Spatial registration pedicle screw navigation method and system based on point cloud structure analysis

Abstract

This invention proposes a spatial registration method and system for pedicle screw navigation based on point cloud structure analysis. First, a reference point cloud dataset for surgical planning and a raw point cloud dataset for intraoperative positioning are acquired. The reference point cloud dataset is obtained from 3D reconstruction of human spinal medical images, while the raw point cloud dataset is obtained by scanning the surgical exposure area using a structured light scanning device or an optical positioning system. Anatomical structure analysis based on the fusion of spatial coordinates and semantic features is performed on the reference point cloud dataset to identify anatomical units at the single vertebral level and construct corresponding spatial structural constraints, resulting in a subset of reference point clouds with anatomical structure information. Combined with the anatomical structure constraints, geometric feature-enhanced point cloud registration is performed to obtain the registered point cloud dataset. Finally, the registration results are integrated into a graphical user interface to assist in displaying spinal spatial pose and screw planning information, providing non-radioactive, high-precision navigation support for pedicle screw placement surgery.

Description

Technical Field

[0001] This invention relates to technologies related to spinal surgery navigation, specifically belonging to the technical field of combining medical image processing, three-dimensional point cloud analysis and intelligent computing, and particularly to a pedicle screw placement surgical navigation system and method for low-overlap point cloud scenarios based on spinal anatomical structure analysis. Background Technology

[0002] Pedicle screw placement is a widely used internal fixation procedure in spinal surgery, commonly used to treat spinal degeneration, trauma, and deformities. Because the pedicles are adjacent to the spinal cord, nerve roots, and important blood vessels, even minor deviations in screw placement can lead to serious complications such as nerve and vascular injury. Therefore, accurately obtaining the spatial orientation of the spine and achieving reliable navigation during the procedure is a key technical issue for improving surgical safety and precision.

[0003] To improve surgical navigation accuracy, existing pedicle screw navigation technologies typically rely on the spatial registration relationship between preoperative medical images and intraoperative sensory data. However, in actual surgical scenarios, there are often discrepancies between the anatomical state reflected in preoperative medical images and the actual exposed state of the spine during surgery. Furthermore, due to limitations in the surgical incision area and soft tissue obstruction, the spinal surface information obtainable intraoperatively is usually limited to a local area. This results in a limited effective overlap between preoperative and intraoperative data, posing significant challenges to the stability and accuracy of the registration process.

[0004] Currently, spinal surgery navigation technologies mainly include those based on medical imaging, visual perception, ultrasound imaging, and augmented reality. Among these, medical imaging-based navigation methods typically rely on intraoperative or repeated image acquisition, posing a risk of radiation exposure. Visual navigation methods based on visible light cameras are limited by imaging conditions and environmental factors, making it difficult to guarantee the accuracy and stability of their three-dimensional spatial positioning. While ultrasound imaging-based navigation methods offer advantages such as being radiation-free and real-time, they are limited by imaging resolution and signal-to-noise ratio, making it difficult to obtain stable and reliable registration results under complex spinal anatomy. Augmented reality-based navigation technologies, although providing intuitive surgical guidance, suffer from complex system integration and still have shortcomings in registration accuracy and display stability.

[0005] Compared to the aforementioned technical approaches, structured light-based 3D sensing technology offers advantages such as non-contact operation, high precision, and no radiation. It can rapidly acquire 3D point cloud information of the spinal surface intraoperatively, thus possessing significant application potential in spinal surgery navigation. However, due to objective factors such as surgical incision limitations and tissue occlusion, the point cloud data acquired intraoperatively is typically localized and structurally incomplete. This results in a low overlap between the acquired point cloud data and the complete spinal point cloud model reconstructed preoperatively from medical images. Consequently, existing point cloud registration methods based on geometric or statistical features struggle to meet clinical application requirements in terms of stability and accuracy.

[0006] Therefore, under the premise of ensuring radiation-free operation and clinical operability, how to achieve stable and accurate automatic registration with a low overlap rate between the preoperative complete point cloud and the intraoperative local point cloud has become a key problem that needs to be solved in current spinal surgery navigation technology. Summary of the Invention

[0007] In view of this, the present invention proposes a spatial registration pedicle screw placement navigation system and method based on the analysis of spinal point cloud anatomical structure, in order to solve the problem of insufficient registration accuracy under the conditions of limited overlap area and incomplete structure between preoperative medical images and intraoperative point cloud data.

[0008] According to an embodiment of the present invention, the pedicle screw placement navigation method includes the following steps: A reference point cloud dataset for preoperative planning and a raw point cloud dataset for intraoperative localization are collected. The reference point cloud dataset is obtained by three-dimensional reconstruction of human spinal medical images, and the raw point cloud dataset is obtained by scanning the surgical exposure area during or simulating the operation using a structured light scanning device or an optical positioning system. The reference point cloud dataset is subjected to anatomical structure parsing processing based on joint constraints of spatial geometric information and semantic features to identify single vertebral body-level anatomical structure units and construct corresponding spatial structure constraint relationships, thereby obtaining a subset of reference point clouds with clear anatomical structure information. Guided by the structural constraints provided by the anatomical structure analysis results, the original point cloud dataset and the reference point cloud subset are subjected to point cloud registration processing with enhanced geometric features to obtain the spatial transformation relationship between the preoperative point cloud and the intraoperative point cloud. Navigation information for pedicle screw placement is generated based on the spatial transformation relationship and visualized in the graphical user interface to assist in completing the pedicle screw placement surgery.

[0009] According to a further embodiment of the present invention, the anatomical structure analysis processing is achieved by interactively modeling the spatial geometric information and semantic features in the reference point cloud. The analysis processing includes constructing a multi-level representation of the point cloud features and introducing spatial positional relationship constraints and feature similarity constraints between points to improve the ability to distinguish the boundaries between different vertebral anatomical structures, thereby obtaining a subset of reference point clouds with single vertebral level anatomical structures.

[0010] According to embodiments of this disclosure, the anatomical structural positional relationship and multi-level feature similarity constraint between single vertebral points are achieved by calculating feature attention weights reflecting feature correlation and coordinate attention weights reflecting spatial positional relationship for the point cloud features and coordinates output by the feature extraction layer, respectively; weighted interaction is performed on the corresponding feature information based on the attention weights, and the point cloud features are fused and enhanced through a cross-self-attention mechanism to obtain enhanced point cloud features containing spatial structural information.

[0011] According to embodiments of this disclosure, the training process for the anatomical structure analysis processing employs a joint optimization objective composed of multiple loss functions, including: negative log-likelihood loss, used to constrain the category prediction results of each point in the point cloud and improve the accuracy of point-level classification; Dice loss, used to measure the degree of overlap between the predicted segmentation results and the actual segmentation results, in order to alleviate the impact of sample imbalance between different vertebral categories on model performance in the point cloud segmentation task; and orthogonal regularization loss, used to constrain the orthogonality of the feature mapping matrix obtained during the feature transformation process, reduce feature redundancy, and enhance the stability and discriminative ability of feature representation.

[0012] According to embodiments of this disclosure, the geometrically enhanced point cloud registration process is implemented based on a progressive kernel feature extraction layer structure. Through a hierarchical local neighborhood feature learning approach, multi-scale geometric features are extracted from the point cloud data. Combined with a local information interaction mechanism and a geometric attention mechanism, representative key point features are enhanced. Based on this, accurate registration between the preoperative and intraoperative point clouds is achieved through key point feature matching and spatial transformation parameter solving.

[0013] According to embodiments of this disclosure, the progressive kernel feature extraction layer is constructed using a hierarchical feature extraction network structure. As the network layers deepen, the number of kernel points used in the feature extraction process is gradually adjusted and increased, enabling the network to learn multi-scale spatial geometric features of point clouds from coarse to fine, thereby enhancing its ability to express anatomical structures at different scales.

[0014] According to embodiments of this disclosure, the local information interaction mechanism and geometric attention mechanism include: calculating the feature association relationship between key points and their local neighboring points, and simultaneously calculating the distance and angle information between point clouds, and introducing it into a weighted fusion method of neighborhood features, so that the key point features can comprehensively reflect their structural association information in geometric space and feature space, thereby improving the discriminability and stability of key point features.

[0015] According to embodiments of this disclosure, the registered target point cloud dataset is integrated into a graphical user interface to provide navigation information for pedicle screw placement surgery. The display methods include: grayscale rendering of the original spinal point cloud to fully present the three-dimensional geometry of the spine; displaying the segmented point cloud using a multi-color encoding method to intuitively distinguish different vertebrae and anatomical regions; supporting visualization and label switching at the single vertebra level to facilitate examination of the anatomical structures of each vertebra; and displaying the registration results of the preoperative and intraoperative point clouds through a red-blue dual-color point cloud overlay method to intuitively reflect the registration accuracy and alignment effect.

[0016] According to embodiments of this disclosure, by performing anatomical structure parsing processing based on joint constraints of spatial coordinate information and semantic features on the reference point cloud dataset, the reference point cloud is divided into a subset of point clouds with single vertebral body-level anatomical semantics. This limits the structural range of the preoperative point cloud at the data organization level, so that it corresponds to the observable area of ​​the intraoperative point cloud in terms of anatomical structure. This effectively increases the effective overlap area between the preoperative point cloud and the intraoperative point cloud, and reduces the adverse effect of low overlap rate on subsequent registration accuracy.

[0017] According to embodiments of this disclosure, under the constraints and guidance of the above-mentioned anatomical structure analysis results, geometric feature enhancement point cloud registration processing is performed on the preoperative point cloud and the intraoperative point cloud. By introducing anatomical structure constraints, local geometric feature enhancement, and spatial context information fusion, the registration process can still stably establish the correspondence between point clouds even when the intraoperative point cloud structure is incomplete and the observation range is limited, thereby achieving high-precision point cloud registration under low overlap conditions.

[0018] According to embodiments of this disclosure, the registered point cloud dataset is integrated into a graphical user interface for three-dimensional visualization. By intuitively presenting the vertebral structural relationships, screw planning paths, and their spatial pose information, it assists doctors in making intraoperative judgments and operations, reduces the burden on doctors to understand the spinal spatial structure, and improves the practicality, reliability, and ease of operation of the navigation system in actual surgical scenarios.

[0019] According to a second aspect of this disclosure, a spatial registration pedicle screw placement navigation system based on spinal point cloud anatomical structure analysis is provided, comprising: The data acquisition module is used to acquire a preoperative reference point cloud dataset and an intraoperative raw point cloud dataset. The preoperative reference point cloud dataset is generated by three-dimensional reconstruction of medical imaging data of the human spine and is used to characterize the complete spinal anatomical structure. The intraoperative raw point cloud dataset is acquired by a structured light scanning device or an optical positioning system to acquire the surgical exposure area and is used to characterize the observable local spinal surface structure during the operation. The structural parsing module is used to perform anatomical structural parsing processing on the reference point cloud dataset based on joint constraints of spatial location information and point cloud feature representation, and to divide the reference point cloud dataset into reference point cloud subsets with vertebral-level anatomical semantic information. The registration processing module is used to perform geometric feature enhancement point cloud registration processing on the intraoperative original point cloud dataset and the reference point cloud subset under the anatomical structural constraints of the reference point cloud subset output by the structural analysis module, so as to determine the spatial transformation relationship between the preoperative point cloud and the intraoperative point cloud and generate a registered point cloud dataset. The navigation display module is used to integrate the registration results into the graphical user interface based on the spatial transformation relationship, so as to provide navigation display information for pedicle screw placement.

[0020] According to embodiments of this disclosure, the pedicle screw placement navigation system adopts a modular client-server architecture, including a front-end interaction module, a back-end processing module, and a communication module. The front-end interaction module is used to import point cloud data, configure surgical parameters, visualize 3D point clouds, and facilitate user interaction. The back-end processing module performs calculations related to point cloud anatomical structure analysis and point cloud registration, and generates segmentation and registration results for navigation display. The front-end interaction module and the back-end processing module interact via the communication module, enabling the structural analysis and registration results to be updated and displayed in real-time or near real-time within the front-end interaction module.

[0021] The technical advantages of this disclosure are as follows: 1. According to the embodiments of this disclosure, intraoperative point cloud data can be obtained by using a structured light scanning device or an optical positioning system, which can acquire three-dimensional spatial information of the spinal surface without the need for intraoperative or repeated medical image acquisition, thereby effectively avoiding the radiation risks brought about by traditional image navigation methods.

[0022] 2. By introducing an anatomical structure parsing network that integrates high-dimensional features of point clouds with spatial coordinate information, single-vertebral-level anatomical structure segmentation is performed on the preoperative spinal reference point cloud. This enables the preoperative reconstruction model to form a correspondence with the local point cloud acquired during surgery at the anatomical semantic level, thereby expanding the effective correspondence area between the preoperative point cloud and the intraoperative point cloud and providing a more stable and reliable structural constraint basis for the subsequent registration process.

[0023] 3. A geometric feature enhancement registration strategy based on a progressive kernel feature extraction layer and a local information interaction mechanism can fully extract local geometric details and, combined with spatial context constraints, stably establish feature correspondences between point clouds, achieving high-precision spatial registration, even when the overlap between preoperative and intraoperative point clouds is low. The registration results are further integrated into a graphical user interface for 3D visualization, intuitively presenting the anatomical relationships of the spine and pedicle screw placement navigation information, facilitating spatial judgment and operational decisions by surgeons during surgery.

[0024] 4. This disclosure reduces reliance on intraoperative radiation imaging while effectively improving the registration accuracy and navigation reliability of preoperative and intraoperative point clouds in low-overlap scenarios, providing a safe, accurate and highly practical navigation solution for pedicle screw placement surgery. Attached Figure Description

[0025] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this disclosure. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is an overall flowchart of a spatial registration pedicle screw placement navigation method based on spinal point cloud anatomical structure analysis in one embodiment of the present disclosure. Figure 2 This is a schematic diagram of a pedicle screw placement navigation system in one embodiment of the present disclosure; Figure 3 This is a schematic diagram of the learning-based anatomical structure analysis processing for spinal point clouds in one embodiment of the present disclosure. Figure 4 This is a schematic diagram of the anatomical structure analysis results of the lumbar spine point cloud in one embodiment of the present disclosure; Figure 5 This is a schematic diagram of the anatomical structure analysis results of a single vertebral point cloud in one embodiment of the present disclosure; Figure 6 This is a schematic diagram of the structure of a single vertebral body and spinous process registration module based on a geometric feature enhancement strategy in one embodiment of this disclosure; Figure 7 This is a schematic diagram of the overall architecture of a spinal surgery navigation point cloud registration system that uses a structured light scanning device or an optical positioning system to acquire intraoperative point clouds, according to one embodiment of this disclosure. Figure 8 This is a schematic diagram of a navigation display module for visualizing results in one embodiment of the present disclosure; Figure 9This is a schematic diagram of the structure of a computer-readable storage medium used to implement the above-described pedicle screw placement navigation method and corresponding system, according to one embodiment of the present disclosure. Detailed Implementation

[0026] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0027] It should be noted that, unless otherwise stated, the technical or scientific terms used in this disclosure shall have the ordinary meaning as understood by one of ordinary skill in the art to which this disclosure pertains.

[0028] The embodiments of this disclosure provide a spinal surgery navigation point cloud registration method based on a structured light scanning device or an optical positioning system. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0029] This invention proposes a spatial registration method and system for pedicle screw navigation based on spinal point cloud anatomical structure analysis, belonging to the field of medical image processing and 3D point cloud intelligent computing technology. Addressing the problem of insufficient registration accuracy caused by the limited overlap between preoperative medical images and intraoperatively available surface point clouds during spinal surgery, and the incomplete anatomical structure, a point cloud registration process constrained by anatomical structure analysis results is proposed.

[0030] This method first acquires a reference point cloud dataset for surgical planning and a raw point cloud dataset for intraoperative localization. The reference point cloud dataset is obtained from 3D reconstruction of human spinal medical images, while the raw point cloud dataset is obtained by scanning the surgically exposed area using a structured light scanning device or an optical positioning system. Anatomical structure analysis based on the fusion of spatial coordinates and semantic features is performed on the reference point cloud dataset to identify anatomical units at the single vertebral level and construct corresponding spatial structural constraints, resulting in a subset of reference point clouds with clear anatomical structural information. This improves the effective overlap between preoperative and intraoperative point clouds at the data level. Based on this, and in conjunction with the anatomical structural constraints, geometric feature enhancement point cloud registration is performed for low-overlap scenarios to obtain a registered point cloud dataset.

[0031] Finally, the registration results are integrated into the graphical user interface to assist in displaying spinal spatial pose and screw planning information, providing non-radioactive, high-precision navigation support for pedicle screw placement surgery.

[0032] This embodiment provides a pedicle screw navigation method and system based on spinal point cloud structure analysis and spatial registration. Figure 1 A schematic diagram of the overall framework is shown. (For example...) Figure 1 As shown, the method of this embodiment includes the following stages: Step S101, Intraoperative Surface Data Acquisition: A high-resolution three-dimensional point cloud of the exposed anatomical area is acquired using a structured light scanning device or optical positioning system as the source data for registration. According to embodiments of this disclosure, preoperative data originates from the patient's medical imaging data, and the medical images are reconstructed in three dimensions to obtain an accurate model of the target vertebral body. Intraoperative data simulation is based on a 3D-printed preoperative reconstruction model, whose surface point cloud data is acquired using a structured light scanning device or optical positioning system to simulate the spinous process point cloud acquired during surgery.

[0033] Step S102, Preoperative Vertebral Anatomy Analysis: Based on the preoperative spinal medical imaging data, the spinal anatomy is reconstructed in three dimensions to generate a preoperative reference point cloud model containing complete spinal geometric information; the preoperative reference point cloud model is input into a point cloud segmentation network that integrates high-dimensional semantic features and spatial coordinate information of the point cloud to analyze and distinguish the point cloud regions corresponding to different vertebrae, thereby obtaining a point cloud model with single-vertebra anatomical semantic information, which is used as a reference model for subsequent point cloud registration processing.

[0034] Step S103: Addressing the situation where the source point cloud acquired intraoperatively only covers a local anatomical region and has limited effective overlap with the preoperative reference model, a geometric feature enhancement point cloud registration model is constructed under the constraint of the single-vertebral-level anatomical structure analysis results obtained in step S102. This model is based on a progressive core point feature extraction layer and a local information interaction mechanism. The registration model uses the single-vertebral-level reference point cloud model as the structural constraint object, extracting and matching features between the intraoperative source point cloud and the corresponding vertebral reference point cloud to avoid interference from non-corresponding features between different vertebrae. Through multi-scale geometric feature learning and enhanced local information interaction, a stable feature correspondence is established between the intraoperative source point cloud and the single-vertebral reference model. Based on this feature correspondence, the spatial transformation parameters between the two are solved, thereby achieving high-precision point cloud registration under low overlap conditions.

[0035] Step S104, Registration Result Generation and Navigation Display: Based on the spatial transformation parameters obtained in step S103, the intraoperative source point cloud and the preoperative reference model are mapped in a unified coordinate system. The registration result is transmitted to the front-end display module through a communication mechanism and presented in a three-dimensional visualization manner, providing real-time navigation reference for the pedicle screw placement process.

[0036] Figure 2 This is a schematic diagram of the overall structure of a pedicle screw placement navigation system according to an embodiment of this disclosure. As shown in the figure, the system includes: an optical acquisition unit for acquiring intraoperative spinal surface point cloud data, which may employ a structured light scanning device or an optical positioning system, for acquiring three-dimensional point cloud data of the spinal surface corresponding to the surgically exposed area during surgery; a reconstruction unit for generating a three-dimensional point cloud model based on preoperative spinal medical imaging data to obtain a preoperative reference point cloud model representing the complete spinal anatomy; and a computing unit for performing point cloud anatomical structure analysis and preoperative and intraoperative point cloud spatial registration processing. The computing unit performs single-vertebral-level structural analysis on the preoperative reference point cloud and completes spatial registration processing between the intraoperative point cloud and the preoperative point cloud under structural constraint guidance, generating registration result data; the registration result data is output to a display unit to visualize the pedicle screw placement navigation information.

[0037] Figure 3 This is a schematic diagram of a model structure used for anatomical analysis of spinal point clouds in one embodiment of this disclosure. Figure 3 As shown, the point cloud anatomical structure analysis model extracts features from the preoperative reference point cloud dataset through a feature extraction layer. Then, by introducing a spatial geometric information and semantic feature interaction enhancement module into the model structure, the point cloud features are jointly modeled to enhance the ability of the point cloud features to constrain the geometric structure of the spine. At the same time, the parameters of the point cloud anatomical structure analysis model are optimized by using a joint optimization strategy of multiple loss functions to improve the ability to distinguish between different vertebral anatomical structures, thereby obtaining vertebral-level point cloud structure analysis results with high accuracy and stability.

[0038] The method in this embodiment includes the following stages: Step a1 involves inputting the keypoint features and their corresponding spatial coordinates generated by the feature extraction layer into the spatial geometric information and semantic feature interaction enhancement module. By projecting and mapping the point cloud feature representation and spatial coordinate information respectively, a correlation between the two is established. Based on this, an interactive feature weighting method is used to jointly model the correlation between feature information and coordinate information, thereby achieving enhanced processing of the keypoint features. Figure 3 The diagram illustrates a spatial geometric information and semantic feature interaction enhancement module according to an embodiment of the present disclosure.

[0039] The point cloud features F and coordinates C obtained from the feature extraction layer are respectively processed... Obtained by mapping the projection matrix and Then, the attention scores and values ​​obtained from the two methods are interacted. The specific calculation process is shown in formulas (1), (2), and (3).

[0040] In the above formula, Indicates to The obtained attention scores are normalized. This represents element-wise multiplication, followed by cross-self-attention operations to obtain enhanced feature information after the interaction between features and coordinates. , Finally, they are added together and merged into new feature information.

[0041] Step a2, according to the embodiments of this disclosure, improves segmentation accuracy through joint optimization of negative log-likelihood loss, Dice loss, and orthogonal regularization loss, including: using negative log-likelihood loss (NLL Loss) to optimize the classification accuracy of each point, ensuring that the model's category prediction for each location in the point cloud is as close as possible to the true label. The NLL Loss function directly acts on the log probability of each point, focusing on point-by-point classification performance, and is calculated as shown in formula (4).

[0042] in, This represents the total number of points in the point cloud. Then it represents the first Each point belongs to its true category. The predicted probability. Because of the class imbalance problem in point cloud segmentation, the model may sometimes focus too much or too little on certain regions. Therefore, this disclosure addresses the class imbalance problem in segmentation by designing the Dice loss function to mitigate the class imbalance issue. By optimizing the overlap ratio between the predicted and real regions, the model's segmentation ability for each category is improved. This avoids excessive or insufficient bias towards a particular category. The calculation process is shown in formula (5).

[0043] in, Indicates the first The predicted probability that a point belongs to a certain category. Corresponding to the One-hot encoding of the real label of each point It is a smoothing coefficient to prevent division by zero errors. Feature vectors are fundamental to point cloud segmentation. An orthogonal regularization loss is added to the feature vectors to constrain the orthogonality of the feature transformation matrix, avoid feature dimension redundancy, and enhance the stability of the feature representation. The calculation process is shown in formula (6).

[0044] in, Indicates the first The feature transformation matrix of each sample Let B represent the identity matrix, and B represent the batch size, which is the total number of datasets fed in each time. Finally, the weighting coefficients of each loss function are used. , and The total loss for joint optimization is shown in formula (7).

[0045]

[0046] Based on the aforementioned point cloud structure analysis network, the anatomical structure of a complete lumbar spine point cloud can be analyzed, and the analysis results are as follows: Figure 4 As shown, different vertebral structures are clearly distinguishable in the point cloud space; based on the analysis results, a subset of point clouds containing single vertebral-level anatomical structures can be extracted, and the structural analysis results are as follows. Figure 5 As shown, this provides a structural constraint reference for the subsequent registration process between the intraoperative point cloud and the preoperative model.

[0047] Figure 6 The diagram schematically illustrates a point cloud registration model diagram involving the interaction of a progressive kernel feature extraction layer and neighborhood local information according to an embodiment of the present disclosure. The method of this embodiment includes the following stages: In the feature encoding stage, the model employs a progressive kernel feature extraction layer module to acquire multi-scale local features of the point cloud. In the keypoint enhancement stage, the local information interaction and geometric attention modules introduce a local information interaction mechanism to improve the feature representation of keypoints, and combine it with a geometric attention mechanism to fuse similarity features between keypoints, thereby strengthening the feature representation of overlapping regions. In the decoding stage, the feature propagation module utilizes backpropagation and skip connections to generate high-quality feature matching pairs, providing support for the accurate solution of the subsequent transformation matrix. The method of this embodiment includes the following stages: Step b1, the progressive kernel feature extraction layer module is the main layer. It consists of multiple sub-layers A network layer composed of stacked layers. Each sublayer... It further comprises three stages (Stage 1–Stage 3), cascaded together, with convolutional kernels configured with 15, 16, and 17 points respectively, thereby achieving multi-level feature capture during the deepening of layers. Specifically, the point cloud is defined as… If the first Layer input features are ,but Layer output is The formula for calculating the overall characteristics is shown in formula (8).

[0048] in belong At the same time Centered on the sphere, calculate its spherical neighborhood as follows: Then it falls within the neighborhood. As neighboring points, find the corresponding kernel points for each layer within the sphere, which will be used for subsequent convolutions. Represents a regular distribution in spherical space. One core point, For learnable projection matrices, This means the first The number of kernel points contained in a convolutional kernel. The kernel function based on distance and angle is shown in formula (9).

[0049] in It is the influence radius of the kernel point. Through multi-layer cumulative convolution, the model can effectively capture local feature information in low-overlap regions.

[0050] Step b2: Integrate the local information interaction mechanism and the geometric attention mechanism to construct a collaborative enhancement framework for multi-scale features. The local information interaction mechanism is used to aggregate high-level feature context information, while the geometric attention mechanism focuses on extracting the local geometric structure features of key points in the point cloud. Through the complementary design of the two, the aggregation and enhancement of multi-scale feature information are achieved. The local information interaction mechanism is introduced into the high-level features to enhance the feature interaction capability between key points. Its feature update formula is as follows (10):

[0051] in Point In the last layer, L represents the final feature identifier after all convolutional layers have processed it. These represent the projection matrices of the query, key, and value, respectively. For feature dimension, For point The local neighborhood, This indicates that the obtained values ​​are normalized. This mechanism improves the ability to express feature information in low-overlap regions and increases the registration accuracy.

[0052] Geometric attention mechanisms utilize key point clouds and The corresponding features are calculated separately, and the point cloud features are obtained through a local information interaction mechanism. The calculation process is shown in equation (11).

[0053] In the above formula, the weighting coefficients By attention scores The attention score is obtained by performing a SoftMax operation. The calculation is as shown in formula (12). (12) in, and They represent point clouds respectively. and The corresponding feature vector, These represent the projection matrices for query, key, value, and geometric information, respectively. It represents the geometric information encoding between key point pairs, which combines angle and distance information to enhance geometric perception capabilities; It refers to the dimension of the feature vector of the key point. Geometric information encoding. The calculation is as shown in formula (13).

[0054] in This represents the Euclidean distance between pairs of interactive keypoints. Indicate key points Except for its surroundings Other Angle information formed by neighboring points. Specifically, angle encoding is achieved by constructing... and The angle between the two connecting lines is represented by the maximum value. Projection matrices, representing distance and angle respectively, are used to map geometric information to the feature space. Finally, in the spatial transformation stage, enhanced keypoint features are mapped back to the original point cloud resolution through backpropagation and skip connections, and high-quality matching pairs are generated based on feature matching. Then, the transformation matrix is ​​solved to achieve high-precision registration between the preoperative single vertebral body point cloud and the intraoperative spinous process point cloud.

[0055] Figure 7This diagram schematically illustrates the architecture of a spinal surgery navigation point cloud registration system based on a structured light scanning device or optical positioning system according to an embodiment of this disclosure. The system adopts a modular client-server architecture, consisting of a front-end interaction module, a back-end processing module, and a communication module. It supports point cloud data processing, anatomical structure analysis, registration, and visualization. The front-end, implemented using HTML, CSS, and JavaScript, is responsible for file upload, 3D visualization, tag management, and user interaction. The back-end integrates a deep learning model to complete the analysis and registration tasks, supporting GPU parallel acceleration. The communication module decouples communication through an API, achieving an efficient and interactive communication process, while also possessing good scalability, cross-platform compatibility, and engineering deployability.

[0056] Figure 8 A schematic diagram illustrates a navigation display module for visualizing results according to an embodiment of the present disclosure; from Figure 8 The interface shows that the visualization navigation display module can display: the original point cloud, segmentation results, single cone point cloud, and registered point cloud. In the control panel on the left, users can select a point cloud file, segment the file, select a registration file, and select the object category. Figure 8 The results show that this disclosure can stably establish the feature correspondence between point clouds and achieve high-precision spatial registration. The registration results are further integrated into a graphical user interface for three-dimensional visualization, presenting the anatomical relationships of the spine and the navigation information for pedicle screw placement in an intuitive way, which facilitates spatial judgment and operational decisions by surgeons during surgery.

[0057] Embodiments of this disclosure also provide a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0058] According to embodiments of this disclosure, intraoperative point clouds are acquired using structured light scanning equipment or optical positioning systems, avoiding radiation risks; preoperative point clouds are segmented using neural networks, increasing the effective overlap area between preoperative and intraoperative data; high-precision registration is achieved under low overlap conditions using a geometrically enhanced registration algorithm; and an integrated navigation display module system provides surgeons with intuitive and convenient navigation information. This disclosure solves the problems of insufficient registration accuracy and reliance on radiation imaging in existing spinal surgery navigation systems, providing precise and safe navigation support for pedicle screw placement.

[0059] According to the embodiments of this disclosure, this method can be applied to a variety of spinal surgery scenarios, including but not limited to pedicle screw implantation, spinal fusion, and intervertebral disc replacement, and has broad clinical application prospects.

[0060] According to embodiments of this disclosure, this method can be used in conjunction with other surgical navigation technologies, such as integration with robot-assisted surgical systems, to further improve surgical accuracy and safety.

[0061] The electronic devices provided in this disclosure are based on the same inventive concept as those provided in this disclosure and have the same beneficial effects as the methods used, operated or implemented therein.

[0062] This disclosure also provides a computer-readable storage medium similar to that provided in the foregoing embodiments; please refer to... Figure 9 The computer-readable storage medium shown is an optical disc 30, on which a computer program (i.e., a program product) is stored, which, when run by a processor, executes the methods provided in any of the foregoing embodiments.

[0063] It should be noted that examples of the computer-readable storage medium may also include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical and magnetic storage media, which will not be elaborated here.

[0064] The computer-readable storage medium provided in the above embodiments of this disclosure and the automatic obstacle avoidance method for surgical robots provided in the embodiments of this disclosure are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0065] It should be noted that: Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this disclosure may be practiced without these specific details. In some instances, well-known structures and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0066] Similarly, it should be understood that, in order to simplify this disclosure and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of this disclosure, various features of this disclosure are sometimes grouped together in a single embodiment, figure, or description thereof. However, this approach to disclosure should not be construed as reflecting a schematic diagram in which the claimed disclosure requires more features than are expressly recited in each claim. Rather, as reflected in the following claims, inventive aspects lie in fewer than all features of a single foregoing disclosed embodiment. Therefore, the claims following the detailed description are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of this disclosure.

[0067] Furthermore, those skilled in the art will understand that although some embodiments described herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of this disclosure and form different embodiments. For example, in the following claims, any of the claimed embodiments can be used in any combination.

[0068] The above description is merely a preferred embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed herein should be included within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims. The embodiments of this disclosure have been described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of this disclosure. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of this disclosure, and all such substitutions and modifications should fall within the scope of this disclosure.

Claims

1. A spatial registration pedicle screw navigation method based on point cloud structure analysis, characterized in that, The method includes: Acquisition steps: Acquire a reference point cloud dataset for preoperative planning and a raw point cloud dataset for intraoperative localization, wherein the reference point cloud dataset is obtained by 3D reconstruction of medical images of the human spine, and the raw point cloud dataset is obtained by scanning the surgical exposure area with a structured light scanning device or an optical positioning system. Anatomical structure analysis steps: Perform anatomical structure analysis processing based on joint constraints of spatial geometric information and semantic features on the reference point cloud dataset, identify single vertebral body-level anatomical units and construct corresponding spatial structure constraint relationships to obtain a subset of reference point clouds with anatomical structure information; Registration step: Guided by the spatial structure constraints obtained from the anatomical structure analysis, the original point cloud dataset and the reference point cloud subset are subjected to geometric feature enhancement point cloud registration processing to obtain the spatial transformation relationship between the preoperative point cloud and the intraoperative point cloud. Navigation display steps: Integrate the registered point cloud dataset into the graphical user interface, generate navigation information for pedicle screw placement based on the spatial transformation relationship, and display it visually in the graphical user interface.

2. The method according to claim 1, characterized in that, The reference point cloud dataset is used to provide complete information on the spinal geometry, while the original point cloud dataset is used to characterize the local spinal surface structures observable during surgery.

3. The method according to claim 1 or 2, characterized in that, When using structured light scanning equipment, three-dimensional point cloud data of the spinal surface during surgery is obtained through non-contact structured light projection and imaging; when using an optical positioning system for scanning, spatial point cloud information of the spinal surface during surgery is indirectly obtained by reconstructing the spatial position of optical markers.

4. The method according to claim 1, characterized in that, The anatomical structure analysis steps include: performing learning-based anatomical structure analysis processing on the reference point cloud dataset, and anatomically distinguishing the point cloud structures corresponding to different vertebrae by jointly utilizing the spatial geometric information and semantic feature information of the point cloud; wherein, the learning-based anatomical structure analysis processing improves the discrimination ability of adjacent vertebral structure boundaries by constructing multi-level representations of point cloud features and introducing spatial positional relationship constraints and structural continuity constraints between vertebrae, thereby obtaining a reference point cloud subset of single vertebral-level anatomical structures.

5. The method according to claim 4, characterized in that, The learning-based anatomical structure analysis process includes: introducing the interaction between the spatial positional relationship of points and feature similarity during the point cloud feature learning process, so as to enhance the sensitivity of point cloud features to local geometric structure changes, thereby improving the distinguishability between adjacent vertebral structures.

6. The method according to claim 4, characterized in that, The learning-based anatomical structure analysis processing is trained by constructing a joint optimization objective for vertebral structure analysis tasks. Through joint optimization of negative log-likelihood loss, Dice loss, and orthogonal regularization loss functions, the overall accuracy and robustness of segmentation are improved.

7. The method according to claim 1, characterized in that, The registration step includes: based on the structural constraints provided by the vertebral anatomical structure analysis results of the reference point cloud subset, and considering the limited overlap between the original point cloud dataset and the reference point cloud subset, constructing a multi-scale local geometric feature representation; based on the multi-scale local geometric feature representation, performing feature enhancement processing on key structural points; establishing the feature correspondence between the preoperative point cloud and the intraoperative point cloud based on the enhanced key point features, and solving the spatial transformation relationship between the preoperative point cloud and the intraoperative point cloud.

8. The method according to claim 7, characterized in that, The multi-scale geometric feature representation is obtained through a progressive kernel point feature extraction layer with hierarchical progression. The feature enhancement process is achieved by interactive enhancement of the key point neighborhood structure relationship and feature similarity through a local information interaction mechanism and a geometric attention mechanism.

9. The method according to claim 1, characterized in that, The process of integrating the registered point cloud dataset into a graphical user interface, generating navigation information for pedicle screw placement based on the spatial transformation relationship, and visualizing it in the graphical user interface includes: three-dimensional visualization of the preoperative reference point cloud and the intraoperative original point cloud; differentiated display of different vertebral body point cloud subsets obtained through anatomical structure analysis; and displaying the point cloud relationship before and after registration in an overlay manner to assist in evaluating the spatial consistency of the point cloud registration results.

10. A spatial registration pedicle screw navigation system based on point cloud structure analysis for implementing the method of any one of claims 1 to 9, characterized in that, It includes a front-end interaction module, a back-end processing module, and a communication module, among which, The front-end interaction module is used to import point cloud data, display 3D visualization, and enable user interaction. The backend processing module is used to perform point cloud anatomical structure analysis and spatial registration processing, and output the result data required for navigation display to the communication module.

Images