A method and system for simulating intraocular light fields based on physiological constraints and pathological origins
By using the PI-KAN model based on physiological constraints and pathological origins, the problems of low computational efficiency, insufficient modeling ability and poor interpretability of existing ophthalmic OCT simulation technology are solved, realizing efficient and accurate personalized intraocular light field simulation, supporting real-time clinical assistance and personalized diagnosis.
Patent Information
- Application Number
- CN202511901578.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-16
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2045-12-16
AI Technical Summary
Existing ophthalmic OCT simulation technology suffers from low computational efficiency, insufficient modeling capabilities, loose physical constraint embedding, poor interpretability, and weak personalization adaptability, failing to meet the needs of real-time clinical assistance and personalized diagnosis.
We employ the Physical Information Kolmogorov-Arnold Network (PI-KAN) model based on physiological constraints and pathological tracing. Through hierarchical eyeball geometry modeling, physical constraint embedding, and adaptive mesh optimization, we construct a five-dimensional input Kolmogorov-Arnold network. Combining the Helmholtz equation and physical laws such as Fresnel reflection and Snell's law, we achieve efficient light field simulation.
It achieves millisecond-level light field simulation inference, significantly improving simulation accuracy and interpretability, supporting personalized eye parameter adaptation, meeting real-time clinical needs, and reducing resource requirements.
Smart Images

Figure CN121366739B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of biomedical imaging and artificial intelligence machine learning. Specifically, it relates to a method and system for simulating intraocular light fields based on physiological constraints and pathological tracing. A novel physical information Kolmogorov-Arnold network (PIKAN) model is designed and combined with optical coherence tomography (OCT) modal data for simulating the light field of the eye. It is applicable to scenarios such as ophthalmic clinical diagnosis assistance, OCT equipment research and development optimization, ophthalmic surgical planning, and medical education and training. It can achieve high-precision simulation, high-efficiency training, and physiological and pathological interpretability analysis of high-dimensional light field propagation in the microenvironment of the eye. Background Technology
[0002] Currently, the field of OCT eye light field simulation mainly relies on three types of technical solutions, as follows: 1. Finite-Difference Time-Domain (FDTD): Directly solves Maxwell's equations and simulates electromagnetic field propagation through three-dimensional spatial discretization. It requires a fine mesh of λ / 20 (λ is the wavelength of light, approximately 1300nm), with a typical computational domain of 30mm×30mm×30mm, requiring millions of mesh elements; 2. Monte Carlo Method (MC): Simulates the propagation trajectory of photons in eye tissue through random sampling, requiring simulation of... Only the above photons can guarantee statistical accuracy, and a single simulation requires several hours to several days; 3. Physical information multilayer perceptron network (MLP-based PINNs): based on fully connected neural networks embedding physical constraints (such as Helmholtz equations), the typical architecture is [5,50,50,1] (input dimension 5: spatial x / y / z, wavelength λ, time t), with more than 15,000 parameters.
[0003] The aforementioned technologies have become routine applications in ophthalmic OCT simulations. For example, FDTD is used for simulating corneal refractive properties, MC is used for simulating retinal scattering effects, and MLP-based PINNs are used to simplify the light field solution of eyeball models.
[0004] The shortcomings of existing technologies: Existing technologies have the following key drawbacks, making it difficult to meet clinical and engineering needs: 1. Low computational efficiency: The computational complexity of FDTD is O(n log n). (N is the number of dimension samples) The inference time for a single OCT B scan (512×512 pixels) reaches 2345ms; MC requires a large number of photon samples, which cannot meet the real-time requirements at all; the training time of MLP-based PINNs reaches 456s and the inference time is 12.3ms, which is still not suitable for clinical real-time auxiliary scenarios; 2. Insufficient high-dimensional modeling capability: OCT light field involves 5-dimensional parameters (x / y / z / λ / t), FDTD and MC face the "curse of dimensionality", and the number of parameters grows exponentially; MLP-based PINNs have parameter explosion in high-dimensional scenarios (more than 15,000 5-dimensional modeling parameters), which is prone to overfitting; 3. Loose embedding of physical constraints: Existing technologies mostly add physical laws (such as Fresnel reflection and Snell's law) as "soft constraints" to the loss function, without deep integration with the model structure, resulting in the simulation accuracy being greatly affected by the training data, for example, MLP-based PINNs 4. Poor interpretability: FDTD and MC are pure numerical black boxes, and the activation functions of MLP-based PINNs are concentrated at nodes, making it impossible to trace the physical mechanism of light field propagation and difficult to locate the source of OCT imaging artifacts (such as lens refraction artifacts); 5. Weak personalization adaptability: There are individual differences in ocular physiological parameters (such as corneal curvature and lens thickness). Existing technologies need to be retrained or have a large number of parameters adjusted to adapt to different individuals, with a generalization error of over 15%, which cannot meet the needs of personalized diagnosis and does not support effective on-site learning and dynamic updates. Summary of the Invention
[0005] The technical problem this invention aims to solve is to overcome the aforementioned shortcomings and key defects of existing technologies. To address these problems, it provides a method and system for simulating the intraocular light field based on physiological constraints and pathological tracing. Its core model is the physical information Kolmogorov-Arnold network model, also known as the PI-KAN network.
[0006] The objective of this invention is achieved through the following method: a method for simulating the intraocular light field based on physiological constraints and pathological origins, the method comprising the following steps:
[0007] S1. Parametric Modeling of the Eyeball: Establish a layered geometric model of the eyeball tissues, dividing the eyeball into the cornea, anterior chamber, iris, lens, vitreous body, retina, Bruch's membrane, and choroid, and configuring the geometric and optical parameters of each layer; wherein the geometric parameters include at least the radius of curvature, thickness, and spatial position, and the optical parameters include at least the wavelength-related refractive index distribution and scattering coefficient; S2. Construction and Configuration of the Physical Information Kolmogorov-Arnold Network: Construct a three-layer Kolmogorov-Arnold network, the input layer of which is configured to receive a five-dimensional input vector, the five-dimensional input vector consisting of spatial coordinates. The network consists of light wavelength λ and time t; each of the two hidden layers contains 11 neurons, and the connections between neurons in adjacent layers are achieved through learnable side functions; the output layer of the network is configured to output the complex amplitude of the light field corresponding to the input vector. S3. Physical Constraint Embedding and Network Training: Defining the Composite Loss Function The network is trained; wherein, The network output is the fitting residual of the Helmholtz equations that govern the propagation of the light field, and the calculation of this residual depends on the network's output throughout the computational domain. The network output is the fitting residual of the theoretical Fresnel reflection or transmission conditions at the interfaces of various tissues in the eye; α, β, γ are preset weight coefficients, and satisfy the weight ratio of α / β / γ as 5:3:2. The training process makes physical laws dominate the optimization of network parameters in the form of hard constraints. For high-quality OCT training data; S4, Personalized light field solution: Input the personalized parameters of the target eyeball to be simulated into the network trained in step S3. The network maps the five-dimensional input vector into the corresponding three-dimensional spatial light field distribution through forward propagation; S5, OCT image synthesis: Based on the light field distribution obtained in step S4... Compare it with the preset reference arm plane wave field Perform interferometric superposition and calculate the interference signal. ; for the interference signal Performing a Fast Fourier Transform along the wavelength dimension yields the depth-resolved OCT A scan signal. This leads to the synthesis of B-scan or C-scan OCT images; S6, Interpretability Analysis and Output: Extracting and visualizing the edge functions in the network associated with specific anatomical layers (such as the cornea and retina). By programmatically fine-tuning a single geometric or optical parameter in the input parameters, the influence of that parameter on the features of the OCT image generated in step S5 is quantitatively analyzed, and a parameter sensitivity report is generated. The OCT image, the side function visualization results, and the parameter sensitivity report are then integrated and output.
[0008] The learnable edge function mentioned in step S2 adopts a cubic B-spline basis function with enhanced physical information, and its expression is: in, and These are learnable weight coefficients used to balance nonlinear activation and spline fitting; Here, G is the cubic B-spline basis function; G is the number of spline intervals, and k=3 is the spline order. Furthermore, the side functions are specifically designed for different physical processes in light field propagation, including: a reflection side function designed based on Fresnel's formula, used to encode interface reflectivity, whose expression is: The refraction side function, designed based on Snell's law, is used to encode light deflection; its expression is: A scattering side function designed based on Mie scattering theory is used to encode tissue scattering effects.
[0009] Step S3 of the network training process further includes an adaptive grid optimization strategy, specifically: during initial training, the number of spline intervals G is set to 3; during training, the loss function L is calculated and analyzed in real time. PDE The distribution within the spatial computational domain; when the residual of a specific region (such as the retinal layer) is detected to be continuously higher than a preset threshold, the number of spline intervals G in that region is automatically expanded to 5~7; after expansion, the spline coefficients on the newly added grid points are reinitialized using the least squares method, and training continues on the expanded grid until the model converges.
[0010] The "quantitative analysis of its impact on OCT image features" mentioned in step S6 specifically includes: establishing "input parameter perturbation" Changes in the response of the edge function Image feature offset The mapping relationship is defined as follows: the image feature F includes the depth position of the interlayer reflection peak, the signal amplitude, and the signal-to-noise ratio; by analyzing this mapping relationship, the artifacts appearing in OCT imaging can be traced back to their source physical processes or anatomical structures.
[0011] An intraocular light field simulation system for implementing the method is provided. The system comprises the following modules interconnected via a data bus: a parameter input and interface module for receiving individual eye parameter data from an external corneal topography instrument and an optical biometer, or for receiving parameters manually input by the user through a graphical interface, and formatting the parameters before transmitting them to a storage module and a model calculation module; a physical information KAN model calculation module, which is the core calculation unit and internally contains a trained physical information Kolmogorov-Arnold network model; this module receives instructions and data from the parameter input module and performs light field calculations; and an OCT signal synthesis and processing module, directly connected to the model calculation module, receiving the output light field data and performing interferometric calculations and frequency domain transformations. The system generates raw OCT image data; the interpretability analysis module is connected to both the model calculation module and the OCT signal processing module, and has permission to access the side function weight data inside the model calculation module, and calls the OCT image data to perform side function visualization, parameter sensitivity calculation and artifact mapping analysis; the unified output and display module is connected to both the OCT signal processing module and the interpretability analysis module, and is used to integrate and display the generated OCT images, dynamic side function curves, parameter sensitivity heatmaps and analysis conclusions in a graphical user interface, and provide interactive functions; the central storage and management module is connected to all the above modules, and is used to store the individual eye parameter database, the parameter set of the pre-trained network model for different pathological states, historical simulation results and customized analysis reports.
[0012] The physical information KAN model calculation module is configured to support heterogeneous computing and can be loaded and run on a central processing unit, a graphics processing unit, or an embedded edge AI chip. When the system is deployed on an RK3588 smart hardware development board or an NVIDIA Jetson AGX Orin edge computing platform, the total time from parameter input to the generation and basic analysis of a single 512x512 pixel OCTB scan image does not exceed 1.2 milliseconds, and the overall power consumption is less than 15 watts.
[0013] The beneficial effects of this invention are: extremely high computational efficiency: By adopting a low-parameter KAN network structure and combining it with an adaptive mesh, the inference time of five-dimensional light field simulation is shortened to the millisecond level, meeting the needs of real-time clinical interaction.
[0014] High modeling accuracy: By "hard-coding" the core physical laws into the network's side functions and loss functions, the physical consistency of simulation training is significantly improved, and the simulation error of interlayer reflection peaks is significantly reduced.
[0015] High interpretability: By utilizing the interpretability of the KAN network side functions, a visual mapping of key physical processes in light field propagation (such as refraction and scattering) is achieved, which can effectively trace OCT imaging artifacts.
[0016] High personalization adaptability: The system supports rapid input of individualized eye parameters, has strong model generalization ability, and can efficiently adapt to different patients or pathological conditions, providing support for personalized medicine and enabling on-site learning and dynamic updates.
[0017] Low resource requirements: The model has a small number of parameters, can be deployed on edge computing devices, and has low power consumption, making it easy to use in primary healthcare institutions or portable devices. Attached Figure Description
[0018] Figure 1 This is a flowchart of the intraocular light field simulation system of the present invention.
[0019] Figure 2 This is a schematic diagram of the physical information Kolmogorov-Arnold network (PI-KAN) algorithm of the present invention.
[0020] Figure 3 This is a hardware diagram of the intraocular light field simulation system of the present invention. Detailed Implementation
[0021] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0022] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same technical meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0023] like Figures 1 to 3 As shown, this invention provides a method and system for simulating the intraocular light field based on physiological constraints and pathological origins. The modeling revolves around the intraocular light field, which primarily consists of eight highly correlated anatomical interfaces: cornea, aqueous humor, iris (pupil), lens, vitreous body, retina, Bruch's membrane, and choroid. From the perspective of physical information laws, this system mainly involves four optical laws: Fresnel's law of reflection, Snell's law of refraction, Mie's law of scattering, and Helmholtz's equations of optics. A novel physical information Kolmogorov-Arnold network, denoted as the PI-KAN network or physical information KAN network, is designed as the core model architecture of this system. From a morphological perspective, the eight intraocular anatomical interfaces serve as the backbone architecture of the PI-KAN network for the following reasons: each of the eight intraocular anatomical interfaces has a clear boundary, and its thickness / curvature is measurable, facilitating the construction of KAN node coordinates.
[0024] The fact that only adjacent structures are connected between layers determines that the number of edge functions is O(number of layers) instead of O(n²), and the number of model parameters drops by 1-2 orders of magnitude.
[0025] The boundary derivative can jump, and the multiple nodes of the B-spline automatically satisfy the characteristic of "discontinuous Fresnel reflection coefficient".
[0026] The method includes the following steps: S1, Parametric modeling of the eyeball: Establishing a layered geometric model of the eyeball tissues, wherein the layered geometric model divides the eyeball into the cornea, anterior chamber, iris, lens, vitreous body, retina, Bruch's membrane, and choroid, and configuring the geometric and optical parameters of each layer; wherein the geometric parameters include at least the radius of curvature, thickness, and spatial position, and the optical parameters include at least the wavelength-related refractive index distribution and scattering coefficient; S2, Construction and configuration of a physical information Kolmogorov-Arnold network: Constructing a three-layer Kolmogorov-Arnold network, wherein the input layer of the network is configured to receive a five-dimensional input vector, wherein the five-dimensional input vector consists of spatial coordinates. The network consists of light wavelength λ and time t; each of the two hidden layers contains 11 neurons, and the connections between neurons in adjacent layers are achieved through learnable side functions; the output layer of the network is configured to output the complex amplitude of the light field corresponding to the input vector. S3. Physical Constraint Embedding and Network Training: Defining the Composite Loss Function The network is trained; wherein, The network output is the fitting residual of the Helmholtz equations that govern the propagation of the light field, and the calculation of this residual depends on the network's output throughout the computational domain. The network output is the fitting residual of the theoretical Fresnel reflection or transmission conditions at the interfaces of various tissues in the eye; α, β, γ are preset weight coefficients, and satisfy the following: the weight ratio of α / β / γ is 5:3:2. The training process makes physical laws dominate the optimization of network parameters in the form of hard constraints. For high-quality OCT training data; S4, Personalized light field solution: Input the personalized parameters of the target eyeball to be simulated into the network trained in step S3. The network maps the five-dimensional input vector into the corresponding three-dimensional spatial light field distribution through forward propagation; S5, OCT image synthesis: Based on the light field distribution obtained in step S4... Compare it with the preset reference arm plane wave field Perform interferometric superposition and calculate the interference signal. ; for the interference signal Performing a Fast Fourier Transform along the wavelength dimension yields the depth-resolved OCT A scan signal. This leads to the synthesis of B-scan or C-scan OCT images; S6. Interpretability Analysis and Output: Extract and visualize the edge functions in the network associated with specific anatomical layers (such as the cornea and retina). By programmatically fine-tuning a single geometric or optical parameter in the input parameters, the influence of that parameter on the features of the OCT image generated in step (S5) is quantitatively analyzed, and a parameter sensitivity report is generated. The OCT image, the side function visualization results, and the parameter sensitivity report are then integrated and output.
[0027] The learnable edge function mentioned in step S2 adopts a cubic B-spline basis function with enhanced physical information, and its expression is: in, and These are learnable weight coefficients used to balance nonlinear activation and spline fitting; Here, G is the cubic B-spline basis function; G is the number of spline intervals, and k=3 is the spline order. Furthermore, the side functions are specifically designed for different physical processes in light field propagation, including: a reflection side function designed based on Fresnel's formula, used to encode interface reflectivity, whose expression is: The refraction side function, designed based on Snell's law, is used to encode light deflection; its expression is: A scattering side function designed based on Mie scattering theory is used to encode tissue scattering effects.
[0028] Step S3 of the network training process further includes an adaptive grid optimization strategy, specifically: during initial training, the number of spline intervals G is set to 3; during training, the loss function L is calculated and analyzed in real time. PDE The distribution within the spatial computational domain; when the residual of a specific region (such as the retinal layer) is detected to be continuously higher than a preset threshold, the number of spline intervals G in that region is automatically expanded to 5~7; after expansion, the least squares method is used to reinitialize only the spline coefficients ci on the newly added grid points, and training continues on the expanded grid until the model converges.
[0029] The "quantitative analysis of its impact on OCT image features" mentioned in step S6 specifically includes: establishing a mapping relationship of "input parameter perturbation ΔP → side function response change Δϕ → OCT image feature offset ΔF"; the image feature F includes the depth position of the interlayer reflection peak, signal amplitude and signal-to-noise ratio; by analyzing this mapping relationship, the artifacts appearing in OCT imaging can be traced back to their source physical processes or anatomical structures.
[0030] An OCT eye light field simulation system for implementing the method, the system includes the following modules interconnected via a data bus: a parameter input and interface module: for receiving individual eye parameter data from an external corneal topography instrument and optical biometry instrument, or for receiving parameters manually input by the user through a graphical interface, and formatting the parameters before transmitting them to a storage module and a model calculation module; a physical information KAN model calculation module: the core calculation unit, which internally contains a trained physical information Kolmogorov-Arnold network model; this module receives instructions and data from the parameter input module and performs light field calculations; and an OCT signal synthesis and processing module: directly connected to the model calculation module, receiving its output light field data, and performing interferometric calculations and... Frequency domain transformation generates raw OCT image data; Interpretability analysis module: connected to both the model calculation module and the OCT signal processing module, with access to the side function weight data within the model calculation module, and can call OCT image data to perform side function visualization, parameter sensitivity calculation, and artifact mapping analysis; Unified output and display module: connected to both the OCT signal processing module and the interpretability analysis module, used to integrate and display the generated OCT images, dynamic side function curves, parameter sensitivity heatmaps, and analysis conclusions in a graphical user interface, and provide interactive functions; Central storage and management module: connected to all the above modules, used to store individual eye parameter databases, pre-trained network model parameter sets for different pathological states, historical simulation results, and analysis reports.
[0031] The physical information KAN model calculation module is configured to support heterogeneous computing and can be loaded and run on a central processing unit, a graphics processing unit, or an embedded edge AI chip. When the system is deployed on the NVIDIA Jetson AGXOrin edge computing platform, the total time from parameter input to the generation and basic analysis of a single 512x512 pixel OCT B scan image does not exceed 1.2 milliseconds, and the overall power consumption is less than 15 watts.
[0032] Deploying the domestically produced RK3588 platform, the total time for generation and analysis is 1.8~3.5 milliseconds, and the overall power consumption is 6~12 watts. Although the computing power of the RK3588 platform is lower than that of Jetson AGX Orin, it can still achieve "millisecond-level inference + low-power operation" due to the lightweight model and heterogeneous computing optimization. It also has the advantages of being fully domestically produced and low-cost (the unit price of the module is only 15~20 US dollars), making it more suitable for deployment in primary medical institutions or portable OCT simulation equipment.
[0033] The technical solution of this invention comprises two parts: a hardware system and a software model. The core is a hierarchical modular architecture based on PIKANs, specifically as follows: Firstly, the system hardware composition of this invention is as follows: 1. Parameter input module: including clinical parameter acquisition equipment (such as a corneal topography instrument or an optical biometer) and a parameter interaction interface, used to input ocular geometric parameters (corneal curvature R1 / R2, anterior chamber depth ACD, etc.) and optical parameters (refractive index n of each layer, scattering coefficient). 1. **Components:** 2. **Computation Module:** Supports CPU (Intel i7-10700K), GPU (NVIDIA RTX 3080), or edge computing chip (NVIDIA Jetson AGX Orin, domestic Rockchip RK3588) for model training and inference; 3. **Result Output Module:** Includes an OCT image display unit, a side function visualization interface, and a parameter sensitivity analysis report generation unit; 4. **Storage Module:** Used to store the eye parameter database (including normal and pathological samples), pre-trained PIKANs (Physics-InformedKANs) model weights, and simulation results.
[0034] Secondly, the system software module of this invention includes a core algorithm model and a module architecture. The software model adopts a four-layer modular design, and the functions and technical details of each layer are as follows: Geometric parameter modeling of the eyeball parameterization module: A segmented spherical model is used to describe the layered structure of the eyeball. Specific parameters include: Cornea: anterior surface curvature R1=7.8±0.5mm, posterior surface curvature R2=6.5±0.5mm, thickness 0.5±0.1mm; anterior chamber depth 3±0.3mm; lens: anterior surface curvature 10±1mm at rest. m, posterior surface curvature 6±1mm, anterior surface curvature minimum 5.3mm during adjustment, thickness 4±0.3mm; vitreous body: vitreous cavity length 16±0.5mm; retina: neuroepithelial layer, 150±20μm; retinal pigment epithelium (RPE), 10±2μm; Bruch's membrane: normal thickness: 2±0.5μm; diameter: matching the retina (approximately 22mm); choroid: thickness: 200±50 (thickest in the macula, thinning at the periphery); vascular density: 50%±10%.
[0035] Optical parameter modeling: Introducing wavelength dependence and spatial distribution characteristics, the specific expression is: Corneal refractive index: =1.376-0.0012×(λ-589) (λ unit: nm); Anterior chamber water refractive index: =1.336-0.0005×(λ-589); Lens gradient refractive index: core =1.40, the cortex linearly decreases to 1.38; Iris: Refractive index model: Iris stroma layer: Pupil area (air): n=1.0; Scattering coefficient: Iris stroma layer: Absorption coefficient: Iris stroma layer: .
[0036] Refractive index gradient of lens: nucleus =1.40, linearly decreasing to 1.38 in the cortex; Refractive index model: Gradient distribution formula: Scattering coefficient: Age-dependent formula: Absorption coefficient: Normal state: Age-related correlation: increases by 0.005 every 10 years.
[0037] Vitreous body: Refractive index model: Normal state: After liquefaction: n=1.336 (close to the aqueous humor); Scattering coefficient: Normal state: After liquefaction: rises to 0.05~0.1; absorption coefficient: under normal conditions: .
[0038] Retina:
[0039]
[0040] Bruch film: The Bruch film is a thin elastic film between RPE and chord membrane. It is extremely thin but has critical optical properties. The parameters are as follows:
[0041]
[0042] Choroid: The choroid is the vascular layer on the outer surface of the retina, containing a large amount of melanin and blood vessels. Its optical properties are mainly strong scattering and strong absorption. The parameters are as follows:
[0043]
[0044] Scattering coefficient: cornea Lens (increases with age) / 10 years), retina Dynamic parameter modeling: considering physiological regulatory functions, pupil diameter (I represents light intensity) (Threshold light intensity), lens accommodation amplitude 15D.
[0045] 2. PIKANs Light Field Solving Module: This module is the core module of the system software architecture. It designs the edge functions, the core components of the PI-KAN network, as follows: The edge functions of the PI-KANs model employ physically enhanced cubic B-spline basis functions, expressed as: in: These are the weighting coefficients. Let k be the number of grid intervals (initially 3, adaptively expanded to 7), and k=3 (spline order). B-spline basis functions; dedicated edge functions designed for different physical processes: reflection edge function: (Encoding Fresnel reflection); Refraction side function: (Encoding Snell's Law); Scattering Side Function: Based on Mie scattering theory, introducing particle size and refractive index difference parameters; Physical Constraint Embedding: Hard constraints are achieved through the residual loss function, with a total loss of: in: Helmholtz equation residuals ( , (Number of points within the domain). Boundary condition residuals ( (Number of boundary points); High-quality OCT training data , , The weight ratio is 5:3:2; Adaptive grid technique: Initially, a coarse grid of G=3 is used for training. Through error distribution analysis, it is expanded to G=5~7 in high gradient regions (such as the retinal layer). The coefficients of the expanded grid are re-initialized using the least squares method to avoid retraining the entire network.
[0046] 3. OCT signal generation module interference signal calculation: Based on the superposition of the light fields of the sample arm and the reference arm, the expression is: in (The reference light is a plane wave,) For amplitude, (for angular frequency); spectral domain processing: for Perform a Fast Fourier Transform (FFT) to obtain the depth-resolved signal: ( (where c is the speed of light); System parameter modeling: Introducing OCT hardware characteristics, including the center wavelength of the light source. =1300nm, bandwidth =100nm (Gaussian spectrum), detector noise level ≤ V, the scanning mode supports switching between A / B / C scans.
[0047] The interpretation and analysis display module, as an important functional module of the system described in this invention, directly serves the analysis and diagnosis of ophthalmologists. Based on evidence-based medicine and interpretable logic, it provides analysis and display of pathological processes in the anterior and posterior segments of the eye.
[0048] Edge function visualization is the main function of the analysis and display module. This function plots the edge functions corresponding to each depth z in real time. The amplitude of the corneal layer boundary function reflects the refractive intensity, while the fluctuation of the retinal layer boundary function reflects the scattering characteristics.
[0049] In PI-KAN, the side function corresponding to the Bruch film is a composite function of **reflection side function (based on Fresnel formula) + scattering side function (based on Mie scattering theory)**, and its amplitude directly reflects two major physical properties: the amplitude of the reflection side function. : Reflection intensity at the interface between the Bruch film and the RPE / vein film (dependent on refractive index difference and interface smoothness); amplitude of the scattering side function. : Encodes the scattering intensity of the Bruch film itself (dependent on thickness, scattering coefficient, and degree of calcification); composite amplitude ( Since interface reflection is the main source of Bruch membrane signal in OCT.
[0050] Changes in the amplitude of the side function → lesion characteristics ( The following is an example of the mapping relationship: First group: Side function amplitude change mode: Composite amplitude Sustained increase (↑20%~60%); corresponding pathological features: Bruch's membrane thickening + mild calcification; matching OCT image features ( Bruch film reflective band thickening (normal) The signal strength increases (SNR↑15%~30%), and the boundaries become clearer.
[0051] Group 2: Side function amplitude variation pattern: Composite amplitude Fluctuations (±10%~30%); corresponding lesion characteristics: local thinning of Bruch membrane + uneven lipid deposition; matching OCT image characteristics (ΔF): uneven thickness of Bruch membrane reflective band, fluctuating signal intensity, alternating dark and bright areas (corresponding to lipid deposition and thinning areas).
[0052] Parameter sensitivity analysis: By perturbing a single parameter (such as retinal thickness ±50μm), the OCT signal change curve (reflection peak depth shift, amplitude attenuation rate) is output; Artifact source tracing: Establish the mapping relationship between abnormal fluctuations of the edge function and imaging artifacts, such as the refraction artifacts corresponding to nonlinear distortion of the lens layer edge function.
[0053] Model Training and Inference Process 1. Data Preparation: Collect 10-20 sets of clinical ocular parameters (including normal and pathological samples) to generate a 5D light field training dataset (sample size) ); 2. Pre-training: Using the Adam optimizer, with an initial learning rate of Adaptive decay, training iterations of 2000 times, reduced MSE to 3. Adaptive Mesh Optimization: Analyze the error distribution of the pre-trained model, expand the mesh in high gradient regions, and retrain 500 times; 4. Inference: Input the eyeball parameters to be simulated, call the pre-trained model, and output the OCT image and interpretability analysis results within 1.2ms.
[0054] Example: The following uses "Simulation of Intraocular Light Field in Healthy Adults" as an example to illustrate the implementation process and method of this invention in detail: 7.1 Implementation Preparation 1. Hardware Configuration: CPU is Intel i7-10700K, GPU is NVIDIA RTX3080, and memory is 32GB; 2. Software Environment: Python 3.9, PyTorch 2.0, FFT library is SciPy, and visualization tool is Matplotlib; 3. Parameter Input: Corneal curvature R1=7.8mm and R2=6.5mm are collected using a corneal topography instrument, and anterior chamber depth of 3mm, lens thickness of 4mm, and total axial length of 24mm are collected using an optical biometer; Optical parameters are set as follows: cornea =1.376, anterior atrial water =1.336, lens =1.40 (nucleus), vitreous body =1.337; Scattering coefficient :cornea Lens retina The retina, Bruch's membrane, and choroid are the three core physiological structures of the fundus in this invention system.
[0055] 2. Model Training Steps 2.1. Dataset Generation: Collect 10-20 sets of clinical ocular parameters (including normal and pathological samples) to generate a 5D light field training dataset (sample size... Based on the input parameters, 5-dimensional light field training data is generated (x: 0~10mm, y: 0~10mm, z: 0~24mm, λ: 1250~1350nm, t: 0~). s), sample size The dataset is divided into training and validation sets in an 8:2 ratio; 2.2. Network initialization: Construct a PIKANs architecture of [5,11,11,1], using cubic B-splines for the side functions (G=3, k=3), and initialize the weights using a Glorot-like scheme; 2.3. Pre-training: Set α=5, β=3, γ=2, and the Adam optimizer learning rate... Training 2000 times, calculating the validation set MSE every 50 iterations, when MSE ≤ Stop at this point; at this time, the training loss L = 8 × 2.4. Mesh Optimization: Analysis of the validation set error distribution revealed a larger error in the retinal layer (z=22~24mm) (MSE=5×). Expand the mesh in this region to G=7, retrain 500 times, and reduce the MSE of the retinal layer to 8× 2.5. Model saving: Save the pre-trained model (9.2MB in size) and record the optimal hyperparameters.
[0056] 3. Simulation Implementation and Result Output 3.1. Inference Trigger: Input the eyeball parameters to be simulated, which are consistent with the training parameters. Call the pre-trained model and output the OCT image and interpretability analysis results within 1.2ms; 3.2. Light Field Solution: The model completes the 5-dimensional light field calculation within 1.2ms. Solving for the complex amplitude distribution; 3.3. OCT signal generation: calculating the interference signal. After FFT, a depth image was obtained, with a foveal depth of 200 μm, which has an error of 1% compared to the clinically measured value (198±2 μm); 3.4. Interpretability analysis: ◦ Side function visualization: The side function amplitude of the corneal layer (z=0~0.5mm) is 0.8~1.0, reflecting strong refraction; the side function fluctuation frequency of the retinal layer (z=22~24mm) is high, reflecting multilayer scattering; ◦ Parameter sensitivity test: Perturbing the lens thickness to 4.3mm (an increase of 7.5%), the depth of the reflection peak at the lens-vitre interface in the OCT image increases by 0.3mm, consistent with theoretical calculations; 3.5. Result output and storage: OCT B scan image (512×512 pixels), side function analysis graph, and parameter sensitivity report are generated and stored in the database.
[0057] 4. Pathological model expansion implementation: Taking the simulation of "macular edema" as an example, only the retinal parameters need to be modified: the retinal thickness is increased from 200μm to 300μm, and the scattering coefficient... from Increase to The model does not require retraining and directly infers and outputs OCT images, showing disordered interlaminar structures and blurred reflectance peaks in the retina. The SSIM value is 0.995 with the clinical pathological images, which verifies the model's pathological adaptability.
[0058] Compared with existing technologies, this invention has the following significant advantages: 1. Efficiency breakthrough: The 5D light field modeling parameters are only 1500, which is 1 / 10 of MLP-based PINNs; the inference time is 1.2ms, which is 1954 times faster than FDTD (2345ms) and 1563 times faster than MC (1876ms), meeting the real-time needs of clinical practice; 2. Improved accuracy: Hard coding of physical constraints ensures that the simulation error of interlayer reflection peaks is ≤3%, which is 67% faster than MLP-based PINNs (10%); the simulation SSIM of pathological models (such as macular edema) is ≥0.997 compared with clinical measurements; 3. Enhanced interpretability: The visualization of side functions realizes the "light field propagation - physical mechanism - physiological structure - pathological features" 4. Direct mapping with artifact tracing accuracy ≥90%, solving the mechanism analysis problem of traditional black box models; 5. Personalized adaptation: Supports rapid input of individual eye parameters, with generalization error ≤8%, a 47% reduction compared to existing technologies (15%), which can be used for personalized planning of refractive surgery; 6. Resource-friendly: Model size ≤10MB, supports deployment on edge devices (power consumption 15W), a 92.5% reduction compared to GPU solutions (200W), making it convenient for application in primary healthcare institutions.
[0059] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several changes and improvements without departing from the overall concept of the present invention, and these should also be considered within the scope of protection of the present invention.
Claims
1. A method for simulating the intraocular light field based on physiological constraints and pathological origins, characterized in that: The method includes the following steps: S1. Parametric modeling of the eyeball: A layered geometric model of the eyeball is established, which divides the eyeball into the cornea, anterior chamber, iris, lens, vitreous body, retina, Bruch's membrane, and choroid, and configures the geometric and optical parameters of each layer; wherein the geometric parameters include at least the radius of curvature, thickness, and spatial position, and the optical parameters include at least the wavelength-related refractive index distribution and scattering coefficient; S2, Physical Information Kolmogorov-Arnold Network Construction and Configuration: A three-layer Kolmogorov-Arnold network is constructed, with its input layer configured to receive a five-dimensional input vector, wherein the five-dimensional input vector consists of spatial coordinates (x, y, z) and light wavelength. λ It consists of time t; The network has two hidden layers, each containing 11 neurons, and the connections between neurons in adjacent layers are achieved through learnable side functions. The output layer of the network is configured to output the complex amplitude of the light field U(x,y,z,λ,t) corresponding to the input vector; S3, Physically Constrained Embedding and Network Training: Define a composite loss function L = aL PDE + bL boundary + gL data Train the network; where L PDE is the fitted residual of the Helmholtz equation followed by the propagation of the light field, the calculation of which relies on the output of the network throughout the computational domain; L boundary outputting for the network a fit residual of a theoretical Fresnel reflection or transmission condition at each tissue interface of the eye; α, β, γ are preset weight coefficients, and satisfy the weight ratio of α / β / γ as 5:3:
2. The training process allows physical laws to dominate the optimization of network parameters in the form of hard constraints. L data For high-quality OCT training data; S4. Solving for personalized light fields: The personalized parameters of the target eyeball to be simulated are input into the network trained in step S3. The network maps the five-dimensional input vector into the corresponding three-dimensional spatial light field distribution in a forward propagation manner. S5 and OCT image synthesis: Based on the light field distribution U obtained in step S4 sample Compare it with the preset reference arm plane wave light field U ref Perform interference superposition and calculate the interference signal I(λ,t)=∣U sample +U ref | 2 ; The interference signal I(λ,t) is subjected to a fast Fourier transform along the wavelength dimension to obtain a depth-resolved OCT A-scan signal A(z), which is then used to synthesize a B-scan or C-scan OCT image. S6. Interpretability Analysis and Output: Extract and visualize the edge functions in the network associated with specific anatomical layers. ; By programmatically fine-tuning a single geometric or optical parameter in the input parameters, the influence of that parameter on the features of the OCT image generated in step S5 is quantitatively analyzed, and a parameter sensitivity report is generated. The OCT image, side function visualization results, and parameter sensitivity report are integrated and output; The learnable edge function mentioned in step S2 adopts a cubic B-spline basis function with enhanced physical information, and its expression is: ; in, and These are learnable weight coefficients used to balance nonlinear activation and spline fitting; Let G be the B-spline basis function; G be the number of spline intervals, and k=3 be the spline order; furthermore, the side functions are specifically designed for different physical processes in light field propagation, including: The reflection edge function, designed based on Fresnel's formula, is used to encode the reflectivity of an interface. Its expression is as follows: ; The refraction side function, designed based on Snell's law, is used to encode light deflection. Its expression is as follows: ; Scattering side functions designed based on Mie scattering theory are used to encode tissue scattering effects.
2. The intraocular light field simulation method based on physiological constraints and pathological tracing according to claim 1, characterized in that: Step S3 of the network training process further includes an adaptive grid optimization strategy, specifically: During initial training, the number of spline intervals G is set to 3; during training, the loss function L is calculated and analyzed in real time. PDE The distribution within the spatial computational domain; when the residual in a specific region is detected to be continuously higher than a preset threshold, the number of spline intervals G in that region is automatically expanded to 5~7; after expansion, the least squares method is used only for the spline coefficients c at the newly added grid points. i Reinitialize the model and continue training on the expanded grid until the model converges.
3. The intraocular light field simulation method based on physiological constraints and pathological origins according to claim 1, characterized in that: The "quantitative analysis of its impact on OCT image features" mentioned in step S6 specifically includes: Establish the change in response from the input parameter perturbation ΔP via the side function. The mapping relationship between the image feature offset ΔF and the image feature F includes the depth position of the interlayer reflection peak, the signal amplitude, and the signal-to-noise ratio. By analyzing this mapping relationship, the artifacts appearing in OCT imaging can be traced back to their source physical processes or anatomical structures.
4. An intraocular light field simulation system for implementing the method according to any one of claims 1 to 3, characterized in that: The system includes the following modules interconnected via a data bus: Parameter input and interface module: used to receive individual eye parameter data from external corneal topography instrument and optical biometer, or to receive parameters manually input by the user through the graphical interface, and to transmit the formatted parameters to the storage module and model calculation module; Physical Information KAN Model Calculation Module: This is the core calculation unit, which contains a trained physical information Kolmogorov-Arnold network model. This module receives instructions and data from the parameter input module and performs light field calculations. OCT signal synthesis and processing module: directly connected to the model calculation module, receiving the light field data output by it, and performing interference calculation and frequency domain transformation to generate raw OCT image data; Interpretability Analysis Module: Connected to the model calculation module and the OCT signal processing module respectively, it has the authority to access the side function weight data inside the model calculation module, and calls OCT image data to perform side function visualization, parameter sensitivity calculation and artifact mapping analysis; Unified Output and Display Module: Connected to the OCT signal processing module and the interpretability analysis module, it integrates and displays the generated OCT image, dynamic side function curve, parameter sensitivity heatmap and analysis conclusions in a graphical user interface and provides interactive functions; Central storage and management module: Connected to all the above modules, it is used to store individual eye parameter databases, pre-trained network model parameter sets for different pathological states, historical simulation results, and customized analysis reports.
5. The system according to claim 4, characterized in that: The physical information KAN model calculation module is configured to support heterogeneous computing and can be loaded and run on a central processing unit, a graphics processing unit, or an embedded edge AI chip. When the system is deployed on an RK3588 smart hardware development board or an NVIDIA Jetson AGX Orin edge computing platform, the total time from parameter input to the generation and basic analysis of a single 512x512 pixel OCT B scan image does not exceed 1.2 milliseconds, and the overall power consumption is less than 15 watts.
Citation Information
Patent Citations
Methods and systems for creating virtual and augmented reality
US20160026253A1
See-through eyeglasses with light sources dispersed over proximal optic
WO2025240681A1
