0036] In view of this, the embodiments of the present invention aim to provide a solution that can move the prediction window forward to before treatment and achieve joint prediction from before treatment to early follow-up, in order to solve or at least partially alleviate the above-mentioned technical problems.

[0037] First Implementation Method This embodiment provides a method for predicting the efficacy of orthokeratology in controlling myopia progression based on multimodal data. See [link to relevant documentation]. Figure 1 As shown, the prediction method provided in this application can be embedded in the clinical workflow. Through two stages, namely pre-treatment screening and early reassessment, the follow-up duration is preset before treatment initiation and after wearing orthokeratology lenses. Then, at these two key time points, prediction results and intervention prompts are output respectively.

[0038] Specifically, see Figure 2 As shown, this predictive method includes a combination of pre-treatment screening and early reassessment steps.

[0039] In the pretreatment screening step, the first corneal topography image and baseline clinical variables of the subjects to be evaluated are obtained before treatment. The first corneal topography image can be a two-dimensional image containing the spatial distribution of the anterior corneal surface curvature or refractive power, acquired by a corneal topography instrument based on the Scheimpflug principle or Placido disk principle, for example. Its input form can be a standardized image generated through subsequent image preprocessing and standardization steps, such as standardizing the color scale and removing system identifiers. Baseline clinical variables can be low-dimensional structured numerical or categorical data such as age, sex, baseline refractive status (spherical or cylindrical lens), baseline axial length (AL), parental myopia history, and corneal topography instrument-derived parameters (such as central corneal curvature, eccentricity).

[0040] The acquired first corneal topographic image and baseline clinical variables are input into a first prediction model, which is configured to process static baseline features. The first prediction model can be constructed, for example, using a convolutional neural network or visual Transformer with ResNet, DenseNet, EfficientNet, ConvNeXt, or ViT as the visual backbone to extract spatial features from the first corneal topographic image; simultaneously, using, for example, a multilayer perceptron (MLP) to encode the baseline clinical variables; and jointly modeling the extracted image features and encoded structured features through feature-level fusion (e.g., feature vector concatenation, gated fusion, attention fusion, or bilinear pooling). The output of the first prediction model can be a one-year axial length growth prediction and an initial risk suggestion. For example, the one-year axial length growth prediction can be a continuous value representing the predicted AL growth in millimeters within one year; the initial risk suggestion can be a classification probability, such as a poor treatment outcome risk probability, representing the risk that the axial length growth of the subject will exceed a preset clinical threshold after one year, for example, greater than 0.3 mm. Since the pre-treatment screening process relies solely on pre-treatment information, risk stratification can be completed before the orthokeratology lens fitting begins, avoiding the economic costs and time delays caused by having to complete fitting, lens manufacturing, and wearing before the efficacy can be assessed.

[0041] In the early reassessment step, after the subject has worn orthokeratology lenses for a preset follow-up period, a second corneal topography image is acquired. Taking a preset follow-up period of one month as an example, at this time, the cornea has initially completed mid-peripheral remodeling under the pressure of the lens, and the changes in corneal morphology can be clearly reflected on the corneal topography. Difference features characterizing corneal remodeling are constructed based on the first and second corneal topography images. One way to construct these difference features is to convert the first and second corneal topography images into refractive power matrices representing the refractive power values ​​at each coordinate point in space through the aforementioned image preprocessing steps. For example, each pixel value corresponds to a two-dimensional array of refractive power D. Then, element-wise subtraction is performed on the two matrices to obtain a pixel-level difference map. The pixel-level difference map is calculated by subtracting the refractive power matrix of the first corneal topography image from the refractive power matrix of the second corneal topography image to obtain a residual matrix. The sign and magnitude of each element in the residual matrix directly quantify the change in refractive power (in diopters, D) of the corresponding corneal coordinate point before and after follow-up. For example, positive values ​​represent areas where the curvature flattens, and negative values ​​represent areas where the curvature steepens. Furthermore, spatial constraints can be imposed on the scope of interest of the pixel-level difference map, limiting its construction to a preset annular region with a radius of 3mm to 5mm in the central cornea. This annular band corresponds precisely to the mid-peripheral defocus area of ​​the orthokeratology lens, removing interference from the central pupillary area and the far peripheral non-treatment area for difference analysis, focusing feature calculations on the annular band that directly characterizes the defocus ring effect produced by the lens. Another approach to constructing differential features is to utilize a feature extraction network based on a convolutional neural network, such as a Siamese network with shared weights or a 3D convolutional early fusion network, to extract high-dimensional feature vectors from the first and second corneal topographic images. The feature vectors from the two time points are then differencing or concatenated to obtain high-dimensional feature vectors representing temporal differences. These high-dimensional feature vectors extracted using convolutional neural networks can capture abstract morphological and textural differences that go beyond pixel-by-pixel linear variations.

[0042] Subsequently, the first corneal topography image, the second corneal topography image, and the constructed differential features are input into the second prediction model to obtain an updated axial elongation prediction value and intervention prompts. The second prediction model is configured to process dynamic features containing temporal variation information. Its architecture can be, for example, an extension of the backbone network of the first prediction model, such as adding a temporal processing branch or introducing a two-stream network architecture. It jointly encodes the original images and differential features at two time points, and after one or more layers of global pooling and fully connected mapping, outputs an updated axial elongation prediction value. It can also be a continuous value, such as the corrected prediction value after one month of wear.

[0043] The logic for generating intervention prompts may include the following steps: First, obtain the probability of poor efficacy risk output by the second predictive model. For example, this probability of poor efficacy risk can be a probability value mapped to the interval between 0 and 1 by the model's terminal through a sigmoid activation function, representing the likelihood of poor efficacy for the subject being evaluated after one year. Then, the probability of poor efficacy risk is compared with a dynamic threshold. Comparison. When the probability of poor treatment outcome exceeds the dynamic threshold. When necessary, suggestions for combined drug interventions are generated, such as suggesting that low-concentration atropine eye drops could be considered as an adjunct intervention in clinical practice; when the risk probability of poor efficacy does not exceed the dynamic threshold. At that time, recommendations for continued routine follow-up with orthokeratology lenses are generated. Dynamic threshold. The optimal threshold is determined based on the validation set. Specifically, the validation set contains several historical samples with known treatment outcomes, where each sample is labeled as either poorly treated (positive) or well-treated (negative). The labeling criteria could be, for example, whether the one-year AL growth is greater than 0.3 mm. The Youden index for the validation set is defined as the sum of sensitivity and specificity minus 1, where sensitivity is the proportion of correctly identified positive samples and specificity is the proportion of correctly identified negative samples. Dynamic threshold. The determination logic is as follows: iterate through several candidate probability thresholds (e.g., 0.1 to 0.9, with a step size of 0.01). For each candidate threshold, calculate the Youden exponent of the validation set at that threshold, and select the candidate probability threshold at which the corresponding Youden exponent reaches its maximum value as the dynamic threshold. This determination method makes It can be recalibrated as new data accumulates (e.g., every 500 new validation samples), thereby dynamically adapting to changes in population characteristic distribution and preventing fixed thresholds from becoming invalid after data characteristics drift.

[0044] A preprocessing step can be included before the pretreatment screening and early reassessment steps to enhance the model's robustness to differences in multi-source devices. This preprocessing step may include periodically synchronizing color scale configuration files released by device manufacturers via an encrypted API interface. Specifically, after two-way authentication via HTTPS, the latest color scale mapping table corresponding to the device model can be downloaded from a cloud server. These mapping tables can be structured data in formats such as XML or JSON, defining the mapping rules from color values ​​(e.g., RGB value ranges) to absolute refractive power values ​​(units of diopters D). Then, based on the mapping function in the color scale configuration file, the pixel color values ​​in the relative color scale mode of the original image are restored to absolute refractive power values. For example, in the relative color scale mode of the original image, the color distribution is automatically scaled according to the minimum and maximum values ​​of the examination; therefore, the same color value may represent different values ​​in different absolute refractive power scenarios. After a unified standard color scale mapping, the color value of each pixel is converted to a value in physical absolute refractive power units, thereby eliminating non-physiological differences introduced by device display settings. Furthermore, when encountering situations where the device model does not match in the local database, the preprocessing steps can also include a cloud-based dynamic template registration step for unknown devices. This step includes: when the device model of the acquired corneal topography image does not match in the local database, the corneal topography image is uploaded to the backend server, and optionally pixel-level desensitization is performed before uploading (e.g., erasing the patient's name and ID information area). After receiving the image, the backend server generates and digitally signs it using its private key, and then pushes a new color-coded configuration file containing the digital signature. After receiving the new color-coded configuration file, the local terminal can use the public key corresponding to the private key to verify the integrity and security of the digital signature, ensuring that the file has not been tampered with and indeed comes from a trusted backend. This mechanism solves the problem of the local preset template library becoming invalid due to rapid updates of device models.

[0045] Based on the aforementioned system architecture and methodology, the first and second prediction models can be deployed as local executable programs, hospital intranet services, or web applications as needed. The local executable program can be configured to read preprocessed corneal topography data and baseline variable files stored on a single machine and directly perform model inference. The hospital intranet service can be deployed on an inference server equipped with a graphics processing unit (GPU), allowing multiple terminals within the hospital to upload data and obtain asynchronous computation results via the internal network. The web application can be integrated into the cloud, allowing authorized users to upload data to the cloud via a browser to complete computation and return a visual report to the interface.

[0046] It is worth mentioning that using systematic ablation experiments to determine the optimal corneal topography image input modality and its corresponding preprocessing strategy can provide a standardized optimal input for the prediction method in this embodiment. Therefore, this embodiment further provides a determination method based on quantitative ablation experiments. This method utilizes an independently constructed retrospective dataset containing image data from 1200 children and adolescents who have completed at least one year of Ortho-K standard follow-up, acquired using different models of corneal topography instruments (such as Pentacam, Sirius, Galilei, etc.) from multiple ophthalmology centers. Each case data includes, for example, the following: Figure 3 , Figure 4 and Figure 5 The four core imaging modalities shown are: refractive tetrad image, iris image, preoperative cataract assessment image, and examination comparison image, along with the true value of axial length (AL) increase one year after lens wear. For each modality, four sets of images with different preprocessing are generated: the first set only undergoes size normalization, retaining the original relative color scale and such... Figure 5 The first group consists of all automatically labeled original images; the second group, based on normalization, maps pixel values ​​to absolute refractive power values ​​according to the color mark configuration file obtained through the encrypted API as described in Example 1, constructing absolute color mark images; the third group, based on normalization, uses a pre-trained semantic segmentation network to erase... Figure 5 The system automatically identifies the images shown in the figure; the fourth group simultaneously performs color standardization and label removal to obtain standardized images that are both physically meaningful and free from display interference. For fair comparison, each input combination (4 modalities × 4 preprocessing schemes = 16) uses the same EfficientNet-B0 structure as the visual backbone network to extract spatial features, paired with an independent multilayer perceptron to process baseline clinical variables. The two are then fused after global pooling to regress the one-year AL growth rate. Under strict 5-fold cross-validation stratified by center, device, and patient, the mean absolute error (MAE) predicting AL growth and the area under the receiver operating characteristic (AUC) for identifying poor treatment outcomes (one-year AL growth > 0.3 mm) were used as evaluation metrics. Key experimental results are shown in the table below.

[0047] The ablation experiment results show that: First, after undergoing dual preprocessing of color standardization and identifier removal, the refractive quadrilateral image significantly outperforms any single preprocessing scheme in both AL prediction accuracy and risk discrimination capability; both preprocessing steps are indispensable. Second, when using the optimal preprocessing scheme, the performance of the refractive quadrilateral image far surpasses that of the iris image, cataract assessment image, and examination comparison image, proving that the spatial morphological information carried by the anterior surface refractive power distribution is most closely related to the efficacy of Ortho-K. Therefore, this embodiment determines that the refractive quadrilateral image undergoing dual preprocessing of color standardization and systematic identifier removal is the sole standard image input for all subsequent prediction processes (including pretreatment screening and early reassessment in Example 1, and model training in Example 2). This selection ensures the highest predictive performance while maximizing the simplification of input dimensions.

[0048] Second Implementation Method This embodiment provides a model training method for training a first prediction model and / or a second prediction model involved in the embodiment. See also Figure 6 As shown, compared with the prior art, the main improvement of the training method in this embodiment is that an adaptive proxy physical constraint architecture is introduced into the training objective, abandoning the purely data-driven loss function. This allows the model to be guided to comply with the creep law of corneal biomechanics while learning data patterns, thereby solving the problem of insufficient generalization ability of the purely data-driven model for extreme cases due to individual differences in corneal mechanical characteristics.

[0049] Specifically, the model training method includes steps for constructing a dataset, defining a loss function, and optimizing parameters.

[0050] In the dataset construction step, two training sets were constructed by collecting cases of children and adolescents from several ophthalmology centers who underwent orthokeratology surgery for the first time and completed at least one year of standardized follow-up. The first training set contained several samples, each consisting of a pre-treatment first corneal topography image (K). inputThe first corneal topography image consists of baseline clinical variables and is labeled with the one-year true axial length growth (AL value) for the corresponding patient. The one-year true axial length growth can be, for example, the increase in axial length measured by a biometer from baseline to one year after treatment, in millimeters. Baseline clinical variables include structured data such as age, sex, baseline spherical power, baseline cylindrical power, baseline axial length, and parental myopia history. The first corneal topography image can be a standardized topography numerical matrix or image after color scaling and system label removal. The second training set also contains several samples. Each sample includes the first corneal topography image and baseline clinical variables, as well as a second corneal topography image obtained after a preset follow-up period (e.g., 1 month), and features characterizing the morphological changes between the first and second corneal topography images. The differential features can be, for example, the pixel-level difference map as described in Embodiment 1, i.e., the residual matrix obtained by subtracting the refractive power matrix of the first corneal topography image from the refractive power matrix of the second corneal topography image; or they can be high-dimensional feature vectors extracted by an independent convolutional neural network. Each sample in the second training set is also labeled with the ground truth value of axial growth over one year corresponding to the follow-up period. By constructing a second training set with differential features, the second prediction model can also be trained using the training architecture of this embodiment.

[0051] In the loss function definition step, a hybrid loss function is defined, which includes at least a data loss term and a surrogate physical loss term. The data loss term characterizes the deviation between the predicted axial growth value output by the prediction model and the true axial growth value over one year. Its specific form can be, for example, mean squared error (MSE) or mean absolute error (MAE). For the first prediction model, the deviation from the first training set is used; for the second prediction model, the deviation from the second training set is used. The surrogate physical loss term is constructed based on the corneal biomechanical creep characteristics and is used to calculate the residual between the predicted topographic map output by the intermediate layer of the prediction model and the theoretical creep topographic map generated based on creep theory. Since the change in corneal morphology after orthokeratology is essentially a long-term creep process under constant lens pressure, empirical formulas describing viscoelastic creep behavior in materials science can be used to construct physical constraints.

[0052] The calculation of this surrogate physical loss term is based on the modified Norton-Bailey creep formula, which is expressed as: in, This represents the predicted creep deformation variable (e.g., the change in refractive power) at location r on the corneal surface. This represents an estimate of the local pressure distribution applied by the lens at that location (which can be obtained through finite element simulation or standard lens design parameters). This indicates the duration of wear, for example, using 365 days to correspond to a one-year prediction; while , , These are the creep coefficient, stress exponent, and time exponent, respectively. Unlike other approaches that treat formula parameters as fixed empirical constants, in this embodiment, the creep coefficient... Stress index and time index These three parameters are set as globally learnable parameters. During training, they are jointly optimized by the backpropagation algorithm along with the network weights of the prediction model, rather than being pre-assigned fixed values. The advantage of this approach is that, due to differences in the arrangement, water content, and cross-linking degree of corneal collagen fibers among individuals, their equivalent creep parameters will inevitably vary. However, as global parameters, the model can fit a set of "optimal compromise equivalent parameters" that minimizes the overall physical loss on a large number of training samples, thus introducing an effective inductive bias into the data. One specific method for calculating the surrogate physical loss term is as follows: i.e., obtain the predicted topographic map Compared with the input topographic map Subtract theoretical creep deformation The square of the Euclidean norm at all spatial locations between the theoretical residual topographic maps obtained later.

[0053] To address the varying degrees of adherence to physical constraints across different corneal morphologies, a Morphological Irregularity Index (MII) and a linked meta-learning weight adapter are introduced during training. Specifically, for samples in the second training set (since the first training set lacks a post-wearing corneal topography image, the surrogate physical loss can be omitted, but the adapter concept can still be used to adjust other regularization terms), the MII of the first corneal topography image within a predefined region of interest is first calculated. This index characterizes the dispersion of corneal curvature distribution. One method for calculating MII is as follows: a ring-shaped band with a central corneal radius of 3mm to 5mm is selected as the region of interest, and the ratio of the standard deviation to the mean of the refractive power values ​​at each coordinate point within this region is calculated. A higher MII value indicates a greater deviation between the corneal morphology and the ideal, regular "bulleye" shape. Then, a meta-learning weight adapter is introduced, which maps the MII to dynamic weight coefficients via a nonlinear transformation. This allows the proportion of the surrogate physical loss term in the hybrid loss function to dynamically fluctuate with the degree of corneal morphological irregularity. Specifically, the meta-learning weight adapter is constructed as a sample-specific lightweight mapping network, such as a single-layer fully connected network. The input vector of this mapping network includes at least the morphological irregularity index (MII), the age of the subject being evaluated, and the baseline axial length, to capture the combined influence of individual factors other than morphological irregularity on the reliability of physical constraints. Optionally, parameters such as average curvature can also be included in the input vector. The output of the mapping network is processed by a sigmoid activation function (e.g., Sigmoid), mapped to a normalized range of 0 to 1, and then multiplied by a preset maximum weight limit. Finally, sample-specific physical loss weights are obtained. Therefore, the final form of the hybrid loss function can be expressed as: when When the value tends to be large, the model is more inclined to believe in physical constraints; when... As the data gets smaller, the model becomes more reliant on fitting the data itself.

[0054] In the parameter optimization step, the aforementioned hybrid loss function is optimized using the backpropagation algorithm. Optimization is performed by jointly updating the network weights of the prediction model and a set of globally learnable physical parameters in the proxy physical loss term. , , To ensure that the learned physical parameters always have biomechanical interpretability and to prevent training instability, physical meaning constraints are imposed on this set of globally learnable physical parameters during the optimization process: for the stress exponent... A first constraint is applied to keep its value within the range of 0.5 to 2.5, which covers the typical nonlinear viscoelastic response range of corneal tissue under lens pressure, for example, when Time represents linear viscoelasticity. Time represents strain hardening effect; for the time exponent A second constraint is applied to keep its value within the range of 0.1 to 1.2 to match the temporal fractal order of creep accumulation effects during nighttime wear; and, regarding the creep coefficient Before backpropagation, a logarithmic transformation is performed, which is the actual storage and update of parameters. When performing forward calculations, take This is to prevent gradient explosion and ensure It is always positive. Numerical range constraints can be implemented using soft constraints (such as adding a penalty term) or hard constraints (such as gradient clipping).

[0055] Preferably, to improve training stability, a two-stage training strategy can be adopted. In the first stage, the parameters of the meta-learning weight adapter are frozen (e.g., their...). (Fixed to 1 or no backpropagation), only the backbone network weights and globally learnable physical parameters of the prediction model are optimized. , , The first stage aims to allow the backbone network to fully learn the data features and initially determine a reasonable set of global physical parameters, unaffected by the instability of the early meta-learning adapter. The second stage unlocks the meta-learning adapter and performs end-to-end joint fine-tuning, enabling the adapter to learn the mapping that adjusts the strength of physical constraints for different sample morphological irregularities. This segmented training significantly reduces the risk of drastic fluctuations in the total loss due to adapter output instability in the early stages of training.

[0056] Corresponding to the first and second prediction models, similar but independent training processes can be used. The second prediction model can also additionally utilize the globally learnable physical parameters saved after the first prediction model has been trained. , , and proxy threshold As an initial prior, parameter cascading is implemented to further improve the prediction accuracy of the 1-month reassessment.

[0057] In this implementation, firstly, by replacing the traditional fixed physical equations with a surrogate creep loss containing learnable parameters, the model no longer needs to artificially assume a specific mechanical constant of the cornea. Instead, it uses a large amount of clinical data to inversely deduce a set of equivalent parameters that best fit the creep assumption under the current data distribution. This is equivalent to allowing the model to discover the corneal deformation pattern itself from the data, while this pattern is forcibly constrained by the loss function within the formal framework of the creep empirical formula, thus balancing physical interpretability and data adaptability. Secondly, by introducing sample-specific weights through a meta-learning weight adapter, the training process achieves a dynamic and smooth switch between data-driven and physical constraints: for patients with regular corneal topography morphology and a typical bull's-eye distribution, their MII is lower, and the adapter output is larger. When the surrogate physical loss term accounts for a high proportion of the total loss, the model relies more heavily on the regularization effect of physical constraints, thus making the prediction results more consistent with biomechanical expectations. However, for corneal morphology that is highly irregular (e.g., with significant asymmetry, abnormal curvature in small areas, or even a tendency towards keratoconus), the MII value is higher, and the adapter output is smaller. The physical loss weights are reduced, and the model relies more on the data features themselves for learning, thus avoiding misleading biases introduced by an inaccurate global physics model for such extreme samples. This adaptive mechanism overcomes the hyperparameter limitations of traditional physics-guided neural networks. The problem requires manual parameter tuning for a uniform and fixed dataset across the entire dataset, eliminating the need for manual global parameter searches. The adapter automatically generates appropriate weights for each sample. Furthermore, the interval constraints and logarithmic transformations applied to the physical parameters ensure the numerical stability of the optimization process and the physiological interpretability of the physical parameters, avoiding physically meaningless solutions such as gradient explosion and negative creep coefficients. Combining these designs, the adaptive surrogate physical constraint training method proposed in this implementation significantly improves the model's generalization ability and robustness in biomedical prediction tasks with scarce data and large individual differences. In particular, when encountering corneal morphological abnormalities or previously rare samples in clinical practice, the system can automatically perceive and reduce the confidence level of the physical assumptions, instead relying on similar patterns in existing data to make more reliable inferences.

[0058] Third Implementation Method This embodiment provides a model training method. Based on the proxy physical constraint training architecture constructed in Embodiment 2, this embodiment makes further improvements, mainly in the introduction of a data tracing and confidence soft transfer mechanism.

[0059] In Implementation Method Two, the globally learnable physical parameters and meta-learning weight adapter endow the model with the ability to dynamically balance data-driven and physical constraints. However, the training process in Implementation Method Two presupposes that the numerical matrix of the input topographic map is accurate and reliable. In real-world multi-center data acquisition scenarios, due to differences in color scales and image rendering caused by different device models, software versions, and even export settings, the refractive power numerical matrix itself may contain varying degrees of error. If these errors are not quantified and participate in the physical loss calculation with the same confidence level, low-quality samples may mislead the estimation of creep residuals in the surrogate physical loss term through inaccurate values, thereby weakening the regularization effect of physical constraints and even causing the globally learnable physical parameters to be compromised. To avoid deviations from the correct approach, this implementation uses a quantitative source verification and confidence soft transfer engine to quantify the reliability of image preprocessing into a continuous confidence score. This confidence score is then directly embedded as a weighting factor into the training loss function to adjust the strength of physical constraints, enabling the model to perceive uncertainties in data quality.

[0060] Specifically, this implementation adds a step of numerically reconstructing the original images using a cloud-based dynamic template library before the dataset construction step, or in the data loading pipeline of the training process. The cloud-based dynamic template library is a database deployed on a cloud server, storing color mark configuration files for numerous different corneal topography device models and software versions. Each color mark configuration file contains a mapping function or lookup table that maps pixel color values ​​(e.g., RGB triples or grayscale values) in the rendered image output by a specific device to absolute refractive power values. The cloud-based dynamic template library continuously synchronizes with the latest color mark data released by device manufacturers through an over-the-air (OTA) update mechanism. The specific OTA update process can be as follows: the model training service framework deployed locally in the hospital or in the cloud sends a query request to the cloud server via an encrypted API interface, such as a two-way certificate authentication link based on HTTPS, during the startup phase or at a preset period. The query request carries a list of version hashes of the color mark configuration files for each model stored locally. After comparing the version information, the cloud server packages the templates with new versions or templates for newly added device models and distributes them to the local machine through the same encrypted channel. The data packet can contain a color code configuration file digitally signed by the developer's private key. After receiving it, the local terminal uses the pre-stored public key paired with the private key to verify the digital signature. Only after confirming that the file is complete and the source is trustworthy can the old configuration be replaced or added to the database as a new template for use in numerical reconstruction.

[0061] Therefore, see Figure 7 As shown, the original corneal topography image is first matched against a local template library. If the device model of a certain original image does not match in the local template library, an unknown device template registration process can be triggered: the original image is uploaded to the backend server after pixel-level desensitization of patient privacy information. After manual and automatic verification by the backend, a new color mark configuration file is generated and pushed to all subscribed clients via OTA. This cloud-based dynamic template registration and OTA collaborative mechanism ensures the long-term timeliness and comprehensiveness of the basic data for numerical reconstruction.

[0062] After obtaining the color mark configuration file corresponding to the original image, the numerical reconstruction step is performed: traversing each pixel within the effective area of ​​the corneal topography in the original image, querying the corresponding absolute refractive power value in the mapping table or mapping function based on its current color value, and replacing it with a floating-point number, ultimately forming a refractive power value matrix corresponding to the pixel grid of the original image. For example, an image output by a Pentacam device in relative color mark mode will have its red to blue range of refractive power dynamically scaled according to the patient's extreme refractive power during that examination; however, after intervention with a color mark profile, each pixel is restored to a value in refractive power (diopters). The absolute physical values ​​are expressed in units of 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 ...

[0063] In obtaining the numerically reconstructed refractive power matrix Next, the data quality confidence level was further calculated. . It is a value range of arrive The continuous values ​​are used to measure the reliability of the valued reconstruction. The calculation can be based on two metrics: reconstruction error. and structural similarity Reconstruction error By comparing the reconstructed refractive power values The expected refractive power value in this coordinate relative to a universal standard physical model applicable to the cornea. It is obtained by the average absolute deviation between them. For example, for a normal cornea, it can be obtained by using a radius of... aspherical coefficient for The expected value is calculated using the sagittal-refractive power conversion relationship of the standard ellipsoidal model. Reconstruction error The formalized expression is as follows: in Effective pixel count These are pixel coordinates. A value close to zero indicates that the reconstruction result highly conforms to the standard corneal physical model, suggesting that the reconstruction process is very reliable; A higher value indicates a deviation from expectations in the reconstruction results, which may stem from incorrect color mark configuration, image artifacts, or extreme irregularity of the cornea itself. Structural similarity. The calculation involves reconstructing the matrix. Normalized and treated as a single image, it is then compared with the original image using a classic structural similarity measure to assess the consistency of spatial structural information under the two representations. Afterwards, and merge into one Fractions. An exemplary fusion formula is as follows: in To achieve the fusion weights, for example, we can take... With a slight emphasis , For example, take the attenuation factor. So that when Reaching the clinically recognized minimum scale for corneal curvature measurement hour, Significantly lowered the confidence level to below .thus, It provides an objective scale for quantifying the quality of input data from a bibliometric perspective.

[0064] After calculating each sample After that, It is incorporated into the training process as a sample weighting factor, and the specific coupling method has two levels. The first level is in the calculation of the data loss term. At that time, As a weighting coefficient: for a single sample Its data loss contribution is its prediction bias multiplied by That is, total data loss. for: in For batch size, It can be, for example, prediction The mean squared error between the increased value and the true value. This weighting effect ensures that the contribution of high-quality reconstructed samples to the data loss is fully preserved, while the contribution of poor-quality reconstructed samples is weakened, thus preventing the model from overfitting or shifting to fit unreliable training labels. The second level, and the key technical loop of this implementation method, It is configured to adjust the weight coefficients of the surrogate physical loss term, thereby achieving an adaptive degradation mechanism where the lower the weight of the physical constraint, the worse the data quality. In implementation method two, the sample-specific weights of the surrogate physical loss term are output by the meta-learning adapter. It is decided. In this embodiment, a method is introduced. Finally, the weights applied to the physical loss term of that sample. Revised to: in The sample weights are the output of the meta-learning adapter. This represents the confidence score of the current sample. When a sample's confidence score is... Approaching At that time, physical constraints are approximated The intensity plays a regularizing role; when Reduced to near due to severe artifacts or mapping errors hour, Also tending to The proxy physical loss term of this sample In effect, it is bypassed, meaning the model degenerates into a purely data-driven mode for that sample, preventing a creep residual calculated based on an incorrect numerical matrix from misleading the global physical parameters. Update. This is achieved through the assessment of data quality. As a multiplicative adjustment factor for the physical constraint weights, a "confidence-physical constraint" coupling interface was constructed, connecting image preprocessing and core model training. This coupling mechanism endows the entire training system with an endogenous uncertainty perception capability: the system can not only perceive the suitability of the sample's own shape to the physical constraints through the meta-learning adapter, but also... The measurement reliability of the input data itself is perceived, and the effective strength of the physical constraint is determined by the product of the two.

[0065] This implementation method reconstructs and continues through metrological traceability. This significantly improves the training robustness and reproducibility of the model when it includes real-world clinical imaging data from multiple sources and of varying quality levels. Compared to traditional preprocessing strategies that directly discard low-quality samples or use hard threshold filtering, this approach employs a flexible weighting mechanism that avoids data waste and sample selection bias. Furthermore, because... The degradation logic ensures that, even in extreme situations such as outdated equipment or human-induced errors, the training process can smoothly suppress the interference of poor-quality data on physical hypotheses without interrupting training. This prevents the globally learnable physical parameters from diverging due to localized erroneous inputs, thus guaranteeing the stable convergence of the model. Furthermore, the OTA update mechanism secured by digital signatures ensures the integrity and credibility of the color-coded configuration file itself, providing a closed-loop guarantee for the security of the entire data traceability chain.

[0066] Fourth Implementation Method This embodiment provides a model training method. Based on the data tracing and confidence soft transfer engine described in Embodiment 3, this embodiment further improves the update and security mechanisms of the cloud-based dynamic template library.

[0067] In Implementation Method 3, the local terminal can obtain color scale configuration files published by the device manufacturer or central administrator from the cloud via Over-the-Top (OTA) update mechanism. However, the completeness of this mechanism depends on whether the external system provides templates for new devices in a timely manner. Clinically, corneal topography instruments undergo frequent model iterations, and the same model may also adjust its rendering color scales between different software versions. If the cloud library consistently fails to cover a specific unknown device model, images from that device will continue to run in the default linear mapping or low-confidence mode, thus failing to fully utilize the system's ability to accurately process all data sources. In view of this, see [link to relevant documentation]. Figure 8 As shown, the core improvement of this implementation lies in introducing a collaborative workflow of crowdsourced verification and strong security digital signature for the cloud-based dynamic template library, enabling the template library to have the ability to self-expand and terminal security verification.

[0068] Specifically, the cloud-based dynamic template library has been improved to not only support periodic configuration synchronization from device manufacturers or central servers, but also support a crowdsourced verification and reporting mechanism on the end-user side. This mechanism allows an authorized user or system to automatically trigger an unknown device reporting process when a client, such as the in-house system of a partner hospital, detects that the device model of the current corneal topography image is not matched in either the local or cloud template libraries during the image numerical reconstruction step. This process first performs pixel-level desensitization operations on the original image frame, such as automatically identifying and erasing protected health information such as patient name, medical record number, and date of birth contained in the image edges or corners, retaining only the pure corneal topography color mark area and necessary device metadata, such as manufacturer name, model string, and software version number, to comply with medical data privacy compliance requirements.

[0069] The anonymized images are uploaded to the backend server. The backend server has an automated verification pipeline, the main steps of which are: segmenting the color mark regions in the uploaded images and extracting an initial estimate of their color-value correspondence, which can be obtained by recognizing color mark bars or inferring from known color mark formats; then, applying this initial mapping to the images to generate a temporary reconstructed numerical matrix. Then With a general standard physical model, such as a central radius of curvature of 7.8 mm and an aspherical coefficient The expected refractive power values ​​at each coordinate point of the standard corneal ellipsoid model of -0.26. Compare and calculate reconstruction error And structural similarity (SSIM). If the automatic verification result meets the preset strict pass criteria, for example... and If the initial mapping curve or lookup table is not found, the backend server can automatically generate a new color code configuration file. If the automatic verification result is in an ambiguous range or fails, the task is transferred to the manual verification queue, where the backend administrator manually labels or reviews the color code mapping relationship of the image. Once approved, the backend server digitally signs the color code configuration file using its private key. For example, the private key can calculate a hash value on the binary content of the configuration file, encrypt the hash using the RSA-SHA256 algorithm to generate a signature file, or the signature can be directly appended to the end of the configuration file. Afterward, the new color code configuration file and its digital signature are stored in the cloud template library and become legitimate content for subsequent OTA (Over-The-Air) pushes.

[0070] On the client side, the digital signature verification mechanism is forcibly embedded into the template loading process. The client has a pre-installed public key certificate or public key hash value that is strictly paired with the backend private key. When the client obtains a new template file and its digital signature pushed from the cloud through periodic OTA queries or proactive retrieval, it first calls the signature verification module to decrypt or verify the digital signature using the public key. If the signature verification passes, it means that the content of the template file has not been tampered with by any third party after signing and indeed comes from a trusted backend private key holder. Only then does the client allow it to be officially loaded into the local template library and replace the old version. If the signature verification fails, even if the file format is correct, the client will refuse to load the template and may issue a security warning to the user or system administrator. This mechanism cryptographically eliminates the risk of malicious color-coded configuration files being implanted into the client due to cloud server intrusion or man-in-the-middle attacks, thereby systematically polluting the numerical reconstruction results.

[0071] More preferably, the release of the new template can adopt a canary release strategy. After the new template is signed, the cloud server does not immediately push it to all clients. Instead, it first marks it as "under canary testing" in the configuration and pushes it only to a preset proportion, such as 5% of randomly selected online clients, during a specific time period, such as a low-load window at night. During the subsequent preset monitoring period, such as 24 hours, the monitoring module collects error logs and abnormal indicators reported by the canary clients. For example, it monitors whether the failure rate of numerical reconstruction of the image of this model or the proportion of DQS below 0.5 is abnormally high compared to the old version or the default mapping. If the indicators are normal, the canary release scope is gradually expanded until it is pushed to all clients; if the indicators are abnormal, the rollback mechanism is automatically triggered, pushing the rollback instruction to the received canary clients, restoring them to the previous version of the template, and marking the new template as "blocked" in the cloud, while notifying the administrator for manual review. This strategy avoids the systemic risk of a large number of clients simultaneously failing to perform numerical reconstruction due to defects in the color code configuration file itself.

[0072] Through the aforementioned dual mechanisms of crowdsourcing verification and digital signatures, the cloud-based dynamic template library has been upgraded from a static, centralized update system to a secure, self-evolving lifecycle management system. Its technical advantages are twofold: First, it addresses the industry pain point that device iteration speed far exceeds the speed of centralized template maintenance, enabling any terminal to act as a "probe" for the system's senses when encountering unknown devices, proactively initiating crowdsourcing data entry and achieving automatic growth in coverage over time. Second, the combination of digital signatures and canary releases ensures security and reliability during the self-evolution process, providing a reliable underlying data foundation for the calculation of Data Quality Confidence (DQS) in Implementation Method Three. Because DQS has a leverage effect on the adjustment of subsequent physical loss weights, if the source color mark mapping itself is tampered with or contains errors, it will systematically lead to DQS distortion for a large number of samples, thereby inhibiting the correct role of physical constraints in training. This implementation method's improvements provide this vulnerable link with high-strength security and fault-tolerant resilience.

[0073] Fifth Implementation Method This embodiment provides a model training method. Based on the adaptive surrogate physics constraint training architecture containing globally learnable physical parameters and a meta-learning weight adapter constructed in Embodiment 2, this embodiment further provides a two-stage training strategy and parameter cascading method to further stabilize the training process and improve the temporal synergy between the first prediction model and the second prediction model.

[0074] Implementation method two describes how to jointly optimize network weights and physical parameters using a hybrid loss function. , , And the meta-learning adapter. In the early stages of training, since both the network weights and the meta-learning adapter are in a randomly initialized state, the adapter outputs sample-specific weights. It may be highly unstable. If the adapter and physical parameters are jointly optimized from scratch, there will be issues with the physical parameters. , , There is a risk of being misled into a biased direction due to unreasonable physical loss weights in the early stages. This implementation overcomes the above-mentioned shortcomings by explicitly executing the training process in stages.

[0075] Specifically, see Figure 9 As shown, in the first training phase, the parameters of the meta-learning weight adapter are frozen or its output is kept at a preset constant value (e.g., let...). (Fixed to 0.5 or 1.0), meaning it does not participate in backpropagation updates. In this stage, only the backbone network weights and globally learnable physical parameters of the prediction model are optimized. , , At this point, the physical loss term participates in the joint optimization with constant weights, forcing the backbone network and physical parameters to learn the main laws of corneal creep implicit in the data under stable physical constraints, and to fit a set of global parameters with reasonable physical meaning. Once the training loss in the first stage converges and tends to stabilize, see [link to relevant documentation]. Figure 10 As shown, the second training phase begins: the parameters of the meta-learning weight adapter are unlocked, and its joint end-to-end fine-tuning with the backbone network and physical parameters is restarted. In this second phase, the meta-learning adapter extracts multivariate information such as MII from the already reasonably optimized feature space of the backbone network, learning how to finely adjust the physical constraint strength for each sample to further improve the model's ability to handle irregular corneal samples. Since the global physical parameters are now located in the neighborhood of the near-optimal solution, the adapter-induced... Fluctuations do not easily cause physical parameters to deviate from the physically meaningful region, thus achieving a synergy between stable convergence and fine-tuning.

[0076] To enhance the temporal continuity between the first predictive model (pre-treatment screening) and the second predictive model (early reassessment), this implementation further introduces a parameter cascading mechanism. In Implementation One, the initial risk recommendations and pre-treatment predicted values ​​output by the first predictive model provide the foundation for the second predictive model; in the model training level of this implementation, this cascading extends to the explicit transfer of internal model parameters. Specifically, firstly, the first predictive model is fully trained using Implementation Two or the aforementioned two-stage training strategy. After the first predictive model training is completed, the globally learnable physical parameters obtained from the training are... , , The final value and the dynamic threshold determined based on the validation set The values ​​are exported from the training checkpoint file. When the second prediction model builds its training initialization, it reads these exported values ​​and directly uses them as the initial values ​​for the corresponding parameters in the second prediction model. For example, the creep coefficient in the surrogate physical loss calculator of the second prediction model. Stress index and time index Training begins with the final value of the first prediction model as the starting point; simultaneously, the second prediction model determines its own dynamic threshold. At that time, the first prediction model can be used. Instead of starting from a completely random range, the candidate thresholds are searched traversally within a narrower range (e.g., ±0.1) around the initialization center of the validation set search. The remaining network parts of the second prediction model (e.g., additional two-stream branches or 3D convolutional layers for handling temporal difference features) are still randomly initialized.

[0077] The rationale and technical motivation behind the aforementioned parameter cascade mechanism lies in the fact that the biomechanical creep characteristics of the cornea are inherent material properties of the corneal tissue itself. Although the corneal morphology changes before and after wearing orthokeratology lenses, the creep coefficient of its underlying material remains constant. Stress index and time index The viscoelastic behavior it represents exhibits temporal continuity over the statistical average of the same individual or population. The first predictive model, learning global physical parameters from large-scale pre-treatment baseline data and one-year outcome data, has already condensed the equivalent pattern information of long-term corneal creep under lens pressure in the current population. The second predictive model faces small sample follow-up data, and one-year outcome labels are typically scarcer than baseline data. If trained from random initialization, its physical parameters may fail to converge to a reasonable physical range due to insufficient data. However, by inheriting physical parameters that have already converged on large datasets, the second predictive model is essentially trained with a "prior knowledge" that conforms to physical common sense. Its training process can converge faster and is less prone to overfitting on small follow-up samples. Regarding dynamic thresholds... In addition, cascading also ensures that the risk decision thresholds of the two stages have statistical consistency over time, avoiding the disconnect between the second prediction model and the first prediction model at the clinical decision node caused by the second prediction model independently searching for thresholds from scratch.

[0078] In summary, this implementation method addresses the early training oscillation issue from an optimization perspective by employing a two-stage training strategy and a parameter cascading method, and establishes a physical knowledge transfer channel between the two prediction nodes—pre-treatment and early follow-up—from a model collaboration perspective. Compared to training the two models completely independently, this approach allows the physical constraints of the first prediction model to directly empower the second prediction model, improving the overall performance of the two-stage prediction framework in terms of training efficiency and cross-time generalization ability.

[0079] Sixth Implementation Method This embodiment provides a computer program product comprising computer program instructions executable by a processor, which, when executed, implement the steps of any of the above embodiments. The program product can be stored on a non-volatile computer-readable medium, such as a solid-state drive of a hospital server or a persistent volume of a cloud virtual machine. It completes preprocessing, numerical reconstruction, forward inference, and clinical suggestion output of the input image by calling the serialized weight files and associated color mark configuration library of the first and second prediction models via software.

[0080] Finally, it should be noted that those skilled in the art will understand that many technical details have been presented in the embodiments of this patent to facilitate a better understanding of the invention. However, even without these technical details and various variations and modifications based on the above embodiments, the technical solutions claimed in the claims of this patent can be substantially achieved. Therefore, in practical applications, various changes can be made to the above embodiments in form and detail without departing from the spirit and scope of this patent.

Claims

1. A method for predicting the efficacy of orthokeratology in controlling myopia progression, characterized in that, include: Pre-treatment screening steps: Obtain the first corneal topography image and baseline clinical variables of the subject to be evaluated before treatment, and input them into the first prediction model to obtain the one-year axial elongation prediction value and initial risk recommendations; Early reassessment steps: After the subject to be assessed has worn orthokeratology lenses for a preset follow-up period, a second corneal topography image is acquired; based on the first and second corneal topography images, differential features characterizing corneal remodeling are constructed, and the first, second, and differential features are input into a second prediction model to obtain updated axial elongation prediction values ​​and intervention prompts; The first prediction model is configured to process static baseline features, while the second prediction model is configured to process dynamic features that include time-series change information; The difference features include at least a pixel-level difference map or a high-dimensional feature vector extracted based on a convolutional neural network. The pixel-level difference map is a residual matrix obtained by subtracting the refractive power matrix of the first corneal topographic image from the refractive power matrix of the second corneal topographic image.

2. The method according to claim 1, characterized in that, In the step of constructing differential features characterizing corneal remodeling, the construction of the pixel-level difference map is restricted to a preset annular region in the center of the cornea; The preset annular area is an annular band with a radius of 3mm to 5mm, and the annular band corresponds to the mid-peripheral defocus area of ​​the orthokeratology lens.

3. The method according to claim 1, characterized in that, The logic for generating the intervention prompt includes: Obtain the probability of poor treatment outcome output by the second prediction model; When the probability of poor therapeutic effect exceeds the dynamic threshold At the same time, suggestions for combined drug-assisted interventions are generated; The dynamic threshold The optimal threshold is determined based on the validation set, which contains several historical samples with known therapeutic outcomes. The Youden index of the validation set is defined as the sum of sensitivity and specificity minus 1. The dynamic threshold When iterating through several candidate probability thresholds, the candidate probability threshold that corresponds to the maximum value of the Youden index is selected.

4. The method according to claim 1, characterized in that, Prior to the pretreatment screening step and the early reassessment step, the procedure also includes a step of periodically synchronizing the color mark configuration file released by the device manufacturer via an encrypted API interface, and restoring the pixel color values ​​in the relative color mark mode of the original image to absolute refractive power values ​​based on the mapping function in the color mark configuration file, so as to eliminate non-physiological differences introduced by the device display settings.

5. The method according to claim 4, characterized in that, It also includes a cloud-based dynamic template registration step for unknown devices, including: when the device model of the acquired corneal topography image does not match in the local database, the corneal topography image is uploaded to the backend server, and a new color mark configuration file is generated by the backend server and digitally signed using its private key and pushed. The new color mark configuration file contains a digital signature and is used to verify the integrity and security of the new color mark configuration file on the local terminal.

6. A model training method, characterized in that, For training a first prediction model and / or a second prediction model in any one of the prediction methods described in claims 1 to 5, the model training method includes: Dataset construction steps: Collect follow-up data of multicenter orthokeratology surgery to construct the first training set and the second training set; The first training set includes the first corneal topography image and baseline clinical variables before treatment, and is labeled with the true value of axial elongation over one year. The second training set includes the first corneal topography image, baseline clinical variables, the second corneal topography image after wearing for a preset duration, and differential features characterizing morphological changes between the first corneal topography image and the second corneal topography image, and is labeled with the true value of axial growth for one year corresponding to the follow-up duration. Loss function definition steps: Define a hybrid loss function, which includes at least a data loss term and a surrogate physical loss term. The data loss term is used to characterize the deviation between the predicted axial growth value output by the prediction model and the true value of the one-year axial growth. The surrogate physical loss term is constructed based on the biomechanical creep characteristics of the cornea and is used to calculate the residual between the predicted topographic map and the theoretical creep topographic map. Parameter optimization steps: Through the backpropagation algorithm, jointly optimize the network weights of the prediction model and a set of globally learnable physical parameters in the proxy physical loss term; For the second training set, the morphological irregularity index of the first corneal topographic image within a preset region of interest is calculated, wherein the morphological irregularity index characterizes the degree of dispersion of the curvature distribution; A univariate learning weight adapter is introduced, which can map the morphological irregularity index into dynamic weight coefficients through nonlinear transformation, so that the proportion of the surrogate physical loss term in the hybrid loss function dynamically fluctuates with the degree of corneal morphological irregularity.

7. The model training method according to claim 6, characterized in that, The set of globally learnable physical parameters includes the creep coefficient. Stress index and time index The surrogate physical loss term is calculated based on a modified Norton-Bailey creep formula, which is expressed as: ； To predict deformation variables, For lens pressure distribution, The creep coefficient is the duration of wear. The stress index and the time index These are globally learnable parameters that are jointly updated via backpropagation during training, rather than pre-set fixed empirical values.

8. The model training method according to claim 7, characterized in that, In the parameter optimization step, a physical constraint is imposed on the globally learnable physical parameter: the stress exponent By pruning its gradient during backpropagation, or by adding a penalty term to its corresponding loss term, a first constraint is applied to keep it within the range of 0.5 to 2.5 to match the nonlinear viscoelastic response of corneal tissue under lens pressure; for the time exponent A second constraint is applied to keep it within the range of 0.1 to 1.2 to match the creep accumulation effect during nighttime wear; and the creep coefficient is adjusted accordingly. Before performing the backpropagation, a logarithmic transformation is performed to prevent gradient explosion and ensure its positive value.

9. The model training method according to claim 6, characterized in that, The meta-learning weight adapter is constructed as a sample-specific mapping network, which is specifically a single-layer fully connected network. Its input vector includes at least the morphological irregularity index, the age of the object to be evaluated, and the baseline axial length of the eye. Its output is processed by a sigmoid activation function to map the output value to a normalized range of 0 to 1 and then multiplied by a preset maximum weight limit. This enables the meta-learning weight adapter to capture the comprehensive impact of individual differences on the credibility of physical constraints.

10. The model training method according to claim 6, characterized in that, Prior to the dataset construction step, a step of numerically reconstructing the original image using a cloud-based dynamic template library is also included; The cloud-based dynamic template library synchronizes color mark configuration files for different device models through an over-the-top update mechanism, mapping the pixel color values ​​of the original image to absolute refractive power values. It also includes the step of calculating the confidence level of data quality based on the numerical reconstruction; The data quality confidence level is calculated by comparing the reconstruction error between the reconstructed refractive power value and the standard physical model. The smaller the reconstruction error, the higher the data quality confidence level. The data quality confidence level is incorporated as a sample weighting factor into the data loss term; When calculating the total loss function, the data quality confidence is configured to adjust the weight coefficient of the surrogate physical loss term, such that the final physical weight is equal to the product of the sample-specific weight and the data quality confidence, thereby realizing an adaptive degradation mechanism in which the worse the data quality, the lower the weight of the physical constraint.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 10.

Images