A method and apparatus for determining the inverse generation model of cancellous bone scaffold based on topology optimization

Abstract

This application relates to the technical field of cancellous bone scaffolds, and discloses a method and apparatus for determining a reverse generation model of cancellous bone scaffolds based on topology optimization. In this method, the three-dimensional thin-layer structure is topology optimized to mimic the complex and irregular microstructural features of cancellous bone tissue. The first and second multi-channel mechanical contour maps represent force-related features. Multimodal data consisting of the three-dimensional thin-layer structure, the initial simplified model, the first multi-channel mechanical contour map, and the second multi-channel mechanical contour map are used as training samples to train a latent graph diffusion model. This allows the model to achieve powerful output performance for cancellous bone scaffolds under multimodal input conditions.

Description

Technical Field

[0001] This application relates to the technical field of cancellous bone scaffolds, and in particular to a method and apparatus for determining a reverse generation model of cancellous bone scaffolds based on topology optimization. Background Technology

[0002] Precise repair of cancellous bone defects in long bones has always been a major challenge in clinical orthopedics. Bone defects refer to the partial or complete loss of bone tissue, usually caused by trauma, bone infection, bone tumors, osteoporosis, and other diseases. Millions of patients worldwide are affected by bone defects each year. If not repaired in time, the defect area may become structurally unstable, leading to a decrease in the local bone tissue's load-bearing capacity. At this point, the bone surrounding the defect becomes fragile, easily suffering secondary damage or even severe fractures from minor external forces. If the defect fails to heal for a long time, it can also lead to articular cartilage degeneration, resulting in osteoarthritis and limb dysfunction, severely impacting the patient's quality of life and work ability.

[0003] The repair of clinical bone defects has long relied on autologous or allogeneic transplantation, metal implants, or stem cell therapy. However, these methods have limitations in terms of donor availability, biocompatibility, and long-term integration. For example, autologous bone transplantation is often accompanied by donor site damage, allogeneic bone transplantation carries the risk of immune rejection and infection, and metal implants, due to significant differences in mechanical properties compared to natural bone tissue, often induce stress shielding effects, affecting new bone regeneration and structural stability. Against this backdrop, the use of 3D printing technology to construct biomimetic scaffolds is gradually emerging as an alternative. These scaffolds, through their controllable porous structures, provide an ideal microenvironment for cells and blood vessels, playing a crucial role in the clinical treatment of bone defects. In particular, the design of cancellous bone scaffolds hinges on achieving a precise match between complex three-dimensional microstructures and target mechanical properties.

[0004] However, most existing scaffold design methods rely on experience-driven forward modeling, often approximating target performance by manually adjusting porosity, pore size, or geometry. This approach is inefficient, difficult to repeat, and fails to fully reveal the nonlinear relationship between microstructure and macroscopic mechanical response. Therefore, obtaining suitable cancellous bone scaffold structures for subsequent application to patients' affected areas has become a pressing technical problem for those skilled in the art. Summary of the Invention

[0005] This application provides a method and apparatus for determining a reverse-generative model of a cancellous bone scaffold based on topology optimization. In this method, the three-dimensional thin-layer structure is topology-optimized to mimic the complex and irregular microstructural features of cancellous bone tissue. The first and second multi-channel mechanical contour maps represent force-related features. Multimodal data consisting of the three-dimensional thin-layer structure, the initial simplified model, the first multi-channel mechanical contour map, and the second multi-channel mechanical contour map are used as training samples to train a latent graph diffusion model. This allows the model to achieve strong performance in outputting cancellous bone scaffold results under multimodal input conditions.

[0006] In a first aspect, embodiments of this application provide a method for determining a reverse generation model of a cancellous bone scaffold based on topology optimization, including:

[0007] Based on various spatial topological structural units, the corresponding initial three-dimensional spatial topological structure is constructed.

[0008] The initial three-dimensional spatial topology is simplified into a corresponding initial simplified model;

[0009] The initial simplified model is used to generate various three-dimensional thin-layer structures by adjusting structural parameters, and intermediate simplified models are determined in the process of generating the three-dimensional thin-layer structures; the structural parameters include the diameter of the support rod, the length of the support rod, and the load applied under the simulation boundary conditions;

[0010] The initial simplified model and the intermediate simplified model are extracted using finite element analysis to obtain the first multi-channel mechanical cloud map and the second multi-channel mechanical cloud map respectively; both the first multi-channel mechanical cloud map and the second multi-channel mechanical cloud map include stress cloud map in the x direction, stress cloud map in the y direction, stress cloud map in the z direction, strain cloud map in the x direction, strain cloud map in the y direction, strain cloud map in the z direction and displacement cloud map.

[0011] The initial simplified model, the first multi-channel mechanical cloud map, the second multi-channel mechanical cloud map, and the three-dimensional thin-layer structure are used as training samples to train the latent graph diffusion model to generate a cancellous bone scaffold inverse generation model.

[0012] In some embodiments, the step of training a latent graph diffusion model to generate a cancellous bone scaffold inverse generation model by using the initial simplified model, the first multi-channel mechanical cloud map, the second multi-channel mechanical cloud map, and the three-dimensional thin-layer structure as training samples includes:

[0013] The stress cloud map in the x-direction, stress cloud map in the y-direction, stress cloud map in the z-direction, strain cloud map in the x-direction, strain cloud map in the y-direction, strain cloud map in the z-direction, and displacement cloud map in the second multi-channel mechanical cloud map are all divided into P image blocks.

[0014] The intermediate condition data is formed by combining the stress cloud map in the x direction, stress cloud map in the y direction, stress cloud map in the z direction, strain cloud map in the x direction, strain cloud map in the y direction, strain cloud map in the z direction, and displacement cloud map in the second multi-channel mechanical cloud map with the P image blocks obtained from each division.

[0015] Using the initial simplified model, the first multi-channel mechanical cloud map, intermediate condition data, and the three-dimensional thin-layer structure, a potential graph diffusion model is trained.

[0016] In some embodiments, prior to the step of training the latent graph diffusion model using the initial simplified model, the first multi-channel mechanical contour map, intermediate conditional data, and the three-dimensional thin-layer structure, the method further includes:

[0017] The first multi-channel mechanical cloud map and the intermediate condition data are masked using a random masking mechanism.

[0018] Replace the masked first multi-channel mechanical cloud map with the first multi-channel mechanical cloud map;

[0019] Replace the intermediate condition data after masking and use it as the intermediate condition data.

[0020] In some embodiments, the step of masking the first multi-channel mechanical cloud map and the intermediate condition data using a random masking mechanism includes:

[0021] The first multi-channel mechanical cloud map is masked using the first mask probability to obtain the masked first multi-channel mechanical cloud map.

[0022] The intermediate condition data is masked using the second mask probability to obtain the masked intermediate condition data.

[0023] In some embodiments, the step of masking the first multi-channel mechanical cloud map and the intermediate condition data using a random masking mechanism includes:

[0024] The stress cloud map in the x-direction, stress cloud map in the y-direction, stress cloud map in the z-direction, strain cloud map in the x-direction, strain cloud map in the y-direction, strain cloud map in the z-direction, and displacement cloud map in the first multi-channel mechanical cloud map are all divided into P image blocks.

[0025] The image block is masked using the third mask probability to obtain the masked image block;

[0026] The image composed of P masked image blocks is determined as the first multi-channel mechanical cloud map after masking.

[0027] Inject the intermediate condition data with Gaussian noise;

[0028] The stress contour maps in the x-direction, y-direction, z-direction, strain contour maps in the x-direction, y-direction, z-direction, and displacement contour maps in the intermediate condition data injected with Gaussian noise are all masked using the fourth mask probability to obtain the masked intermediate condition data.

[0029] In some embodiments, the latent graph diffusion model adopts a symmetrical U-shaped structure; the latent graph diffusion model includes a downsampling encoder and an upsampling decoder, and employs skip connections;

[0030] The steps for training the latent graph diffusion model using the initial simplified model, the first multi-channel mechanical cloud map, intermediate condition data, and the three-dimensional thin-layer structure include:

[0031] The initial simplified model is encoded using the upsampling encoder of the latent graph diffusion model to obtain latent spatial features;

[0032] Based on the aforementioned potential spatial features, a three-dimensional thin-layer structure is obtained using a downsampling decoder.

[0033] In some embodiments, the downsampling encoder includes a first graph convolutional network layer, a first self-attention layer, a first cross-attention layer, and a downsampling layer connected in sequence; the upsampling decoder includes an upsampling layer, a second self-attention layer, a second cross-attention layer, and a second graph convolutional network layer connected in sequence, wherein the first cross-attention layer and the second cross-attention layer introduce a conditional mapping mechanism, which is guided by the first multi-channel mechanical cloud map and intermediate conditional data.

[0034] Secondly, embodiments of this application provide a device for determining a reverse generation model of a cancellous bone scaffold based on topology optimization, comprising:

[0035] The building unit is used to construct the corresponding initial three-dimensional spatial topology based on various spatial topology units;

[0036] A simplification unit is used to simplify the initial three-dimensional spatial topology into a corresponding initial simplified model;

[0037] The generation unit is used to generate various three-dimensional thin-layer structures from the initial simplified model by adjusting the structural parameters, and to determine the intermediate simplified model in the process of generating the three-dimensional thin-layer structures; the structural parameters include the support rod diameter, support rod length, and load applied under simulation boundary conditions;

[0038] The extraction unit is used to extract the initial simplified model and the intermediate simplified model using finite element analysis, respectively, to obtain a first multi-channel mechanical cloud map and a second multi-channel mechanical cloud map; both the first multi-channel mechanical cloud map and the second multi-channel mechanical cloud map include stress cloud maps in the x-direction, stress cloud maps in the y-direction, stress cloud maps in the z-direction, strain cloud maps in the x-direction, strain cloud maps in the y-direction, strain cloud maps in the z-direction, and displacement cloud maps;

[0039] The training unit is used to train the latent graph diffusion model using the initial simplified model, the first multi-channel mechanical cloud map, the second multi-channel mechanical cloud map, and the three-dimensional thin-layer structure as training samples, so as to generate a cancellous bone scaffold inverse generation model.

[0040] Thirdly, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method for determining the inverse generation model of cancellous bone scaffold based on topology optimization.

[0041] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the step of determining the inverse generation model of the cancellous bone scaffold.

[0042] The above embodiments provide a method and apparatus for determining a reverse generation model of cancellous bone scaffold based on topology optimization. In this method, the three-dimensional thin-layer structure is topology optimized to mimic the complex and irregular microstructural features of cancellous bone tissue. The first and second multi-channel mechanical contour maps represent force-related features. Multimodal data consisting of the three-dimensional thin-layer structure, the initial simplified model, the first multi-channel mechanical contour map, and the second multi-channel mechanical contour map are used as training samples to train the latent graph diffusion model. This allows the model to achieve strong performance in outputting cancellous bone scaffold results under multimodal input conditions. The method includes: constructing an initial three-dimensional spatial topology based on multiple spatial topological structural units; simplifying the initial three-dimensional spatial topology into a corresponding initial simplified model; generating multiple three-dimensional thin-layer structures from the initial simplified model by adjusting structural parameters, and determining an intermediate simplified model in the process of generating the three-dimensional thin-layer structures; the structural parameters include the diameter of the support rod, the length of the support rod, and the load applied under the simulation boundary conditions; extracting the initial simplified model and the intermediate simplified model using finite element analysis to obtain a first multi-channel mechanical cloud map and a second multi-channel mechanical cloud map; both the first and second multi-channel mechanical cloud maps include stress cloud maps in the x-direction, stress cloud maps in the y-direction, stress cloud maps in the z-direction, strain cloud maps in the x-direction, strain cloud maps in the y-direction, strain cloud maps in the z-direction, and displacement cloud maps; using the initial simplified model, the first multi-channel mechanical cloud map, the second multi-channel mechanical cloud map, and the three-dimensional thin-layer structure as training samples to train a latent graph diffusion model to generate a cancellous bone scaffold inverse generation model. Attached Figure Description

[0043] Figure 1 An exemplary flowchart illustrates a method for determining a reverse generation model of a cancellous bone scaffold based on topology optimization, according to some embodiments.

[0044] Figure 2 The illustration provides a schematic diagram of the stepwise simplification process of four spatial topology units provided according to some embodiments from an initial three-dimensional spatial topology result to an initial simplified model;

[0045] Figure 3 Schematic diagrams of an initial simplified model and an intermediate simplified model provided according to some embodiments are illustrated;

[0046] Figure 4 The illustration shows a schematic diagram of experimental stress-strain and finite element simulation results of a reverse-generated model of a cancellous bone scaffold provided according to some embodiments under different orientations and structural parameters;

[0047] Figure 5An exemplary flowchart is shown of a device for determining a reverse generation model of a cancellous bone scaffold based on topology optimization, according to some embodiments. Detailed Implementation

[0048] To make the objectives and implementation methods of this application clearer, the exemplary implementation methods of this application will be clearly and completely described below with reference to the accompanying drawings of the exemplary embodiments of this application. Obviously, the exemplary embodiments described are only some embodiments of this application, and not all embodiments.

[0049] It should be noted that the brief descriptions of terms in this application are only for the convenience of understanding the embodiments described below, and are not intended to limit the embodiments of this application. Unless otherwise stated, these terms should be understood in their ordinary and common meaning.

[0050] The terms "first," "second," "third," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar or related objects or entities, and do not necessarily imply a specific order or sequence, unless otherwise specified. It should be understood that such terms are interchangeable where appropriate.

[0051] The terms “comprising” and “having”, and any variations thereof, are intended to cover but not exclude inclusion, for example, a product or device that includes a range of components is not necessarily limited to all of the components that are clearly listed, but may include other components that are not clearly listed or that are inherent to such product or device.

[0052] To address the aforementioned technical problems, this application provides a method and apparatus for determining a reverse-generative model of a cancellous bone scaffold based on topology optimization. In this method, the three-dimensional thin-layer structure undergoes topology optimization, mimicking the complex and irregular microstructural features of cancellous bone tissue. The first and second multi-channel mechanical contour maps represent force-related features. Multimodal data consisting of the three-dimensional thin-layer structure, the initial simplified model, the first multi-channel mechanical contour map, and the second multi-channel mechanical contour map are used as training samples to train a latent graph diffusion model. This allows the model to achieve powerful output performance for cancellous bone scaffold results under multimodal input conditions.

[0053] Figure 1 An exemplary flowchart is shown of a method for determining a reverse generation model of a cancellous bone scaffold based on topology optimization, according to some embodiments, the method comprising steps S100-S500.

[0054] S100. Based on various spatial topology units, construct the corresponding initial three-dimensional spatial topology.

[0055] In this embodiment, the various spatial topology units may include four types: Voronoi spatial units, cellular spatial units, triangular grid spatial units, and parallelogram spatial units. By randomly distributing the spatial topology units, the corresponding initial three-dimensional spatial topology can be constructed using engineering modeling software (e.g., nTopology Inc.).

[0056] Four types of spatial topology units can be used to construct corresponding initial three-dimensional spatial topologies, specifically the initial three-dimensional spatial topologies constructed by Voronoi spatial units, cellular spatial units, triangular grid spatial units, and parallelogram spatial units.

[0057] For example, such as Figure 2 As shown, Figure 2 The initial volume models (i.e., the initial three-dimensional spatial topology) in the image, from top to bottom, are the initial three-dimensional spatial topology constructed by Voronoi spatial units, the initial three-dimensional spatial topology constructed by cellular spatial units, the initial three-dimensional spatial topology constructed by triangular grid spatial units, and the initial three-dimensional spatial topology constructed by parallelogram spatial units.

[0058] S200. Simplify the initial three-dimensional spatial topology into a corresponding initial simplified model;

[0059] In this embodiment, because the initial three-dimensional spatial topology is complex and not conducive to subsequent latent graph diffusion model learning, the initial three-dimensional spatial topology is simplified to a corresponding initial simplified model. Specifically, this may include first simplifying the initial three-dimensional spatial topology to an initial surface structure, and then simplifying it to an initial simplified structure. See again... Figure 2 , Figure 2 The initial three-dimensional spatial topology includes four types of initial three-dimensional spatial topologies: the initial three-dimensional spatial topology constructed by Voronoi spatial units, the initial three-dimensional spatial topology constructed by cellular spatial units, the initial three-dimensional spatial topology constructed by triangular grid spatial units, and the initial three-dimensional spatial topology constructed by parallelogram spatial units. The initial surface structure and the initial simplified structure are obtained by simplifying these four initial three-dimensional spatial topologies.

[0060] S300. Generate various three-dimensional thin-layer structures from the initial simplified model by adjusting the structural parameters, and determine the intermediate simplified model in the process of generating the three-dimensional thin-layer structure; the structural parameters include the support rod diameter, support rod length, and load applied under simulation boundary conditions.

[0061] This three-dimensional thin-layer structure was obtained by adjusting the structural parameters, or by topology optimization.

[0062] In this embodiment, the initial simplified model is composed of different support rods connected at the ends. The specific method for adjusting the load applied under the simulation boundary conditions of the initial simplified model is to apply a fixed constraint to the bottom surface of the initial simplified model and apply a vertical downward load (load applied under the simulation boundary conditions) to the top surface.

[0063] In this embodiment of the application, the step of generating multiple three-dimensional thin-layer structures from the initial simplified model by adjusting the structural parameters may include adjusting the structural parameters of the initial simplified model multiple times within a preset range of adjustable structural parameters to obtain multiple three-dimensional thin-layer structures, that is, three-dimensional thin structures obtained after three-dimensional topology optimization.

[0064] In one example, the preset range of adjustable structural parameters includes a range for the average support rod diameter, a range for the average support rod length, and a range for the load applied under simulation boundary conditions. The step of generating various three-dimensional thin-layer structures from the initial simplified model by adjusting the structural parameters may include: adjusting different support rod diameters of the initial simplified model within the range for the average support rod diameter; adjusting different support rod lengths of the initial simplified model within the range for the average support rod length; and adjusting the load applied under the simulation boundary conditions of the initial simplified model within the range for the load applied under the simulation boundary conditions. See again. Figure 2 , Figure 2 The length l of the support rod is marked in the middle.

[0065] For example, the average support rod diameter is set within the range of [0.2, 1.0] mm with a step size of 0.01 mm; the average support rod length is set within the range of [1.0, 3.0] mm with a step size of 0.1 mm; and the applied load under the simulation boundary conditions is set within the range of [5.0, 10.0] N with a step size of 0.2 N.

[0066] In this embodiment, the generation of various three-dimensional thin-layer structures from the initial simplified model by adjusting structural parameters is a step-by-step process. For example, the initial simplified model can be transformed into a three-dimensional thin-layer structure with the target structural parameters through 100 iterations. The initial simplified model can be the initial simplified model at iteration 0, and the intermediate simplified model can be any initial simplified model between iterations 0 and 100. For instance, the initial simplified model can be iterated 10 times to obtain the intermediate simplified model. Figure 3 As shown, Figure 3 The image shows an initial simplified model and an intermediate simplified model.

[0067] In this embodiment, the three-dimensional thin-layer model is combined with the potential graph diffusion model to obtain a three-dimensional cancellous bone scaffold generation process that integrates topology optimization design priors, multi-channel finite element physical field simulation and graph diffusion generation mechanism. This realizes the automatic generation of bone scaffold structures based on target mechanical properties and opens up a full-chain modeling path for design, performance and manufacturing.

[0068] S400. The initial simplified model and the intermediate simplified model are extracted using finite element analysis to obtain the first multi-channel mechanical cloud map and the second multi-channel mechanical cloud map respectively. The first multi-channel mechanical cloud map and the second multi-channel mechanical cloud map both include stress cloud map in the x direction, stress cloud map in the y direction, stress cloud map in the z direction, strain cloud map in the x direction, strain cloud map in the y direction, strain cloud map in the z direction, and displacement cloud map.

[0069] See again Figure 3 , Figure 3 The image shows a first multi-channel mechanical contour plot extracted from the initial simplified model using finite element analysis (FEM), and a second multi-channel mechanical contour plot extracted from the intermediate simplified model using finite element analysis. Both the first and second multi-channel mechanical contour plots consist of seven contour plots, representing, from front to back, the stress in the x-direction, stress in the y-direction, stress in the z-direction, strain in the x-direction, strain in the y-direction, strain in the z-direction, and displacement.

[0070] S500. Using the initial simplified model, the first multi-channel mechanical cloud map, the second multi-channel mechanical cloud map, and the three-dimensional thin-layer structure as training samples, train the latent graph diffusion model to generate a cancellous bone scaffold inverse generation model.

[0071] In this embodiment, an initial simplified model, a first multi-channel mechanical cloud map, a second multi-channel mechanical cloud map, and a three-dimensional thin-layer structure are used as training samples. By employing a multi-modal fusion approach, the accuracy of the model can be further improved.

[0072] In some embodiments, the step of training a latent graph diffusion model to generate a cancellous bone scaffold inverse generation model by using the initial simplified model, the first multi-channel mechanical cloud map, the second multi-channel mechanical cloud map, and the three-dimensional thin-layer structure as training samples includes:

[0073] The stress cloud map in the x-direction, stress cloud map in the y-direction, stress cloud map in the z-direction, strain cloud map in the x-direction, strain cloud map in the y-direction, strain cloud map in the z-direction, and displacement cloud map in the second multi-channel mechanical cloud map are all divided into P image blocks.

[0074] The intermediate condition data is formed by combining the stress cloud map in the x direction, stress cloud map in the y direction, stress cloud map in the z direction, strain cloud map in the x direction, strain cloud map in the y direction, strain cloud map in the z direction, and displacement cloud map in the second multi-channel mechanical cloud map, as well as the P image blocks obtained from each division.

[0075] In this embodiment of the application, the stress cloud map, stress cloud map, stress cloud map, strain cloud map, strain cloud map, and displacement cloud map in the x-direction, y-direction, and z-direction of the second multi-channel mechanical cloud map, as well as the P image blocks obtained by dividing the stress cloud map in the x-direction, the stress cloud map in the y-direction, the stress cloud map in the z-direction, the strain cloud map in the x-direction, the strain cloud map in the y-direction, the strain cloud map in the z-direction, and the displacement cloud map, are combined into intermediate condition data.

[0076] Using the initial simplified model, the first multi-channel mechanical cloud map, intermediate condition data, and the three-dimensional thin-layer structure, a potential graph diffusion model is trained.

[0077] In some embodiments, prior to the step of training the latent graph diffusion model using the initial simplified model, the first multi-channel mechanical contour map, intermediate conditional data, and the three-dimensional thin-layer structure, the method further includes:

[0078] The first multi-channel mechanical cloud map and the intermediate condition data are masked using a random masking mechanism.

[0079] In this embodiment, in order to enhance the robustness of the cancellous bone scaffold inverse generation model, especially when dealing with incomplete data or occlusion, the training samples are masked when training the latent graph diffusion model, thereby improving the robustness and generalization ability of the cancellous bone scaffold inverse generation model to different types of missing information scenarios.

[0080] In some embodiments, there are two ways to mask the first multi-channel mechanical cloud map and the intermediate condition data using a random masking mechanism.

[0081] In the first method, the step of masking the first multi-channel mechanical cloud map and the intermediate condition data using a random masking mechanism includes:

[0082] The first multi-channel mechanical cloud map is masked using a first mask probability to obtain a masked first multi-channel mechanical cloud map; the intermediate condition data is masked using a second mask probability to obtain masked intermediate condition data.

[0083] By conducting multiple comparative studies, the impact of setting the first and second mask probabilities on the robustness and generation performance of the cancellous bone scaffold inverse generation model was evaluated. A comprehensive analysis of the adaptability of the cancellous bone scaffold inverse generation model to missing inputs under various probability combinations was conducted. Ultimately, a first mask probability of 0.2 and a second mask probability of 0.1 were selected. This results in a 20% probability masking of the stress contour maps in the x, y, and z directions, as well as the strain contour maps in the x, y, and z directions, and the displacement contour maps in the first multi-channel mechanical contour map. The seven channels in the intermediate conditional data—stress in the x, y, and z directions, strain in the x, y, and z directions, strain in the x, and displacement directions—are each masked with a 10% probability.

[0084] The second method involves masking the first multi-channel mechanical cloud map and the intermediate condition data using a random masking mechanism, which includes:

[0085] The stress cloud map in the x-direction, stress cloud map in the y-direction, stress cloud map in the z-direction, strain cloud map in the x-direction, strain cloud map in the y-direction, strain cloud map in the z-direction, and displacement cloud map in the first multi-channel mechanical cloud map are all divided into P image blocks.

[0086] The image block is masked using the third mask probability to obtain the masked image block;

[0087] The image composed of P masked image blocks is determined as the first multi-channel mechanical cloud map after masking.

[0088] Inject the intermediate condition data with Gaussian noise;

[0089] In this embodiment, Gaussian noise is injected into the intermediate condition data to simulate disturbances.

[0090] The stress contour maps in the x-direction, y-direction, z-direction, strain contour maps in the x-direction, y-direction, z-direction, and displacement contour maps in the intermediate condition data injected with Gaussian noise are all masked using the fourth mask probability to obtain the masked intermediate condition data.

[0091] In this way, the inverse generation model of the cancellous bone scaffold obtained after training can complete the data generation task in scenarios where the input structure is incomplete or contains noise, thereby improving its recovery ability and robustness in practical applications.

[0092] Replace the masked first multi-channel mechanical cloud map with the first multi-channel mechanical cloud map;

[0093] Replace the intermediate condition data after masking and use it as the intermediate condition data.

[0094] Thus, the cancellous bone scaffold inverse generation model, which is generated by training the latent graph diffusion model using the first multi-channel mechanical cloud map after masking and the intermediate condition data after masking, has high robustness and recovery ability.

[0095] In some embodiments, the latent graph diffusion model adopts a symmetrical U-shaped structure; the latent graph diffusion model includes a downsampling encoder and an upsampling decoder, and employs skip connections.

[0096] In this embodiment, a symmetrical U-shaped structure is used as the core framework, which includes a downsampling encoder and an upsampling decoder, and information transmission and fusion between features of different scales are maintained through skip connections.

[0097] The steps for training the latent graph diffusion model using the initial simplified model, the first multi-channel mechanical cloud map, intermediate condition data, and the three-dimensional thin-layer structure include:

[0098] The initial simplified model is encoded using the upsampling encoder of the latent graph diffusion model to obtain latent spatial features. Specifically, a multilayer perceptron can be used to encode the initial simplified model to obtain latent spatial features.

[0099] In some embodiments, the downsampling encoder includes a first graph convolutional network layer, a first self-attention layer, a first cross-attention layer, and a downsampling layer connected in sequence; the upsampling decoder includes an upsampling layer, a second self-attention layer, a second cross-attention layer, and a second graph convolutional network layer connected in sequence, wherein the first cross-attention layer and the second cross-attention layer introduce a conditional mapping mechanism, which is guided by the first multi-channel mechanical cloud map and intermediate conditional data.

[0100] In this embodiment, the latent graph diffusion model is built on a 3D mesh graph. Both the encoder and decoder modules contain four key layers: a Graph Convolutional Network (GCN) layer, a self-attention layer, a cross-attention layer, and a downsampling or upsampling layer. The GCN layer extracts local topological features from the 3D cancellous bone mesh structure by aggregating information from each node and its neighborhood. Stacking multiple GCN layers effectively expands the network's receptive field, thereby enhancing the model's understanding of the global structure. The self-attention layer enhances the model's ability to model long-distance dependencies between nodes, promoting the propagation of global information on the 3D mesh surface. The cross-attention layer further introduces external conditional vectors, dynamically adjusting the generated features through a conditional mapping mechanism, giving the network greater output flexibility and customization capabilities.

[0101] Through downsampling and upsampling operations, the latent graph diffusion model achieves compression and reconstruction of graph structural features in the latent space. Skip connections in the U-shaped structure effectively preserve shallow feature information, avoiding information loss common in deep networks. Simultaneously, the fusion of GCN and attention mechanisms enables the network to model both local fine structures and grasp overall morphology and performance constraints. Furthermore, the conditional mapping mechanism in the latent graph diffusion model introduces target performance parameters into the generation process, providing clear control signals for the generation of 3D thin-layer structures. The synergistic effect of these mechanisms enables the latent graph diffusion model to demonstrate excellent generalization and expressive power in tasks involving the generation of cancellous bone graph structures with strong structural-performance coupling and complex geometric morphology.

[0102] In this embodiment, the condition mapping mechanism, combined with the aforementioned random masking mechanism, can further enhance the robustness of the model. The introduction of a multimodal condition mapping mechanism with random masking perturbs the first multi-channel mechanical contour map and intermediate condition data during the input phase, improving the model's generalization ability and generation stability under incomplete or heterogeneous conditional inputs.

[0103] The GCN layer extracts local graph structure features by aggregating information from vertices in the 3D mesh graph, capturing relationships between nodes and information about neighboring nodes. The self-attention mechanism propagates global information in the 3D mesh graph, enabling each node to adjust its features according to the global context. The self-attention layer captures global information by calculating and weighting the similarity between nodes. The cross-attention mechanism injects external conditional information into the generative model; that is, through a conditional mapping mechanism, it provides additional contextual support for the generation process, allowing the model to adjust its output according to different conditions. The upsampling layer restores the number of nodes and feature dimensions through deconvolution or interpolation operations, reconstructing structural details.

[0104] Based on the aforementioned potential spatial features, a three-dimensional thin-layer structure is obtained using a downsampling decoder.

[0105] In this embodiment, by using the latent diffusion model to perform iterative sampling and denoising in the latent space, the nonlinear mapping relationship between the structure and performance of the three-dimensional thin-layer structure used to represent the cancellous bone scaffold structure can be effectively captured, demonstrating good performance and stability in the generation of complex biological structures.

[0106] The model in this embodiment outputs a three-dimensional thin-layer structure in practical applications. The three-dimensional thin-layer structure can be obtained through engineering modeling software to obtain the target three-dimensional spatial topology, and the target three-dimensional spatial topology can be used for printing.

[0107] The following experimental analysis further illustrates the effectiveness of the reverse generation model of the cancellous bone scaffold in this application.

[0108] To systematically evaluate the performance of the proposed inverse generation model for cancellous bone scaffolds (a conditional diffusion model) in the 3D cancellous bone scaffold generation task, particularly its effectiveness in dimensions such as geometric fidelity, topology preservation, and physical property representation and controllability, this study conducted multi-dimensional comparative experiments to quantify model performance. Furthermore, to verify the statistical significance of the experimental results, all experiments were repeated under 10 different random seeds and compared at a significance level. A paired t-test was used at a value of 0.05.

[0109] First, the performance of the cancellous bone scaffold inverse generation model was compared with that of three other models (Mesh-VAE model, ConditionalGAN model, and Uncond-DDPM model, the latter being an unconditional diffusion model). The evaluation metrics included Chamfer distance (CD), Hausdorff distance (HD), voxel crossover ratio (IoU), topology preservation score (TPS), and topology preservation score (SSIM). The specific results are shown in Table 1.

[0110] Table 1

[0111]

[0112] The results show that the cancellous bone scaffold inverse generation model outperforms the other three models in all evaluation metrics, demonstrating comprehensive and significant performance advantages, especially in topology preservation, where it achieved the highest score of 0.711. In terms of geometric accuracy, the cancellous bone scaffold inverse generation model has a CD of 0.136 and a HD of 0.754, significantly outperforming all other methods.

[0113] Then, to verify the impact of key modules on model performance, a module ablation experiment was conducted, sequentially removing the conditional mapping mechanism (w / oCondMap), the random masking mechanism (w / oMask), and the multimodal fusion mechanism (w / oMultiCond). Specific results are shown in Table 2.

[0114] Table 2

[0115]

[0116] Removing any module resulted in a decrease in generation quality, particularly in terms of geometric consistency and topology preservation. The CondMap module had the most significant impact on performance, increasing CD from 0.136 to 0.181 and decreasing TPS to 0.633, indicating that this module plays a crucial role in achieving conditional semantic alignment and structure-physical mapping. It significantly improves the model's ability to model physically perceived attributes by encoding and mapping multi-source conditional information.

[0117] The removal of the Mask module also led to a decrease in model robustness. Although its impact on geometric accuracy was slightly weaker than that of CondMap, the HD rose to 0.855, indicating that the model's ability to handle occluded or missing inputs was significantly weakened. This verifies the importance of occlusion modeling strategies in improving the model's generalization and anti-interference capabilities.

[0118] In contrast, the impact of the MultiCond module is relatively mild, but it still shows a steady decline in metrics such as IoU, TPS, and SSIM, indicating that multi-source conditional co-coding helps to extract more representative cross-modal structural features and is an important condition for achieving high-fidelity generation.

[0119] Furthermore, the model's generative control capability under different input combinations was verified. Four experimental configurations were designed: the first configuration input only the seven-channel stress-strain map of the initial simplified structure (i.e., the first multi-channel mechanical contour map); the second configuration input only the seven-channel stress-strain map of the intermediate structure (i.e., the second multi-channel mechanical model); the third configuration input the key physical channels of both configurations simultaneously (i.e., the first and second multi-channel mechanical contour maps); and the fourth configuration used a random masking mechanism for both the first and second multi-channel mechanical contour maps. The experimental results are shown in Table 3.

[0120] Table 3

[0121]

[0122] The results show that combining the key physical channels of the initial and intermediate structures with a single conditional input significantly improves the generation quality, increasing IoU from 0.784 and 0.792 to 0.835 and TPS from 0.641 and 0.659 to 0.686. Further combining the CondMap module with a random masking mechanism to form a complete conditional input configuration further enhances model performance, achieving an IoU of 0.869 and a TPS of 0.711, fully demonstrating the strong adaptability and generalization ability of the cancellous bone scaffold inverse generation model to complex and heterogeneous conditional inputs.

[0123] The following section presents a mechanical verification and simulation comparison analysis of the inverse generation model of the cancellous bone scaffold.

[0124] To verify the practical feasibility of the proposed method in the fields of bone tissue engineering and regenerative medicine, this study used biodegradable polymer materials and an extrusion-based 3D printing device to complete the in vitro fabrication experiment of the designed cancellous bone scaffold. Polylactic acid (PLA) was chosen as the printing material, widely used in bone repair due to its excellent biocompatibility. After inputting multiple target mechanical property profiles, the model generated corresponding thin-layer structures and recorded key structural parameters, such as the diameter of the support rods. With length Subsequently, the generated thin-layer structure was reconstructed using engineering modeling software, and scaffold samples were constructed in the X, Y, and Z directions, followed by uniaxial compression tests.

[0125] Mechanical property testing was performed using a universal testing machine, and its stress-strain curves are shown below. Figure 4 As shown in the figure. Experimental results show that the stress-strain curves of the structure in different directions have a high degree of consistency: in the initial stage, the stress increases linearly with the strain, indicating that the structure is in the elastic deformation stage; subsequently, the slope of the curve decreases, indicating that the structure yields and enters the topological collapse stage; finally, the curve tends to flatten, reflecting that the internal pores gradually close and enter the plastic deformation stage. This indicates that, under the same structural parameters, the generated three-dimensional scaffold exhibits isotropic mechanical response in all directions (X, Y, and Z), verifying the effectiveness of the cancellous bone scaffold inverse generation model in the generation of isotropic scaffold structures.

[0126] Furthermore, to verify the structure-performance consistency of the TriTopo-LGDM method, this study conducted finite element simulation (FEM) on the generated samples and obtained simulation results. Figure 4 The simulation results are represented by black dashed lines and compared with experimental curves. The results show that in the low-strain region, the simulation curves and experimental data are highly consistent, indicating that the proposed method has good predictive ability and modeling effect in the elastic stage. Both maintain a high degree of consistency in the overall trend, especially in the initial linear deformation stage, demonstrating that the structural design method proposed in this study has good design-manufacturing consistency and engineering feasibility.

[0127] To further evaluate the practical applicability of the scaffold performance, this study compared the experimental results with typical mechanical parameters of human cancellous bone. The compressive strength of human cancellous bone is approximately 2 MPa–20 MPa, and its elastic modulus is 50 MPa–200 MPa. In this study, the printed scaffold exhibited a compressive strength of 2 MPa–11 MPa and an elastic modulus of 30 MPa–270 MPa, with overall performance comparable to natural cancellous bone. This indicates that this method can achieve structure-performance design that matches natural tissue.

[0128] In terms of morphological characteristics, natural cancellous bone typically exhibits a Voronoi spatial grid pattern. The Voronoi, honeycomb, triangular grid, and parallelogram structural units selected in this study can effectively reconstruct the morphology of cancellous bone in different locations through structural parameter adjustments. Simultaneously, the printed scaffolds meet the dimensional requirements of cancellous bone, demonstrating the ability to construct diverse scaffolds to meet different mechanical needs, anatomical locations, and scales, showcasing broad application potential in the field of biomanufacturing.

[0129] In summary, the reverse generation model for cancellous bone scaffolds successfully generated cancellous bone scaffold structures that meet mechanical performance requirements. The output structure has good isotropic mechanical properties and engineering adaptability, providing theoretical support for the high-quality design and manufacturing of personalized cancellous bone scaffolds.

[0130] The method in this application embodiment is based on a data-driven mechanism of "structure-physics-optimization" triple alignment, which directly maps mechanical performance requirements to personalized 3D scaffold generation, thus constructing a new intelligent reverse design process for bone defect repair tasks. Relying on the structure-physics-optimization triple alignment dataset, this method realizes reverse structure generation guided by target mechanical performance, effectively constructing a cancellous bone scaffold structure that is highly similar to natural bone tissue in terms of mechanics and morphology.

[0131] The above embodiments provide a method and apparatus for determining a reverse generation model of cancellous bone scaffold based on topology optimization. In this method, the three-dimensional thin-layer structure is topology optimized to mimic the complex and irregular microstructural features of cancellous bone tissue. The first and second multi-channel mechanical contour maps represent force-related features. Multimodal data consisting of the three-dimensional thin-layer structure, the initial simplified model, the first multi-channel mechanical contour map, and the second multi-channel mechanical contour map are used as training samples to train the latent graph diffusion model. This allows the model to achieve strong performance in outputting cancellous bone scaffold results under multimodal input conditions.

[0132] This application also provides a device for determining a reverse generation model of a cancellous bone scaffold based on topology optimization. Figure 5 An exemplary schematic diagram of a device for determining a reverse generation model of a cancellous bone scaffold based on topology optimization, according to some embodiments, is shown.

[0133] A device for determining a reverse generation model of a cancellous bone scaffold based on topology optimization includes:

[0134] Construction unit 501 is used to construct the corresponding initial three-dimensional spatial topology based on a variety of spatial topology units;

[0135] Simplification unit 502 is used to simplify the initial three-dimensional spatial topology into a corresponding initial simplified model;

[0136] The generation unit 503 is used to generate various three-dimensional thin-layer structures from the initial simplified model by adjusting the structural parameters, and to determine the intermediate simplified model in the process of generating the three-dimensional thin-layer structures; the structural parameters include the support rod diameter, support rod length, and load applied under simulation boundary conditions;

[0137] Extraction unit 504 is used to extract the initial simplified model and the intermediate simplified model using finite element analysis, respectively, to obtain a first multi-channel mechanical cloud map and a second multi-channel mechanical cloud map; the first multi-channel mechanical cloud map and the second multi-channel mechanical cloud map both include stress cloud map in the x direction, stress cloud map in the y direction, stress cloud map in the z direction, strain cloud map in the x direction, strain cloud map in the y direction, strain cloud map in the z direction and displacement cloud map;

[0138] Training unit 505 is used to train the latent graph diffusion model by using the initial simplified model, the first multi-channel mechanical cloud map, the second multi-channel mechanical cloud map, and the three-dimensional thin-layer structure as training samples to generate a cancellous bone scaffold inverse generation model.

[0139] In this embodiment of the application, a computer device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the method for determining the inverse generation model of cancellous bone scaffold based on topology optimization.

[0140] In this embodiment of the application, a computer-readable storage medium is also provided, which stores a computer program that, when executed by a processor, implements the steps of determining the inverse generation model of the cancellous bone scaffold.

[0141] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platform, or it can be implemented by hardware.

[0142] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the units or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the units in the apparatus of the embodiment can be distributed within the apparatus of the embodiment as described, or can be located in one or more apparatuses different from this embodiment, with corresponding changes. The units of the above-described embodiment can be combined into one unit, or further divided into multiple sub-units.

[0143] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any variations conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A method for determining a reverse generation model of a cancellous bone scaffold based on topology optimization, characterized in that, The method comprises the following steps: Based on a plurality of spatial topology units, an initial three-dimensional spatial topology corresponding to the initial three-dimensional spatial topology is constructed; The initial three-dimensional spatial topology is simplified into an initial simplified model corresponding to the initial three-dimensional spatial topology; A plurality of three-dimensional thin layer structures are generated by adjusting structure parameters of the initial simplified model, and intermediate simplified models in the process of generating the three-dimensional thin layer structures are determined; the structure parameters include support rod diameter, support rod length and load applied under simulation boundary conditions; The initial simplified model and the intermediate simplified model are extracted by finite element analysis, and first and second multi-channel mechanical cloud maps are obtained; the first and second multi-channel mechanical cloud maps each include a stress cloud map in the x direction, a stress cloud map in the y direction, a stress cloud map in the z direction, a strain cloud map in the x direction, a strain cloud map in the y direction, a strain cloud map in the z direction and a displacement cloud map; The initial simplified model, the first multi-channel mechanical cloud map, the second multi-channel mechanical cloud map and the three-dimensional thin layer structure are used as training samples to train a latent graph diffusion model to generate a cancellous bone scaffold reverse generation model.

2. The method of claim 1, wherein, The step of training the latent graph diffusion model to generate the cancellous bone scaffold reverse generation model using the initial simplified model, the first multi-channel mechanical cloud map, the second multi-channel mechanical cloud map and the three-dimensional thin layer structure as training samples comprises: The stress cloud map in the x direction, the stress cloud map in the y direction, the stress cloud map in the z direction, the strain cloud map in the x direction, the strain cloud map in the y direction, the strain cloud map in the z direction and the displacement cloud map in the second multi-channel mechanical cloud map are each divided into P image blocks; The stress cloud map in the x direction, the stress cloud map in the y direction, the stress cloud map in the z direction, the strain cloud map in the x direction, the strain cloud map in the y direction, the strain cloud map in the z direction and the displacement cloud map in the second multi-channel mechanical cloud map are combined with the P image blocks obtained by division to form intermediate condition data; The initial simplified model, the first multi-channel mechanical cloud map, the intermediate condition data and the three-dimensional thin layer structure are used to train the latent graph diffusion model.

3. The method of claim 2, wherein, Before the step of training the latent graph diffusion model using the initial simplified model, the first multi-channel mechanical cloud map, the intermediate condition data and the three-dimensional thin layer structure, the method further comprises: The first multi-channel mechanical cloud map and the intermediate condition data are subjected to mask processing using a random mask mechanism; The first multi-channel mechanical cloud map after mask processing is replaced and used as the first multi-channel mechanical cloud map; The intermediate condition data after mask processing is replaced and used as the intermediate condition data.

4. The method of claim 3, wherein, The step of subjecting the first multi-channel mechanical cloud map and the intermediate condition data to mask processing using a random mask mechanism comprises: The first multi-channel mechanical cloud map is subjected to mask processing using a first mask probability to obtain the first multi-channel mechanical cloud map after mask processing; The intermediate condition data is subjected to mask processing using a second mask probability to obtain the intermediate condition data after mask processing.

5. The method of claim 3, wherein, The step of subjecting the first multi-channel mechanical cloud map and the intermediate condition data to mask processing using a random mask mechanism comprises: The stress cloud map in the x direction, the stress cloud map in the y direction, the stress cloud map in the z direction, the strain cloud map in the x direction, the strain cloud map in the y direction, the strain cloud map in the z direction and the displacement cloud map in the first multi-channel mechanical cloud map are all divided into P image blocks; The image blocks are masked by using the third mask probability to obtain the masked image blocks; An image composed of the P masked image blocks is determined as the first multi-channel mechanical cloud map after mask processing; The intermediate conditional data is injected with Gaussian noise; The stress cloud map in the x direction, the stress cloud map in the y direction, the stress cloud map in the z direction, the strain cloud map in the x direction, the strain cloud map in the y direction, the strain cloud map in the z direction and the displacement cloud map in the intermediate conditional data injected with Gaussian noise are all masked by using the fourth mask probability to obtain the intermediate conditional data after mask processing.

6. The method of claim 2, wherein, The latent graph diffusion model adopts a symmetric U-shaped structure; the latent graph diffusion model comprises a down-sampling encoder and an up-sampling decoder, and adopts a skip connection; The step of training the latent graph diffusion model by using the initial simplified model, the first multi-channel mechanical cloud map, the intermediate conditional data and the three-dimensional thin layer structure comprises: The initial simplified model is encoded by using the up-sampling encoder of the latent graph diffusion model to obtain latent space features; Based on the latent space features, the three-dimensional thin layer structure is obtained by using the down-sampling decoder.

7. The method of claim 6, wherein, The down-sampling encoder comprises a first graph convolutional network layer, a first self-attention layer, a first cross-attention layer and a down-sampling layer connected in sequence; the up-sampling decoder comprises an up-sampling layer, a second self-attention layer, a second cross-attention layer and a second graph convolutional network layer connected in sequence, the first cross-attention layer and the second cross-attention layer introduce a conditional mapping mechanism, and the conditional mapping mechanism is guided by the first multi-channel mechanical cloud map and the intermediate conditional data.

8. A device for determining a reverse generation model of a cancellous bone scaffold based on topology optimization, characterized in that, It comprises: A construction unit is configured to construct an initial three-dimensional spatial topological structure corresponding to a plurality of spatial topological structure units; A simplification unit is configured to simplify the initial three-dimensional spatial topological structure into a corresponding initial simplified model; A generation unit is configured to generate a plurality of three-dimensional thin layer structures by adjusting structure parameters of the initial simplified model, and determine an intermediate simplified model in the process of generating the three-dimensional thin layer structures; the structure parameters include a support rod diameter, a support rod length and a load applied under a simulation boundary condition; An extraction unit is configured to extract the initial simplified model and the intermediate simplified model by using finite element analysis to obtain a first multi-channel mechanical cloud map and a second multi-channel mechanical cloud map; the first multi-channel mechanical cloud map and the second multi-channel mechanical cloud map each include a stress cloud map in the x direction, a stress cloud map in the y direction, a stress cloud map in the z direction, a strain cloud map in the x direction, a strain cloud map in the y direction, a strain cloud map in the z direction and a displacement cloud map; A training unit is configured to train a latent graph diffusion model by using the initial simplified model, the first multi-channel mechanical cloud map, the second multi-channel mechanical cloud map and the three-dimensional thin layer structure as training samples to generate a cancellous bone scaffold reverse generation model.

Images