Intraocular light field simulation method and system based on physiological constraint and pathological traceability

Abstract

The invention discloses an intraocular light field simulation method and system based on physiological constraint and pathology traceability, the system takes a Physics Informed Kolmogorov-Arnold Net (PI-KAN) network as a core, namely physical information KAN, the method comprises the following steps: obtaining personalized parameters of eyeballs; 5-dimensional light field parameters including space, wavelength and time are input into a pre-trained PI-KAN model for light field solving, a three-layer network architecture including an input layer, a hidden layer and an output layer is established, a physical information edge function is constructed, and training is performed through fusion of a Helmholtz equation and a loss function of boundary conditions; generating an OCT image based on the light field solved by simulation; by analyzing side function mapping, visualization and pathological traceability of a light field propagation physical mechanism are realized. According to the method, the sparsity and interpretability of PI-KAN are utilized, the problems that a traditional method is low in calculation efficiency, difficult in high-dimensional modeling, weak in physical constraint and poor in interpretability are solved, millisecond-level, high-precision and interpretable simulation of the eye light field is achieved, and the method is suitable for ophthalmic clinical auxiliary diagnosis, surgical planning and equipment optimization.

Description

TECHNICAL FIELD

[0001] The application belongs to the cross field of biomedical imaging and artificial intelligence machine learning, and particularly relates to an intraocular light field simulation method and system based on physiological constraints and pathological tracing, a new physical information Kolmogorov-Arnold network (PIKAN) model is designed, combined with optical coherence tomography (OCT) modal data for eye light field simulation, suitable for ophthalmic clinical diagnosis assistance, OCT device research and development optimization, ophthalmic surgery planning and medical education training scenes, and can realize high-precision simulation, high-efficiency training and physiological and pathological explainability analysis of high-dimensional light field propagation in the eyeball microenvironment. BACKGROUND

[0002] Current OCT eye light field simulation field mainly relies on three types of technical solutions, as follows: 1. Finite difference time domain method (FDTD): directly solving Maxwell's equation set, simulating electromagnetic field propagation through three-dimensional space discretization, needing to set a fine grid of λ / 20 (λ is the wavelength of light, about 1300 nm), and the typical calculation domain is 30mm×30mm×30mm, which needs millions of grid units; 2. Monte Carlo method (MC): simulating the propagation trajectory of photons in the eyeball tissue through random sampling, needing to simulate The above photons can ensure statistical accuracy, and a single simulation needs several hours to several days; 3. Physical information multilayer perceptron network (MLP-based PINNs): based on embedding physical constraints (such as Helmholtz equation) in a fully connected neural network, the typical architecture is [5, 50, 50, 1] (input dimension 5: space x / y / z, wavelength λ, time t), and the number of parameters is up to 15,000.

[0003] The above technologies have formed a routine application in ophthalmic OCT simulation, for example, FDTD is used for corneal refractive property simulation, MC is used for retinal scattering effect simulation, and MLP-based PINNs is used for light field solution of a simplified eyeball model.

[0004] The existing technology has the following key defects, which are difficult to meet the clinical and engineering needs: 1. Low computational efficiency: the computational complexity of FDTD is O ( ) (N is the number of dimensional sampling), the inference time of a single OCT B scan (512*512 pixels) is 2345ms; MC needs a large number of photon sampling, and the real-time performance cannot be met at all; the training time of MLP-based PINNs is 456s, and the inference time is 12.3ms, which still cannot adapt to the clinical real-time assistance scene; 2. The modeling ability of high dimension is insufficient: the OCT light field involves 5-dimensional parameters (x / y / z / lambda / t), FDTD and MC face the "dimensional curse", and the number of parameters increases exponentially; MLP-based PINNs have parameter explosion (5-dimensional modeling parameters exceed 15,000) in high-dimensional scenarios, and are prone to overfitting; 3. The embedding of physical constraints is loose: the existing technology adds the physical law (such as Fresnel reflection and Snell's law) as a "soft constraint" to the loss function, and does not deeply integrate with the model structure, resulting in that the simulation accuracy is greatly affected by the training data, for example, the error of MLP-based PINNs in simulating the inter-retinal reflection peak is more than 10%; 4. Poor interpretability: FDTD and MC are pure numerical black boxes, and MLP-based PINNs focus on nodes with activation functions, which cannot trace the physical mechanism of light field propagation, and it is difficult to locate the source of OCT imaging artifacts (such as lens refraction artifacts); 5. Weak individual adaptation: there are individual differences in physiological parameters of eyeballs (such as corneal curvature and lens thickness), and the existing technology needs to retrain or adjust a large number of parameters to adapt to different individuals, the generalization error is more than 15%, which cannot meet the individual diagnosis demand, and does not support effective on-site learning and dynamic updating. SUMMARY

[0005] The technical problem to be solved by the present application is to overcome the above-mentioned deficiencies and key defects of the prior art. In order to solve the above problems, an intraocular light field simulation method and system based on physiological constraints and pathological tracing are provided. The core model is a physical information Kolmogorov-Arnold network model, namely PI-KAN network.

[0006] The purpose of the present application is achieved in the following manner: an intraocular light field simulation method based on physiological constraints and pathological tracing, the method comprising the following steps:

[0007] S1, eyeball parameterized modeling: a layered geometric model of eyeball tissue is established, which divides the eyeball into cornea, anterior chamber, iris, lens, vitreous, retina, Bruch's membrane and choroid, and configures the geometric parameters and optical parameters of each layer; wherein the geometric parameters at least include the radius of curvature, thickness and spatial position, and the optical parameters at least include the wavelength-dependent refractive index distribution and scattering coefficient; S2, physical information Kolmogorov-Arnold network construction and configuration: a three-layer Kolmogorov-Arnold network is constructed, the input layer of the network is configured to receive a five-dimensional input vector, the five-dimensional input vector is composed of spatial coordinates , light wavelength λ and time t; each of the two hidden layers of the network contains 11 neurons, and the connection between the neurons of adjacent layers is realized by a learnable edge function; 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: define a composite loss function to train the network; wherein, is the fitting residual of the Helmholtz equation followed by the network output for light field propagation, which depends on the output of the network in the entire calculation domain; is the fitting residual of the theoretical Fresnel reflection or transmission condition at the interface of each tissue of the eyeball by the network output; α, β, γ are preset weight coefficients, and satisfy, the weight ratio of α / β / γ is 5:3:2, and the training process makes the physical law dominate the network parameter optimization in the form of hard constraint; is high-quality OCT training data; S4, personalized light field solving: input the personalized parameters of the target eyeball to be simulated into the network trained in step S3, and the network maps the five-dimensional input vector to the corresponding three-dimensional spatial light field distribution in a forward propagation manner; S5, OCT image synthesis: based on the light field distribution obtained in step S4 , interference superposition is carried out with the preset reference arm plane wave light field , and the interference signal is calculated; the interference signal is subjected to fast Fourier transform along the wavelength dimension to obtain the depth-resolved OCT A-scan signal , and then B-scan or C-scan OCT images are synthesized; S6, explainability analysis and output: the edge function associated with a specific anatomical layer (such as cornea, retina) in the network is extracted and visualized ; by programmatic fine-tuning of the single geometric or optical parameter in the input parameter, the influence of the 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 edge function visualization result and the parameter sensitivity report are integrated and output.

[0008] The learnable edge function in step S2 adopts a cubic B-spline basis function enhanced by physical information, and its expression is: wherein, and are learnable weight coefficients for balancing the nonlinear activation and the spline fitting; is a cubic B-spline basis function; G is the number of spline intervals, and k=3 is the order of spline; and the edge function is specially designed for different physical processes in light field propagation, including: a reflection edge function designed based on the Fresnel formula for encoding the interface reflectivity, and its expression is: a refraction edge function designed based on Snell's law for encoding the light deflection, and its expression is: and a scattering edge function designed based on the Mie scattering theory for encoding the tissue scattering effect.

[0009] The network training process of step S3 further includes an adaptive grid optimization strategy, specifically: during initial training, set the number of spline intervals G=3; during the training process, the loss function L PDE is calculated and analyzed in real time; when it is detected that the residual error of a specific region (such as the retinal layer) is continuously higher than a preset threshold, the number of spline intervals G of the region is automatically expanded to 5-7; after expansion, the least square method is used to re-initialize the spline coefficients on the newly added grid points, and the training is continued on the expanded grid until the model converges.

[0010] The “quantitative analysis of its influence on the OCT image features” in step S6 specifically includes: establishing a mapping relationship of “input parameter perturbation Edge function response change Image feature offset ”; the image features F include the depth position, signal amplitude and signal-to-noise ratio of the interlayer reflection peak; by analyzing the mapping relationship, the correlation traceability of the artifacts appearing in OCT imaging to the source physical process or anatomical structure is realized.

[0011] An intraocular light field simulation system for implementing the method, the system comprising the following modules interconnected by a data bus: a parameter input and interface module: for receiving individual eyeball parameter data from external corneal topographers, optical biometric instruments, or receiving parameters manually input by a user through a graphical interface, and transmitting the formatted parameters to a storage module and a model calculation module; a physical information KAN model calculation module: a core calculation unit, which internally solidifies a trained physical information Kolmogorov-Arnold network model; the module receives instructions and data from the parameter input module, and performs light field solving calculation; an OCT signal synthesis and processing module: directly connected with the model calculation module, receives the output light field data of the model calculation module, and performs interference calculation and frequency domain transformation to generate OCT image raw data; an explainability analysis module: connected with the model calculation module and the OCT signal processing module, respectively, has the authority to access the edge function weight data inside the model calculation module, and calls the OCT image data to perform edge function visualization, parameter sensitivity calculation and artifact mapping analysis; a unified output and display module: connected with the OCT signal processing module and the explainability analysis module, used for integrating and displaying the generated OCT image, dynamic edge function curve, parameter sensitivity heat map and analysis conclusion in a graphical user interface, and providing interactive functions; a central storage and management module: connected with all the above-mentioned modules, used for storing individual eyeball parameter database, pre-trained network model parameter set of 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 processor, a graphics processor or an embedded edge artificial intelligence chip; when the system is deployed based on an RK3588 intelligent hardware development board or an NVIDIA Jetson AGX Orin edge computing platform, the total time from parameter input to completion of generation of a single 512x512 pixel OCTB scan image and basic analysis is not more than 1.2 milliseconds, and the overall power consumption is less than 15 watts.

[0013] The present application has the following advantages: high calculation efficiency: the use of a low-parameter KAN network structure combined with adaptive grids shortens the inference time of five-dimensional light field simulation to milliseconds, meeting the clinical real-time interaction requirements.

[0014] High modeling accuracy: by "hard coding" the core physical laws into the network edge function and the loss function, the physical consistency of simulation training is significantly improved, and the simulation error of interlayer reflection peaks is significantly reduced.

[0015] Strong explainability: the explainability of the KAN network edge function is used to realize the visualization mapping of the key physical processes (such as refraction and scattering) of light field propagation, which can effectively trace the OCT imaging artifacts.

[0016] Good personalization adaptability: the system supports rapid input of individualized eyeball parameters, has strong model generalization capability, can efficiently adapt to different patients or pathological states, provides support for personalized medicine, and realizes on-site learning and dynamic updating.

[0017] Low resource demand: the model has small parameter quantity, can be deployed on edge computing devices, has low power consumption, and is convenient for application in primary medical institutions or portable devices. BRIEF DESCRIPTION OF DRAWINGS

[0018] Fig. 1 is the flow chart of the intraocular light field simulation system of the present application.

[0019] Fig. 2 is the principle architecture diagram of the physical information Kolmogorov-Arnold network (PI-KAN) algorithm of the present application.

[0020] Fig. 3 is the hardware composition diagram of the intraocular light field simulation system of the present application. DETAILED DESCRIPTION

[0021] The present application will be further described in detail below in combination with the drawings and specific embodiments.

[0022] It should be pointed out that the following detailed description is exemplary and is intended to provide further description of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same technical meaning as understood by ordinary skilled persons in the technical field to which the present application belongs.

[0023] As shown in Figs. 1 to 3 The present application provides an intraocular light field simulation method and system based on physiological constraints and pathological tracing, which models the intraocular light field. The eye has eight highly correlated anatomical interfaces, namely: cornea, anterior chamber, iris (pupil), lens, vitreous, retina, Bruch's membrane, and choroid. From the dimension of physical information law, the system mainly involves four optical laws, namely: reflection Fresnel formula, refraction Snell's law, scattering Mie theory, and optical Helmholtz equation. A new type of physical information Kolmogorov-Arnold network is designed, denoted as PI-KAN network, or physical information KAN network, as the core model architecture of the system. From the morphological dimension analysis, the eight anatomical interfaces in the eye are used as the backbone architecture of the PI-KAN network, and the specific reasons are as follows: the eight anatomical interfaces in the eye have clear boundaries and measurable thickness / curvature, which is beneficial to the construction of KAN node coordinates.

[0024] The connection between layers is only adjacent structures, which determines that the number of edge functions = O (number of layers) rather than O (n²), and the model parameter quantity is reduced by 1-2 orders of magnitude.

[0025] The boundary derivative can jump, and the B-spline multiple node satisfies the feature of "Fresnel reflection coefficient discontinuity" automatically.

[0026] The method comprises the following steps: S1, eyeball parameterized modeling: a layered geometric model of eyeball tissue is established, the layered geometric model divides the eyeball into cornea, anterior chamber, iris, lens, vitreous body, retina, Bruch's membrane and choroid, and geometric parameters and optical parameters of each layer are configured; wherein the geometric parameters at least include curvature radius, thickness and spatial position, and the optical parameters at least include wavelength-dependent refractive index distribution and scattering coefficient; S2, Kolmogorov-Arnold network construction and configuration: a three-layer Kolmogorov-Arnold network is constructed, an input layer of the network is configured to receive a five-dimensional input vector, the five-dimensional input vector is composed of spatial coordinates , light wavelength λ and time t; two hidden layers of the network each contain 11 neurons, and the connection between neurons of adjacent layers is realized through a learnable edge function; an output layer of the network is configured to output a complex amplitude of an optical field corresponding to the input vector ; S3, physical constraint embedding and network training: a composite loss function is defined, and the network is trained; wherein, is a fitting residual of the Helmholtz equation followed by the network output for light field propagation, and the calculation of the residual depends on the output of the network in the entire calculation domain; is a fitting residual of the network output for the theoretical Fresnel reflection or transmission condition at the interface of each tissue of the eyeball; α, β, γ are preset weight coefficients, and satisfy, the weight ratio of α / β / γ is 5:3:2, and the training process makes the physical law dominate the network parameter optimization in the form of a hard constraint; is high-quality OCT training data; S4, personalized light field solving: the personalized parameters of a target eyeball to be simulated are input into the network trained in step S3, the network maps the five-dimensional input vector to the corresponding three-dimensional spatial light field distribution in a forward propagation manner; S5, OCT image synthesis: based on the light field distribution obtained in step S4 , interference superposition is performed with a preset reference arm plane wave light field , and an interference signal is calculated; the interference signal is subjected to fast Fourier transform along the wavelength dimension to obtain a depth-resolved OCT A-scan signal , and then B-scan or C-scan OCT images are synthesized; S6, explainability analysis and output: the edge function associated with a specific anatomical layer (such as cornea, retina) in the network is extracted and visualized ; generating a parameter sensitivity report by perturbing a single geometry or optical parameter in the input parameters through a programmatic tuning, quantitatively analyzing its impact on the OCT image features generated in step (S5); and integrating the OCT image, the edge function visualization result and the parameter sensitivity report for output.

[0027] The learnable edge function in step S2 adopts a cubic B-spline basis function enhanced by physical information, and its expression is: wherein, and are learnable weight coefficients for balancing the nonlinear activation and the spline fitting; is a cubic B-spline basis function; G is the number of spline intervals, and k=3 is the order of spline; and the edge function is specially designed for different physical processes in the light field propagation, including: a reflection edge function designed based on the Fresnel formula, used for encoding the interface reflectivity, and its expression is: ; a refraction edge function designed based on Snell's law, used for encoding the light deflection, and its expression is: ; and a scattering edge function designed based on the Mie scattering theory, used for encoding the tissue scattering effect.

[0028] The network training process in step S3 further includes an adaptive grid optimization strategy, specifically: in the initial training, the number of spline intervals G is set to 3; in the training process, the loss function L PDE is calculated and analyzed in real time; when it is detected that the residual error of a specific region (such as the retinal layer) continuously exceeds a preset threshold, the number of spline intervals G of the region is automatically expanded to 5-7; after the expansion, the least square method is used to re-initialize the spline coefficients ci only on the newly added grid points, and the training is continued on the expanded grid until the model converges.

[0029] The "quantitative analysis of its impact on the OCT image features" in step S6 specifically includes: establishing a mapping relationship of "input parameter perturbation ΔP→ edge function response change Δϕ→ OCT image feature offset ΔF"; the image features F include the depth position, signal amplitude and signal-to-noise ratio of the interlayer reflection peak; by analyzing the mapping relationship, the correlation between the artifacts appearing in the OCT imaging and the source physical process or anatomical structure is traced.

[0030] An OCT eye light field simulation system for implementing the method, the system comprising the following modules interconnected through a data bus: a parameter input and interface module: for receiving individual eyeball parameter data from external corneal topography, optical biometry, or receiving parameters manually input by the user through the graphical interface, and transmitting the formatted parameters to the storage module and model calculation module; a physical information KAN model calculation module: a core calculation unit, which internally solidifies a trained physical information Kolmogorov-Arnold network model; the module receives instructions and data from the parameter input module, and performs light field solving calculation; an OCT signal synthesis and processing module: directly connected with the model calculation module, receives the light field data output by the model calculation module, and performs interference calculation and frequency domain transformation to generate OCT image raw data; an explainability analysis module: connected with the model calculation module and the OCT signal processing module, respectively, has the right to access the edge function weight data inside the model calculation module, and calls the OCT image data, performs edge function visualization, parameter sensitivity calculation and artifact mapping analysis; a unified output and display module: connected with the OCT signal processing module and the explainability analysis module, used for integrating and displaying the generated OCT image, dynamic edge function curve, parameter sensitivity heat map and analysis conclusion in a graphical user interface, and providing interactive functions; a central storage and management module: connected with all the above-mentioned modules, used for storing individual eyeball parameter database, pre-trained network model parameter set of 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 processor, a graphics processor or an embedded edge artificial intelligence chip; when the system is deployed based on the NVIDIA Jetson AGX Orin edge computing platform, the total time from parameter input to completion of generation and basic analysis of a single 512x512 pixel OCT B scan image is not more than 1.2 milliseconds, and the overall power consumption is less than 15 watts.

[0032] The total time of generation and analysis is 1.8~3.5 milliseconds, and the overall power consumption is 6~12 watts. Although the RK3588 platform has lower computing power than the Jetson AGX Orin, it still can realize "millisecond-level inference + low-power operation" by virtue of the lightweight model and heterogeneous computing optimization, and has the advantages of nationalization and low cost (the module price is only 15~20 US dollars), and is more suitable for deployment in primary medical institutions or portable OCT simulation devices.

[0033] The technical scheme of the application includes two parts of hardware system and software model, and the core is a layered modular architecture based on PIKANs, which is specifically as follows: in the first aspect, the system hardware composition of the application is specifically as follows: 1. parameter input module: including clinical parameter acquisition equipment (such as corneal topography, optical biological measurement instrument) and parameter interaction interface, which is used for inputting eyeball geometric parameters (corneal curvature R1 / R2, anterior chamber depth ACD, etc.), optical parameters (refractive index n of each layer, scattering coefficient , etc.); 2. calculation module: supporting CPU (Intel i7-10700K), GPU (NVIDIA RTX 3080) or edge computing chip (NVIDIA Jetson AGX Orin, domestic Ruikai micro RK3588), which is used for model training and inference; 3. result output module: including OCT image display unit, edge function visualization interface and parameter sensitivity analysis report generation unit; 4. storage module: used for storing eyeball parameter database (including normal and pathological samples), pre-trained PIKANs (Physics-Informed KANs) model weight and simulation results.

[0034] In the second aspect, the system software module of the application includes core algorithm model and module architecture, and the software model adopts four-layer modular design, and the functions and technical details of each layer are as follows: eyeball parameterization module geometric parameter modeling: a segmented spherical surface model is used to describe the layered structure of the eyeball, and 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, posterior surface curvature 6±1mm when static, minimum anterior surface curvature 5.3mm when adjusting, thickness 4±0.3mm; vitreous body: vitreous cavity length 16±0.5mm; retina: neural epithelial layer, 150±20μm; retinal pigment epithelial layer (RPE), 10±2μm; Bruch's membrane: normal thickness: 2±0.5μm; diameter: matched with the retina (about 22mm); choroid: thickness: 200±50 (thickest in the macular region, thinnest at the periphery); blood vessel density: 50%±10%.

[0035] Optical parameter modeling: wavelength dependence and spatial distribution characteristics are introduced, and the specific expression is as follows: corneal refractive index: =1.376-0.0012×(λ-589) (λ unit: nm); anterior chamber refractive index: =1.336-0.0005×(λ-589); lens gradient refractive index: nucleus =1.40, the cortex part linearly decreases to 1.38; iris: refractive index model: iris stroma layer: ; pupil zone (air): n = 1.0; scattering coefficient: iris stroma: ; absorption coefficient: iris stroma: .

[0036] lens gradient refractive index: nucleus = 1.40, cortex linearly decreases to 1.38; refractive index model: gradient distribution formula: ; scattering coefficient: age-dependent formula: ; absorption coefficient: normal state: ; age correlation: increase by 0.005 per 10 years.

[0037] vitreous: refractive index model: normal state: ; liquefaction: n = 1.336 (close to anterior chamber fluid); scattering coefficient: normal state: ; liquefaction: increases to 0.05~0.1; absorption coefficient: normal state: .

[0038] retina:

[0039]

[0040] Bruch's membrane: Bruch's membrane is a thin elastic membrane between RPE and choroid, with extremely thin thickness but key optical properties, parameters as follows:

[0041]

[0042] choroid: choroid is the vascular layer outside the retina, containing a large amount of melanin and blood vessels, with strong scattering and strong absorption as the main optical properties, parameters as follows:

[0043]

[0044] scattering coefficient: cornea , lens (increases by / 10 years with age), retina ; dynamic parameter modeling: considering physiological regulation function, pupil diameter (I is the light intensity, is the threshold light intensity), lens adjustment amplitude 15D.

[0045] 2, PIKANs light field solving module, this module is the core module of the system software architecture, this module is designed around the core component edge function of PI-KAN network, as follows: the edge function design of PI-KANs model adopts physical information enhanced cubic B-spline basis function, the expression is: wherein: is a weight coefficient, is a grid interval number (initial 3, adaptively extended to 7), k=3 (spline order), is a B-spline basis function; special edge functions are designed for different physical processes: reflection edge function: (encoding Fresnel reflection); refraction edge function: (encoding Snell's law); scattering edge function: based on Mie scattering theory, particle size and refractive index difference parameters are introduced; physical constraints are embedded: hard constraints are realized through residual loss function, and the total loss is: wherein: : Helmholtz equation residual, , is a domain point number); : boundary condition residual, is a boundary point number); : high-quality OCT training data, , , The weight ratio is 5:3:2; adaptive grid technology: initially, a coarse grid of G=3 is trained, and through error distribution analysis, it is extended to G=5~7 in the high gradient area (such as the retinal layer), and the least square method is used to re-initialize the extended grid coefficient, avoiding full network retraining.

[0046] 3, OCT signal generation module interference signal calculation: based on the superposition of sample arm and reference arm light field, the expression is: wherein (the reference light is a plane wave, is an amplitude, is an angular frequency); spectral domain processing: the fast Fourier transform (FFT) is performed on to obtain the depth resolution signal: , and c is the speed of light); system parameter modeling: the OCT hardware characteristics are introduced, including the light source center wavelength =1300nm, bandwidth =100nm (Gaussian spectrum), detector noise level≤ V, and the scanning mode supports A / B / C scanning switching.

[0047] The explanation and analysis display module, as an important functional module of the system described in the application, directly serves the analysis and diagnosis of ophthalmologists, provides analysis and display of anterior and posterior segment pathological processes based on the explainable logic of evidence-based medicine.

[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 attenuation, training iteration 2000 times, MSE is reduced to ; 3. adaptive grid optimization: analyze the error distribution of the pre-trained model, expand the grid in the high gradient area, retrain 500 times; 4. inference: input the parameters of the eyeball to be simulated, call the pre-trained model, output the OCT image and the explainability analysis result within 1.2ms.

[0054] Embodiment: The following takes "intraocular light field simulation of healthy adult" as an example to explain the implementation process and implementation mode of the present application in detail: 7.1 Implementation preparation 1. Hardware configuration: CPU is Intel i7-10700K, GPU is NVIDIA RTX3080, memory is 32GB; 2. Software environment: Python 3.9, PyTorch 2.0, FFT library is SciPy, visualization tool is Matplotlib; 3. Parameter input: corneal curvature R1=7.8mm, R2=6.5mm is collected by corneal topographer, anterior chamber depth 3mm, lens thickness 4mm, total eye axis length 24mm is collected by optical biometer; optical parameters are set as: cornea =1.376, anterior chamber =1.336, lens =1.40 (nucleus), vitreous =1.337; scattering coefficient : cornea , lens , retina . Among them, the retina, Bruch's membrane and choroid are the three core fundus physiological structures of the system.

[0055] 2, model training steps 2.1. Dataset generation: 10-20 groups of clinical eyeball parameters (including normal and pathological samples) are collected, and a 5-dimensional light field training dataset (sample size ) is generated based on the input parameters, 5-dimensional light field training data (x: 0-10mm, y: 0-10mm, z: 0-24mm, λ: 1250-1350nm, t: 0- s) is generated, sample size , divided into training set and validation set according to 8:2; 2.2. Network initialization: construct PIKANs architecture of [5,11,11,1], use cubic B-spline (G=3, k=3) as edge function, and use Glorot-like scheme for weight initialization; 2.3. Pre-training: set α=5, β=3, γ=2, Adam optimizer learning rate , train 2000 times, calculate the validation set MSE every 50 iterations, when MSE≤ When stopped, the training loss L = 8 x ; 2.4. Grid optimization: analyze the validation set error distribution, find that the retinal layer (z = 22~24mm) error is larger (MSE = 5 x ), expand the grid in this area to G = 7, retrain 500 times, and reduce the retinal layer MSE to 8 x ; 2.5. Model saving: save the pre-trained model (size 9.2MB), and record the optimal hyperparameters.

[0056] 3. Simulation implementation and result output 3.1. Reasoning trigger: input the parameters of the eyeball to be simulated, consistent with the training parameters, call the pre-trained model, and output the OCT image and explainability analysis results within 1.2ms; 3.2. Light field solving: the model completes the solving of the 5-dimensional light field within 1.2ms, and outputs the complex amplitude distribution; 3.3. OCT signal generation: calculate the interference signal , and get the depth image after FFT, with the foveal depth of 200μm, with an error of 1% compared with the clinically measured value (198±2μm); 3.4. Explainability analysis: o Edge function visualization: the edge function amplitude of the corneal layer (z = 0~0.5mm) is 0.8~1.0, reflecting strong refraction; the edge function frequency of the retinal layer (z = 22~24mm) is high, reflecting multi-layer scattering; o Parameter sensitivity test: perturb the lens thickness to 4.3mm (increase by 7.5%), and the depth of the lens-glass interface reflection peak in the OCT image increases by 0.3mm, consistent with the theoretical calculation; 3.5. Result output and storage: generate OCT B-scan images (512 x 512 pixels), edge function analysis graphs, and parameter sensitivity reports, and store them in the database.

[0057] 4. Pathological model expansion implementation Take "macular edema" simulation as an example, only need to modify the retinal parameters: increase the retinal thickness from 200μm to 300μm, and increase the scattering coefficient from to , the model does not need to be retrained, and directly outputs the OCT image, showing the disorder of the inter-retinal structure and the blur of the reflection peak, with an SSIM of 0.995 compared with the clinical pathological image, verifying the pathological adaptation ability of the model.

[0058] Compared with the prior art, the present application has the following remarkable beneficial effects: 1. Efficiency breakthrough: 5-dimensional 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 higher than FDTD (2345ms) and 1563 times higher than MC (1876ms), meeting the clinical real-time demand; 2. Precision improvement: physical constraint hard coding makes the simulation error of interlayer reflection peak ≤3%, which is 67% higher than MLP-based PINNs (10%); the simulation of pathological models (such as macular edema) and the clinically measured SSIM are ≥0.997; 3. Enhanced interpretability: edge function visualization realizes the direct mapping of "light field propagation - physical mechanism - physiological structure - pathological characteristics", and the accuracy of artifact tracing is ≥90%, solving the mechanism analysis problem of traditional black box models; 4. Individual adaptation: supports the rapid input of individual eyeball parameters, and the generalization error is ≤8%, which is 47% lower than the prior art (15%), and can be used for individual planning of refractive surgery; 5. Resource-friendly: the model size is ≤10MB, supporting edge device deployment (power consumption 15W), which is 92.5% lower than the GPU scheme (200W), and is convenient for application in primary medical institutions.

[0059] The above only describes the preferred embodiments of the present application, and it should be noted that for those skilled in the art, without departing from the overall concept of the present application, a number of changes and improvements can be made, which should also be considered as the protection scope of the present application.

Claims

1. An intraocular light field simulation method based on physiological constraints and pathological tracing, characterized in that: The method comprises the following steps: S1, parametric modeling of the eyeball: Establish a layered geometric model of the eyeball tissue, which divides the eyeball into cornea, anterior chamber, iris, lens, vitreous, retina, Bruch's membrane and choroid, and configures the geometric parameters and optical parameters of each layer; wherein the geometric parameters at least include the radius of curvature, thickness and spatial position, and the optical parameters at least include the wavelength-dependent refractive index distribution and scattering coefficient; S2, physical information Kolmogorov-Arnold network construction and configuration: a three-layer Kolmogorov-Arnold network is constructed, the input layer of the network is configured to receive a five-dimensional input vector composed of spatial coordinates (x, y, z), light wavelength λ and time t; each of the two hidden layers of the network contains 11 neurons, and the connection between adjacent layer neurons is realized by a learnable edge function; the output layer of the network is configured to output the light field complex amplitude U(x, y, z, λ, t) corresponding to the input vector; S3, physical constraint embedding and network training: define a composite loss function L = aL PDE + bL boundary + gL data Train the network; wherein, L PDE is the fitting residual of the network output to the Helmholtz equation followed by the propagation of the light field, the calculation of which depends on the output of the network throughout the entire calculation domain; L boundary is the fitting residual of the network output to the theoretical Fresnel reflection or transmission conditions at the interfaces of the various tissues of the eyeball; a, b, g are preset weight coefficients, and a / b / g weight ratio is 5:3:2, and the training process makes the physical law dominate the network parameter optimization in the form of hard constraint; L data For high-quality OCT training data; S4, personalized light field solving: input the personalized parameters of the target eyeball to be simulated into the network trained in step S3, and the network maps the five-dimensional input vector to the corresponding three-dimensional spatial light field distribution in a forward propagation manner; S5, OCT image synthesis: based on the light field distribution U obtained in step S4 sample , interference addition is performed with a preset reference arm plane wave light field U ref , and an interference signal I(λ, t) = |U sample + U ref | is calculated 2 ; a fast Fourier transform is performed on the interference signal I(λ, t) along the wavelength dimension to obtain a depth-resolved OCT A-scan signal A(z), and further to synthesize a B-scan or C-scan OCT image; S6, explainability analysis and output: extract and visualize the edge function ϕ(z) associated with a specific anatomical layer (such as cornea, retina) in the network; by programmatic fine-tuning of a single geometric or optical parameter in the input parameter, the influence of the parameter on the OCT image features generated in step S5 is quantitatively analyzed, and a parameter sensitivity report is generated; integrate and output the OCT image, edge function visualization result and parameter sensitivity report.

2. The method of claim 1, wherein the method is based on physiological constraints and pathologic provenance of intraocular light field simulation. The learnable edge function in step S2 adopts a physical information enhanced cubic B-spline basis function, and its expression is: ; wherein, and are learnable weight coefficients for balancing the non-linear activation and the spline fitting; are B-spline basis functions; G is the number of spline intervals, k = 3 is the order of spline; and, for different physical processes in light field propagation, the edge functions are specially designed, including: The reflection edge function is designed based on the Fresnel formula, which is used to encode the interface reflectivity, and its expression is: ； The refraction edge function is designed based on Snell's law, which is used to encode the light deflection, and its expression is: ； The scattering edge function is designed based on Mie scattering theory, which is used to encode the tissue scattering effect.

3. The method of claim 2, wherein the physiological constraints and pathologies are determined based on a patient's medical history. The network training process of step S3 further comprises an adaptive grid optimization strategy, specifically: during initial training, set the spline interval number G=3; during the training process, real-time calculation and analysis of the loss function L PDE Distribution in the spatial calculation domain; when it is detected that the residual of a specific area (such as the retinal layer) continuously exceeds a preset threshold, the spline interval number G of the area is automatically expanded to 5-7; after expansion, the least square method is used to only train the spline coefficients c i of the newly added grid points, and the model is reinitialized and continues to be trained on the expanded grid until the model converges.

4. The method of claim 1, wherein: The "quantitative analysis of its influence on OCT image features" in step S6 specifically includes: establishing a mapping relationship of "input parameter perturbation ΔP→edge function response change Δϕ→OCT image feature offset ΔF"; the image features F include the depth position, signal amplitude and signal-to-noise ratio of the interlayer reflection peak; by analyzing the mapping relationship, the correlation between the artifacts appearing in OCT imaging and the source physical process or anatomical structure is traced.

5. An intraocular light field simulation system for implementing the method of any one of claims 1 to 4, characterized by: The system comprises the following modules interconnected through a data bus: Parameter input and interface module: used for receiving individual eyeball parameter data from external corneal topography instrument and optical biometric instrument, or receiving parameters manually input by user through graphical interface, and transmitting the formatted parameters to storage module and model calculation module; Physical information KAN model calculation module: the core calculation unit, which is internally solidified with the trained physical information Kolmogorov-Arnold network model; the module receives instructions and data from the parameter input module, and performs light field solving calculation; OCT signal synthesis and processing module: directly connected with the model calculation module, receives the output light field data of the module, and performs interference calculation and frequency domain transformation to generate OCT image original data; Interpretability analysis module: connected with the model calculation module and OCT signal processing module, respectively, has the right to access the edge function weight data inside the model calculation module, and calls the OCT image data to perform edge function visualization, parameter sensitivity calculation and artifact mapping analysis; Unified output and display module: connected with the OCT signal processing module and interpretability analysis module, used for integrating and displaying the generated OCT image, dynamic edge function curve, parameter sensitivity heat map and analysis conclusion in a graphical user interface, and providing interactive functions; Central storage and management module: connected with all the above modules, used for storing individual eyeball parameter database, pre-trained network model parameter set of different pathological states, historical simulation results and customized analysis report.

6. The system of claim 5, wherein: The physical information KAN model calculation module is configured to support heterogeneous computing and can be loaded and run on a central processor, a graphics processor or an embedded edge artificial intelligence chip; when the system is deployed based on an RK3588 intelligent hardware development board or an NVIDIA Jetson AGX Orin edge computing platform, the total time from parameter input to completion of generation and basic analysis of a single 512x512 pixel OCT B scan image is not more than 1.2 milliseconds, and the overall power consumption is less than 15 watts.

Images