Intelligent Intraocular Lens Selection System for Cataract Surgery Based on Ocular Biometrics

Abstract

This application relates to the field of intelligent intraocular lens (IOL) selection, specifically disclosing an intelligent IOL selection system for cataract surgery based on ocular bioparameters. The system is not limited to the acquisition of traditional static geometric data such as axial length and corneal curvature, but further utilizes a deep convolutional neural network to extract texture features and perform dimensionality reduction mapping on anterior segment tomographic images, thereby resolving the capsular bag compliance feature vector characterizing the biomechanical properties of the ocular soft tissue. Based on this, the geometric parameters of the patient's eye, capsular bag compliance features, and the mechanical and physical properties of the IOL are combined. By introducing a physical constraint mechanism, the mechanical equilibrium state after lens implantation is simulated, thus accurately predicting the effective lens position. This effectively addresses the fitting challenges of non-standard eyeballs or special lens models, significantly improving the accuracy and stability of postoperative refractive prediction.

Description

Technical Field

[0001] This application relates to the field of intelligent selection of intraocular lenses, and more specifically, to an intelligent selection system for intraocular lenses in cataract surgery based on ocular biological parameters. Background Technology

[0002] Cataract surgery has gradually shifted from simply restoring sight to precise refractive surgery, with patients increasingly expecting higher rates of vision improvement and better visual quality after surgery. In preoperative planning for cataract surgery, the precise selection of the intraocular lens (IOL) type and power is crucial to the success of the procedure. The most critical technical challenge lies in accurately predicting the effective position (ELP) of the IOL within the eye postoperatively. The accuracy of the ELP prediction directly determines the accuracy of the IOL refractive power calculation; even a small deviation in position prediction can lead to significant postoperative refractive errors. Therefore, developing an intelligent selection system capable of deeply analyzing ocular biological parameters and fully considering ocular anatomy and lens physical properties is of vital clinical value and practical significance for improving the refractive predictability of cataract surgery and avoiding postoperative refractive complications.

[0003] However, existing IOL selection and calculation schemes, whether based on traditional Gaussian optical formulas (such as Barrett Universal II) or emerging pure data-driven machine learning models (such as Hill-RBF), all have significant technical limitations. These existing technologies essentially rely on static geometric parameters such as axial length, corneal curvature, and anterior chamber depth to estimate the postoperative lens position, simplifying the complex physiological environment of the eye into a static, rigid optical platform. This prediction model based on pure geometric optics severely neglects the biomechanical characteristics of ocular tissues and fails to deconstruct the physical nature of the interaction forces between the implant and biological tissues. In reality, the final position of the implanted intraocular lens is the result of the mechanical rebound tension of the lens haptic and the elasticity and contractile force of the lens capsule reaching a mechanical equilibrium. Existing technologies lack effective characterization of the patient's capsular compliance, such as capsular tightness and texture features, and do not incorporate the intraocular lens's own mechanical parameters, such as material modulus and compressive displacement characteristics, into the calculation model. Therefore, when dealing with high myopia (long axial length with large capsular bag) or special lens designs, existing models cannot perceive the mechanical matching state between the capsular bag and the lens, leading to severe distortion in the prediction of the effective lens position, often causing unexplained hyperopia drift or refractive errors. In addition, pure algorithmic models lacking physical and mechanical constraints have insufficient generalization ability when dealing with extreme values ​​of samples, making it difficult to achieve stable and accurate predictions in the face of complex clinical individual differences.

[0004] Therefore, there is currently no intelligent system in the technology that can integrate geometric morphology, tissue biomechanical characteristics and crystal mechanical properties, and perform effective position prediction and crystal selection under physical constraints. Summary of the Invention

[0005] To address the aforementioned technical challenges, this application is proposed. According to this application, a smart intraocular lens selection system for cataract surgery based on ocular biometrics includes: a surgical data acquisition module for acquiring the subject's raw biometric data stream, anterior segment tomographic image sequence, and intraocular lens specification data; a surgical data preprocessing module for performing signal quality verification, region of interest (ROI) cropping, and specification parameter extraction on the raw biometric data stream, anterior segment tomographic image sequence, and intraocular lens specification data to obtain geometric parameter vectors, ROI image tensors, and mechanical parameter vectors; and a texture feature extraction and dimensionality reduction mapping module for inputting the ROI image tensor into a pre-trained deep convolutional neural network. Texture features are extracted and dimensionality reduced via a network to obtain a capsule compliance feature vector characterizing the biomechanical properties of ocular soft tissue. An effective position prediction module is used to predict the effective position of the lens under physical constraints based on the geometric parameter vector, mechanical parameter vector, and capsule compliance feature vector. A lens candidate module is used to perform optical ray tracing and degree quantization on the geometric parameter vector and the predicted effective lens position to obtain a list of candidate lenses containing degree information. A preferred sorting queue generation module is used to evaluate the stability of the candidate lens list and perform preferred sorting based on the capsule compliance feature vector to obtain a preferred sorting queue.

[0006] Compared with existing technologies, this application provides an intelligent intraocular lens (IOL) selection system for cataract surgery based on ocular bioparameters, aiming to solve the problem of inaccurate prediction of effective lens position caused by neglecting the biomechanical characteristics of the eye in existing technologies. This system is not limited to the acquisition of traditional static geometric data such as axial length and corneal curvature, but further utilizes a deep convolutional neural network to extract texture features and perform dimensionality reduction mapping on anterior segment tomographic images, thereby resolving the capsular bag compliance feature vector characterizing the biomechanical properties of ocular soft tissue. Based on this, the system combines the patient's ocular geometric parameters, capsular bag compliance features, and the mechanical and physical properties of the IOL, introducing a physical constraint mechanism to simulate the mechanical equilibrium state after lens implantation, thus accurately predicting the effective lens position. This coupled analysis model of capsular biomechanics and lens mechanical properties overcomes the limitations of pure geometric optics, effectively addressing the adaptation challenges of non-standard eyeballs or special lens models. By outputting the optimal lens solution through ray tracing and stability assessment, it significantly improves the accuracy and stability of postoperative refractive prediction. Attached Figure Description

[0007] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0008] Figure 1 This is a block diagram of an intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters, according to an embodiment of this application.

[0009] Figure 2 This is a schematic diagram of the data flow of an intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters, according to an embodiment of this application.

[0010] Figure 3 This is a block diagram of the surgical data preprocessing module in the intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters, according to an embodiment of this application.

[0011] Figure 4 This is a block diagram of the effective position prediction module in the intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters, according to an embodiment of this application.

[0012] Figure 5 This is a schematic diagram of the data flow in the candidate lens module of the intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters, according to an embodiment of this application. Detailed Implementation

[0013] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0014] This application is made in response to the problems of the prior art. Figure 1 This is a block diagram of an intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters, according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow in an intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters, according to an embodiment of this application. Specifically, as... Figure 1 and Figure 2As shown, the intelligent intraocular lens selection system 100 for cataract surgery based on ocular bioparameters according to an embodiment of this application includes: a surgical data acquisition module 110, used to acquire the subject's raw biometric data stream, anterior segment tomographic image sequence, and intraocular lens specification data; a surgical data preprocessing module 120, used to perform signal quality verification, region of interest cropping, and specification parameter extraction on the raw biometric data stream, anterior segment tomographic image sequence, and intraocular lens specification data to obtain geometric parameter vectors, region of interest image tensors, and mechanical parameter vectors; and a texture feature extraction and dimensionality reduction mapping module 130, used to input the region of interest image tensor into a pre-trained deep learning database. A convolutional neural network is used to extract texture features and perform dimensionality reduction mapping to obtain a capsule compliance feature vector that characterizes the biomechanical properties of ocular soft tissue; an effective position prediction module 140 is used to predict the effective position of the lens under physical constraints based on the geometric parameter vector, mechanical parameter vector, and capsule compliance feature vector to obtain the predicted effective lens position; a lens candidate module 150 is used to perform optical ray tracing and degree quantization on the geometric parameter vector and the predicted effective lens position to obtain a candidate lens list containing degree information; and a preferred sorting queue generation module 160 is used to perform stability evaluation and preferred sorting on the candidate lens list based on the capsule compliance feature vector to obtain a preferred sorting queue.

[0015] Specifically, the surgical data acquisition module 110 is used to acquire the subject's raw biometric data stream, anterior segment tomographic image sequence, and intraocular lens (IOL) specification data. It is understood that in existing cataract surgery planning technologies, the prediction of the effective postoperative IOL position mainly relies on the static geometric parameters of the eyeball, such as axial length, corneal curvature, and anterior chamber depth. However, the human eyeball is not a rigid optical dark box, but a complex physiological system full of bioactivity and mechanical properties. After the IOL is implanted into the capsular bag, its final stable position is actually the result of the antagonistic interaction between the mechanical expansion force of the lens haptic and the elastic contraction force of the lens capsular bag, ultimately achieving mechanical equilibrium. Traditional measurement methods only capture the morphology of the eyeball, neglecting the texture of the eye tissue and the mechanical properties of the lens. This leads to the prediction model failing due to the lack of physical boundary conditions when facing non-standard cases such as large capsular bags in high myopia or lax suspensory ligaments. To overcome this limitation, this application not only aims to accurately record the geometric dimensions of the eyeball, but also to deeply acquire tomographic texture information that reflects the biomechanical properties of soft tissue, as well as the mechanical and physical parameters of the artificial lens itself. This simultaneous acquisition and in-depth analysis of multidimensional data aims to provide complete physical constraints for subsequent mechanical models, thereby enabling accurate prediction of the effective lens position.

[0016] In a specific implementation scheme, the surgical data acquisition module 110 operates as follows: The surgical data acquisition module communicates with medical equipment and a database through a standardized data interface, acquiring raw biometric data streams, anterior segment tomographic image sequences, and intraocular lens specification data in parallel. The raw biometric data stream primarily originates from an optical biometer. This module directly reads the raw data packets generated by the measuring instrument through a digital interface. These data packets contain key anatomical parameters such as the subject's axial length, corneal curvature readings, anterior chamber depth, lens thickness, and corneal white-to-white distance. It is worth noting that, to ensure the robustness of subsequent calculations, the acquired data stream is not merely a single average value, but includes raw values ​​from multiple measurements and their corresponding signal-to-noise ratio (SNR) markers. For example, for axial length data, the module reads the quality scores of the interference waveform signals generated from several measurements. If the SNR of a measurement is lower than a preset verification threshold (e.g., SNR less than 15.0), the data point will be marked or removed in subsequent steps, retaining only high-confidence geometric parameters. This process ensures that the geometric boundary conditions of the input model have extremely high physical realism, avoiding initial state deviations caused by measurement errors.

[0017] After acquiring information about the geometric dimensions, the surgical data acquisition module further controls anterior segment optical coherence tomography (OCT) or ultrasound biomicroscopy to perform high-resolution imaging, obtaining a sequence of anterior segment tomographic images. Unlike conventional corneal topography scans, this acquisition focuses on the anatomical details of the lens equator, suspensory ligament region, and ciliary body. The module is configured with a specific scanning protocol, controlling the scanning light source to perform multi-directional radial scans centered on the subject's pupil, generating a sequence of grayscale images containing different angular sections. These images have extremely high spatial resolution, clearly showing the edge morphology of the lens capsule, the attachment position of the suspensory ligaments, and the structural texture of the ciliary processes. For example, the image sequence may contain a tomographic image acquired every 15 degrees from 0 to 180 degrees, each image being a two-dimensional pixel matrix, where the grayscale value of the pixels corresponds to the intensity of light reflection or scattering by the ocular tissue. These high-resolution tomographic image sequences are not only used for visual observation, but also serve as carriers of biomechanical information of the ocular soft tissues. The thickness variations of the capsule, the sparseness of the suspensory ligaments, and the morphological characteristics of the ciliary body are all implied in the texture distribution of the images.

[0018] Meanwhile, to introduce another crucial parameter required for physical equilibrium calculations, the surgical data acquisition module is also responsible for retrieving intraocular lens (IOL) specification data from a pre-built database. This dataset fundamentally differs from traditional databases that only contain optical parameters such as the A constant and spherical power; its core innovation lies in the introduction of detailed mechanical and physical properties. Specifically, this data includes the Young's modulus of the material for specific IOL models, characterizing the lens material's resistance to deformation; overall diameter data, defining the geometric span of the lens haptic after unfolding; and digital codes for the haptic design type, used to distinguish different mechanical support structures such as plate haptic, C-type haptic, or modified L-type haptic. More importantly, the data acquired by this module includes haptic compression mechanics curves, a set of functional relationships describing the reaction forces generated by the lens haptic during compression. For example, this curve records the magnitude of the radial support force generated by the haptic as the lens diameter is compressed to different sizes, in the form of discrete data points. This data allows for precise quantification of the mechanical performance of the IOL within capsular pockets of different diameters.

[0019] Specifically, the surgical data preprocessing module 120 is used to perform signal quality verification, region of interest cropping, and specification parameter extraction on the raw biometric data stream, anterior segment tomographic image sequence, and intraocular lens specification data to obtain geometric parameter vectors, region of interest image tensors, and mechanical parameter vectors. Correspondingly, during the acquisition of raw biological data, due to subject eye tremors, tear film instability, or electronic thermal noise from the device itself, the directly acquired data stream often contains non-negligible random errors and signal-to-noise ratio fluctuations. Simultaneously, ophthalmic clinical data involves multiple physical dimensions; for example, axial length is measured in millimeters, corneal curvature in diopters, and lateral dimensions such as corneal white-to-white distance have different statistical distributions due to anatomical differences. If such multi-source, heterogeneous data with vastly different numerical scales is directly used for subsequent high-precision regression prediction, it can easily lead to model oscillations during gradient descent or the masking effect of large numerical features causing the neglect of crucial micro-biomechanical variables. Furthermore, tomographic image sequences contain a large amount of redundant background information, and the mechanical properties of intraocular lenses are usually presented in unstructured descriptions or curves that computers cannot directly understand. Based on this, this application constructs a preprocessing module capable of cleaning, standardizing, aligning, and structuring features of multimodal raw data. The aim is to eliminate environmental noise interference, map physical quantities of different dimensions to a unified dimensionless feature space, and extract machine-readable, precise geometric and mechanical features from complex images and specifications.

[0020] Figure 3 This is a block diagram of the surgical data preprocessing module in the intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters, according to an embodiment of this application. Figure 3As shown, in a specific implementation, the surgical data preprocessing module 120 includes: a biometric data preprocessing unit 121, used to clean and standardize the raw biometric data stream to obtain a geometric parameter vector; a region of interest extraction unit 122, used to filter keyframes from the anterior segment tomographic image sequence and extract the region of interest from the keyframes to obtain a region of interest image tensor; an intraocular lens specification data retrieval unit 123, used to retrieve intraocular lens specification data using a pre-selected model as an index to obtain pre-selected intraocular lens specification data; and a mechanical parameter vector generation unit 124, used to extract static physical properties and compressive force displacement curve data from the pre-selected intraocular lens specification data and serially stitch the discretized curve features to obtain a mechanical parameter vector.

[0021] The surgical data preprocessing module 120 operates as follows: First, the biometric data preprocessing unit 121 begins with a signal-to-noise ratio (SNR) verification step. The unit internally sets a validity threshold based on clinical statistics, for example, 15.0. This threshold represents the minimum ratio of the peak intensity of the measured waveform to the background noise level. The unit iterates through each discrete measurement sample in the data stream, reading its associated SNR marker. If the SNR of a measurement is less than the validity threshold, it is determined that the measurement may be affected by tear film breakage or poor fixation, and the system marks it as invalid noise and removes it directly from the queue. For example, in a measurement sequence for axial length (AL) {26.02, 25.98, 31.05, 26.00}, if the SNR corresponding to the third data point 31.05 is only 5.0, far below the threshold of 15.0, this outlier will be physically removed. Subsequently, the unit performs an arithmetic mean calculation on the remaining valid data after screening, thereby obtaining cleaned biometric data that represents the subject's true anatomical state as 26.0 mm. To eliminate the dimensional differences in physical units such as millimeters (mm) and diopters (D) between different anatomical parameters and ensure that weight updates in subsequent neural network models are not affected by the absolute magnitude of the values, this unit further introduces a Z-score normalization mapping mechanism. The unit internally stores a statistical distribution parameter library built based on a large-scale healthy population (e.g., 100,000 samples), containing the population mean and standard deviation for each biometric item. For each cleaned biometric value, a Z-score normalization transformation is performed. For example, if a subject's axial length (AL) is 26.0 mm after cleansing, while the population mean in the statistical library is 23.5 mm and the standard deviation is 1.2 mm, then the normalized axial length feature value is (26.0 - 23.5) / 1.2 = 2.083. This value represents that the patient's axial length deviates from the average level by approximately 2 standard deviations. Similarly, this operation is performed for parameters such as corneal curvature (K), anterior chamber depth (ACD), lens thickness (LT), and corneal white-to-white distance (WTW). Finally, the unit concatenates these standardized dimensionless eigenvalues ​​in a specific topological order and appends a one-hot encoding or numerical index of the pre-selected intraocular lens model at the end, generating a high-dimensional one-dimensional array, namely the geometric parameter vector. This vector not only preserves the relative differences in anatomical structures but also possesses uniform numerical distribution characteristics.

[0022] Meanwhile, the region of interest extraction unit 122 performs parallel processing on the input anterior segment tomographic image sequence. Since the original sequence contains multiple frames of scanned images from different angles, and each frame is filled with numerous non-critical regions such as the cornea, iris, and lens nucleus, direct processing would introduce significant computational redundancy and distract the attention mechanism. Therefore, this unit first executes a keyframe selection strategy. The Otsu algorithm is used to calculate the full-image contrast of each frame in the sequence. The Otsu algorithm iterates through all possible grayscale thresholds. Find the optimal threshold To maximize the inter-class variance between foreground and background : ,in , These represent the percentage of foreground and background pixels, respectively. , This corresponds to the average gray level. A larger variance indicates a clearer texture structure and more distinct organizational boundaries in the image. (Unit selection) The frame containing the largest and most clearly detectable pupil center reflection is selected as the keyframe. After selecting the keyframe, the unit uses the Canny edge detection operator to perform feature localization on the image. This process smooths image noise with a Gaussian filter, calculates the magnitude and direction of the gradient, and uses double threshold detection and hysteresis boundary tracking techniques to accurately depict the contour edges of the anterior segment tissue. The algorithm searches for geometric feature points of the scleral spur and iris root in the image coordinate system to determine their pixel coordinates. The scleral spur, as the attachment point of the ciliary muscle, is a key anatomical landmark for determining the location of the suspensory ligaments. The unit is defined by the identified coordinates. Using the anchor point, construct a rectangular clipping window with a physical size covering the equatorial portion of the lens and the potential area of ​​the suspensory ligament, with a pixel size of [missing information]. For example, if the keyframe resolution is 1024×1024 and the anchor point coordinates are (300, 450), and the preset window size corresponds to a region with a physical size of 4mm×4mm, the unit will crop out this local image patch. Subsequently, a bilinear interpolation algorithm is used to uniformly scale the cropped irregular or differently resolution image patches to the standard input size required by the pre-trained neural network, such as 224×224 pixels. To enhance feature representation, the unit will also perform grayscale normalization on the image, mapping pixel values ​​to the [0,1] interval, and may increase the depth dimension by copying channels or superimposing adjacent slices, ultimately generating a region of interest image tensor of the form (224,224,3). This tensor eliminates irrelevant background and highly focuses on the biomechanical key regions that determine the crystal position, namely the equatorial part of the capsule and the suspensory ligament region.

[0023] In parallel, the intraocular lens (IOL) specification data retrieval unit 123 and the mechanical parameter vector generation unit 124 are responsible for processing the implant data. The system uses the doctor's pre-selected IOL model, such as a brand and model like Model-H1, as the database primary key and performs a fast hash index on the pre-set IOL specification dataset. The retrieval unit extracts a series of static physical properties of this model of IOL, including optical surface diameter (e.g., 6.0 mm), overall diameter (e.g., 13.0 mm), haptic equivalent Young's modulus (e.g., 2.5 MPa), and haptic natural unfolding angle. More crucially, it handles nonlinear mechanical characteristics. After the IOL is implanted into the capsular bag, its haptic is subjected to centripetal compressive force at the equator of the capsular bag; this compressive behavior is usually not linear.

[0024] The mechanical parameter vector generation unit 124 extracts the compression force-displacement curve data corresponding to the model from the database. This curve describes the functional relationship between the amount of crystal diameter compression and the generated reaction force. Since neural networks cannot directly process inputs in functional form, the unit discretizes and samples the curve. In specific implementation, within the possible compression range of the crystal, such as a diameter from 10.0 mm to 13.0 mm, the unit sets k sampling points at a fixed step size (e.g., 0.1 mm) to extract the corresponding reaction force values, forming a mechanical feature sequence. For example, for a crystal with a diameter of 13.0 mm, an expansion force of 0.5 mN is generated when compressed to 12.0 mm, and an expansion force of 1.2 mN is generated at 11.0 mm. These discrete values ​​accurately characterize the mechanical compliance of the crystal. Finally, the unit serially concatenates the static physical properties with the discretized curve sequence to construct a high-dimensional mechanical parameter vector.

[0025] Specifically, the texture feature extraction and dimensionality reduction mapping module 130 is used to input the tensor of the region of interest image into a pre-trained deep convolutional neural network for texture feature extraction and dimensionality reduction mapping to obtain a capsule compliance feature vector characterizing the biomechanical properties of ocular soft tissue. It should be understood that in the preoperative assessment system for cataract surgery, although traditional optical measurement methods can obtain macroscopic geometric parameters of the eyeball with micron-level precision, such as axial length and anterior chamber depth, these static geometric values ​​cannot touch the biomechanical essence of the microstructure inside the eyeball tissue. As the direct carrier of the artificial lens, the fibrous density of the lens capsule, the integrity of the suspensory ligaments, and the degree of relaxation of the ciliary muscle directly determine its elastic deformation ability and long-term contraction trend when resisting lens haptic tension postoperatively. These soft biomechanical properties are not directly expressed as explicit numerical indicators, but are hidden in the high-dimensional pixel matrix of the anterior segment tomographic image as implicit texture patterns, gray-level gradients, and spatial structural distributions. Human visual observation alone is insufficient to quantify these complex texture features, let alone establish a definite mapping relationship between them and physical and mechanical parameters. Therefore, this application introduces a deep computing architecture with high-order feature extraction and manifold learning capabilities, aiming to extract deep semantic features strongly related to biomechanics from seemingly messy tomographic images, and to reduce the dimensionality of unstructured visual information into mathematical vectors that are understandable to computers and have physical meaning.

[0026] In one specific implementation, the texture feature extraction and dimensionality reduction mapping module 130 includes: a texture feature encoding unit 131, used to input the image tensor of the region of interest into the backbone layer of a deep convolutional neural network for multi-scale texture feature encoding to obtain a high-dimensional feature map; a spatial topology aggregation unit 132, used to perform global spatial topology aggregation on the high-dimensional feature map to obtain a global feature vector; and a mechanical manifold embedding and mapping unit 133, used to perform mechanical manifold embedding and dimensionality reduction mapping on the global feature vector to obtain a bag-like compliance feature vector.

[0027] The texture feature extraction and dimensionality reduction mapping module 130 is executed as follows: First, the texture feature encoding unit 131 receives the region of interest (ROI) image tensor as input. To effectively capture multi-scale information from fine edges to macroscopic textures, this unit employs a deep convolutional neural network finely tuned by a specific transfer learning strategy as the backbone extraction network. In this embodiment, the EfficientNet-B0 architecture, which balances computational efficiency and feature representation capability, is selected. This network is a deep hierarchical structure composed of a series of moving-flipping-bottom convolutional modules (MBConv). When the ROI image tensor is input into the network, it first passes through the bottom standard convolutional layer, using small-sized convolutional kernels, such as 3×3, to perform stride-sliding scanning on the image. The convolution operation extracts basic edge, corner, and gradient information in specific directions by calculating the sum of the convolutional kernel weights and local pixel dot products. Subsequently, the data flow enters the deep MBConv modules. These modules utilize depthwise separable convolution and channel attention mechanisms to adaptively amplify key feature channels related to suspensory ligament tension or capsular fibrosis, while suppressing irrelevant noise such as light reflection. Following each convolutional operation in the network is batch normalization and the Swish activation function. Compared to the traditional ReLU function, the smooth and non-monotonic nature of the Swish activation function allows the network to retain more negative gradient information when processing natural images of biological soft tissues with continuous and complex deformations, thus fitting non-linear texture features more delicately. As the network depth increases, the spatial resolution of the feature map gradually decreases, but the channel dimension increases layer by layer, transforming the originally concrete pixel array into an abstract semantic representation. For example, a shallow network may identify the sharpness of the ciliary bag edge, while a deep network can encode the overall gray-level entropy of the ciliary body region or the sparse texture pattern of the suspensory ligament. Finally, this unit outputs the high-dimensional feature map generated by the last convolutional stage of the backbone network. If the network downsamples the input by a factor of 32, for a 224×224 input, the output feature map size will be 7×7×1280. Each 7×7 matrix slice in this tensor represents the spatial response distribution of a specific texture pattern throughout the region of interest.

[0028] Next, the data stream enters the spatial topology aggregation unit 132. While the resulting high-dimensional feature map contains rich semantic information, it still retains a 7×7 spatial structure with a dimension as high as 1280. Direct expansion would lead to overfitting and computational disaster. More importantly, when evaluating pocket compliance, the focus is more on the presence and intensity of specific pathological textures, such as fibrotic plaques or ligament rupture signs, within the overall region, rather than their specific pixel coordinates. That is, the features should possess translation invariance and rotation robustness. To this end, this unit performs a global average pooling operation on the high-dimensional feature map. This operation is performed independently on each feature channel, calculating the arithmetic mean of the activation values ​​at all spatial locations within that channel. Through this process, the 7×7 spatial dimension is collapsed to 1×1, which physically equates to statistically analyzing the concentration or average response intensity of each texture feature throughout the entire region of interest. After performing this operation on all D channels, the unit reassembles the resulting D scalars to generate a one-dimensional real vector of length D, such as 1280, i.e., the global feature vector. This vector highly condenses the overall visual state of the image, completely stripping away the constraints of specific spatial coordinates, allowing the features to purely express the material properties of the eye tissue.

[0029] Finally, the mechanical manifold embedding mapping unit 133 completes the process. Although the global feature vector is compact, its physical meaning remains similar to diagonal texture responses in computer vision, and it contains a large amount of redundant information unrelated to biomechanics, such as the noise floor patterns of the imaging device. To extract biomechanical properties that can directly guide crystal position prediction, this unit constructs a multilayer perceptron (MLP) as the projection head. This MLP includes an input layer, hidden layers, and an output layer. In specific implementation, the global feature vector... First, the weights are multiplied by the first-layer weight matrix and a bias vector is added. After non-linear activation, this is mapped to a lower-dimensional hidden space, such as 512-dimensional space. Then, it passes through an intermediate layer containing a bottleneck. This forces the model to perform extreme feature compression during information flow, compelling the network to retain only the discriminative features that contribute most to the final prediction target (effective location), automatically filtering out irrelevant visual noise. Finally, the output layer weight matrix... A linear transformation maps features to a predefined low-dimensional manifold space, such as 16-dimensional space. The mathematical model of this operation can be expressed as: ;in The activation function is ReLU or GELU, while the output uses the hyperbolic tangent function. . The function strictly limits the values ​​of each dimension of the output vector to the range [-1, 1]. This normalization is crucial: on the one hand, it prevents numerical divergence; on the other hand, it gives the output vector a clear relative physical meaning. For example, a certain dimension of the output vector may correspond to the capsular elastic modulus index, with a value close to 1 indicating extreme capsular rigidity (fibrosis), and a value close to -1 indicating extreme laxity. These weight matrices and biases were jointly optimized using backpropagation during the offline training phase of the model, based on a large amount of clinical data annotated with postoperative true lens position (ELP) and refractive results. By minimizing the position prediction error, the network learns how to map visually high-density textures to physically high-stiffness vectors. Ultimately, this unit outputs a compact, purely biomechanically characterizing capsular compliance feature vector representing the soft tissue biomechanical properties of the eye. This vector is no longer a simple statistical representation of image data, but rather a digital physical portrait of the patient's internal soft environment after nonlinear transformation and manifold embedding via a deep neural network; it quantifies the capsular tightness, elasticity, and contraction potential.

[0030] Specifically, the effective position prediction module 140 is used to predict the effective position of the intraocular lens (IOL) under physical constraints by analyzing the geometric parameter vector, mechanical parameter vector, and capsular compliance feature vector. Correspondingly, the geometric parameter vector, mechanical parameter vector, and capsular compliance feature vector reside in independent and heterogeneous feature spaces: geometric parameters describe morphology, mechanical parameters describe stiffness, and capsular compliance describes the environment. This isolated feature representation cannot directly answer the core question of where the IOL will ultimately remain after implantation. This is because the postoperative effective lens position (ELP) is not a simple geometric superposition, but a dynamic physical equilibrium endpoint—the position where the radial expansion force of the IOL haptic, the centripetal contraction force of the lens capsular, and the tension of the suspensory ligaments antagonize each other and ultimately reach a mechanically stable state. Existing single models often fail to integrate these three heterogeneous information types, leading to prediction failures in complex cases. Therefore, this module can not only align and deeply fuse multi-source heterogeneous data in terms of feature dimensions, but more importantly, it needs to introduce physical constraint mechanisms in the inference process to simulate the mechanical interaction between the crystal and the capsule, and map the abstract high-dimensional feature vector back to the specific physical coordinate space, thereby outputting prediction results that conform to physiological laws and have extremely high accuracy.

[0031] Figure 4 This is a block diagram of the effective position prediction module in the intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters, according to an embodiment of this application. Figure 4As shown, in a specific implementation, the effective position prediction module 140 includes: a multi-physics feature fusion unit 141, used to perform multi-physics feature alignment and fusion on geometric parameter vectors, mechanical parameter vectors and bag compliance feature vectors to obtain a panoramic feature vector; a mechanical equilibrium constraint inference unit 142, used to input the panoramic feature vector into a physical perception regression network embedded with a mechanical equilibrium loss function to perform deep inference based on mechanical equilibrium constraints to obtain a normalized prediction value; and a physical space mapping and restoration unit 143, used to perform physical space mapping and numerical restoration on the normalized prediction value based on preset statistical distribution parameters to obtain the predicted effective crystal position.

[0032] The effective location prediction module 140 operates as follows: First, the multiphysics feature fusion unit 141 receives a geometric parameter vector, a mechanical parameter vector, and a bag compliance feature vector. These three vectors differ significantly in physical meaning and numerical dimension. For example, the geometric parameter vector might be a short vector containing 8-dimensional standard fractional values ​​such as axial length and curvature; the mechanical parameter vector is a 50-dimensional high-dimensional vector containing the material's Young's modulus and discretized compressive force curve, focusing on describing the implant's physical properties; while the bag compliance feature vector is a 16-dimensional manifold embedding vector, highly condensing the softness and hardness of biological tissue. To achieve information interaction, this unit first performs a serial splicing operation in the feature channel dimension. The spliced ​​original mixed vector has a dimension of 8 + 50 + 16, which is a 74-dimensional mixed vector. However, simple splicing does not achieve intrinsic feature fusion. Therefore, this unit then uses a fully connected layer to perform linear transformation and feature recombination on the original mixed vector. In this application, the fully connected layer is configured with a weight matrix and bias vector of dimension 74×512. Through matrix multiplication, heterogeneous features are uniformly mapped to a high-dimensional shared latent space, generating a 512-dimensional panoramic feature vector. During this process, the introduction of activation functions such as GELU enables the model to capture the nonlinear relationships between different physical quantities, such as the combined effect of long axial length (geometric) and high-modulus crystals (mechanical) in the relaxed pocket (compliance). This step ensures that the subsequent inference network can simultaneously perceive anatomical structures, crystal properties, and tissue states at a unified semantic level.

[0033] Next, the panoramic feature vector is fed into the core mechanical equilibrium constraint inference unit 142. This unit runs a specially designed physical perception regression network. Unlike general regression models, this network incorporates physical constraint mechanisms in both its architecture design and parameter training phases. The main architecture of the network consists of several stacked residual modules, each containing a normalization layer, a linear layer, and an activation layer, designed to simulate the iterative process of a mechanical system finding its energy minimum. The loss function is fundamentally changed during the network training phase. Traditional networks optimize only the prediction error (MSE), while the network used in this unit incorporates a mechanical equilibrium loss function. .in, It is a penalty term constructed based on Hooke's Law. Specifically, during training, the model not only fits the ELP values ​​of historical data, but also considers the input crystal mechanical parameters such as stiffness coefficient. and predicted compression Calculate the outward expansion force generated by the crystal. Simultaneously, the contractile force of the capsule is derived using the compliance characteristics of the capsule. The loss function will penalize... Larger predictions necessitate that the mapping learned by the network must satisfy the mechanical equilibrium condition (i.e., the net force is zero). After extensive training on large-scale clinical data (including real postoperative ELP and complete preoperative parameters), the weights of the hidden layer neurons in the network have internalized the physical laws of ocular biomechanical interactions. During inference, the input 512-dimensional panoramic feature vector undergoes layer-by-layer nonlinear transformations in the deep layers of the network, simulating the physical process of axial displacement of the lens implantation capsule under juxtaposed forces. The final layer of the network uses a linear activation function, outputting a dimensionless scalar value, i.e., the normalized prediction value. This value statistically represents the degree of deviation of the predicted effective lens position from the population average. For example, a normalized prediction value of 0.5 does not mean the position is 0.5 mm, but rather that the patient's postoperative lens position is expected to be 0.5 standard deviations behind the average. This normalized output design avoids the gradient instability problem that may occur when the network directly regresses large numerical physical quantities and enhances the model's adaptability to ocular data from different ethnic groups.

[0034] Finally, the physical space mapping and reconstruction unit 143 transforms the abstract output of the model into clinically usable numerical values. This unit pre-stores statistical distribution parameters based on large-scale retrospective cohort studies, primarily including the mean of the actual effective lens position after surgery in the target population. For example, 4.50mm and standard deviation For example, 0.80 mm. These parameters form the benchmark for anchoring dimensionless predictions back to physical reality. The reduction process first performs inverse normalization calculations. According to the formula... The unit uses the normalized predicted values ​​of the input. Calculate the preliminary physical location. (As shown in the model output.) =-1.2, and preset parameters =4.50mm, =0.80mm, then the preliminary predicted position is calculated as 4.50 + (-1.2 × 0.80) = 3.54mm. This indicates that the model judges that the patient's ELP is significantly lower than the average due to a tight capsular bag or anterior lens displacement. However, simple statistical reduction may produce outliers that violate physiological structure under extreme inputs (e.g., the lens position is located anterior to the cornea or posterior to the retina). To ensure the safety of surgical planning, this unit introduces a physical space clamping mechanism based on anatomical constraints. The unit reads the patient's original anatomical parameters: anterior chamber depth (ACD) and axial length (AL). According to the safety axioms set by this technical solution, the artificial lens must be located within the anatomical space of the original lens, that is, the anterior surface cannot cross the iris plane, approximately posterior to the ACD, and the posterior surface cannot contact the retina. The specific implementation formula is as follows: In this formula, A physiological boundary for the lens position was established. It should be noted that the value of 0.5 here is an empirical safety constant (in mm) based on the distance between the lens thickness and the iris plane, ensuring that the predicted position lies at least within a reasonable lens pocket plane after the anatomical anterior chamber depth. A physiological posterior boundary was set to prevent the predicted location from excessively shifting backward and encroaching on the vitreous cavity space or causing numerical divergence. For example, if a highly myopic patient has ACD = 3.5 mm and AL = 30.0 mm, the model's preliminary calculations... =6.8mm (due to the extremely long axial length, the model predicts the capsule to be extremely loose, causing a posterior displacement). Performing a front bound check: max(6.8,3.5+0.5)=max(6.8,4.0)=6.8mm. Performing a back bound check: min(6.8,30.0-10.0)=min(6.8,20.0)=6.8mm. The final predicted effective lens position is 6.8mm. Conversely, if the model outputs outliers... =2.0mm, then max(2.0,4.0)=4.0mm. The predicted value is forcibly corrected to the minimum value of 4.0mm that conforms to the anatomical structure, thereby avoiding the possibility of impossible physical predictions misleading doctors.

[0035] In a specific preferred embodiment, the effective position prediction module 140 includes: a multi-physics feature fusion unit, used to align and fuse geometric parameter vectors, mechanical parameter vectors, and bag compliance feature vectors to obtain a panoramic feature vector; a neural implicit solution unit, used to perform neural implicit solution based on energy potential surface on the panoramic feature vector to obtain a normalized predicted value; and a physical space mapping and restoration unit, used to perform physical space mapping and numerical restoration on the normalized predicted value based on preset statistical distribution parameters to obtain the predicted effective crystal position. Since the implementation of the multi-physics feature fusion unit and the physical space mapping and restoration unit has been explained above, it will not be repeated here, but the focus will be on the specific implementation of the neural implicit solution unit.

[0036] It's understandable that in traditional cataract surgery planning, the prediction of the effective position of the intraocular lens (ELP) after surgery typically relies on statistical regression models based on large datasets. However, these models are essentially probabilistic fitting tools; their training objective is to minimize the Euclidean distance between the predicted position and the actual historical data, while physical laws are only introduced as a soft constraint into the loss function. This means that after the model is trained and deployed to actual clinical applications (i.e., the inference phase), the predicted value output by the network, while statistically reasonable, cannot strictly guarantee that this position satisfies the physical equilibrium conditions. This is like someone mimicking the action of throwing a stone; it looks very similar, but they haven't truly understood the force principles of a parabola. For the vast majority of standard eyeballs, this approximation is acceptable; however, in extreme and rare cases such as high myopia with a large capsular bag, abnormal laxity of the suspensory ligaments due to Marfan syndrome, or plate-loop intraocular lenses with complex mechanical designs, simple probability fitting often produces severe physical illusions—that is, outputting a position that, while numerically reasonable, does not result in zero net force in actual intraocular force analysis, leading to unexplained hyperopia drift or refractive errors postoperatively. To fundamentally solve this statistically reasonable but physically distorted engineering problem, a computational architecture that can directly embed physical laws as hard constraints into the reasoning process is introduced. The neural implicit solution mechanism based on energy potential surface proposed in this scheme is precisely designed to address this challenge. This mechanism does not force the neural network to directly memorize complex mechanical equilibrium points, but cleverly utilizes the principle of minimum potential energy in physics: for any stable mechanical system, its final equilibrium position must be in the state of minimum total potential energy. Therefore, the system's task becomes predicting the potential energy field parameters of eye tissue and naturally deriving the true physical equilibrium position by finding the potential energy minimum point, thereby ensuring that the output strictly follows the laws of physics under any extreme conditions.

[0037] Based on this, in a specific implementation, the neural implicit solver unit is used for: Mapping the panoramic feature vector to the capsule's mechanical parameters yields the capsule's nonlinear elastic parameter vector. The background for this step is that, as an industrially standardized product, the mechanical properties of an intraocular lens, such as Young's modulus and geometric dimensions, are completely known constants; however, the lens capsule, as biological tissue, has mechanical properties such as elastic modulus and relaxation coefficient that vary from person to person and cannot be directly measured, making it the only unknown variable in this mechanical system. Therefore, utilizing the powerful nonlinear mapping capabilities of deep neural networks to deconstruct parameters characterizing the capsule's physical properties from the panoramic feature vector, which incorporates multimodal information, is a prerequisite for constructing an accurate potential energy model. Specifically, this deep neural network employs a dedicated architecture called Residual Fully Connected Network (Res-FCN), designed to handle the deep entanglement of high-dimensional features and the decoupling of physical parameters. This network is configured with an input layer, such as 512 nodes, matching the dimension of the panoramic feature vector, followed by cascading several residual interaction modules, such as 4 to 6. Each module consists of a linear fully connected layer, layer normalization, and a GELU activation function. This residual skip connection design not only accelerates gradient backpropagation but also enables the network to keenly capture subtle but crucial changes in biomechanical features. At the network's end is a physically constrained projection head. Positive activation functions such as Softplus or Exponential force the output stiffness coefficient and hardening exponent to strictly fall within the positive real number domain, thus avoiding predictions that violate physical principles, such as negative stiffness. During training, since the actual capsule mechanical parameters cannot be directly observed as labels in vivo, the network is jointly optimized through an end-to-end differential physics rendering mechanism: the error between the calculated final effective crystal position and the clinically true value is driven by a differentiable implicit solver and a chain-like backpropagation of the total potential energy equation, driving the network to automatically adjust weights, thereby learning which texture features correspond to which physical parameters. Specifically, the 512-dimensional panoramic feature vector generated in the preceding steps... Input into a specially designed parameter generation network In this network, instead of outputting position coordinates, a set of dimensionless parameters describing the nonlinear elasticity of the bag is mapped. If the elastic behavior of the sac conforms to a two-parameter exponential hardening model, then the output parameter vector... It may contain two components: ,in Represents the basic stiffness coefficient of the capsule. This represents the nonlinear hardening exponent. Its mathematical expression is: For example, for a highly myopic patient with extremely loose capsular pockets, the network might output a small stiffness coefficient. =0.2 and a large hardening index =2.5, which means that the cyst is easily deformed when initially subjected to crystal expansion force, but hardens rapidly after reaching a certain expansion amount. This step not only achieves a semantic leap from abstract features to physical parameters, but more importantly, it transforms uncertain biological tissue characteristics into computable mathematical variables, laying the foundation for the subsequent construction of potential energy equations.

[0038] The total potential energy of the system is calculated based on the known mechanical parameter vectors and the capsule's mechanical parameters to obtain the total potential energy equation. Now that the known mechanical parameters of the crystal and the predicted mechanical parameters of the capsule are obtained, according to the principles of physics, the total potential energy of the entire crystal-capsule system is... This is equal to the sum of the energy stored during the elastic deformation of the crystal and the energy stored during the elastic deformation of the capsule. The purpose of constructing this equation is to provide a global energy topography map for finding the equilibrium position. In specific implementation, a normalized effective crystal position is introduced. As the independent variable, taking values ​​in the range [0,1], the total potential energy function is constructed. : The first term in the formula represents the elastic potential energy of the crystal. .in From mechanical parameter vector The compressive restoring force function (a known quantity) of the crystal loop extracted from it. This is the integration variable. This term, through integration over the force-displacement curve, precisely calculates the position to which the crystal is compressed. The potential energy accumulated over time. The second term in the formula represents the elastic potential energy of the bag. These are parameters predicted using a neural network. And the unknowns calculated using a pre-defined constitutive model (such as a spring model or an exponential model). For example, if an exponential model is used, By superimposing these two factors, the system defines a value related to position. The scalar field. In a physical sense, each point on this potential energy surface represents a possible system state, and the lower the potential energy, the more stable the state. For example, in the case of the relaxed sac mentioned above, the calculated total potential energy curve might show a broad, gentle trough, and the minimum point would be significantly biased towards the rear. The value is relatively large, which is completely consistent with the physical intuition that the relaxation bag has a weak constraint on the crystal, leading to the crystal shifting backward.

[0039] The total potential energy equation is solved implicitly to obtain a normalized predicted value. This is the core decision-making step in the entire process. Since the true physical equilibrium position necessarily corresponds to the minimum point of the system's total potential energy, solving the ELP is equivalent to solving the mathematical problem of finding the zero first derivative of the total potential energy equation with respect to position (i.e., the net force). This step utilizes a differentiable optimization layer to directly solve the physical equation during the inference process, completely eliminating the errors caused by rote memorization in the network. Specifically, the system uses the Newton-Raphson iterative method to find... The local minimum point. The iterative formula is as follows: ; and the final output: ;in, It is the first The position estimate for the next iteration; It is the first derivative of the total potential energy with respect to position, and its physical meaning is the resultant force on the system (the difference between the crystal expansion force and the capsule contraction force). It is the second derivative, and its physical meaning is the tangent stiffness matrix of the system; The step size factor, if set to 1.0, controls the convergence speed. In actual operation, the system sets an initial guess position. For example, if the value is 0.5, calculate the resultant force and stiffness at that point, and use the formula to update it to obtain a value closer to the equilibrium point. This cycle continues until the net force approaches zero, for example... The final converged value It refers to the normalized predicted value that strictly satisfies the mechanical equilibrium conditions. For example, even for extremely complex asymmetric bags, this implicit solution can precisely pinpoint the unique force equilibrium point of the system. =0.68, corresponding to a physical location of 5.04mm, rather than simply a statistical average. This design ensures the absolute self-consistency of the output in terms of physical nature, effectively avoiding the risk of prediction drift that violates the laws of physics, and significantly improving the reliability and safety of surgical planning.

[0040] Specifically, the candidate lens module 150 is used to perform optical ray tracing and diopter quantification on the geometric parameter vector and the predicted effective lens position to obtain a list of candidate lenses containing diopter information. Correspondingly, the effective lens position (ELP) only completes the task of geometric positioning and does not yet address the core requirement of optical correction. The ultimate goal of the surgery is to select an artificial lens with appropriate diopter so that external light, after being refracted by the cornea and the artificial lens located at a specific ELP, can be precisely focused on the fovea centralis of the retina. Since the eyeball is not a standard, simple combination of lenses, and the diopter of commercially available artificial lenses is distributed in discrete steps (e.g., 0.5D), the optimal lens model cannot be directly determined based solely on positional data. This application combines static geometric parameters, predicted physical position, and the optical properties of the lens material, using a rigorous ray tracing algorithm to solve for the theoretical ideal diopter, and further evaluates the prognostic residual error caused by discrete-size lenses. This process aims to translate biomechanical predictions into clinically feasible lens implantation plans, ensuring that the finally selected lens is not only positionally stable but also optically capable of precisely offsetting the patient's refractive error.

[0041] Figure 5 This is a schematic diagram of the data flow in the candidate lens module of the intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters, according to an embodiment of this application. Figure 5 As shown, in a specific implementation, the crystal candidate module 150 includes: a virtual eye optical model data construction unit 151, used to reconstruct the optical surface coordinate system using curvature and axial length data in the geometric parameter vector, and load crystal refractive index parameters at the predicted effective crystal position to construct virtual eye optical model data; an ideal refractive power generation unit 152, used to perform zero-point iterative solution based on ray tracing on the incident virtual eye optical model data to obtain the theoretical ideal refractive power; and a specification discretization ray tracing unit 153, used to map discretized specifications centered on the theoretical ideal refractive power, and perform forward ray tracing on each specification to calculate the equivalent spherical residual relative to the retina to obtain a candidate crystal list.

[0042] The lens candidate module 150 operates as follows: First, the virtual eye optical model data construction unit 151 receives the geometric parameter vector output by the surgical data preprocessing module and the predicted effective lens position output by the effective position prediction module. For accurate ray tracing, a global optical surface coordinate system based on the visual axis is first established. The unit extracts the corneal curvature reading K (in diopters D) from the geometric parameter vector and combines it with a preset corneal refractive index. Take 1.3375 or 1.336, and use the formula. Convert it to the radius of curvature of the anterior corneal surface For example, if the patient With a value of 43.0D, the calculated radius of curvature is approximately 7.85 mm. Subsequently, the unit defines the vertex of the anterior corneal surface as the physical origin (0,0,0) of a three-dimensional Cartesian coordinate system, and the optical axis direction as the positive Z-axis. Within this coordinate system, the unit uses the predicted effective lens position, such as 4.50 mm, and its corresponding physical meaning (the position of the principal plane of the lens) to determine the spatial coordinates of the intraocular lens optical center on the Z-axis. =4.50. Simultaneously, using the axial length AL in the geometric parameter vector, such as 24.00 mm, the intercept coordinate of the fovea centralis on the Z-axis is determined to be 24.00. Besides establishing the geometric surface, the unit also needs to fill the optical properties of the medium between each optical interface. Based on the Gullstrand model eye or customized parameters, the refractive index constants of the cornea, anterior chamber fluid, lens material, and vitreous body are set respectively. In particular, the refractive index of the artificial lens... The specifications of the intraocular lens (IOL) from the previous steps are directly inherited, such as the hydrophobic acrylic material being 1.49. Through this process, the unit constructs a complete virtual eye optical model data containing multiple refractive interfaces (air-cornea, cornea-aqueous humor, aqueous humor-IOL, IOL-vitreous) and the endpoint receiving screen (retina). This model provides accurate geometric and physical boundaries for subsequent optical calculations, and its core mathematical description implicitly addresses residual refractive errors. The error is defined as measuring the optical distance between the actual focal point and the ideal retinal plane: ;in, This represents the refractive index of the glass, which is 1.336. axial length, This represents the defocusing of the actual focal point of the light rays relative to the retinal plane. This formula indicates that the fundamental purpose of the model construction is to adjust the crystal parameters to achieve... Approaching zero.

[0043] Next, the ideal refractive power generation unit 152, based on the constructed virtual eye model, uses numerical analysis to solve for the theoretically optimal lens power. Since this power is a continuous variable and has no analytical solution, the unit employs a ray-tracing-based zero-point iterative solution strategy. The unit first sets a simulated parallel ray (simulating imaging of an object at infinity) and incident it onto the virtual corneal surface in a concentric ring configuration. This concentric ring design can simulate spherical aberration effects under different pupil diameters, such as a 3.0mm photosensitive zone. The light strictly follows Snell's law as it passes through each optical interface: ;in, , These are the refractive indices of the media on either side of the interface. , These are the angle of incidence and the angle of refraction, respectively. The light rays pass sequentially through the anterior and posterior surfaces of the cornea (if the model includes corneal thickness), the anterior chamber fluid, and finally reach the area located at... An artificial lens was developed to precisely focus light onto the surface of the retina. The system initializes a crystalline refractive power estimate using the Newton-Raphson iterative algorithm. For example, 20.0D, and calculate the focal coordinates of the light rays after passing through the system at that degree. At this point, the system will detect the focus error. .like If the error exceeds the preset precision tolerance range, such as >0.01mm, the algorithm will adjust according to the error gradient. Value. The iterative formula aims to minimize the objective function: , here This represents the composite mapping of the entire ray tracing function. After multiple rapid iterations, the algorithm converges, outputting a high-precision continuous value, namely the theoretical ideal diopter. For example, the calculated result might be 21.345D. This value represents the theoretical lens power required to achieve perfect emmetropia given the current anatomy and predicted location.

[0044] Finally, the specification discretization ray tracing unit 153 translates theoretical calculations into clinically viable products. Due to limitations in industrial production and the supply chain, the diopter of commercially available intraocular lenses is typically discrete, with a standard step size of 0.5D, and possibly 0.25D in some ranges. Therefore, a 21.345D lens does not exist in reality. This unit uses the calculated theoretical ideal diopter of 21.345D as the center, rounding up and down (left and right) to select several commercially available specifications covering a certain range, such as ±1.5D, forming a candidate set. For example, the selected set might include {20.5D, 21.0D, 21.5D, 22.0D}. For each discrete lens diopter in the set... The unit then re-introduces the theoretical parameters into the virtual eye optical model, replacing them, and performs another complete forward ray tracing simulation. This time, the calculation is no longer about finding the diopter, but rather about calculating the optical effect at a fixed diopter. Light rays pass through... After a 21.5D crystal is formed, the particles will converge at a new focal point. The unit calculates the difference in physical distance between the focal point and the retina, and converts it into a postoperative estimated residual power using the aforementioned residual refractive error formula. For example, for a 21.5D lens, the calculated focal point is located at a small distance in front of the retina, resulting in an equivalent spherical residual error of -0.12D (mild myopia); while for a 21.0D lens, the residual error is +0.31D (mild hyperopia). The unit encapsulates each set of calculated data—including the lens type, discrete nominal power, and corresponding estimated residual refractive power—into a structured data object. Finally, the unit outputs a list of candidate lenses containing multiple records, clearly showing the expected optical results after implantation of different lens sizes.

[0045] Specifically, the preferred sorting queue generation module 160 is used to perform stability evaluation and preferred sorting of the candidate lens list based on the capsular bag compliance feature vector to obtain the preferred sorting queue. In other words, after completing optical ray tracing and diopter calculation, although the system obtains a series of candidate lenses that optically meet the refractive correction requirements, the complexity of clinical decision-making goes far beyond optical matching. Different types of lenses have drastically different degrees of dependence on capsular bag stability. For example, while multifocal intraocular lenses can provide vision at both near and far distances, their complex optical surface design is extremely sensitive to eccentricity and tilt. Once the capsular bag contracts postoperatively or the suspensory ligaments loosen, causing even a small displacement of the lens, the patient's visual quality will suffer a catastrophic blow, resulting in severe glare and decreased contrast. Conversely, traditional monofocal spherical lenses, although functionally singular, have a strong tolerance for positional deviations. Existing selection systems often make recommendations based solely on minimizing optical residuals, ignoring the hidden risk factor of capsular bag biomechanical stability. This leads to the erroneous implantation of high-end sensitive lenses in patients with poor capsular bag conditions (such as high myopia or pseudoexfoliation syndrome), causing postoperative disputes. Therefore, the biomechanical risks of the capsules are quantified from the previously extracted features, and this risk is combined with the crystal type sensitivity to construct a comprehensive evaluation system that includes optical benefits and risk costs, thereby intelligently adjusting the recommendation order.

[0046] In one specific implementation, the preferred sorting queue generation module 160 includes: a bag stability risk quantification unit 161, used to quantify the bag stability risk of the bag compliance feature vector to obtain the bag instability index; and a comprehensive cost calculation and sorting unit 162, used to match position sensitivity coefficients for different crystal types in the candidate crystal list, calculate a comprehensive cost function by combining the bag instability index and optical residual error, and sort them in ascending order of comprehensive cost to obtain the preferred sorting queue.

[0047] The preferred sorting queue generation module 160 executes as follows: First, the capsular stability risk quantification unit 161 receives the capsular compliance feature vector. This vector is a real-valued vector in a low-dimensional manifold space, such as 16-dimensional vectors, whose numerical distribution implicitly reflects the mechanical properties of the ocular soft tissue. In a specific implementation, the capsular stability risk quantification unit 161 is used to: input the capsular compliance feature vector into a risk assessment model for scalar mapping to obtain the capsular instability index. That is, in order to map this vector into a single scalar index, a lightweight risk assessment model is deployed within the unit. This model can adopt a support vector machine (SVM) or logistic regression architecture, the core of which is to construct a decision hyperplane. During the offline training phase of the model, a large number of clinical samples labeled with postoperative complications (such as lens displacement, capsular contraction syndrome) are used. The model parameters are optimized through supervised learning, enabling the model to identify the feature vector patterns corresponding to high-risk capsules. Specifically, the unit inputs the capsular compliance feature vector into the model. For the logistic regression model, the calculation formula is: in, It is the feature vector of bag compliance. The weight vector obtained during training has a dimension of 1×16. It is a bias term. It is the sigmoid function. The output of this function is defined as the bag instability index. Its value is strictly limited to the range [0,1]. The closer the value is to 0, the more stable the capsule structure and the normal tension of the suspensory ligaments; the closer the value is to 1, the more relaxed the capsule, the greater the risk of suspensory ligament rupture, or the greater the tendency for fibrosis. For example, for a patient with mild lens tremor but long-term pseudo-exfoliation syndrome, the calculated eigenvector might be: =0.8, indicating a high risk of postoperative positional drift.

[0048] Subsequently, the data stream enters the comprehensive cost calculation and sorting unit 162. This unit first iterates through each crystal option i in the candidate list. The system has a built-in crystal attribute knowledge base, which automatically matches the corresponding position sensitivity coefficient based on the functional type of each crystal. This coefficient is a dimensionless constant that quantifies the sensitivity of the optical properties of this type of crystal to positional deviations. The specific rules are as follows: For multifocal / trifocal IOLs: due to their complex optical diffraction ring design, even small eccentricities can lead to severe coma and image quality degradation; therefore, the highest sensitivity coefficient is assigned, for example... =1.0. For extended depth-of-field IOLs (EDOF IOLs) or aspheric IOLs: their positional requirements are moderate, so the coefficient is set to 0.5. For spherical monofocal IOLs: their optical properties have good tolerance for tilt and eccentricity, so the coefficient is set to the lowest possible value, for example, 0.1. After matching the coefficients, the unit calculates the comprehensive cost function for each candidate crystal. This function innovatively unifies the mathematical formula for optical residual error and biomechanical risk penalty as follows: The first term is the absolute value (in diopters, D) of the estimated postoperative residual refractive power calculated by the dimensional discretized ray tracing unit. This represents the optical imperfection. The second term is the risk penalty term, in which... This is a weight hyperparameter used to balance this dimension; for example, if it is set to 1.0, This represents the expected potential risk of a specific crystal under specific patient pocket conditions. Specifically, it could be the pocket instability index calculated for the patient. =0.8 (high-risk pocket). The system has two candidate crystals: 1. Crystal A (multifocal): Optical calculations show a residual power of 0.05D (excellent optics), sensitivity coefficient = 1.0. 2. Crystal B (monofocal): Optical calculations show a residual power of 0.25D (moderate optics), sensitivity coefficient = 0.8 (high-risk pocket). =0.1. Calculate the overall cost: For crystal A: =0.05 + 1.0 × 1.0 × 0.8 = 0.85; For crystal B: =0.25 + 1.0 × 0.1 × 0.8 = 0.33. Even though lens A is far superior to lens B (0.25D) in a purely optical sense (0.05D), the patient's capsular area is extremely unstable (…). =0.8), and after calculation using the comprehensive cost function, the high-risk penalty (0.8) of lens A causes its total cost to soar to 0.85, far exceeding the total cost of lens B (0.33). This means that the system intelligently determines that for this patient, although implanting a multifocal lens theoretically provides better visual acuity, the risk is too high and not worth the effort; while implanting a monofocal lens, although resulting in some residual refractive error, is safer and more reliable. Finally, the unit, based on the calculated... The system sorts all candidate crystals in ascending order (lower cost is preferred). In the example above, crystal B (monofocal) will be ranked before crystal A (multifocal). The system ultimately generates and outputs this intelligently rearranged preferred ranking queue. This result is a comprehensive treatment recommendation that incorporates biomechanical safety considerations, truly embodying the patient-centered, safety-first approach to intelligent clinical decision-making.

[0049] In summary, the intelligent intraocular lens selection system 100 for cataract surgery based on ocular biomechanical parameters, as described in this application, aims to solve the problem of prediction deviation of effective lens position caused by neglecting the biomechanical characteristics of the eye in existing technologies. This system is not limited to the acquisition of traditional static geometric data such as axial length and corneal curvature, but further utilizes a deep convolutional neural network to extract texture features and perform dimensionality reduction mapping on anterior segment tomographic images, thereby resolving the capsular bag compliance feature vector characterizing the biomechanical properties of ocular soft tissue. Based on this, the geometric parameters of the patient's eye, capsular bag compliance features, and the mechanical and physical properties of the intraocular lens are combined. By introducing a physical constraint mechanism to simulate the mechanical equilibrium state after lens implantation, the effective lens position can be accurately predicted. This coupled analysis model of capsular biomechanics and lens mechanical properties overcomes the limitations of pure geometric optics, effectively addressing the adaptation challenges of non-standard eyeballs or special lens models. Through ray tracing and stability assessment, the optimal lens solution is output, significantly improving the accuracy and stability of postoperative refractive prediction.

Claims

1. A smart intraocular lens selection system for cataract surgery based on ocular biological parameters, characterized in that, include: The surgical data acquisition module is used to acquire the subject's raw biometric data stream, anterior segment tomographic image sequence, and intraocular lens specification data; The surgical data preprocessing module is used to perform signal quality verification, region of interest cropping, and specification parameter extraction on the raw biometry data stream, anterior segment tomographic image sequence, and intraocular lens specification data to obtain geometric parameter vectors, region of interest image tensors, and mechanical parameter vectors. The texture feature extraction and dimensionality reduction mapping module is used to input the image tensor of the region of interest into a pre-trained deep convolutional neural network for texture feature extraction and dimensionality reduction mapping to obtain the bag compliance feature vector characterizing the biomechanical properties of ocular soft tissue. The effective position prediction module is used to predict the effective position of the crystal under physical constraints by analyzing the geometric parameter vector, mechanical parameter vector, and bag compliance feature vector. The crystal candidate module is used to perform optical ray tracing and degree quantization on the geometric parameter vector and the predicted effective crystal positions to obtain a list of candidate crystals containing degree information. The preferred sorting queue generation module is used to perform stability evaluation and preferred sorting of the candidate crystal list based on the bag compliance feature vector to obtain the preferred sorting queue.

2. The intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters according to claim 1, characterized in that, The surgical data preprocessing module includes: The biometric data preprocessing unit is used to clean and standardize the raw biometric data stream to obtain a geometric parameter vector. The region of interest extraction unit is used to filter keyframes from the anterior segment tomographic image sequence and extract the region of interest from the keyframes to obtain the region of interest image tensor. The intraocular lens specification data retrieval unit is used to retrieve intraocular lens specification data by using a pre-selected model as an index to obtain the pre-selected intraocular lens specification data. The mechanical parameter vector generation unit is used to extract static physical properties and compressive force-displacement curve data from pre-selected artificial crystal specification data, and to serially stitch the discretized curve features to obtain the mechanical parameter vector.

3. The intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters according to claim 1, characterized in that, The texture feature extraction and dimensionality reduction mapping module includes: The texture feature encoding unit is used to input the image tensor of the region of interest into the backbone layer of the deep convolutional neural network to perform multi-scale texture feature encoding to obtain a high-dimensional feature map. Spatial topology aggregation unit, used to perform global spatial topology aggregation on high-dimensional feature maps to obtain global feature vectors; The mechanical manifold embedding mapping unit is used to perform mechanical manifold embedding and dimensionality reduction mapping on the global feature vector to obtain the bag compliance feature vector.

4. The intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters according to claim 1, characterized in that, The effective location prediction module includes: The multi-physics feature fusion unit is used to perform multi-physics feature alignment and fusion on geometric parameter vectors, mechanical parameter vectors and bag compliance feature vectors to obtain panoramic feature vectors. The mechanical equilibrium constraint inference unit is used to input the panoramic feature vector into a physical perception regression network with an embedded mechanical equilibrium loss function to perform deep inference based on mechanical equilibrium constraints to obtain normalized prediction values. The physical space mapping and restoration unit is used to perform physical space mapping and numerical restoration on the normalized predicted values ​​based on preset statistical distribution parameters to obtain the predicted effective crystal position.

5. The intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters according to claim 1, characterized in that, The effective location prediction module includes: The multi-physics feature fusion unit is used to perform multi-physics feature alignment and fusion on geometric parameter vectors, mechanical parameter vectors and bag compliance feature vectors to obtain panoramic feature vectors. The neural implicit solver unit is used to perform neural implicit solver on the panoramic feature vector based on the energy potential surface to obtain normalized prediction values. The physical space mapping and restoration unit is used to perform physical space mapping and numerical restoration on the normalized predicted values ​​based on preset statistical distribution parameters to obtain the predicted effective crystal position.

6. The intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters according to claim 5, characterized in that, The neural implicit solver unit is used for: The capillary mechanical parameters are mapped onto the panoramic feature vector to obtain the capillary nonlinear elastic parameter vector; The total potential energy of the system is calculated based on the known mechanical parameter vector and the mechanical parameters of the bag to obtain the total potential energy equation; The total potential energy equation is solved implicitly to obtain normalized predictions.

7. The intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters according to claim 1, characterized in that, The crystal candidate module includes: The virtual eye optical model data construction unit is used to reconstruct the optical surface coordinate system using the curvature and axis length data in the geometric parameter vector, and to load the crystal refractive index parameter at the predicted effective crystal position to construct the virtual eye optical model data. The ideal refractive power generation unit is used to perform zero-point iterative solving based on ray tracing on the incident virtual eye optical model data to obtain the theoretical ideal refractive power; The specification discretization ray tracing unit is used to map discretized specifications centered on the theoretical ideal refractive power and perform forward ray tracing on each specification to calculate the equivalent spherical residual relative to the retina to obtain a list of candidate lenses.

8. The intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters according to claim 1, characterized in that, The preferred sorting queue generation module includes: The bag stability risk quantification unit is used to quantify the bag stability risk of the bag compliance feature vector to obtain the bag instability index. The comprehensive cost calculation and sorting unit is used to match the position sensitivity coefficients for different crystal types in the candidate crystal list, and calculate the comprehensive cost function by combining the bag instability index and optical residual error, and sort them in ascending order of comprehensive cost to obtain the preferred sorting queue.

9. The intelligent intraocular lens selection system for cataract surgery based on ocular biological parameters according to claim 8, characterized in that, The bag stability risk quantification unit is used to: input the bag compliance feature vector into the risk assessment model for scalar mapping to obtain the bag instability index.

Images