Fracture reduction path planning system and method based on image data

Abstract

The invention discloses a fracture reduction path planning system and method based on image data, and the method comprises the steps: carrying out the fine anatomical structure segmentation of high-resolution preoperative image data, firstly constructing a three-dimensional bone block model and a three-dimensional high-risk soft tissue model, carrying out the meshing of the three-dimensional bone block model and the three-dimensional high-risk soft tissue model based on anatomical map data, and carrying out the risk assignment, a predictive static probability risk field network is obtained; in the operation process, 2D-3D registration and tracking are conducted through intraoperative imaging data, real-time displacement information of a key tracking point is obtained, then the static probability risk field network is dynamically updated, and an updated probability risk field network capable of reflecting current soft tissue real deformation is formed; and finally, based on the updated risk field and the current pose of the robot, self-adaptive adjustment of a local path is carried out, so that a safe and efficient fracture reduction path is accurately planned. In this way, the precision and safety of the fracture reduction surgery are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent planning, and more specifically, to a fracture reduction path planning system and method based on image data. BACKGROUND

[0002] Fracture reduction is a key operation in orthopedic surgery, which aims to restore the displaced fracture fragments to the normal anatomical position and alignment, create favorable conditions for bone healing, and maximize the recovery of patients' limb function. For a long time, fracture reduction has mainly relied on the experience, anatomical knowledge and skilled surgical techniques of surgeons. However, there are many challenges in manual reduction, such as the limited operation precision due to the stability of human hands, and the difficulty of reducing complex fractures or deep fractures. In order to improve the accuracy and safety of fracture reduction, robot-assisted fracture reduction and computer navigation technology have emerged, bringing revolutionary changes to orthopedic surgery and indicating a new direction of precision medicine.

[0003] Although computer-assisted and robotic technology has shown great potential in the field of fracture reduction, existing reduction path planning schemes still face many technical difficulties and challenges in practical application. Most current fracture reduction path planning is often based on preoperative CT or MRI image data to construct a three-dimensional model and perform virtual planning. However, the anatomical model relied on by such planning is usually static and rigid, while the soft tissues in the human body have complex biomechanical properties such as viscoelasticity, nonlinearity and anisotropy. When the instrument or bone fragment moves and contacts the soft tissue during the reduction process, the soft tissue will undergo unpredictable dynamic deformation, which makes the pre-planned path deviate easily, and even directly leads to collision with high-risk soft tissues (such as blood vessels and nerves), causing serious complications. This means that existing methods lack effective mechanisms to perceive and dynamically update the risks brought by soft tissue deformation in real time, and cannot adaptively adjust the reduction path according to the actual situation during surgery, making it difficult to cope with complex and uncertain environmental changes during the operation.

[0004] Therefore, an optimized fracture reduction path planning scheme based on image data is expected. SUMMARY

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a fracture reduction path planning system and method based on image data. It first constructs a three-dimensional bone block model and a three-dimensional high-risk soft tissue model by performing fine anatomical structure segmentation on high-resolution preoperative image data. These models are then meshed based on anatomical atlas data and risk-assigned to obtain a predictive static probabilistic risk field network. During the surgery, intraoperative imaging data is used for 2D-3D registration and tracking to obtain real-time displacement information of key tracking points. This dynamically updates the static probabilistic risk field network, forming an updated probabilistic risk field network that reflects the actual deformation of the soft tissue. Finally, based on this updated risk field and the robot's current pose, adaptive adjustments are made to the local path, thereby accurately planning a safe and efficient fracture reduction path. This approach significantly improves the accuracy and safety of fracture reduction surgery.

[0006] According to one aspect of this application, a fracture reduction path planning method based on imaging data is provided, comprising: Anatomical structure segmentation was performed on the acquired high-resolution preoperative CT data and high-resolution preoperative MRI data to obtain three-dimensional bone block models and three-dimensional high-risk soft tissue models. Based on anatomical atlas data, a three-dimensional high-risk soft tissue model is meshed and risk-assigned to obtain a static probability risk field network. Based on the current position and target pose of the bone block, key tracking points are defined and initial path planning is performed on the 3D bone block model and static probabilistic risk field network to obtain the initial reset path and key tracking point set; 2D-3D registration and tracking of intraoperative imaging data were performed to obtain the coordinates of the tracking point after displacement; Based on the coordinates of the tracking points after displacement and the set of key tracking points, the static probability risk field network is updated to obtain the updated probability risk field network. Local path adaptive adjustment is performed based on the updated probability risk field network and the robot's current pose to obtain the adjusted short reset path.

[0007] According to another aspect of this application, a fracture reduction path planning system based on imaging data is provided, comprising: The anatomical structure segmentation module is used to segment the acquired high-resolution preoperative CT data and high-resolution preoperative MRI data into anatomical structures to obtain three-dimensional bone block models and three-dimensional high-risk soft tissue models. The meshing and risk assignment module is used to mesh and assign risk values ​​to three-dimensional high-risk soft tissue models based on anatomical atlas data to obtain a static probability risk field network. The key tracking point definition and path planning module is used to define key tracking points and perform initial path planning on the 3D bone block model and static probabilistic risk field network based on the current position and target pose of the bone block to obtain the initial reset path and key tracking point set. The 2D-3D registration and tracking module is used to perform 2D-3D registration and tracking on intraoperative imaging data to obtain the coordinates of the tracking point after displacement. The risk field update module is used to update the static probability risk field network based on the coordinates of the tracking points after displacement and the set of key tracking points to obtain the updated probability risk field network. The local path adaptive adjustment module is used to perform local path adaptive adjustment based on the updated probability risk field network and the robot's current pose to obtain the adjusted short reset path.

[0008] Compared with existing technologies, this application provides a fracture reduction path planning system and method based on image data. It first constructs a three-dimensional bone block model and a three-dimensional high-risk soft tissue model by performing fine anatomical segmentation on high-resolution preoperative image data. These models are then meshed and risk-assigned based on anatomical atlas data to obtain a predictive static probabilistic risk field network. During the operation, intraoperative imaging data is used for 2D-3D registration and tracking to obtain real-time displacement information of key tracking points. This dynamically updates the static probabilistic risk field network, forming an updated probabilistic risk field network that reflects the actual deformation of the soft tissue. Finally, based on this updated risk field and the robot's current pose, adaptive adjustments are made to the local path, thereby accurately planning a safe and efficient fracture reduction path. This approach significantly improves the accuracy and safety of fracture reduction surgery. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a fracture reduction path planning method based on image data according to an embodiment of this application; Figure 2 This is a data flow diagram of the fracture reduction path planning method based on image data according to an embodiment of this application; Figure 3 This is a block diagram of a fracture reduction path planning system based on image data according to an embodiment of this application. Detailed Implementation

[0011] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0013] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0016] The technical solution of this application proposes a fracture reduction path planning method based on image data. Figure 1 This is a flowchart of a fracture reduction path planning method based on image data according to an embodiment of this application. Figure 2 This is a system architecture diagram of a fracture reduction path planning method based on image data according to an embodiment of this application. Figure 1 and Figure 2As shown, the fracture reduction path planning method based on image data according to an embodiment of this application includes the following steps: S1, performing anatomical structure segmentation on the acquired high-resolution preoperative CT data and high-resolution preoperative MRI data to obtain a three-dimensional bone block model and a three-dimensional high-risk soft tissue model; S2, based on anatomical atlas data, performing meshing and risk assignment on the three-dimensional high-risk soft tissue model to obtain a static probability risk field network; S3, based on the current position and target pose of the bone block, defining key tracking points and planning an initial path for the three-dimensional bone block model and the static probability risk field network to obtain an initial reduction path and a set of key tracking points; S4, performing 2D-3D registration and tracking on intraoperative imaging data to obtain the coordinates of the tracking points after displacement; S5, based on the coordinates of the tracking points after displacement and the set of key tracking points, updating the risk field of the static probability risk field network to obtain an updated probability risk field network; S6, performing local path adaptive adjustment based on the updated probability risk field network and the current pose of the robot to obtain an adjusted short-segment reduction path.

[0017] Specifically, in S1, the acquired high-resolution preoperative CT and MRI data are segmented into anatomical structures to obtain a three-dimensional bone block model and a three-dimensional high-risk soft tissue model. It should be understood that an accurate three-dimensional model enables surgeons and robots to clearly identify and differentiate fracture fragments and surrounding critical soft tissues (such as blood vessels and nerves), thereby allowing for precise reduction path planning and potential collision detection in a virtual environment. This is a crucial starting point for ensuring surgical safety and effectiveness, providing a patient-individualized, high-precision anatomical map for the entire reduction planning and surgical navigation.

[0018] High-resolution preoperative CT data refers to patient imaging data acquired before surgery via computed tomography (CT). CT images clearly display bone structures and have good spatial and density resolution. High-resolution preoperative MRI data refers to patient imaging data acquired before surgery via magnetic resonance imaging (MRI). MRI images are advantageous in displaying soft tissue structures and can clearly distinguish high-risk soft tissues such as muscles, blood vessels, and nerves. Anatomical segmentation refers to the process of separating different anatomical regions (such as bones, organs, and soft tissues) from the background in medical images. Its goal is to label each voxel (or pixel) with its specific anatomical structure category. A three-dimensional bone block model is a three-dimensional digital representation of the complete bone structure, including fractured bone fragments, obtained by three-dimensional reconstruction of the bone based on CT data segmentation results. This model can reflect the precise shape and spatial location of bone fragments. A three-dimensional high-risk soft tissue model is a three-dimensional digital representation of surrounding important soft tissues (such as major blood vessels, nerves, important ligaments, and organs) obtained by three-dimensional reconstruction based on MRI data (possibly combined with CT data) segmentation results. This model is used to identify and avoid potentially damaged sensitive areas in surgical planning.

[0019] In practice, firstly, using the high-resolution preoperative CT data as a reference, the high-resolution preoperative MRI data is aligned to the CT coordinate system to obtain registered high-resolution preoperative MRI data. That is, the data from two different imaging modalities (CT and MRI) are precisely aligned spatially to achieve information complementarity. Here, it should be understood that CT images have advantages in displaying bone structures with high resolution and high contrast, while MRI excels in distinguishing different soft tissues. To fully utilize the information from both and ensure that they are processed in the same three-dimensional coordinate system, the system uses the high-resolution preoperative CT data as a spatial reference and aligns the high-resolution preoperative MRI data with the CT data spatially using an image registration algorithm. This typically involves calculating a geometric transformation (such as a rigid transformation or affine transformation) so that points on the MRI image can be accurately mapped to their corresponding anatomical locations on the CT image. In this way, registered high-resolution preoperative MRI data can be obtained in the CT coordinate system, laying the foundation for subsequent multimodal image fusion and segmentation.

[0020] Next, semantic segmentation was performed on the high-resolution preoperative CT data and the registered high-resolution preoperative MRI data to obtain bone mask data and risk tissue mask data. That is, after spatial alignment of the image data, advanced image processing technology was used to identify and distinguish the bone structures to be operated on and the high-risk soft tissues to be avoided from complex medical images, providing a precise anatomical basis for subsequent 3D modeling of bone blocks, risk field construction, and path planning.

[0021] In practical implementation, taking high-resolution preoperative CT data as an example, the process for processing high-resolution preoperative CT data first involves inputting the high-resolution preoperative CT data into the visual encoder of a pre-trained 3D U-Net deep learning segmentation network to obtain preoperative CT voxel-level visual feature maps. During this process, the 3D U-Net visual encoder (i.e., the shrinking path) gradually captures and compresses contextual information from the input high-resolution CT voxel data through a series of downsampling and convolution operations, transforming the original grayscale value of each voxel and its surrounding spatial relationships into a preoperative CT voxel-level visual feature map containing rich semantics. The generated preoperative CT voxel-level visual feature map is a highly condensed representation of the original image, effectively capturing details such as bone texture and edges, laying the foundation for subsequent accurate segmentation. The 3D U-Net visual encoder is a convolutional neural network architecture, particularly suitable for pixel-level segmentation of 3D medical images, efficiently capturing contextual information and local details through its encoder-decoder structure and skip connections.

[0022] Next, the preoperative CT voxel-level visual feature map is visually salientized to obtain a salientized preoperative CT voxel-level visual feature map. Specifically, first, the preoperative CT voxel-level visual feature map is flattened to obtain a flattened vector of preoperative CT voxel-level visual features. Then, the flattened vector of preoperative CT voxel-level visual features is salientized to obtain a salientized flattened vector of preoperative CT voxel-level visual features. Finally, the salientized flattened vector of preoperative CT voxel-level visual features is reshaped to obtain a salientized preoperative CT voxel-level visual feature map.

[0023] Specifically, in the process of feature saliency enhancement of the preoperative CT voxel-level visual feature flattening vector, firstly, feature redundancy is removed from the preoperative CT voxel-level visual feature flattening vector to obtain the preoperative CT voxel-level visual redundancy-removed feature vector. It should be understood that although high-resolution CT scans can provide accurate bone structure information, due to limitations in equipment physical characteristics and individual patient differences, the original voxel-level features often contain a large number of anatomical details and imaging artifacts unrelated to fracture reduction. Traditional fixed-depth feature extraction networks struggle to adaptively distinguish clinically significant bone microstructural features from interfering information, which can lead to edge blurring or topological errors in subsequent 3D reconstruction models, thus affecting the safety assessment of reduction path planning. Therefore, in the technical solution of this application, a high-order representation with anatomical discriminative power is established through a dynamic feature redundancy removal mechanism. This process uses the KL divergence of the feature vectors before and after redundancy removal as a feedback signal. When the feature distribution change generated by continuous iteration is lower than a preset threshold, it indicates that the feature space has converged to a state that can stably represent key geometric properties such as cortical continuity, fracture line orientation, and trabecular spatial arrangement. This adaptive computational depth mechanism ensures that feature extraction for complex fracture types (such as comminuted fractures or intra-articular fractures) is targeted, avoiding underfitting or overfitting problems caused by fixed network depth, thereby significantly improving the topological fidelity of subsequent 3D bone block model reconstruction.

[0024] More specifically, in a specific example of this application, the preoperative CT voxel-level visual feature flattening vector is subjected to feature redundancy removal using the following formula to obtain the preoperative CT voxel-level visual redundancy removed feature vector: , , in, The vector representing the preoperative CT voxel-level visual features flattened out. For learnable weight matrix, For learnable bias vectors, For random disturbance coefficients, Introducing random noise to improve robustness for Activation function For the feature redundancy removal function, This is the preoperative CT voxel-level visual redundancy removal feature vector.

[0025] Next, the redundancy of the visual feature distribution relative to the preoperative CT voxel-level visual feature flattening vector is calculated. It should be understood that due to significant differences in voxel features of different anatomical structures in CT images—for example, the grayscale distribution in dense cortical bone and cancellous bone regions exhibits a clear gradient change—traditional fixed-number feature transformations are unable to adaptively capture this non-uniform distribution characteristic, leading to the loss of important edge details in subsequent 3D reconstruction models. Therefore, in the technical solution of this application, the degree of modification of the original feature distribution by the redundancy of the visual feature distribution is evaluated by calculating the redundancy of the visual feature distribution. By measuring the difference in statistical properties of the feature vectors before and after redundancy removal (such as KL divergence), it is determined whether the current feature representation has reached a stable state. In fracture reduction scenarios, this mechanism ensures sufficient redundancy removal of key geometric features around the fracture line while avoiding meaningless repetitive calculations in uniform cortical bone regions. When the feature distribution changes generated by continuous iteration tend to level off, it indicates that the network has fully extracted the key anatomical features affecting reduction path planning, such as the fracture surface geometry of the fracture fragments and important spatial information such as cortical bone continuity. This redundancy-based dynamic deredundancy mechanism significantly improves the targeting of fracture area feature extraction.

[0026] More specifically, in a specific example of this application, the feature distribution redundancy of the first qualified record encoding de-redundancy feature vector relative to the feature engineering vector of the first qualified record is calculated using the following formula: , in, This refers to the mutual information between the preoperative CT voxel-level visual redundancy removal feature vector and the preoperative CT voxel-level visual feature flattening vector. The information entropy of the flattened vector of preoperative CT voxel-level visual features is given. To ensure the first qualified record information guarantee, The first qualified record redundancy compression ratio, for Divergence, which divides the preoperative CT voxel-level visual redundancy removal feature vector into... Subvectors , This indicates that each group is calculated. Conditional distribution With marginal distribution Between divergence, To control the weight of redundant terms, This indicates the redundancy of the characteristic distribution.

[0027] Subsequently, in response to the redundancy of the visual feature distribution meeting the preset condition, the preoperative CT voxel-level visual de-redundancy feature vector is set as the significant preoperative CT voxel-level visual feature flattening vector. Because fracture areas exhibit complex spatial heterogeneity in CT images, including sharp gradient changes on the cortical fracture surface, disordered texture in the cancellous bone region, and discrete distribution of small bone fragments, traditional fixed-layer neural networks struggle to adaptively determine the sufficiency of feature extraction. Especially in handling comminuted fractures, excessive de-redundancy can lead to overly smoothed features and loss of key edge information, while premature termination fails to fully capture the spatial relationships between bone fragments. This dilemma necessitates the use of quantitative indicators to objectively evaluate the convergence state of feature optimization. Therefore, in the technical solution of this application, a redundancy threshold is used to achieve precise control of feature space optimization. During this process, when the difference between the de-redundancy feature vector generated by continuous iterations and the feature distribution of the previous state is lower than a preset threshold, it indicates that the network has fully extracted the anatomical features related to fracture reduction in the current CT data, including key information such as the geometric morphology of bone fragments, fracture surface topology, and bone density distribution. This dynamic termination mechanism based on redundancy thresholds significantly improves the clinical applicability of fracture feature extraction.

[0028] More specifically, in a specific example of this application, in response to the visual feature distribution redundancy meeting a preset condition, the preoperative CT voxel-level visual de-redundancy feature vector is set as the significant preoperative CT voxel-level visual feature flattening vector, expressed by the formula: , in, For target redundancy, For tolerance, It is the absolute value; If the condition is met, output The flattened vector represents the significant preoperative CT voxel-level visual features; otherwise, it will be... As a new preoperative CT voxel-level visual feature, a cyclic redundancy removal operation is performed. It should be understood that if the cyclic feature redundancy removal process lacks a clear termination condition, it will fall into infinite computation or use a fixed, suboptimal number of iterations, which violates the original intention of allocating computational resources on demand. Therefore, in response to the distance between the feature distribution redundancy and the target redundancy meeting a preset tolerance, a set operation is performed, providing an automated, data-driven exit for this dynamic refinement process. In this process, the feature distribution redundancy calculated in the previous step is used as a feedback signal, and by comparing it with a target redundancy representing the ideal convergence state, it is determined whether the feature representation has converged to the representational fixed point described in the corpus.

[0029] Subsequently, the saliency-enhanced preoperative CT voxel-level visual feature map is input into the decoder of a pre-trained 3D U-Net deep learning segmentation network to obtain the bone mask data. In this stage, the saliency-enhanced visual feature map is input into the 3D U-Net visual decoder (i.e., the extended path). During this process, the decoder gradually restores the abstract feature map to the same spatial resolution as the original CT image through upsampling and deconvolution operations, and fuses the detailed information passed from the encoder through skip connections, thereby accurately predicting whether each voxel belongs to the bone on a voxel-by-voxel basis. The bone mask data is a three-dimensional binary image where the value of each voxel indicates whether it is classified as a bone region, with bone regions labeled with a specific value (e.g., 1) and non-bone regions labeled with another value (e.g., 0).

[0030] Similarly, the specific process of semantic segmentation of registered high-resolution preoperative MRI data to obtain risk tissue mask data includes: inputting the registered high-resolution preoperative MRI data into the visual encoder of a pre-trained 3D U-Net deep learning segmentation network to obtain a preoperative MRI voxel-level visual feature map; performing visual feature saliency on the preoperative MRI voxel-level visual feature map to obtain a saliency preoperative MRI voxel-level visual feature map; and inputting the saliency preoperative MRI voxel-level visual feature map into the decoder of a pre-trained 3D U-Net deep learning segmentation network to obtain the risk tissue mask data. Here, risk tissue mask data refers to the binary image (or three-dimensional voxel data) obtained after semantic segmentation, where high-risk soft tissue regions are marked with a specific value (e.g., 1), and non-high-risk soft tissue regions are marked with another value (e.g., 0).

[0031] Subsequently, 3D reconstruction is performed on the bone mask data and risk tissue mask data to obtain 3D bone block models and 3D high-risk soft tissue models. It should be understood that the bone mask data and risk tissue mask data obtained through semantic segmentation are essentially discrete, voxel-based representations. To enable visualization, measurement, and interactive operations in 3D space, and to provide geometrically manipulable input for subsequent path planning algorithms, this technical solution transforms these mask data into continuous geometric surface models. Specifically, commonly used algorithms in the 3D reconstruction process include Marching Cubes, which can generate 3D surface models composed of triangular meshes based on the boundary information of the voxel masks. For bone masks, each individual fracture fragment can be identified and separated to generate a 3D bone block model; for risk tissue masks, a 3D high-risk soft tissue model is generated. These 3D models are precise digital representations of real anatomical structures, enabling surgeons and robotic systems to intuitively observe, analyze, and manipulate them in a virtual environment, thereby achieving high-precision preoperative planning and intraoperative navigation.

[0032] Specifically, S2, based on anatomical atlas data, involves meshing and risk-assigning a three-dimensional high-risk soft tissue model to obtain a static probabilistic risk field network. It should be understood that simply possessing a three-dimensional high-risk soft tissue model is insufficient to guide robot path planning. The robot needs a mechanism capable of accurately assessing the potential risk of any point or segment of the path within its workspace. Through meshing and risk assignment, the continuous geometric model can be discretized into a series of mesh points, and each mesh point or its region can be assigned a quantified risk value, thus forming a static probabilistic risk field network covering the surgical area. This quantification allows the risk to be understood and optimized by the algorithm. Furthermore, soft tissue models obtained solely from segmenting individual patient MRI data may be incomplete or noisy. Anatomical atlas data contains a wealth of demographic average anatomical information, anatomical variability, and common surgical risk zone definitions. By combining this prior, universally accepted anatomical knowledge with patient-specific imaging data, a more robust and comprehensive risk assessment system can be constructed, compensating for the potential deficiencies of individual patient data and ensuring the accuracy and reliability of risk assessment.

[0033] Anatomical atlas data is a standardized, predefined dataset of anatomical structures, typically containing information on the average morphology, positional relationships, and variability of normal human bones, organs, blood vessels, nerves, and other structures. This data can exist as statistical shape models, probabilistic maps, or regions of interest (ROIs) for specific anatomical areas, providing prior knowledge for understanding and locating individual anatomical structures. Meshing refers to dividing a continuous three-dimensional space into a series of discrete, regular or irregular units (e.g., voxels or grid points) and storing data on these units. In this context, it maps information from high-risk soft tissue models onto a unified base grid; risk assignment refers to assigning a quantified risk value to each unit or point in the grid space. This value represents the probability or severity of biological tissue damage that may be encountered when performing operations at that spatial location.

[0034] In practical implementation, firstly, a standardized spatial mesh is defined for the three-dimensional high-risk soft tissue model to obtain the base mesh. The base mesh refers to the discretized spatial structure obtained by standardizing the spatial definition of the three-dimensional high-risk soft tissue model; it serves as the computational and storage carrier for the entire risk field network. It is typically a uniform voxel mesh. In the technical solution of this application, a unified spatial framework for carrying risk information is constructed by standardizing the spatial mesh definition of the three-dimensional high-risk soft tissue model. In this process, firstly, the extent of the surgical area is determined; then, within this three-dimensional spatial range, a uniform voxel mesh is defined as the base mesh, with each voxel representing a discrete unit in space. The mesh density (i.e., the size of the voxels) depends on the required accuracy of the risk assessment and computational resources. A typical implementation involves defining the region of interest using an axis-aligned bounding box and then uniformly dividing it into... Voxels, each voxel is the size of .

[0035] Next, based on the three-dimensional bone block model, the anatomical atlas data is registered to obtain registered atlas data. Atlas registration refers to the process of aligning general anatomical atlas data spatially with the patient's specific image data (here, the three-dimensional bone block model is used as a reference) through geometric transformation. This allows the prior anatomical information in the atlas to be accurately applied to the individual patient. In the technical solution of this application, atlas registration is performed on the anatomical atlas data to align general anatomical knowledge with the patient's own anatomical structure. In this process, because bone structures are relatively stable in the human body and clearly displayed in CT images, they are often chosen as the reference for registration. Specifically, the atlas registration process generally includes the following steps: First, identifying the corresponding reference bone structures in the anatomical atlas from the patient's 3D bone block model, such as the femur and tibia; next, establishing a correspondence between the patient's bone block model and the atlas bone model by manually marking key anatomical landmarks or using an automatic matching algorithm based on image features (such as local shape descriptors); then, employing a non-rigid registration algorithm, which optimizes a deformation field to ensure that the bones in the atlas match the patient's bones with minimal deformation. The deformation field is typically represented by a series of basis functions or elastic deformation based on a physical model. The optimization objective is to minimize the distance or similarity metric between the deformed bones in the atlas and the patient's bones. In this way, all information about soft tissues, blood vessels, nerves, etc., in the anatomical atlas data changes accordingly with the deformation of the bones, resulting in registered atlas data that is spatially consistent with the patient's specific anatomical structures.

[0036] Subsequently, the 3D high-risk soft tissue model is discretized into the base mesh, and a Euclidean distance transformation (EDT) is performed on the base mesh to obtain a distance field. The distance field, obtained after the Euclidean distance transformation of the base mesh, stores the minimum distance from each mesh point to the surface of the high-risk soft tissue model. The distance field is the basis for generating a deterministic risk field. In this process, firstly, the 3D high-risk soft tissue model is projected or mapped onto a predefined base mesh. For each voxel in the mesh, it is determined whether it is occupied by the soft tissue model. Occupied voxels are marked as 1, and unoccupied voxels are marked as 0. Then, an Euclidean distance transformation (EDT) is performed on this binary voxel mesh. The EDT algorithm calculates the Euclidean distance from each "0" voxel in the mesh to the nearest "1" voxel (i.e., the high-risk soft tissue surface). For the "1" voxel itself, the distance is typically 0.

[0037] Then, a risk function is applied to each distance value in the distance field to obtain a high-determinism risk field. The risk function maps the distance values ​​in the distance field to their corresponding risk values. The high-determinism risk field is a risk field directly calculated based on the individual patient's high-risk soft tissue model and Euclidean distance transformation. That is, in the technical solution of this application, the distance values ​​in the distance field are mapped to a risk value (typically between 0 and 1). Specifically, when the distance value is close to or less than 0 (within or on the surface of high-risk soft tissue), the risk value should be close to or reach its maximum value (e.g., 1); as the distance value increases, the risk value should decrease rapidly until a certain safe distance is reached, after which the risk value drops to 0 or very close to 0. In one example, common risk function forms include step functions, linear decay functions, and exponential decay functions. By substituting each distance value in the distance field into the risk function, a high-determinism risk field is obtained, which accurately reflects the risk distribution based on the patient's own anatomical data.

[0038] Furthermore, a priori risk field is generated based on the registered atlas data and the base grid. This priori risk field is generated from the registered anatomical atlas data and the base grid. This risk field reflects a pre-defined anatomical risk distribution based on collective experience and is typically used to supplement or modify a high-determinism risk field. In this process, firstly, it should be understood that anatomical atlases usually contain functional areas of different anatomical structures, common pathways of neurovascular bundles, and priori high-risk areas that are easily damaged in surgeries such as fracture reduction. This information can be stored in the atlas as a probability map, representing the probability of finding a specific risk structure at a certain location. Next, this priori risk information from the registered atlas data is mapped onto the previously defined base grid. If the atlas exists in the form of a probability map, the corresponding probability values ​​are directly assigned to the grid points; if the atlas only defines the boundaries of high-risk areas, a Euclidean distance-like transformation can be performed on these areas and a risk function can be applied, but the input is the risk area defined by the atlas, not the patient's own soft tissue model. Finally, a priori risk field is generated, reflecting a risk distribution based on general anatomical experience.

[0039] Finally, the prior risk field and the high-determinism risk field are fused using a multi-source risk field fusion method to obtain the static probabilistic risk field network. The high-determinism risk field is based on actual patient images and has high detail; the prior risk field provides general knowledge to compensate for potential deficiencies or noise in individual data. In the technical solution of this application, multi-source risk field fusion of the prior risk field and the high-determinism risk field is used to integrate risk information from different sources to form a final, comprehensive risk field. Specifically, the static probabilistic risk field network can be obtained by weighted summation of the two risk fields. It is worth noting that the weights can be set based on experience or learned through machine learning methods; typically, the patient-specific high-determinism risk field is given higher weights. The generated static probabilistic risk field network is a pre-calculated, probabilistic risk distribution map in three-dimensional space, which associates each spatial location within the surgical area with a risk probability, representing the likelihood of soft tissue damage when surgical instruments are used at that location.

[0040] Specifically, in step S3, based on the current position and target pose of the bone fragments, key tracking points are defined and initial path planning is performed on the 3D bone fragment model and the static probabilistic risk field network to obtain the initial reduction path and a set of key tracking points. The current position and target pose of the bone fragments refer to the actual 3D coordinates and rotational orientation of each fractured bone fragment in 3D space after the fracture occurs (current position), and the ideal, anatomically normal final 3D coordinates and rotational orientation that these bone fragments should achieve after preoperative planning (target pose). The current position reflects the actual anatomical state after the fracture, while the target pose is the ideal shape determined through preoperative planning that conforms to normal anatomical reduction. By obtaining the current position and target pose of the bone fragments, a clear starting and ending point is provided for subsequent path planning. Furthermore, it should be understood that in robot-assisted fracture reduction scenarios, the set of key tracking points will serve as a benchmark for intraoperative image tracking and pose estimation, enabling the robot to perceive the movement of the bone fragments in real time and correlate it with the preset initial reduction path to achieve precise navigation and control. Key tracking points are geometric points with specific significance selected on the 3D bone fragment model. These are feature points that are easily identifiable and tracked during surgery, and their coordinates change synchronously with the movement of the bone fragment. By tracking the spatial position of these points in real time, the current pose of the entire corresponding bone fragment can be determined. The initial reduction path is a three-dimensional motion trajectory planned by an algorithm from the current (fracture) pose of the bone fragment to the target (reduction) pose. This path not only considers the geometric movement of the bone fragment but also integrates the requirements for safety and efficiency through a cost function, serving as the initial navigation route for the surgical robot to perform the reduction operation. Before the actual surgery, planning the initial reduction path based on a static probabilistic risk field network can assess and avoid potential surgical risks in advance. This risk field includes the spatial distribution and risk level of high-risk soft tissues within the surgical area, enabling the planned path to avoid critical structures such as nerves and blood vessels to the greatest extent possible while ensuring successful reduction of the bone fragment, thereby improving the safety of the surgery. In addition, optimizing the path length and risk cost can also improve the overall efficiency of the reduction process.

[0041] In specific implementation, firstly, tracking reference points are selected on the three-dimensional bone block model to obtain the key tracking point set. In the technical solution of this application, the selection of tracking reference points is to enable the real-time determination of the position and orientation of the bone block by imaging equipment (such as a C-arm X-ray machine, optical tracking system, or electromagnetic tracking system) during actual surgery. These points must have good identifiability and stability, and be able to accurately reflect the six degrees of freedom pose of the bone block. In this process, firstly, several regions with distinctive features, such as bony prominences, bony ridges, specific bone foramina, or geometric feature points uniformly distributed on the surface of the bone block, can be manually or automatically identified on the three-dimensional bone block model. If markers are pre-implanted, the centers of these markers can be directly used as tracking points; then, in order to accurately calculate the six degrees of freedom pose of the bone block, at least three non-collinear points are usually required. To improve robustness and accuracy, more points are often selected, and it is ensured that they have sufficient space distribution on the surface of the bone block (e.g., not on the same plane or the same axis). These points should not be obstructed during robot operation and should be clearly visible under X-ray (if X-ray tracking is used). Subsequently, shape feature analysis algorithms on the 3D model, such as Gaussian curvature analysis, principal component analysis (PCA), or feature point detection algorithms combined with geometric context information (such as 3D SIFT, Spin Images, etc.), are used to automatically identify highly recognizable and stable feature points under different viewpoints as candidate tracking reference points. From these candidate points, points that constitute a good tracking point set are selected. Through this process, a key tracking point set containing several points is finally obtained.

[0042] Subsequently, using the path length and the risk integral of the static probability risk field network along the path as the cost function, based on A The algorithm plans the initial reset path from the current position and target pose of the bone block. Wherein, A The algorithm is a popular and efficient pathfinding algorithm. It finds the shortest path from the starting point to the destination in a graph structure. Its unique feature is that it combines the actual cost of the currently traversed path with a heuristically estimated cost from the current point to the destination, guiding the search direction by minimizing the sum of the two. This characteristic makes it excellent in fields such as medical route planning, enabling it to find a near-optimal path while maintaining efficiency. In this process, firstly, A... The algorithm needs to search within a discretized state space. For the motion of the bone block, the state space is the six-degree-of-freedom pose of the bone block in three-dimensional space. This includes three translational degrees of freedom and three rotational degrees of freedom. This six-dimensional space needs to be discretized into a series of feasible nodes, each node representing a specific pose of the bone block, with adjacent nodes connected by small translations and / or rotations; where the starting node is the current position and pose of the bone block, and the target node is the target pose of the bone block; then, for A... Each step in the algorithm (i.e., moving from one pose node to another) requires calculating its cost, where the cost function integrates two key factors: path length and risk integral. Path length measures the geometric "distance" from the current pose to the next pose, and can include a weighted sum of translation distance and rotation angle; the risk integral measures the potential risk of the area traversed by the bone block or its key tracking points during the movement from the current pose to the next pose, which is obtained by querying and integrating a static probabilistic risk field network. The final cost function is a weighted sum of these two factors. During execution, A... The algorithm starts with the initial node and maintains an open list (nodes to be expanded) and a closed list (nodes that have been expanded). For each node n in the open list, its priority function is calculated, and the specific formula is as follows: , in, This represents the actual minimum total cost from the starting node to the current node n. This represents the heuristic cost estimate from the current node n to the target node; each time, a node is selected from the open list. The node with the smallest value is expanded by adding its adjacent valid pose nodes to the open list and updating their g values. This process continues iteratively until the target node is found, or the open list is empty (indicating no solution). Finally, A is... The algorithm finds a series of intermediate pose nodes from the current pose of the bone block to the target pose and connects them to obtain the initial reset path.

[0043] Specifically, in step S4, 2D-3D registration and tracking of intraoperative imaging data are performed to obtain the coordinates of the tracking points after displacement. It should be understood that the three-dimensional bone block model, initial reduction path, and related risk assessment generated during preoperative planning are all based on a virtual patient model. However, in the operating room, the surgeon and robot face a real, dynamically changing patient. Therefore, in the technical solution of this application, the actual spatial pose of the bone block is obtained in real time through intraoperative imaging data, thereby accurately mapping the planning in the virtual world to the real world, ensuring the accuracy of the operation. Furthermore, fracture reduction is a dynamic process, and the bone block will continuously move during the operation. To enable the robot to accurately guide the bone block along the planned path, or to adjust the path when necessary, the system can continuously update the coordinates of the tracking points after displacement of the bone block through 2D-3D registration and tracking, thereby calculating the complete three-dimensional pose of the bone block and providing data input for the robot's precise navigation and feedback control.

[0044] Intraoperative imaging data refers to medical images acquired in real time during surgery. Common examples include two-dimensional fluoroscopic images acquired by a C-arm X-ray machine and intraoperative CT scans. 2D-3D registration is a medical image processing technique that aims to determine the precise three-dimensional pose of a three-dimensional model at the time of two-dimensional image acquisition by spatially aligning two-dimensional images (such as intraoperative X-rays) with three-dimensional models (such as bone models reconstructed from preoperative CT scans).

[0045] In practice, the coordinates of the tracked points after displacement can be obtained by projecting the set of key tracking points onto the intraoperative imaging data. Specifically, before performing 2D-3D registration, precise geometric calibration of the intraoperative imaging equipment (e.g., a C-arm X-ray machine) is required. This is to determine its internal parameters (such as focal length, principal point, pixel size, etc.) and external parameters (the pose of the camera coordinate system relative to the operating room or patient reference coordinate system), which are crucial for 3D-to-2D projection. During the operation, equipment such as a C-arm X-ray machine acquires two-dimensional projected images of the fracture area in real time or periodically. To improve the robustness of registration, multiple X-ray images are usually acquired from different angles. In the acquired intraoperative two-dimensional images, two-dimensional feature points corresponding to the set of key tracking points on the pre-operated three-dimensional bone block model are identified and extracted. These feature points can be pre-implanted radiopaque markers (such as small steel balls) or obvious anatomical landmarks formed by the bone itself on X-ray images (such as the edges of bony prominences, cortical bone lines, etc.). Then, using calibrated camera parameters, a pre-manipulated 3D bone block model, and feature points identified in intraoperative 2D images, the precise 3D pose of the 3D bone block model at the current moment is calculated. After determining the current 3D pose of the bone block, the set of key tracking points can be transformed to the common reference coordinate system of the operating room by applying this 3D pose transformation matrix, thereby obtaining their actual 3D coordinates at the current moment, i.e., the coordinates of the tracking points after displacement. It is worth mentioning that this process is continuous; each time new intraoperative imaging data is acquired, registration and calculation are performed again, thus achieving real-time tracking of the bone block and its key tracking points.

[0046] Specifically, in step S5, based on the coordinates of the tracking points after displacement and the set of key tracking points, the static probability risk field network is updated to obtain an updated probability risk field network. It should be understood that the surgical area is not entirely rigid; movement of bone fragments, pressure from surgical instruments, changes in muscle tension, and even gravity can cause deformation of surrounding soft tissues, making the pre-constructed static probability risk field network no longer accurately reflect the current true risk distribution of the organism. Therefore, to ensure the continuous safety and accuracy of the surgical procedure, in the technical solution of this application, the static probability risk field network is updated based on the coordinates of the tracking points after displacement and the set of key tracking points. By updating the risk field in real time, the system can dynamically perceive the latest risk distribution of the surgical area, enabling risk assessment and avoidance capabilities to adaptively adjust with the movement of bone fragments and changes in tissue.

[0047] In practice, firstly, the coordinates of the displaced tracking points and the set of key tracking points are input into a rapid deformation model to obtain a three-dimensional displacement vector field of the surgical area. The rapid deformation model is a mathematical or computational model used to simulate and predict how biological tissue deforms under external forces (such as bone movement or instrument compression). These models are typically based on physical principles (such as elasticity) or data-driven learning methods. The three-dimensional displacement vector field of the surgical area is a three-dimensional spatial vector field, in which each voxel (or grid point) within the surgical area is associated with a three-dimensional displacement vector. This vector represents the predicted displacement direction and magnitude of the voxel (or grid point) relative to its original position.

[0048] It is worth mentioning that physics-based deformation models include finite element models and mass-spring models. The finite element model discretizes the soft tissue region into a series of small elements, each with specific physical properties such as elasticity and viscosity. The displacement of the bone block serves as a boundary condition or a force applied to it, and the model solves the equations of elasticity (such as the Navier-Stokes equations or linear elasticity equations) to calculate the deformation of the entire voxel mesh. The mass-spring model discretizes the soft tissue into a network of mass points connected by springs. The movement of the bone block causes nearby springs to be compressed or stretched, resulting in force transmission and deformation of the entire network. Data-driven deformation models learn the mapping relationship between bone block movement and soft tissue deformation from large amounts of tissue deformation data through machine learning (such as neural networks). These include regression models and conditional generative adversarial networks (GANs). The regression model constructs a regression model whose input is the pose change of the bone block, and whose output is the displacement vector field of the entire region. The conditional generative adversarial network generates the corresponding soft tissue displacement field by taking the initial and displacement pose of the bone block as input.

[0049] Subsequently, the three-dimensional displacement vector field of the surgical area is applied to the static probabilistic risk field network to obtain the updated probabilistic risk field network. In a specific example, the original risk field can be resampled or interpolated based on the predicted soft tissue deformation (three-dimensional displacement vector field of the surgical area) to generate a risk field reflecting the current state.

[0050] Specifically, in step S6, a local path adaptive adjustment is performed based on the updated probabilistic risk field network and the robot's current pose to obtain an adjusted short-segment reduction path. It should be understood that in the entire chain of fracture reduction path planning, previous steps have already constructed the patient's anatomical model, established a dynamically changing risk assessment network, and planned an initial reduction path from the fracture pose to the target pose. However, the real surgical environment is full of variables, such as the possibility of further soft tissue deformation during surgery, or individual differences that cannot be fully predicted preoperatively. To address these uncertainties and ensure that the robot can continuously perform reduction operations safely and accurately, the technical solution of this application performs a local path adaptive adjustment based on the updated probabilistic risk field network and the robot's current pose to obtain an adjusted short-segment reduction path, ensuring real-time safety and flexibility throughout the entire surgical process.

[0051] In practice, the first step is to check the risk integral of the next path segment the robot is about to execute in the updated probabilistic risk field network. It should be understood that the system continuously and proactively assesses the safety of the next short path segment to be executed during robot movement. This path is typically a fixed-length segment in the initial reset path, extending from the robot's current pose towards the target direction. During this process, the robot performs dense sampling along this path segment, obtaining a series of intermediate poses for the bone blocks. For each sampling point on the path, the bone block model is placed in that pose. To more accurately assess the risk, the volume swept by the bone block as it moves from one sampling point to the next can be calculated. For the volume within these sampling poses, or for the volume swept by the bone block, grid points in the updated probabilistic risk field network are queried to obtain the risk values ​​for these points; all queried risk values ​​are then integrated or weighted and accumulated. For example, the average risk value of all sampling points on the path and several surrounding voxels can be calculated, or the risk of the bone block swept within the path segment can be integrally determined.

[0052] Next, in response to the risk score being lower than a preset threshold, the robot moves along the next path segment to be executed; in response to the risk score exceeding or equal to the preset threshold, a fast optimization algorithm is used in the updated probabilistic risk field network, starting from the robot's current pose, to search for the adjusted short reset path segment. That is, if the calculated risk score is lower than a preset safety threshold, it indicates that the short path segment to be executed is safe and will not conflict with critical high-risk soft tissues. In this case, the robot will continue moving according to the short planned path segment corresponding to the initial reset path. Once the risk score detects a potential risk (exceeding the threshold), the system immediately stops executing the current short initial path segment and, starting from the robot's current pose, defines a sufficiently small search space centered on the robot's current pose. Subsequently, within this local search space, the system calls a fast optimization algorithm to replan an alternative path from the robot's current pose to the end point of the original path segment. The goal of this algorithm is to find a new adjusted short reset path segment that has the lowest possible risk score in the updated probabilistic risk field network, the shortest possible path length, and a smooth connection back to the initial reset path. The cost function remains a weighted sum of path length and risk integral, but the weights may be biased towards risk avoidance. In one example, this can be achieved by calculating the gradient of the risk field near the current bone pose, taking small steps along the direction of fastest risk reduction, and simultaneously optimizing the path length.

[0053] Finally, the robot is controlled to move along the adjusted short reset path. That is, after obtaining the adjusted short reset path, the system generates corresponding robot motion commands (e.g., joint angle sequences or end effector trajector trajectories) and sends them to the robot controller, driving the robot to move smoothly along this corrected path. After this corrected path is completed, the robot will reassess its current state and attempt to rejoin the subsequent segments of the initial reset path, or make further local adjustments, until the entire reset process is successfully completed or manually interrupted.

[0054] In summary, the fracture reduction path planning method based on image data according to the embodiments of this application is explained. It first constructs a three-dimensional bone block model and a three-dimensional high-risk soft tissue model by performing fine anatomical structure segmentation on high-resolution preoperative image data. These models are then meshed based on anatomical atlas data and risk-assigned to obtain a predictive static probabilistic risk field network. During the operation, intraoperative imaging data is used for 2D-3D registration and tracking to obtain real-time displacement information of key tracking points. This dynamically updates the static probabilistic risk field network, forming an updated probabilistic risk field network that reflects the actual deformation of the soft tissue. Finally, based on this updated risk field and the robot's current pose, adaptive adjustments are made to the local path, thereby accurately planning a safe and efficient fracture reduction path. This approach significantly improves the accuracy and safety of fracture reduction surgery.

[0055] Furthermore, a fracture reduction path planning system based on image data is also provided.

[0056] Figure 3 This is a block diagram of a fracture reduction path planning system based on image data according to an embodiment of this application. Figure 3 As shown, the wind and solar power generation and energy storage management system 300 according to an embodiment of this application includes: an anatomical structure segmentation module 310, used to segment the acquired high-resolution preoperative CT data and high-resolution preoperative MRI data into a three-dimensional bone block model and a three-dimensional high-risk soft tissue model; a meshing and risk assignment module 320, used to mesh and assign risk values ​​to the three-dimensional high-risk soft tissue model based on anatomical atlas data to obtain a static probability risk field network; and a key tracking point definition and path planning module 330, used to define the three-dimensional bone block model and the static probability risk field network based on the current position and target pose of the bone block. The network defines key tracking points and plans the initial path to obtain the initial reset path and key tracking point set; the 2D-3D registration and tracking module 340 is used to perform 2D-3D registration and tracking on the intraoperative imaging data to obtain the coordinates of the tracking points after displacement; the risk field update module 350 is used to update the static probability risk field network based on the coordinates of the tracking points after displacement and the key tracking point set to obtain the updated probability risk field network; the local path adaptive adjustment module 360 ​​is used to perform local path adaptive adjustment based on the updated probability risk field network and the robot's current pose to obtain the adjusted short reset path.

[0057] As described above, the fracture reduction path planning system 300 based on image data according to the embodiments of this application can be implemented in various wireless terminals, such as servers with fracture reduction path planning algorithms based on image data. In one possible implementation, the fracture reduction path planning system 300 based on image data according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the fracture reduction path planning system 300 based on image data can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the fracture reduction path planning system 300 based on image data can also be one of many hardware modules of the wireless terminal.

[0058] Alternatively, in another example, the image-based fracture reduction path planning system 300 and the wireless terminal can also be separate devices, and the image-based fracture reduction path planning system 300 can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0059] The various embodiments of this disclosure 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. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A fracture reduction path planning method based on image data, characterized in that, include: Anatomical structure segmentation was performed on the acquired high-resolution preoperative CT data and high-resolution preoperative MRI data to obtain three-dimensional bone block models and three-dimensional high-risk soft tissue models. Based on anatomical atlas data, a three-dimensional high-risk soft tissue model is meshed and risk-assigned to obtain a static probability risk field network. Based on the current position and target pose of the bone block, key tracking points are defined and initial path planning is performed on the 3D bone block model and static probabilistic risk field network to obtain the initial reset path and key tracking point set; 2D-3D registration and tracking of intraoperative imaging data were performed to obtain the coordinates of the tracking point after displacement; Based on the coordinates of the tracking points after displacement and the set of key tracking points, the static probability risk field network is updated to obtain the updated probability risk field network. Local path adaptive adjustment is performed based on the updated probability risk field network and the robot's current pose to obtain the adjusted short reset path.

2. The fracture reduction path planning method based on image data according to claim 1, characterized in that, Anatomical segmentation was performed on the acquired high-resolution preoperative CT and MRI data to obtain three-dimensional bone block models and three-dimensional high-risk soft tissue models, including: Using the high-resolution preoperative CT data as a reference, the high-resolution preoperative MRI data is aligned to the CT coordinate system to obtain the registered high-resolution preoperative MRI data. Semantic segmentation was performed on high-resolution preoperative CT data and registered high-resolution preoperative MRI data to obtain bone mask data and risk tissue mask data. Three-dimensional reconstruction was performed on the bone mask data and the risk tissue mask data to obtain a three-dimensional bone block model and a three-dimensional high-risk soft tissue model.

3. The fracture reduction path planning method based on image data according to claim 3, characterized in that, Semantic segmentation was performed on high-resolution preoperative CT data and registered high-resolution preoperative MRI data to obtain bone mask data and risk tissue mask data, including: The high-resolution preoperative CT data is input into the visual encoder of a pre-trained 3D U-Net deep learning segmentation network to obtain a voxel-level visual feature map of the preoperative CT. Visual feature saliency is performed on the preoperative CT voxel-level visual feature map to obtain a saliency preoperative CT voxel-level visual feature map. The saliency-enhanced preoperative CT voxel-level visual feature map is input into the decoder of a pre-trained 3D U-Net deep learning segmentation network to obtain the bone mask data.

4. The fracture reduction path planning method based on image data according to claim 3, characterized in that, Visual feature saliency is performed on the preoperative CT voxel-level visual feature map to obtain a saliency preoperative CT voxel-level visual feature map, including: The preoperative CT voxel-level visual feature map is flattened to obtain the preoperative CT voxel-level visual feature flattening vector. The preoperative CT voxel-level visual feature flattening vector is saliency-enhanced to obtain significant preoperative CT voxel-level visual feature flattening vector; The feature shape of the flattened vectors of significant preoperative CT voxel-level visual features is reshaped to obtain significant preoperative CT voxel-level visual feature maps.

5. The fracture reduction path planning method based on image data according to claim 1, characterized in that, Based on anatomical atlas data, a three-dimensional high-risk soft tissue model is meshed and risk-assigned to obtain a static probabilistic risk field network, including: A standardized spatial mesh definition is performed on a three-dimensional high-risk soft tissue model to obtain the basic mesh; Based on the three-dimensional bone block model, the anatomical atlas data is registered to obtain the registered atlas data. The three-dimensional high-risk soft tissue model is discretized into the base mesh, and the base mesh is subjected to Euclidean distance transformation to obtain the distance field; Apply a risk function to each distance value in the distance field to obtain a highly deterministic risk field; Based on the registered map data and the basic grid, a priori risk field is generated. The prior risk field and the high-determinism risk field are fused using a multi-source risk field fusion to obtain the static probability risk field network.

6. The fracture reduction path planning method based on image data according to claim 1, characterized in that, Based on the current position and target pose of the bone block, key tracking points are defined and initial path planning is performed on the 3D bone block model and static probabilistic risk field network to obtain the initial reset path and key tracking point set, including: The key tracking point set is obtained by selecting tracking reference points from the three-dimensional bone block model; Using the path length and the risk integral of the static probability risk field network along the path as the cost function, based on A The algorithm plans the initial reset path from the current position and target pose of the bone block.

7. The fracture reduction path planning method based on image data according to claim 1, characterized in that, Performing 2D-3D registration and tracking on intraoperative imaging data to obtain the coordinates of the tracking points after displacement includes: projecting the key tracking point set onto the intraoperative imaging data to obtain the coordinates of the tracking points after displacement.

8. The fracture reduction path planning method based on image data according to claim 1, characterized in that, Based on the displacement of the tracking point coordinates and the key tracking point set, the static probability risk field network is updated to obtain an updated probability risk field network, including: Input the coordinates of the displaced tracking points and the set of key tracking points into the fast deformation model to obtain the three-dimensional displacement vector field of the surgical area; The three-dimensional displacement vector field of the surgical area is applied to the static probability risk field network to obtain the updated probability risk field network.

9. The fracture reduction path planning method based on image data according to claim 1, characterized in that, Local path adaptive adjustment is performed based on the updated probabilistic risk field network and the robot's current pose to obtain an adjusted short reset path, including: Check the risk integral of the next path that the robot is about to execute in the updated probability risk field network; In response to the risk score falling below a preset threshold, the movement proceeds along the next segment of the path to be executed. In response to the risk score exceeding or equaling a preset threshold, a fast optimization algorithm is used to search for the adjusted short reset path under the updated probability risk field network, starting from the robot's current pose. The robot is controlled to move along the adjusted short reset path.

10. A fracture reduction path planning system based on image data, characterized in that, include: The anatomical structure segmentation module is used to segment the acquired high-resolution preoperative CT data and high-resolution preoperative MRI data into anatomical structures to obtain three-dimensional bone block models and three-dimensional high-risk soft tissue models. The meshing and risk assignment module is used to mesh and assign risk values ​​to three-dimensional high-risk soft tissue models based on anatomical atlas data to obtain a static probability risk field network. The key tracking point definition and path planning module is used to define key tracking points and perform initial path planning on the 3D bone block model and static probabilistic risk field network based on the current position and target pose of the bone block to obtain the initial reset path and key tracking point set. The 2D-3D registration and tracking module is used to perform 2D-3D registration and tracking on intraoperative imaging data to obtain the coordinates of the tracking point after displacement. The risk field update module is used to update the static probability risk field network based on the coordinates of the tracking points after displacement and the set of key tracking points to obtain the updated probability risk field network. The local path adaptive adjustment module is used to perform local path adaptive adjustment based on the updated probability risk field network and the robot's current pose to obtain the adjusted short reset path.

Images