A three-dimensional organ-oriented geometric data processing method and device

Abstract

The application provides a three-dimensional organ-oriented geometric data processing method and device, and relates to the technical field of three-dimensional image processing. The method comprises the following steps: acquiring an original organ three-dimensional mesh model, performing feature extraction and topological optimization, and constructing an initial embryo library with anatomical representation; extracting features of a three-dimensional organ target geometric model to be registered, calculating the similarity of the features of the three-dimensional organ target geometric model and the features of the embryos in the initial embryo library, and selecting a corresponding target embryo according to the similarity; fusing the features of the three-dimensional organ target geometric model and the features of the target embryo, predicting and acquiring a plurality of candidate registration strategy paths based on the fused features; deforming and registering the target embryo to the three-dimensional organ target geometric model in parallel according to the plurality of candidate registration strategy paths, generating a plurality of candidate registration geometric results, and screening a target three-dimensional organ geometric model meeting a quality threshold based on the plurality of candidate registration geometric results.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional image processing technology, and in particular to a geometric data processing method and apparatus for three-dimensional organs. Background Technology

[0002] In the field of traditional medical 3D geometric modeling, early methods typically relied on voxel-to-mesh technology or statistical shape models for surface mesh extraction. Traditional techniques based on algorithms such as Marching Cubes can directly extract surfaces from 3D voxel data, but the generated meshes generally suffer from problems such as chaotic topology, redundant vertex numbers, and inconsistent element quality, failing to meet the stringent requirements of biomechanical simulation and numerical computation. While methods based on statistical shape models (SSM) can construct shape representations with consistent topology and regular morphology, they heavily rely on a large number of high-quality, manually annotated medical image templates. This not only results in high annotation costs and extremely cumbersome processes but also makes it difficult to achieve efficient expansion and practical application in large-scale unannotated medical image data.

[0003] With the development of deep learning technology, existing 3D reconstruction has gradually evolved into end-to-end reconstruction methods driven by graph convolutional networks or implicit neural representations. For example, Pixel2Mesh or models based on continuous signed distance functions achieve direct deformation and representation from images to 3D surfaces, greatly improving the automation of reconstruction. However, these end-to-end models are prone to problems such as mesh self-intersection, topological breaks, and local morphological distortion when processing medical images with strict anatomical constraints (such as thin-walled organs and complex vascular branches). Especially in the absence of multi-level registration and morphological constraints, simple end-to-end mesh deformation cannot guarantee the mesh cell quality and geometric rationality necessary for subsequent biomechanical analysis.

[0004] While numerous tools have emerged in the field of medical image processing capable of fully automated segmentation of large-scale images, a significant technological gap remains between the raw segmented images and the high-quality geometric models directly usable for mechanical simulation. Existing medical modeling solutions often only address local organ reconstruction or 2D image data synthesis, and the generated synthetic data largely lacks topological consistency with real anatomical structures. Therefore, current technology lacks a closed-loop pipeline method that can completely eliminate reliance on manual annotation and achieve fully automated conversion from massive amounts of raw medical images into standardized, topologically consistent, and anatomically faithful simulateable 3D geometric models. This results in a severe deficiency in the large-scale production chain of high-quality medical 3D mesh datasets. Summary of the Invention

[0005] This invention provides a geometric data processing method and apparatus for three-dimensional organs, which addresses the shortcomings of existing technologies that lack a closed-loop pipeline method that can completely eliminate the dependence on manual annotation and achieve fully automatic conversion from massive amounts of raw medical images into standardized, topologically consistent, and anatomically faithful simulateable three-dimensional geometric models, thereby realizing the large-scale and automated production of high-quality medical three-dimensional mesh datasets.

[0006] This invention provides a geometric data processing method for three-dimensional organs, comprising: Obtain the original organ 3D mesh model, perform feature extraction and topology optimization, and construct an initial embryonic body library with anatomical representativeness; Features of the three-dimensional organ target geometric model to be registered are extracted, the similarity between the features of the three-dimensional organ target geometric model and the features of embryos in the initial embryo database is calculated, and the corresponding target embryo is selected based on the similarity. The features of the three-dimensional organ target geometric model are fused with the features of the target embryo, and multiple candidate registration strategy paths are obtained based on the fused features. The target embryo is deformed and registered to the three-dimensional organ target geometric model in parallel according to multiple candidate registration strategy paths, generating multiple candidate registration geometric results. Based on the multiple candidate registration geometric results, the target three-dimensional organ geometric model that meets the quality threshold is selected.

[0007] The present invention also provides a geometric data processing device for three-dimensional organs, comprising: The module is used to acquire the original organ 3D mesh model, perform feature extraction and topology optimization, and build an initial embryonic body library with anatomical representativeness; The calculation module is used to extract the features of the three-dimensional organ target geometric model to be registered, calculate the similarity between the features of the three-dimensional organ target geometric model and the features of embryos in the initial embryo database, and select the corresponding target embryo based on the similarity. The fusion module is used to fuse the features of the three-dimensional organ target geometric model with the features of the target embryo, and to obtain multiple candidate registration strategy paths based on the fused features. The processing module is used to perform deformation registration of the target embryo to the three-dimensional organ target geometric model in parallel according to multiple candidate registration strategy paths, generate multiple candidate registration geometric results, and select the target three-dimensional organ geometric model that meets the quality threshold based on the multiple candidate registration geometric results.

[0008] The present invention also provides an electronic 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 geometric data processing method for three-dimensional organs as described above.

[0009] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the geometric data processing method for three-dimensional organs as described above.

[0010] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the geometric data processing method for three-dimensional organs as described above.

[0011] The geometric data processing method and apparatus for three-dimensional organs provided by this invention reduces the reliance on manually labeled templates by constructing an initial embryonic body library with anatomical representativeness and automatically selecting target embryonic bodies using feature similarity matching. It fuses the features of the target geometric model with the features of the target embryonic body to predict multiple candidate registration strategy paths, providing various deformation guidance directions for different anatomical structures and reducing the risk of mesh self-intersection or topological breakage in complex organs (such as thin-walled tissues and branching blood vessels) using a single registration method. Parallel deformation registration according to multiple candidate registration strategy paths and selection of the target three-dimensional organ geometric model based on a quality threshold help improve the success rate of the registration process and the fidelity of the output geometry, thus forming an automated processing link from the original mesh to a simulateable geometric model. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating the geometric data processing method for three-dimensional organs provided by the present invention.

[0014] Figure 2 This is a schematic diagram of the overall structure of the decision network model provided by the present invention.

[0015] Figure 3 This is an overall flowchart of the geometric data processing method for three-dimensional organs provided by the present invention.

[0016] Figure 4This is a comparison chart of the three-dimensional reconstruction effects of the research method provided by this invention and the traditional Marching Cubes algorithm on various organs.

[0017] Figure 5 This is a schematic diagram of the geometric data processing device for three-dimensional organs provided by the present invention.

[0018] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] In existing medical 3D geometric modeling practices, methods based on voxel-to-mesh conversion (such as the MarchingCubes algorithm), while capable of directly extracting surface meshes from segmentation masks, generally suffer from irregular topological structures, redundant vertex counts, and an excessive number of elongated triangular faces. Without extensive post-processing repair, these meshes are difficult to use directly for finite element calculations in biomechanical simulations, as simulation solvers have stringent requirements for element quality (such as Jacobian and aspect ratio). On the other hand, while methods based on statistical shape models can guarantee topological consistency, their construction heavily relies on manually labeled template sets—each organ category requires dozens or even hundreds of precisely labeled images, resulting in extremely high labeling costs. Furthermore, when encountering novel anatomical variants or pathological deformities, the expansion of the template library also requires manual intervention. In recent years, deep learning methods such as Pixel2Mesh and Voxel2Mesh have attempted to perform end-to-end deformation of initial shapes (such as spheres) through graph convolutional networks. However, this is essentially a "black box" optimization, lacking explicit constraints on anatomical structures. When processing thin-walled organs (such as the gallbladder and bladder) or complex branching structures (such as intrahepatic vessels and bronchial trees), unacceptable geometric distortions such as mesh self-intersections, surface breaks, or local bulges / depressions are often produced. Furthermore, while existing automatic segmentation tools (such as TotalSegmentator) can output organ masks, the conversion from masks to simulated meshes still relies on manual or semi-automatic repair processes, resulting in a significant technological gap in the construction of large-scale, high-quality geometric datasets.

[0021] To address the aforementioned issues, this application provides a geometric data processing method for three-dimensional organs. The method first constructs an initial embryonic body library offline, representing anatomical structures. Each embryonic body in this library possesses a uniform topology and high-quality mesh cells. Then, for the input target geometric model of the organ to be registered, its features are extracted and matched with similarity data from embryonic bodies in the library, automatically selecting the most suitable target embryonic body. Next, the features of the target geometric model and the target embryonic body are deeply fused, and a trained decision model is used to predict multiple candidate registration strategy paths. Finally, the deformation registration under these paths is performed in parallel, and the optimal target three-dimensional organ geometric model that meets a preset quality threshold is selected from multiple registration results. Through this approach, this method can automatically transform raw, chaotic three-dimensional organ meshes into topologically consistent, high-quality, and anatomically accurate simulateable geometric models without relying on manual annotation.

[0022] Before describing the technical solutions of the embodiments of the present invention, the terms and concepts involved in the embodiments of the present invention will be explained illustratively.

[0023] The original 3D organ mesh model refers to the initial 3D surface mesh directly extracted from medical image segmentation masks, such as the mesh reconstructed using the Marching Cubes algorithm. This type of mesh typically exhibits inherent characteristics such as redundant vertex counts (potentially reaching tens or even hundreds of thousands of vertices), uneven distribution of triangular faces (containing elongated or degenerate triangles), potential topological noise (e.g., isolated small holes or non-manifold edges), or non-watertightness (the mesh surface is not closed). Due to these issues, the original mesh generally cannot be directly used for biomechanical simulations and requires simplification and quality optimization.

[0024] Initial embryonic body library: This refers to an offline collection of 3D organ models containing multiple units with a unified topology and high-quality mesh elements. Each embryonic body (i.e., a representative organ geometry) in the library possesses excellent Jacobian quality, mesh aspect ratio, and other metrics at a fixed vertex size (e.g., 1000 vertices), and all embryonic bodies share the same triangular facet connectivity (i.e., consistent vertex index topological patterns). This embryonic body library serves as a baseline template library for subsequent deformation and registration, and can be repeatedly used, avoiding the need to rebuild templates every time new samples are processed.

[0025] 3D organ target geometry model: This refers to the 3D mesh of the target organ to be processed, usually obtained from image data of a new patient through segmentation and surface reconstruction. This model is an instance of the original 3D mesh model of the organ, and its mesh quality may have defects (such as vertex redundancy, topological irregularities, etc.), but it needs to be registered with embryos in the embryo library to generate the final standardized geometry that meets the simulation requirements.

[0026] Candidate registration strategy paths refer to the combinations of different registration methods or parameter configuration sequences used in the process of deforming the target embryo into a three-dimensional organ target geometric model. Examples include rigid alignment based on PCA, feature matching based on RANSAC, and similarity transformation with scale transformation. Each strategy corresponds to an independent registration path. Since different organ morphologies (such as spheroidal kidneys and tubular blood vessels) may be suitable for different registration strategies, providing multiple candidate paths can increase the probability of finding the optimal registration scheme and avoid the failure of a single strategy.

[0027] Candidate registration geometry results: These refer to the 3D mesh models generated after performing deformation registration according to a specific candidate registration strategy path. Multiple candidate paths can generate multiple candidate results in parallel, and each result may differ in registration accuracy (e.g., whether it fits the target shape) and mesh quality (e.g., whether there is self-intersection or distortion). Subsequently, a multi-dimensional quality evaluation is conducted to select the best candidate as the final output.

[0028] Edge collapse random iteration processing: During mesh simplification, different simplification paths are generated by randomly changing the order and step size weight of edge collapse operations multiple times. In each iteration, an edge is randomly selected, and its two endpoints are merged into a single vertex (i.e., collapsed), thereby reducing one vertex and two triangular faces. By repeatedly executing and recording the simplification results under different orders, the system can compare the mesh quality obtained from different paths and select the optimal simplification scheme, avoiding the problem of traditional greedy algorithms easily getting trapped in local optima.

[0029] Simplified meshes refer to low-resolution meshes obtained by reducing the number of vertices and faces in an original high-resolution 3D mesh. Simplified meshes retain key geometric features (such as the overall outline of organs and major anatomical landmarks) while having significantly fewer vertices (e.g., reduced from tens of thousands to 1000) and a unified topological connectivity. This lightweight representation facilitates rapid subsequent registration calculations, finite element simulations, and deep learning model training.

[0030] Jacobian quality: This refers to the degree of distortion of a mesh element (triangle or tetrahedron) from its ideal shape. It is typically evaluated by calculating the determinant of the element's Jacobian matrix, which reflects the volume scaling factor of the element during local coordinate transformations. Ideally, an equilateral triangle has a Jacobian value of 1; when the element is extremely stretched or compressed, the value approaches 0; if the element is flipped (i.e., the normal direction is reversed), the value becomes negative. The closer the Jacobian quality is to 1, the better the element quality, and the higher the stability and accuracy of the simulation calculation.

[0031] Mesh aspect ratio: This refers to the ratio of the longest side length to the shortest side length of a mesh element. For triangular elements, the aspect ratio reflects their slenderness: the aspect ratio of an equilateral triangle is 1, while the aspect ratio of an extremely narrow needle-shaped triangle can reach tens or even hundreds. The closer the aspect ratio is to 1, the more regular the element shape, which is beneficial to the stability and convergence of numerical calculations; an excessively large aspect ratio can lead to ill-conditioned stiffness matrix in finite element analysis, thereby reducing computational accuracy or even causing simulation failure.

[0032] Geometric fidelity refers to the degree of similarity in anatomical morphology between the simplified or deformed mesh and the original mesh. It is typically quantified by comparing the positional deviations of key anatomical landmarks (such as organ extrema, bifurcation points, and corners), the average chamfer distance between the two mesh surfaces, or the consistency of surface curvature distribution. Higher geometric fidelity indicates that the processed mesh better preserves the true anatomical features of the original organ.

[0033] Local curvature difference: This refers to the difference in the degree of surface curvature between two mesh models in corresponding local regions (such as the apex, edge, or branching point of an organ). Curvature describes the degree of curvature of a surface at a point and can take various forms, such as principal curvature, Gaussian curvature, or mean curvature. This difference can be used to measure the matching accuracy of the mesh on local anatomical structures, such as evaluating whether the deformed embryonic mesh accurately reproduces the fine features of the target organ, such as depressions, protrusions, or tubular bifurcations.

[0034] Decision network models refer to deep learning-based neural network models used to predict the scores or optimal probabilities of various registration strategies based on the input target geometric model features and embryonic features. The decision network in this application adopts a Transformer-like architecture, integrating a sparse multi-head attention mechanism (allowing interaction only between different types of features, rather than full connectivity) and a hybrid expert structure (MoE, replacing traditional MLP layers with multiple expert subnetworks, dynamically selected for activation by a routing mechanism). This model acquires the ability to understand organ geometry through multi-task training (including self-supervised reconstruction, supervised classification, and reinforcement learning).

[0035] Feature fusion information refers to the comprehensive feature representation obtained during the feature interaction process of the decision network. This representation incorporates the correspondence between the high-dimensional features of the target geometric model, the high-dimensional features of the target embryo, and the classification label features through cross-attention computation. This information not only includes their independent geometric descriptions but, more importantly, implies their matching relationships in shape, pose, and local structure. This fused feature representation is then used by the subsequent prediction branch to output the registration strategy score and deformation displacement.

[0036] Self-supervised reconstruction task: This refers to a training task that does not require manual labeling. In this task, the model uses the target embryo as the initial shape and reconstructs it into a target geometric model by predicting deformation displacement. The similarity between the reconstructed mesh and the original target geometry (such as chamfer distance) is used as the loss function. This task allows the model to be pre-trained using a large amount of unlabeled organ geometry data, thereby learning a general 3D shape representation ability, providing a good feature foundation for subsequent supervised tasks.

[0037] Supervised learning classification tasks refer to the task of training a decision network model to classify input geometric pairs using pre-labeled optimal registration strategy labels. Specifically, for each input pair (target embryo, target geometric model), multiple candidate registration strategies are pre-executed for actual registration testing. Based on the quantitative evaluation score, the optimal registration strategy is determined and used as the true label. The network model outputs the probability distribution of each registration strategy and calculates a negative log-likelihood loss with the true label, thereby learning the ability to select the optimal registration strategy for different anatomical morphologies.

[0038] Reinforcement learning policy selection task: This refers to modeling registration policy selection as a reinforcement learning problem, where the state is the input geometric features (i.e., the feature vectors of the target geometric model and the target embryo), the action is different registration policies (e.g., policies 1-8), and the reward is the geometric deformation evaluation score obtained after executing the policy (combining registration accuracy and mesh quality). Algorithms such as PPO (Proximal Policy Optimization) are used to train the model, allowing it to learn through interaction with the environment (i.e., the registration process): striking a balance between exploring new policies and utilizing known good policies, thereby optimizing the long-term cumulative reward.

[0039] Geometric deformation evaluation score: This refers to the score obtained by quantitatively evaluating the geometric results of candidate registrations generated after implementing a certain registration strategy. Evaluation dimensions typically include registration accuracy (e.g., chamfer distance, average deviation of anatomical landmarks) and mesh quality (e.g., Jacobian quality mean, aspect ratio variance, whether the mesh is watertight, presence of self-intersecting or non-manifold edges). The scores of each dimension can be weighted and summed according to preset weights to obtain the total score. This score can serve as a reward signal in reinforcement learning and can also be used to determine the quality threshold for ultimately selecting the optimal registration result.

[0040] The method provided in this invention can be executed by any electronic device with data storage and computing capabilities. For example, it can be a standalone personal computer, workstation, server (including a single server or server cluster), a virtual computing instance deployed on a cloud computing platform, or a dedicated processing module integrated into medical imaging equipment. This electronic device typically includes at least: one or more processors (such as CPU, GPU, or dedicated AI acceleration chip), a memory (such as RAM, ROM, hard disk, or solid-state drive) for storing data and program instructions, a communication interface for input / output data (such as USB interface, network interface), and an internal communication bus. The processor executes the steps described in this method (including feature extraction, cluster matching, deep learning model inference, parallel registration, and quality assessment) by calling computer program instructions stored in the memory. In specific implementations, different stages of the method (such as offline embryo library construction and online registration inference) can be executed on the same physical device or distributed across different devices (e.g., building the embryo library on an offline server and performing the registration task on an online service node), and this application does not limit this. Those skilled in the art will understand that the technical solution of the present invention can be achieved as long as sufficient computing and storage resources are available and the logical flow disclosed in the embodiments of the present invention is followed.

[0041] The specific implementation steps of the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be noted that the processes, parameters, and model structures described in the following embodiments are merely illustrative examples, and those skilled in the art can make adjustments according to actual application scenarios without departing from the core ideas of this application.

[0042] Figure 1 This is one of the flowcharts illustrating a geometric data processing method for three-dimensional organs provided in an embodiment of the present invention. The method includes the following: Step 101: Obtain the original organ three-dimensional mesh model, perform feature extraction and topology optimization, and construct an initial embryonic body library with anatomical representativeness.

[0043] In this embodiment, the raw data comes from an open-source medical imaging dataset (e.g., TotalSegmentator), selecting image data from 500 patients, covering 100 types of anatomical organs, resulting in approximately 50,000 anatomical geometric samples. First, the Marching Cubes algorithm is used to reconstruct the initial 3D mesh model of each organ from the original segmentation mask. To ensure the quality of the subsequent embryonic body library construction, this embodiment performs a rigorous data cleaning process: through semi-automated geometric integrity verification, topological noise, non-watertight structures (i.e., meshes with holes or boundaries), and incomplete organ geometry due to limitations in the image scanning field of view are removed. After cleaning, multidimensional statistical analysis is performed on the remaining high-quality geometric subset to calculate indicators such as organ volume distribution, surface area consistency, and curvature complexity, providing a basis for balanced sampling of the subsequent embryonic body library.

[0044] In this embodiment, a batch of original organ 3D mesh models are first obtained. These models can be obtained by surface reconstruction from the segmentation mask of medical image data (such as CT and MRI), for example, by extracting the mesh from the segmentation results using the Marching Cubes algorithm. Since directly reconstructed meshes usually have problems such as too many vertices, uneven distribution of triangular facets, local noise, or non-manifold structures, they cannot be directly used for subsequent deformation registration and simulation calculations. Therefore, they need to be processed to construct a high-quality initial embryonic body library.

[0045] Specifically, features are extracted from each original mesh and mapped to a high-dimensional feature space to characterize its global shape properties. Then, based on the extracted features, meshes with anatomical representativeness are selected from a large sample as candidate embryos. Topology optimization is performed on these candidate embryos, simplifying the number of vertices and faces of the mesh while preserving the main morphological features of the organ, resulting in lightweight models with a unified topology and reasonable unit quality. These optimized models are stored in an embryo database, ensuring that each embryo in the database represents the commonalities of a certain type of anatomical morphology while possessing good mesh quality, providing a stable and standardized initial template for subsequent registration.

[0046] Step 102: Extract the features of the three-dimensional organ target geometric model to be registered, calculate the similarity between the features of the three-dimensional organ target geometric model and the features of embryos in the initial embryo database, and select the corresponding target embryo based on the similarity.

[0047] In this embodiment, when a new 3D organ target geometric model to be registered is received (e.g., a mesh reconstructed from image data of a new patient), the features of the target geometric model are first extracted. The feature extraction method is consistent with that used when constructing the embryonic body library, so as to allow for comparison in the same feature space. Subsequently, the similarity between the feature vector of the target geometric model and the feature vectors of each embryonic body in the initial embryonic body library is calculated. The similarity can be measured using cosine similarity, Euclidean distance, or other applicable metrics. Based on the calculated similarity, an embryonic body that is closest to the target geometric model is selected from the embryonic body library as the target embryonic body. This target embryonic body will serve as the starting point for deformation registration in subsequent steps.

[0048] This feature similarity matching mechanism can automatically find the most suitable initial template for input organs with different shapes, avoiding the subjectivity and inefficiency of manually specifying templates.

[0049] Step 103: The features of the three-dimensional organ target geometric model and the features of the target embryo are fused together, and multiple candidate registration strategy paths are obtained based on the fused features.

[0050] In this embodiment, the features of the three-dimensional organ target geometric model obtained in step 102 are fused with the features of the selected target embryo. The purpose of fusion is to allow sufficient information exchange between the two features, enabling the model to understand the morphological differences and correspondences between the target shape and the template shape. The fusion process can employ feature concatenation, weighted summation, or more complex interaction methods (e.g., mutual attention based on attention mechanisms). Based on the comprehensive feature representation obtained after fusion, multiple candidate registration strategy paths are further predicted and obtained. Each candidate registration strategy path represents a specific scheme for deforming the target embryo to the target geometric model, such as different rigid alignment methods, different non-rigid deformation parameters, or different optimization sequences.

[0051] The advantage of providing multiple candidate paths instead of a single path is that different anatomical structures may be suitable for different registration strategies. Executing multiple paths in parallel can increase the likelihood of finding high-quality registration results and avoid registration failure due to the inapplicability of a single strategy.

[0052] Step 104: Perform deformation registration of the target embryo to the three-dimensional organ target geometric model in parallel according to multiple candidate registration strategy paths to generate multiple candidate registration geometric results. Based on the multiple candidate registration geometric results, select the target three-dimensional organ geometric model that meets the quality threshold.

[0053] In this embodiment, for each candidate registration strategy path predicted in step 103, the system synchronously initiates an independent registration process. Each process performs deformation registration on the target preform from its initial pose to the target geometric model according to the registration scheme defined by the corresponding path. The registration process typically includes global rigid body transformations (such as rotation, translation, and scaling) to align the overall position and orientation, and local non-rigid deformations to conform to the fine surface structure of the target geometry. After each path is executed, a corresponding candidate registration geometry result is generated.

[0054] Because the registration results of different paths may vary (for example, some paths have high registration accuracy but average mesh quality, while others have the opposite), a multi-dimensional quality evaluation is required for all candidate results. Evaluation metrics may include the surface distance between the registered mesh and the target geometry, the shape quality of mesh cells (such as whether there is excessive distortion or self-intersection), and whether watertightness is met. Based on a preset quality threshold, the geometric model that meets the requirements is selected from multiple candidate results as the final target 3D organ geometric model.

[0055] If multiple results meet the threshold, the one with the highest overall score can be selected as the output; if no result meets the threshold, an anomaly handling mechanism is triggered, such as recording failed samples for subsequent analysis or manual intervention. This combination of parallel registration and quality screening effectively improves the registration success rate and the reliability of the output geometry.

[0056] The geometric data processing method for three-dimensional organs provided in this invention reduces reliance on manually labeled templates by constructing an initial embryonic body library with anatomical representativeness and automatically selecting target embryonic bodies using feature similarity matching. It fuses the features of the target geometric model with the features of the target embryonic body to predict multiple candidate registration strategy paths, providing various deformation guidance directions for different anatomical structures and reducing the risk of mesh self-intersection or topological breaks in complex organs (such as thin-walled tissues and branching blood vessels) using a single registration method. Parallel deformation registration according to multiple candidate registration strategy paths and selection of the target three-dimensional organ geometric model based on a quality threshold help improve the success rate of the registration process and the fidelity of the output geometry, thus forming an automated processing link from the original mesh to a simulateable geometric model.

[0057] In this embodiment, step 101 specifically includes the following implementation methods: First, the raw 3D mesh models of the organs, obtained after data cleaning, are input into a pre-trained feature extraction network. This network can employ a graph convolutional or autoencoder-based architecture to map each 3D mesh model into a high-dimensional feature vector. This feature vector encodes the overall shape information of the organ, such as global curvature distribution, volume proportions, and the orientation of major anatomical axes, thus characterizing the global topological consistency of the raw organ 3D mesh models. Through this mapping, organs with similar morphologies are grouped closer together in the feature space, laying the foundation for subsequent clustering and selection.

[0058] Secondly, all extracted high-dimensional feature vectors are subjected to iterative clustering. This embodiment employs a multi-strategy clustering algorithm (e.g., an iterative search method combining K-Means and hierarchical clustering). In each iteration, the system dynamically adjusts the cluster boundaries based on the distribution of feature vectors and calculates the geometric centroid of all samples in each cluster, i.e., the average position in the feature space. Then, the original organ 3D mesh model closest to this geometric centroid is used as the representative sample of that group and stored in the seed bank. In this way, each sample in the seed bank is located at the center of its group, which can better represent the typical anatomical morphology of that organ and avoids selecting marginal or extreme samples as templates.

[0059] Finally, under preset topological constraints, multi-path edge collapse random iteration is performed on each representative sample in the seed library. The preset topological constraints include maintaining the manifold structure of the mesh and prohibiting the generation of non-manifold edges or degenerate faces. Specifically, the system randomly permutes the execution order of edge collapse and the step size weight of each collapse multiple times, generating multiple different simplified paths. After each path is completed, the quality of the resulting simplified mesh in terms of topological distribution is compared, such as whether the positions of key feature points are preserved and whether excessive distortion or elongated triangles are generated. The globally optimal topological distribution scheme is selected, and a embryo with a fixed vertex size is generated based on this scheme. Embedded bodies that meet the quality requirements are stored in the initial embryo library, thus completing the transformation from the original mesh to a standardized, high-quality embryo.

[0060] The above three sub-steps are executed sequentially, together realizing the process of extracting representative samples from the original three-dimensional mesh model of organs and performing topology optimization to construct the initial embryonic body library.

[0061] Furthermore, in the process of performing multi-path edge collapse random iteration processing on representative samples in the seed bank, this embodiment specifies the specific implementation method of each random iteration path and how to select the optimal result based on multi-dimensional quality evaluation indicators.

[0062] Specifically, for each representative sample in the seed bank, the system initiates multiple parallel random iterative paths under preset topological constraints. Within each path, the system randomly permutes the execution order of edge collapse and the step size weight for each collapse. The order of edge collapse determines which edges are merged first, while the step size weight affects the magnitude of change in the local mesh region after each collapse. Through this random permutation, different paths generate distinct simplification sequences, thus exploring diverse topological simplification possibilities. After each path is completed, a corresponding simplified mesh is generated.

[0063] Next, the system performs a multi-dimensional quality evaluation on each simplified mesh. The evaluation metrics include three aspects: first, Jacobian quality, which measures the degree of distortion of mesh cells from their ideal shape; a value closer to 1 indicates better cell quality; second, mesh aspect ratio, the ratio of the longest side to the shortest side of a cell; a value closer to 1 indicates a more regular cell shape; and third, geometric fidelity to key anatomical landmarks, the positional deviation between predefined key points in the simplified mesh (such as organ extrema, corners, or bifurcation points) and their corresponding points in the original mesh. The system calculates these three metrics to obtain a comprehensive quality score for each simplified mesh.

[0064] Finally, the system uses the evaluation results of the aforementioned multi-dimensional quality assessment indicators as reward feedback to guide the optimal selection of search paths. Specifically, the higher the simplified mesh quality score of each path, the greater the reward value obtained by that path. The system sorts all paths according to the reward value and selects the topology distribution scheme corresponding to the path with the highest reward. Since this scheme performs optimally in all three dimensions—Jacobi quality, mesh aspect ratio, and geometric fidelity—it is considered the globally optimal topology distribution scheme. Based on this scheme, the system generates embryos with a fixed vertex size, ensuring that the embryos entering the database meet the lightweight requirements while possessing excellent element quality and high anatomical fidelity.

[0065] (a) Random iterative processing of multi-path edge collapse.

[0066] Specifically, for a given representative sample grid M0 (with an initial number of vertices N0), the system sets the target vertex size N. target (e.g. N) target =1000). Under preset topological constraints (e.g., maintaining the manifold structure, prohibiting the generation of non-manifold edges or degenerate patches, and maintaining boundary integrity), the system initiates multiple parallel random iterative paths. Let the path indices be r = 1, 2, ..., R. In each path r, the system randomly generates an edge collapse sequence E. r ={e1,e2,…,e N0-Ntarget}, and the step-weight sequence W corresponding to each collapse. r={w1,w2,…}. The edge collapse operation is defined as collapsing an edge e=(v... i ,v j The two endpoints of ) are merged into a new vertex v. new Simultaneously, the edge and its two adjacent triangles are deleted, and the adjacency relationships are updated. The step size weight w affects how the position of the new vertex after merging is calculated (e.g., the weighted average weight). By randomly permuting the collapse order and step size weight, different simplification strategies were explored for different paths. After each path is executed, a simplified mesh M is generated. r Its number of vertices is N target .

[0067] (ii) Calculation of multi-dimensional quality evaluation indicators.

[0068] For each simplified grid M r The system calculates quality evaluation indicators in the following three dimensions.

[0069] (1) Jacobian mass: For each triangular cell t in the mesh, let its three vertices be a, b, c ∈ R. 3 Construct the Jacobian matrix Jt=[ba,ca]∈R 3×2 The Jacobian mass Q of this unit Jac (t) is defined as: ; This value ranges between 0 and 1; the closer it is to 1, the closer the triangle is to equilateral, and the better the element quality. For degenerate triangles (area 0), Q... Jac =0. The Jacobian mass of the entire mesh is the average of all cells: ; Among them, T r For grid M r A set of triangular units.

[0070] (2) Mesh aspect ratio: For each triangular element t, calculate the length l of its longest side. max (t) and the length of the shortest side l min The ratio of (t): ; An aspect ratio AR(t) ≥ 1, with a value closer to 1 indicating a more regular shape. The aspect ratio quality Q of the entire mesh. AR (M r Defined as the average of the reciprocals of the aspect ratios of all units: .

[0071] This value also falls between 0 and 1, with the closer it is to 1 indicating a more ideal overall grid aspect ratio.

[0072] (3) Geometric fidelity of key anatomical landmarks: Suppose a set of key anatomical landmarks P={p1,p2,…,p...} are predefined on the original mesh M0. m (e.g., extreme points, corners, or bifurcation points of organs). For simplified mesh M r By using nearest-point projection or parametric mapping, each p is found i In M r The corresponding point p on i (r) Geometric fidelity is defined as the normalized average distance error: ; Where D max Q is the diagonal length of the bounding box of the original mesh, used for normalization. geo The closer the value is to 1, the stronger the ability of the simplified mesh to preserve anatomical landmarks.

[0073] (III) Rewards and Feedback and Path Optimization.

[0074] The three quality evaluation indicators mentioned above are weighted and fused to obtain the comprehensive reward value R for each path r. r : ; Where α, β, and γ are preset weighting coefficients, satisfying α + β + γ = 1. In this embodiment, α = 0.4, β = 0.3, and γ = 0.3 can be set to reflect the emphasis on Jacobian quality. This comprehensive reward value serves as a feedback signal to evaluate the merits of each random iteration path.

[0075] The system assigns reward values ​​{R1, R2, ..., R} to all paths. R} Compare and select the path r with the highest reward value. =argmax r R r The edge collapse sequence and step size weights corresponding to this path constitute the globally optimal topology distribution scheme. Finally, the system generates a fixed vertex size N based on this scheme. target embryonic mesh And store it in the initial embryo bank.

[0076] Through the above-mentioned multi-path random iteration and reward feedback optimization mechanism, this embodiment can effectively overcome the defect of traditional greedy simplification algorithms that are prone to getting trapped in local optima, and generate embryo models of higher quality that are more suitable for subsequent registration and simulation.

[0077] In this embodiment, step 102, which involves extracting features from the geometric model of the three-dimensional organ target to be registered, calculating similarity, and matching the selected target embryo, is implemented using a two-stage screening mechanism of global recall and local rearrangement.

[0078] First, the system extracts high-dimensional feature vectors from the geometric model of the 3D organ target to be registered. These feature vectors reside in the same high-dimensional space as the features output by the Embedding model used to construct the embryo database. Through cosine similarity calculation, multiple high-dimensional feature vectors most similar to the target feature vector are quickly retrieved within the feature space. Then, a candidate set is recalled based on the embryo database index corresponding to these feature vectors to obtain a group of candidate embryos (e.g., Top-K matched embryos).

[0079] Secondly, the system enters the secondary screening stage, namely local rearrangement. For each candidate embryo, the local curvature difference between it and the target 3D organ geometric model to be registered at key anatomical locations (such as the organ's apex, edges, and bifurcation points), as well as the geometric distance between their mesh surfaces, are obtained. Based on these two indicators, the overall similarity score is recalculated. The higher the overall similarity score, the closer the candidate embryo is to the target geometric model in local morphology.

[0080] Finally, the system selects the candidate embryo with the highest overall similarity score as the target embryo. Through this two-level screening mechanism of "global recall + local rearrangement", it ensures both retrieval efficiency (fast global filtering) and matching accuracy of local anatomical features (fine local rearrangement), thus providing the most suitable initial template for subsequent deformation registration.

[0081] In this embodiment, the process of fusing the features of the three-dimensional organ target geometric model with the features of the target embryo in step 103, and obtaining multiple candidate registration strategy paths based on the fused features, is specifically implemented through a deep learning decision network model.

[0082] First, the high-dimensional features of the three-dimensional organ target geometric model and the high-dimensional features of the target embryo are concatenated with a preset classification label feature (cls_token) to form a combined feature sequence in the form of "cls_token+template_feature_tokens+input_feature_tokens".

[0083] Secondly, feature interaction processing is performed on the combined feature sequence. Specifically, a sparse multi-head attention module is adopted, where cls_token needs to perform attention calculation with all tokens (to aggregate global information for subsequent classification), and the input features and template features perform mutual attention calculation, but do not perform their own internal attention calculation, and a random deletion operation is introduced to ensure the robustness of attention calculation.

[0084] Then, the classification label feature (cls_token) containing feature fusion information is separated and extracted from the combined sequence after feature interaction processing. This feature is input into the network prediction branch of the pre-defined decision network model. The decision network model contains two parallel output branches: one branch outputs the score estimate of each registration strategy (regression task), and the other branch calculates the probability of each registration strategy being selected as the optimal one using softmax (classification task). Based on the score estimate or probability value, the system selects the top-ranked candidate registration strategy paths (e.g., Top-3).

[0085] In this way, the model can intelligently recommend multiple reliable registration paths for different organ geometries, providing diverse options for subsequent parallel deformation registration.

[0086] Figure 2 This is a schematic diagram of the overall structure of the decision network model provided in an embodiment of the present invention. Figure 2 As shown, the left side of the model has two parallel input branches, receiving geometric sampling point clouds of the organ to be registered and geometric sampling point clouds of the embryo, respectively. Each branch first performs geometric feature embedding through a set of convolutional neural networks (CNNs) with shared weights. This CNN structure includes multiple one-dimensional convolutional layers, batch normalization layers, ReLU activation layers, and max pooling layers, outputting high-dimensional geometric feature vectors. Subsequently, the two feature vectors are concatenated with a learnable classification label (cls_token) to form a combined feature sequence. This sequence is input into a sparse multi-head attention module for feature fusion and interaction. Specifically, the cls_token interacts with all features, and the input features and template features interact with each other but not with themselves. After the interaction, the feature sequence is separated, and the fused cls_token is extracted. The cls_token is fed into three parallel output branches: the first branch outputs the score estimate (value) of each registration strategy from the fully connected layer (MLP); the second branch outputs the probability (action) of each strategy being selected as the best from the fully connected layer and Softmax; and the third branch outputs the deformation displacement (Output Deformation) of the fused feature sequence (or part of the features) through a multi-layer Transformer decoder and a linear layer. This displacement is applied to the embryo grid to obtain the predicted registered geometry. Figure 2 The entire process from input to three outputs is fully demonstrated, with the size and number of layers of each module being for illustrative purposes only.

[0087] In this embodiment, the decision network model needs to be trained before performing feature fusion and prediction in step 103. This embodiment uses a multi-task hybrid paradigm for training, specifically including a self-supervised reconstruction task, a supervised learning classification task, and a reinforcement learning policy selection task. The reinforcement learning policy selection task is implemented as follows.

[0088] First, the feature data of the geometric model to be registered and the corresponding embryo from the sample library are used as the initial state and input into the decision network model. Specifically, each training sample consists of a set of paired data: the feature vector of the geometric model of the three-dimensional organ target to be registered, and the feature vector of the target embryo selected from the embryo library. These two parts of features together constitute the state S in the reinforcement learning framework.

[0089] Secondly, different registration strategies are presented as optional actions. The decision network model predicts the probability of choosing each optional action based on the current input state. These actions correspond to various possible registration strategy paths (e.g., rigid alignment based on PCA, feature matching based on RANSAC, similarity transformation with scaling, etc.). The action probability distribution output by the model reflects the model's preference for each strategy in the current state.

[0090] Then, the geometric deformation evaluation score after completing the registration by executing each optional action is obtained, and this score is used as the actual action feedback value for the corresponding optional action. Specifically, the system actually executes a certain registration strategy (action) to deform the target preform to the geometric model to be registered, obtaining the registered geometric result. Through multi-dimensional quality evaluation (including registration accuracy and mesh quality), a comprehensive score is calculated as the reward for this action. This reward value reflects the true effect of the strategy on the current sample.

[0091] Finally, based on the actual action feedback values ​​and the model's predicted selection probabilities, the decision network model parameters are dynamically updated with the goal of maximizing the actual action feedback values. This embodiment employs the PPO (Proximal Policy Optimization) reinforcement learning algorithm, using the aforementioned reward signals and policy probabilities to iteratively update the model parameters, making the model tend to choose actions that yield higher rewards. Simultaneously, to ensure the stability of the training process, KL loss is added to control the similarity between the model's output action probabilities and the actual action probability distribution, and entropy loss is used to control the model's uncertainty, thus balancing exploration and exploitation in reinforcement learning training.

[0092] Through training on the reinforcement learning strategy selection task described above, the decision network model can further explore strategies that, while not uniquely optimal, can still produce good registration results, building upon supervised classification. This enhances the model's generalization ability and robustness in practical applications. The reinforcement learning training, along with the self-supervised reconstruction task and the supervised classification task, constitutes a multi-task hybrid training paradigm, collaboratively optimizing model parameters.

[0093] In this embodiment, the combined feature sequence is processed for feature interaction and classification label features are extracted in the above embodiments. Specifically, a sparse multi-head attention mechanism is used, which includes: First, the classification label feature (cls_token) is interactively computed with the high-dimensional features of the 3D organ target geometric model and the target embryonic body, respectively. Specifically, in the sparse multi-head attention module, cls_token needs to perform attention computation with all tokens in the combined feature sequence (including target geometric features, embryonic body features, and other labels). Through this global interaction, cls_token can aggregate global morphological information from the two input sources, laying the foundation for subsequent classification and prediction tasks.

[0094] Secondly, cross-interaction computation is performed on the high-dimensional features of the 3D organ target geometric model and the high-dimensional features of the target embryo. Specifically, input features (from the target geometric model) and template features (from the target embryo) perform mutual attention computation, meaning each input feature needs to pay attention to all template features, and vice versa. However, they do not perform their own internal attention computation (i.e., input features do not perform attention computation with each other, and template features do not perform attention computation with each other). Simultaneously, to improve the model's robustness, a random dropout mechanism is introduced during the attention computation process, randomly ignoring some attention connections. This sparsity-based interaction significantly reduces parameters irrelevant to the task, enhancing the model's generalization ability.

[0095] Finally, after the aforementioned feature interaction calculations are completed, the updated classification label feature (cls_token) is separated and extracted from the combined feature sequence. At this point, the cls_token contains feature fusion information from the target geometric model and the target embryo, simultaneously encoding the correspondence between the two in terms of global shape and local structure. This fused classification label feature will be fed into the subsequent prediction branch to output the score estimate and probability distribution of each registration strategy.

[0096] Through the above-described feature interaction processing, this embodiment can efficiently fuse the morphological information between the target geometry and the embryo body, providing sufficient and accurate feature representation for intelligent selection of the optimal registration strategy.

[0097] In this embodiment, step 104, which involves parallel deformation registration of the target embryo to the three-dimensional organ target geometric model according to multiple candidate registration strategy paths, specifically follows a hierarchical optimization strategy of "global first, local second; coarse first, fine second." Each candidate path performs deformation registration in the order of global rigid initial registration to local non-rigid deformation. The following detailed explanation uses one path as an example; the other paths execute the same process in parallel.

[0098] (a) Formalizing the problem.

[0099] Let the source mesh (i.e. the target embryo) be M. S The target mesh (i.e., the three-dimensional geometric model of the organ target) is M. T The goal of registration is to find an optimal transformation operator. :M S →M T This minimizes the composite loss function L: ; Where T is the allowed transformation space.

[0100] (ii) Global rigid initial registration.

[0101] First, the source and target meshes are centered and normalized to eliminate translation and scale differences. Then, based on the registration strategy specified by the current candidate path, the global spatial transformation matrix between the target embryo and the 3D organ target geometric model is calculated. This embodiment supports multiple rigid registration methods, including but not limited to the following four: 1. PCA Rigid Alignment: Solving for the Optimal Rotation Matrix This minimizes the following expression: ; Where V S V T These are the vertex coordinate matrices of the source mesh and the target mesh, respectively. It is the Frobenius norm. The analytical solution to this problem is obtained by considering the covariance matrix H=(V S -mean(V S )) T (V T -mean(V T Singular value decomposition yields: ; U and V are the left and right singular matrices obtained by performing singular value decomposition on the covariance matrix H, respectively, and Σ is the corresponding singular value diagonal matrix.

[0102] 2. Rotation Monte Carlo Search: Aligning PCA results M PCA Based on the previous step, search for the optimal rotation. : ; Where T rot It is a predefined set of discrete rotations.

[0103] 3. Feature-based RANSAC registration: Solving for rigid transformations using the random sampling consensus algorithm. ; in, p represents the optimal rigid transformation result obtained by the feature-based RANSAC algorithm. i ,q i C represents the corresponding feature points extracted from the source and target meshes. match Let I(·) be the candidate matching point set, and let I(·) be the indicator function. t is the distance threshold, and t is the translation vector.

[0104] 4. RANSAC registration with scaling transformation: Introducing a scaling factor s∈R + : in, The optimal similarity transformation result obtained by the RANSAC algorithm with scaling transformation is given; s is the scaling transformation factor, and s belongs to the set of positive real numbers R. + The remaining parameters have the same meaning as those in the feature-based RANSAC registration described above. A global spatial transformation matrix (rotation R0, translation t0, optional scale s0) is obtained using any of the above methods, transforming the target embryo to the pose of the three-dimensional organ target geometric model, thus completing the initial global alignment.

[0105] (iii) Local fine-tuning of ICP.

[0106] Based on the global initial alignment, the Iterative Closest Point (ICP) algorithm is used for local refinement. ICP continuously updates the local alignment pose by iteratively minimizing the mean square error of the nearest point pair between the source and target meshes. In the k-th iteration, the solution is: ; Where C k The corresponding point set R is obtained through nearest neighbor search. k and t k Let be the rotation matrix and translation vector to be optimized during the k-th iteration, respectively. and These are the optimal rotation matrix and optimal translation vector obtained in the k-th iteration, respectively. The ICP algorithm has fast convergence characteristics and can obtain a stable local alignment pose within dozens of iterations, achieving a smooth transition from coarse registration to fine registration.

[0107] (iv) Non-rigid deformation.

[0108] Starting from the target preform mesh after updating the local alignment pose (denoted as...) Furthermore, non-rigid deformation displacement parameters are applied to conform to the fine surface structure of the target geometry. Let the coordinates of the deformed mesh vertices be M'. deform = +d, where d is the displacement vector to be solved. The optimization problem is: ; Among them, L deform (d) is the non-rigid deformation composite loss function to be minimized, L CD R(·) is the chamfer distance, used to measure the degree of surface matching between the deformed mesh and the target geometry; R(·) is a regularization term (such as a Laplacian smoothing term), used to maintain the smoothness of the deformation and prevent mesh self-intersection; w CD and w reg represents the weighting coefficients. In this embodiment, the gradient descent method is used to iteratively optimize the above loss function until convergence, thereby completing the non-rigid deformation optimization and obtaining the corresponding candidate registration geometric results.

[0109] After multiple candidate registration strategy paths execute the complete deformation registration process described above simultaneously, multiple candidate registration geometric results are generated. Then, the target three-dimensional organ geometric model that meets the requirements is selected by using a quality threshold.

[0110] Furthermore, Figure 3 A flowchart illustrating the overall process of geometric data processing for three-dimensional organs provided in an embodiment of the present invention is shown. Figure 3 As shown, the method includes the following four main steps.

[0111] Step 1: Automatic Template Construction.

[0112] First, 3D mesh models of the original organs are acquired, and feature extraction and topology optimization are performed to construct an initial embryonic body library with anatomical representativeness. Specifically, an encoder-decoder structure is used to perform implicit template learning on a large number of original organ meshes, and after explicit construction and processing, a template library containing multiple high-quality embryonic bodies is formed. Each embryonic body in this library has a uniform topology and excellent mesh quality, providing a standardized initial shape for subsequent registration.

[0113] Step 2: Multi-strategy Parallel Registration.

[0114] Then, for each patient to be registered (Patient-1, Patient-2, …, Patient-x) and its corresponding organ category (Organ-1, Organ-2, …, Organ-y), features of the 3D target organ geometric model are extracted. The similarity between these features and embryonic features in the initial embryonic body library is calculated, and the corresponding target embryonic body is selected based on the similarity. After selection, multiple registration strategies (Strategy-1, Strategy-2, …, Strategy-z) are executed in parallel to deform and register the target embryonic body to the 3D target organ geometric model. This step generates multiple candidate registration geometric results through hierarchical optimization of "global first, then local; coarse first, then fine".

[0115] Step 3: Intelligent Strategy Decision.

[0116] Building upon parallel registration, this invention further introduces a deep learning decision model to achieve intelligent selection of registration strategies. Specifically, the input mesh of the organ to be registered and a template mesh selected from a template library are input into a pre-trained "Mesh-to-StrategyModel". This model outputs confidence scores for each candidate strategy (e.g., Strategy-A: 0.45, Strategy-B: 0.95, Strategy-N: 0.12), thereby predicting one or more optimal candidate registration strategy paths. This step, by fusing the features of the 3D organ target geometric model with the features of the target embryo, and predicting multiple candidate registration strategy paths based on the fused features, significantly reduces computational overhead and improves decision-making efficiency.

[0117] Step 4: Inference Pipeline.

[0118] Finally, in the practical inference application phase, the system executes according to the following pipeline: receiving the input mesh, selecting the best-matching template mesh from the template library, inputting the selected template mesh and the input mesh into the policy model to obtain the recommended policy (e.g., Strategy-C), performing parallel registration according to this policy, and finally outputting the registered geometric result. This step performs deformation registration on the target embryo in parallel according to multiple candidate registration policy paths, generating multiple candidate registration geometric results, and selecting the target 3D organ geometric model that meets the requirements based on a quality threshold. Simultaneously, the system is equipped with anomaly detection and closed-loop feedback mechanisms to ensure the robustness and reliability of the simulation model output.

[0119] Through the organic integration of the above four steps, the embodiments of the present invention realize the fully automatic conversion from raw medical images to standardized, high-quality, simulable three-dimensional geometric models, filling the technical gap between segmentation masks and simulable geometric models in the prior art.

[0120] Figure 4 This paper compares the results of our proposed method with the traditional Marching Cubes algorithm in the 3D reconstruction of various organs. Experimental results show that, under strict control of the vertex size (1000 vertices for all vertices), our method can still reproduce the anatomical morphology of organs with high fidelity. This lightweight and high-quality geometric representation provides an ideal data foundation for subsequent biomechanical simulations, surgical path planning, and deep learning model training.

[0121] Table 1 compares the average success rates of five different geometric 3D registration schemes. These five schemes cover a variety of technical paths from basic to advanced: 1) single rigid registration based on PCA; 2) a cascaded scheme of PCA rigid registration and ICP non-rigid registration; 3) a refined non-rigid deformation scheme that introduces vertex-wise displacement based on the above; 4) the best-performing fixed single-strategy registration scheme; and 5) the multi-strategy parallel registration scheme proposed in this study. The data show that the method in this study performs best, with an average registration success rate of 86%, significantly improving the registration reliability of complex anatomical structures.

[0122] Furthermore, to further verify the contribution of each core component in the three-stage deep learning framework to the selection of registration strategies, this study conducted systematic ablation experiments, the results of which are detailed in Table 2. Experimental analysis shows that the Feature Fusion Module significantly enhances the model's ability to identify complex anatomical structures by integrating multi-scale morphological features; the introduction of the Self-supervised Reconstruction task effectively utilizes the latent distribution of unlabeled data and strengthens the robustness of feature representation; and the Reinforcement Learning task optimizes the policy search space through a dynamic reward mechanism.

[0123] Experimental data clearly demonstrate the synergistic effect of the aforementioned modules: compared to the baseline model, the Top-1 and Top-3 accuracies of the model showed a steady upward trend after adding each component. This strongly proves the effectiveness and necessity of the multi-task learning architecture proposed in this study in improving registration decision accuracy.

[0124] Table 1. Average registration success rate of 3D registration using different registration methods Table 2 Ablation Experiment Results of Deep Learning Model In summary, the geometric data processing method for three-dimensional organs presented in this invention achieves efficient production of simulable three-dimensional geometric models through the organic integration of offline pre-construction of a high-quality embryonic body library and online intelligent matching. The system first utilizes Embedding feature clustering and MCTS-QEM topology optimization to construct an initial embryonic body library with anatomical representativeness, ensuring extremely high mesh quality for the embryonic bodies within a fixed vertex scale. Subsequently, through a hierarchical optimization strategy of "global first, local second; coarse first, fine second," parallel multi-path registration including PCA, RANSAC, ICP, and non-rigid deformation is performed, increasing the average registration success rate from 52% in traditional cascade schemes to 86%, significantly enhancing the registration reliability of complex anatomical structures. Finally, a deep learning decision model and a multi-dimensional quality quantification evaluation mechanism are introduced to ensure the robustness of the output target three-dimensional organ geometric model, effectively filling the technological gap between raw medical images and high-quality simulable geometric models. This method not only provides an ideal data foundation for biomechanical simulation and numerical computation but also offers a reliable technical path for the fully automated standardized processing of large-scale medical images.

[0125] The geometric data processing apparatus for three-dimensional organs provided in the embodiments of the present invention will be described below. The geometric data processing apparatus for three-dimensional organs described below can be referred to in correspondence with the geometric data processing method for three-dimensional organs described above.

[0126] This invention provides a geometric data processing device for three-dimensional organs, see [link to related document]. Figure 5 ,include: Module 510 is used to acquire the original organ three-dimensional mesh model, perform feature extraction and topology optimization, and construct an initial embryonic body library with anatomical representativeness. The calculation module 520 is used to extract the features of the three-dimensional organ target geometric model to be registered, calculate the similarity between the features of the three-dimensional organ target geometric model and the features of embryos in the initial embryo database, and select the corresponding target embryo based on the similarity. The fusion module 530 is used to fuse the features of the three-dimensional organ target geometric model with the features of the target embryo, and to obtain multiple candidate registration strategy paths based on the fused features. Processing module 540 is used to perform deformation registration of the target embryo to the three-dimensional organ target geometric model in parallel according to multiple candidate registration strategy paths, generate multiple candidate registration geometric results, and select the target three-dimensional organ geometric model that meets the quality threshold based on the multiple candidate registration geometric results.

[0127] Figure 6 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6 As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute a geometric data processing method for three-dimensional organs.

[0128] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0129] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the geometric data processing methods for three-dimensional organs provided by the above methods.

[0130] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the geometric data processing methods for three-dimensional organs provided by the methods described above.

[0131] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0132] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A geometric data processing method for three-dimensional organs, characterized in that, include: Obtain the original organ 3D mesh model, perform feature extraction and topology optimization, and construct an initial embryonic body library with anatomical representativeness; Features of the three-dimensional organ target geometric model to be registered are extracted, the similarity between the features of the three-dimensional organ target geometric model and the features of embryos in the initial embryo database is calculated, and the corresponding target embryo is selected based on the similarity. The features of the three-dimensional organ target geometric model are fused with the features of the target embryo, and multiple candidate registration strategy paths are obtained based on the fused features. The target embryo is deformed and registered to the three-dimensional organ target geometric model in parallel according to multiple candidate registration strategy paths, generating multiple candidate registration geometric results. Based on the multiple candidate registration geometric results, the target three-dimensional organ geometric model that meets the quality threshold is selected. The process of obtaining a three-dimensional mesh model of the original organ, performing feature extraction and topology optimization, and constructing an initial embryonic body library with anatomical representativeness includes: The obtained original organ 3D mesh model is mapped to a high-dimensional feature vector to characterize the global topological consistency of the original organ 3D mesh model; The high-dimensional feature vectors are subjected to iterative clustering to calculate the geometric centroid of the corresponding feature group, and the original organ 3D mesh model that is closest to the geometric centroid is stored as a representative sample in the seed library. Under preset topological constraints, multi-path edge collapse random iteration processing is performed on representative samples in the seed library. Based on the optimal result of the topological distribution, embryos with a fixed vertex size are generated and stored in the initial embryo library. Under preset topological constraints, the process of performing multi-path edge collapse random iteration on representative samples in the seed bank, and generating a fixed-vertex embryo based on the optimal result of the topological distribution, includes: In each path of random iteration, a simplified grid is generated by randomly permuting the edge collapse sequence and the step size weight. The Jacobian quality, mesh aspect ratio, and geometric fidelity of key anatomical landmarks of the simplified mesh are calculated as multi-dimensional quality evaluation indicators. The evaluation results of the multi-dimensional quality evaluation indicators are used as reward feedback to optimize the search path and select the globally optimal topology distribution scheme to generate a embryo with a fixed number of vertices.

2. The method according to claim 1, characterized in that, The process of extracting features from the three-dimensional organ target geometric model to be registered, calculating the similarity between the features of the three-dimensional organ target geometric model and the features of embryos in the initial embryo database, and selecting the corresponding target embryo based on the similarity includes: High-dimensional feature vectors of the geometric model of the three-dimensional organ target to be registered are extracted, and embryos corresponding to multiple similar high-dimensional feature vectors are retrieved in the feature space as candidate embryos by cosine similarity calculation. The local curvature differences and geometric distances between the three-dimensional organ target geometric model and each candidate embryo at key anatomical sites are obtained respectively. A comprehensive similarity score is calculated based on the local curvature differences and the geometric distances, and the candidate embryo with the highest score is determined as the target embryo.

3. The method according to claim 1, characterized in that, The process of fusing the features of the three-dimensional organ target geometric model with the features of the target embryo, and predicting multiple candidate registration strategy paths based on the fused features, includes: The high-dimensional features of the three-dimensional organ target geometric model, the high-dimensional features of the target embryo, and the preset classification label features are concatenated to form a combined feature sequence; The combined feature sequence is subjected to feature interaction processing to extract the classification label features after feature interaction processing; The extracted classification label features are input into the network prediction branch of the preset decision network model to predict and obtain score estimates for various registration strategies, and the top candidate registration strategy paths are selected based on the score estimates.

4. The method according to claim 3, characterized in that, The step of performing feature interaction processing on the combined feature sequence to extract the classification label features after feature interaction processing includes: The classification marker features are interactively calculated with the high-dimensional features of the three-dimensional organ target geometric model and the high-dimensional features of the target embryo, respectively. Cross-interactive calculations are performed on the high-dimensional features of the three-dimensional organ target geometric model and the high-dimensional features of the target embryo. After feature interaction calculation, the classification label features containing feature fusion information are separated and extracted from the combined feature sequence.

5. The method according to claim 3, characterized in that, Before fusing the features of the three-dimensional organ target geometric model with the features of the target embryo, the process further includes: The decision network model is pre-trained using a multi-task hybrid paradigm that includes self-supervised reconstruction, supervised learning classification, and reinforcement learning policy selection tasks, wherein the reinforcement learning policy selection task includes: The feature data of the geometric model to be registered and the corresponding embryo in the sample library are used as the initial state input into the decision network model; Different registration strategies are treated as optional actions, and the decision network model predicts the selection probability for each optional action based on the initial state. Obtain the geometric deformation evaluation score after registration is completed by executing each of the optional actions, and use the geometric deformation evaluation score as the actual action feedback value for the corresponding optional action; The parameters of the decision network model are dynamically updated based on the actual action feedback value and the selection probability, with the goal of maximizing the actual action feedback value.

6. The method according to claim 1, characterized in that, The process involves performing parallel deformation registration of the target embryo to the three-dimensional organ target geometric model according to multiple candidate registration strategy paths, generating multiple candidate registration geometric results, including: Multiple candidate registration strategy paths are executed simultaneously, and each candidate registration strategy path is deformed and registered according to a hierarchical registration strategy from global rigid initial registration to local non-rigid deformation. Calculate and obtain the global spatial transformation matrix between the target embryo and the three-dimensional organ target geometric model, and complete the global initial alignment of the target embryo and the three-dimensional organ target geometric model based on the global spatial transformation matrix; Based on the global initial alignment, the nearest point pair between the target embryo and the mesh surface of the three-dimensional organ target geometric model is iteratively searched, and the spatial distance error of the nearest point pair is minimized to update the local alignment pose; Starting with the target preform after updating the local alignment posture, non-rigid deformation displacement parameters are calculated and applied to complete non-rigid deformation optimization and obtain the corresponding candidate registration geometry results.

7. A geometric data processing device for three-dimensional organs, characterized in that, include: The module is used to acquire the original three-dimensional mesh model of the organ, perform feature extraction and topology optimization, and build an initial embryonic body library with anatomical representativeness; The calculation module is used to extract the features of the three-dimensional organ target geometric model to be registered, calculate the similarity between the features of the three-dimensional organ target geometric model and the features of embryos in the initial embryo database, and select the corresponding target embryo based on the similarity. The fusion module is used to fuse the features of the three-dimensional organ target geometric model with the features of the target embryo, and to obtain multiple candidate registration strategy paths based on the fused features. The processing module is used to perform deformation registration of the target embryo to the three-dimensional organ target geometric model in parallel according to multiple candidate registration strategy paths, generate multiple candidate registration geometric results, and select the target three-dimensional organ geometric model that meets the quality threshold based on the multiple candidate registration geometric results; The process of obtaining a three-dimensional mesh model of the original organ, performing feature extraction and topology optimization, and constructing an initial embryonic body library with anatomical representativeness includes: The obtained original organ 3D mesh model is mapped to a high-dimensional feature vector to characterize the global topological consistency of the original organ 3D mesh model; The high-dimensional feature vectors are subjected to iterative clustering to calculate the geometric centroid of the corresponding feature group, and the original organ 3D mesh model that is closest to the geometric centroid is stored as a representative sample in the seed library. Under preset topological constraints, multi-path edge collapse random iteration processing is performed on representative samples in the seed library. Based on the optimal result of the topological distribution, embryos with a fixed vertex size are generated and stored in the initial embryo library. Under preset topological constraints, the process of performing multi-path edge collapse random iteration on representative samples in the seed bank, and generating a fixed-vertex embryo based on the optimal result of the topological distribution, includes: In each path of random iteration, a simplified grid is generated by randomly permuting the edge collapse sequence and the step size weight. The Jacobian quality, mesh aspect ratio, and geometric fidelity of key anatomical landmarks of the simplified mesh are calculated as multi-dimensional quality evaluation indicators. The evaluation results of the multi-dimensional quality evaluation indicators are used as reward feedback to optimize the search path and select the globally optimal topology distribution scheme to generate a embryo with a fixed number of vertices.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the geometric data processing method for three-dimensional organs as described in any one of claims 1 to 6.

Images