Variable-focal-depth femtosecond laser cornea crosslinking parameter optimization method and system based on artificial intelligence and cornea crosslinking device

Abstract

The invention discloses a variable-focal-depth femtosecond laser cornea cross-linking parameter optimization method and system based on artificial intelligence and a cornea cross-linking device, and relates to the field of artificial intelligence and ophthalmology treatment. The method comprises the following steps: reconstructing a personalized three-dimensional cornea model according to a multi-modal cornea image of a keratoconus patient, and obtaining a cross-linking path parameter by using a deep neural network; according to the personalized three-dimensional cornea model, stress-strain simulation analysis is carried out on the cornea subjected to femtosecond laser crosslinking according to the crosslinking path parameters, and biomechanical and refractive indexes of the cornea after laser treatment are obtained; evaluating a cross-linking effect by using the trained regression model; if the crosslinking effect does not reach the preset standard, adaptively updating the deep neural network, and regenerating the crosslinking path parameters; and if the cross-linking effect reaches a preset standard, determining the cross-linking path parameter as an optimal cross-linking scheme. According to the method, closed-loop self-optimization of femtosecond laser cornea cross-linking parameters can be realized, and the accuracy of femtosecond laser directional cornea cross-linking is improved.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and ophthalmic treatment, and in particular to an artificial intelligence-based method, system and device for optimizing corneal crosslinking parameters of a femtosecond laser with variable depth of focus, which is suitable for delaying the progression of keratoconus and assisting in the correction of refractive status. Background Technology

[0002] Keratoconus is a progressive corneal ectasia that commonly affects adolescents. As the disease progresses, the cornea gradually thins and protrudes forward, leading to severe irregular astigmatism and decreased vision. Without intervention, it can result in corneal perforation and scarring, severely impacting visual function.

[0003] Corneal cross-linking (CXL) is an effective treatment for corneal ectasias such as keratoconus. Currently, most clinically used corneal cross-linking methods are based on a photochemical reaction induced by ultraviolet A (UVA) light to produce riboflavin infiltrating the cornea. While this can improve corneal biomechanical strength, it has limitations such as restricted UVA energy depth control and inability to achieve spatially targeted enhancement of corneal lesions. Femtosecond-Corneal Cross-Linking (FS-CXL) can precisely control the energy focusing range within the three-dimensional space of the cornea and achieves spatially targeted corneal cross-linking by stimulating riboflavin infiltrating the cornea through the multiphoton absorption characteristics of the femtosecond laser, thus overcoming the above limitations and representing a potentially ideal treatment. However, current FS-CXL still lacks clinical tools to automatically optimize cross-linking parameters based on the individual patient's corneal structure and mechanical properties. While Corvis ST can quantitatively assess overall corneal mechanics, it does not link this to laser energy distribution. Summary of the Invention

[0004] The purpose of this application is to provide an artificial intelligence-based method, system, and device for optimizing the cross-linking parameters of femtosecond laser cornea with variable depth of focus, which can realize closed-loop self-optimization of femtosecond laser cornea cross-linking parameters and improve the accuracy of femtosecond laser directional cornea cross-linking.

[0005] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides an artificial intelligence-based method for optimizing corneal crosslinking parameters of a variable depth-of-focus femtosecond laser, including: Acquire multimodal corneal images from patients with keratoconus; Based on the multimodal corneal images, a personalized three-dimensional corneal model for keratoconus patients is reconstructed; Based on the multimodal corneal images, crosslinking path parameters are obtained using a deep neural network; the crosslinking path parameters include laser scanning path, depth of focus control parameters, and pulse energy. Based on the personalized three-dimensional corneal model, stress-strain simulation analysis was performed on the cornea cross-linked by femtosecond laser according to the cross-linking path parameters to obtain the biomechanical and refractive indices of the cornea after laser treatment. Based on the biomechanical and refractive indices of the cornea after laser treatment, and the cross-linking path parameters, a trained regression model is used to evaluate the cross-linking effect; If the cross-linking effect does not meet the preset standard, the deep neural network is adaptively updated using the spatial location where the cross-linking effect does not meet the preset standard and the deviation information between the cross-linking effect and the preset standard. The deep neural network is then replaced with the adaptively updated deep neural network, and the process returns to the step "According to the multimodal corneal image, the cross-linking path parameters are obtained using the deep neural network". If the crosslinking effect reaches the preset standard, then the crosslinking path parameters are determined as the optimal crosslinking scheme.

[0006] Secondly, this application provides an artificial intelligence-based variable depth-of-focus femtosecond laser corneal crosslinking parameter optimization system, including: a data acquisition module, an AI path generation module, a simulation feedback module, and a clinical output module; The data acquisition module is used to acquire multimodal corneal images of patients with keratoconus; The AI ​​path generation module is used to reconstruct a personalized three-dimensional corneal model for keratoconus patients based on the multimodal corneal images; and to obtain cross-linking path parameters using a deep neural network based on the multimodal corneal images; the cross-linking path parameters include laser scanning path, depth of focus control parameters, and pulse energy. The simulation feedback module is used to perform stress-strain simulation analysis on the cornea cross-linked by femtosecond laser according to the cross-linking path parameters based on the personalized three-dimensional corneal model, and to obtain the biomechanical and refractive indices of the cornea after laser treatment; based on the biomechanical and refractive indices of the cornea after laser treatment and the cross-linking path parameters, a trained regression model is used to evaluate the cross-linking effect; if the cross-linking effect does not meet the preset standard, the deep neural network is adaptively updated using the spatial location where the cross-linking effect does not meet the preset standard and the deviation information between the cross-linking effect and the preset standard, and the deep neural network is replaced with the adaptively updated deep neural network, and the process returns to execute the process of obtaining the cross-linking path parameters based on the multimodal corneal images using the deep neural network; The clinical output module is used to determine the crosslinking path parameters as the optimal crosslinking scheme if the crosslinking effect meets the preset standard.

[0007] Thirdly, this application provides a corneal crosslinking device that employs the aforementioned artificial intelligence-based variable depth-of-focus femtosecond laser corneal crosslinking parameter optimization system.

[0008] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides an artificial intelligence-based method, system, and device for optimizing corneal crosslinking parameters using a variable depth-of-focus femtosecond laser. It reconstructs a personalized three-dimensional corneal model for keratoconus patients. Based on this personalized three-dimensional corneal model, stress-strain simulation analysis is performed on the cornea to obtain biomechanical and refractive indices, thereby evaluating the crosslinking effect. When the crosslinking effect does not meet the preset standard, the crosslinking path parameters are optimized to achieve closed-loop self-optimization of the crosslinking parameters. The crosslinking path parameters include depth-of-focus control parameters and pulse energy, which are linked to the laser energy distribution and enable precise directional crosslinking at different corneal depths, improving the accuracy of directional crosslinking. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A flowchart illustrating an artificial intelligence-based method for optimizing corneal crosslinking parameters using a femtosecond laser with variable depth of focus, provided for an embodiment of this application; Figure 2 This is a schematic diagram of personalized three-dimensional corneal structure reconstruction provided in an embodiment of this application; Figure 3 A schematic diagram of the path containing x and y spatial coordinates and the z-axis focal depth control point provided for the embodiments of this application; Figure 4 This is a schematic diagram of the dynamic control process for laser focal depth provided in an embodiment of this application; Figure 5 This is a schematic diagram of the femtosecond laser crosslinking principle experiment provided in the embodiments of this application; Figure 6 This is a schematic diagram of the distribution of elastic modulus variation in the finite element simulation prediction results provided in the embodiments of this application; Figure 7 A schematic diagram illustrating the principle of an artificial intelligence-based method for optimizing corneal crosslinking parameters using femtosecond lasers with variable depth of focus, provided in this application embodiment; Figure 8 This is a schematic diagram of the functional modules of an artificial intelligence-based variable depth-of-focus femtosecond laser corneal crosslinking parameter optimization system provided in an embodiment of this application. Detailed Implementation

[0011] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0013] In one exemplary embodiment, such as Figure 1 As shown, an artificial intelligence-based method for optimizing the crosslinking parameters of a femtosecond laser cornea with variable depth of focus is provided. This method is executed by a computer device, specifically by a computer device such as a terminal or a server alone, or by a terminal and a server together. In the embodiments of this application, it includes the following steps 101 to 107.

[0014] Step 101: Acquire multimodal corneal images of patients with keratoconus.

[0015] Step 102: Reconstruct a personalized three-dimensional corneal model for the keratoconus patient based on the multimodal corneal images.

[0016] Step 103: Based on the multimodal corneal images, obtain cross-linking path parameters using a deep neural network; the cross-linking path parameters include laser scanning path, depth of focus control parameters, and pulse energy.

[0017] Step 104: Based on the personalized three-dimensional corneal model, perform stress-strain simulation analysis on the cornea cross-linked by femtosecond laser according to the cross-linking path parameters to obtain the biomechanical and refractive indices of the cornea after laser treatment.

[0018] Step 105: Based on the biomechanical and refractive indices of the cornea after laser treatment, and the cross-linking path parameters, use the trained regression model to evaluate the cross-linking effect.

[0019] Step 106: If the cross-linking effect does not meet the preset standard, the deep neural network is adaptively updated using the spatial location where the cross-linking effect does not meet the preset standard and the deviation information between the cross-linking effect and the preset standard. The deep neural network is then replaced with the adaptively updated deep neural network, and the process returns to step 103.

[0020] Step 107: If the crosslinking effect reaches the preset standard, then the crosslinking path parameters are determined as the optimal crosslinking scheme.

[0021] By implementing steps 101 to 107 above, and combining artificial intelligence with optical simulation, the matching of the three-dimensional stress field inside the cornea with laser depth of focus control was achieved for the first time, providing a new approach for personalized corneal cross-linking treatment. It is suitable for delaying the progression of keratoconus and for assisting in the correction of refractive errors.

[0022] In another exemplary embodiment of this application, the multimodal corneal images include: multimodal images of the anterior corneal surface and multimodal images of the posterior corneal surface. The multimodal images include surface topography maps, OCT images, densitograms, and CorvisST Dynamic Corneal Response (DCR) images. The CorvisST dynamic corneal response image sequence is used to characterize the instantaneous deformation response of the cornea under air pulse loading, providing time-resolved displacement-stress coupling information.

[0023] In another exemplary embodiment of this application, such as Figure 2 As shown, the general process of reconstructing a personalized 3D corneal model for keratoconus patients is as follows: Multimodal data is acquired, including anterior and posterior corneal surface topography, high-resolution OCT images, densitograms, and Corvis ST data, which are then preprocessed to improve image quality. Next, image segmentation techniques are used to extract different layers of the cornea. Deep learning methods (such as the U-Net network) are used to efficiently and accurately segment the corneal tissue layers. Then, image registration is performed, aligning images from different data sources to a unified coordinate system to ensure accurate fusion of images from different sources. Non-rigid registration is used to better handle variations in corneal morphology. Finally, based on the segmented and registered data, interpolation and point cloud reconstruction techniques are used to generate a personalized 3D corneal model, optimizing its geometry to conform to biomechanical properties. This model provides an important foundation for subsequent femtosecond laser corneal crosslinking path optimization and personalized treatment. Therefore, step 102 can be replaced by steps 201-204.

[0024] Step 201: Preprocess the multimodal corneal image to obtain a preprocessed multimodal image; the preprocessing includes normalization, noise reduction, resolution unification and time series synchronization.

[0025] Preprocessing operations such as normalization, denoising, resolution unification, and time series synchronization are performed on various types of data in multimodal corneal images to improve the accuracy of subsequent segmentation, registration, and dynamic feature extraction.

[0026] Step 202: Based on the preprocessed multimodal image, the corneal multilayer structure is segmented using image segmentation technology to obtain segmented multi-source data.

[0027] Applications and Methods of Segmentation Technology: Image segmentation technology is mainly used for high-resolution OCT images and Corvis ST dynamic corneal response images to extract the multi-layered structure of the cornea (epithelium, stroma, endothelium, etc.) and its transient deformation contours under air pulses. The segmentation network adopts deep learning methods (such as the U-Net structure). The input is multimodal preprocessed image data, including OCT images, topographic maps, optical density maps, and DCR dynamic frame sequences. The feature fusion module achieves joint segmentation of static structural information and dynamic response information.

[0028] Step 203: Perform three-dimensional non-rigid registration on the preprocessed multimodal images to obtain registered multi-source data.

[0029] Image registration and fusion: Three-dimensional non-rigid registration was performed on the anterior and posterior corneal surface topographic maps, high-resolution OCT images, densitograms, and DCR images. All data sources were aligned to a unified spatial and temporal coordinate system, ensuring geometric and temporal consistency between images from different sources and at different times, thus achieving spatiotemporal fusion of multi-source information. The dynamic frame sequence of the DCR images was used to provide real-time deformation field information of the cornea during stress loading, serving as an important constraint for subsequent simulation modeling.

[0030] Step 204: Generate a personalized three-dimensional corneal model for keratoconus patients based on the segmented and registered multi-source data.

[0031] 3D Model Reconstruction and Biomechanical Optimization: Based on segmented and registered static and dynamic multi-source data, a personalized 3D corneal model is generated using interpolation algorithms, point cloud fusion, and deformation field reconstruction techniques. By incorporating time-series deformation information from DCR images, the elasticity distribution in different regions of the model is initially estimated and corrected, thereby optimizing the biomechanical consistency of the corneal model.

[0032] The personalized 3D corneal model output by this step not only includes the geometric structure, but also integrates time-series deformation features, providing mechanical boundary conditions for subsequent finite element simulations.

[0033] In another exemplary embodiment of this application, a deep neural network (DNN+Transformer) is used to extract corneal structure, density, and biomechanical features. The input is multimodal data (OCT, topographic maps, optical density maps, DCR dynamic images), and the output is a path P(x,y,z) containing x, y, and z spatial coordinates, along with the corresponding pulse energy, scanning speed, and depth-of-focus control parameter fz. The deep neural network includes: an input layer, a convolutional neural network (CNN), a Transformer model, a feature fusion layer, and an output layer.

[0034] The input layer is used to input multimodal corneal images, such as anterior and posterior corneal surface topographic maps, high-resolution OCT images, densitograms, and Corvis ST dynamic response (DCR) images. These data undergo preprocessing and normalization steps to prepare them for depth feature extraction.

[0035] Convolutional Neural Networks (CNNs) are used to extract local features of the cornea from multimodal corneal images. Specifically, DNN feature extraction involves extracting basic features from corneal image data, such as corneal curvature, surface morphology, and interlayer structure, through CNN layers. Each CNN layer extracts local information through convolution operations and enhances the abstraction of features through pooling and activation functions. Subsequently, the DNN uses fully connected layers to further nonlinearly combine the extracted features to capture the complexity of corneal morphology.

[0036] The Transformer model is used to extract global features of the cornea from multimodal corneal images. Specifically, Transformer feature extraction uses a self-attention mechanism to weight features from different regions to capture the global correlations of the internal structures of the cornea. This mechanism can handle the dependencies between the corneal surface and internal structures, especially subtle changes under the influence of factors such as corneal deformation and optical density variations. The Transformer part uses a multi-head attention mechanism for feature learning, with each head capturing relevant information at different levels, thereby enhancing the model's adaptability to complex structures. The deep neural network and Transformer structure described in this invention both use multimodal corneal data as input, including but not limited to anterior and posterior corneal surface topography maps, high-resolution OCT images, optical density maps, and Corvis ST dynamic corneal response images. After unified preprocessing and feature fusion, these data are input to the DNN and Transformer branches respectively for local and global feature extraction and relationship modeling.

[0037] By combining DNN with Transformer, the model can extract key features from both local and global corneal data. These extracted features are then used in subsequent optimization modules to generate personalized crosslinking pathways, thereby providing precise treatment plans for femtosecond laser corneal crosslinking surgery.

[0038] The feature fusion layer is used to fuse local and global features of the cornea to obtain fused features.

[0039] The output layer is used to output crosslinking path parameters based on the fusion characteristics. The output layer output includes three-dimensional coordinates (x, y, z), depth-of-focus control parameter fz, pulse energy, and scan speed. These outputs will be used to generate personalized femtosecond laser paths, optimizing the depth and distribution of the laser interaction area. The output path containing x and y spatial coordinates and the z-axis depth-of-focus control point is shown below. Figure 3 As shown.

[0040] As can be seen, the deep neural network, an artificial intelligence model, includes a combination of DNN and Transformer structures.

[0041] In another exemplary embodiment of this application, the crosslinking path parameter further includes: scan speed.

[0042] The spatial coordinates of the laser scanning path characterize the spatial positioning points of the laser cross-linking path within the corneal tissue, corresponding to the position (x, y) in the projection plane of the corneal surface and the depth (z) in the thickness direction, respectively, clearly specifying where and in which layer of the cornea the laser should release energy.

[0043] Depth of focus (fz) refers to the focal depth of a femtosecond laser at each spatial positioning point, i.e., the effective depth of action of the laser focus along the axial direction (z-axis). The value of the fz parameter reflects the focal depth range during laser focusing and can be set to different values ​​as needed to adjust the energy distribution thickness at each laser point of action. A smaller focal depth (smaller fz value) results in a more concentrated energy distribution, suitable for precise targeting of specific corneal layers; a larger focal depth (larger fz value) results in a wider energy distribution, capable of covering thicker corneal tissue layers. The fz parameter does not have a simple one-to-one correspondence with the z-coordinate but rather describes the spatial range of laser energy distribution along the z-axis. Artificial intelligence models can automatically generate optimal focal depth control parameters for different three-dimensional points of action based on corneal lesion distribution, biomechanical characteristics, and treatment requirements, thereby enabling multi-layered, variable-width cross-linked treatment schemes.

[0044] Pulse energy is the unit energy of the laser at each spatial positioning point. Scanning speed is the speed at which the laser moves along the path segment.

[0045] Depth-of-focus adjustment mechanism: Based on the depth-of-focus control parameter fz generated by AI, the femtosecond laser depth of focus is controlled through objective lens focusing or liquid lens. The system's focusing response time is less than 10ms, ensuring precise control of the energy application volume and enabling directional focusing and cross-linking of the laser's focal point at different corneal depths. Dynamic control of laser depth of focus is as follows: Figure 4 As shown.

[0046] The principle of femtosecond laser corneal crosslinking is as follows: Femtosecond laser pulses excite riboflavin molecules in the corneal tissue through multiphoton absorption, causing them to transition from the ground state to an excited state. This generates free radicals and singlet oxygen, leading to the formation of new covalent bonds between corneal collagen fibers, thereby significantly enhancing the mechanical strength and stability of the corneal tissue. It is evident that femtosecond laser corneal crosslinking excites riboflavin molecules through a multiphoton absorption mechanism, inducing covalent crosslinking of corneal stromal collagen. Unlike traditional UVA-induced crosslinking, femtosecond lasers achieve localized energy deposition within the tissue through a nonlinear absorption process, forming a crosslinking reaction zone deep within the cornea without surface irradiation. Figure 5 Part 'a' in the diagram represents a schematic of riboflavin fluorescence generated by femtosecond laser (variable depth of focus) excitation. Figure 5 Part b in the text indicates that the spectrometer detected excitation fluorescence at a wavelength of 532 nm produced by femtosecond laser excitation of riboflavin. Figure 5 The "c" part indicates that the principle of femtosecond laser excitation of riboflavin is based on two-photon absorption characteristics.

[0047] In another exemplary embodiment of this application, step 104 above uses the finite element method to construct a Neo-Hookean corneal simulation model to predict stress and elasticity changes. The finite element simulation establishes a Neo-Hookean corneal material model based on the reconstructed model and path parameters, performs stress-strain simulation analysis, and outputs indicators such as elastic modulus change ΔE, refractive power distribution Φ, and corneal deformation displacement U. The finite element simulation process can be replaced by steps 301 to 306.

[0048] Step 301: The nonlinear mechanical behavior of corneal tissue is described using the Neo-Hookean hyperelastic material model.

[0049] The stress-strain relationship is defined as follows: σ=μ(BI)+λln(J)I ; in: σ It is the stress tensor; μ It is the shear modulus; B For the left Cauchy-Green strain tensor, J This represents the volume change ratio; λ It is Lamé's constant; I It is a unit tensor, typically a 3D identity matrix. Material parameters are obtained from experimental data or through Corvis ST DCR inversion.

[0050] Step 302: Set the boundary conditions for the corneal simulation.

[0051] Boundary condition settings: Fix the limbus node to simulate the constraint of the corneal-scleral transition zone; apply the equivalent internal stress distribution caused by laser energy to the central cornea or lesion area; apply air pulses or equivalent pressure boundaries to the outer surface of the model to verify the consistency between the simulation model and the real Corvis ST measurement.

[0052] Step 303: Discretize the personalized three-dimensional corneal model.

[0053] Step 304: Based on the Neo-Hookean hyperelastic material model, boundary conditions, and discretized personalized three-dimensional corneal model, a nonlinear finite element iterative method is used to perform stress-strain simulation analysis on the cornea cross-linked by femtosecond laser according to the cross-linking path parameters, and to obtain the stress distribution and strain distribution of the cornea after laser treatment.

[0054] The reconstructed 3D corneal model is discretized into a finite element mesh, and the stress and strain values ​​of each small element are calculated using the finite element method (FEM). Nonlinear finite element analysis is then used to calculate the stress and strain distribution of the cornea after laser treatment. The solution is iteratively solved by updating the displacement value of each element until convergence.

[0055] In the finite element simulation step, the "element" refers to the basic volume unit obtained by discretizing the reconstructed three-dimensional corneal model. Each element contains its corresponding node coordinates, material parameters, and initial state. During the simulation calculation, a global mechanical equilibrium equation is constructed based on the Neo-Hookean material model, the energy loading distribution defined by the path parameter P(x,y,z), the corneal boundary conditions, and the external loads.

[0056] The solution process employs a nonlinear finite element iterative method. By calculating and updating the displacement value *u* of each element node, internal and external forces are balanced until the residuals converge. The nodal displacement updates are based on the current stress-strain state and the material constitutive relation (…). σ=μ(BI)+λln(J)I The process continues until a steady-state distribution is achieved. Ultimately, the displacement field of each element reflects the true deformation pattern of the cornea under laser crosslinking, providing data support for assessing elasticity improvement, refractive changes, and optimizing personalized treatment parameters.

[0057] Step 305: Based on the stress and strain distribution of the cornea after laser treatment, use the formula... and Determine the distribution of elastic modulus and corneal refractive power; where, E For elastic modulus, σ For stress, Φ represents strain; Φ represents corneal refractive power. It is the ratio of the refractive index of the cornea to that of air (usually taken as 1.376). The radius of curvature of the cornea is denoted as .

[0058] Refractive power distribution Φ: Based on the curvature change of the cornea, the refractive power distribution in different regions is calculated using optical formulas.

[0059] Step 306: The distribution of changes in elastic modulus and the distribution of corneal refractive power before and after corneal cross-linking are used as biomechanical and refractive indices of the cornea after laser treatment.

[0060] Change in elastic modulus ΔE: By calculating the elastic modulus of the cornea before and after laser crosslinking, the impact of crosslinking on the biomechanical properties of the cornea is evaluated. The distribution of the change in elastic modulus in the finite element simulation prediction results is as follows: Figure 6 As shown.

[0061] The effects of biomechanical enhancement and refractive changes were then evaluated by comparing the changes in ΔE and Φ before and after crosslinking.

[0062] In another exemplary embodiment of this application, the results are analyzed and feedback is provided as follows: After the simulation results are output, the crosslinking effect is evaluated based on indicators such as changes in elastic modulus and refractive power distribution. If certain indicators do not meet the preset standards (e.g., insufficient improvement in elastic modulus), feedback is provided, and the crosslinking path parameters are regenerated to achieve closed-loop optimization. Through these steps, finite element simulation can not only predict the effect of laser crosslinking but also provide a scientific basis for the development of personalized treatment plans.

[0063] The effect of crosslinking on lesion delay and refractive improvement is evaluated using a trained regression model (such as XGBoost). If the indicators do not meet the preset standards, feedback is provided, and an optimized path is regenerated. The specific implementation process is as follows: 1. Evaluation of the effect of delaying disease progression: The effect of femtosecond laser cross-linking on delaying the progression of keratoconus was evaluated. This was primarily assessed through changes in corneal biomechanics, such as changes in elastic modulus and a reduction in corneal protrusion. The formula for calculating changes in elastic modulus is as follows: ; Where, Δ E This represents the change in elastic modulus before and after crosslinking. E post and E pre These are the elastic modulus before and after surgery, respectively.

[0064] 2. Evaluation of refractive improvement effect: The improvement in refractive error (e.g., myopia, astigmatism) was assessed by changes in refractive error before and after surgery.

[0065] The formula for calculating the change in refractive power is as follows: ; ΔRefractive Error represents the change in refractive error before and after surgery.

[0066] The "change in refractive error" here actually reflects the overall change in focal power of the ocular optical system after surgery. Its physical basis is that changes in corneal refractive power Φ lead to changes in the overall refractive state of the eye. If only the cornea as the dominant factor is considered, it can be approximated as follows: .

[0067] The cross-linking effect is evaluated using a trained regression model (such as XGBoost) based on the aforementioned assessment indicators (elastic modulus change, refractive power distribution, corneal morphological changes, etc.). If the cross-linking effect fails to meet the preset standard (e.g., the lesion delay effect is not obvious or the refractive improvement is insufficient), the assessment results are fed back to the AI ​​optimization module to regenerate a personalized cross-linking path and adjust laser parameters (such as pulse energy, scanning speed, depth of focus control, etc.) to further optimize the treatment effect.

[0068] In the prediction and optimization phase of this application, a pre-trained regression model (such as XGBoost) is used to comprehensively evaluate the cross-linking effect. The pre-trained regression model takes the elastic modulus change ΔE, refractive power distribution Φ, corneal morphological parameters (thickness T, curvature r, displacement U) output from the simulation module, and AI-generated path parameters (pulse energy Ep, scanning speed v, focal depth parameter fz) as input. By learning the nonlinear mapping relationship between multi-dimensional parameters, it outputs comprehensive prediction indicators for the cross-linking efficacy, ΔE_pred and ΔRef_pred. If ΔE_pred < ΔE_target or ΔRef_pred < ΔRef_target, indicating insufficient efficacy, the prediction residual information is fed back to the AI ​​optimization module to regenerate the cross-linking path and laser parameters. This model can comprehensively consider the coupling effect of structural features, energy distribution, and biomechanical response, reflecting the cross-linking effect more accurately than simple indicator threshold judgment. If the model prediction result does not meet the preset standard, the prediction residual and the unmet standard spatial region are fed back to the AI ​​optimization module to update the deep learning network weights and regenerate personalized cross-linking path parameters, achieving closed-loop optimization. As can be seen, the AI ​​model dynamically updates the network parameters based on simulation feedback results, achieving adaptive optimization.

[0069] Through this process, machine learning models can comprehensively evaluate the cross-linking effect based on multiple factors (such as corneal elasticity, refractive power, and corneal deformation), thereby optimizing treatment plans and ensuring personalized and precise treatment.

[0070] In another exemplary embodiment of this application, in the closed-loop optimization mechanism, the AI ​​optimization module compares the results data output by the finite element simulation analysis, including indicators such as the elastic modulus change ΔE(x,y,z) and refractive power distribution Φ(x,y) of various regions of the cornea, with preset target standards, and calculates various deviations and areas that do not meet the standards. For areas with insufficient elasticity improvement, incomplete deformation recovery, or unsatisfactory refractive improvement, the AI ​​optimization module extracts the index and deviation information of these spatial locations to form a feedback dataset containing residual matrices, optimization constraints, and local energy adjustment suggestions.

[0071] The feedback dataset is fed into the DNN+Transformer model to guide its adaptive updates. The DNN+Transformer model receives raw multimodal corneal data (OCT images, topographic maps, densitograms, and Corvis ST dynamic response images) while integrating simulation evaluation features from the AI ​​optimization module, enabling a readjustment of the network's weights and feature distributions. The DNN part improves its sensitivity to morphological and thickness features by updating the weights of the convolutional kernels; the Transformer part strengthens global correlation modeling of non-compliant regions through dynamic reallocation of self-attention weights. The feature fusion layer integrates structural and biomechanical feedback information to optimize the model's multimodal feature representation.

[0072] The adaptively updated model regenerates a new set of crosslinking path parameters. This includes optimized 3D path coordinates and depth-of-focus control parameters. Pulse energy and scanning speed The new path parameters can achieve personalized energy distribution and layer depth control while satisfying corneal structural characteristics and biomechanical objectives. After verification through finite element simulation, if the results meet expectations, the final optimized path is output; if it still fails to converge, the next round of closed-loop feedback iteration is initiated until the optimal crosslinking scheme is obtained.

[0073] In another exemplary embodiment of this application, the results are output to the clinical interface: the final crosslinking path, laser parameters, depth of focus adjustment parameters, and postoperative prediction report are output, and can be imported into a femtosecond laser surgery device to complete automatic path execution.

[0074] This application integrates corneal topography, OCT images, densitograms, Corvis ST DCR dynamic images, and refractive parameters. Through an artificial intelligence model, it outputs a personalized cross-linking scheme containing x, y, and z spatial information, laser energy, and depth-of-focus control parameters. The system combines finite element simulation and machine learning algorithms to predict the effects of different cross-linking strategies on corneal biomechanics and refractive changes, and achieves closed-loop feedback optimization. This method can achieve multi-layer deep directional cross-linking, is suitable for personalized treatment of keratoconus, and has promising clinical prospects.

[0075] The effectiveness of the method of this application will be verified by a specific embodiment below.

[0076] For a rabbit model of keratoconus with a central corneal thickness of 425 μm and a Kmax value of 57.8 D: Acquire Pentacam topographic maps, Corvis ST DCR dynamic images, and diopter -4.50DS / -2.00DC×85. ° AI model analysis determined that the output crosslinking path covers the underlying eccentric lesion area, with the path focal depth set to a gradient range of 160μm–290μm. The simulation module predicted a 29% increase in postoperative elasticity and a 1.1D reduction in corneal astigmatism. The surgical laser control module controls the laser's scanning path, energy parameters, and fz parameters, enabling automatic adjustment of focusing and crosslinking operations. The system was validated through in vitro corneal experiments, showing that the simulated predicted elasticity increase had an error of less than 10% compared to the actual measurement.

[0077] The method described in this application can be widely applied to femtosecond laser-assisted corneal cross-linking surgery. Through AI intelligent optimization and adjustable depth of focus technology, it promotes corneal cross-linking into a new stage of high precision, personalization, and predictability, providing a new intelligent solution for the treatment of keratoconus.

[0078] The principle of the method in this application is as follows: Figure 7 As shown. Figure 7 Each step can be summarized as follows: S1. Acquire the three-dimensional structure and refractive parameters of the cornea to reconstruct an individualized corneal model; S2. Input the data into the AI ​​model and output the cross-linking path containing three-dimensional coordinates and focal depth parameters; S3. Control the laser system to focus according to the focal depth parameters to achieve cross-linking at different levels; S4. Evaluate the cross-linking effect based on simulation and machine learning modules and provide feedback for optimization.

[0079] This application integrates corneal structure, optical density, biomechanics, and refractive information to generate personalized three-dimensional cross-linking pathways and regulate focal depth, thereby slowing the progression of keratoconus, improving biomechanical stability, and also partially improving refractive status. The beneficial effects of this method are as follows: 1. Introducing a z-axis depth-of-focus control mechanism to achieve precise cross-linking at different depths of the cornea; 2. By integrating preoperative multimodal data through AI, personalized output of three-dimensional cross-linking paths and laser energy parameters is generated; 3. By combining simulation and machine learning models, closed-loop self-optimization of crosslinking parameters is achieved; 4. Based on the principle of multiphoton absorption, deep nonlinear cross-linking of the cornea is achieved, improving energy utilization efficiency and biosafety; 5. It can effectively slow down the progression of keratoconus, reduce irregular astigmatism, and improve visual quality; 6. The system can be imported into femtosecond laser surgery platforms and has good potential for clinical translation.

[0080] Based on the same inventive concept, this application also provides an AI-based variable depth-of-focus femtosecond laser corneal crosslinking parameter optimization system for implementing the aforementioned AI-based variable depth-of-focus femtosecond laser corneal crosslinking parameter optimization method. The solution provided by this device is similar to the implementation described in the above method. Therefore, the specific limitations of the one or more AI-based variable depth-of-focus femtosecond laser corneal crosslinking parameter optimization system embodiments provided below can be found in the limitations of the AI-based variable depth-of-focus femtosecond laser corneal crosslinking parameter optimization method described above, and will not be repeated here.

[0081] In one exemplary embodiment, such as Figure 8 As shown, an artificial intelligence-based variable depth-of-focus femtosecond laser corneal crosslinking parameter optimization system is provided, comprising: a data acquisition module, an AI path generation module, a simulation feedback module, and a clinical output module.

[0082] The data acquisition module is used to acquire multimodal corneal images of keratoconus patients. The AI ​​path generation module is used to reconstruct a personalized three-dimensional corneal model of the keratoconus patient based on the multimodal corneal images; and to obtain cross-linking path parameters using a deep neural network based on the multimodal corneal images. The cross-linking path parameters include laser scanning path, depth-of-focus control parameters, and pulse energy.

[0083] The simulation feedback module performs stress-strain simulation analysis on the cornea cross-linked by femtosecond laser according to the cross-linking path parameters based on the personalized 3D corneal model, obtaining the biomechanical and refractive indices of the cornea after laser treatment. Based on the biomechanical and refractive indices of the cornea after laser treatment, and the cross-linking path parameters, a trained regression model is used to evaluate the cross-linking effect. If the cross-linking effect does not meet the preset standard, the deep neural network is adaptively updated using the spatial location of the failure to meet the preset standard and the deviation information between the cross-linking effect and the preset standard. The updated deep neural network is then replaced, and the process returns to execute the cross-linking path parameters obtained using the deep neural network based on the multimodal corneal images. The clinical output module determines the cross-linking path parameters as the optimal cross-linking scheme if the cross-linking effect meets the preset standard.

[0084] As an optional implementation, the depth-of-focus control module controls the objective lens or electronic control platform to dynamically focus based on the fz parameters generated by AI.

[0085] As an alternative implementation, the system can export a crosslinking path file and automatically execute the path via the surgical device interface.

[0086] As an optional implementation, the system also includes a dynamic feedback module that can receive Corvis ST dynamic images in real time and correct the boundary conditions of the simulation model to improve prediction accuracy.

[0087] In one example, this application embodiment also provides a corneal crosslinking device that employs the aforementioned artificial intelligence-based variable depth-of-focus femtosecond laser corneal crosslinking parameter optimization system.

[0088] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0089] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for optimizing crosslinking parameters of variable depth-of-focus femtosecond laser cornea based on artificial intelligence, characterized in that, include: Acquire multimodal corneal images from patients with keratoconus; Based on the multimodal corneal images, a personalized three-dimensional corneal model for keratoconus patients is reconstructed; Based on the multimodal corneal images, crosslinking path parameters are obtained using a deep neural network; the crosslinking path parameters include laser scanning path, depth of focus control parameters, and pulse energy. Based on the personalized three-dimensional corneal model, stress-strain simulation analysis was performed on the cornea cross-linked by femtosecond laser according to the cross-linking path parameters to obtain the biomechanical and refractive indices of the cornea after laser treatment. Based on the biomechanical and refractive indices of the cornea after laser treatment, and the cross-linking path parameters, a trained regression model is used to evaluate the cross-linking effect; If the cross-linking effect does not meet the preset standard, the deep neural network is adaptively updated using the spatial location where the cross-linking effect does not meet the preset standard and the deviation information between the cross-linking effect and the preset standard. The deep neural network is then replaced with the adaptively updated deep neural network, and the process returns to the step "According to the multimodal corneal image, the cross-linking path parameters are obtained using the deep neural network". If the crosslinking effect reaches the preset standard, then the crosslinking path parameters are determined as the optimal crosslinking scheme.

2. The method for optimizing corneal crosslinking parameters based on artificial intelligence using femtosecond laser with variable depth of focus, as described in claim 1, is characterized in that... The multimodal corneal images include: multimodal images of the anterior corneal surface and multimodal images of the posterior corneal surface; The multimodal images include surface topography maps, OCT images, optical density maps, and Corvis ST dynamic corneal response images.

3. The method for optimizing corneal crosslinking parameters based on artificial intelligence using femtosecond laser with variable depth of focus, as described in claim 1, is characterized in that... Based on the multimodal corneal images, a personalized three-dimensional corneal model for a keratoconus patient is reconstructed, specifically including: The multimodal corneal images are preprocessed to obtain preprocessed multimodal images; the preprocessing includes normalization, noise reduction, resolution unification, and time series synchronization. Based on the preprocessed multimodal images, image segmentation technology is used to segment the multi-layer structure of the cornea to obtain segmented multi-source data; Three-dimensional non-rigid registration is performed on the preprocessed multimodal images to obtain registered multi-source data; Based on the segmented and registered multi-source data, a personalized three-dimensional corneal model for keratoconus patients is generated.

4. The method for optimizing corneal crosslinking parameters based on artificial intelligence using femtosecond laser with variable depth of focus, as described in claim 1, is characterized in that... The deep neural network includes: an input layer, a convolutional neural network, a Transformer model, a feature fusion layer, and an output layer; The input layer is used to input multimodal corneal images; Convolutional neural networks are used to extract local features of the cornea from multimodal corneal images; The Transformer model is used to extract global features of the cornea based on multimodal corneal images; The feature fusion layer is used to fuse local and global features of the cornea to obtain fused features; The output layer is used to output crosslinking path parameters based on the fusion characteristics.

5. The method for optimizing corneal crosslinking parameters based on artificial intelligence using a femtosecond laser with variable depth of focus, as described in claim 1 or 4, is characterized in that... The crosslinking path parameters also include: scanning speed; The spatial coordinates of the laser scanning path characterize the spatial location of the laser cross-linking path within the corneal tissue; The depth-of-focus control parameter refers to the focal depth of the femtosecond laser at each spatial positioning point. The pulse energy is the unit energy of the laser at each spatial positioning point; The scanning speed is the rate at which the laser moves along the path segment.

6. The method for optimizing corneal crosslinking parameters based on artificial intelligence using femtosecond laser with variable depth of focus according to claim 1, characterized in that, Based on the personalized three-dimensional corneal model, stress-strain simulation analysis was performed on the cornea cross-linked by femtosecond laser according to the cross-linking path parameters to obtain the biomechanical and refractive indices of the cornea after laser treatment, specifically including: The nonlinear mechanical behavior of corneal tissue is described using a Neo-Hookean hyperelastic material model; Set the boundary conditions for the corneal simulation; Discretize the personalized three-dimensional corneal model; Based on the Neo-Hookean hyperelastic material model, boundary conditions, and discretized personalized three-dimensional corneal model, a nonlinear finite element iterative method is used to perform stress-strain simulation analysis on the cornea cross-linked by femtosecond laser according to the cross-linking path parameters, and to obtain the stress and strain distribution of the cornea after laser treatment. Based on the stress and strain distribution of the cornea after laser treatment, using the formula and Determine the distribution of elastic modulus and corneal refractive power; where, E For elastic modulus, σ For stress, Φ represents strain; Φ represents corneal refractive power. The ratio of the refractive index of the cornea to that of air. The radius of curvature of the cornea; The distribution of changes in elastic modulus and the distribution of corneal refractive power before and after corneal cross-linking were used as biomechanical and refractive indices of the cornea after laser treatment.

7. The method for optimizing corneal crosslinking parameters based on artificial intelligence using femtosecond laser with variable depth of focus according to claim 4, characterized in that, Adaptive updates in deep neural networks include: updating the weights of convolutional feature kernels in convolutional neural networks, dynamically reallocating self-attention weights in Transformer models, and optimizing the multimodal feature representation of feature fusion layers.

8. The method for optimizing corneal crosslinking parameters based on artificial intelligence using femtosecond laser with variable depth of focus according to claim 1, characterized in that, Femtosecond lasers use multiphoton absorption properties to excite riboflavin to achieve localized energy deposition within corneal tissue, forming a cross-linking reaction zone deep within the cornea without the need for surface irradiation.

9. A variable depth-of-focus femtosecond laser corneal crosslinking parameter optimization system based on artificial intelligence, characterized in that, include: Data acquisition module, AI path generation module, simulation feedback module, and clinical output module; The data acquisition module is used to acquire multimodal corneal images of patients with keratoconus; The AI ​​path generation module is used to reconstruct a personalized three-dimensional corneal model for keratoconus patients based on the multimodal corneal images; and to obtain cross-linking path parameters using a deep neural network based on the multimodal corneal images; the cross-linking path parameters include laser scanning path, depth of focus control parameters, and pulse energy. The simulation feedback module is used to perform stress-strain simulation analysis on the cornea cross-linked by femtosecond laser according to the cross-linking path parameters based on the personalized three-dimensional corneal model, and to obtain the biomechanical and refractive indices of the cornea after laser treatment; based on the biomechanical and refractive indices of the cornea after laser treatment and the cross-linking path parameters, the trained regression model is used to evaluate the cross-linking effect. If the crosslinking effect does not meet the preset standard, the deep neural network is adaptively updated using the spatial location where the crosslinking effect does not meet the preset standard and the deviation information between the crosslinking effect and the preset standard. The deep neural network is then replaced with the adaptively updated deep neural network, and the process returns to the point where the crosslinking path parameters are obtained using the deep neural network based on the multimodal corneal image. The clinical output module is used to determine the crosslinking path parameters as the optimal crosslinking scheme if the crosslinking effect meets the preset standard.

10. A corneal crosslinking device, characterized in that, The corneal crosslinking device uses the artificial intelligence-based variable depth-of-focus femtosecond laser corneal crosslinking parameter optimization system described in claim 9.

Images