Statistical shape atlas-based method for generating fracture data

Abstract

The application discloses a kind of based on statistical shape atlas's fracture data generation method, comprising: the statistical shape atlas of target skeleton is constructed, statistical shape atlas includes the statistical shape model based on healthy skeleton image data set is constructed, for generating anatomically reasonable diversified healthy skeleton form;Fracture mode library independent of specific bone shape is constructed, including by the fracture mode extracted by mapping clinical real fracture case to template, and / or by the synthetic fracture mode generated by geometric segmentation algorithm;Sample from statistical shape atlas to generate target healthy skeleton form, and select fracture mode from fracture mode library and map to target healthy skeleton form;Geometric perturbation is carried out to the fragment boundary after mapping, and fracture simulation data with multidimensional label is generated.The application can automatically and controllably generate a large amount of fracture simulation data with multidimensional, high-precision label.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, and more specifically, to a method for generating fracture data based on statistical shape atlases. Background Technology

[0002] The diagnosis and surgical treatment of fractures are among the most critical and complex clinical aspects of modern trauma orthopedics. With the development of imaging and computer-aided surgical techniques, technologies such as three-dimensional fracture reconstruction, reduction planning, internal fixation design, and intraoperative navigation are becoming increasingly common in clinical practice. However, the complexity and diversity of fracture types, the varying number of fragments, and the different fracture morphologies present significant challenges to preoperative planning and intraoperative procedures. Especially in complex fractures (such as pelvic, tibial plateau, or intra-articular fractures), precise reduction and internal fixation strategies directly impact postoperative functional recovery and complication rates; therefore, the intelligent research of related algorithms and systems has become an important direction in this field.

[0003] Currently, intelligent analysis and automated planning for fractures mainly rely on medical image processing and deep learning technologies, such as fracture region segmentation based on CT images, fragment registration, reduction posture prediction, screw path planning, and robotic surgical control algorithms. Although deep learning algorithms have achieved significant results in general medical image analysis tasks, their application in fracture scenarios remains severely limited. The core bottleneck lies in the lack of sufficient high-quality fracture data and accurately labeled preoperative planning samples.

[0004] Acquiring fracture imaging data is extremely difficult. Firstly, the uneven distribution of clinical cases slows the accumulation of real fracture CT data. Secondly, the "gold standard" annotation process for fracture data is complex. For example, in preoperative planning, surgeons need to segment, register, infer reduction posture, and design screw paths for each fragment—a process that often requires several hours for experienced experts. For reduction and internal fixation planning, annotation is highly dependent on individual anatomical structures and surgeon experience, making it subjective and difficult to standardize. Furthermore, in applications such as robot-assisted surgery, limitations imposed by clinical ethics and experimental conditions make it difficult to effectively train and validate algorithms using real data.

[0005] Therefore, how to simulate and generate large-scale fracture datasets in a high-quality and controllable manner to support algorithm training, validation, and evaluation has become a key issue that urgently needs to be addressed in the field of intelligent orthopedic research. Fracture data simulation generation technology can not only overcome the bottleneck of scarce real data, but also provide significant advantages in terms of ethical safety, repeatability, and structural consistency, providing a unified foundation for subsequent tasks such as automatic segmentation, morphological reconstruction, reduction planning, screw optimization design, and robot control. Summary of the Invention

[0006] The purpose of this invention is to propose a fracture data generation method based on statistical shape atlas, which can automatically and controllably generate a large amount of fracture simulation data with multi-dimensional and high-precision labels, providing a unified foundation for subsequent tasks such as automatic segmentation, morphological reconstruction, reduction planning, screw optimization design and robot control.

[0007] To achieve the above objectives, this invention proposes a method for generating fracture data based on statistical shape atlases, comprising:

[0008] S1: Construct a statistical shape atlas of the target bone, the statistical shape atlas including a statistical shape model constructed based on a healthy bone image dataset, used to generate anatomically reasonable and diverse healthy bone morphologies, and to annotate the template of the statistical shape model with clinical knowledge tags in a parametric form;

[0009] S2: Construct a fracture pattern library independent of specific bone shapes, the fracture pattern library including fracture patterns extracted by mapping real clinical fracture cases to the template, and / or synthetic fracture patterns generated by geometric segmentation algorithms;

[0010] S3: Sample and generate the target healthy bone morphology from the statistical shape atlas, and select a fracture pattern from the fracture pattern library and map it onto the target healthy bone morphology; perform geometric perturbation on the boundary of the mapped fragments to generate fracture simulation data with multi-dimensional labels; wherein, in the process of generating fracture simulation data, the clinical knowledge labels are automatically mapped onto the current fracture morphology and fragments based on the deformation field provided by the statistical shape model.

[0011] Optionally, constructing the statistical shape model in step S1 includes the following steps:

[0012] S101: Obtain a dataset of healthy medical images of the target skeleton, and preprocess the raw images to obtain a high-quality 3D skeleton model;

[0013] S102: Use a rigid registration algorithm to initially register 3D skeleton models of the same type to a common coordinate system to eliminate pose differences;

[0014] S103: Establish precise point-to-point correspondence for the registered skeletal model;

[0015] S104: Based on the point correspondence, principal component analysis is performed on the morphological changes of the skeletal model to obtain a shape parameter vector representing the morphological changes, thereby constructing the statistical shape model; wherein, by adjusting the shape parameter vector, diverse healthy skeletal morphologies with anatomical rationality are generated.

[0016] Optionally, step S101 specifically includes:

[0017] Acquire CT medical imaging data of the target bone;

[0018] The CT images are cropped, background areas are removed, and they are classified and archived according to bone type.

[0019] The cropped image is segmented into skeletons using an automatic segmentation algorithm, and the segmentation results are manually verified and corrected to generate accurate 3D skeleton segmentation labels.

[0020] The 3D segmentation labels of the skeleton are converted into a 3D mesh model, which is used as input data for constructing a statistical shape model.

[0021] Optionally, in step S102, the rigid registration algorithm is an iterative nearest point algorithm.

[0022] Optionally, in step S103, the point correspondence is established through a point cloud matching model based on deep learning;

[0023] The point cloud matching model is trained using a joint loss function, which includes a Chamfer distance loss to ensure spatial accuracy and a contrast loss to ensure pose robustness.

[0024] Optionally, in step S1, the clinical knowledge tags are labeled in a parametric form, including:

[0025] On the template, based on bony landmarks or anatomical specifications, determine and record the spatial coordinates and direction vectors of the start and end points of the safety screw channel, and / or the spatial coordinates of the major muscle attachment points.

[0026] Optionally, in step S2, the geometric segmentation algorithm includes the Voronoi diagram algorithm and the watershed algorithm.

[0027] Optionally, in step S3, geometric perturbation is applied to the mapped fragment boundaries to generate fracture simulation data with multi-dimensional labels, specifically including:

[0028] Geometric perturbations are applied to the mapped fragment boundaries to simulate a multi-fragment fracture model with irregular fracture surfaces, and the following multi-dimensional labels are automatically generated:

[0029] Segmentation labels: Automatically generate segmentation labels for each fragment, and accurately distinguish and label the fracture surface from the normal anatomical outer surface based on geometric rules;

[0030] Reset Labels: Define or generate the displacement of each fragment relative to its anatomical location to provide the target pose and path required for reset planning;

[0031] Screw planning tag: Maps the predefined screw channel information in the atlas to the current fracture morphology, performs feasibility screening and path optimization based on the distribution of fragments, and provides individualized screw implantation solutions;

[0032] Surgical scene planning tags: Map the muscle attachment point information marked in the atlas to the corresponding fragments to realize the spatial mounting of muscle soft tissue, providing an anatomical basis for mechanical constraint-based repositioning simulation, traction force calculation or virtual surgical scene simulation.

[0033] Optionally, in step S3, the geometric perturbation of the fragment boundary includes:

[0034] Apply a distance-field-based shear transformation to the fragment boundaries, and / or perform random vertex deletion or normal-vector-based random offset perturbation on the region near the boundaries to simulate irregular fracture surfaces.

[0035] Optionally, it also includes:

[0036] Step S5: Based on the mechanical data and motion trajectory collected by optical and force sensors during robot-assisted reduction surgery, the mechanical model parameters of the soft tissue mapped onto the fracture simulation data are optimized using a parameter fitting algorithm.

[0037] The parameter fitting algorithm employs least squares optimization or gradient descent methods, taking the error between the force output of the muscle model and the measured force as the optimization objective, and iteratively adjusts the elastic coefficient and contractile force parameters of the muscle.

[0038] The beneficial effects of this invention are as follows:

[0039] 1. Breaking through the bottleneck of data scarcity: By adopting a decoupled generation paradigm of "first generating healthy forms, then mapping fractures and knowledge", large-scale, automated, and on-demand production of fracture data has been achieved, completely eliminating the complete dependence on expensive, scarce, and difficult-to-label real clinical data, and providing almost unlimited data support for artificial intelligence algorithms.

[0040] 2. Multi-dimensional and multi-label support: The generated data is not a single geometric model, but an integrated, multi-dimensional labeled data body encompassing the entire process, from "geometric morphology (fragments, fracture surfaces) to surgical planning (screw channels, reduction posture) to physical physiology (attachment points, mechanical parameters)." A single sample can simultaneously support the research and verification of multiple algorithms such as segmentation, registration, path planning, and mechanical simulation, greatly improving the utilization value and efficiency of the data.

[0041] 3. High degree of anatomical and clinical rationality: Based on statistical shape models, the anatomical rationality of the generated skeleton is ensured; by integrating clinical fracture patterns and expert knowledge (screw channels, muscle attachment points), the simulation data is guaranteed to be highly consistent with clinical practice.

[0042] 4. Elevating simulation realism from the geometric level to the physical and physiological level: By introducing mechanical parameter optimization based on inversion from real surgical robot data, the generated fracture model can simulate real biomechanical responses. This provides an indispensable and safe training and testing environment for advanced applications such as surgical robot control algorithms and intraoperative force feedback prediction.

[0043] 5. High controllability, scalability, and security: The generation process is parameterized and modularized, allowing flexible control over the diversity, complexity, and focus of the generated data (such as specific fracture types). The framework is easily extended to different bone sites, and the entire process is completed in a virtual environment, completely avoiding clinical ethical risks and limitations of physical experiments. The conditions are repeatable, facilitating fair comparison and iteration of the algorithm.

[0044] The system of the present invention has other features and advantages that will be apparent from or will be set forth in detail in the accompanying drawings and following detailed description, which together serve to explain the particular principles of the invention. Attached Figure Description

[0045] The above and other objects, features and advantages of the present invention will become more apparent from the accompanying drawings, in which like reference numerals generally denote like parts.

[0046] Figure 1 This is a flowchart illustrating the steps of a fracture data generation method based on statistical shape atlas according to the present invention.

[0047] Figure 2 This is a schematic diagram of the statistical shape map construction process in an embodiment of the present invention.

[0048] Figure 3 This is a schematic diagram of the fracture data generation process in an embodiment of the present invention.

[0049] Figure 4 This is a schematic diagram of the muscle biomechanical parameter calibration and optimization process in an embodiment of the present invention.

[0050] Figure 5 This is a schematic diagram of the fracture data generation (five stages) in an embodiment of the present invention.

[0051] Figure 6 This is a schematic diagram of the generation and optimization of fracture screw channels in an embodiment of the present invention.

[0052] Figure 7 This is a schematic diagram illustrating the generation of a fractured muscle model in an embodiment of the present invention. Detailed Implementation

[0053] The invention will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0054] like Figure 1 As shown, a method for generating fracture data based on statistical shape atlas according to the present invention includes:

[0055] S1: Construct a statistical shape atlas of the target bone, the statistical shape atlas including a statistical shape model constructed based on a healthy bone image dataset, used to generate anatomically reasonable and diverse healthy bone morphologies, and to annotate the template of the statistical shape model with clinical knowledge tags in a parametric form;

[0056] The construction of the statistical shape model in this step includes the following steps:

[0057] S101: Obtain a dataset of healthy medical images of the target skeleton, and preprocess the raw images to obtain a high-quality 3D skeleton model, specifically including:

[0058] Acquire CT medical imaging data of the target bone;

[0059] The CT images are cropped, background areas are removed, and they are classified and archived according to bone type.

[0060] The cropped image is segmented into skeletons using an automatic segmentation algorithm, and the segmentation results are manually verified and corrected to generate accurate 3D skeleton segmentation labels.

[0061] The 3D segmentation labels of the skeleton are converted into a 3D mesh model, which is used as input data for constructing a statistical shape model.

[0062] S102: Use a rigid registration algorithm to initially register 3D skeleton models of the same type to a common coordinate system to eliminate pose differences;

[0063] The rigid registration algorithm can be selected as the iterative nearest point algorithm;

[0064] S103: Establish precise point-to-point correspondence for the registered skeletal model;

[0065] The point correspondence is established through a point cloud matching model based on deep learning. The point cloud matching model is trained using a joint loss function, which includes a Chamfer distance loss to ensure spatial accuracy and a contrast loss to ensure pose robustness.

[0066] S104: Based on the point correspondence, principal component analysis is performed on the morphological changes of the skeletal model to obtain a shape parameter vector representing the morphological changes, thereby constructing the statistical shape model; wherein, by adjusting the shape parameter vector, diverse healthy skeletal morphologies with anatomical rationality are generated.

[0067] In step S1, the parametric labeling of clinical knowledge tags includes: on the template, based on bony landmarks or anatomical norms, determining and recording the spatial coordinates and direction vectors of the start and end points of the safety screw channel, and / or the spatial coordinates of the major muscle attachment points.

[0068] In step S1, the geometric segmentation algorithm includes the Voronoi diagram algorithm and the watershed algorithm.

[0069] Specifically, this step constructs a statistical shape atlas of the target bone, which not only generates various anatomically reasonable healthy bone morphologies but also pre-integrates parameterized clinical knowledge labels. This step includes:

[0070] a. Anatomical Shape Model: Based on a dataset of healthy skeletal images, a statistical shape model is constructed by establishing precise point-to-point correspondences. This model can generate a large number of anatomically plausible, point-to-point corresponding, and diverse healthy skeletal morphologies by adjusting its shape parameters.

[0071] b. Clinical Knowledge Labeling: Key clinical knowledge is labeled in parametric form (such as spatial coordinates and direction vectors) on the average or representative template of the statistical shape model. This includes, but is not limited to, the start and end points of the safety screw channel and the attachment points of major muscles. These knowledge labels are bound to the geometric structure of the model template. When the model morphology changes, the labels can be automatically and accurately mapped to the newly generated individual morphology based on the deformation field, maintaining its anatomical accuracy.

[0072] S2: Construct a fracture pattern library independent of specific bone shapes, the fracture pattern library including fracture patterns extracted by mapping real clinical fracture cases to the template, and / or synthetic fracture patterns generated by geometric segmentation algorithms;

[0073] The geometric segmentation algorithm includes the Voronoi diagram algorithm and the watershed algorithm.

[0074] Specifically, this step constructs a reusable fracture pattern library independent of specific shapes. Each fracture pattern in the library defines a fragmentation method. Its construction sources include: mapping real clinical fracture cases onto a template of a statistical shape atlas using a registration algorithm to extract their fragmentation distribution patterns; and / or generating synthetic fracture patterns using geometric segmentation algorithms (such as Voronoi diagrams or watershed algorithms) to increase diversity.

[0075] S3: Sample and generate the target healthy bone morphology from the statistical shape atlas, and select a fracture pattern from the fracture pattern library and map it onto the target healthy bone morphology; perform geometric perturbation on the boundary of the mapped fragments to generate fracture simulation data with multi-dimensional labels; wherein, in the process of generating fracture simulation data, the clinical knowledge labels are automatically mapped onto the current fracture morphology and fragments based on the deformation field provided by the statistical shape model.

[0076] In this step, geometric perturbation is applied to the mapped fragment boundaries to generate fracture simulation data with multi-dimensional labels. Specifically, this includes:

[0077] Geometric perturbations are applied to the mapped fragment boundaries to simulate a multi-fragment fracture model with irregular fracture surfaces, and the following multi-dimensional labels are automatically generated:

[0078] Segmentation labels: Automatically generate segmentation labels for each fragment, and accurately distinguish and label the fracture surface from the normal anatomical outer surface based on geometric rules;

[0079] Reset Labels: Define or generate the displacement of each fragment relative to its anatomical location to provide the target pose and path required for reset planning;

[0080] Screw planning tag: Maps the predefined screw channel information in the atlas to the current fracture morphology, performs feasibility screening and path optimization based on the distribution of fragments, and provides individualized screw implantation solutions;

[0081] Surgical scene planning tags: Map the muscle attachment point information marked in the atlas to the corresponding fragments to realize the spatial mounting of muscle soft tissue, providing an anatomical basis for mechanical constraint-based repositioning simulation, traction force calculation or virtual surgical scene simulation.

[0082] The geometric perturbation of the fragment boundary includes: applying a distance-field-based shear transformation to the fragment boundary, and / or performing random vertex deletion or random offset perturbation based on the normal vector in the region near the boundary to simulate an irregular fracture surface.

[0083] Specifically, this step is the controllable fracture data generation step, which includes: sampling and generating a target healthy bone morphology from the statistical shape atlas, then selecting a fracture pattern from the fracture pattern library and mapping it onto the healthy morphology. By geometrically perturbing the fragment boundaries (such as deformation or point loss), a multi-fragment fracture model with irregular fracture surfaces is simulated and generated, and the following multi-dimensional labels are automatically generated:

[0084] a. Segmentation Labels: Automatically generate segmentation labels for each fragment and accurately distinguish and label the fracture surface from the normal anatomical outer surface based on geometric rules (such as distance and normal comparison).

[0085] b. Reset Labels: These can define or randomly generate the displacement of each fragment relative to its anatomical position, providing the target pose and path required for reset planning.

[0086] c. Screw planning label: Maps the predefined screw channel information in the atlas to the current fracture morphology, and performs feasibility screening (such as eliminating channels that seriously conflict with the fracture line) and path optimization (such as adjusting the channel direction on the complete bone morphology to seek the maximum bone volume) based on the distribution of fragments, providing an individualized screw implantation plan for internal fixation surgery.

[0087] d. Surgical Scene Planning Tag: This tag maps information such as muscle attachment points marked in the atlas to corresponding fragments, enabling spatial mounting of soft tissues such as muscles. This tag provides the anatomical basis for mechanically constrained repositioning simulation, intraoperative traction force calculation, and virtual surgical scene simulation, thereby supporting the planning and verification of the entire surgical environment.

[0088] Optionally, the method further includes step S5: based on the mechanical data and motion trajectory collected from the robot-assisted reduction surgery by optical sensors and force sensors, the mechanical model parameters of the soft tissue mapped onto the fracture simulation data are optimized by a parameter fitting algorithm.

[0089] The parameter fitting algorithm employs least squares optimization or gradient descent methods, taking the error between the force output of the muscle model and the measured force as the optimization objective, and iteratively adjusts the elastic coefficient and contractile force parameters of the muscle.

[0090] Specifically, this step is a physical simulation enhancement step. Based on the mechanical data and motion trajectory collected in a real robot-assisted reduction surgery, the mechanical model parameters of muscles and other soft tissues mapped onto the fracture model are optimized through a parameter fitting algorithm, so that the simulated environment has biomechanical realism.

[0091] The present invention is illustrated below through an example of pelvic fracture data simulation, but it should not be regarded as a limitation on the scope of protection of the present invention.

[0092] Example

[0093] This embodiment provides a method for generating fracture data based on statistical shape atlases, including the following steps:

[0094] Step 1. Preparation and preprocessing of healthy bone data, the specific process is as follows: Figure 2 As shown:

[0095] a. Collect CT or other medical imaging data of healthy bones, covering the target bone region. Crop the original images to remove background areas unrelated to the bones, reducing computation and improving model accuracy. Classify and archive the data according to bone type (e.g., ilium, femur, tibia, etc.) for subsequent processing.

[0096] b. Use existing automatic segmentation algorithms (such as TotalSegmentor or other deep learning-based skeletal segmentation tools) to perform skeletal segmentation on the cropped image to generate a preliminary 3D skeletal structure. Manually check and correct the automatic segmentation results as necessary to ensure accurate bone labeling.

[0097] c. Convert the segmented bone data into a 3D mesh to provide high-quality input for anatomical morphology modeling.

[0098] Step 2. Construction of the anatomical shape model, continue to refer to Figure 2 :

[0099] a. Initially register bones of the same type to a common coordinate system, and use the Iterative Closest Point (ICP) algorithm to make all sample bones overlap as much as possible in space, thereby eliminating the influence of pose differences on subsequent modeling.

[0100] b. Establish point correspondences for the registered healthy skeletal data to ensure accurate alignment of key points in the skeletons of different individuals. These point correspondences are implemented using a deep learning-based point cloud matching model (Point2SSM++). The model's input consists of the vertex coordinates and corresponding normal vectors of the downsampled 3D skeletal mesh. A Graph Convolutional Point Cloud Network (DGCNN) extracts the local features of each point, and an attention module generates an attention map between points. This map is then multiplied with the input point features to obtain the corresponding position of each input point on the reference skeleton, thus achieving a complete point correspondence.

[0101] c. The model training uses a joint loss function, including the Chamfer distance loss (CD loss) between the input point cloud and the output point cloud to ensure the spatial accuracy of the points; it also includes a contrastive loss, which uses the mean square error (MSE) of the corresponding position of the output point to measure the same skeleton input into the model under different poses, ensuring that the model maintains a stable correspondence under different poses.

[0102] d. After completing the point mapping, the point cloud data of each bone sample is vectorized using Principal Component Analysis (PCA). By analyzing the morphological variation patterns of the training data, the main shape components and individual differences are extracted to construct a Statistical Shape Model (SSM). This model can generate the geometric morphology of different individual bones by adjusting the shape parameters, while maintaining the rationality of the anatomical structure, providing a basic bone model for subsequent fracture pattern mapping and fracture data generation.

[0103] In other embodiments, non-rigid registration methods (such as non-rigid iterative nearest-point algorithms based on point cloud meshes, or differential homeomorphic registration algorithms based on voxel image grayscale) can also be used to obtain point-to-point correspondences. Furthermore, other deep learning frameworks besides the networks described in the embodiments (such as point cloud matching networks based on graph attention mechanisms) can also be used to establish point correspondences. Regarding the construction of variable shape models, Gaussian process deformable models can be used as the mathematical basis for generating diverse and anatomically sound morphologies. A more direct alternative is to treat the large set of healthy skeletal models with established precise point correspondences as a discrete shape space, from which existing shapes can be directly sampled, or new transitional morphologies can be synthesized by interpolation or weighted averaging between different individual shapes.

[0104] Step 3. Clinical label construction; refer to the detailed process. Figure 3 :

[0105] All labels are established on a statistical shape model and maintain a one-to-one correspondence with the model's morphological parameters. When the statistical shape model undergoes geometric deformation, all label points, channels, and regions will be updated synchronously with the deformation field, thereby maintaining their relative positions and physiological significance within the anatomical structure and ensuring the stability of anatomical correspondences under different individual morphologies.

[0106] a. Fracture Pattern Labeling: Clinical fracture fragments are rigidly registered and aligned with a statistical shape template. A bidirectional nearest neighbor algorithm is used to determine the correspondence between template points and fragment points, and the morphological differences between the template and the main fragments are calculated. The deformation parameters of the statistical shape model are dynamically updated based on the difference results, allowing the template to gradually approximate the morphology of the clinical fragments. Through multiple iterations, the optimal set of corresponding points between the template and fragments is obtained, and the regions on the template corresponding to different fragments are labeled with different categories, thus forming a clinically representative fracture pattern library. To enhance the diversity and coverage of fracture patterns, Voronoi region partitioning, a watershed algorithm based on Signed Distance Function (SDF), and other geometrically heuristic segmentation methods can be combined to automatically partition the skeleton, generating richer fracture samples.

[0107] b. Screw Channel Marking: The entry point of the screw channel is usually located in a relatively fixed anatomical position and can be determined based on clinical standards and surgical experience. For example, common screw channels in the pelvic region include anterior iliac column screws, posterior iliac column screws, iliopectine screws, and sacroiliac screws. Their entry points and channel paths can be accurately marked on a healthy bone model based on bony landmarks. Orthopedic experts mark the start and end areas of each channel on the 3D model and determine the channel direction and feasible range by combining screw diameter, bone density, and local bone thickness. The established screw channel information maintains consistent anatomical correspondence under different bone morphologies and is used for fracture formation and preoperative planning simulation.

[0108] c. Muscle Attachment Point Labeling: The attachment points of major muscles are labeled on the skeletal model to support subsequent reduction simulation and biomechanical analysis. Muscle attachment points are typically located in anatomically stable bone surfaces, and their locations can be determined based on standard anatomical data and expert experience. Orthopedic experts label the attachment points of major muscle groups on the 3D model and record the direction vectors and their spans, forming a set of muscle attachment labels. These established muscle attachment points can be mapped to the corresponding fragment surfaces during fracture formation, achieving spatial binding between muscle attachments and fragments. This supports biomechanically driven fracture reduction, intraoperative force data fitting, and virtual surgical scenario simulation.

[0109] In other embodiments, besides mapping real clinical fracture cases to atlas templates, fracture pattern generation can also achieve a variety of pattern libraries through the following methods:

[0110] Physical simulation-based methods: such as using finite element analysis to simulate the bone fracture process under stress, generating fracture patterns with biomechanical basis.

[0111] Algorithm-based methods: These methods do not rely on clinical data and can quickly and cost-effectively generate diverse fragmentation patterns. Examples include random segmentation based on geometric features, segmentation based on bone density, or the use of physics-inspired algorithms specifically designed to simulate brittle fractures (such as "Breaking Bad").

[0112] Step 4. Fracture data generation, continue to refer to Figure 3 :

[0113] After completing the anatomical morphology model construction and clinical labeling, fracture data are automatically generated based on the statistical shape model. This process takes a morphological model with anatomical labels, screw channels, and muscle attachment point information as input. Through steps such as parameter sampling, fracture pattern mapping, fragment generation, and posture perturbation, it generates diverse fracture samples with anatomical consistency and clinical rationality. The overall workflow is as follows: Figure 5 As shown, it includes five stages: anatomical shape sampling, fragment segmentation generation, fracture surface annotation, screw channel screening, and reset posture annotation.

[0114] a. Anatomical Shape Sampling: Shape parameter vectors are randomly sampled from the statistical shape model. Based on the principal component distribution law during the model training phase, normal distribution sampling is performed in the parameter space to generate healthy bone morphology samples S for different individuals. The sampled morphology samples are the anatomical shape label data of healthy bones.

[0115] b. Fragment Label Generation: Based on the mapped fracture pattern labels, local non-rigid deformation is applied to the boundaries of each label region to simulate the irregular fracture morphology in real fractures. Specifically, firstly, a distance-field-based shear transformation is applied to the label boundaries between different fragments, causing deformation perturbation in the region near the fracture boundary; secondly, random point loss or surface perturbation is applied to some regions near the boundary to simulate the fragment and defect characteristics in clinical fractures. Through this segmentation method based on geometric perturbation, a multi-fragment fracture model with reasonable morphological continuity and natural fracture boundaries can be generated. Finally, the label regions of each fragment serve as training labels for the fracture segmentation task.

[0116] c. Fracture Surface and Anatomical Surface Labeling: After generating fracture fragments, the fracture surfaces between fragments and the outer surface regions of the bone are automatically identified through geometric neighborhood relationships. Specifically, the method is as follows: First, based on the spatial distance between fragments, vertex pairs with a distance less than a preset adjacency threshold are selected to determine neighboring regions where fracture relationships may exist. Then, a healthy template is used for further filtering: if a candidate point is too close to the template surface and the angle between its normal direction and the template normal is small (e.g., the normal cosine is greater than 0.9), then the point is determined to belong to the outer surface of the bone rather than the fracture surface; otherwise, it is marked as a fracture surface region. This criterion effectively distinguishes fracture surfaces from normal anatomical surfaces, providing high-precision labels for fracture surface segmentation.

[0117] d. Screw channel generation: such as Figure 6 As shown, firstly, all screw channels generated during the annotation stage are mapped onto the currently generated fracture shape based on the deformation field of the anatomical morphology model, ensuring that the channel position is consistent with the current individual bone morphology. Based on this, channels intersecting the current fracture surface are selected as the effective screw channel set for that fracture sample. Subsequently, to ensure the clinical feasibility of the screw path, channels are optimized on the original unfractured complete anatomical model. During optimization, based on the spatial relationship between the screw channel and the cortical bone thickness and outer surface, while maintaining the insertion point and overall spatial position, the channel direction and diameter are adjusted to place it within the range of being furthest from the outer surface, having the longest channel length, and the largest feasible diameter. Finally, the optimized screw channels on the unfractured morphology can be directly used as reference labels for screw planning, their spatial position corresponding to the current fracture sample, thus ensuring the stable usability of the screw path under individualized morphology.

[0118] e. Muscle attachment mapping and generation: such as Figure 7 The system pre-labels muscle attachment points in the skeletal anatomy model, mapping the deformation field of the statistical shape model to the current fracture morphology model, enabling adaptive updates of the spatial positions of attachment points under individualized morphologies. When a bone fractures and is segmented, the bone surface to which the original attachment points belong is divided into corresponding fragment surfaces. The system automatically identifies the fragment number and local coordinates corresponding to each attachment point, achieving spatial binding between muscle attachment points and fragments. Subsequently, based on the pose transformation matrix of each fracture fragment, the positions of the attachment points on the corresponding fragments are updated synchronously to ensure that their relative spatial relationships in the fragment coordinate system remain consistent. The resulting muscle attachment point labels stably reflect the distribution of muscle origin and insertion points under different fracture states, providing accurate physiological data support for subsequent fracture reduction simulation based on mechanical constraints, intraoperative traction direction calculation, and virtual surgery training.

[0119] Furthermore, it should be noted that statistical shape maps achieve stable mapping of clinical labels through inherent point correspondences, which is the most direct and efficient method. However, the core idea of ​​this invention—"mapping parameterized clinical knowledge from a template to a new morphology"—is also applicable to scenarios without explicit point correspondences. For example, for a new, external skeletal morphology data (such as an independent 3D model not involved in the map construction), a non-rigid registration algorithm can be used to align the integrated clinical knowledge template with it with high precision. The resulting continuous deformation field can then be used to map parameterized labels such as screw channels and muscle attachment points defined on the template to the new morphology in a one-time, holistic manner.

[0120] Step 5. Muscle biomechanical parameter calibration and optimization; refer to the detailed procedure. Figure 5 :

[0121] After generating the fracture data, to achieve fracture reduction simulation based on mechanical constraints, it is necessary to assign physiologically reasonable mechanical parameters to each muscle attachment point. The specific implementation method is as follows:

[0122] a. Setting initial muscle biomechanical parameters: Based on literature and anatomical data, linear or nonlinear elastic models (such as linear spring models, Hill-type muscle models, etc.) are used to describe muscle biomechanical properties. Parameters such as initial elastic coefficients, maximum elongation lengths, and contractile force ranges are assigned to each major muscle group to construct a preliminary biomechanical model.

[0123] b. Generate clinically corresponding samples: Using the fracture samples generated in "Step 4. Fracture Data Generation," including bone morphology, fracture location, and the posture of each fragment, a fracture model consistent with real robotic fracture surgery is established. The muscle attachment points on each fracture fragment maintain a corresponding relationship with the fracture posture.

[0124] c. Robotic Reduction Data Acquisition and Mapping: During the actual robotic fracture reduction process, the motion trajectory of the fracture fragments along the reduction path and the resulting changes in muscle force are recorded (using NDI optical sensors and ATI force sensors). The robot's operating path is then mapped onto the fracture model, allowing the model to simulate muscle mechanical responses along the same path.

[0125] d. Muscle biomechanical parameter optimization: Based on the mechanical data collected by the robot, the muscle biomechanical parameters are optimized using parameter fitting methods to make the simulated muscle force changes as close as possible to the actual measured mechanical curve. Specifically, least squares optimization or gradient descent methods can be used, with the error between the muscle model force output and the robot's measured force as the optimization objective function. The muscle elastic coefficient and related parameters are iteratively adjusted until convergence to obtain the best fitting parameters.

[0126] In other embodiments, in addition to the least squares fitting described above, more advanced optimization algorithms, such as Bayesian optimization, can be used to address the problems of complex parameter spaces and high evaluation costs; or a reinforcement learning framework can be used to allow the muscle model to autonomously learn the optimal parameters through interaction with the simulation environment.

[0127] Application of Results: The optimized muscle mechanics parameters correspond one-to-one with the fracture fragment model and attachment point labels, and can be used for subsequent fracture reduction simulation based on mechanical constraints, intraoperative traction force calculation, and virtual surgery training, providing accurate mechanical basis for individualized fracture surgery planning.

[0128] Based on the above, it can be seen that the core idea of ​​this invention lies in decoupling the fracture data generation process into two core stages: controllable anatomical morphology generation and multi-dimensional clinical knowledge mapping. In this way, the core bottlenecks of scarce real fracture data and difficulty in annotation are effectively solved. Several key points of the method of this invention are as follows:

[0129] 1. Generative Basis Based on Statistical Shape Maps: The key lies in using variable models (such as statistical shape models) that can establish stable point correspondences or provide continuous deformation fields as the basis for anatomical morphology generation. This ensures that all generated skeletal morphologies have anatomical rationality and provides a geometric consistency guarantee for the accurate mapping of subsequent knowledge.

[0130] 2. Construction and Application of Fracture Pattern Library: Fracture patterns serve as mappable templates. Whether learned through registration with real clinical cases or generated by geometric algorithms, this pattern library is independent of specific bone shapes and can be flexibly mapped onto newly generated anatomical morphologies. Geometric perturbations enhance the realism of fracture surfaces, thereby enabling the diversified generation of fracture types.

[0131] 3. Parametric and Automatic Mapping Mechanism of Clinical Planning Tags: The key lies in pre-labeling core clinical knowledge such as screw safety channels and muscle attachment points in parametric form (e.g., coordinates, direction vectors) onto the anatomical atlas template. When the atlas generates a new morphology, these tags can automatically and accurately map to the correct anatomical position of the new morphology based on the deformation field, ensuring that the generated data is directly applicable to the preoperative planning algorithm.

[0132] 4. Enhanced Physical Simulation Based on Real Data Inversion: A key enhancement feature is the use of mechanical data collected during real surgeries (such as robot repositioning paths and force feedback) to inversely calibrate the mechanical parameters of soft tissues in the simulated environment through parameter optimization algorithms. This elevates data generation from the geometric level to the physical and physiological level, providing a crucial training and validation foundation for advanced applications such as surgical robot control.

[0133] 5. Multi-dimensional data label generation: As a complete "digital twin", the dataset's label system covers the entire process from diagnosis and planning to surgical simulation. Specifically, it includes: morphological labels (such as fragment segmentation and fracture surface identification), planning labels (such as individualized safety screw channels, reset paths and target postures), and physiological and physical labels (such as muscle attachment points and mechanical parameters derived from real data).

[0134] In summary, this invention seeks to protect not only a specific algorithm, but also a complete technical solution for generating high-quality fracture data through the systematic integration of anatomical morphology, clinical knowledge, and physical simulation. The organic combination of these key points constitutes the core competitiveness that distinguishes this invention from existing technologies.

[0135] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A method for generating fracture data based on statistical shape atlas, characterized in that, include: S1: Construct a statistical shape atlas of the target bone, the statistical shape atlas including a statistical shape model constructed based on a healthy bone image dataset, used to generate anatomically reasonable and diverse healthy bone morphologies, and to annotate the template of the statistical shape model with clinical knowledge tags in a parametric form; S2: Construct a fracture pattern library independent of specific bone shapes, the fracture pattern library including fracture patterns extracted by mapping real clinical fracture cases to the template, and / or synthetic fracture patterns generated by geometric segmentation algorithms; S3: Sample and generate the target healthy bone morphology from the statistical shape map, and select a fracture pattern from the fracture pattern library and map it onto the target healthy bone morphology; Geometric perturbation is applied to the mapped fragment boundaries to generate fracture simulation data with multi-dimensional labels; wherein, during the generation of fracture simulation data, the clinical knowledge labels are automatically mapped onto the current fracture morphology and fragments based on the deformation field provided by the statistical shape model.

2. The method according to claim 1, characterized in that, The construction of the statistical shape model in step S1 includes the following steps: S101: Obtain a dataset of healthy medical images of the target skeleton, and preprocess the raw images to obtain a high-quality 3D skeleton model; S102: Use a rigid registration algorithm to initially register 3D skeleton models of the same type to a common coordinate system to eliminate pose differences; S103: Establish precise point-to-point correspondence for the registered skeletal model; S104: Based on the point correspondence, principal component analysis is performed on the morphological changes of the skeletal model to obtain a shape parameter vector representing the morphological changes, thereby constructing the statistical shape model; wherein, by adjusting the shape parameter vector, diverse healthy skeletal morphologies with anatomical rationality are generated.

3. The method according to claim 2, characterized in that, Step S101 specifically includes: Acquire CT medical imaging data of the target bone; The CT images are cropped, background areas are removed, and they are classified and archived according to bone type. The cropped image is segmented into skeletons using an automatic segmentation algorithm, and the segmentation results are manually verified and corrected to generate accurate 3D skeleton segmentation labels. The 3D segmentation labels of the skeleton are converted into a 3D mesh model, which is used as input data for constructing a statistical shape model.

4. The method according to claim 2, characterized in that, In step S102, the rigid registration algorithm is an iterative nearest point algorithm.

5. The method according to claim 2, characterized in that, In step S103, the point correspondence is established through a point cloud matching model based on deep learning; The point cloud matching model is trained using a joint loss function, which includes a Chamfer distance loss to ensure spatial accuracy and a contrast loss to ensure pose robustness.

6. The method according to claim 1, characterized in that, In step S1, the clinical knowledge tags are labeled in a parametric form, including: On the template, based on bony landmarks or anatomical specifications, determine and record the spatial coordinates and direction vectors of the start and end points of the safety screw channel, and / or the spatial coordinates of the major muscle attachment points.

7. The method according to claim 1, characterized in that, In step S2, the geometric segmentation algorithm includes the Voronoi diagram algorithm and the watershed algorithm.

8. The method according to claim 1, characterized in that, In step S3, geometric perturbation is applied to the mapped fragment boundaries to generate fracture simulation data with multi-dimensional labels, specifically including: Geometric perturbations are applied to the mapped fragment boundaries to simulate a multi-fragment fracture model with irregular fracture surfaces, and the following multi-dimensional labels are automatically generated: Segmentation labels: Automatically generate segmentation labels for each fragment, and accurately distinguish and label the fracture surface from the normal anatomical outer surface based on geometric rules; Reset Labels: Define or generate the displacement of each fragment relative to its anatomical location to provide the target pose and path required for reset planning; Screw planning tag: Maps the predefined screw channel information in the atlas to the current fracture morphology, performs feasibility screening and path optimization based on the distribution of fragments, and provides individualized screw implantation solutions; Surgical scene planning tags: Map the muscle attachment point information marked in the atlas to the corresponding fragments to realize the spatial mounting of muscle soft tissue, providing an anatomical basis for mechanical constraint-based repositioning simulation, traction force calculation or virtual surgical scene simulation.

9. The method according to claim 8, characterized in that, In step S3, the geometric perturbation of the mapped fragment boundaries includes: Apply a distance-field-based shear transformation to the fragment boundaries, and / or perform random vertex deletion or normal-vector-based random offset perturbation on the region near the boundaries to simulate irregular fracture surfaces.

10. The method according to claim 1, characterized in that, Also includes: Step S5: Based on the mechanical data and motion trajectory collected by optical and force sensors during robot-assisted reduction surgery, the mechanical model parameters of the soft tissue mapped onto the fracture simulation data are optimized using a parameter fitting algorithm. The parameter fitting algorithm employs least squares optimization or gradient descent methods, taking the error between the force output of the muscle model and the measured force as the optimization objective, and iteratively adjusts the elastic coefficient and contractile force parameters of the muscle.

Images