Sclera biomechanical evaluation and myopia risk prediction method based on artificial intelligence

Abstract

A sclera biomechanical evaluation and myopia risk prediction method based on artificial intelligence comprises the following steps: acquiring a plurality of groups of sample data through an OCTB scanning mode, and performing Gaussian smoothing denoising; based on an Otsu threshold segmentation algorithm, carrying out segmentation extraction on a fundus tissue ROI region; stacking and weighting the segmented two-dimensional images to build a three-dimensional alum model surface model; generating a displacement field u by using an image registration technology; calculating a strain tensor E based on the deformation gradient tensor F; calculating an effective strain belonging to eff based on the characteristic value lambda i of the strain; constructing a deformation field prediction model; a baseline image is obtained, a deformation field prediction model is used for prediction, and a fundus deformation field phi is obtained; quantifying the deformation percentage; constructing a compliance classification prediction model; and performing prediction based on the deformation field phi by using the compliance classification prediction model, and outputting a binary classification compliance prediction result, namely high compliance or low compliance. According to the method, accurate quantification and pathological risk prediction of the posterior sclera deformation of the eye can be realized.

Description

Technical Field

[0001] The present invention belongs to the field of medical image processing and computer vision technology, and specifically is a method for scleral biomechanics assessment and myopia risk prediction based on artificial intelligence. Background Art

[0002] Myopia is the most common refractive error and has become a major public health problem worldwide. Numerous studies have shown that the earlier myopia develops, the faster it progresses, and the higher the probability of developing high myopia or pathological myopia. Pathological myopia is often accompanied by a series of serious fundus complications, including retinal detachment, choroidal neovascularization, macular degeneration, etc., and is an important cause of low vision and blindness worldwide. Therefore, in clinical intervention and public health strategies, how to accurately assess the risk of myopia and identify potential pathological changes as early as possible is of great urgency.

[0003] The morphology and biomechanical properties of the posterior sclera play a key role in the progression of myopia. Traditional evaluation methods usually rely only on indirect observation of axial length or fundus images, making it difficult to quantitatively analyze the posterior sclera. Because the sclera is a hard connective tissue located deep in the back of the eyeball, there is a lack of non-invasive and convenient detection methods in clinical practice to directly measure its biomechanical properties. With the advancement of medical imaging technology, the use of multimodal imaging methods such as OCT and MRI to obtain three-dimensional data has made it possible to more accurately analyze the deformation and mechanical properties of the posterior sclera, but it still faces challenges such as insufficient reconstruction accuracy, noise interference, and complex multimodal registration.

[0004] In recent years, artificial intelligence (AI) technology has achieved significant breakthroughs in medical image analysis, demonstrating superior performance in multi-scale and multi-dimensional feature extraction and learning. Therefore, there is an urgent need for an AI-based scleral biomechanical assessment and myopia risk prediction method to accurately quantify posterior scleral deformation and accurately predict pathological risk. Summary of the Invention

[0005] In response to the problems existing in the above-mentioned prior art, the present invention provides an artificial intelligence-based scleral biomechanical assessment and myopia risk prediction method. Based on conventional imaging, this method can achieve accurate quantification of scleral deformation at the back of the eye and pathological risk prediction through three-dimensional reconstruction, fusion of tensor analysis and deep learning algorithms. It can also achieve accurate assessment of the degree of fundus deformation and accurate prediction of myopia progression risk in a non-invasive and convenient manner.

[0006] To achieve the above objectives, the present invention provides an artificial intelligence-based method for scleral biomechanical assessment and myopia risk prediction, comprising the following steps:

[0007] Step 1: Data collection and preprocessing;

[0008] S11: Select several patients as volunteers and collect eye images through OCTB scanning. Two types of images are collected for each eye of each patient: one type of image is a baseline image, and the other type is a mechanical loading image. Several sets of sample data are obtained through the massive collection process;

[0009] S12: Data preprocessing: Perform Gaussian smoothing and denoising on the image I in several groups of sample data to obtain the pixel value I of the smoothed image at the coordinate (x, y) smoothed (x,y);

[0010] S13: Segment and extract the fundus tissue ROI area based on the Otsu threshold segmentation algorithm;

[0011] S14: reconstructing the three-dimensional sclera surface model by stacking the segmented two-dimensional image layers, thereby obtaining the three-dimensional sclera surface models of several volunteers;

[0012] Step 2: Use image registration technology combined with mechanical algorithms to calculate the effective strain of fundus tissue;

[0013] S21: Calculation of displacement field; Generate displacement field u using image registration technology;

[0014] S22: Calculation of strain tensor; Calculate the strain tensor E based on the deformation gradient tensor F;

[0015] S23: Effective strain calculation; calculate the effective strain ∈ based on the strain eigenvalue λi eff , and then obtain the effective strain sample data of several volunteers;

[0016] Step 3: Use the deformation field prediction network to extract the fundus deformation field and quantify the deformation degree;

[0017] S31: Construct a deformation field prediction model; use an encoder-decoder structure, combine a three-dimensional SwinTransformer with a convolutional neural network to construct a deformation field prediction model, and use valid strain sample data from several volunteers for training to obtain a trained deformation field prediction model;

[0018] S32: For new patients, a baseline image is obtained through OCTB scanning, followed by data preprocessing, segmentation and extraction of the fundus tissue ROI area, reconstruction of the three-dimensional scleral mold surface model, and calculation of the effective strain. The effective strain is then input as input data into the deformation field prediction model for prediction, and the fundus deformation field φ is output;

[0019] S33: Deformation percentage quantification; the deformation percentage is calculated based on the ratio of the average deformation value to the reference value;

[0020] Step 4: Compliance classification prediction;

[0021] S41: using a deformation field prediction model to predict effective strain sample data of a number of volunteers to obtain deformation fields of the number of volunteers, then quantifying the deformation percentages of the deformation fields of the number of volunteers to obtain deformation percentage data of the number of volunteers, and then classifying the deformation percentage data of each volunteer as low compliance or high compliance based on a set threshold, thereby obtaining compliance sample data sets of the number of volunteers;

[0022] S42: Constructing a compliance classification prediction model; constructing a compliance classification prediction model based on 3D ResNet18, and training the compliance classification prediction model using compliance sample datasets of several volunteers to obtain a trained compliance classification prediction model;

[0023] S43: The deformation percentage of the new patient is input as input data to the compliance classification prediction model for prediction, and a binary compliance prediction result is output, i.e., high compliance or low compliance.

[0024] As a preferred embodiment, in step 1 S12, the pixel value I of the smoothed image at the coordinate (x, y) is obtained according to formula (1): smoothed (x,y);

[0025]

[0026] Where I(x-u0,y-v0) is the pixel value of the original image I at the coordinate (x-u0,y-v0); G(u0,v0) is the weight value of the Gaussian kernel at the offset (u0,v0), u0 is the horizontal offset from the kernel center to the current pixel, v0 is the vertical offset from the kernel center to the current pixel, k is the radius of the Gaussian kernel, σ is the standard deviation of the Gaussian kernel.

[0027] As a preferred embodiment, in step 1 S13, the process of extracting the fundus tissue ROI region is as follows:

[0028] S13-1: Calculate the grayscale histogram of the entire image and the occurrence probability p of each grayscale level i according to formula (2) i ;

[0029]

[0030] Where, the number of pixels n at each gray level i is i ; N is the total number of pixels in the image;

[0031] S13-2: Obtain the proportion ω(t) of pixels in the image whose grayscale value is less than or equal to t according to formula (3), and obtain the average grayscale value μ(t) of these pixels according to formula (4);

[0032]

[0033] S13-3: According to formula (5), the inter-class variance between the background and foreground is maximized And traverse all possible thresholds t, select the t that makes σ the largest as the optimal threshold to binarize the image into foreground and background, and then perform segmentation and extraction of the fundus tissue ROI area based on the optimal threshold t;

[0034]

[0035] As a preferred embodiment, in step 1 S14, the process of performing three-dimensional reconstruction of the fundus image is as follows:

[0036] S14-1: Superimpose the segmented 2D image layers, align the 3D primitives, and then use 3D rendering technology to generate a continuous 3D surface;

[0037] S14-2: Surface extraction technology based on the MarchingCubes algorithm is used to extract continuous surfaces from three-dimensional voxel data to obtain a continuous and clear three-dimensional scleral mold surface model.

[0038] As a preference, in step 2 S21, the generation process of the displacement field u is as follows:

[0039] S21-1: Perform image registration using the baseline image as the fixed image for image registration and the mechanical loading image as the moving image for image registration. During the image registration process, the transformation parameters of the moving image are adjusted to maximize the MMI metric I(X; Y), and the transformation parameters that maximize I(X; Y) are used as the final transformation parameters. The MMI metric I(X; Y) is calculated using formula (6).

[0040]

[0041] Where X is the baseline image, Y is the mechanical loading image, p(x,y) is the joint probability density function of X and Y, which represents the probability of observing x and y at the same time, p(x) and p(y) are the marginal probability density functions, which represent the probability of observing x and y, respectively.

[0042] S21-2: Apply the final transformation parameters to all pixels of the moving image, calculate the new positions of these pixels relative to the fixed image and generate a displacement field u, where the displacement vector D(x, y, z) of each pixel point (x, y, z) in the three-dimensional space is calculated according to formula (7);

[0043] D(x,y,z)=T(x,y,z)-O(x,y,z) (7);

[0044] Where T(x,y,z) is the new coordinates of the point (x,y,z) after the transformation function is applied to it, and O(x,y,z) represents the original coordinates of the point (x,y,z) in the moving image.

[0045] As a preference, in step 2 S22, the calculation process of the strain tensor E is as follows:

[0046] S22-1: Calculate the deformation gradient tensor F according to formula (8);

[0047]

[0048] Where, is the spatial gradient of the displacement field u; I is the identity matrix;

[0049] S22-2: Calculate the strain tensor E according to formula (9);

[0050]

[0051] Where, F T is the transpose of the deformation gradient tensor F.

[0052] As a preferred embodiment, in step 2 S23, the effective strain ∈ eff The calculation process is as follows:

[0053] S23-1: First, solve formula (10) to obtain the eigenvalue λi of the strain tensor symmetric to the principal strain, and then obtain the principal strain;

[0054] det(E-λi)=0 (10);

[0055] S23-2: Comprehensively consider all principal strain contributions in three-dimensional space and calculate the effective strain ∈ according to formula (11) eff ;

[0056]

[0057] Where,∈ i are the components of the principal strains.

[0058] As a preferred embodiment, in step 3 S31, the specific process of constructing the deformation field prediction network is as follows:

[0059] S31-1: Build encoder;

[0060] First, move the image I m and fixed image I f The input tensor is concatenated into 2×H×W×D along the channel dimension, and the input is split into non-overlapping 3D patches, each of which is 2×P×P×P in size, where P is set to 4 and the total number of patches is And flatten each patch into a vector x i , where i∈{1,…,N}; each patch is mapped to the feature dimension C through a linear projection layer to obtain the initial token sequence z0, as shown in formula (12), where the dimension of z0 is N×C;

[0061] z0=[x 1p E;x 2p E;…;x Np E] (11);

[0062] Where E represents linear projection,

[0063] Second, we add the position embedding E in the linear projection layer. pos , to integrate the location information into the token and get z0+E pos ;

[0064] Then, after the linear projection layer, four stages of Swin Transformer blocks and three stages of patch merging are applied, and the patch merging of each stage is followed by the Swin Transformer block of the previous stage; for the Swin Transformer block, each stage contains multiple Swin Transformer blocks, and an alternating window partitioning strategy is adopted, and rectangular parallelepiped windows are used to adapt to non-square images, and the size of each window is M x ×M y ×M z At each resolution, the first Swin Transformer block uses the conventional window partitioning method, starting from the voxel in the upper left corner, to evenly divide the feature map into M x ×M y ×M z non-overlapping windows of , and calculate self-attention inside each window, as shown in formula (12); the subsequent Swin Transformer block is moved from the configuration in the previous Swin Transformer block by offsetting the window of the previous Swin Transformer block by voxels to achieve this. At the same time, window-based attention is used to calculate two consecutive Swin Transformer blocks according to formulas (13), (14) and (15). For patch merging, each stage reduces the spatial resolution by 8 times by merging 2×2×2 adjacent tokens, and increases the feature dimension to 2C. The output of the encoder is

[0065]

[0066] Where, are query, key, and value matrices respectively, d represents the dimension of query and key features, M x M y M z is the number of tokens in the 3D window, B represents the relative position of the token in each window; W-MSA and SW-MSA represent the multi-head self-attention module based on the conventional window and the multi-head self-attention module based on the moving window, respectively; MLP represents the multi-layer perceptron module; LN represents layer normalization; and z l They represent the output features of W-MSA and SW-MSA of the l-th Swin Transformer block respectively;

[0067] S31-2: Construct a decoder; use multiple consecutive upsampling, convolutional layers and skip connections to form a decoder, where each upsampling is followed by two consecutive convolutional layers, and the final stage is followed by an additional convolutional layer. Except for the last convolutional layer, each convolutional layer is followed by a Leaky Rectified Linear Unit activation, with a convolution kernel size of 3×3, and each upsampled feature map in the decoding stage is connected to the corresponding feature map in the encoding path through a skip connection, and then processed by two consecutive convolutional layers to capture local information while generating a high-resolution feature map to generate a deformation field φ; finally, a spatial transformation function is used to apply nonlinear deformation to the moving image I m and deformation field φ;

[0068] S31-3: Constructing the loss function of the deformation prediction network

[0069]

[0070] Where, Represents the image fidelity metric, and the similarity metric used is the mean square error. Where p represents the voxel position and Ω represents the image domain; represents the regularization of the deformation field, Where u(p) is the spatial gradient of the displacement field u, which is expressed by Perform calculations.

[0071] As a preferred embodiment, in step 3 S33, the process of calculating the deformation percentage is as follows:

[0072] S33-1: Calculate the Euclidean norm of each point of the deformation field u according to formula (17);

[0073]

[0074] Where u x ,u y ,u z are the displacements of each vector u at position x, y, and z in three-dimensional space;

[0075] S33-2: Calculate the average shape of all voxel shapes according to formula (18);

[0076]

[0077] Where N is the total number of valid voxels, |u i | is the shape of the i-th voxel;

[0078] S33-3: Calculate the diagonal length d of the maximum size of the image in the three-dimensional space according to formula (19);

[0079]

[0080] S33-4: Calculate the deformation percentage according to formula (20);

[0081]

[0082] As a preferred embodiment, in step 4, the process of constructing the compliance classification prediction model is as follows:

[0083] S42-1: Apply the 3D ResNet18 structure to 3D data to form a 3D residual network; 3D ResNet18 is composed of residual basic blocks, which consist of two convolutional layers, each followed by a batch normalization and a ReLU, with the top of the block and the layer before the last ReLU in the block connected via a residual connection;

[0084] S42-2: Calculate the cross entropy loss of the three-dimensional residual network according to formula (121);

[0085]

[0086] Where N is the total number of categories, y i Represents the probability distribution of the actual category. If the sample belongs to the i-th category, then y i =1, otherwise 0. represents the probability of the i-th category predicted by the model.

[0087] The present invention provides a method for scleral biomechanical assessment and myopia risk prediction based on medical imaging and artificial intelligence. First, the original image is preprocessed using a Gaussian denoising algorithm, which can effectively reduce noise interference and enhance the accuracy of subsequent analysis, thereby improving the clarity of the original image and ensuring that the three-dimensional morphology of the posterior scleral structure formed by subsequent reconstruction is clearer. Then, the segmentation and extraction of the fundus tissue ROI area is performed based on the Otsu threshold segmentation algorithm, which can effectively reduce the subsequent computational complexity and, at the same time, enhance tissue contrast, providing a good foundation for subsequent lesion detection or analysis. The MarchingCubes surface extraction technology is used to extract the three-dimensional surface of multimodal or sequential fundus images to obtain a continuous and clear three-dimensional scleral surface model. This can ensure the integrity of the image structure while minimizing the impact of imaging differences on subsequent deformation analysis. At the same time, the eye tissue structure can be intuitively displayed. Based on the obtained three-dimensional scleral model, a three-dimensional deformation calculation framework that integrates image registration and tensor analysis is designed and adopted, which can obtain important biomechanical indicators such as effective strain. By precisely registering image data at different time points or viewing angles, deformation information at corresponding locations is extracted, and key biomechanical parameters such as effective strain are calculated using tensor analysis methods. This process effectively reveals the local and global mechanical changes in deep scleral tissue under physiological or pathological conditions, providing quantitative indicators for subsequent risk prediction and ensuring subsequent prediction accuracy. To address the resolution differences and noise issues in the three-dimensional deformation field of the posterior sclera structure, the present invention constructs a deformation field prediction model based on a three-dimensional Swin Transformer and a convolutional neural network, and directly performs smooth regression and prediction on the deformation field. Leveraging the global correlation modeling advantages of this deep neural network, it effectively suppresses noise and captures subtle deformations, enabling accurate prediction of posterior scleral deformation. Furthermore, the present invention innovatively proposes a new quantitative metric—"deformation percentage"—for scleral deformation analysis and risk assessment, which comprehensively measures the degree of scleral deformation under external forces or physiological changes, providing a more intuitive and accurate measurement method for myopia risk assessment. Finally, different levels of eye compliance are divided according to the biomechanical properties of the sclera as reflected by the deformation percentage. The present invention further utilizes a three-dimensional residual network (ResNet) to construct a compliance classification prediction model to classify and learn the grading results, and accurately predict the degree of myopia progression and pathological risks that individuals may face. By combining the deformation percentage with other anatomical indicators (such as axial length, retinal thickness, etc.), the model can provide high-risk warnings for early potential pathological myopia and provide comprehensive data support for clinical decision-making.

[0088] The present invention utilizes a technical approach combining three-dimensional reconstruction, deformation field analysis, and deep learning to achieve quantitative evaluation of the biomechanical properties of the posterior sclera and accurate prediction of the risk of myopia progression based on conventional imaging. On the one hand, the present invention can effectively solve the technical pain point of the current lack of intuitive quantitative indicators for the biomechanical measurement of the posterior sclera. On the other hand, it can provide a scientific basis for the early clinical identification of pathological myopia and its related complications, thereby enabling timely prevention and control of high myopia and fundus lesions at the public health level. The present invention can provide a new technical approach for the early screening and diagnosis of ophthalmic diseases, and help promote the widespread application of conventional imaging and artificial intelligence in the detection of myopia and other fundus diseases. At the same time, through innovations in non-invasive detection and quantitative analysis, it is also beneficial to open up broader prospects for subsequent medical research combining multimodal imaging with deep learning, and make greater contributions to clinical practice and public health.

[0089] Compared with the prior art, the present invention has the following technical advantages:

[0090] 1. Non-invasive quantitative assessment: By acquiring multimodal posterior ocular images using conventional imaging equipment, a three-dimensional model can be constructed and the biomechanical properties of the sclera can be quantified, avoiding the burden of invasive testing methods on patients.

[0091] 2. Accurate deformation analysis: By combining image registration, tensor analysis, and the 3D Swin Transformer network, the ability to identify subtle local deformations of the posterior sclera is significantly improved, enabling early detection of signs of pathological myopia progression.

[0092] 3. New quantitative indicator: "Deformation percentage" as a new mechanical quantitative indicator can achieve good robustness and comparability under different imaging conditions, providing an effective means to further subdivide and evaluate the myopia risk of different individuals.

[0093] 4. Deep learning prediction model: Uses a three-dimensional residual network to classify ocular compliance levels and directly provide clinical risk warnings. It is highly automated and accurate, facilitating early intervention and the development of personalized treatment plans.

[0094] Based on conventional imaging, this method can achieve accurate quantification of scleral deformation at the back of the eye and prediction of pathological risks through three-dimensional reconstruction, fusion of tensor analysis and deep learning algorithms. It can non-invasively and conveniently achieve accurate assessment of the degree of fundus deformation and accurate prediction of the risk of myopia progression. It aims to provide reliable technical support for the early identification and intervention of myopia, and has important application value in the field of medical image analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 It is the denoising effect diagram of the present invention;

[0096] Figure 2 It is the segmentation effect diagram of the present invention;

[0097] Figure 3 is a reconstruction rendering of the present invention;

[0098] Figure 4 is a flow chart of the present invention;

[0099] Figure 5 It is a structural diagram of the deformation field prediction model in the present invention;

[0100] Figure 6 It is a visualization diagram of the deformation effect in the present invention;

[0101] Figure 7 It is a structural diagram of the residual block in the present invention. DETAILED DESCRIPTION

[0102] In order to more clearly illustrate the objectives, technical solutions and advantages of the present invention, the present invention will be further described comprehensively and clearly with reference to the accompanying drawings and implementation steps.

[0103] like Figures 1 to 7 As shown, the present invention provides a method for scleral biomechanics assessment and myopia risk prediction based on artificial intelligence, comprising the following steps:

[0104] Step 1: Data collection and preprocessing;

[0105] S11: Several patients were selected as volunteers and eye images were collected using OCTB scanning. To more comprehensively understand the morphological changes of the eyeball under different physiological states, two types of images were collected for each eye of each patient: one type of image was a baseline image (Baseline) and the other type was a mechanical loading image (Loading). Several sets of sample data were obtained through the massive collection process;

[0106] As a preference, 100 patients were selected as volunteers; when collecting baseline images for each patient, a routine OCTB scan was performed under the condition that the patient was naturally relaxed and the intraocular pressure was normal, and 920 cross-sectional two-dimensional images were obtained, covering the complete area from the front to the back of the retina, and the size of each cross-sectional image was 512×512, which could capture the details of the eye tissue at a higher resolution, providing a complete and detailed imaging basis for subsequent three-dimensional reconstruction and mechanical deformation analysis, and saved as a set of baseline image sample data; when collecting mechanical loading images for each patient, on the premise of ensuring the safety and comfort of the patient, , using mild external pressure or controlling the appropriate increase in intraocular pressure (IOP), a certain mechanical load is applied to the patient's eyeball before a conventional OCTB scan is performed. This loading process will cause a certain degree of elastic deformation of the posterior sclera, providing more differentiated data for subsequent mechanical analysis, and obtaining 920 cross-sectional two-dimensional images, covering the complete area from the front to the back of the retina, and the size of each cross-sectional image is 512×512, which can capture the details of the eye tissue at a higher resolution, providing a complete and detailed imaging basis for subsequent three-dimensional reconstruction and mechanical deformation analysis, and saved as a set of mechanical loading image sample data. Preferably, the above images are all stored in DICOM format.

[0107] S12: Data preprocessing: Perform Gaussian smoothing and denoising on the image I in several groups of sample data to obtain the pixel value I of the smoothed image at the coordinate (x, y) smoothed (x,y);

[0108] S13: Segment and extract the fundus tissue ROI area of ​​the image data based on the Otsu threshold segmentation algorithm; the core of this method is to calculate the threshold by statistically analyzing the grayscale distribution of the entire image.

[0109] S14: reconstructing the three-dimensional sclera surface model by stacking the segmented two-dimensional image layers, thereby obtaining the three-dimensional sclera surface models of several volunteers;

[0110] Step 2: Use image registration technology combined with mechanical algorithms to calculate the effective strain of fundus tissue;

[0111] S21: Calculation of displacement field; Generate displacement field u using image registration technology;

[0112] S22: Calculation of strain tensor; Calculate the strain tensor E based on the deformation gradient tensor F;

[0113] S23: Effective strain calculation; calculate the effective strain ∈ based on the strain eigenvalue λi eff , and then obtain the effective strain sample data of several volunteers;

[0114] Step 3: Use the deformation field prediction network to extract the fundus deformation field and quantify the deformation degree;

[0115] S31: Construct a deformation field prediction model; use an encoder-decoder structure, combine a three-dimensional SwinTransformer with a convolutional neural network to construct a deformation field prediction model, and use valid strain sample data from several volunteers for training to obtain a trained deformation field prediction model;

[0116] S32: For new patients, a baseline image is obtained through OCTB scanning, followed by data preprocessing, segmentation and extraction of the fundus tissue ROI area, reconstruction of the three-dimensional scleral mold surface model, and calculation of the effective strain. The effective strain is then input as input data into the deformation field prediction model for prediction, and the fundus deformation field φ is output;

[0117] S33: Deformation Percentage Quantification; Deformation percentage is calculated based on the ratio of the average deformation value to the reference value, converted into a percentage; Deformation percentage provides a way to quantify the strength of the deformation field in the deformation of the object. This metric is very important for evaluating algorithm performance and ensuring the physical feasibility of the deformation field. In medical imaging, such as MRI or CT scan registration, deformation percentage can help evaluate the ability of the registration algorithm to accurately align an image at one time point to an image at another time point. A lower deformation percentage may indicate a smaller registration error, while a higher deformation percentage may indicate a more significant spatial change or registration error.

[0118] Step 4: Compliance classification prediction;

[0119] S41: using a deformation field prediction model to predict effective strain sample data of a number of volunteers to obtain deformation fields of the number of volunteers, then quantifying the deformation percentages of the deformation fields of the number of volunteers to obtain deformation percentage data of the number of volunteers, and then classifying the deformation percentage data of each volunteer as low compliance or high compliance based on a set threshold, thereby obtaining compliance sample data sets of the number of volunteers;

[0120] S42: Construct a compliance classification prediction model; construct a compliance classification prediction model based on 3D ResNet18, and use the compliance sample data set of several volunteers to train the compliance classification prediction model to obtain a trained compliance classification prediction model; in this way, by constructing a compliance classification prediction model through a three-dimensional residual network, only the baseline image can be used as the network input, and the output is a binary classification task, namely the level of compliance.

[0121] S43: The deformation percentage of the new patient is input as input data to the compliance classification prediction model for prediction, and a binary compliance prediction result is output, i.e., high compliance or low compliance.

[0122] As a preferred embodiment, in step 1 S12, the pixel value I of the smoothed image at the coordinate (x, y) is obtained according to formula (1): smoothed (x,y);

[0123]

[0124] Where I(x-u0,y-v0) is the pixel value of the original image I at the coordinate (x-u0,y-v0); G(u0,v0) is the weight value of the Gaussian kernel at the offset (u0,v0), u0 is the horizontal offset from the kernel center to the current pixel, v0 is the vertical offset from the kernel center to the current pixel, k is the radius of the Gaussian kernel, and k usually depends on the value of σ, ensuring that the convolution kernel covers enough area to contain enough pixel information. σ is the standard deviation of the Gaussian kernel, which controls the degree of smoothing. The recursive Gaussian filter implements Gaussian filtering in an efficient computational manner. Its core advantage is that the computational complexity is independent of the filter size (determined by σ), which allows it to maintain high computational efficiency even for large values ​​of σ.

[0125] As a preferred embodiment, in step 1 S13, the process of extracting the fundus tissue ROI region is as follows:

[0126] S13-1: Assume that the image has L gray levels. Calculate the gray histogram of the entire image and the probability p of each gray level i according to formula (2): i ;

[0127]

[0128] Where, the number of pixels n at each gray level i is i ; N is the total number of pixels in the image;

[0129] S13-2: Obtain the proportion ω(t) of pixels in the image whose grayscale value is less than or equal to t according to formula (3), and obtain the average grayscale value μ(t) of these pixels according to formula (4);

[0130]

[0131] S13-3: To find the optimal threshold, the Otsu algorithm considers maximizing the inter-class variance between background and foreground According to formula (5), the maximum inter-class variance between background and foreground is obtained And traverse all possible thresholds t (from 0 to L-1), calculate the corresponding value of each threshold The optimal threshold t that maximizes σ is selected to binarize the image into foreground and background. This threshold ensures maximum separation between the background and foreground, resulting in optimal segmentation. Then, segmentation and extraction of the fundus tissue ROI region is performed based on the optimal threshold t.

[0132]

[0133] As a preferred embodiment, in step 1 S14, the process of performing three-dimensional reconstruction of the fundus image is as follows:

[0134] S14-1: Superimpose the segmented 2D image layers, align the 3D primitives, and then use 3D rendering technology to generate a continuous 3D surface;

[0135] S14-2: Surface extraction technology based on the MarchingCubes algorithm is used to extract continuous surfaces from three-dimensional voxel data to obtain a continuous and clear three-dimensional scleral mold surface model.

[0136] This algorithm iterates through each voxel (cube) in the dataset and determines how the isosurface should truncate these cubes based on the comparison of the scalar values ​​of each cube vertex with a preset threshold. Each cube consists of eight vertices, which can be separated by a threshold into those above or below the threshold. The state of each vertex (i.e., above or below the threshold) can be represented as an 8-bit binary number, which defines 256 possible cube configurations. The algorithm precomputes and stores a table of isosurface geometries corresponding to these 256 configurations, called the "triangle table." For each cube, the algorithm performs the following steps: 1. Determines whether each vertex is above or below a given isovalue threshold. 2. Uses these eight Boolean values ​​to create an index and uses it to retrieve the specific triangle configuration from the triangle table. 3. Inserts vertices along the cube edges as needed to form triangles. The specific vertex positions are calculated using linear interpolation to ensure they lie precisely on the data's isosurface. This step converts the 3D voxel data into a mesh-like surface model. For example, if the scalar value of a vertex is below the threshold and the value of the adjacent vertex is above the threshold, the isosurface vertex will be placed somewhere between the two vertices. This position can be determined by the linear interpolation formula Calculated as follows, where V1 and V2 are the positions of the two vertices on the edge of the cube. value1 and value2 are the scalar values ​​of the two vertices. iso_value is the threshold.

[0137] As a preference, in step 2 S21, the generation process of the displacement field u is as follows:

[0138] S21-1: First, the baseline image and the mechanical loading image are registered using a registration framework based on Mattes Mutual Information (MMI) and Gradient Descent optimizer. MMI is a popular image registration metric used to quantify the statistical correlation between two image sets. The baseline image is used as the fixed image for image registration, and the mechanical loading image is used as the moving image for image registration. During the image registration process, the transformation parameters of the moving image are adjusted to maximize the MMI metric I(X; Y), and the transformation parameters that maximize I(X; Y) are used as the final transformation parameters; where the MMI metric I(X; Y) is calculated using formula (6);

[0139]

[0140] Where X is the baseline image, Y is the mechanically loaded image, p(x, y) is the joint probability density function of X and Y, representing the probability of observing both x and y simultaneously, and p(x) and p(y) are the marginal probability density functions, representing the probabilities of observing x and y, respectively. MMI uses a nonparametric approach to estimate probability distributions, specifically histograms to approximate the joint and marginal distributions, thus avoiding the bias that may be introduced by traditional parametric methods. The baseline image serves as the fixed image for image registration, and the mechanically loaded image serves as the moving image for image registration. During the image registration process, the MMI metric is maximized by adjusting the transformation parameters (such as rotation and translation) of the moving image. Once the optimal transformation parameters are found, the relative position and orientation between the moving and fixed images can be adjusted to the desired state.

[0141] S21-2: The calculation of the displacement field first relies on the transformation parameters obtained after registration. Specifically, the displacement field is generated by applying the final transformation parameters to all pixels in the moving image and calculating their new positions relative to the fixed image. The displacement vector of each point is determined by the difference between its original position in the moving image and its new position in the fixed image. This difference directly reflects the required displacement.

[0142] Apply the final transformation parameters to all pixels of the moving image, calculate the new positions of these pixels relative to the fixed image and generate a displacement field u, where the displacement vector D(x, y, z) of each pixel (x, y, z) in three-dimensional space is calculated according to formula (7);

[0143] D(x,y,z)=T(x,y,z)-O(x,y,z) (7);

[0144] Where T(x,y,z) is the new coordinates of the point (x,y,z) after the transformation function is applied to it, and O(x,y,z) represents the original coordinates of the point (x,y,z) in the moving image, which are the coordinates of the point before any transformation.

[0145] As a preference, in step 2 S22, the calculation process of the strain tensor E is as follows:

[0146] S22-1: The deformation gradient tensor is obtained by calculating the gradient of the displacement field and adding the unit matrix. Specifically, the deformation gradient tensor F is calculated according to formula (8); this tensor describes the local deformation of the object in detail, including stretching and compression.

[0147]

[0148] Where, is the spatial gradient of the displacement field u; I is the identity matrix;

[0149] S22-2: Strain tensor, usually used to describe the internal deformation of materials or tissues, is calculated from the deformation gradient tensor. Specifically, the strain tensor E is calculated according to formula (9);

[0150]

[0151] Where, F T It is the transpose of the deformation gradient tensor F. The strain tensor provides important information about the deformation state of the material and can reflect the degree and mode of deformation of the material after being subjected to force.

[0152] As a preferred embodiment, in step 2 S23, the effective strain ∈ eff The calculation process is as follows:

[0153] S23-1: First, solve formula (10) to obtain the eigenvalue λi of the strain tensor symmetrical to the principal strain, and then obtain the principal strain; λ i Describes the maximum and minimum strains along specific principal axes. These values ​​physically represent the maximum tension and compression a material can experience along those principal axes and are key parameters for understanding a material’s response to external loads.

[0154] det(E-λi)=0 (10);

[0155] S23-2: Comprehensively consider all principal strain contributions in three-dimensional space and calculate the effective strain ∈ according to formula (11) eff ; Effective strain∈ eff, is a scalar value that measures the overall deformation strength. It combines all principal strain contributions in three-dimensional space and is commonly used to assess material fatigue and failure. This metric helps understand the overall degree of deformation. This formula unifies the strains in different directions into a single deformation scale by taking the root sum of squares.

[0156]

[0157] Where,∈ i are the components of the principal strains.

[0158] As a preferred embodiment, in step 3 S31, the specific process of constructing the deformation field prediction network is as follows:

[0159] S31-1: Build encoder;

[0160] First, move the image I m and fixed image I f Splice the input tensor into 2×H×W×D along the channel dimension, split the input into non-overlapping 3D patches, each of size 2×P×P×P, P is usually set to 4, the total number of patches And flatten each patch into a vector x i , where i∈{1,…,N}; each patch is mapped to the feature dimension C through a linear projection layer to obtain the initial token sequence z0, as shown in formula (12), where the dimension of z0 is N×C;

[0161] z0=[x 1p E;x 2p E;…;x Np E] (11);

[0162] Where E represents linear projection,

[0163] Second, since the linear projection operation is performed on image patches and does not preserve the position of the token relative to the entire image, a position embedding E is added in the linear projection layer. pos , to integrate the location information into the token and get z0+E pos ;

[0164] Then, after the linear projection layer, four stages of Swin Transformer blocks and three stages of patch merging are applied, and the patch merging of each stage is followed by the Swin Transformer block of the previous stage; for the Swin Transformer block, the output of the Swin Transformer block is the same as the number of input tokens, which is used to extract spatial information. Each stage contains multiple Swin Transformer blocks and adopts an alternating window partitioning strategy. A rectangular parallelepiped window is used to adapt to non-square images. The size of each window is M x ×M y ×M z At each resolution, the first SwinTransformer block uses the conventional window partitioning method, starting from the upper left corner of the voxel, to evenly divide the feature map into M x ×M y ×M z non-overlapping windows of , and calculate self-attention inside each window, as shown in formula (12); the subsequent Swin Transformer block is moved from the configuration in the previous Swin Transformer block by offsetting the window of the previous Swin Transformer block by For example, the input feature map has 4×8×12 voxels. Using a window size of 2×4×6, the feature map is evenly divided into 2×2×2=8 windows in the first SwinTransformer block. Then, in the next block, the windows are divided by The number of windows becomes 3×3×3=27. We extend the original 2D efficient batch processing calculation (i.e., cyclic shift) to 3D and apply it to these 27 shifted windows, keeping the final number of attention calculation windows at 8. At the same time, based on the window-based attention, two consecutive Swin Transformer blocks are calculated according to formulas (13), (14), and (15); for patch merging, the number of tokens is reduced by 8 times by merging 2×2×2 adjacent tokens at each stage, and the feature dimension is increased to 2C; the output of the encoder is

[0165]

[0166] Where, are query, key, and value matrices respectively, d represents the dimension of query and key features, M x M y M zis the number of tokens in the 3D window, and B represents the relative position of the token in each window; because the relative position of the token along each axis (i.e., x, y, z) can only be [-M x,y,z +1,M x,y,z -1], the values ​​in B are taken from a smaller bias matrix W-MSA and SW-MSA represent the multi-head self-attention module based on regular window and the multi-head self-attention module based on moving window, respectively; MLP represents the multi-layer perceptron module; LN represents layer normalization; and z l They represent the output features of W-MSA and SW-MSA of the l-th Swin Transformer block respectively;

[0167] S31-2: Construct a decoder; the decoder is composed of multiple consecutive upsampling, convolutional layers and skip connections, where each upsampling is followed by two consecutive convolutional layers, and the final stage is followed by an additional convolutional layer. Except for the last convolutional layer, each convolutional layer is followed by a Leaky Rectified Linear Unit activation, with a convolution kernel size of 3×3, and each upsampled feature map in the decoding stage is connected to the corresponding feature map in the encoding path through a skip connection, and then processed by two consecutive convolutional layers to capture local information while generating a high-resolution feature map to generate a deformation field φ; the output deformation field φ is generated by applying sixteen 3×3 convolutions. Finally, the nonlinear deformation is applied to the moving image I using a spatial transformation function. m and deformation field φ;

[0168] S31-3: Constructing the loss function of the deformation prediction network Optimizing only the similarity metric will lead to As close to I as possible visually f However, the resulting deformation field φ may not be smooth or realistic. In order to introduce smoothness into the deformation field, a regularization term is added to the loss function It is encouraged that the displacement value of a position is similar to the values ​​of its neighboring positions.

[0169]

[0170] Where, Represents the image fidelity metric, and the similarity metric used is the mean square error (MSE) to calculate Where p represents the voxel position and Ω represents the image domain; represents the regularization of the deformation field, Where u(p) is the spatial gradient of the displacement field u, which is expressed by Perform calculations.

[0171] As a preferred embodiment, in step 3 S33, the process of calculating the deformation percentage is as follows:

[0172] S33-1: Calculate the Euclidean norm of each point in the deformation field u according to formula (17); the result of this step is an array with the same dimension as the deformation field, and each element represents the deformation variable of the corresponding voxel.

[0173]

[0174] Where u x ,u y ,u z are the displacements of each vector u at position x, y, and z in three-dimensional space;

[0175] S33-2: Next, the average deformation can be calculated. The average deformation of the deformation field is the arithmetic mean of all voxel deformations. Specifically, the average deformation of all voxel deformations is calculated according to formula (18);

[0176]

[0177] Where N is the total number of valid voxels, |u i | is the shape of the i-th voxel;

[0178] S33-3: To convert the shape variable into a relative metric, the diagonal length of the maximum image dimension is used as a reference value. For a three-dimensional image with dimensions D×H×W, specifically, the diagonal length d of the maximum image dimension in three-dimensional space is calculated according to formula (19);

[0179]

[0180] S33-4: The deformation percentage is calculated by converting the ratio of the average deformation value to the reference value into a percentage. Specifically, the deformation percentage is calculated according to formula (20);

[0181]

[0182] As a preferred embodiment, in step 4, the process of constructing the compliance classification prediction model is as follows:

[0183] S42-1: Apply the 3D ResNet18 structure to 3D data to form a 3D residual network; 3D ResNet18 consists of residual basic blocks, which consist of two convolutional layers, each followed by BN and ReLU, and the top of the block and the layer before the last ReLU in the block are connected through residual connections; ResNet-18 adopts basic blocks and uses identity connections and zero padding as shortcuts for basic blocks to avoid increasing the number of parameters of these relatively shallow networks.

[0184] S42-2: Calculate the cross-entropy loss of the 3D residual network according to formula (121). Cross-entropy loss is a commonly used loss function, especially suitable for classification problems. Its principle is to evaluate the performance of the model by measuring the difference between the predicted probability distribution and the true distribution.

[0185]

[0186] Where N is the total number of categories, y i Represents the probability distribution of the actual category. If the sample belongs to the i-th category, then y i =1, otherwise 0. represents the probability of class i predicted by the model. The purpose of cross-entropy loss is to optimize model parameters by maximizing the similarity between the predicted probability distribution and the true probability distribution, thereby improving classification accuracy. By minimizing cross-entropy loss, the model is able to better adjust its parameters so that the predicted probability distribution is as close as possible to the true distribution, thereby improving classification performance.

[0187] This paper proposes a method for scleral biomechanical assessment and myopia risk prediction based on medical imaging and artificial intelligence. This method aims to address the inability of traditional fundus image analysis to directly quantify the biomechanical properties of the posterior sclera, and provides a non-invasive, convenient, and accurate technical approach. The significant advantages of this method include: first, by combining conventional imaging techniques with advanced deep learning algorithms, it successfully quantifies the biomechanical properties of the posterior sclera, avoiding the burden of invasive testing methods. Second, based on image registration and tensor analysis, combined with the Swin Transformer deep neural network, it accurately identifies subtle scleral deformations, providing a reliable basis for early diagnosis. Furthermore, compliance classification prediction using a three-dimensional residual network (ResNet) can accurately predict the risk of myopia progression based on deformation percentage, which is of great significance for the early identification of patients with potential pathological myopia. This innovative method not only improves the accuracy of image analysis but also demonstrates good robustness under different imaging conditions, making it suitable for all types of conventional ophthalmic imaging equipment. As the technology continues to improve, it is expected to be widely used in the early screening, diagnosis, and intervention of ophthalmic diseases, and promote the deep integration of medical imaging and artificial intelligence in the field of ophthalmology. In general, the present invention provides a scientific, accurate and efficient technical solution for the prevention and control of myopia and its related complications, and has great clinical application prospects and social value.

[0188] The present invention validated the effectiveness of the proposed deformation analysis and quantitative indicators by testing them on three-dimensional models of the posterior sclera from different patients. The results demonstrate that, compared with traditional methods that rely solely on axial length or simple morphological observations, the present invention can identify patients with potential pathological myopia earlier and more precisely, providing an important basis for personalized myopia prevention and control programs. Therefore, the present invention has significant practical value and social benefits, and can make a significant contribution to reducing the risk of myopia-related vision impairment.

Claims

1. A method for scleral biomechanical assessment and myopia risk prediction based on artificial intelligence, characterized in that: The following steps are involved: Step 1: Data collection and preprocessing; S11: Select several patients as volunteers and collect eye images through OCTB scanning. Two types of images are collected for each eye of each patient: one type of image is a baseline image, and the other type is a mechanical loading image. Several sets of sample data are obtained through the massive collection process; S12: Data preprocessing: Perform Gaussian smoothing and denoising on the image I in several groups of sample data to obtain the pixel value I of the smoothed image at the coordinate (x, y) smoothed (x,y); S13: Segment and extract the fundus tissue ROI area based on the Otsu threshold segmentation algorithm; S14: reconstructing a three-dimensional sclera surface model by stacking the segmented two-dimensional image layers, thereby obtaining three-dimensional sclera surface models of several volunteers; Step 2: Use image registration technology combined with mechanical algorithms to calculate the effective strain of fundus tissue; S21: Calculation of displacement field; Generate displacement field u using image registration technology; S22: Calculation of strain tensor; Calculate the strain tensor E based on the deformation gradient tensor F; S23: Effective strain calculation; calculate the effective strain ∈ based on the strain eigenvalue λi eff , and then obtain the effective strain sample data of several volunteers; Step 3: Use the deformation field prediction network to extract the fundus deformation field and quantify the deformation degree; S31: Construct a deformation field prediction model; use an encoder-decoder structure, combine a three-dimensional SwinTransformer with a convolutional neural network to construct a deformation field prediction model, and use valid strain sample data from several volunteers for training to obtain a trained deformation field prediction model; S32: For new patients, a baseline image is obtained through OCTB scanning, followed by data preprocessing, segmentation and extraction of the fundus tissue ROI area, reconstruction of the three-dimensional scleral mold surface model, and calculation of the effective strain. The effective strain is then input as input data into the deformation field prediction model for prediction, and the fundus deformation field φ is output; S33: quantification of deformation percentage; The deformation percentage was calculated based on the ratio of the average deformation value to the reference value converted into a percentage; Step 4: Compliance classification prediction; S41: using a deformation field prediction model to predict effective strain sample data of a number of volunteers to obtain deformation fields of the number of volunteers, then quantifying the deformation percentages of the deformation fields of the number of volunteers to obtain deformation percentage data of the number of volunteers, and then classifying the deformation percentage data of each volunteer as low compliance or high compliance based on a set threshold, thereby obtaining compliance sample data sets of the number of volunteers; S42: Constructing a compliance classification prediction model; constructing a compliance classification prediction model based on 3D ResNet18, and training the compliance classification prediction model using compliance sample datasets of several volunteers to obtain a trained compliance classification prediction model; S43: The deformation percentage of the new patient is input as input data to the compliance classification prediction model for prediction, and a binary compliance prediction result is output, i.e., high compliance or low compliance.

2. The method for scleral biomechanics assessment and myopia risk prediction based on artificial intelligence according to claim 1, characterized in that: In step 1 S12, the pixel value I of the smoothed image at the coordinate (x, y) is obtained according to formula (1): smoothed (x,y); Where I(x-u0,y-v0) is the pixel value of the original image I at the coordinate (x-u0,y-v0); G(u0,v0) is the weight value of the Gaussian kernel at the offset (u0,v0), u0 is the horizontal offset from the kernel center to the current pixel, v0 is the vertical offset from the kernel center to the current pixel, k is the radius of the Gaussian kernel, σ is the standard deviation of the Gaussian kernel.

3. The method for scleral biomechanics assessment and myopia risk prediction based on artificial intelligence according to claim 1, characterized in that: In step 1 S13, the process of extracting the fundus tissue ROI region is as follows: S13-1: Calculate the grayscale histogram of the entire image and the occurrence probability p of each grayscale level i according to formula (2) i ; Where, the number of pixels n at each gray level i is i ; N is the total number of pixels in the image; S13-2: Obtain the proportion ω(t) of pixels in the image whose grayscale value is less than or equal to t according to formula (3), and obtain the average grayscale value μ(t) of these pixels according to formula (4); S13-3: According to formula (5), the inter-class variance between the background and foreground is maximized And traverse all possible thresholds t, select the t that makes σ the largest as the optimal threshold to binarize the image into foreground and background, and then perform segmentation and extraction of the fundus tissue ROI area based on the optimal threshold t; 4. The method for scleral biomechanics assessment and myopia risk prediction based on artificial intelligence according to claim 1, characterized in that: In step 1 S14, the process of performing three-dimensional reconstruction of the fundus image is as follows: S14-1: Superimpose the segmented 2D image layers, align the 3D primitives, and then use 3D rendering technology to generate a continuous 3D surface; S14-2: Surface extraction technology based on the MarchingCubes algorithm is used to extract continuous surfaces from three-dimensional voxel data to obtain a continuous and clear three-dimensional scleral mold surface model.

5. The method for scleral biomechanics assessment and myopia risk prediction based on artificial intelligence according to claim 1, characterized in that: In step 2 S21, the generation process of the displacement field u is as follows: S21-1: Perform image registration using the baseline image as the fixed image for image registration and the mechanical loading image as the moving image for image registration. During the image registration process, the transformation parameters of the moving image are adjusted to maximize the MMI metric I(X; Y), and the transformation parameters that maximize I(X; Y) are used as the final transformation parameters. The MMI metric I(X; Y) is calculated using formula (6). Where X is the baseline image, Y is the mechanical loading image, p(x,y) is the joint probability density function of X and Y, which represents the probability of observing x and y at the same time, p(x) and p(y) are the marginal probability density functions, which represent the probability of observing x and y, respectively. S21-2: Apply the final transformation parameters to all pixels of the moving image, calculate the new positions of these pixels relative to the fixed image and generate a displacement field u, where the displacement vector D(x, y, z) of each pixel point (x, y, z) in the three-dimensional space is calculated according to formula (7); D(x,y,z)=T(x,y,z)-O(x,y,z) (7); Where T(x,y,z) is the new coordinates of the point (x,y,z) after the transformation function is applied to it, and O(x,y,z) represents the original coordinates of the point (x,y,z) in the moving image.

6. The artificial intelligence-based scleral biomechanical assessment and myopia risk prediction method according to claim 1, characterized in that: In step 2 S22, the calculation process of the strain tensor E is as follows: S22-1: Calculate the deformation gradient tensor F according to formula (8); Where, is the spatial gradient of the displacement field u; I is the identity matrix; S22-2: Calculate the strain tensor E according to formula (9); Where, F T is the transpose of the deformation gradient tensor F.

7. The method for scleral biomechanics assessment and myopia risk prediction based on artificial intelligence according to claim 1, characterized in that: In step 2 S23, the effective strain ∈ eff The calculation process is as follows: S23-1: First, solve formula (10) to obtain the eigenvalue λi of the strain tensor symmetric to the principal strain, and then obtain the principal strain; det(E-λi)=0 (10); S23-2: Comprehensively consider all principal strain contributions in three-dimensional space and calculate the effective strain ∈ according to formula (11) eff ; Where,∈ i are the components of the principal strains.

8. The artificial intelligence-based scleral biomechanical assessment and myopia risk prediction method according to claim 1, characterized in that: In step 3 S31, the specific process of constructing the deformation field prediction network is as follows: S31-1: Build encoder; First, move the image I m and fixed image I f The input tensor is concatenated into 2×H×W×D along the channel dimension, and the input is split into non-overlapping 3D patches, each of which is 2×P×P×P in size, where P is set to 4 and the total number of patches is And flatten each patch into a vector x i , where i∈{1,…,N}; each patch is mapped to the feature dimension C through a linear projection layer to obtain the initial token sequence z0, as shown in formula (12), where the dimension of z0 is N×C; z0=[x 1p E;x 2p E;…;x Np E] (11); Where E represents linear projection, Second, we add the position embedding E in the linear projection layer. pos , to integrate the location information into the token and get z0+E pos ; Then, after the linear projection layer, four stages of Swin Transformer blocks and three stages of patch merging are applied, and the patch merging of each stage is followed by the Swin Transformer block of the previous stage; for the Swin Transformer block, each stage contains multiple Swin Transformer blocks, and an alternating window partitioning strategy is adopted, and rectangular parallelepiped windows are used to adapt to non-square images, and the size of each window is M x ×M y ×M z At each resolution, the first Swin Transformer block uses the conventional window partitioning method, starting from the voxel in the upper left corner, to evenly divide the feature map into M x ×M y ×M z non-overlapping windows of , and calculate self-attention inside each window, as shown in formula (12); the subsequent Swin Transformer block is moved from the configuration in the previous Swin Transformer block by offsetting the window of the previous Swin Transformer block by voxels to achieve this. At the same time, window-based attention is used to calculate two consecutive Swin Transformer blocks according to formulas (13), (14) and (15). For patch merging, each stage reduces the spatial resolution by 8 times by merging 2×2×2 adjacent tokens, and increases the feature dimension to 2C. The output of the encoder is Where, are query, key, and value matrices respectively, d represents the dimension of query and key features, M x M y M z is the number of tokens in the 3D window, B represents the relative position of the token in each window; W-MSA and SW-MSA represent the multi-head self-attention module based on the conventional window and the multi-head self-attention module based on the moving window, respectively; MLP represents the multi-layer perceptron module; LN represents layer normalization; and z l They represent the output features of W-MSA and SW-MSA of the l-th Swin Transformer block respectively; S31-2: Construct a decoder; use multiple consecutive upsampling, convolutional layers and skip connections to form a decoder, where each upsampling is followed by two consecutive convolutional layers, and the final stage is followed by an additional convolutional layer. Except for the last convolutional layer, each convolutional layer is followed by a Leaky Rectified Linear Unit activation, with a convolution kernel size of 3×3, and each upsampled feature map in the decoding stage is connected to the corresponding feature map in the encoding path through a skip connection, and then processed by two consecutive convolutional layers to capture local information while generating a high-resolution feature map to generate a deformation field φ; finally, a spatial transformation function is used to apply nonlinear deformation to the moving image I m and deformation field φ; S31-3: Constructing the loss function of the deformation prediction network Where, Represents the image fidelity metric, and the similarity metric used is the mean square error. Where p represents the voxel position and Ω represents the image domain; represents the regularization of the deformation field, Where u(p) is the spatial gradient of the displacement field u, which is expressed by Perform calculations.

9. The method for scleral biomechanics assessment and myopia risk prediction based on artificial intelligence according to claim 1, characterized in that: In step 3 S33, the process of calculating the deformation percentage is as follows: S33-1: Calculate the Euclidean norm of each point of the deformation field u according to formula (17); Where u x ,u y ,u z are the displacements of each vector u at position x, y, and z in three-dimensional space; S33-2: Calculate the average shape of all voxel shapes according to formula (18); Where N is the total number of valid voxels, |u i | is the shape of the i-th voxel; S33-3: Calculate the diagonal length d of the maximum size of the image in the three-dimensional space according to formula (19); S33-4: Calculate the deformation percentage according to formula (20); 10. The artificial intelligence-based scleral biomechanical assessment and myopia risk prediction method according to claim 1, characterized in that: In step 4, the process of building a compliance classification prediction model is as follows: S42-1: Apply the 3D ResNet18 structure to 3D data to form a 3D residual network; 3D ResNet18 is composed of residual basic blocks, which consist of two convolutional layers, each followed by a batch normalization and a ReLU, with the top of the block and the layer before the last ReLU in the block connected via a residual connection; S42-2: Calculate the cross entropy loss of the three-dimensional residual network according to formula (121); Where N is the total number of categories, y i Represents the probability distribution of the actual category. If the sample belongs to the i-th category, then y i =1, otherwise 0. represents the probability of the i-th category predicted by the model.

Images