Big data-based orthokeratology lens aided design method and system

Abstract

The application belongs to the technical field of orthokeratology lens design, and discloses an orthokeratology lens auxiliary design method and system based on big data, which comprises the following steps: first, obtaining historical fitting data sets containing preoperative / postoperative corneal parameter maps and lens design target values; then, calculating the local response quantity of each grid point of each sample corneal map based on the preoperative and postoperative corneal parameter maps to represent the change of the neighborhood corneal parameters; then, obtaining the original sensitivity score of each grid point according to the correlation strength of the local response quantity and the lens design target value; performing spatial smoothing processing and mapping to a preset weight range to generate a local area sensitivity weight map; subsequently, weighting the current patient's corneal characteristics with the weight map to obtain a weighted feature set; finally, inputting the pre-trained height difference regression model to predict the height difference design parameters of the current patient's orthokeratology lens. The application can accurately adapt to the local irregular cornea and improve the prediction accuracy and reliability of the design parameters.

Description

Technical Field

[0001] This invention relates to the field of big data-based orthokeratology lens design-assisted methods and systems, specifically to a big data-based orthokeratology lens design-assisted method and system. Background Technology

[0002] Orthokeratology lenses, a non-surgical medical device for correcting myopia, reshape the cornea overnight to improve uncorrected visual acuity during the day and are widely used in clinical myopia control. Their core design lies in precisely matching the patient's corneal shape and optimizing the height difference between the lens and the cornea to ensure both effective reshaping and wearing safety.

[0003] However, the human cornea is not an ideal, regular curved surface. Clinically, many patients exhibit localized irregularities such as mild astigmatism, corneal asymmetry, and non-uniform response of the astigmatic axis. Traditional orthokeratology lens design methods employ an "averaging" parameter processing logic, treating the entire corneal surface as a uniform whole. This fails to consider the varying impacts of different areas on the reshaping effect, leading to insufficient accuracy in height difference prediction. This is particularly problematic for patients with locally irregular corneas, easily resulting in issues such as poor lens fit, unsatisfactory reshaping effects, and even corneal damage.

[0004] In existing technologies, some design assistance methods rely on conventional regression models or simple feature weighting strategies, failing to establish a correlation mechanism between local corneal regions and reshaping effects. They also cannot effectively characterize regional sensitivity differences caused by factors such as astigmatism axis direction and palpebral fissure compression. Furthermore, the lack of a mechanism to assess the reliability of weight estimation further affects the stability and accuracy of the design parameters.

[0005] With the application of big data and machine learning technologies in the medical field, it has become possible to mine the correlation between corneal morphology and design parameters based on historical case data. However, how to specifically address the prediction bias caused by local irregularities of the cornea and construct a scientific regional sensitivity quantitative model remains an urgent problem to be solved in the field of orthokeratology lens assisted design. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for assisting in the design of orthokeratology lenses based on big data, so as to solve the problems mentioned in the background art.

[0007] According to one aspect of this application, a method for assisting in the design of orthokeratology lenses based on big data is provided, comprising the following steps:

[0008] Obtain a historical lens fitting dataset, which includes multiple samples, each of which includes a preoperative corneal parameter map, a postoperative corneal parameter map, and a lens design target value;

[0009] Based on preoperative and postoperative corneal parameter maps, the local response of each sample at each grid point on the corneal map is calculated. The local response represents the change of corneal parameters in the neighborhood centered on that grid point.

[0010] Based on the correlation strength between the local response of all samples and the lens design target value, the original sensitivity score of each grid point is calculated;

[0011] The original sensitivity score is spatially smoothed and mapped to a preset weight range to generate a local area sensitivity weight map with the same size as the corneal parameter map.

[0012] The corneal features of the current patient are weighted based on the local region sensitivity weight map to obtain a weighted feature set;

[0013] The weighted feature set is input into a pre-trained height difference regression model to predict the height difference design parameters for the orthokeratology lens used for the current patient.

[0014] Preferably, the calculation of the raw sensitivity score for each grid point includes:

[0015] Calculate the statistical correlation coefficient between the local response of all samples at each grid point and the lens design target value;

[0016] In addition, calculate the robust regression coefficients of the local response of all samples at each grid point on the lens design target value;

[0017] The absolute value of the statistical correlation coefficient and the normalized value of the robust regression coefficient are weighted and fused to obtain the original sensitivity score.

[0018] Preferably, calculating the local response of each sample at each grid point includes:

[0019] Define a physical neighborhood centered on the current grid point;

[0020] A Gaussian kernel function is used to assign weights to the grid points in the neighborhood, with grid points closer to the center grid point having higher weights.

[0021] The difference aggregation between the preoperative and postoperative corneal parameter maps within the weighted neighborhood is calculated and used as the local response value of that grid point.

[0022] Preferably, based on the confidence level of the statistics used in calculating the original sensitivity score, an uncertainty map corresponding to the local region sensitivity weight map is generated, wherein the uncertainty map characterizes the estimation reliability of the sensitivity weight of each grid point.

[0023] Preferably, the corneal features of the current patient are weighted using the local region sensitivity weight map, including:

[0024] Read the uncertainty map corresponding to the local region sensitivity weight map;

[0025] Based on the uncertainty value of each grid point in the uncertainty graph, the weighting effect of the local region sensitivity weight graph is adaptively adjusted. For grid points with uncertainty values ​​higher than a preset threshold, the weighting effect of the sensitivity weight is weakened or ignored.

[0026] The adaptively adjusted weights are combined with the corneal features of the current patient to generate the final feature set used as input to the regression model.

[0027] Preferably, the pre-trained height difference regression model is a tree model, which is trained using the following method:

[0028] The historical feature set weighted by the local region sensitivity weight map and its corresponding lens design target value are used as training data;

[0029] Model optimization is performed using a loss function robust to outliers;

[0030] During training, the weighted regional features are concatenated with the unweighted original clinical scalar variables and used together as the model input features.

[0031] In another aspect, this application also provides a big data-based orthokeratology lens design-assisted system, comprising:

[0032] The data acquisition module is used to acquire historical lens fitting datasets, which include multiple samples. Each sample includes a preoperative corneal parameter map, a postoperative corneal parameter map, and lens design target values.

[0033] The local response quantity calculation module is used to calculate the local response quantity of each sample at each grid point on the corneal map based on the preoperative and postoperative corneal parameter maps. The local response quantity represents the change of corneal parameters in the neighborhood centered on the grid point.

[0034] The original sensitivity score calculation module is used to calculate the original sensitivity score of each grid point based on the correlation strength between the local response quantity of all samples and the lens design target value.

[0035] The weight map generation module is used to perform spatial smoothing on the original sensitivity score and map it to a preset weight range to generate a local area sensitivity weight map with the same size as the corneal parameter map.

[0036] The weighted processing module is used to perform weighted processing on the corneal features of the current patient based on the local region sensitivity weight map to obtain a weighted feature set;

[0037] The design parameter prediction module is used to input the weighted feature set into the pre-trained height difference regression model to predict the height difference design parameters for the orthokeratology lens used for the current patient.

[0038] In another aspect, this application also provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the big data-based orthokeratology lens assisted design method as described above.

[0039] In another aspect, this application provides a computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor to implement the big data-based orthokeratology lens assisted design method described above.

[0040] Another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the big data-based orthokeratology lens-assisted design method described above.

[0041] This application utilizes Local Sensitivity Weighted Map (LSWM) to mine the differentiated sensitivity of different corneal regions to height difference based on big data, accurately characterizing the impact of local features such as astigmatism axis and palpebral fissure compression, thus solving the problem of non-uniform response. Simultaneously, it adaptively adjusts the weights using an uncertainty map, quantifying weight reliability through statistical confidence metrics to avoid interference from unreliable weights and improve design stability. Furthermore, it employs a tree model regression that integrates LSWM weighted features and clinical scalar variables, coupled with a Huber loss function to combat outliers, balancing local feature focus with global information integration, significantly improving the accuracy of height difference prediction. The overall solution overcomes the limitations of traditional regression models and provides personalized design support for patients with irregular or asymmetrical corneas, combining clinical practicality with technological innovation, significantly reducing the risk of poor fit, and ensuring both orthokeratology effect and wearing safety. Attached Figure Description

[0042] Figure 1 A schematic diagram of a big data-based orthokeratology lens design method provided in an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the local response calculation process provided in an embodiment of the present invention;

[0044] Figure 3 This is a schematic diagram of the original sensitivity score calculation process provided in an embodiment of the present invention;

[0045] Figure 4 This is a schematic diagram of the LSWM generation process provided in an embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram of the uncertainty graph generation process provided in an embodiment of the present invention;

[0047] Figure 6 This invention provides a schematic diagram of the LSWM adaptive weighted processing flow;

[0048] Figure 7 A schematic diagram of a big data-based orthokeratology lens design-assisted system provided in an embodiment of the present invention;

[0049] Figure 8 This is a schematic diagram of the structure of a device provided in an embodiment of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] It should be noted that all user information (including but not limited to user device information, user personal information, object information corresponding to device usage data, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, device usage data, etc.) involved in all embodiments of this disclosure are information and data authorized by the user or fully authorized by all parties.

[0052] like Figure 1 As shown in the diagram, this invention discloses a big data-based method for assisting in the design of orthokeratology lenses, comprising the following steps:

[0053] S1. Obtain historical lens fitting dataset, which includes multiple samples. Each sample includes a preoperative corneal parameter map, a postoperative corneal parameter map, and lens design target values.

[0054] S2, based on preoperative and postoperative corneal parameter maps, calculate the local response of each sample at each grid point on the corneal map. The local response represents the change of corneal parameters in the neighborhood centered on that grid point.

[0055] S3, Calculate the original sensitivity score for each grid point based on the correlation strength between the local response of all samples and the lens design target value;

[0056] S4, Spatial smoothing is performed on the original sensitivity score, and it is mapped to a preset weight range to generate a local area sensitivity weight map with the same size as the corneal parameter map.

[0057] S5, The corneal features of the current patient are weighted based on the local region sensitivity weight map to obtain a weighted feature set;

[0058] S6, input the weighted feature set into the pre-trained height difference regression model to predict the height difference design parameters for the orthokeratology lens used for the current patient.

[0059] In some embodiments, for step S1, according to the present invention, a historical lens fitting dataset is first obtained. Specifically, the historical lens fitting dataset needs to contain multiple valid samples, each sample fully covering three types of core data: preoperative corneal parameter map, postoperative corneal parameter map, and lens design target value.

[0060] For example, both the preoperative and postoperative corneal parameter maps are gridded data. The coordinate system is set to a two-dimensional rectangular coordinate system centered on the corneal optical axis or pupillary center. The coverage area is preferably an area with a central corneal diameter of 8mm (i.e., the x-axis and y-axis values ​​are both -4mm to 4mm). The grid spacing Δ is preferably 0.25mm, corresponding to a grid size of approximately 33×33. This grid resolution ensures the accuracy of parameter representation while avoiding efficiency issues caused by excessive computation. Optionally, the grid spacing can also be adjusted to any value between 0.2mm and 0.5mm according to the actual application scenario, as long as the grid parameters of all samples remain consistent.

[0061] Regarding the specific types of corneal parameters, the preoperative and postoperative corneal parameter maps may include key parameters related to orthokeratology lens design, such as corneal height, corneal curvature, and corneal thickness. Corneal height is preferred because it directly reflects changes in corneal surface morphology and has the strongest correlation with lens height difference design. The target value for lens design is the height difference parameter ultimately determined from successful fitting cases, and the unit can be μm or mm, with a standardized approach based on clinical data recording practices, such as uniformly converting to μm to improve calculation accuracy.

[0062] Optionally, during data collection, samples are initially screened to remove invalid samples with missing or abnormal data. Specifically, for a single sample, if the number of missing grid points in its corneal parameter map exceeds 25% of the total number of grid points, it is determined to be an invalid sample and is removed. For samples with a missing grid point ratio not exceeding 25%, nearest neighbor interpolation or bilinear interpolation is used to fill in the missing grid points to ensure data integrity. Simultaneously, the corneal parameter maps of all samples are rotated and corrected to ensure that their optical axis / pupil center is aligned with the coordinate system center, avoiding subsequent calculation errors caused by coordinate system offset.

[0063] In some embodiments, for step S2, after acquiring and preprocessing the historical eyeglasses dataset, the calculation of the local response quantity is performed. This step is the basis for the subsequent sensitivity score calculation, and the accuracy of its calculation result affects the reliability of the Local Sensitivity Weight Map (LSWM).

[0064] According to an embodiment of the present invention, please refer to Figure 2 , Figure 2 This is a schematic diagram of the local response calculation process provided in an embodiment of the present invention. The calculation of the local response of each sample at each grid point on the corneal map specifically includes the following sub-steps:

[0065] In S201, the neighborhood is defined, specifically, for each grid point j on the corneal map (whose coordinates are...). A physical neighborhood is defined centered on this grid point. Specifically, the neighborhood is a circular region with a radius of [missing information]. The default setting is 1.0 mm; optionally, the neighborhood radius is adjusted according to the actual range of local corneal effects. It can also be adjusted to any value between 0.5mm and 1.5mm. For example, for areas where the corneal shape changes relatively gently, the neighborhood radius can be increased appropriately, while for areas with drastic changes, the neighborhood radius can be decreased to more accurately capture changes in local parameters.

[0066] In S202, a Gaussian kernel function is used for weighting. Specifically, a Gaussian kernel function is applied to assign weights to all grid points within the neighborhood to reflect the degree of influence of different grid points on the local response of the central grid point. Specifically, the expression for the Gaussian kernel function is: Where d is any grid point k in the neighborhood (with coordinates of k, k, k) The Euclidean distance between ) and the central grid point j, i.e. ; The standard deviation of the Gaussian kernel is preferably set to [value]. This setting ensures that grid points closer to the central grid point have higher weights and grid points farther away have lower weights, which is consistent with the characteristics of the physical continuity of the cornea, that is, the local changes of the central grid point are mainly affected by the surrounding nearby grid points.

[0067] For example, when the neighborhood radius hour, For a grid point 0.25 mm from the center grid point, its weight is... For a grid point 1.0 mm from the center grid point, its weight is... This weighting effectively highlights the contribution of parameter changes in the nearby region surrounding the central grid point to the local response.

[0068] In S203, the difference aggregation amount is calculated. Specifically, the difference aggregation amount between the preoperative and postoperative corneal parameter maps within the weighted neighborhood is calculated as the local response amount of that grid point. Specifically, the preoperative corneal parameter values ​​for each sample i at each grid point k within the neighborhood are first calculated. Postoperative corneal parameter values The difference, i.e. Then, the difference is multiplied by the Gaussian kernel weight K(d) of the corresponding grid point to obtain the weighted difference; finally, the weighted differences of all grid points in the neighborhood are aggregated to obtain the local response of the central grid point j. .

[0069] Preferably, the aggregation calculation uses a weighted average method, the expression of which is:

[0070] in, This represents the set of all grid points within the neighborhood of grid point j. Using a weighted average aggregation method can integrate neighborhood parameter variations while preserving the weight differences between different grid points, thus more accurately characterizing the corneal parameter variation features of a local region. Optionally, in some embodiments, a simple average aggregation method can also be used. ,in This method is simpler to calculate than the weighted average method, as it represents the number of grid points in the neighborhood. However, its accuracy is slightly lower than that of the weighted average method.

[0071] It should be noted that before calculating the local response, the preoperative and postoperative corneal parameter maps for each sample need to be standardized in terms of dimensions. For example, all parameter values ​​should be converted to the same unit (such as μm) to avoid calculation errors caused by inconsistent dimensions.

[0072] In some embodiments, for step S3, after obtaining the local response values ​​of all samples at each grid point, the original sensitivity score of each grid point is calculated based on the correlation strength between the local response values ​​and the lens design target value. The core of this method is to quantify the influence of each region of the cornea on the lens height difference design through statistical analysis.

[0073] Please see Figure 3 , Figure 3 This is a schematic diagram of the original sensitivity score calculation process provided by an embodiment of the present invention. According to an embodiment of the present invention, calculating the original sensitivity score for each grid point specifically includes the following sub-steps:

[0074] In S301, the statistical correlation coefficient is calculated. Specifically, the statistical correlation coefficient between the local response of all samples at each grid point j and the target value of the lens design is calculated. Preferably, the statistical correlation coefficient is the Spearman correlation coefficient. It can effectively measure the monotonic correlation between two variables and has strong robustness to outliers, making it suitable for clinical data such as corneal parameters that may have a small number of outliers.

[0075] Specifically, for grid point j, collect the local response values ​​of all N samples. and the corresponding lens design target value ,in Design a target value for the lens of sample i. Spearman correlation coefficient. The calculation process is as follows: First, for Sort the results to get each rank ;right Sort the results to get each rank Then calculate according to the following formula. : Spearman correlation coefficient The value range of is [-1, 1], where The closer the absolute value is to 1, the stronger the monotonic correlation between the local response and the lens design target value. The closer the value is to 0, the weaker the monotonic correlation between the two factors. Since the sensitivity score only considers the magnitude of the association, not its direction, the absolute value of the Spearman correlation coefficient is used in subsequent calculations. .

[0076] Optionally, in some embodiments, the Pearson correlation coefficient can also be used as the statistical correlation coefficient, but it has high requirements for the normality of the data and is more sensitive to outliers, so it is only suitable for scenarios where the data distribution is relatively ideal.

[0077] In S302, robust regression coefficients are calculated. Specifically, at each grid point j, the robust regression coefficients of the local response of all samples on the lens design target value are calculated. Preferably, the robust regression coefficients are Huber regression coefficients or Theil-Sen regression coefficients. Both of these regression methods have strong resistance to outliers, can avoid the interference of outlier samples on the regression results, and ensure that the regression coefficients can truly reflect the linear correlation strength between the local response and the lens design target value.

[0078] Taking Huber regression as an example, its objective function is: ,in, For the intercept term, The Huber regression coefficients (i.e., local response quantities) for grid point j For target value (regression slope) The Huber loss function is expressed as follows: ,in, The threshold parameter for the Huber loss function is preferably set to 1.345. This value allows Huber regression to approach the efficiency of ordinary least squares regression on normally distributed data while maintaining robustness to outliers. The robust regression coefficients can be obtained by minimizing the above objective function. , The larger the absolute value, the greater the change in the lens design target value for each unit change in the local response, meaning that the region where the grid point is located has a more significant impact on the height difference design.

[0079] Alternatively, in some embodiments, Theil–Sen regression can be used to calculate robust regression coefficients. It calculates the median slope between all sample point pairs as regression coefficients, which is simpler to calculate and more robust, but its computational efficiency is slightly lower than Huber regression.

[0080] In S303, the absolute value of the statistical correlation coefficient and the normalized value of the robust regression coefficient are weighted and fused to obtain the original sensitivity score. By integrating two different statistical indicators, this study considers both the monotonic correlation between the local response and the target value (through the Spearman correlation coefficient) and the linear contribution strength between the two (through the robust regression coefficient), thus quantifying the sensitivity of grid points more comprehensively and accurately.

[0081] Specifically, the calculation process for weighted fusion is as follows:

[0082] Robust regression coefficient normalization: First, calculate the absolute value of the robust regression coefficients for all grid points. Where M is the total number of grid points in the corneal map; then calculate the median(|β|) and interquartile range (IQR)(|β|) of the absolute value sequence, where IQR(|β|) = Q3(|β|) - Q1(|β|), where Q1 and Q3 are the first and third quartiles of the absolute value sequence, respectively; finally, calculate the absolute value of the robust regression coefficient for each grid point j according to the following formula. Normalization is performed:

[0083] This normalization process eliminates the dimensional differences in regression coefficients between different grid points, resulting in normalized coefficients. They are comparable.

[0084] Weighted fusion calculation: A weighted summation method is used to calculate the weighted summation of ... and The fusion is performed using the following formula:

[0085] in, and These are weighting coefficients, and The preferred setting is , This setting allows the weight of monotonic correlation to be slightly higher than that of linear contribution, which better reflects the influence of local corneal regions on the height difference design; optionally, based on the characteristics of the actual data, and It can also be adjusted to other ratios, for example , ,or , You only need to ensure that the sum of the two is 1.

[0086] Score normalization: The merged score Normalization to the [0,1] interval is achieved using linear scaling, specifically: ,in, The minimum score after merging all grid points. This is the maximum value. This normalization process ensures that the original sensitivity scores of all grid points are within the same numerical range, facilitating subsequent spatial smoothing and weight mapping.

[0087] In some embodiments, for step S4, after obtaining the original sensitivity score for each grid point, the original sensitivity score is spatially smoothed and mapped to a preset weight range to generate an LSWM with the same size as the corneal parameter map. This step transforms the statistical analysis results into a practical tool that can be used for feature weighting.

[0088] Please see Figure 4 , Figure 4 This is a schematic diagram of the LSWM generation process provided in an embodiment of the present invention. According to an embodiment of the present invention, the generation of LSWM specifically includes the following sub-steps:

[0089] In step S401, the normalized original sensitivity score is... Spatial smoothing is performed to eliminate abrupt changes in scores in space, ensuring that the generated LSWM has smooth continuity and conforms to the physical laws of corneal morphological changes. Preferably, the spatial smoothing process employs a two-dimensional Gaussian smoothing method, the core of which lies in using a Gaussian kernel function to perform a weighted average of the scores of each grid point, so that the score of each grid point is influenced by the scores of its surrounding grid points, thereby achieving spatial smoothing.

[0090] Specifically, the expression for two-dimensional Gaussian smoothing is: ,in, The score of grid point j after smoothing; G(·,·) is the smooth neighborhood of grid point j, preferably with the same radius as the neighborhood in the local response calculation (i.e., radius 1.0 mm), but can be adjusted to 0.5 mm to 1.5 mm according to actual needs; G(·,·) is the two-dimensional Gaussian kernel function, and its expression is: , where Δx = Δy = , For smoothing parameters, a setting of 0.75 mm is preferred, as this setting effectively eliminates local abrupt changes while maintaining the score space resolution; alternatively, It can also be adjusted to any value between 0.5mm and 1.0mm. The larger the value, the stronger the smoothing effect, but the lower the spatial resolution. The smaller the value, the weaker the smoothing effect and the higher the spatial resolution.

[0091] In step S402, the score after spatial smoothing is... Mapping to a preset weight range yields the sensitivity weight for each grid point. This leads to the generation of a complete LSWM. Specifically, the preset weight range is... The preferred setting is , The purpose of this range setting is to avoid the feature of a region being ignored due to a low weight (close to 0) or the feature of a region being over-amplified due to a high weight (much greater than 1). At the same time, it maps the median weight to 1, ensuring that the weight of most grid points is close to the baseline level, and only the weight of special regions (such as the astigmatic axis direction) deviates from the baseline.

[0092] The weight mapping uses a combination of linear scaling and clipping, and the specific steps are as follows:

[0093] Linear scaling: First, according to the following formula... Linear mapping to Interval: Where a and b are linear mapping coefficients, determined by: achieving a smooth score for all grid points. the median of Mapped to 1, i.e. At the same time minimum value Mapped to , maximum value Mapped to ,Right now , By solving the above system of equations, we can obtain the specific values ​​of a and b: ,

[0094] Pruning process: For the weights after linear mapping Cut to remove excess material. The weight values ​​for the range are adjusted to the corresponding boundary values, that is:

[0095]

[0096] By pruning, we ensure that the weights of all grid points are within a preset range, thus avoiding the adverse effects of extreme weights on subsequent feature weighting.

[0097] After completing the weight calculation for all grid points, then all Arranged according to their corresponding grid coordinates, an LSWM with the same size as the corneal parameter map is generated. The value of each grid point in this weight map represents the sensitivity weight of the region to the lens height difference design, which can be directly used for subsequent feature weighting processing.

[0098] In some embodiments, for step S5, according to a preferred embodiment of the present invention, after generating LSWM, an uncertainty map U(x,y) corresponding to LSWM can be generated based on the confidence level of the statistics used in calculating the original sensitivity score. This step is a supplement and optimization of LSWM, which can provide a reliability reference for subsequent feature weighting and further improve the accuracy of model prediction.

[0099] Please see Figure 5 , Figure 5 This is a schematic diagram of the uncertainty diagram generation process provided in an embodiment of the present invention. Specifically, the uncertainty diagram generation process is as follows:

[0100] In S501, for each grid point j, the statistic used in calculating the original sensitivity score (i.e., the Spearman correlation coefficient) is used. and robust regression coefficient The confidence level of the grid point is used to determine the uncertainty value of that grid point. Preferably, the confidence level is calculated using one of the following two methods:

[0101] Based on the p-value of the correlation coefficient: For Spearman correlation coefficient The corresponding p-value is calculated through hypothesis testing. The smaller the p-value, the stronger the... The higher the statistical significance, the more reliable the correlation between the local response and the lens design target value; the larger the p-value, the more reliable the correlation. The lower the statistical significance, the less reliable the association.

[0102] Confidence interval width based on regression coefficients: for robust regression coefficients Calculate its 95% confidence interval. ,in This represents the lower limit of the confidence interval. The upper limit is given, and the width of the confidence interval is given. The narrower the confidence interval, the stronger the confidence level. The higher the estimation accuracy, the stronger the reliability; the wider the width, the more reliable the estimation. The lower the accuracy of the estimate, the weaker its reliability.

[0103] In S502, the confidence index is mapped to the [0,1] interval to obtain the uncertainty value for each grid point. ,in The closer the value is to 1, the lower the reliability of the sensitivity weight estimate for that grid point. The closer it is to 0, the higher the reliability of the estimate.

[0104] The specific mapping method is as follows:

[0105] If we use the p-value of the correlation coefficient: firstly... Mapping to the [0,1] interval, since the range of p values ​​is [0,1], it can be directly used. This serves as the initial uncertainty value; then it is processed in reverse, i.e. This is because the larger the p-value, the lower the reliability and the higher the uncertainty.

[0106] If using the confidence interval width of the regression coefficients: first calculate the confidence interval width for all grid points. Then, linear scaling is used to... Mapped to the interval [0,1], i.e.: ,in, The minimum value for the confidence interval width for all grid points. This is the maximum value.

[0107] Alternatively, the two confidence indices can be weighted and fused to obtain the final uncertainty value. ,For example:

[0108]

[0109] Wherein, γ is the weighting coefficient, preferably set to 0.5. This method can integrate the information of the two confidence indices, making the uncertainty value more valuable for reference.

[0110] After calculating the uncertainty values ​​for all grid points, then all Arrange the corresponding grid coordinates to generate an uncertainty map U(x,y) with the same size as the LSWM. This map can be stored together with the LSWM for subsequent feature weighting adaptive adjustment.

[0111] After generating the LSWM and an optional uncertainty map, the LSWM is used to weight the corneal features of the current patient to obtain a weighted feature set. The core of this step is to adjust the importance of features in each region of the cornea through sensitivity weights, so that the regression model pays more attention to the feature information of high-sensitivity regions.

[0112] According to embodiments of the present invention, the corneal feature weighting processing specifically includes the following two implementation methods:

[0113] First, the basic weighted processing method (without considering the uncertainty map) directly uses LSWM to weight the corneal features of the current patient. This method is suitable for scenarios where the reliability of the weight estimation does not need to be considered. The specific steps are as follows:

[0114] Corneal feature extraction: Obtain a corneal parameter map of the current patient, ensuring that the coordinate system and grid resolution of this map are consistent with those of historical corneal parameter maps (i.e., central diameter 8mm, grid spacing 0.25mm, etc.); then extract corneal features from this map to form the original feature vector F = ,in M represents the corneal feature value (such as corneal height) of the current patient at grid point j, and M is the number of grid points.

[0115] Feature weighting calculation: The original feature vector F is compared with the LSWM weight vector W = Perform a dot product operation to obtain the weighted eigenvector. ,Right now: ,

[0116] Where ⊙ represents the dot product operation. This operation is used to determine the high-sensitivity region (…). Larger eigenvalues ​​are amplified, while low-sensitivity regions ( Smaller eigenvalues ​​are suppressed, allowing subsequent regression models to focus more on regional features that significantly affect the height difference design.

[0117] Preferably, secondly, an adaptive weighting processing method (considering the uncertainty diagram), specifically, please refer to [link to relevant documentation]. Figure 6 , Figure 6 This invention provides a schematic diagram of the LSWM adaptive weighting process. This method combines the uncertainty diagram to adaptively adjust the weighting effect of LSWM, which can effectively reduce the impact of unreliable weights on feature weighting. It is a preferred embodiment of this invention. The specific steps are as follows:

[0118] In S601, the uncertainty map is read, and the uncertainty map U(x,y) corresponding to LSWM is read to obtain the uncertainty value of each grid point j. .

[0119] In S602, the weights are adaptively adjusted according to... Size of sensitivity weight Adaptive adjustments are made to reduce or ignore the weighted impact of sensitivity weights on grid points with uncertainty values ​​higher than a preset threshold.

[0120] The specific adjustment methods are as follows:

[0121] First, set the uncertainty threshold. The preferred setting is 0.7, that is, when At that time, it is determined that the reliability of the sensitivity weight estimate for that grid point is low, and it needs to be adjusted; optionally, It can also be adjusted to any value between 0.6 and 0.8 depending on the actual application scenario.

[0122] Then, according to the following formula... Adjustments will be made:

[0123] in, The adjusted sensitivity weights; The baseline weight is preferably set to 1.0. This means that for grid points with high uncertainty, their weight is adjusted to the baseline value, effectively ignoring the sensitivity weighting effect of that grid point and using the default feature importance. To adjust the coefficient, The smaller the value, the closer the adjustment coefficient is to 1. The closer to the original This ensures that the weighting effect of reliable weights is not affected.

[0124] Alternatively, the following adjustment methods may also be adopted: Where k is the attenuation coefficient, preferably set to 2.0. This method allows for a more significant weight attenuation at grid points with higher uncertainty values, resulting in a smoother adjustment process.

[0125] In S603, the final feature set is generated, and the adjusted weight vector is... Perform a dot product operation with the original feature vector F of the current patient to obtain the adaptively weighted feature vector. = F ⊙ .

[0126] Furthermore, according to an embodiment of the present invention, when generating the final feature set, the weighted regional features are also concatenated with the unweighted original clinical scalar variables.

[0127] Specifically, the original clinical scalar variables include clinical parameters related to lens design, such as the current patient's astigmatic axis, baseline refractive power, corneal thickness, and age. These variables are global information and not regional, therefore they do not require LSWM weighting. The weighted regional feature vectors... With the clinical scalar variable vector C = (K represents the number of clinical scalar variables) are concatenated to obtain the final model input feature set. This feature set contains both region-sensitive weighted local features and global clinical information, providing more comprehensive input data for the regression model.

[0128] In some embodiments, for step S6, the weighted feature set is input into a pre-trained height difference regression model to predict the current patient's corneal reshaping lens height difference design parameters. The core of this method is to predict the height difference parameters through a pre-trained regression model.

[0129] According to an embodiment of the present invention, the pre-trained height difference regression model is a tree model, preferably the XGBoost model, which has strong nonlinear fitting ability, is friendly to feature engineering, has high training efficiency and a certain robustness to outliers, and is suitable for complex clinical data modeling scenarios such as corneal reshaping lens height difference prediction; alternatively, other tree models such as LightGBM and random forest can also be used, and their training process is similar to XGBoost.

[0130] Model training specifically includes the following sub-steps:

[0131] The training data consists of a historical feature set weighted by LSWM and its corresponding lens design target values. Specifically, for each historical sample i, its corneal feature vector is first extracted. The weighted region feature vector is obtained by weighting it using the weight vector W of LSWM. ⊙ W; then With the clinical scalar variable vector of this sample By concatenating the vectors, we obtain the complete training feature vector of sample i. Finally, construct the training dataset D = ,in The target value is designed for the lens of sample i.

[0132] Model optimization is performed using a loss function robust to outliers, with the Huber loss function being the preferred choice. This function can effectively suppress the interference of outliers on model training while ensuring model fitting accuracy, thus preventing the model from overfitting to outliers.

[0133] The expression for the Huber loss function is: ,

[0134] In the diagram, y represents the actual target value for lens design ( ), The predicted value of the model. The threshold parameter is preferably set to 1.345.

[0135] Optionally, in some embodiments, the MAE (mean absolute error) loss function or the MSE (mean squared error) loss function may be used as auxiliary monitoring indicators to more comprehensively evaluate model performance, but the core loss function for model optimization is still preferably the Huber loss function.

[0136] To ensure the model's fitting and generalization abilities, the hyperparameters of the tree model need to be set appropriately. The following are the preferred ranges for the initial values ​​of the hyperparameters of the XGBoost model:

[0137] Learning rate ( The value is 0.03-0.1, preferably 0.05. This parameter controls the weight update step size of each tree. Too small a value will result in slow training speed, while too large a value may lead to model overfitting.

[0138] Maximum depth of the tree ( ): 4-7, preferably set to 5. This parameter controls the complexity of each tree. Too deep a tree will lead to overfitting, while too shallow a tree will lead to underfitting.

[0139] Number of estimators ( ): 200-1000, this parameter controls the number of trees, and the final number needs to be determined in conjunction with the early stopping strategy;

[0140] Subsample rate: 0.7-0.9. This parameter controls the sample sampling ratio during training of each tree and is used to prevent overfitting.

[0141] Feature sampling rate ( The value is 0.6-0.9, which controls the feature sampling ratio during training of each tree to prevent overfitting.

[0142] Regularization parameters ( ): 1e-3-1.0, where For L1 regularization parameters, This is the L2 regularization parameter, used to suppress model overfitting.

[0143] The model training uses K-fold cross-validation (K=5) to ensure the model's generalization ability. The specific process is as follows:

[0144] 1. Randomly divide the training dataset D into K non-overlapping subsets. The number of samples in each subset is approximately equal;

[0145] 2. For each fold number k (k=1,2,…,K), with The union of the remaining K-1 subsets serves as the validation set and as the training set.

[0146] 3. Train the XGBoost model based on the current training set, employing an early stopping strategy during training. This means that after each training epoch (adding a new tree), calculate the model's loss function (e.g., Huber loss) on the validation set. If the validation set loss function value does not decrease for 50 consecutive epochs (patience=50), stop training and record the current number of trees as the optimal value. ;

[0147] Repeat steps 2-3 to obtain K trained models;

[0148] Integrate the prediction results of K models (e.g., take the average) to obtain the final pre-trained model, or select the model with the smallest loss on the validation set as the final pre-trained model.

[0149] In addition, feature selection is performed during training. By calculating feature importance (such as the gain value of XGBoost), features that contribute significantly to the prediction results are selected, while redundant features that contribute less are removed. This simplifies the model structure and improves training efficiency and generalization ability. Simultaneously, the performance comparison between the weighted model (using LSWM weighted features) and the baseline model (not using LSWM weighted features) is recorded to verify the effectiveness of LSWM.

[0150] After completing model training, the pre-trained height difference regression model is used to predict the design parameters of the corneal reshaping lens height difference for the current patient. The specific steps are as follows:

[0151] Current patient data acquisition and preprocessing: Acquire the preoperative corneal parameter map and clinical scalar variables (astigmatic axis, baseline refractive power, etc.) of the current patient, and perform the same preprocessing on the corneal parameter map as on historical samples (such as rotation correction, missing value filling, etc.) to ensure that the data format is consistent with the training data;

[0152] Feature extraction and weighting: The original feature vector F is extracted from the current patient's corneal parameter map. LSWM (and optional uncertainty map) is used to weight F, resulting in a weighted region feature vector. ;Will Concatenate the input feature set with the clinical scalar variable vector C to obtain the final input feature set. ;

[0153] Model prediction: The input is fed into a pre-trained height difference regression model, which outputs the design parameters for the current patient's orthokeratology lens height difference. ;

[0154] Output: The predicted height difference design parameters The output serves as the basis for the current design of orthokeratology lenses for patients. Doctors can further fine-tune the design based on this parameter and their clinical experience to ultimately determine a personalized orthokeratology lens design plan.

[0155] It is important to note that during the prediction process, it is crucial to ensure that the coordinate system, grid resolution, and other parameters of the current patient's corneal parameter map are completely consistent with historical samples. Otherwise, feature dimension mismatch will occur, affecting the accuracy of the prediction results. Furthermore, if the current patient's cornea exhibits specific local irregularities (such as severe astigmatic axis shift), LSWM will automatically adjust the weights to highlight the features of the relevant areas, making the prediction results more closely reflect the actual condition of the patient's cornea.

[0156] In summary, this invention achieves accurate prediction of orthokeratology lens height difference design parameters based on big data through a series of steps, including acquiring historical lens fitting data, calculating local response quantities, quantifying the original sensitivity score, generating LSWM and uncertainty maps, weighting corneal features, and using pre-trained regression models for prediction. It has significant advantages, especially for scenarios with irregular or asymmetrical corneas, and can provide more scientific and personalized orthokeratology lens design assistance for clinicians.

[0157] Figure 7 This paper illustrates a big data-based orthokeratology lens design-assisted system. The system implementation is similar to... Figure 1 Corresponding to the illustrated method embodiments, this system can be specifically applied to various electronic devices. The system specifically includes:

[0158] The data acquisition module 701 is used to acquire historical lens fitting datasets, which include multiple samples, each of which includes a preoperative corneal parameter map, a postoperative corneal parameter map, and lens design target values.

[0159] The local response calculation module 702 is used to calculate the local response of each sample at each grid point on the corneal map based on the preoperative and postoperative corneal parameter maps. The local response represents the change of corneal parameters in the neighborhood centered on the grid point.

[0160] The original sensitivity score calculation module 703 is used to calculate the original sensitivity score of each grid point based on the correlation strength between the local response quantity of all samples and the lens design target value.

[0161] The weight map generation module 704 is used to perform spatial smoothing on the original sensitivity score and map it to a preset weight range to generate a local area sensitivity weight map with the same size as the corneal parameter map.

[0162] The weighted processing module 705 is used to perform weighted processing on the corneal features of the current patient based on the local region sensitivity weight map to obtain a weighted feature set;

[0163] The design parameter prediction module 706 is used to input the weighted feature set into the pre-trained height difference regression model to predict the height difference design parameters for the corneal reshaping lens of the current patient.

[0164] Based on the same inventive concept, this application also provides an electronic device. The method corresponding to the electronic device can be the method in the foregoing embodiments, and its problem-solving principle is similar to that method. The electronic device provided in this application includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the methods and / or technical solutions of the foregoing embodiments of this application.

[0165] The electronic device can be a user device, or a device formed by integrating user devices and network devices through a network, or it can be an application running on the aforementioned devices. The user device includes, but is not limited to, various terminal devices such as computers, mobile phones, tablets, smartwatches, and smart bands. The network device includes, but is not limited to, network hosts, single network servers, multiple network server sets, or cloud computing-based computer sets, and can be used to implement some processing functions when setting an alarm clock. Here, the cloud consists of a large number of hosts or network servers based on cloud computing. Cloud computing is a type of distributed computing, consisting of a virtual computer composed of a group of loosely coupled computer sets.

[0166] Figure 8 The diagram illustrates the structure of an apparatus suitable for implementing the methods and / or technical solutions in the embodiments of this application. The apparatus includes a central processing unit (CPU) 801, which can perform various appropriate actions and processes based on a program stored in a read-only memory (ROM) 802 or a program loaded from a storage portion 808 into a random access memory (RAM) 803. The RAM 803 also stores various programs and data required for system operation. The CPU 801, ROM 802, and RAM 803 are interconnected via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0167] The following components are connected to I / O interface 805: an input section 806 including a keyboard, mouse, touchscreen, microphone, infrared sensor, etc.; an output section 807 including a cathode ray tube (CRT), liquid crystal display (LCD), LED display, OLED display, etc., and speakers, etc.; a storage section 808 including one or more computer-readable media such as hard disk, optical disk, magnetic disk, semiconductor memory, etc.; and a communication section 809 including a network interface card such as a LAN (local area network) card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet.

[0168] In particular, the methods and / or embodiments in this application can be implemented as computer software programs. For example, the embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by a central processing unit (CPU) 801, it performs the functions defined in the methods of this application.

[0169] Another embodiment of this application provides a computer-readable storage medium having computer program instructions stored thereon, which can be executed by a processor to implement the methods and / or technical solutions of any one or more embodiments of this application described above.

[0170] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0171] The flowcharts or block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of devices, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-specific system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0172] Furthermore, the inclusion of a single word does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in a system claim may also be implemented by a single unit or device through software or hardware. Terms such as "first," "second," etc., are used to indicate names and do not indicate any particular order.

Claims

1. A method for assisting in the design of orthokeratology lenses based on big data, characterized in that, The methods and steps include the following: Obtain a historical lens fitting dataset, which includes multiple samples, each of which includes a preoperative corneal parameter map, a postoperative corneal parameter map, and a lens design target value; Based on preoperative and postoperative corneal parameter maps, the local response of each sample at each grid point on the corneal map is calculated. The local response represents the change of corneal parameters in the neighborhood centered on that grid point. Based on the correlation strength between the local response of all samples and the lens design target value, the original sensitivity score of each grid point is calculated; The original sensitivity score is spatially smoothed and mapped to a preset weight range to generate a local area sensitivity weight map with the same size as the corneal parameter map. The corneal features of the current patient are weighted based on the local region sensitivity weight map to obtain a weighted feature set; The weighted feature set is input into a pre-trained height difference regression model to predict the height difference design parameters for the orthokeratology lens used for the current patient. The calculation of the local response of each sample at each grid point includes: Centered on the current grid point, define a neighborhood in physical space, wherein the neighborhood is a circular region; A Gaussian kernel function is used to assign weights to the grid points in the neighborhood, with grid points closer to the center grid point having higher weights. The difference aggregation between the preoperative and postoperative corneal parameter maps in the weighted neighborhood is calculated and used as the local response of that grid point. The calculation of the original sensitivity score for each grid point includes: Calculate the statistical correlation coefficient between the local response of all samples at each grid point and the lens design target value; In addition, calculate the robust regression coefficients of the local response of all samples at each grid point on the lens design target value; The absolute value of the statistical correlation coefficient and the normalized value of the robust regression coefficient are weighted and fused to obtain the original sensitivity score.

2. The method for assisting in the design of orthokeratology lenses based on big data according to claim 1, characterized in that, Based on the confidence level of the statistics used in calculating the original sensitivity score, an uncertainty map corresponding to the local region sensitivity weight map is generated, and the uncertainty map characterizes the estimation reliability of the sensitivity weight of each grid point.

3. The method for assisted design of orthokeratology lenses based on big data according to claim 2, characterized in that, The corneal features of the current patient are weighted using the local region sensitivity weight map, including: Read the uncertainty map corresponding to the local region sensitivity weight map; Based on the uncertainty value of each grid point in the uncertainty graph, the weighting effect of the local region sensitivity weight graph is adaptively adjusted. For grid points with uncertainty values ​​higher than a preset threshold, the weighting effect of the sensitivity weight is weakened or ignored. The adaptively adjusted weights are combined with the corneal features of the current patient to generate the final feature set used as input to the regression model.

4. A big data-based method for assisting in the design of orthokeratology lenses according to any one of claims 1-3, characterized in that, The pre-trained height difference regression model is a tree model, which is trained using the following method: The historical feature set weighted by the local region sensitivity weight map and its corresponding lens design target value are used as training data; Model optimization is performed using a loss function robust to outliers; During training, the weighted regional features are concatenated with the unweighted original clinical scalar variables and used together as the model input features.

5. A corneal reshaping lens assisted design system based on big data, characterized in that, include: The data acquisition module is used to acquire historical lens fitting datasets, which include multiple samples. Each sample includes a preoperative corneal parameter map, a postoperative corneal parameter map, and lens design target values. The local response quantity calculation module is used to calculate the local response quantity of each sample at each grid point on the corneal map based on the preoperative and postoperative corneal parameter maps. The local response quantity represents the change of corneal parameters in the neighborhood centered on the grid point. The original sensitivity score calculation module is used to calculate the original sensitivity score of each grid point based on the correlation strength between the local response quantity of all samples and the lens design target value. The weight map generation module is used to perform spatial smoothing on the original sensitivity score and map it to a preset weight range to generate a local area sensitivity weight map with the same size as the corneal parameter map. The weighted processing module is used to perform weighted processing on the corneal features of the current patient based on the local region sensitivity weight map to obtain a weighted feature set; The design parameter prediction module is used to input the weighted feature set into the pre-trained height difference regression model to predict the height difference design parameters for the corneal reshaping lens of the current patient. The calculation of the local response of each sample at each grid point includes: Centered on the current grid point, define a neighborhood in physical space, wherein the neighborhood is a circular region; A Gaussian kernel function is used to assign weights to the grid points in the neighborhood, with grid points closer to the center grid point having higher weights. The difference aggregation between the preoperative and postoperative corneal parameter maps in the weighted neighborhood is calculated and used as the local response of that grid point. The calculation of the original sensitivity score for each grid point includes: Calculate the statistical correlation coefficient between the local response of all samples at each grid point and the lens design target value; In addition, calculate the robust regression coefficients of the local response of all samples at each grid point on the lens design target value; The absolute value of the statistical correlation coefficient and the normalized value of the robust regression coefficient are weighted and fused to obtain the original sensitivity score.

6. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-4.

7. A computer-readable medium having computer program instructions stored thereon, characterized in that, The computer program instructions can be executed by a processor to implement the method as described in any one of claims 1-4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-4.

Images