Implant Generation Method and System Based on CT Imaging and Generative AI

By using intelligent preprocessing and multi-target synchronous segmentation based on CT images, combined with generative AI-generated implants, the problems of long time consumption and low accuracy in the design of personalized implants in existing technologies are solved. This achieves efficient and accurate generation of personalized implants that meet the requirements of anatomical fit, mechanical compatibility and manufacturability.

CN123118950APending Publication Date: 2026-11-17NORTHWESTERN POLYTECHNICAL UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611228249.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-13
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing personalized implant design methods are time-consuming and have low precision, and cannot simultaneously meet the requirements of anatomical fit, mechanical adaptability and manufacturability. Furthermore, they lack a unified mechanism for dynamic physiological strain and manufacturing process constraints.

Method used

By using intelligent preprocessing based on CT images, multi-target synchronous segmentation, three-dimensional surface reconstruction and personalized design constraint analysis, combined with generative AI to generate implants, the entire process from CT images to the final implant model is automatically generated, including the introduction of multi-task loss functions and the use of biomechanical proxy models.

Benefits of technology

It enables high-precision and rapid generation of personalized implants, ensuring anatomical fit, mechanical compatibility and manufacturability, reducing error accumulation, and improving design controllability and clinical compliance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118950A_ABST
    Figure CN123118950A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for implant generation based on CT images and generative AI, belonging to the field of medical image processing and personalized implant design technology. The method includes: acquiring a region of interest (ROI) image based on CT image data of the target anatomical site; determining a segmentation mask based on the ROI image, including at least the segmentation mask for the target anatomical structure, the defect area, and the cross-section of the bone fragment; reconstructing a three-dimensional surface model based on the segmentation mask, calculating the geometric parameters of the defect area and the mechanical contact parameters of the stump cross-section, and determining personalized implant design constraints by combining prior pathological diagnosis results and physiological parameters; extracting the defect boundary curve and stump cross-section contact surface data from the three-dimensional surface model, generating an external contour model, and generating a microscopic topology and lattice structure within its spatial envelope to obtain an initial implant model; validating the initial implant model, and determining the final implant model based on the validation results. This invention achieves integrated collaborative optimization of the entire process from CT images to implant models, improving the accuracy and efficiency of implant design.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of medical image processing and personalized implant design technology, specifically to an implant generation method and system based on CT images and generative AI. Background Technology

[0002] The key to personalized implant design lies in achieving a deep fit between the implant and the individual's anatomical features based on the patient's CT images. Current clinical practice involves three steps: image segmentation, 3D reconstruction, and implant design. Image segmentation relies on manual layer-by-layer drawing, which takes several hours and suffers from blurred boundaries due to experience differences. 3D reconstruction requires repeated manual repair of geometric defects such as ruptures and holes. Implant design employs a trial-and-error approach, making it difficult to reproduce individualized anatomical features and focusing only on static mechanical indicators, failing to adapt to the cyclical strain of dynamic tissues such as the chest wall, and easily leading to fatigue failure. Furthermore, each step is independent; insufficient segmentation accuracy will be transmitted to the reconstructed model, and geometric defects in the reconstructed model will amplify morphological deviations in the implant design. The cumulative error between steps severely affects the final implant fitting accuracy.

[0003] In recent years, although deep learning technology has been introduced into the automatic segmentation of bone CT images, it only covers a single stage, and the AI ​​models for each stage are trained and infer independently, lacking a unified feature representation and data flow coordination mechanism. If the independent models are simply spliced ​​together, the mask boundary error output by the segmentation network will be amplified in the 3D reconstruction stage. The topological defects of the reconstruction model will further interfere with the spatial constraints of generative design, making it difficult to achieve a balance between anatomical fit, mechanical adaptability, and manufacturability in the final implant. At the same time, existing AI methods lack a mechanism to directly embed design constraints (such as dynamic physiological strain range and additive manufacturing process limitations) into the generation process. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides an implant generation method and system based on CT images and generative AI. By performing intelligent preprocessing, multi-target synchronous segmentation, three-dimensional surface reconstruction and personalized design constraint analysis on the patient's CT images, and macro-micro collaborative implant structure creation, the entire process of integrated automatic generation from CT images to the final implant model is realized.

[0005] This invention is achieved through the following technical solution: Secondly, this application provides a method for implant generation based on CT images and generative AI, including the following steps: Step 1: Obtain the region of interest image based on CT image data of the target anatomical site; Step 2: Determine the segmentation mask based on the region of interest image; The segmentation mask includes at least a segmentation mask for the target anatomical structure, a segmentation mask for the defect area, and a segmentation mask for the cross-section of the bone fracture end; Step 3: Reconstruct a three-dimensional surface model of the target anatomical site based on the segmentation mask, calculate the geometric parameters of the defect area and the mechanical contact parameters of the stump section based on the model, and determine the personalized implant design constraints in combination with the prior pathological diagnosis results and corresponding physiological parameters. Step 4: Extract the boundary curve of the defect area and the contact surface data of the residual section from the three-dimensional surface model, generate an outline model that fits the remaining anatomical structure, and generate a micro-topology and lattice structure within the spatial envelope defined by the outline model to obtain the initial model of the implant. Step 5: Verify the initial implant model and determine the final implant model based on the verification results.

[0006] Preferably, in step 1, acquiring the region of interest image based on CT image data of the target anatomical site includes: The CT image data is sequentially denoised and artifact corrected to obtain the corrected CT image. The region of interest is extracted from the corrected CT image to obtain the region of interest image.

[0007] Preferably, determining the segmentation mask based on the region of interest image includes: The image of the region of interest is input into a trained multi-target segmentation network to obtain segmentation masks for the target anatomical structure, the defect area, and the cross-section of the bone fracture.

[0008] Preferably, in step 3, reconstructing the three-dimensional surface model of the target anatomical region based on the segmentation mask includes: The segmentation mask is converted into a three-dimensional surface model using the moving cube algorithm; The Poisson reconstruction algorithm is used to automatically repair the holes and broken surfaces in the three-dimensional surface model, so as to obtain a three-dimensional surface model with complete water tightness.

[0009] Preferably, in step 3, the geometric parameters of the defect region include the defect volume, surface area, and average curvature; the mechanical contact parameters of the residual section include the macroscopic contact area and the coefficient of variation of curvature; and the design constraints of the personalized implant include the target volume fraction and the expected physiological dynamic strain range.

[0010] Preferably, in step 4, generating an external contour model that conforms to the remaining anatomical structure includes: A shape fitting network is constructed, and a shape contour model is generated by taking the boundary curve of the defective region and the contact surface data of the residual section as input; the shape contour model serves as the maximum spatial envelope design boundary constraint for generating the micro-topology and lattice structure.

[0011] Preferably, in step 4, generating the microscopic topology and lattice structure within the spatial envelope defined by the outline model includes: Construct a generative artificial intelligence model, which includes a forward diffusion process and a reverse denoising process with an embedded bidirectional attention mechanism module; The personalized implant design constraints are input into the trained generative artificial intelligence model, and implant structure data is generated through a reverse denoising process. The implant structure data is post-processed to obtain the initial model of the implant.

[0012] Preferably, during the training process of the generative artificial intelligence model, a biomechanical proxy model based on a graph neural network architecture is introduced, using a multi-task loss function to integrate three constraints: static compliance, dynamic strain deviation, and manufacturing feasibility; the multi-task loss function is expressed as:

[0013] in: The primary objective is to minimize the loss term related to structural static compliance; To dynamically adapt the target loss term; For manufacturing-related loss items; , , Hyperparameters are used to balance the importance of various losses.

[0014] Preferably, the step of validating the initial implant model and determining the final implant model based on the validation results includes: The geometric accuracy of the initial model of the implant is verified, and the Dice similarity coefficient and Hausdorff distance between it and the three-dimensional surface model are calculated. The initial model of the implant was verified by mechanical simulation. A simulated physiological dynamic cyclic alternating load spectrum was applied to calculate the maximum stress and dynamic strain deviation inside the model. The optimal implant model is determined based on the Dice similarity coefficient, Hausdorff distance, maximum stress, and dynamic strain deviation. The optimal implant model is then locally geometrically reconstructed based on the geometric deviation heatmap and the stress distribution results of the mechanical simulation to obtain the final implant model.

[0015] Secondly, this application provides an implant generation system based on CT images and generative AI, comprising: The preprocessing module is used to acquire region-of-interest images based on CT image data of the target anatomical site; A segmentation module is used to determine a segmentation mask based on the region of interest image; the segmentation mask includes at least a segmentation mask for the target anatomical structure, a segmentation mask for the defect area, and a segmentation mask for the cross-section of the bone fracture end; The reconstruction and analysis module is used to reconstruct a three-dimensional surface model of the target anatomical site based on the segmentation mask, calculate the geometric parameters of the defect area and the mechanical contact parameters of the stump section based on the model, and determine the personalized implant design constraints in combination with the prior pathological diagnosis results and corresponding physiological parameters. The design generation module is used to extract the boundary curve of the defect area and the contact surface data of the residual section from the three-dimensional surface model, generate the outline model that fits the remaining anatomical structure, generate the micro-topology and lattice structure within the spatial envelope defined by the outline model, and obtain the initial model of the implant. The verification and interaction module is used to verify the initial model of the implant and determine the final implant model based on the verification results.

[0016] Compared with the prior art, the present invention has the following beneficial technical effects: This application presents an implant generation method based on CT images and generative AI. First, denoising and artifact correction preprocessing of CT images effectively suppresses radial artifact interference caused by metal internal fixation devices, restoring the true anatomical structure of artifact-occluded areas. Second, a multi-target segmentation network is used to simultaneously extract segmentation masks of target anatomical structures, defect areas, and bone fragment sections from the region of interest image. This avoids the experience differences and boundary ambiguities caused by manual layer-by-layer delineation in traditional methods, while also overcoming the error accumulation caused by step-by-step segmentation. Then, a watertight 3D surface model is reconstructed based on the segmentation mask, calculating the geometric parameters of the defect area and the mechanical contact parameters of the stump section. Combined with prior pathological diagnosis results and patient physiological parameters, personalized implant design constraints are determined, bridging the semantic gap between segmentation results and design inputs, and realizing the transformation of design parameters from generalized to personalized. Furthermore, defect boundary curves and contact surface data of the residual section are extracted from the 3D surface model. First, an external contour model that conforms to the remaining anatomical structure is generated. Then, the external contour model is used as a spatial envelope boundary constraint to generate microscopic topology and lattice structures. This ensures the coordinated optimization of macroscopic shape and microscopic structure, avoiding the problem of shape deviations being amplified into internal structures in traditional trial-and-error models. Finally, the initial implant model is evaluated through a dual mechanism of geometric accuracy verification and mechanical simulation verification. The optimal model is selected using objective quantitative indicators, ensuring the clinical compliance and design controllability of the output results.

[0017] This application also proposes an implant generation system based on CT images and generative AI, an electronic device, and a computer storage medium, which possess all the advantages of the aforementioned implant generation method based on CT images and generative AI. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the implant generation method based on CT images and generative AI of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0022] See Figure 1 An implant generation method based on CT images and generative AI includes the following steps: Step 1: Obtain the original CT image data of the patient's target anatomical location and perform preprocessing to obtain the optimized region of interest image.

[0023] In some embodiments, the preprocessing method in step 1 is as follows: S1.1 Acquire raw CT image data of the patient's target anatomical location, wherein the raw CT image data is in DICOM format; S1.2. The original CT image data is denoised to preserve the details of the anatomical edges and defect boundaries while suppressing the scanning noise introduced during image acquisition, thus obtaining a denoised CT image. S1.3 Perform artifact correction on the denoised CT images to eliminate radial artifacts caused by metal internal fixation devices, restore the clarity of anatomical structures in the artifact-covered areas, and obtain the corrected CT images. S1.4 Extract the region of interest from the corrected CT image and remove irrelevant background areas to obtain the optimized region of interest image.

[0024] In other embodiments, in S1.2, the denoising process employs at least one of the following: nonlocal mean filtering algorithm, three-dimensional block matching filtering algorithm, anisotropic diffusion filtering algorithm, deep learning-based image denoising network, or wavelet transform denoising method.

[0025] In a further embodiment, in S1.3, the artifact correction employs at least one of the following: a metal artifact correction algorithm based on dual-energy CT, a metal artifact correction network based on deep learning, a metal artifact correction method based on the projection domain, or a linear interpolation method based on a sine graph.

[0026] In a further embodiment, in S1.4, the region of interest extraction is based on automatically detecting the thoracic area using rib contour features and then cropping the boundary, or using a deep learning-based object detection network to automatically identify and crop the region of interest, or using an interactive segmentation method based on a region growing algorithm or a graph cut algorithm.

[0027] By establishing a processing procedure from raw CT images to optimized regions of interest (ROI) images, high-quality, low-redundancy input data is provided for the segmentation network. Denoising processing employs algorithms such as nonlocal mean filtering to suppress scanning noise while maintaining the clarity of anatomical edges, avoiding the edge blurring problem of traditional Gaussian filtering methods. Artifact correction specifically addresses radial artifacts caused by metal internal fixation devices, restoring the true anatomical structure in artifact-occluded areas. ROI extraction significantly reduces the computational resource consumption for subsequent deep learning network inference by automatically identifying and cropping irrelevant background regions.

[0028] Step 2: Input the region of interest image into the trained segmentation network for forward inference calculation to obtain a multi-target segmentation mask.

[0029] In some embodiments, the method for obtaining the multi-target segmentation mask in step 2 is as follows: S2.1 Construct a multi-target segmentation network, which is used to simultaneously segment the target anatomical structure, the defect area, and the cross-section of the bone fracture from the input CT image; S2.2 Using a CT image dataset labeled with healthy bone regions, cartilage boundaries, tissue defect regions, and the upper and lower edges of bone fracture sections, the segmentation network is iteratively trained until convergence to obtain a trained segmentation network. S2.3. Input the region of interest image obtained in step 1 into the trained segmentation network for forward inference calculation to obtain a multi-target segmentation mask; The segmentation mask includes at least a segmentation mask for the target anatomical structure, a segmentation mask for the defect area, and a segmentation mask for the cross-section of the bone fracture.

[0030] In a further embodiment, the segmentation network is a segmentation network based on an improved U-Net++ architecture. This improved U-Net++ architecture introduces an attention mechanism module and a multi-scale feature fusion module on the basis of standard U-Net++ to enhance the network's ability to recognize anatomical structures and local boundary features of different sizes.

[0031] The Dice similarity coefficient of the trained network output mask is greater than or equal to 0.95 compared with the gold standard mask manually annotated by radiology experts. By combining sub-pixel 3D interpolation algorithm, the Hausdorff distance of its surface geometric boundary is controlled within an accuracy range of less than or equal to 0.5 mm.

[0032] As a preferred approach, the training dataset can also be augmented by data augmentation operations such as random rotation, scaling, elastic deformation, and grayscale perturbation to generate augmented samples.

[0033] By constructing a multi-target synchronous segmentation network, end-to-end inference from a single CT image input to multi-class anatomical structure mask output was achieved. By introducing an attention mechanism module and a multi-scale feature fusion module on top of U-Net++, the network's ability to synchronously recognize anatomical structures of different sizes was enhanced. Regarding accuracy control methods, a sub-pixel 3D interpolation algorithm was used to overcome the physical resolution limitation of the inherent slice thickness of CT, providing reliable geometric boundary data for subsequent accurate 3D reconstruction and implant design.

[0034] Step 3: Reconstruct a watertight three-dimensional surface model of the target anatomical site based on the segmentation mask, calculate the geometric and mechanical boundary parameters of the anatomical remnants based on the watertight three-dimensional surface model, and determine the design constraints of the implant by combining the patient's physiological parameters.

[0035] In some embodiments, the method for obtaining the design constraints in step 3 is as follows: S3.1. Based on segmentation masking, convert the anatomical structure mask and the defect area mask into a three-dimensional surface model; S3.2 Automatically repair the holes and broken surfaces in the three-dimensional surface model to obtain a three-dimensional surface model with complete water tightness; S3.3. Based on the repaired watertight three-dimensional surface model, calculate the macroscopic geometric parameters of the defect area, and calculate the mechanical contact parameters of the residual section based on the fracture interface mask. S3.4 Input the prior pathological diagnosis results, the geometric parameters, and the mechanical contact parameters into the constraint solution model; the constraint solution model determines the design constraints of the implant based on the prior classification results and the patient's specific physiological parameters.

[0036] The prior pathological diagnosis result is a diagnostic conclusion determined by the clinician before surgery based on the patient's medical history, imaging examinations and pathological analysis. The diagnostic conclusion is used to indicate whether the defect type is a traumatic defect or a tumor defect.

[0037] The design constraints constitute the multi-source raw data input for the subsequent generative artificial intelligence model, including the target volume fraction, the desired physiological dynamic strain range, the maximum allowable deformation of the implant, fatigue life requirements, or stress shielding control parameters.

[0038] In one possible implementation, the implant design constraints may also include a bone integration interface micromotion threshold, a stress shielding rate upper limit, or a vascularized pore connectivity requirement.

[0039] In other embodiments, at least one of the moving cube algorithm, the moving tetrahedron algorithm, or the direct volume rendering method is used in S3.1 to convert the segmentation mask into a three-dimensional surface model in STL format.

[0040] For tumor defects, the cross-section of the bone fracture ends is determined by the boundary of the virtual osteotomy surface planned preoperatively; For traumatic defects, the cross-section of the fracture ends is directly determined by the boundary of the original fracture remnant.

[0041] In some embodiments, S3.2 employs one of the Poisson reconstruction algorithm, radial basis function interpolation method, or least squares-based mesh completion method to close holes and repair broken surfaces in the three-dimensional surface model.

[0042] In some embodiments, the macroscopic geometric parameters in S3.3 include the defect volume, surface area, and average curvature; the mechanical contact parameters include the macroscopic contact area of ​​the residual section and the curvature variation coefficient characterizing the geometric complexity of its cross-section.

[0043] By establishing a design link from segmentation mask to design constraints, the automatic conversion from medical images to engineering design parameters was achieved. First, a traveling cube algorithm combined with Poisson reconstruction was used to transform the discrete segmentation mask into a watertight continuous surface model, resolving the inter-slice staircase effect and porosity issues caused by CT slice thickness limitations. Second, the macroscopic geometric features of the defect area and the mechanical contact conditions of the remnant were comprehensively quantified. Finally, prior pathological diagnostic results were introduced as a classification basis, combined with patient-specific physiological parameters, to adaptively output differentiated design constraints.

[0044] Step 4: Based on the boundary data of the defect area, the cross-sectional data of the residual end, and the design constraints of the implant, a macroscopic shape contour model of the implant is first generated using a two-level generative network. Then, the optimal microscopic topology and gradient lattice structure are generated within the spatial envelope defined by the shape contour model to obtain an initial implant model containing macroscopic shape and microscopic structure parameters.

[0045] In some embodiments, the method for generating the initial model of the implant in step 4 is as follows: S4.1 Extract the boundary curves of the defective area and the contact surface data of the residual section from the repaired watertight three-dimensional surface model; S4.2 Based on the boundary curves and contact surface data, generate an external contour model that macroscopically fits the remaining anatomical structure, which serves as the maximum spatial envelope design boundary constraint for the generative model. S4.3 Construct a generative artificial intelligence model based on an improved topological diffusion model. This generative artificial intelligence model includes a forward diffusion process and a reverse denoising process with an embedded bidirectional attention mechanism module. S4.4 Perform multi-task training with physical constraints on the generative artificial intelligence model to obtain a trained model; input the personalized implant design constraints output in step 3 into the trained model, and generate multiple candidate implant structure data through a reverse denoising process.

[0046] S4.5. Post-process the implant structure data to obtain the corresponding initial model of the implant.

[0047] In a further embodiment, in S4.2, the shape contour model is generated by a shape fitting network, which is a model based on a variational autoencoder, a generative adversarial network, or a deep learning regression network, and the resulting fitting gap is controlled to be less than or equal to 0.2 mm.

[0048] As an alternative, the shape fitting network can also employ an iterative nearest-point algorithm based on point cloud registration combined with a free deformation method or a shape completion network based on implicit neural representation.

[0049] In some embodiments, S4.4, the physically constrained multi-task training includes: First, data from multiple historical clinical cases were collected to extract geometric boundary conditions such as defect location, area, and geometry. Then, human biomechanical simulation software was used to simulate periodic load spectra under specific physiological activities, defining dynamic mechanical environment parameters. Manufacturing process constraints, including minimum wall thickness and maximum allowable suspended material ratio for additive manufacturing, were set. Initial topological or lattice structure samples were generated in batches using a solid isotropic material penalty model or progressive structural optimization method. Their static compliance, stress, and periodic dynamic strain time history curves were calculated through finite element simulation, and manufacturing feasibility labels were assigned. A multimodal training dataset was then constructed through structured correlation. Secondly, a multi-task loss function L that integrates physical and manufacturing constraints is designed and minimized. A biomechanical AI proxy model based on a graph neural network architecture is introduced, which pre-learns and solidifies the end-to-end mapping from structural morphology to the three-dimensional strain field offline. During backpropagation training, this proxy model predicts the various loss values ​​in the multi-task loss function in real time. The multi-task loss function is expressed as follows:

[0050] In the formula: This is the primary objective loss term used to minimize the static compliance of the structure; To dynamically adapt the target loss term, it is used to control the deviation between the predicted microscopic dynamic strain distribution of the structure during the physiological cycle and the expected physiological strain range, which does not exceed a preset threshold. This is a manufacturing target loss item used to penalize unmanufacturable behaviors such as violating minimum wall thickness, exceeding process limit overhang angles, or the presence of suspended materials. , , To balance the importance of various loss terms, the weights of each loss term are automatically adjusted using a dynamic dimensional normalization algorithm; and the network parameters are iteratively updated using an adaptive moment estimation optimizer until convergence. Then, the specific personalized constraints of the current target patient output in step 3 are used as guiding conditions to input the completed generative artificial intelligence model. The model performs joint spatial evolution of macroscopic contour boundary density and internal porous structure within the step size through the reverse denoising process, and quickly infers and generates multiple candidate implant structure data.

[0051] In one possible implementation, in S4.5, the post-processing includes: The implant structure data is sequentially subjected to isolated voxel removal, smoothing filtering, and parametric transformation to eliminate non-physical micro-material islands or pores, convert voxelized stepped boundaries into smooth continuous surfaces, and convert the processed structure data into an editable computer-aided design format.

[0052] Optionally, the computer-aided design format is at least one of STEP, IGES, STP, or STL formats.

[0053] In a further embodiment, the internal structure of the initial implant model obtained after post-processing includes at least one of the following: a scheme for spatial gradient optimization of a spring-like lattice structure, a biomimetic honeycomb-spring composite structure, or a three-period minimum surface structure with functional gradient.

[0054] A highly fitting external contour is inferred directly from defect boundary data using variational autoencoders or generative adversarial networks. A biomechanical surrogate model based on graph neural networks is introduced, reducing the computation time of traditional finite element simulation from hours to seconds, enabling the diffusion model to perceive the mechanical response of the structural morphology in real time during training and inference. Static compliance, dynamic strain deviation, and manufacturing feasibility constraints are integrated into the multi-task loss function to ensure that the generated structure meets mechanical performance requirements while possessing additive manufacturing feasibility.

[0055] Step 5: Perform automatic geometric accuracy verification and mechanical simulation verification on multiple candidate implant initial models, and determine the final implant model based on the verification results.

[0056] In some embodiments, the method for obtaining the final implant model in step 5 is as follows: S5.1 For the multiple candidate implant initial models generated in step 4, calculate the Dice similarity coefficient and Hausdorff distance between each candidate implant initial model and the repaired watertight three-dimensional surface model. S5.2 Apply simulated physiological dynamic cyclic alternating load spectrum to the multiple candidate implant initial models generated in step 4, and calculate the maximum stress and dynamic strain deviation inside each candidate implant initial model. Furthermore, high-precision finite element analysis or combined equivalent static analysis and multiphysics coupling analysis are used for mechanical simulation; the maximum dynamic alternating stress value is required to be less than or equal to the high-cycle fatigue limit of the implant material under a specific stress ratio.

[0057] S5.3 Select the optimal implant model from the multiple candidate implant initial models, which has the highest Dice similarity coefficient, the smallest Hausdorf distance, and the maximum stress that meets the preset safety threshold; render the optimal implant model, geometric deviation heatmap, and mechanical simulation stress distribution results through a three-dimensional visualization interactive interface; As an example, the Dice similarity coefficient is required to be greater than or equal to 0.98 and the Hausdorff distance is required to be less than or equal to 0.2 mm.

[0058] S5.4. Based on the clinician's fine-tuning instructions, perform local geometric reconstruction on the optimal implant model, and perform transient mechanical re-verification and finite element simulation verification on the reconstructed implant model. Once the verification is successful, the final implant model is obtained.

[0059] Furthermore, the fine-tuning instructions include at least one of the following: edge rounding adjustment, local thickening, local rod diameter or porosity fine-tuning, edge chamfering or fixed hole position adjustment; The transient mechanical re-verification is controlled within 1 minute.

[0060] By performing dual verification through geometric similarity and fatigue mechanical simulation on multiple candidate implant initial models, the optimal implant model is automatically selected based on the highest morphological fit and best mechanical performance. A biomechanical surrogate model is introduced for transient re-verification, ensuring that each fine-tuning operation by the physician receives mechanical feedback within one minute, avoiding the hours-long waiting time required by traditional finite element simulations and achieving a highly efficient human-machine collaborative optimization closed loop. The model verified by high-precision finite element analysis is used as the final implant model to meet the mechanical safety standards for clinical implantation.

[0061] Step 6: Convert the final implant model into a physical format recognizable by the additive manufacturing equipment and transmit it to the printing equipment. Based on the set material process parameters, perform solid printing and post-processing manufacturing.

[0062] In some embodiments, step 6, the method for printing the entity, is as follows: S6.1 Convert the final implant model into standard STL format or 3D manufacturing format; S6.2. Wirelessly transmit the converted model physical file to the additive manufacturing equipment, and adapt and set the additive manufacturing process parameters for the selected medical high-performance engineering material. Furthermore, the medical high-performance engineering material includes at least one of pure titanium, titanium alloy, cobalt-chromium alloy, or biodegradable metal material; The additive manufacturing process parameters include laser power or electron beam power, scanning speed, powder thickness, and scanning strategy.

[0063] S6.3 Based on additive manufacturing process parameters, selective laser melting, electron beam powder bed melting, or binder jetting are used to perform solid 3D printing to obtain a personalized metal implant skeleton physical entity. As an example, the laser power of the selective laser melting process is 150 to 250 watts, the scanning speed is 800 to 1200 millimeters per second, and the powder layer thickness is 20 to 40 micrometers.

[0064] S6.4. Standardize and process the printed physical entity to obtain a final implantable product that can be used clinically. The standardized post-processing includes: placing the physical entity in a high-vacuum forming furnace for hot isostatic pressing to eliminate residual stress and microscopic internal pores caused by forming, thereby obtaining a dense microstructure; subsequently, sandblasting the hot isostatically pressed physical entity to remove unmelted powder particles adhering to the surface; then, anodizing or micro-arc oxidation surface treatment is performed on the sandblasted physical entity to improve surface biocompatibility; finally, sterilization is performed on the surface-treated physical entity to obtain a final implantable product that can be used clinically.

[0065] Preferably, the pressure of the hot isostatic pressing is set to 100 to 140 MPa, the temperature is set to 900 to 930 degrees Celsius, and the holding time is not less than 2 hours.

[0066] In an optional embodiment, the following steps are also included: S6.5. Extract solid samples from the same batch for standard mechanical and physical performance testing to obtain the measured breaking strength, elastic modulus and strain retention rate after fatigue; combine the real physical test data, the original simulation prediction data and the corresponding personalized input conditions into a closed-loop feedback sample, and asynchronously import it into the offline accumulated multimodal physical experiment database for periodic batch incremental training or weight evolution update of generative artificial intelligence model and mechanical agent model.

[0067] The final implant model is converted into a physical format directly usable by additive manufacturing equipment and transmitted wirelessly to the printing equipment, avoiding accuracy loss during data conversion and transmission. Residual stress and internal porosity caused by rapid solidification during selective laser melting are eliminated through hot isostatic pressing. Combined with sandblasting, surface oxidation, and sterilization, this ensures that the physical implant meets the mechanical performance and biosafety standards for clinical implantation.

[0068] Correspondingly, this application also provides an implant generation system based on CT images and generative AI, including: The preprocessing module is used to acquire region-of-interest images based on CT image data of the target anatomical site; A segmentation module is used to determine a segmentation mask based on the region of interest image; the segmentation mask includes at least a segmentation mask for the target anatomical structure, a segmentation mask for the defect area, and a segmentation mask for the cross-section of the bone fracture end; The reconstruction and analysis module is used to reconstruct a three-dimensional surface model of the target anatomical site based on the segmentation mask, calculate the geometric parameters of the defect area and the mechanical contact parameters of the stump section based on the model, and determine the personalized implant design constraints in combination with the prior pathological diagnosis results and corresponding physiological parameters. The design generation module is used to extract the boundary curve of the defect area and the contact surface data of the residual section from the three-dimensional surface model, generate the outline model that fits the remaining anatomical structure, generate the micro-topology and lattice structure within the spatial envelope defined by the outline model, and obtain the initial model of the implant. The verification and interaction module is used to verify the initial model of the implant and determine the final implant model based on the verification results.

[0069] Example 1 This embodiment uses the repair of a large-area rib defect caused by trauma to the left chest wall as an application scenario to provide a detailed description of the implant generation method based on CT images and generative AI provided by the present invention.

[0070] Step 1: Obtain the original CT image data of the patient's target anatomical location and perform intelligent preprocessing to obtain optimized region of interest images.

[0071] S1.1, The patient's chest raw CT image data was acquired using multi-slice spiral CT with a slice thickness of 0.625 mm. The raw CT image data was in DICOM format. The patient had severe bone tissue defects in the left 3rd to 6th rib area due to trauma, and strong radial artifacts were present in the conventional CT scan images due to the placement of a temporary internal fixation plate for the fracture during previous surgery.

[0072] S1.2, The acquired raw CT image data is subjected to adaptive denoising processing using a nonlocal mean filtering algorithm. While preserving the details of anatomical edges and defect boundaries, the scanning noise introduced during image acquisition is suppressed to obtain denoised CT images.

[0073] S1.3, the denoised CT images are corrected using a deep learning-based metal artifact correction network to eliminate radial artifacts caused by the metal internal fixation plate, restore the clarity of the anatomical structure in the artifact-covered area, and obtain the corrected CT images.

[0074] S1.4, Based on the macroscopic anatomical features of the ribs, the region of interest of the 3rd to 6th ribs on the left side is automatically identified and cut out, irrelevant background areas are removed, and an optimized high-quality, high signal-to-noise ratio three-dimensional image block is output as the optimized region of interest image.

[0075] Step 2: Input the optimized region of interest image into the trained segmentation network for forward inference calculation to obtain a multi-target segmentation mask.

[0076] S2.1 The segmentation network is a 5-layer improved U-Net++ architecture segmentation network. The improved U-Net++ architecture introduces a dense cross-attention mechanism module between each layer to enhance the network's ability to recognize anatomical structures of different sizes and local boundary features.

[0077] S2.2 uses a dataset of CT images labeled with healthy rib areas, sternal boundaries, cartilage boundaries, tissue defect areas, and the upper and lower edges of bone fracture sections to iteratively train the segmentation network until convergence, resulting in a well-trained segmentation network. The Dice similarity coefficient of the trained network's output mask reaches 0.96 compared to the gold standard mask manually annotated by radiologists. Through sub-pixel 3D interpolation, the reconstructed digital boundary geometric overlap Hausdorff distance is 0.28 mm.

[0078] In step S2.3, the optimized region of interest image is input into the trained segmentation network for forward inference calculation. During forward inference, the network automatically and synchronously delineates independent masks for healthy ribs, sternum, cartilage boundaries, and bone fracture sections, obtaining a multi-target segmentation mask. The entire inference process takes approximately 3 minutes, significantly reducing the traditional manual delineation time of several hours.

[0079] Step 3: Reconstruct the watertight three-dimensional surface model of the target anatomical site based on the segmentation mask, calculate the geometric and anatomical remnant mechanical boundary parameters based on the model, and output the implant design constraints in combination with the patient's specific physiological parameters.

[0080] S3.1, based on the segmentation mask, the moving cube algorithm is used to convert the bone and defect boundary mask into a 3D surface model in STL format. The isosurface threshold of the moving cube algorithm is set to 128. In this example, the defect is a traumatic defect, and its fracture end section is directly determined by the boundary of the original fracture remnant.

[0081] S3.2 The Poisson reconstruction algorithm is used to automatically repair the holes and broken surfaces in the three-dimensional surface model, and automatically achieves more than 99% hole closure repair, thus obtaining a watertight and complete three-dimensional surface model.

[0082] S3.3, based on the repaired watertight three-dimensional surface model, calculate the macroscopic geometric parameters of the defect area: the volume of the defect entity envelope is 85.3 cubic centimeters, the macroscopic outer surface area is 120.5 square centimeters, and the average curvature is 0.15 per millimeter; based on the fracture end interface mask, calculate the macroscopic contact area of ​​the residual section as 4.8 square centimeters, and the curvature variation coefficient characterizing the geometric complexity of its cross-section.

[0083] S3.4 Receive the clinically confirmed prior pathological diagnosis result, which indicates that the defect type is a traumatic defect. Input this prior pathological diagnosis result, along with the defect volume, surface area, average curvature, contact area, and coefficient of variation of curvature, into the constraint solution model.

[0084] S3.5, the constraint solution model adaptively outputs the personalized implant design constraints for this case based on the prior classification results of traumatic defects and the mechanical compliance requirements of the chest wall under respiratory physiological conditions: the target volume fraction needs to be controlled at 60%, and the expected respiratory physiological dynamic micro-strain constraint range is 800 microstrain to 1500 microstrain.

[0085] Step 4: Based on the boundary data of the defect area, the cross-sectional data of the residual end, and the design constraints of the personalized implant, a macroscopic outline model of the implant is first generated using a two-level generative network. Then, the optimal microscopic topology and gradient lattice structure are generated within the spatial envelope defined by the outline model to obtain the initial model of the implant.

[0086] S4.1 Extract the three-dimensional spatial boundary curve of the defective edge and the contact surface data of the residual section from the repaired watertight three-dimensional surface model.

[0087] S4.2, a shape fitting network based on a deep learning regression network is constructed. Using the extracted boundary curves and contact surface data as input, a macroscopic shape contour model that highly conforms to the patient's anatomical morphology is inferred and constructed. The maximum relative mathematical fit gap between the outer contour of this shape contour model and the defect edge is only 0.18 mm, achieving extremely high macroscopic anatomical adaptation. This shape contour model serves as the maximum spatial envelope design boundary constraint for subsequent generative models.

[0088] S4.3, constructing a generative AI model based on an improved topology diffusion model as the core generative network. This model includes a forward diffusion process and a reverse denoising process with an embedded bidirectional attention mechanism module, used to explicitly focus on key frequency bands of the dynamic load spectrum and the local geometric features of the manufacturing constraints.

[0089] S4.4, the constructed generative AI model is trained using multi-task training with physical constraints to obtain the trained model. This model has converged on a multimodal optimization dataset of 10,000 physically labeled datasets. The loss function of the multimodal training dataset integrates a biomechanical AI agent model based on a graph neural network architecture. The personalized implant design constraints output from step 3 (target volume fraction 60%, expected physiological microstrain 800 to 1500 microstrains) and the minimum wall thickness of 0.3 mm for the selected area laser melting process are used as guiding conditions input to the trained model. Through a reverse denoising process, the model simultaneously performs joint spatial evolution of the macroscopic contour boundary density and the internal porous structure within the step size, autonomously creating multiple candidate implant structure data containing asymmetric gradient lattices and biomimetic honeycomb-spring composite configurations. This AI inference and creation process takes approximately 10 minutes and 25 seconds in total.

[0090] S4.5 Post-processes multiple candidate implant structural data: Isolated voxel removal is performed to eliminate non-physical micromaterial islands or pores; Laplacian mesh smoothing is performed to convert voxelized stepped boundaries into smooth, continuous surfaces; implicit surface parameterization is performed to convert the processed structural data into an editable STEP format, resulting in initial models of multiple candidate implants. The total specific surface area of ​​the internal porous, multi-scale skeleton of the post-processed initial implant models reaches 5800 square centimeters, with a micro-rod diameter spatial gradient distribution ranging from 0.3 mm to 0.8 mm and a porosity gradient distribution ranging from 50% to 70%.

[0091] Step 5: Perform automatic geometric accuracy verification and mechanical simulation verification on multiple candidate implant initial models. Determine the optimal implant model based on the verification results and provide a visual interactive interface for doctors to perform closed-loop interactive fine-tuning to obtain the final implant model.

[0092] S5.1 For multiple candidate implant initial models, calculate the Dice similarity coefficient and Hausdorff distance between each candidate implant initial model and the watertight 3D surface model repaired in step 3. A biomimetic honeycomb-spring composite structure was selected for verification. This structure achieved a relative mathematical Dice similarity coefficient of 0.98 with the defective 3D surface, and a Hausdorff distance of 0.19 mm.

[0093] S5.2, a periodic cyclic tensile alternating load simulating deep thoracic breathing is applied to the biomimetic honeycomb-spring composite structure at a frequency of 0.2 Hz. Mechanical simulation accurately calculates the maximum von Mises stress within the structure to be 150 MPa, far below the material's high-cycle fatigue limit safety threshold of 360 MPa. The dynamic strain deviation is controlled within ±40 micro-strains, perfectly matching physiological skeletal compliance.

[0094] S5.3, The biomimetic honeycomb-spring composite structure is selected as the optimal implant model. A three-dimensional visualization interactive interface is output on the main control display unit to intuitively render and display the three-dimensional color stress cloud map and geometric deviation heat map of the optimal implant model.

[0095] In step S5.4, after viewing the interface, the attending physician, based on clinical experience with stress, inputs a fine-tuning command for the edge rounding, adjusting the local rounding of the lateral fixed area from 0.5 mm to 1.0 mm. The system performs local geometric reconstruction of the optimal implant model according to the fine-tuning command and automatically invokes a biomechanical AI proxy model based on a graph neural network architecture for transient mechanical re-verification, completing both the local geometric reconstruction and transient mechanical re-verification within 45 seconds. After multiple rounds of confirmation by the physician that all indicators remain qualified, the system triggers the export process. The system completes the final high-precision finite element analysis verification and confirms safety within 3 minutes, approving the output of the final implant model. The total time for the physician's fine-tuning and re-verification is less than 15 minutes.

[0096] Step 6: Convert the final implant model obtained in Step 5 into a physical format recognizable by the additive manufacturing equipment and transmit it to the printing equipment. Based on the set material and process parameters, perform solid printing and post-processing manufacturing. At the same time, feed the measured physical data back to the database to realize the asynchronous evolution and update of the model throughout its entire life cycle.

[0097] S6.1 converts the final implant model to the standard STL format.

[0098] S6.2 wirelessly transmits the converted STL model file to an industrial-grade selective laser melting 3D printing device. Medical-grade high-purity titanium alloy powder is used as the substrate, and the additive manufacturing process parameters are set as follows: laser power 200 watts, scanning speed 1000 mm / s, and powder layer thickness 30 micrometers.

[0099] S6.3, based on the set process parameters, a selective laser melting process is used to perform high-precision melting and molding to obtain a physical entity of a personalized metal rib implant skeleton.

[0100] S6.4, Standardized post-processing of the printed physical entity: The physical entity is placed in a high-vacuum forming furnace for hot isostatic pressing (HIP), with a pressure set at 120 MPa and a temperature set at 920 degrees Celsius. After holding in high vacuum for 2 hours, it is cooled under controlled temperature to completely eliminate residual stress and microscopic internal pores from the printing process, obtaining a fine α+β two-phase structure. Subsequently, the HIP-treated physical entity is sandblasted to remove unmelted powder particles adhering to the surface. Then, the sandblasted physical entity undergoes anodizing surface treatment to improve surface biocompatibility. Finally, the surface-treated physical entity is sterilized with medical-grade ethylene oxide to obtain a final implantable product suitable for clinical use.

[0101] In step S6.5, offline physical tensile and tension-tensile fatigue tests were conducted on solid samples from the same batch. The quasi-static crush strength was measured to be 120 MPa, and the micro-dynamic strain retention rate was as high as 92% after 10 million fatigue cycles. This high-confidence physical test result, along with the original design boundary conditions, was asynchronously fed back into the offline multimodal training database for subsequent closed-loop periodic fine-tuning and evolution of the model.

[0102] Example 2 This embodiment provides a system for implementing the above-described implant generation method based on CT images and generative AI. The system includes: The preprocessing module receives raw DICOM format CT image data transmitted from the medical image acquisition device, performs online denoising, artifact suppression, and region of interest (ROI) cropping on the raw CT image data, and outputs optimized ROI images.

[0103] The preprocessing module incorporates a nonlocal mean filtering algorithm and a metal artifact correction network. The medical image acquisition equipment includes a spiral CT scanner or a multimodal magnetic resonance imaging (MRI) scanner.

[0104] The segmentation module is connected to the preprocessing module and is used to receive the optimized region of interest image. It is internally deployed with weight parameters of a multi-target image segmentation network pre-trained using the gold standard. It performs parallel inference on the region of interest image and outputs multi-dimensional segmentation mask data of target anatomical structure, defect area and bone fracture section.

[0105] The segmentation network is a multi-objective segmentation network based on an improved U-Net++ architecture or a Transformer architecture.

[0106] The reconstruction and analysis module is connected to the segmentation module and is used to receive the segmentation mask data, reconstruct the segmentation mask into a watertight three-dimensional curved surface geometric model, automatically and quantitatively analyze and calculate the morphological characteristics of the defect area, and, in combination with prior clinical pathological classification, output personalized implant design constraint parameters to the subsequent stage.

[0107] The reconstruction and analysis module integrates a moving cube algorithm, a Poisson mesh automatic surface repair algorithm, and a constraint solution model. The morphological features include defect volume, surface area, curvature, and the coefficient of variation of contact surface curvature. The personalized implant design constraint parameters include the target volume fraction and the expected physiological strain range.

[0108] The design generation module is connected to the reconstruction and analysis module and is used to receive the personalized implant design constraint parameters. It has a two-level generative network model loaded inside. First, it generates an outline model that macroscopically fits the remaining anatomical structure. Then, it creates a microscopic topology and gradient lattice structure within the spatial envelope defined by the outline model and outputs the initial model of the implant.

[0109] The two-level generative network model includes a shape fitting network, a generative artificial intelligence framework based on a topology diffusion model, and a biomechanical artificial intelligence surrogate model prediction component. The shape fitting network is used to generate the macroscopic outer contour, the generative artificial intelligence framework is used to directly generate the optimal microscopic matrix topology structure with spatial gradient distribution of microscopic stiffness and porosity within minutes, and the biomechanical artificial intelligence surrogate model prediction component is used to predict the mechanical response of the structural morphology in real time during the generative process.

[0110] The verification and interaction module is connected to the design generation module and is used to receive the initial model of the implant, automatically calculate the geometric similarity index and mechanical performance index, render and output the verified model, geometric deviation heat map and stress cloud map to the display unit in real time, capture the fine-tuning input instructions of the clinician through the graphical interactive interface, update the model in real time and trigger the transient mechanical re-verification based on the biomechanical artificial intelligence agent model and the final verification based on high-precision finite element analysis, and output the final implant model.

[0111] The geometric similarity indices include the Dice similarity coefficient and the Hausdorff distance; the mechanical performance indices include the maximum von Mises stress and the dynamic strain response. The fine-tuning input commands include edge fillet adjustment, local thickening, local rod diameter or porosity fine-tuning, edge chamfering, or fixed hole position adjustment.

[0112] The communication module is connected to the verification and interaction module and supports universal serial bus interface, high-definition multimedia interface, wireless fidelity network or low power Bluetooth communication protocol to realize high-speed bidirectional data interaction with external 3D printing equipment or hospital image archiving and communication system database.

[0113] The communication module will wirelessly transmit the implant model and associated printing process parameter file, which are ultimately verified through interaction, to the selective laser melting additive manufacturing equipment for production scheduling. At the same time, it will receive physical and mechanical measurement data asynchronously fed back from the production quality inspection end, and rewrite the measurement data into the offline database to trigger the closed-loop periodic incremental learning and weight evolution optimization of the neural network.

[0114] The processor includes a heterogeneous parallel computing architecture comprising a central processing unit and a graphics processing unit, and is connected to the preprocessing module, segmentation module, reconstruction and analysis module, design generation module, verification and interaction module, and communication module, respectively, for coordinating data flow and computing task scheduling among the modules.

[0115] The memory is connected to the processor and is used to store computer program instructions, high-precision finite element analysis simulation software, biomechanical artificial intelligence agent model evaluation components, geometric similarity evaluation matrix, and offline accumulated multimodal physical experiment database.

[0116] The display unit is connected to the processor and is used to display a three-dimensional visualization interface, intuitively rendering and displaying the implant model, geometric deviation heatmap, and mechanical simulation stress distribution results. The display unit is a high-resolution display panel or a laser scanning display.

[0117] It should be noted that, in the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another device, or some features may be ignored or not executed. The modules described as separate components may or may not be physically separated. The components shown as modules may be one or more physical units, that is, they may be located in one place or distributed in multiple different places. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs.

[0118] Furthermore, in the various embodiments of the present invention, the modules can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The integrated unit described above can be implemented in hardware or as a software functional unit.

[0119] An electronic device provided in this application includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the implant generation method based on CT images and generative AI as described in any of the above embodiments.

[0120] Another electronic device provided in this application embodiment may further include: an input port connected to a processor for transmitting multimodal data collected by an external acquisition device to the processor; a display unit connected to the processor for displaying the processor's processing results to the outside world; and a communication module connected to the processor for enabling communication between the electronic device and the outside world. The display unit may be a display panel, a laser scanning display, etc.; the communication method adopted by the communication module includes, but is not limited to, Mobile High Definition Link (HML), Universal Serial Bus (USB), High Definition Multimedia Interface (HDMI), and wireless connection (including Wi-Fi, Bluetooth, Bluetooth Low Energy, and IEEE 802.11s-based communication technology).

[0121] This application provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps of the implant generation method based on CT images and generative AI as described in any of the above embodiments.

[0122] For descriptions of relevant parts of the implant generation system, electronic device, and computer-readable storage medium based on CT images and generative AI provided in this application, please refer to the detailed descriptions of the corresponding parts in the implant generation method based on CT images and generative AI provided in this application, which will not be repeated here. Furthermore, parts of the technical solutions provided in this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0123] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A method for implant generation based on CT images and generative AI, characterized in that, Includes the following steps: Step 1: Obtain the region of interest image based on CT image data of the target anatomical site; Step 2: Determine the segmentation mask based on the region of interest image; The segmentation mask includes at least a segmentation mask for the target anatomical structure, a segmentation mask for the defect area, and a segmentation mask for the cross-section of the bone fracture end; Step 3: Reconstruct a three-dimensional surface model of the target anatomical site based on the segmentation mask, calculate the geometric parameters of the defect area and the mechanical contact parameters of the stump section based on the model, and determine the personalized implant design constraints in combination with the prior pathological diagnosis results and corresponding physiological parameters. Step 4: Extract the boundary curve of the defect area and the contact surface data of the residual section from the three-dimensional surface model, generate an outline model that fits the remaining anatomical structure, and generate a micro-topology and lattice structure within the spatial envelope defined by the outline model to obtain the initial model of the implant. Step 5: Verify the initial implant model and determine the final implant model based on the verification results.

2. The implant generation method based on CT images and generative AI according to claim 1, characterized in that, In step 1, acquiring the region of interest image based on CT image data of the target anatomical site includes: The CT image data is sequentially denoised and artifact corrected to obtain the corrected CT image. The region of interest is extracted from the corrected CT image to obtain the region of interest image.

3. The implant generation method based on CT images and generative AI according to claim 1, characterized in that, The step of determining the segmentation mask based on the region of interest image includes: The image of the region of interest is input into a trained multi-target segmentation network to obtain segmentation masks for the target anatomical structure, the defect area, and the cross-section of the bone fracture.

4. The implant generation method based on CT images and generative AI according to claim 1, characterized in that, Step 3, reconstructing the three-dimensional surface model of the target anatomical region based on the segmentation mask, includes: The segmentation mask is converted into a three-dimensional surface model using the moving cube algorithm; The Poisson reconstruction algorithm is used to automatically repair the holes and broken surfaces in the three-dimensional surface model, so as to obtain a three-dimensional surface model with complete water tightness.

5. The implant generation method based on CT images and generative AI according to claim 1, characterized in that, In step 3, the geometric parameters of the defect region include the defect volume, surface area, and average curvature; the mechanical contact parameters of the residual section include the macroscopic contact area and the coefficient of variation of curvature; and the design constraints of the personalized implant include the target volume fraction and the expected physiological dynamic strain range.

6. The implant generation method based on CT images and generative AI according to claim 1, characterized in that, Step 4, generating an external contour model that fits the remaining anatomical structure, includes: A shape fitting network is constructed, and a shape contour model is generated by taking the boundary curve of the defective region and the contact surface data of the residual section as input; the shape contour model serves as the maximum spatial envelope design boundary constraint for generating the micro-topology and lattice structure.

7. The implant generation method based on CT images and generative AI according to claim 1, characterized in that, In step 4, generating the microscopic topology and lattice structure within the spatial envelope defined by the outline model includes: Construct a generative artificial intelligence model, which includes a forward diffusion process and a reverse denoising process with an embedded bidirectional attention mechanism module; The personalized implant design constraints are input into the trained generative artificial intelligence model, and implant structure data is generated through a reverse denoising process. The implant structure data is post-processed to obtain the initial model of the implant.

8. The implant generation method based on CT images and generative AI according to claim 1, characterized in that, During the training process of the generative artificial intelligence model, a biomechanical proxy model based on a graph neural network architecture is introduced, which integrates three constraints—static compliance, dynamic strain deviation, and manufacturing feasibility—using a multi-task loss function. The multi-task loss function is expressed as follows: in: The primary objective is to minimize the loss term related to structural static compliance; To dynamically adapt the target loss term; For manufacturing-related loss items; , , Hyperparameters are used to balance the importance of various losses.

9. The implant generation method based on CT images and generative AI according to claim 1, characterized in that, The process of validating the initial implant model and determining the final implant model based on the validation results includes: The geometric accuracy of the initial model of the implant is verified, and the Dice similarity coefficient and Hausdorff distance between it and the three-dimensional surface model are calculated. The initial model of the implant was verified by mechanical simulation. A simulated physiological dynamic cyclic alternating load spectrum was applied to calculate the maximum stress and dynamic strain deviation inside the model. The optimal implant model is determined based on the Dice similarity coefficient, Hausdorff distance, maximum stress, and dynamic strain deviation. The optimal implant model is then locally geometrically reconstructed based on the geometric deviation heatmap and the stress distribution results of the mechanical simulation to obtain the final implant model.

10. An implant generation system based on CT images and generative AI, characterized in that, include: The preprocessing module is used to acquire region-of-interest images based on CT image data of the target anatomical site; A segmentation module is used to determine a segmentation mask based on the region of interest image; the segmentation mask includes at least a segmentation mask for the target anatomical structure, a segmentation mask for the defect area, and a segmentation mask for the cross-section of the bone fracture end; The reconstruction and analysis module is used to reconstruct a three-dimensional surface model of the target anatomical site based on the segmentation mask, calculate the geometric parameters of the defect area and the mechanical contact parameters of the stump section based on the model, and determine the personalized implant design constraints in combination with the prior pathological diagnosis results and corresponding physiological parameters. The design generation module is used to extract the boundary curve of the defect area and the contact surface data of the residual section from the three-dimensional surface model, generate the outline model that fits the remaining anatomical structure, generate the micro-topology and lattice structure within the spatial envelope defined by the outline model, and obtain the initial model of the implant. The verification and interaction module is used to verify the initial model of the implant and determine the final implant model based on the verification results.