A method and system for reconstruction of bone defects

Abstract

The application provides a bone defect reconstruction method and system, and relates to the technical field of medical image processing. The method comprises the following steps: obtaining three-dimensional data of a to-be-repaired affected bone and a pre-constructed statistical shape model; extracting an anatomical landmark point based on the three-dimensional data of the to-be-repaired affected bone to construct a corresponding local coordinate system, and performing voxel sampling under the corresponding local coordinate system to obtain an observed truncated signed distance field of the to-be-repaired affected bone; projecting the observed truncated signed distance field into the statistical shape model to detect a defect, and determining reliable observed voxels in the observed truncated signed distance field; fitting the statistical shape model based on the reliable observed voxels to obtain shape coefficients, and reconstructing the complete morphology of the to-be-repaired affected bone according to the shape coefficients, so that non-anatomical distortion of the reconstruction result is prevented, and the accuracy and anatomical rationality of the three-dimensional reconstruction of the bone morphology under a complex defect scenario are ensured.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method and system for bone defect reconstruction. Background Technology

[0002] In orthopedic clinical settings such as joint replacement, trauma repair, or tumor resection, local bone defects often occur in the glenoid fossa and surrounding bone structures (such as the glenoid cavity of the scapula and the acetabulum) due to degeneration or trauma. These glenoid fossae have complex curvatures and bear significant joint contact stresses; the loss of their original anatomical morphology can severely affect the initial stability of subsequent prosthesis implantation. Therefore, to meet the urgent needs of preoperative planning, osteotomy guide design, and precise matching of personalized implants, obtaining a complete three-dimensional bone model that closely approximates the patient's true anatomical morphology for high-quality, individualized three-dimensional reconstruction of the local defect area has significant clinical application value.

[0003] Currently, 3D reconstruction of the glenoid fossa and surrounding bone defect areas mainly relies on methods such as manual experience-based repair, contralateral image mirroring replacement, and statistical shape models (SSM) based on surface meshes. Among these, the statistical model method based on surface meshes is widely used. Its core implementation process involves creating a triangular mesh from bone samples of healthy individuals, establishing a dense correspondence at the vertex level among all samples using registration algorithms such as Non-rigid Iterative Closest Point (NICP), and then performing Principal Component Analysis (PCA) on the corresponding mesh to construct a statistical shape model. When reconstructing the affected side, the mean mesh of this model is non-rigidly registered to the surface of the residual bone on the affected side, and the shape coefficients are solved to predict and complete the complete bone morphology mesh.

[0004] However, the aforementioned reconstruction method based on surface mesh correspondence forces the establishment of a one-to-one vertex topology among all sample meshes. This process is not only highly sensitive to initial alignment and computationally expensive, but also prone to serious topological correspondence errors such as mesh flipping, folding, or misalignment in glenoid regions with drastic curvature changes or significant individual morphological differences. Once the basic correspondence is deviated, the quality of the statistical model and the subsequent fitting accuracy will decrease significantly; especially when registering bone on the affected side containing defects, unreliable geometric observations at the defect boundary often pull the shape base to the wrong position, ultimately causing the reconstruction result to deviate from the true anatomical morphology, making it difficult to achieve robust bone defect repair. Summary of the Invention

[0005] This invention provides a method and system for bone defect reconstruction, which solves the problems in the prior art where the reconstruction of glenoid bone defects is highly dependent on the correspondence of dense mesh vertices, which is prone to topological errors, and where directly fitting the defect area can cause the shape base to be pulled to the wrong position, resulting in non-anatomical morphology. The invention achieves highly robust and accurate individualized complete bone morphology restoration that conforms to the real anatomical structure.

[0006] This invention provides a method for bone defect reconstruction, comprising: The three-dimensional data of the affected bone to be repaired and a pre-constructed statistical shape model are obtained. The statistical shape model is constructed based on the truncated symbolic distance field of a healthy bone sample in a local coordinate system. Anatomical landmarks are extracted based on the three-dimensional data of the bone to be repaired to construct a corresponding local coordinate system. Voxel sampling is performed in the corresponding local coordinate system to obtain the observation truncated symbol distance field of the bone to be repaired. The observed truncated symbol range field is projected onto the statistical shape model for defect detection to determine the reliable observation voxels in the observed truncated symbol range field; The statistical shape model is fitted with the reliable observation voxels to obtain shape coefficients, and the complete morphology of the affected bone to be repaired is reconstructed based on the shape coefficients.

[0007] According to the bone defect reconstruction method provided by the present invention, the step of extracting anatomical landmarks based on the three-dimensional data of the affected bone to be repaired to construct a corresponding local coordinate system specifically includes: Multiple anatomical landmarks are identified and extracted from the three-dimensional data of the affected bone to be repaired. These anatomical landmarks include a target landmark and multiple directional landmarks. The target landmark is used as the origin of the coordinate system. Two sets of directional seed vectors are determined based on the multiple directional landmarks, and three orthogonal coordinate axes are obtained by vector cross product and normalization. Based on the origin of the coordinate system and the three orthogonal coordinate axes, the rigid body transformation relationship from the world coordinate system to the corresponding local coordinate system is determined to construct the corresponding local coordinate system.

[0008] According to the bone defect reconstruction method provided by the present invention, the construction step of the statistical shape model includes: transforming the three-dimensional data of multiple healthy bone samples to a set local coordinate system, and generating a regular voxel mesh in a determined region of interest; The truncated ... The truncation symbol distance field of each healthy bone sample is dimensionality reduced, and the principal component basis and mean vector are extracted. Based on the truncated symbol distance field of each healthy bone sample, the occupancy frequency of each voxel is statistically analyzed to divide the fitting domain, and the principal component basis, the mean vector, and the fitting domain are used together as the constructed statistical shape model.

[0009] According to the bone defect reconstruction method provided by the present invention, the step of performing dimensionality reduction processing on the truncated symbolic distance field of each healthy bone sample and extracting the principal component basis and mean vector includes: The truncated symbolic distance field of the healthy bone sample is converted into a one-dimensional vector and stacked to generate a data matrix; Calculate the mean of the data in the data matrix to obtain the mean vector; The data matrix is ​​centered using the mean vector. Principal component analysis is performed on the centered data matrix to obtain multiple eigenvalues ​​and corresponding eigenvectors. The eigenvectors are filtered based on the magnitude of the eigenvalues, and the retained eigenvectors are used as the principal component basis.

[0010] According to the bone defect reconstruction method provided by the present invention, the step of statistically analyzing the occupancy frequency of each voxel to divide the fitting domain based on the truncated symbol distance field of each healthy bone sample includes: For each voxel in the voxel grid of the rule, the proportion of samples occupied by bones in the truncated symbol distance field of multiple healthy bone samples is statistically analyzed to obtain the occupancy frequency field. Voxels in the occupied frequency field whose occupied frequency is not lower than a set threshold are marked as support domains; The voxels whose occupancy frequencies fall within a set interval in the occupancy frequency field are labeled as the fitting domain.

[0011] According to the bone defect reconstruction method provided by the present invention, the step of projecting the observed truncated symbolic distance field onto the statistical shape model for defect detection and determining the reliable observation voxels in the observed truncated symbolic distance field includes: The observed truncated symbolic range field is projected onto the statistical shape model to obtain the initial reconstructed reference range field; Calculate the difference between the observed truncated symbol range field and the initial reconstructed reference range field; The larger value between the set quantile of the difference and the preset fixed threshold is taken as the effective threshold. Voxels whose difference is less than or equal to the effective threshold are marked as the reliable observation voxels.

[0012] According to the bone defect reconstruction method provided by the present invention, the step of fitting the statistical shape model based on the reliable observation voxels to obtain shape coefficients specifically includes: Extract the subset of model feature basis contained in the statistical shape model on the reliable observation voxel, the subset of mean data on the reliable observation voxel, and the subset of observation truncated symbolic distance field on the reliable observation voxel; By combining the preset regularization term, the initial shape coefficient is obtained by performing least squares fitting calculation using each subset corresponding to the reliable observation voxel. The initial shape coefficient is truncated and trimmed according to the preset truncation parameters to obtain the shape coefficient.

[0013] According to the bone defect reconstruction method provided by the present invention, the step of combining a preset regularization term and performing least squares fitting calculation using the subsets corresponding to the reliable observation voxels to obtain the initial shape coefficients specifically includes: Soft weights are assigned to the reliable observation voxels based on the distance from the reliable observation voxel to the defect boundary, wherein the soft weights decrease as the distance from the corresponding voxel to the defect boundary decreases, and the values ​​of the soft weights are limited to a preset range. Using the assigned soft weights and the preset regularization term, a weighted least squares fitting calculation is performed on each subset corresponding to the reliable observation voxel to obtain the initial shape coefficient of the current round. The reference shape is reconstructed using the initial shape coefficients obtained in the current round, and defect detection is performed again to update the trusted observation voxels and the soft weights. The process then returns to the step of performing the weighted least squares fitting calculation until a preset number of iterations is reached or the trusted observation voxels remain stable.

[0014] According to the bone defect reconstruction method provided by the present invention, the step of reconstructing the complete morphology of the affected bone to be repaired based on the shape coefficient includes: Based on the shape coefficient and the statistical shape model, a complete voxel-level symbolic distance field is reconstructed; voxels with values ​​less than or equal to zero in the complete voxel-level symbolic distance field are extracted as reconstructed occupied voxels, and morphological cleaning is performed on the reconstructed occupied voxels to obtain local occupied data; the local occupied data is inversely transformed to the world coordinate system through the corresponding local coordinate system, and a three-dimensional morphological mask is output as the complete shape of the affected bone to be repaired.

[0015] The present invention also provides a bone defect reconstruction system, comprising: The data acquisition module is used to acquire three-dimensional data of the bone to be repaired on the affected side and a pre-constructed statistical shape model, wherein the statistical shape model is constructed based on the truncated symbolic distance field of a healthy bone sample in a local coordinate system. The sampling processing module is used to extract anatomical landmarks based on the three-dimensional data of the bone to be repaired, to construct a corresponding local coordinate system, and to perform voxel sampling in the corresponding local coordinate system to obtain the observation truncated symbol distance field of the bone to be repaired. The defect detection module is used to project the observation truncated symbol range field onto the statistical shape model to perform defect detection and determine the reliable observation voxels in the observation truncated symbol range field; The morphological reconstruction module is used to fit the statistical shape model based on the reliable observation voxels to obtain shape coefficients, and to reconstruct the complete morphology of the affected bone to be repaired based on the shape coefficients.

[0016] 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 bone defect reconstruction method as described above.

[0017] 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 bone defect reconstruction method as described above.

[0018] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the bone defect reconstruction method as described above.

[0019] This invention provides a method and system for bone defect reconstruction. It acquires a statistical shape model pre-constructed based on a truncated symbolic distance field in a local coordinate system using healthy bone samples. Anatomical landmarks of the affected bone to be repaired are extracted to construct a corresponding local coordinate system for sampling, resulting in an observed truncated symbolic distance field. The implicit distance field and unified local coordinate system achieve voxel-by-voxel alignment between samples, avoiding dense correspondence calculation errors caused by non-rigid registration. Simultaneously, the observed truncated symbolic distance field is projected onto the statistical shape model for defect detection to determine reliable observation voxels. The statistical shape model is then fitted only based on these reliable observation voxels to obtain shape coefficients. The complete morphology of the affected bone to be repaired is reconstructed based on these shape coefficients. This mechanism eliminates interference from invalid observation data in the defect area on the model fitting process, ensuring that the missing bone segment is generated primarily by the prior of the healthy statistical shape. This prevents non-anatomical distortions in the reconstruction results and guarantees the accuracy and anatomical rationality of the three-dimensional bone morphology reconstruction in complex defect scenarios. Attached Figure Description

[0020] 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.

[0021] Figure 1 This is a schematic flowchart of the bone defect reconstruction method provided by the present invention.

[0022] Figure 2 This is a schematic diagram of the input data and statistical shape modeling framework provided in this embodiment of the invention. ① represents the healthy target bone segmentation mask for constructing the statistical shape model, serving as a training sample for the statistical shape model construction; ② represents anatomical landmarks used to establish a unified local coordinate system for the target bone; ③ represents the segmentation mask of the affected side bone to be repaired, serving as input observation data for the inference stage; ④ represents a schematic diagram of local truncated symbolic distance field statistical shape modeling, demonstrating the process of sampling the truncated symbolic distance field of the training samples and constructing the statistical shape model in the local coordinate system.

[0023] Figure 3 This is a schematic diagram of truncated symbolic distance field sampling provided in an embodiment of the present invention. ⑤ represents the healthy target bone segmentation mask serving as the sampling source; ⑥ represents the local truncated symbolic distance field calculation result obtained after sampling the segmentation mask onto a regular voxel grid in the local coordinate system, showing the distribution of signed distance values ​​for each voxel near the bone surface. Figure 3 The blue area represents the inside of the bone (negative distance value), the red area represents the outside of the bone (positive distance value), and the black line represents the zero level set, i.e., the position of the bone surface.

[0024] Figure 4 This is a schematic diagram of the defect reconstruction result provided by an embodiment of the present invention. ⑦ represents the target bone segmentation mask on the affected side to be repaired, where there is a significant bone defect in the glenoid fossa region; ⑧ represents the effect after defect reconstruction using this method, demonstrating the complete bone morphology after the defect area has been properly filled.

[0025] Figure 5 This is a schematic diagram of the bone defect reconstruction system provided by the present invention.

[0026] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0027] 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.

[0028] In orthopedic clinical preoperative planning, joint replacement prosthesis selection, and personalized implant design, bony sockets and surrounding bone structures often suffer from varying degrees of local bone defects due to degenerative changes, traumatic fractures, tumor resection, or revision surgery. This makes it impossible to directly and completely obtain the original anatomical morphology of the articular surface from the imaging data of the affected side. How to reasonably restore the bone morphology of the missing area based on partial observation information of the residual bone on the affected side, so as to make it as close as possible to the patient's true anatomical structure, is a key technical problem that urgently needs to be solved in digital orthopedic diagnosis and treatment.

[0029] Currently, the main technical solutions for three-dimensional reconstruction of the glenoid fossa and surrounding bone defect areas are as follows: One method is manual experience-based reconstruction. This involves doctors or engineers manually filling in the missing areas on CT images or 3D bone models based on their clinical experience and anatomical knowledge, or adjusting the articular surface contours, glenoid fossa center, and edge positions using interactive software to ultimately obtain a complete bone structure model. This method is relatively straightforward in clinical practice and requires no additional algorithmic support or training data.

[0030] The second method is the contralateral mirror reconstruction method. This method segments the bone on the unaffected side of the same patient, then mirrors and flips it. The mirrored bone morphology is used as a substitute for the defect area on the affected side. Rigid or non-rigid registration is then used to place it in the coordinate system of the affected side, thereby obtaining the complete bone morphology of the affected side. This method typically requires prior acquisition of CT, CBCT, or other 3D image data of the contralateral bone, and completion of the contralateral bone segmentation operation.

[0031] The third method is based on a statistical shape model corresponding to a surface mesh. This method first establishes a triangular mesh for bone samples from healthy individuals. Then, using a non-rigid iterative nearest-point registration algorithm, it establishes dense correspondences at the vertex level among all samples. Principal component analysis is performed on the mesh with established correspondences to construct a statistical shape model. When facing reconstruction of a defect on the affected side, the mean mesh of the model is non-rigidly registered to the surface of the residual bone on the affected side. The predicted morphology of the intact bone is obtained by solving for the low-dimensional shape coefficients.

[0032] The fourth method is the template library or average model matching method. This method selects a template from a pre-established bone model library of healthy individuals that is most similar to the morphology of the residual bone on the affected side, and generates a complete bone model that approximates the affected side through rigid body registration or scaling operations. The effectiveness of this method is highly dependent on the size of the template library and the representativeness of the stored templates.

[0033] The fifth method is a deep learning-based completion method. This method utilizes neural networks to learn the end-to-end mapping relationship between missing bone and complete bone, and the network directly predicts the geometry of the defect area or the complete 3D mesh of the skeleton. In recent years, it has gradually attracted attention with the development of deep learning technology.

[0034] However, all of the above-mentioned existing technical solutions have limitations and defects to varying degrees in practical applications: Manual experience reconstruction methods are highly dependent on the operator's experience level. Different doctors or engineers may give completely different completion results for the same case. The consistency and repeatability of the plan are poor, and the manual adjustment process is time-consuming, making it difficult to establish stable quantitative quality control and standardized operating procedures.

[0035] While the simple contralateral mirror method utilizes individualized information from the same patient, it essentially replaces or covers the affected side with the complete contralateral bone morphology. However, the left and right glenoid fossae are not perfectly symmetrical in terms of orientation, tilt, radius of curvature, and articular surface dimensions. Directly using a mirror model for completion may lead to glenoid fossa center displacement, articular surface orientation deviation, and subsequent implant matching errors. Furthermore, this method typically requires additional contralateral imaging data and complete contralateral bone segmentation, increasing the patient's radiation exposure and the complexity of preoperative preparation.

[0036] Statistical shape modeling methods based on surface mesh correspondences face a fundamental difficulty in establishing dense correspondences. These methods require establishing a one-to-one vertex topology for all sample meshes during the training phase. This process is highly sensitive to initial alignment, computationally expensive, and prone to errors such as flipping, folding, or misalignment in glenoid regions with drastic curvature changes or significant morphological differences. Inaccurate correspondences significantly degrade the quality of the statistical model and subsequent fitting accuracy. Furthermore, during inference, the topological correspondence at the defect boundaries is inherently unreliable, easily pulling the shape base to incorrect locations and affecting the anatomical plausibility of the reconstruction results.

[0037] Template library matching methods lack continuity and statistical constraints in shape space. They only select the most similar templates from a discrete set and scale them, making it difficult to accurately match the individual morphological differences of patients. This creates a strong dependence on the size and representativeness of the template library.

[0038] Deep learning-based completion methods typically require large amounts of training data with complete ground truth annotations, and the interpretability of the models is insufficient. In orthopedic preoperative planning scenarios, clinicians need to clearly understand whether the reconstruction results meet key anatomical indicators such as articular surface curvature, glenoid fossa center, and orientation angle. However, the prediction results provided by black-box networks are difficult to provide stable parameter constraints and uncertainty hints, and it is also difficult to provide quantitative guarantees that the reconstruction results fall within the reasonable morphological range of the population. Therefore, they have inherent shortcomings in meeting the needs of regulatory and clinical review.

[0039] In summary, existing technologies, when reconstructing local articular surface bone defects such as glenoid fossae, struggle to simultaneously achieve robust representation of volume occupancy and articular surface geometry, reasonable completion of statistical priors for defect areas, and convenient integration with voxel imaging workflows without establishing error-prone dense mesh vertex correspondences.

[0040] To address the aforementioned technological limitations, this invention provides a method and system for bone defect reconstruction. The core concept of this method lies in employing a truncated symbolic distance field as a unified implicit representation of bone morphology. By sampling training samples on a regular voxel grid within a local coordinate system defined by anatomical landmarks, all samples are naturally aligned voxel-by-voxel on the same grid, thus completely avoiding the need to establish vertex-level non-rigid correspondences between samples. Based on this, principal component analysis is performed on the local truncated symbolic distance field of the training samples to construct a low-dimensional statistical shape model, providing a group anatomical prior for the local morphology of the glenoid fossa. In the inference stage, the affected side containing the defect is sampled on the same local coordinate system and grid as the observed truncated symbolic distance field. By projecting the observations onto the statistical model for defect voxel detection, reliable observed voxels and defect voxels are identified. Then, only reliable observed voxels are used to perform masked least-squares fitting in a low-dimensional shape space to solve for the shape coefficients, allowing the morphology of the defect area to be generated primarily by statistical priors. Finally, a complete local symbolic distance field is reconstructed, and voxel-level repair results are output. The entire method achieves accurate, robust, and anatomically sound three-dimensional reconstruction of bony glenoid defects without the need for dense correspondence by organically combining three major mechanisms: implicit distance field representation, voxel-by-voxel alignment of local coordinate systems, and masked partial observation fitting.

[0041] The technical solution of the present invention will be further described below with reference to the accompanying drawings and a detailed description of specific embodiments.

[0042] It should be noted that the method provided in the embodiments of the present invention can be executed by a computing device with medical image data processing capability. The computing device can be an independent image processing workstation, a post-processing terminal matched with image acquisition equipment (such as CT, CBCT), a data processing system deployed on a hospital local server or a cloud computing platform, or a functional module integrated in an orthopedic preoperative planning software, an implant design system or a surgical navigation system. The computing device usually includes at least one or more processors, a memory storing relevant computer programs, a data input-output interface and an image display device, which can read and process three-dimensional image data in DICOM format or other standard medical image data formats, execute the modeling operation in the training stage and the reconstruction operation in the inference stage described in the present invention, and output the reconstruction result in the form of three-dimensional mask, triangular mesh or voxel distance field for use in subsequent links such as clinical planning, prosthesis design or 3D printed guide plate manufacturing. For ease of understanding, the embodiments of the present invention will hereinafter collectively refer to the subject executing the method as a reconstruction device, which can be implemented in software, hardware, or a combination of software and hardware.

[0043] Before describing the technical solutions of the embodiments of the present invention, schematic explanations of the nouns and terms involved in the embodiments of the present invention are provided.

[0044] Affected bone to be repaired: refers to the target bone with local bone defect in the joint fossa and its surrounding bone structure caused by degenerative lesions, traumatic fracture, bone tumor resection or revision surgery. The original anatomical shape of the target area in its three-dimensional image data is incomplete, and the shape of the defective area needs to be reconstructed and completed by the method provided in the embodiments of the present invention.

[0045] Statistical shape model: refers to a shape prior model obtained by performing statistical modeling on a set of healthy bone samples under a unified representation framework. In the embodiments of the present invention, it specifically refers to a statistical shape model constructed based on truncated signed distance fields. The model includes at least a mean vector, a principal component basis and corresponding eigenvalues, which can represent any bone shape as a coefficient vector in a low-dimensional shape space, so that the reasonable variation range of bone shape is constrained within the statistical distribution of healthy people.

[0046] Truncated signed distance field: refers to a scalar field defined on a regular voxel grid. The value of each voxel represents the signed distance from the voxel to the bone surface. It is agreed that the inside of the bone is negative and the outside of the bone is positive, and the distance value is truncated to a preset threshold range to suppress the interference of voxel information far away from the bone surface. The truncated signed distance field is a unified implicit representation form of bone shape in the embodiments of the present invention.

[0047] Confidential observation voxels refer to voxels in the observed truncated symbolic distance field of the bone to be repaired that are determined by defect detection to be unaffected by the defect and whose values ​​can accurately reflect the actual geometry of the patient's remaining bone. Confidential observation voxels are used in subsequent model fitting, while voxels in the defect area are excluded from the fitting.

[0048] Shape coefficients: These are low-dimensional coefficient vectors obtained when projecting the skeletal morphology onto the low-dimensional shape space of a statistical shape model. Within the expressive framework of the statistical shape model, the shape coefficients uniquely determine a reconstructed bone morphology. Solving for the shape coefficients is the core task of the inference stage, as their values ​​determine the specific geometric shape of the final reconstructed bone morphology.

[0049] Anatomical landmarks are identifiable and localizable feature points on the target bone with clear anatomical definitions, such as the glenoid center, superior pole, inferior pole, anterior border, and posterior border of the glenoid fossa. Anatomical landmarks are used to construct the local coordinate system of the target bone, which is the basis for aligning training samples and affected side samples in the same coordinate system.

[0050] World coordinate system: refers to the coordinate system used in the raw 3D medical image data, usually associated with the spatial positioning of the image acquisition equipment. The segmentation mask and triangular mesh of the skeleton are represented in 3D coordinates under the world coordinate system in the raw input data.

[0051] Local coordinate system: refers to a local orthogonal coordinate system defined based on the anatomical landmarks of the target skeleton, consisting of an origin and three orthogonal coordinate axes. All training samples and samples of the affected side to be repaired are transformed to a unified local coordinate system through rigid body transformation, thereby achieving pose alignment between samples.

[0052] Principal component basis: refers to the high-dimensional feature vectors obtained after performing principal component analysis on the truncated symbolic distance field of the training samples. Each dimension of the basis represents a major shape change direction in the training data. The principal component basis constitutes the low-dimensional shape space of the statistical shape model. The complete truncated symbolic distance field can be reconstructed by adding the mean vector to any linear combination of shape coefficients on the principal component basis.

[0053] Mean vector: The mean distance field is obtained by averaging the truncated sign distance field vectors of all training samples on a voxel-by-voxel basis. It represents the average anatomical morphology of the training population. The mean vector is an important component of the statistical shape model, serving as the origin of the shape space.

[0054] Occupied frequency field: refers to the frequency distribution of each voxel occupied by the skeleton, statistically obtained on the training sample set. Specifically, for each voxel position in a regular voxel grid, the truncated sign distance field value of that voxel in all training samples is statistically represented as the proportion of samples occupied by the skeleton, thus forming a frequency field of the same size as the voxel grid.

[0055] Fitting domain: refers to the set of voxels selected from the region of interest based on the occupancy frequency field, used for defect detection and shape coefficient fitting. The fitting domain is usually selected from voxels whose occupancy frequency falls within a set range, that is, those voxels that are sometimes occupied and sometimes not occupied in the population. These voxels carry the main individual shape differences and are the main area of ​​application for defect detection and model fitting in this invention.

[0056] Soft weights: These are weight coefficients between 0 and 1 assigned to reliable observation voxels during masked least squares fitting. The magnitude of the soft weights is determined by the distance of the voxel from the defect boundary. Voxels closer to the defect boundary are assigned smaller weights due to their lower measurement reliability, while voxels farther from the defect boundary are assigned weights close to 1 to reduce the influence of unreliable voxels near the defect boundary on the fitting results.

[0057] Voxel-level symbolic distance field: This refers to the symbolic distance field data stored voxel by voxel on a regular voxel grid. Each voxel location records a distance value, and the entire field forms a three-dimensional array with the same shape as the grid. The voxel-level symbolic distance field serves as both the basic data form for statistical modeling during the training phase and the final output form of the reconstruction results during the inference phase, and can be easily integrated with downstream voxel image processing workflows.

[0058] Local occupancy data refers to the voxel-level binary occupancy information obtained on a regular voxel mesh in a local coordinate system after zero-level set extraction and morphological cleaning. Specifically, voxels with distance values ​​less than or equal to zero are extracted from the reconstructed voxel-level symbolic distance field as voxels occupied by the skeleton. Then, small holes are filled by morphological closing operations, isolated noise points are removed by opening operations, and discrete abnormal regions are excluded by retaining the largest connected component. The clean voxel occupancy result obtained after the above cleaning is the local occupancy data.

[0059] Figure 1 This is a flowchart illustrating the bone defect reconstruction method provided by the present invention, which includes: Step 101: Obtain the three-dimensional data of the affected bone to be repaired and a pre-constructed statistical shape model, wherein the statistical shape model is constructed based on the truncated symbolic distance field of a healthy bone sample in a local coordinate system.

[0060] Before performing defect reconstruction, the reconstruction device first needs to acquire three-dimensional data of the affected bone to be repaired, as well as a pre-constructed statistical shape model. For example... Figure 2As shown, the input data in this embodiment of the invention includes ① a healthy target bone segmentation mask for constructing a statistical shape model, ② anatomical landmarks, ③ a segmentation mask for the affected bone to be repaired, and ④ a statistical shape model of the local truncated symbolic distance field. The affected bone to be repaired refers to the target bone with local bone defects in the glenoid fossa and surrounding bone structures due to degenerative diseases, traumatic fractures, bone tumor resection, or revision surgery. Its three-dimensional data is usually provided in the form of a medical image segmentation mask or a triangular mesh extracted from the segmentation mask. The statistical shape model is pre-constructed based on the truncated symbolic distance field of the target bone sample from the healthy population in the local coordinate system. This model includes at least a mean vector, principal component basis, and corresponding eigenvalues, used to represent any bone morphology as a coefficient vector in a low-dimensional shape space.

[0061] During the training phase, the reconstruction device processes segmentation masks or triangular meshes of multiple healthy bone samples, sampling the truncated signed distance field of each sample on a unified local coordinate system and regular voxel mesh. Principal component analysis is then used to construct the statistical shape model. The statistical shape model obtained in step 101 can be either pre-trained locally and stored in the reconstruction device, or imported from an external modeling system after training.

[0062] Step 102: Extract anatomical landmarks based on the three-dimensional data of the affected bone to be repaired, construct a corresponding local coordinate system, and perform voxel sampling in the corresponding local coordinate system to obtain the observation truncated symbolic distance field of the affected bone to be repaired.

[0063] After acquiring the 3D data and statistical shape model of the affected side, the reconstruction device extracts anatomical landmarks based on the 3D data of the bone to be repaired on the affected side, and uses the extracted anatomical landmarks to construct a local coordinate system isomorphic to the statistical shape model. Anatomical landmarks are characteristic points on the target bone with clear anatomical definitions. Taking the glenoid cavity of the scapula as an example, these may include the center of the glenoid cavity, the superior pole, the inferior pole, the anterior edge, and the posterior edge. For other locations such as the acetabulum, these can be replaced with the corresponding identifiable anatomical landmarks. The reconstruction device uses one of the anatomical landmarks as the origin of the local coordinate system, and uses the remaining landmarks to determine two sets of direction seed vectors. Through vector cross product and normalization, three mutually orthogonal coordinate axes are calculated, thus constructing a rigid body transformation relationship from the world coordinate system to the local coordinate system. The coordinates of all the bones to be repaired on the affected side are transformed to the local coordinate system through this rigid body transformation.

[0064] After the local coordinate system is constructed, the reconstruction device samples voxels on the same regular voxel grid defined by the statistical shape model to obtain the observed truncated symbolic distance field. Specifically, the reconstruction device transforms the segmentation mask or triangular mesh of the affected bone to the local coordinate system, calculates its symbolic distance field, and samples it on the regular voxel grid. The distance values ​​are truncated to a preset threshold range, and voxels outside the region of interest are uniformly set to the upper limit of the truncated threshold. After sampling, the observed truncated symbolic distance field is flattened into a one-dimensional vector, which serves as the input for subsequent defect detection. Since there is bone defect in and around the glenoid fossa of the affected bone to be repaired, the voxels in the defect area show positive values ​​in the observed truncated symbolic distance field, which provides a basis for the discrimination of defect voxel detection in subsequent steps. By sampling on the same local coordinate system and the same voxel grid, the observed data of the affected side and the statistical shape model are naturally aligned within the same spatial framework, without the need for additional non-rigid registration operations.

[0065] Step 103: Project the observed truncated symbol distance field onto the statistical shape model to perform defect detection and determine the reliable observation voxels in the observed truncated symbol distance field.

[0066] After acquiring the observation truncated symbolic distance field of the affected side, the reconstruction device projects it onto the statistical shape model for defect detection. It should be noted that the observation truncated symbolic distance field includes both voxels that can accurately reflect the actual geometry of the residual bone on the affected side and voxels that cannot reflect the true anatomical shape due to bone loss caused by the defect; if all voxels are used directly for model fitting, invalid observations of the defect area will pull the shape base to the wrong position.

[0067] To this end, the reconstruction device first projects the observed truncated symbolic distance field onto the statistical shape model to obtain an initial reconstructed reference distance field. Then, it calculates the difference between the observed distance field and this reference distance field within the fitting domain. In the defect region, the observed value is biased positive due to bone loss, while the reference value is biased negative because the model anticipates the presence of bone; therefore, the difference is significantly larger. Based on the statistical distribution of this difference, the reconstruction device adaptively determines an effective threshold and marks voxels within the fitting domain whose difference is less than or equal to this effective threshold as reliable observed voxels. Voxels whose difference is greater than this effective threshold are determined to be defective voxels and do not participate in subsequent model fitting.

[0068] Through the above-described defect detection operation, the reconstruction device identifies a set of reliable observation voxels that can be used for reliable fitting from all voxels in the observed truncated symbolic distance field.

[0069] Step 104: Fit the statistical shape model based on the reliable observation voxels to obtain shape coefficients, and reconstruct the complete morphology of the affected bone to be repaired based on the shape coefficients.

[0070] After determining the reliable observation voxels, the reconstruction device fits the statistical shape model to solve for shape coefficients based solely on these reliable observation voxels. Specifically, the reconstruction device extracts subsets of the principal component basis and mean vectors in the statistical shape model on the reliable observation voxels, and simultaneously extracts subsets of the observation truncated sign distance field on the reliable observation voxels. Based on this, a masked least squares fitting is performed with a preset regularization term to obtain low-dimensional shape coefficients. These shape coefficients are the unique parameter representations that determine the complete bone morphology in the low-dimensional shape space. Since the defective voxels are excluded from the fitting process, the specific morphology of the missing bone segment is not determined by invalid observations of the defect area, but is primarily generated by the morphological priors of the healthy population carried by the statistical shape model, thus ensuring the anatomical rationality of the reconstruction results. After obtaining the shape coefficients, the reconstruction device reconstructs the complete morphology of the bone on the affected side to be repaired based on these shape coefficients.

[0071] Specifically, the fitted shape coefficients are linearly combined with the principal component basis and a mean vector is added to reconstruct a complete voxel-level symbolic distance field; the zero level set of this symbolic distance field is the reconstructed complete local bone-occupying region. After morphological cleaning of the reconstructed voxels, the reconstruction device inversely transforms the reconstruction results in the local coordinate system back to the world coordinate system, outputting a complete voxel-level three-dimensional morphological mask, which serves as the final three-dimensional bone model after repairing the bone defect area on the affected side.

[0072] Through steps 101-104, this embodiment of the invention utilizes the population anatomical prior provided by a statistical shape model based on a truncated symbolic distance field, combined with voxel-level alignment and masked partial observation fitting mechanisms in a local coordinate system, to achieve accurate, robust, and anatomically plausible three-dimensional reconstruction of bony glenoid defect areas without establishing dense mesh correspondences between samples. The reconstruction results can be directly used for downstream clinical applications such as preoperative planning, implant design, and osteotomy guide fabrication.

[0073] Furthermore, in step 102, anatomical landmarks are extracted from the three-dimensional data of the bone to be repaired to construct a corresponding local coordinate system. The reconstruction device identifies and extracts multiple anatomical landmarks from the three-dimensional data of the bone to be repaired. These anatomical landmarks include a target landmark and multiple directional landmarks. Taking the glenoid cavity of the scapula as an example, the target landmark can be the registration center of the glenoid cavity, and the directional landmarks can include the superior pole, inferior pole, anterior edge, and posterior edge of the glenoid cavity. For other locations such as the acetabulum, they can be replaced with the identifiable anatomical landmarks corresponding to that location. The above-mentioned anatomical landmarks can be obtained by manually annotating each point on the three-dimensional image, or they can be automatically identified and located by a trained deep learning network or a traditional image feature detection algorithm. This embodiment of the invention does not limit the specific method of obtaining the landmarks.

[0074] After extracting the anatomical landmarks, the reconstruction device uses the target landmarks as the origin O of the local coordinate system and determines two sets of direction seed vectors based on multiple direction landmarks. Taking the glenoid fossa as an example, the vector between the anterior and posterior edges is taken as the first direction seed vector x. seed Take the vector from the upper pole to the lower pole as the second direction seed vector y. seed First normalize x seed Obtain the x-axis, then use the x-axis and y-axis seed The z-axis is obtained by cross product and normalization. Finally, the z-axis and x-axis cross product is used to obtain the y-axis, which is strictly orthogonal to both x and z. The z-axis is then compared with the x-axis. seed The inner product sign is unified with the y-axis and z-axis orientation to ensure consistent chirality of the coordinate system. For other anatomical sites, an orthogonal local coordinate system is constructed using the corresponding landmark points of that site in an equivalent manner.

[0075] After obtaining the origin and three orthogonal coordinate axes, the reconstruction device determines the rigid body transformation relationship from the world coordinate system to the local coordinate system: for any world coordinate point p world via p local =(p world Transform to the local coordinate system using -O)•A, where p local Let A be the coordinates of the point in the local coordinate system after the world coordinate transformation, and let A be the direction matrix composed of three orthogonal coordinate axes; the inverse transformation is p. world= O+p local •A T In this way, all training samples and the affected bone to be repaired are aligned in pose under the same anatomical definition. This is the basis for the present invention to achieve voxel-by-voxel alignment and avoid dense mesh correspondence.

[0076] Furthermore, regarding the pre-construction of the statistical shape model, the reconstruction device first transforms the 3D data of multiple healthy bone samples to a set local coordinate system and generates a regular voxel mesh within a defined region of interest (ROI). The ROI focuses on the glenoid fossa and its surrounding local structures rather than the entire bone, thus achieving high spatial resolution and statistical efficiency with a smaller voxel scale. The specific method for determining the ROI range is as follows: several training samples are transformed to the local coordinate system, and the range of each axis is statistically analyzed according to a set quantile within the neighborhood of the local origin (e.g., surface points less than the selected radius from the origin), and then extended outwards with a certain margin to obtain the minimum and maximum boundaries of the ROI; within the boundaries, a regular 3D mesh is generated with a set voxel spacing to obtain the query point set and mesh shape. The ROI preferably adopts a spherical region, that is, a spherical space with the origin of the local coordinate system as the center and a set radius. Voxels in the mesh whose distance from the origin exceeds the radius are marked as outside the ROI, making the modeling area more closely fit the approximate spherical distribution of the glenoid fossa and reducing irrelevant voxels.

[0077] Secondly, the reconstruction device calculates the symbolic distance field of each healthy bone sample and samples it on the aforementioned regular voxel grid to obtain the truncated symbolic distance field of each healthy bone sample. Figure 2 Section ④ illustrates the overall framework for statistical shape modeling. Figure 3 This visually demonstrates the sampling process from the segmented mask ⑤ to the truncated symbol distance field ⑥. For example... Figure 3 As shown, taking a healthy bone sample as an example, its segmentation mask ⑤, after sampling with a truncated symbolic distance field, yields a distance field ⑥ on a regular voxel grid in the local coordinate system, clearly displaying the distribution of signed distance values ​​for each voxel near the bone surface. The symbolic distance field defines negative values ​​inside the bone and positive values ​​outside, with its absolute value representing the distance to the bone surface. The distance values ​​are truncated to a preset threshold range, such as 6 mm or 8 mm. Voxel values ​​outside the region of interest are uniformly set to the upper limit of the truncated threshold, considered as being far outside the bone. The truncated symbolic distance field of each sample is flattened into a one-dimensional vector, serving as a row in the data matrix.

[0078] Next, the reconstruction device performs dimensionality reduction on the truncated symbolic distance field of each healthy bone sample, extracting the principal component basis and mean vector. Since the number of training samples is much smaller than the voxel dimension, the reconstruction device uses snapshot principal component analysis to achieve dimensionality reduction. Finally, the reconstruction device uses the truncated symbolic distance field of each healthy bone sample to statistically analyze the occupancy frequency of each voxel to divide the fitting domain, and uses the principal component basis, mean vector, and fitting domain together as the constructed statistical shape model.

[0079] Regarding the dimensionality reduction of the truncated symbolic distance field to extract the principal component basis and mean vectors, the reconstruction device stacks the truncated symbolic distance field vectors of all healthy bone samples into a data matrix X∈R. N×D Where N is the number of training samples, D is the voxel dimension, and R is the set of real numbers. The mean vector μ is obtained by calculating the voxel-by-voxel mean of the data matrix, and this mean vector is used to center the data matrix, resulting in the centered matrix X. c The reconstruction device then calculates the N×N Gram matrix G=X. c X c TThe eigenvalues ​​and eigenvectors of the low-dimensional eigenvectors are obtained by performing eigenvalue decomposition on the eigenvectors / (N-1). These eigenvalues ​​are then sorted in descending order, retaining positive eigenvalues. The high-dimensional principal component basis Φ is recovered from the low-dimensional eigenvectors, with each column representing a principal component direction, retaining the first few principal modes. Simultaneously, the shape score for each sample is obtained. The final statistical shape model includes at least: the mean vector μ, the principal component basis Φ, the eigenvalues ​​Λ, and the local coordinate system and mesh definition, the region of interest mask, the support domain mask, and the fitting domain mask. Through this dimensionality reduction process, the skeletal morphological changes in the original high-dimensional voxel space are compressed into a low-dimensional shape space, allowing subsequent shape coefficient calculations to be performed only in a few dimensions, significantly reducing computational complexity.

[0080] To define the fitting domain, the reconstruction device uses the truncation sign distance field of each healthy bone sample to statistically analyze the occupancy frequency of each voxel. Specifically, for each voxel in a regular voxel grid, the proportion of samples occupied by bones in all training samples is calculated, resulting in an occupancy frequency field. Specifically, for each voxel, the proportion of samples in all training samples where the truncation sign distance field value of that voxel is less than or equal to zero (i.e., occupied by bones) is calculated. Based on this occupancy frequency field, the reconstruction device marks voxels with an occupancy frequency not lower than a set threshold (e.g., 0.02) and located within the region of interest as the support domain, indicating that the voxel was once occupied by bones in the population and has modeling significance. Voxels with an occupancy frequency falling within a set interval (e.g., [0.05, 0.95]) and located within the region of interest are marked as the fitting domain, i.e., voxels that are sometimes occupied and sometimes not in the population. These voxels carry the main individual shape differences and are the main areas of application for subsequent defect detection and shape coefficient fitting. When the number of voxels in the fitting domain is too small, it can revert to the support domain. By dividing the support domain and the fitting domain, the reconstruction device excludes voxels that are almost always occupied or almost always unoccupied, effectively focusing on the area near the joint surface where individual differences are significant, thus improving the tightness of the statistical model and the robustness of subsequent fitting.

[0081] Furthermore, regarding the projection of the observed truncated symbolic range field onto the statistical shape model for defect detection to determine reliable observation voxels, the reconstruction device first projects the observed truncated symbolic range field onto the statistical shape model to obtain an initial reconstruction reference range field. Specifically, the observation vector x is first... obs Directly projected onto the statistical model, b0=Φ•(x obs -μ), b0 is the initial shape coefficient obtained by projecting the observation vector, and the reference range field x is then reconstructed using this initial shape coefficient. ref After obtaining the reference range field, the reconstruction device calculates the difference Δ=x between the observed vector and the initial reconstructed reference range field within the fitting domain. obs -x refIn the defect area, the distance field value of the truncated sign is positive due to the absence of bone on the affected side, while the reference distance field is negative due to the statistical prior expectation of bone presence. Therefore, the difference is significantly larger.

[0082] Subsequently, the reconstruction device takes the larger of the set quantile of the difference and the preset fixed threshold as the effective threshold: τeff = max (percentile(Δ fit ,95%),τ0), where Δ fit This represents the subset of values ​​for the difference Δ in the fitted domain, where τ0 is a preset fixed threshold (e.g., 2.5 mm). A combination of adaptive quantile thresholding and fixed thresholding is employed to balance detection sensitivity and robustness across different defect sizes.

[0083] Finally, the reconstruction device marks voxels with differences within the fitting domain less than or equal to the effective threshold as reliable observation voxels, and voxels with differences greater than the effective threshold as defective voxels, and outputs statistics such as the number of observation voxels, the number of defective voxels, and the effective threshold.

[0084] Through the above defect detection operation, the reconstruction device effectively identifies and eliminates unreliable observation data of the defect area, ensuring that only reliable voxels that reflect the true residual bone morphology participate in subsequent fitting. This is a key prerequisite to prevent the shape coefficient from being pulled to the wrong position.

[0085] Furthermore, let Ω be the index set of reliable observation voxels, and let the reconstruction device extract the subset of principal component basis on Ω. A subset of the mean vector on Ω μ Ω and the subset of the observed distance field on Ω Then, a least-squares fit is performed using a pre-defined regularization term to solve for the initial shape coefficients. b The optimization objective is: ; in, The initial shape coefficient for the current round. W The diagonal weight matrix is ​​composed of the soft weights of each reliable observation voxel. λ For regularization weights, For the first k The eigenvalues ​​of the principal components Initial shape factor b The kth component, K The total number of principal components to be retained. This optimization problem has a closed-ended solution: ; ; in, The transpose of the subset of the principal component basis on the reliable observation voxels. Let the difference vector be the observation range field and the mean vector on the reliable observation voxels. It is a diagonal matrix composed of the reciprocals of the eigenvalues. The initial shape coefficients are directly calculated by the reconstruction device.

[0086] After obtaining the initial shape coefficients, the reconstruction device truncates and trims the initial shape coefficients according to preset truncation parameters: ,in Shape factor The k One portion, For the first k The standard deviation of each principal component c This is the preset cropping factor. Truncation and cropping prevents the shape coefficient from deviating excessively from the statistical distribution range of healthy individuals, avoids non-anatomical morphological distortions in the reconstruction results due to observation noise or residual errors at defect boundaries, and further ensures the anatomical rationality of the reconstruction results.

[0087] Furthermore, regarding the introduction of soft weights and iterative refinement mechanisms in least squares fitting, in weighted least squares fitting, the reconstruction device assigns soft weights to each reliable observation voxel based on its distance from the defect boundary. Specifically, the smaller the distance from the reliable observation voxel to the defect boundary, the closer the voxel is to the defect region, and the lower the reliability of its measurement value; therefore, it is assigned a smaller soft weight. Conversely, the larger the distance from the reliable observation voxel to the defect boundary, the farther the voxel is from the defect region, and the higher the reliability of its measurement value; therefore, it is assigned a larger soft weight. The values ​​of the soft weights are limited to a preset range (e.g., between 0.1 and 1.0). After assigning the soft weights, the reconstruction device uses the assigned soft weights and a preset regularization term to perform weighted least squares fitting calculations on each subset corresponding to the reliable observation voxels to solve for the initial shape coefficients of the current round. in, W It is a diagonal weight matrix composed of soft weights.

[0088] In terms of iterative refinement, the reconstruction device reconstructs the reference morphology using the initial shape coefficients obtained in the current iteration and re-detects the defect to update the reliable observation voxels and soft weights. It then returns to the step of performing weighted least squares fitting calculations until a preset number of iterations is reached or the reliable observation voxels remain stable. Through multiple iterations, the delineation of the defect boundary and the solution of the shape coefficients are alternately optimized and gradually converged, effectively reducing the impact of initial defect detection bias on the reconstruction results. This is particularly suitable for complex cases with blurred or irregular defect boundaries, further improving the robustness and accuracy of the reconstruction method in complex defect scenarios.

[0089] Regarding the reconstruction of the complete morphology of the affected bone to be repaired based on the shape coefficient, the reconstruction device will iteratively refine the termination at the time of termination. As the final shape coefficient, the complete voxel-level symbolic distance field is reconstructed based on this final shape coefficient: This yields a complete voxel-level symbolic distance field covering the entire region of interest mesh. The value of each voxel in this symbolic distance field represents the signed distance from that location to the surface of the reconstructed skeleton, with its zero level being the surface of the reconstructed skeleton.

[0090] After obtaining the complete voxel-level symbolic distance field, the reconstruction device extracts voxels with values ​​less than or equal to zero as reconstructed occupied voxels, and performs morphological cleaning on these voxels to obtain local occupied data. Morphological cleaning specifically includes: using morphological closing operations to fill small holes, using morphological opening operations to remove isolated noise regions, and eliminating discrete outlier regions by preserving the largest connected components. After the above cleaning, clean and continuous local occupied data is obtained.

[0091] Finally, the reconstruction device inversely transforms the local occupancy data to the world coordinate system using the corresponding local coordinate system, outputting a three-dimensional morphological mask as the complete shape of the affected bone to be repaired. Specifically, the reconstruction results on the local mesh are inversely transformed to the local coordinate system p world =O+p local •A T Return to world coordinates and resample back to the original image coordinate system, outputting a voxel-level mask; if a mirroring operation was performed on the affected side of the bone to be repaired due to inconsistency between the left and right sides during preprocessing, the reconstruction result needs to be mirrored back to the original side. For example... Figure 4 As shown, after the target bone segmentation mask ⑦ on the affected side to be repaired is reconstructed according to the embodiment of the present invention, the complete bone morphology ⑧ with the defect area filled is obtained. The output three-dimensional morphology mask is the complete three-dimensional bone model after the defect area is repaired, which can be directly used for downstream clinical applications such as preoperative planning, implant design, osteotomy guide fabrication, or 3D printing. The above method realizes a complete conversion from statistical shape space to a clinically usable three-dimensional model, avoiding the accuracy loss that may be introduced during the voxelization process of traditional mesh models.

[0092] Compared with the prior art, the present invention has the following main technical advantages: First, it avoids dense grid correspondence. This invention uses a truncated symbolic distance field as a unified implicit shape representation. By sampling training samples on a regular voxel grid in a local coordinate system, all samples are naturally aligned voxel-by-voxel on the same grid. There is no need to establish vertex-level non-rigid correspondences between samples, which fundamentally avoids the correspondence errors and high computational costs caused by dense registration.

[0093] Second, the defect area is generated primarily by statistical priors. This invention uses a defect voxel detection and masked projection fitting mechanism to allow only reliable observation voxels to participate in the shape coefficient solution. The defect area is excluded from the fitting process, and its morphology is generated entirely by statistical shape priors from healthy individuals. This effectively prevents non-anatomical distortions in the reconstruction results and ensures the rationality of key anatomical indicators such as articular surface curvature and glenoid fossa orientation.

[0094] Third, it exhibits high robustness in reconstruction under complex defect scenarios. This invention combines adaptive threshold defect detection, boundary soft weighting mechanism, and multi-round iterative refinement, enabling alternating optimization and gradual convergence of defect boundary division and shape coefficient solution, effectively addressing complex clinical scenarios such as large-area defects, blurred or irregular defect boundaries.

[0095] Fourth, the voxel-level output facilitates seamless integration with downstream applications. This invention uses a voxel-level symbolic distance field as the representation of the reconstruction results, and the output three-dimensional morphological mask can be directly used for downstream clinical applications such as preoperative planning, implant design and selection, osteotomy guide fabrication, or 3D printing.

[0096] Fifth, location-independent universality. This invention does not rely on anatomical assumptions about specific bones. As long as a stable local coordinate system and identifiable anatomical landmarks can be defined for the target site, it can be applied to the reconstruction of defects in various articular surfaces or local bone structures, such as the glenoid fossa of the scapula, the acetabulum, the tibial plateau, and the proximal humerus.

[0097] The bone defect reconstruction system provided in the embodiments of the present invention is described below. The bone defect reconstruction system described below can be referred to in correspondence with the bone defect reconstruction method described above.

[0098] This invention provides a bone defect reconstruction system, see [link to relevant documentation]. Figure 5 ,include: The data acquisition module 510 is used to acquire three-dimensional data of the bone to be repaired on the affected side and a pre-constructed statistical shape model, wherein the statistical shape model is constructed based on the truncated symbolic distance field of a healthy bone sample in a local coordinate system. The sampling processing module 520 is used to extract anatomical landmarks based on the three-dimensional data of the bone to be repaired, to construct a corresponding local coordinate system, and to perform voxel sampling in the corresponding local coordinate system to obtain the observation truncated symbol distance field of the bone to be repaired. The defect detection module 530 is used to project the observation truncated symbol distance field onto the statistical shape model to perform defect detection and determine the reliable observation voxels in the observation truncated symbol distance field; The morphological reconstruction module 540 is used to fit the statistical shape model based on the reliable observation voxels to obtain shape coefficients, and reconstruct the complete morphology of the affected bone to be repaired based on the shape coefficients.

[0099] 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 communications bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other through the communications bus 840. The processor 810 can call logical instructions in the memory 830 to execute a bone defect reconstruction method. This method includes: acquiring three-dimensional data of the affected bone to be repaired and a pre-constructed statistical shape model, the statistical shape model being constructed based on a truncated symbolic distance field of a healthy bone sample in a local coordinate system; extracting anatomical landmarks based on the three-dimensional data of the affected bone to be repaired to construct a corresponding local coordinate system, and performing voxel sampling in the corresponding local coordinate system to obtain an observed truncated symbolic distance field of the affected bone to be repaired; projecting the observed truncated symbolic distance field onto the statistical shape model for defect detection, and determining reliable observed voxels in the observed truncated symbolic distance field; fitting the statistical shape model based on the reliable observed voxels to obtain shape coefficients, and reconstructing the complete morphology of the affected bone to be repaired based on the shape coefficients.

[0100] 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.

[0101] 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 can execute the bone defect reconstruction method provided by the above methods. The method includes: acquiring three-dimensional data of the affected bone to be repaired and a pre-constructed statistical shape model, wherein the statistical shape model is constructed based on the truncated symbolic distance field of a healthy bone sample in a local coordinate system; extracting anatomical landmarks based on the three-dimensional data of the affected bone to be repaired to construct a corresponding local coordinate system, and performing voxel sampling in the corresponding local coordinate system to obtain the observed truncated symbolic distance field of the affected bone to be repaired; projecting the observed truncated symbolic distance field onto the statistical shape model for defect detection, and determining the reliable observed voxels in the observed truncated symbolic distance field; fitting the statistical shape model based on the reliable observed voxels to obtain shape coefficients, and reconstructing the complete morphology of the affected bone to be repaired based on the shape coefficients.

[0102] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a bone defect reconstruction method provided by the above methods. The method includes: acquiring three-dimensional data of the affected bone to be repaired and a pre-constructed statistical shape model, the statistical shape model being constructed based on a truncated symbolic distance field of a healthy bone sample in a local coordinate system; extracting anatomical landmarks based on the three-dimensional data of the affected bone to be repaired to construct a corresponding local coordinate system, and performing voxel sampling in the corresponding local coordinate system to obtain an observed truncated symbolic distance field of the affected bone to be repaired; projecting the observed truncated symbolic distance field onto the statistical shape model for defect detection, and determining reliable observed voxels in the observed truncated symbolic distance field; fitting the statistical shape model based on the reliable observed voxels to obtain shape coefficients, and reconstructing the complete morphology of the affected bone to be repaired based on the shape coefficients.

[0103] 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.

[0104] 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.

[0105] 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 method for bone defect reconstruction, characterized in that, include: The three-dimensional data of the affected bone to be repaired and a pre-constructed statistical shape model are obtained. The statistical shape model is constructed based on the truncated symbolic distance field of a healthy bone sample in a local coordinate system. Anatomical landmarks are extracted based on the three-dimensional data of the bone to be repaired to construct a corresponding local coordinate system. Voxel sampling is performed in the corresponding local coordinate system to obtain the observation truncated symbol distance field of the bone to be repaired. The observed truncated symbol range field is projected onto the statistical shape model for defect detection to determine the reliable observation voxels in the observed truncated symbol range field; The statistical shape model is fitted with the reliable observation voxels to obtain shape coefficients, and the complete morphology of the affected bone to be repaired is reconstructed based on the shape coefficients.

2. The method according to claim 1, characterized in that, The step of extracting anatomical landmarks based on the three-dimensional data of the affected bone to be repaired, in order to construct a corresponding local coordinate system, specifically includes: Multiple anatomical landmarks were identified and extracted from the three-dimensional data of the affected bone to be repaired. The anatomical landmarks included target landmarks and multiple directional landmarks. The target marker point is taken as the origin of the coordinate system. Two sets of direction seed vectors are determined based on the multiple direction markers, and three orthogonal coordinate axes are obtained by vector cross product and normalization. Based on the origin and the three orthogonal coordinate axes, determine the rigid body transformation relationship from the world coordinate system to the corresponding local coordinate system, so as to construct the corresponding local coordinate system.

3. The method according to claim 1, characterized in that, The steps for constructing the statistical shape model include: The three-dimensional data of multiple healthy bone samples are transformed into a set local coordinate system, and a regular voxel mesh is generated in the defined region of interest. The truncated ... The truncation symbol distance field of each healthy bone sample is dimensionality reduced, and the principal component basis and mean vector are extracted. Based on the truncated symbol distance field of each healthy bone sample, the occupancy frequency of each voxel is statistically analyzed to divide the fitting domain, and the principal component basis, the mean vector, and the fitting domain are used together as the constructed statistical shape model.

4. The method according to claim 3, characterized in that, The dimensionality reduction of the truncated symbolic distance field of each healthy bone sample, and the extraction of the principal component basis and mean vector, includes: The truncated symbolic distance field of the healthy bone sample is converted into a one-dimensional vector and stacked to generate a data matrix; Calculate the mean of the data in the data matrix to obtain the mean vector; The data matrix is ​​centered using the mean vector. Principal component analysis is performed on the centered data matrix to obtain multiple eigenvalues ​​and corresponding eigenvectors. The eigenvectors are filtered based on the magnitude of the eigenvalues, and the retained eigenvectors are used as the principal component basis.

5. The method according to claim 3, characterized in that, The truncated symbol distance field based on each healthy bone sample, and the statistical analysis of the occupancy frequency of each voxel to delineate the fitting domain, includes: For each voxel in the voxel grid of the rule, the proportion of samples occupied by bones in the truncated symbol distance field of multiple healthy bone samples is statistically analyzed to obtain the occupancy frequency field. Voxels in the occupied frequency field whose occupied frequency is not lower than a set threshold are marked as support domains; The voxels whose occupancy frequencies fall within a set interval in the occupancy frequency field are labeled as the fitting domain.

6. The method according to claim 1, characterized in that, The step of projecting the observed truncated symbolic range field onto the statistical shape model for defect detection and determining the reliable observation voxels in the observed truncated symbolic range field includes: The observed truncated symbolic range field is projected onto the statistical shape model to obtain the initial reconstructed reference range field; Calculate the difference between the observed truncated symbol range field and the initial reconstructed reference range field; The larger value between the set quantile of the difference and the preset fixed threshold is taken as the effective threshold. Voxels whose difference is less than or equal to the effective threshold are marked as the reliable observation voxels.

7. The method according to claim 1, characterized in that, The process of fitting the statistical shape model based on the reliable observation voxels to obtain shape coefficients specifically includes: Extract the subset of model feature basis contained in the statistical shape model on the reliable observation voxel, the subset of mean data on the reliable observation voxel, and the subset of observation truncated symbolic distance field on the reliable observation voxel; By combining the preset regularization term, the initial shape coefficient is obtained by performing least squares fitting calculation using each subset corresponding to the reliable observation voxel. The initial shape coefficient is truncated and trimmed according to the preset truncation parameters to obtain the shape coefficient.

8. The method according to claim 7, characterized in that, The initial shape coefficients are obtained by combining a preset regularization term and performing least-squares fitting calculations using the subsets corresponding to the reliable observation voxels. Specifically, this includes: Soft weights are assigned to the reliable observation voxels based on the distance from the reliable observation voxel to the defect boundary, wherein the soft weights decrease as the distance from the corresponding voxel to the defect boundary decreases, and the values ​​of the soft weights are limited to a preset range. Using the assigned soft weights and the preset regularization term, a weighted least squares fitting calculation is performed on each subset corresponding to the reliable observation voxel to obtain the initial shape coefficient of the current round. The reference shape is reconstructed using the initial shape coefficients obtained in the current round, and defect detection is performed again to update the trusted observation voxels and the soft weights. The process then returns to the step of performing the weighted least squares fitting calculation until a preset number of iterations is reached or the trusted observation voxels remain stable.

9. The method according to claim 1, characterized in that, The process of reconstructing the complete morphology of the affected bone to be repaired based on the shape coefficient includes: Based on the shape coefficients and the statistical shape model, the complete voxel-level symbolic distance field is reconstructed; Voxels with values ​​less than or equal to zero in the complete voxel-level symbolic distance field are extracted as reconstructed occupied voxels, and morphological cleaning is performed on the reconstructed occupied voxels to obtain local occupied data. The local occupancy data is inversely transformed to the world coordinate system through the corresponding local coordinate system, and a three-dimensional morphological mask is output as the complete morphology of the affected bone to be repaired.

10. A bone defect reconstruction system, characterized in that, include: The data acquisition module is used to acquire three-dimensional data of the bone to be repaired on the affected side and a pre-constructed statistical shape model, wherein the statistical shape model is constructed based on the truncated symbolic distance field of a healthy bone sample in a local coordinate system. The sampling processing module is used to extract anatomical landmarks based on the three-dimensional data of the bone to be repaired, to construct a corresponding local coordinate system, and to perform voxel sampling in the corresponding local coordinate system to obtain the observation truncated symbol distance field of the bone to be repaired. The defect detection module is used to project the observation truncated symbol range field onto the statistical shape model to perform defect detection and determine the reliable observation voxels in the observation truncated symbol range field; The morphological reconstruction module is used to fit the statistical shape model based on the reliable observation voxels to obtain shape coefficients, and to reconstruct the complete morphology of the affected bone to be repaired based on the shape coefficients.

Images