CN-121839089-B - Cornea crosslinking postoperative prediction model construction method based on physical information neural network

CN121839089BCN 121839089 BCN121839089 BCN 121839089BCN-121839089-B

Abstract

A cornea cross-linking operation prediction model construction method based on a physical information neural network comprises the steps of establishing a superelastic constitutive model, characterizing key mechanical parameters of the superelastic constitutive model as a spatial distribution function related to cross-linking treatment parameters, constructing physical priori constraints, acquiring and preprocessing a preoperative cornea front surface height map and a full-thickness map, constructing a multi-mode input tensor and a truth value label through data enhancement, dividing the multi-mode input tensor and the truth value label into a training set and a test set, constructing a neural network integrating a coding-decoding structure and a circulating feedback mechanism, wherein the neural network comprises branches for adaptively predicting individual mechanical parameters, embedding the physical priori constraints into a training process in a differential loss function mode, and utilizing training data to jointly optimize the network to learn a mapping relation from preoperative morphology to postoperative morphology for predicting the cornea morphology of new data. The invention realizes the deep fusion of physical constraint and data driving, and remarkably improves the prediction precision, efficiency and physical credibility.

Inventors

MI SHENGLI
CHANG LONG
HE LINXIN
ZHAO YAO
Sun beichen
Yin Daiyu

Assignees

清华大学深圳国际研究生院

Dates

Publication Date: 20260508
Application Date: 20260313

Claims (10)

1. The method for constructing the cornea crosslinking post-operation prediction model based on the physical information neural network is characterized by comprising the following steps of: S1, establishing a superelastic constitutive model for describing cornea biomechanical behaviors, and characterizing key mechanical parameters in the superelastic constitutive model as a spatial distribution function related to crosslinking treatment parameters so as to construct physical priori constraints; S2, acquiring clinical data comprising a preoperative cornea front surface height map and a full-thickness map, preprocessing and enhancing the clinical data, constructing a multi-mode input tensor and a corresponding postoperative morphology truth value label, and dividing the multi-mode input tensor into a training set and a testing set; S3, constructing a cyclic physical information neural network prediction model which fuses a coding-decoding structure and a cyclic feedback mechanism and comprises branches for adaptively predicting individual mechanical parameters from an input image, and embedding the physical prior constraint into a network training process in the form of a differentiable loss function, wherein the cyclic physical information neural network prediction model constructed in the step S3 comprises the following steps: the double-channel encoder is used for extracting multi-scale space features from the input double-channel tensor; A cyclic feedback decoder which operates in a multi-time-step manner, receives the current cornea state at each time step and predicts a deformation increment, accumulates the increment to the current state and feeds back as an input to the next time step, and approximates the post-operative morphology by iterative successive approximation; the self-adaptive physical parameter prediction branch is connected with bottleneck characteristics output by the encoder, and automatically regresses a learnable proportionality coefficient representing the mechanical properties of the cornea of an individual through carrying out global compression and nonlinear mapping on the high-dimensional characteristics; The physical constraint loss calculation module predicts the proportion coefficient of branch output by utilizing the self-adaptive physical parameters, combines the rigidity distribution model in the step S1 to generate a rigidity field with spatial variation, couples the rigidity field with a deformation field predicted by a network, and calculates the physical regularization loss following the minimum potential energy principle; S4, training the cyclic physical information neural network prediction model by using training data, and enabling the model to learn a mapping relation from preoperative morphology and treatment parameters to postoperative morphology through joint optimization of image reconstruction loss, physical constraint loss and clinical parameter loss, so as to be used for predicting the novel preoperative cornea morphology by using the new preoperative data.
2. The method for constructing a prediction model after cornea crosslinking operation based on a physical information neural network as set forth in claim 1, wherein the step S1 specifically includes: selecting a superelastic constitutive model suitable for fiber reinforced biological tissues as a mechanical description basis of cornea; the total strain energy function of the constitutive model consists of superelastic energy representing nonlinear response of the matrix and the anisotropic fiber, viscoelastic energy representing time dependence and viscous energy representing rate-dependent dissipation; Establishing a mapping relation between the crosslinking treatment parameters and the key rigidity parameters in the constitutive model, so that the matrix rigidity and the fiber rigidity of the cornea after operation become rigidity functions which change along with the treatment light intensity, time and space position; and simplifying the rigidity function into an explicit rigidity field model which is distributed along the radial direction and tangential direction of the cornea respectively, and defining the proportionality coefficient in the model as a parameter which can be learned by a neural network, so that the physical rule of the constitutive model is converted into a loss calculation basis which can be embedded into network training.
3. The method for constructing a post-corneal crosslinking prediction model based on a physical information neural network as claimed in claim 1, wherein the step S2 specifically comprises: cutting, denoising and color-physical value decoding are carried out on the clinically collected corneal topography data before and after the pairing operation, so as to obtain a corneal anterior surface height matrix and a full-thickness matrix; based on the rotation invariance of cornea, carrying out rotation resampling on the height matrix and the thickness matrix to realize data enhancement so as to expand training samples and improve model robustness; stacking the elevation map of the front surface of the cornea before operation and the full thickness map of the cornea before operation along the channel dimension to form a dual-channel input tensor; And (3) strictly separating data of different cases into a training set and a testing set by adopting a case-based reserving method, so as to ensure clinical effectiveness of model generalization capability assessment.
4. The method for constructing a post-corneal crosslinking prediction model based on a physical information neural network as claimed in claim 1, wherein the specific method for performing color-physical value decoding on the corneal topography in step S2 comprises the following steps: Establishing a mapping relation table of color values of the terrain legend color bars and corresponding physical height values through interactive sampling; during decoding, calculating the distance between the color feature vector of each pixel in the topographic map and all sample color vectors in the mapping relation table; And selecting a physical value corresponding to the sample closest to the pixel as a decoding height value of the pixel.
5. The method for constructing a cornea crosslinking post-operation prediction model based on a physical information neural network according to claim 1, wherein the dual-channel encoder is composed of a plurality of cascaded convolution modules, each module comprises a convolution layer, a normalization layer and an activation function, and downsampling is performed through a pooling operation; And the cyclic feedback decoder restores the resolution of the feature map through up-sampling operation and is in jump connection with the features of the corresponding layer of the encoder.
6. The method for constructing a prediction model after cornea crosslinking operation based on a physical information neural network as claimed in claim 1, wherein the working mechanism of the adaptive physical parameter prediction branch comprises: Carrying out global average pooling on the high-dimensional feature map output by the encoder, and stripping the spatial position information of the high-dimensional feature map to obtain a global description vector representing the morphological features of the whole cornea; inputting the global description vector into a multi-layer perceptron comprising a full connection layer for nonlinear transformation; And applying a monotonically increasing activation function ensuring that an output value is positive to an output layer of the multi-layer perceptron, and finally outputting two learnable scale coefficients respectively corresponding to matrix hardness correction and fiber anisotropy correction, so that the network can adaptively infer individualized mechanical properties according to an input image.
7. The method for constructing a cornea crosslinking post-operation prediction model based on a physical information neural network as claimed in claim 1, wherein the specific process of the physical constraint loss calculation module comprises the following steps: constructing a normalized space coordinate grid according to the size of the input image, and calculating the radial distance and the cornea depth corresponding to each pixel point; Utilizing the self-adaptive physical parameters to predict the proportionality coefficient of branch output, and generating radial stiffness field tensor and tangential stiffness field tensor which are the same as the input image in size through a broadcasting mechanism by combining the radial stiffness distribution function and the tangential stiffness distribution function defined in the step S1; Correspondingly multiplying the deformation displacement field predicted by the network in the current time step by the rigidity field in a pixel level, and calculating the local elastic strain energy of each pixel point; The local strain energy of all pixel points is summed and averaged, and the local strain energy is used as a physical constraint loss term, wherein the deformation amplitude predicted by a constraint model in training is matched with the spatial rigidity distribution, namely, the deformation is restrained in a region with high rigidity, and larger deformation is allowed in a region with low rigidity.
8. The method for constructing a cornea crosslinking post-operation prediction model based on a physical information neural network as claimed in claim 1, wherein the model training in the step S4 adopts a composite loss function, and the composite loss function is a weighted sum of an image reconstruction loss term, a physical constraint loss term and a clinical parameter regression loss term; wherein, the image reconstruction loss term introduces dynamic weight based on deformation amplitude in calculation to focus on a significant deformation area caused by surgery; the clinical parameter regression loss term extracts the maximum curvature clinical index from the predicted height map through a differentiable operation module and minimizes the difference between the maximum curvature clinical index and the true value.
9. The method for constructing a post-corneal crosslinking prediction model based on a physical information neural network according to claim 1, wherein in step S4, the cyclic physical information neural network prediction model is trained and predicted by a multi-time-step recursive residual mechanism: inputting the preoperative cornea morphology as an initial state into a network; Predicting a first deformation increment according to the initial state in a first time step by the network, and adding the increment and the initial state to obtain a state after the first update; Taking the updated state as input, entering a second time step, predicting a second deformation increment by the network, and accumulating the updated state again; Repeating the above process until reaching the preset time step number, and outputting the finally accumulated state as the predicted postoperative cornea morphology.
10. A computer program product comprising a computer program, characterized in that the computer program when executed by a processor implements the method for constructing a post-corneal cross-linking prediction model based on a physical information neural network as claimed in any one of claims 1 to 9.

Description

Cornea crosslinking postoperative prediction model construction method based on physical information neural network Technical Field The invention relates to the crossing field of biomedical engineering and artificial intelligence, in particular to a method for constructing a cornea crosslinking postoperative prediction model based on a physical information neural network. Background Keratoplasty is a key method of treating keratolytic disorders (such as keratoconus) by enhancing cross-linking between collagen fibers of the cornea through photochemical reactions, and increasing the rigidity of the cornea to delay or prevent lesion progression. However, the influence of operation parameters (such as ultraviolet light irradiation intensity, time, spot shape and the like) on the cornea morphology after operation is still lack of an accurate theoretical prediction model, the formulation of the current clinical treatment scheme mainly depends on the experience of doctors, the evaluation of the postoperative effect is also limited to tracking observation, and personalized and quantitative preoperative prediction and curative effect optimization are difficult to realize. In recent years, studies have been attempted to predict post-operative morphology based on pre-operative corneal topography. One class of methods relies entirely on data-driven deep learning models, such as convolutional neural network-based image mapping relationships that directly fit preoperative post-operative topography. Although the method can capture a certain morphological statistical rule, the prediction result lacks of physical interpretability and is easily influenced by training data noise, so that morphological output which does not conform to the cornea biomechanics principle can be possibly generated, and reliable basis is difficult to provide in clinical decision. Another category of research focuses on computational mechanics methods based on finite element simulation, etc., by building a constitutive model of cornea and combining with a physicochemical equation of the crosslinking process, the mechanical response and morphological change of cornea after surgery are simulated. Cornea is a multi-layer collagen fiber structure, the mechanical behavior of which is complex, and the existing constitutive model mainly comprises superelastic model, viscoelastic model, superviscoelastic model and the like, and the table 1 is referred to. The HGO (Holzapfel-Gasser-Ogden) model is taken as a typical super-elastic constitutive model, can effectively describe the mechanical response of cornea anisotropy by separating the strain energy functions of matrix and fiber and combining with fiber dispersion parameters, and is suitable for simulating the directional enhancement effect of collagen fiber in the cornea crosslinking process. In contrast, the Maxwell, kelvin-like viscoelastic model can describe time-dependent stress relaxation or creep behavior, but is more suitable for slow loading or small deformation processes, and is difficult to describe nonlinear large deformation characteristics of cornea under intraocular pressure and crosslinking, while the Ogden-like super-viscoelastic model can describe large strain and time effect at the same time, but has numerous parameters, high calculation cost and weak physical interpretability. Therefore, the HGO model is often selected as the basis of cornea mechanics modeling due to the advantages of clear structure, definite physical meaning of parameters, convenience for coupling with crosslinking parameters and the like. However, although the method has a definite physical mechanism, the calculation cost is high, the modeling process is complex, the accurate acquisition of the individual cornea geometry and material parameters is highly dependent, and the rapid and efficient clinical real-time prediction and scheme optimization are difficult to realize. TABLE 1 classification and characterization of constitutive models commonly used for cornea The physical information neural network is used as an emerging modeling paradigm integrating a physical rule and data driving, and the physical consistency of the generalization capability and the prediction result of the model can be improved under the condition of data scarcity by embedding physical constraints such as a control equation, boundary conditions and the like into a loss function of the neural network. Successful application in the fields of fluid mechanics, solid mechanics and the like shows that PINN has the potential of solving the problems of complex nonlinearity and multi-physical field coupling. However, how to effectively embed the constitutive model describing the cornea anisotropic and superelastic mechanical behaviors into a deep learning framework and realize end-to-end, interpretable and efficient prediction from preoperative multi-modal images to postoperative morphologies is still a technical problem which is not solved in the current