Feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery

Abstract

The present invention relates to a feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery, which extracts source point cloud features and target point cloud features; calculates point-to-distribution correspondence and overlap fraction vectors based on source point cloud features and target point cloud features; constructs source point cloud mixed feature model and target point cloud mixed feature model based on source point cloud and target point cloud; trains source point cloud mixed feature model based on point-to-distribution correspondence to obtain source point cloud joint mixed model; trains target point cloud mixed feature model based on point-to-distribution correspondence and overlap fraction vector to obtain target point cloud joint mixed model; performs bidirectional registration calculation on source point cloud joint mixed model and target point cloud joint mixed model, and optimizes to obtain the optimal transformation matrix. The present invention can significantly improve the robustness of the algorithm in the face of complex situations such as partial overlap, noise interference and outliers, and the model has good generalization ability.

Description

Technical Field

[0001] The present invention belongs to the field of computer-assisted surgery, and in particular relates to a feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] In the field of computer vision and graphics, point cloud registration is a key technology that is widely used in many industries such as medicine, robotics, and autonomous driving. In intelligent surgical robot systems, accurate point cloud registration is the basis of surgical planning and navigation, and directly affects surgical results and patient safety. Especially in computer-assisted orthopedic surgery, 3D bone registration is crucial, and the transformation relationship between the complete bone point cloud before surgery and the partial point cloud during surgery must be resolved to ensure accurate implementation of the surgery.

[0004] Traditional registration algorithms based on feature matching or ICP are limited in their performance when dealing with partial coverage, noise interference, and outlier problems of intraoperative point clouds. Intraoperative point clouds usually only cover the surgical area, which is significantly different from the complete preoperative point cloud, and both may be affected by noise, reducing the registration accuracy. In addition, traditional methods often ignore the geometric features of point clouds such as normal vectors, which limits the robustness of registration.

[0005] In recent years, although deep learning-based 3D point cloud registration methods have made progress, most methods only use position information and ignore geometric features such as normal vectors associated with point clouds. At the same time, the mixed Gaussian model has limitations in processing global to local registration, because the model parameters differ due to the different coverage areas of the point clouds before and after surgery. Existing registration methods often assume that one side’s point cloud is perfect, but in reality both sides are affected by noise and outliers, and innovative algorithms are urgently needed to improve registration accuracy and robustness. Summary of the invention

[0006] In order to solve at least one technical problem existing in the above-mentioned background technology, the present invention proposes a feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery, which covers the use of deep neural networks to predict the overlapping area between point clouds, and based on the overlapping area, the joint mixed probability model is predicted to more accurately estimate the correspondence between point clouds. At the same time, the method adopts a bidirectional optimization strategy, which optimizes the transformation from the preoperative complete point cloud to the intraoperative partial point cloud, and simultaneously optimizes the transformation from the intraoperative point cloud to the preoperative point cloud, so as to deal with the noise problem in the two spaces of preoperative and intraoperative, and comprehensively improve the robustness and accuracy of the registration.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery, comprising:

[0009] Feature extraction is performed on the source point cloud and the target point cloud respectively to obtain source point cloud features and target point cloud features;

[0010] Calculate the point-to-distribution correspondence and the overlap score vector based on the source point cloud features and the target point cloud features; construct a source point cloud mixed feature model and a target point cloud mixed feature model based on the source point cloud and the target point cloud respectively;

[0011] Based on the point-to-distribution correspondence, the source point cloud mixed feature model is trained to obtain the source point cloud joint mixed model; based on the point-to-distribution correspondence and the overlap score vector, the target point cloud mixed feature model is trained to obtain the target point cloud joint mixed model;

[0012] Based on the joint hybrid model of source point cloud and target point cloud, bidirectional registration calculation is performed to optimize the optimal transformation matrix.

[0013] Preferably, the extracting features of the source point cloud and the target point cloud respectively to obtain source point cloud features and target point cloud features specifically includes:

[0014] Acquire three-dimensional information of a preoperative source point cloud and an intraoperative target point cloud; the three-dimensional information includes three-dimensional coordinate information and three-dimensional normal information;

[0015] The three-dimensional information of the source point cloud and the target point cloud are respectively input into the feature extraction module to obtain the source point cloud features and the target point cloud features; the feature extraction module is constructed based on the multi-layer perceptron of Pointnet combined with the Transformer self-attention mechanism.

[0016] Preferably, the point-to-distribution correspondence relationship includes a source point cloud point-to-distribution correspondence relationship and a target point cloud point-to-distribution correspondence relationship; and the step of calculating the point-to-distribution correspondence relationship according to the source point cloud features and the target point cloud features specifically includes:

[0017] Cluster the source point cloud features and the target point cloud features respectively to obtain the cluster centers;

[0018] The probability of each point in the source point cloud to its cluster center is calculated through the Softmax layer to obtain the corresponding relationship between the source point cloud point and the distribution; and the probability of each point in the target point cloud to its cluster center is calculated through the Softmax layer to obtain the corresponding relationship between the target point cloud point and the distribution.

[0019] Preferably, the calculating the overlap score vector according to the source point cloud features and the target point cloud features specifically includes:

[0020] The cosine similarity is used to calculate the similarity between the source point cloud features and the target point cloud features, and a point-by-point feature similarity matrix is ​​constructed;

[0021] The point-by-point feature similarity matrix is ​​processed by the first multi-layer perceptron network, and the point-by-point feature similarity vector is obtained by using the Sigmoid function;

[0022] The source point cloud features are passed through the global maximum pooling layer to obtain the source point cloud global feature descriptor, which is combined with the target point cloud features to form a global feature matrix; the global feature matrix is ​​passed through the second multi-layer perceptron network to obtain a global similarity vector;

[0023] Multiply the point-wise feature similarity vector with the global similarity vector to obtain the final overlap score vector.

[0024] Preferably, the point-by-point feature similarity matrix is:

[0025]

[0026] Among them, F xn and F ym Represent the high-dimensional feature vectors of the nth and mth points in the source point cloud and target point cloud respectively;

[0027] The point-by-point feature similarity vector is specifically:

[0028]

[0029] Among them, MLP S Indicates that it has been processed by the first multi-layer perceptron network;

[0030] The final overlap score vector is:

[0031]

[0032] in, is the global similarity vector; O q is an M×1 vector, each element of which is qm Represents the similarity score of the overlapping area between the mth point in the target point cloud and the source point cloud.

[0033] Preferably, the step of constructing a source point cloud mixed feature model and a target point cloud mixed feature model based on the source point cloud and the target point cloud respectively comprises:

[0034]

[0035] Wherein, X represents a source point cloud set or a target point cloud set containing 3D coordinate information and 3D normal information; is a parameter, N(x n |μ j,∑ j ) represents the Gaussian mixture model, represents the Fisher mixture model.

[0036] Preferably, the forward prediction transformation matrix T is inversely proportional to the reverse reference transformation matrix The forward prediction error L is calculated by multiplying the product and the difference between the identity matrix I forward ,for:

[0037]

[0038] By inverse prediction transformation matrix T -1 With the reference transformation matrix T gt The reverse prediction error L is calculated by multiplying the product and the difference between the identity matrix I backward ,for:

[0039] L backward =||T gt ·T -1 -I|| 2

[0040] The two deviations are weighted and integrated to obtain the final two-way deviation result L reg ,for:

[0041] L reg =L forward +α·L backward

[0042] Among them, α is the reverse prediction bias weight value, and α≥0.

[0043] In a second aspect, the present invention provides a feature-guided rigid point cloud registration system for computer-assisted orthopedic surgery, comprising:

[0044] A feature extraction module is used to extract features from the source point cloud and the target point cloud respectively to obtain source point cloud features and target point cloud features;

[0045] An intermediate quantity calculation module is used to calculate the point-to-distribution correspondence and the overlap score vector according to the source point cloud features and the target point cloud features; and to construct a source point cloud mixed feature model and a target point cloud mixed feature model based on the source point cloud and the target point cloud respectively;

[0046] The model parameter calculation module is used to train the source point cloud mixed feature model based on the point-to-distribution correspondence to obtain the source point cloud joint mixed model; and to train the target point cloud mixed feature model based on the point-to-distribution correspondence and the overlap fraction vector to obtain the target point cloud joint mixed model;

[0047] The bidirectional registration module is used to perform bidirectional registration calculation based on the joint hybrid model of the source point cloud and the joint hybrid model of the target point cloud to optimize the optimal transformation matrix.

[0048] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery described in the first aspect.

[0049] In a fourth aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery described in the first aspect are implemented.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. In the present invention, by fusing the position and normal information of the point cloud and adopting a bidirectional optimization strategy, the method of the present invention significantly improves the accuracy of point cloud registration and its robustness to noise and outliers.

[0052] 2. The present invention takes into account the normal information of the point cloud. Compared with the traditional method that only uses the position information, it better utilizes the geometric features of the point cloud and enhances the geometric invariance and accuracy of the registration result.

[0053] 3. The designed dual-branch network structure effectively predicts and quantifies the overlapping areas between point clouds, provides key geometric information for the registration process, and enhances the accuracy of registration.

[0054] 4. The bidirectional optimization strategy adopted in the present invention improves the robustness to the noise in both preoperative and intraoperative spaces by simultaneously optimizing the transformation from the preoperative complete point cloud to the intraoperative partial point cloud and the transformation from the intraoperative point cloud to the preoperative point cloud, thereby improving the accuracy of the registration results and significantly enhancing the robustness of the algorithm in complex situations such as partial overlap, noise interference and outliers, ensuring that stable and reliable registration effects can be obtained in various clinical application scenarios.

[0055] 5. Extensive experimental verification on a large number of human skeleton data sets shows that the method of the present invention has achieved ideal results under different overlap ratios, noise levels and the presence of outliers, showing important potential application value in the field of computer-assisted orthopedic surgery.

[0056] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The accompanying drawings in the specification, which constitute a part of the present invention, are used for further understanding of the present invention. The exemplary embodiments of the present invention and their description are used to explain the present invention and do not constitute improper limitations on the present invention.

[0058] Figure 1 This is a main flow chart of a feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery provided by an embodiment of the present invention;

[0059] Figure 2 is a flow chart of calculating the point cloud feature overlap score provided by an embodiment of the present invention;

[0060] Figure 3 is a calculation flow chart of a joint mixed probability model provided by an embodiment of the present invention;

[0061] Figure 4 It is a flow chart of a bidirectional registration calculation module provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0063] Embodiment 1

[0064] like Figure 1 As shown, this embodiment discloses a feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery, comprising the following steps:

[0065] S1: Extract features from the source point cloud and the target point cloud respectively to obtain source point cloud features and target point cloud features;

[0066] S2: Calculate the point-to-distribution correspondence and the overlap score vector based on the source point cloud features and the target point cloud features; construct a source point cloud hybrid feature model and a target point cloud hybrid feature model based on the source point cloud and the target point cloud respectively;

[0067] S3: Train the source point cloud mixed feature model based on the point-to-distribution correspondence to obtain the source point cloud joint mixed model; train the target point cloud mixed feature model based on the point-to-distribution correspondence and the overlap score vector to obtain the target point cloud joint mixed model;

[0068] S4: Based on the joint hybrid model of the source point cloud and the joint hybrid model of the target point cloud, a bidirectional registration calculation is performed to optimize the optimal transformation matrix.

[0069] Next, combine Figure 1 , a feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery disclosed in this embodiment is described in detail.

[0070] In S1, feature extraction is performed on the preoperative source point cloud and the intraoperative target point cloud to obtain source point cloud features and target point cloud features, including:

[0071] S101: Acquire three-dimensional information of a preoperative source point cloud and an intraoperative target point cloud; the three-dimensional information includes three-dimensional coordinate information and three-dimensional normal information;

[0072] S102: Input the three-dimensional information of the source point cloud and the target point cloud into the feature extraction module respectively to obtain the source point cloud features and the target point cloud features; the feature extraction module is constructed based on the multi-layer perceptron of Pointnet combined with the Transformer self-attention mechanism.

[0073] Among them, CT equipment is used to scan and obtain source point cloud data before surgery, and the point cloud set has complete point information; during surgery, the target point cloud information is obtained by real-time scanning with NDI equipment. Due to the limited exposure area of ​​organs during surgery, errors in scanning equipment, and complex environmental factors, the point cloud set is usually incomplete and has large noise / outliers. In order to extract effective features, this embodiment constructs a feature extraction module based on Pointnet's multi-layer perceptron combined with the Transformer self-attention mechanism.

[0074] Preferably, the feature extraction module of this embodiment receives point cloud data containing three-dimensional coordinates and normal information, and performs preliminary extraction of local features for each point through the Pointnet architecture. Subsequently, the multi-layer perceptron further deepens these features and enhances the model's understanding of the geometric properties of the point cloud. On this basis, the Transformer self-attention mechanism is introduced to capture the global dependencies between point clouds, thereby realizing the perception of the global structure of the point cloud at the feature level. Through this method of combining local and global features, the feature extraction module can output a high-dimensional feature representation that comprehensively considers the geometric and topological properties of the point cloud, providing a basis for subsequent point cloud registration tasks.

[0075] As a specific implementation, the Pointnet is a deep learning network for directly learning features from point cloud data, which can process disordered point cloud data. In this embodiment, Pointnet, as the basis of the feature extraction module, encodes the position information of each point through its multi-layer perceptron to extract the local features of each point in the point cloud. The advantage of this embodiment is that it can ignore the order of points in the point cloud and directly learn features of the point cloud data, thereby capturing the geometric structure information of the point cloud.

[0076] Preferably, in the Pointnet architecture, a multi-layer perceptron is used to extract features from each point in the point cloud. The MLP consists of multiple fully connected layers and can learn nonlinear representations of points in the point cloud. In this way, the MLP can extract richer features from local areas of the point cloud, enhancing the model's ability to understand the geometric properties of the point cloud.

[0077] Preferably, the Transformer self-attention mechanism implemented in this method allows the model to consider the information of the entire point cloud when processing each point, and to perceive the global features of the point cloud by calculating the relationship between points. This mechanism is particularly suitable for processing point cloud data because it does not depend on the order of the points and can flexibly capture long-distance dependencies in the point cloud.

[0078] This embodiment performs deep feature extraction on the source point cloud and the target point cloud to obtain feature representations of each point cloud. These feature representations capture the unique attributes of each point in the point cloud, including its position and normal.

[0079] It is necessary to explain in addition that Figure 2 The two source point cloud features have the same meaning and appear twice only to reduce the ambiguity caused by multiple branches in the flowchart; the same is true for the two target point cloud features.

[0080] In S2, the point-to-distribution correspondence is calculated based on the source point cloud features and the target point cloud features, specifically:

[0081] The source point cloud features and the target point cloud features are clustered, and then the corresponding probability of the point to the distribution is estimated through the Softmax layer.

[0082] In this embodiment, the source point cloud features and the target point cloud features are clustered to obtain the cluster centers of the source point cloud and the target point cloud, respectively. The purpose of clustering is to identify and integrate similar point cloud features, thereby providing a structured feature representation for the subsequent point cloud registration process.

[0083] The clustering process involves assigning each point in the point cloud to the nearest cluster center based on its feature vector. These cluster centers represent different regions or patterns in the point cloud data. Through this unsupervised learning method, the model can automatically discover the intrinsic structure of the point cloud data and provide a basis for establishing the correspondence between point clouds. It should be understood that obtaining the cluster center is achievable by those skilled in the art.

[0084] After clustering is completed, the probability of estimating the correspondence between the point and the distribution is achieved by quantifying the similarity between each point cloud feature vector and the distribution center obtained by clustering. Specifically, the Softmax layer is used as a mechanism for estimating probability. This layer receives the feature vector of each source point cloud point and outputs the probability distribution corresponding to each target point cloud cluster center to obtain the correspondence between the source point cloud point and the distribution; it is also used to receive the feature vector of each target point cloud point and output the probability distribution corresponding to each source point cloud cluster center to obtain the correspondence between the target point cloud point and the distribution.

[0085] The output of the Softmax layer can be interpreted as the probability that each point in the point cloud belongs to each cluster center, which reflects the degree of match between the point cloud features and each distribution center. In this way, the model can assign a probability corresponding to the most likely cluster center to each point cloud point, thereby providing key statistical information for point cloud registration.

[0086] In this embodiment, a dual-branch network is designed to calculate the overlap score. The first branch calculates the point feature similarity matrix to estimate the overlap area, and uses cosine similarity to quantify the similarity between feature pairs. The second branch captures the global difference between point cloud features, which is achieved through global maximum pooling and feature concatenation. The results of the two branches are combined to provide an accurate estimate of the overlap area. The specific steps are:

[0087] S201: using cosine similarity to calculate the similarity between the source point cloud features and the target point cloud features, and constructing a point-by-point feature similarity matrix;

[0088] S202: Processing the point-by-point feature similarity matrix through a first multi-layer perceptron network, and obtaining a point-by-point feature similarity vector using a Sigmoid function;

[0089] S203: Passing the source point cloud features through a global maximum pooling layer to obtain a source point cloud global feature descriptor, broadcasting and concatenating the source point cloud features with the target point cloud features to form a global feature matrix; passing the global feature matrix through a second multi-layer perceptron network to obtain a global similarity vector;

[0090] S204: Performing Hadamard product on the point-by-point feature similarity vector and the global similarity vector to obtain a final overlap score vector.

[0091] Specifically, the point feature extraction network is used to extract the features of the source point cloud and the target point cloud to obtain their respective feature representations. Then, in the feature space, the cosine similarity is used to quantify the similarity between each pair of point features and construct a point-by-point feature similarity matrix M F ,for:

[0092]

[0093] Among them, F xn and F ym The high-dimensional feature vectors representing the nth and mth points in the source point cloud and the target point cloud, respectively. The pair of point features refers to two points selected from the source point cloud and the target point cloud, which are paired in the feature space to calculate the similarity. This pairing process is based on the feature representation obtained by the point feature extraction network.

[0094] Next, the feature similarity matrix M F Processed through a multi-layer perceptron network and applying the Sigmoid function to obtain the point-by-point feature similarity vector It is defined as:

[0095]

[0096] The source point cloud features are deeply encoded, and a global maximum pooling layer is used to obtain a source point cloud global feature descriptor with highly generalized information. It is broadcasted and concatenated with the target point cloud features to form a new feature matrix. After being processed by another multi-layer perceptron network, a global similarity vector is obtained.

[0097] The point-wise feature similarity vector With the global similarity vector Through the Hadamard product, the final overlap score vector O is formed q ,for:

[0098]

[0099] Here q is an M×1 vector, each element of which is qm Represents the similarity score of the overlapping area between the mth point in the target point cloud and the source point cloud.

[0100] In S2, a source point cloud mixed feature model and a target point cloud mixed feature model are constructed based on the source point cloud and the target point cloud respectively.

[0101] As an implementation method, the combination of Gaussian mixture model and Fisher mixture model constitutes a mixed feature model. It integrates the position vector and normal vector information of the point cloud feature set, specifically including:

[0102] The joint mixed probability model expresses the probability distribution of point cloud data as a weighted sum, which contains multiple components, and expresses the probability distribution in 6D space as a weighted sum of J∈NJ components, each of which is the product of a Gaussian distribution and a Fisher distribution, which is:

[0103]

[0104] Wherein, X represents a source point cloud set or a target point cloud set containing three-dimensional position information (coordinate information) and three-dimensional normal information, that is, the source point cloud and the target point cloud respectively obtain corresponding mixed feature components based on the formula; represents the model parameters, N(x n |μ j ,∑ j ) represents the Gaussian mixture model, represents the Fisher mixture model.

[0105] like Figure 3 As shown, the mixed feature model calculation process includes the following steps:

[0106] S211: Aggregating the 3D point cloud position information and the 3D point cloud normal information to obtain a point cloud set information representation in a 6D space;

[0107] S212: performing Gaussian distribution and Fisher distribution solutions on the point cloud set in the 6D space respectively to obtain a 6D space Gaussian mixture model and a 6D space Fisher mixture model;

[0108] S213: Perform a product operation on the above two mixed models, and perform a weighted sum element by element to obtain a mixed feature model.

[0109] Among them, in S211, when it is necessary to obtain a source point cloud mixed feature model, the three-dimensional point cloud position information and three-dimensional point cloud normal information of the source point cloud are input; similarly, when it is necessary to obtain a target point cloud mixed feature model, the three-dimensional point cloud position information and three-dimensional point cloud normal information of the target point cloud are input.

[0110] In S3, the source point cloud mixed feature model is trained based on the point-to-distribution correspondence to obtain the source point cloud joint mixed model; the target point cloud mixed feature model is trained based on the point-to-distribution correspondence and the overlap score vector to obtain the target point cloud joint mixed model; specifically, it includes:

[0111] S301: Based on the corresponding relationship between the source point cloud and the distribution, a source point cloud joint feature model including position and normal information is input, and model parameter training is performed to obtain a source point cloud joint hybrid model;

[0112] S302: Based on the target point cloud point-to-distribution correspondence and the overlap score vector, a target point cloud joint feature model including position and normal information is input to perform model parameter training to obtain a target point cloud joint hybrid model.

[0113] In this embodiment, the training process is achieved by minimizing the KL divergence of the probability distribution, thereby obtaining the optimal model parameters, and then the source point cloud joint hybrid model and the target point cloud joint hybrid model. In this way, the model can take into account the overlapping areas between the point clouds, so as to perform registration more accurately. The joint hybrid model is used to obtain the corresponding transformation matrix according to the point-to-distribution correspondence.

[0114] In S4, bidirectional registration calculation is performed based on the joint hybrid model of the source point cloud and the joint hybrid model of the target point cloud, and the optimal transformation matrix is ​​optimized.

[0115] As an implementation method, the reference transformation matrix is ​​the real correspondence matrix from the source point cloud to the target point cloud, and the reverse reference transformation matrix is ​​the real correspondence matrix from the target point cloud to the source point cloud. Both real correspondence matrices are given as prior knowledge and are only used in the training process of the point cloud conversion matrix. The forward prediction transformation matrix is ​​obtained through the source point cloud joint hybrid model, and the reverse prediction transformation matrix is ​​obtained through the target point cloud joint hybrid model. The forward prediction transformation matrix represents the transformation from the source point cloud to the target point cloud, and the reverse prediction transformation matrix represents the inverse transformation from the target point cloud to the source point cloud, and the best match is found by mapping to the other party's space.

[0116] like Figure 4 As shown, the bidirectional registration parameter calculation considers the prediction deviation of the forward transformation matrix and the reverse transformation matrix in the calculation of the loss function. Both deviations use the deviation from the reference transformation matrix as the evaluation criterion, allowing the algorithm to optimize the point cloud registration from two directions at the same time, including:

[0117] S401: The forward prediction transformation matrix T is inversely proportional to the reverse reference transformation matrix T. The forward prediction error L is calculated by multiplying the product and the difference between the identity matrix I forward ,for:

[0118]

[0119] S402: Inverse T of the reverse prediction transformation matrix -1 With the reference transformation matrix T gt The reverse prediction error L is calculated by multiplying the product and the difference between the identity matrix I backward ,for:

[0120] L backward =||T gt ·T -1 -I|| 2

[0121] S403: Weighted integration of the two deviations to obtain a final two-way deviation result L reg,for:

[0122] L reg =L forward +α·L backward

[0123] Among them, α is the reverse prediction bias weight value, and α≥0.

[0124] In the bidirectional registration method of this embodiment, the two transformation matrices are not calculated independently, but influence each other and are optimized together. This bidirectional optimization strategy effectively reduces the impact caused by noise and incomplete data, and improves the accuracy and robustness of registration.

[0125] By minimizing the bidirectional deviation result, the optimal forward prediction transformation matrix and the optimal reverse prediction transformation matrix can be obtained, and the source point cloud and the target point cloud can be aligned through the transformation matrix.

[0126] In the bidirectional registration method, the two transformation matrices are not calculated independently, but influence each other and are optimized together. The purpose of this bidirectional optimization strategy is to reduce the impact caused by noise and incomplete data and improve the accuracy and robustness of the registration.

[0127] This embodiment uses a deep neural network to predict the overlapping area between point clouds, and fuses the position and normal information of the overlapping area into a joint mixed probability model to more accurately estimate the correspondence between point clouds. At the same time, this method adopts a bidirectional optimization strategy. On the basis of optimizing the transformation from the preoperative complete point cloud to the intraoperative partial point cloud, it also optimizes the transformation from the intraoperative point cloud to the preoperative point cloud to deal with the noise problem in both the preoperative and intraoperative spaces, and comprehensively improves the robustness and accuracy of the registration.

[0128] Embodiment 2

[0129] This embodiment provides a feature-guided rigid point cloud registration system for computer-assisted orthopedic surgery, including:

[0130] A feature extraction module is used to extract features from the source point cloud and the target point cloud respectively to obtain source point cloud features and target point cloud features;

[0131] An intermediate quantity calculation module is used to calculate the point-to-distribution correspondence and the overlap score vector according to the source point cloud features and the target point cloud features; and to construct a source point cloud mixed feature model and a target point cloud mixed feature model based on the source point cloud and the target point cloud respectively;

[0132] The model parameter calculation module is used to train the source point cloud mixed feature model based on the point-to-distribution correspondence to obtain the source point cloud joint mixed model; and to train the target point cloud mixed feature model based on the point-to-distribution correspondence and the overlap fraction vector to obtain the target point cloud joint mixed model;

[0133] The bidirectional registration module is used to perform bidirectional registration calculation based on the joint hybrid model of the source point cloud and the joint hybrid model of the target point cloud to optimize the optimal transformation matrix.

[0134] Embodiment 3

[0135] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the steps in the feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery as described in the first embodiment above are implemented.

[0136] Embodiment 4

[0137] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery as described in the first embodiment above are implemented.

[0138] The steps or modules involved in the above embodiments 2 to 4 correspond to those in embodiment 1. For the specific implementation, please refer to the relevant description of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0139] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery, characterized in that: include: Feature extraction is performed on the source point cloud and the target point cloud respectively to obtain source point cloud features and target point cloud features; Calculate the point-to-distribution correspondence and the overlap score vector based on the source point cloud features and the target point cloud features; construct a source point cloud mixed feature model and a target point cloud mixed feature model based on the source point cloud and the target point cloud respectively; Based on the point-to-distribution correspondence, the source point cloud mixed feature model is trained to obtain the source point cloud joint mixed model; based on the point-to-distribution correspondence and the overlap score vector, the target point cloud mixed feature model is trained to obtain the target point cloud joint mixed model; Based on the joint hybrid model of the source point cloud and the joint hybrid model of the target point cloud, a bidirectional registration calculation is performed to optimize the optimal transformation matrix; The overlap score vector is calculated based on the source point cloud features and the target point cloud features, including: The cosine similarity is used to calculate the similarity between the source point cloud features and the target point cloud features, and a point-by-point feature similarity matrix is ​​constructed; The point-by-point feature similarity matrix is ​​processed by the first multi-layer perceptron network, and the point-by-point feature similarity vector is obtained by using the Sigmoid function; The source point cloud features are passed through the global maximum pooling layer to obtain the source point cloud global feature descriptor, which is combined with the target point cloud features to form a global feature matrix; the global feature matrix is ​​passed through the second multi-layer perceptron network to obtain a global similarity vector; Multiply the point-wise feature similarity vector with the global similarity vector to obtain the final overlap score vector.

2. A feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery as claimed in claim 1, characterized in that: The feature extraction of the source point cloud and the target point cloud is performed respectively to obtain the source point cloud features and the target point cloud features, specifically including: Acquire three-dimensional information of a preoperative source point cloud and an intraoperative target point cloud; the three-dimensional information includes three-dimensional coordinate information and three-dimensional normal information; The three-dimensional information of the source point cloud and the target point cloud are respectively input into the feature extraction module to obtain the source point cloud features and the target point cloud features; the feature extraction module is constructed based on the multi-layer perceptron of Pointnet combined with the Transformer self-attention mechanism.

3. The feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery according to claim 1, characterized in that: The point-to-distribution correspondence relationship includes a source point cloud point-to-distribution correspondence relationship and a target point cloud point-to-distribution correspondence relationship; the step of calculating the point-to-distribution correspondence relationship according to the source point cloud features and the target point cloud features specifically includes: Cluster the source point cloud features and the target point cloud features respectively to obtain the cluster centers; The probability of each point in the source point cloud to its cluster center is calculated through the Softmax layer to obtain the corresponding relationship between the source point cloud point and the distribution; and the probability of each point in the target point cloud to its cluster center is calculated through the Softmax layer to obtain the corresponding relationship between the target point cloud point and the distribution.

4. The feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery according to claim 1, characterized in that: The point-by-point feature similarity matrix is: in, and Represent the high-dimensional feature vectors of the nth and mth points in the source point cloud and target point cloud respectively; The point-by-point feature similarity vector is specifically: in, Indicates that it has been processed by the first multi-layer perceptron network; The final overlap score vector is: in, is the global similarity vector; is a A vector where each element Represents the similarity score of the overlapping area between the mth point in the target point cloud and the source point cloud.

5. The feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery according to claim 1, characterized in that: The method of constructing a source point cloud mixed feature model and a target point cloud mixed feature model based on the source point cloud and the target point cloud respectively includes: Wherein, X represents a source point cloud set or a target point cloud set containing 3D coordinate information and 3D normal information; As parameters, represents the Gaussian mixture model, represents the Fisher mixture model.

6. The feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery according to claim 1, characterized in that: The method performs bidirectional registration calculation based on the source point cloud joint hybrid model and the target point cloud joint hybrid model to optimize and obtain the optimal transformation matrix; specifically includes: By forward prediction transformation matrix Inverse of the inverse basis transformation matrix The product between the two and the identity matrix The forward prediction deviation is calculated by the difference between ,for: By inverse prediction transformation matrix With the basis transformation matrix The product between the two and the identity matrix The difference between the two is used to calculate the reverse prediction deviation ,for: The two deviations are weighted and integrated to obtain the final two-way deviation result ,for: in, is the reverse prediction bias weight value, and .

7. A feature-guided rigid point cloud registration system for computer-assisted orthopedic surgery, characterized in that: include: A feature extraction module is used to extract features from the source point cloud and the target point cloud respectively to obtain source point cloud features and target point cloud features; An intermediate quantity calculation module is used to calculate the point-to-distribution correspondence and the overlap score vector according to the source point cloud features and the target point cloud features; and to construct a source point cloud mixed feature model and a target point cloud mixed feature model based on the source point cloud and the target point cloud respectively; The model parameter calculation module is used to train the source point cloud mixed feature model based on the point-to-distribution correspondence to obtain the source point cloud joint mixed model; and to train the target point cloud mixed feature model based on the point-to-distribution correspondence and the overlap fraction vector to obtain the target point cloud joint mixed model; The bidirectional registration module is used to perform bidirectional registration calculation based on the source point cloud joint hybrid model and the target point cloud joint hybrid model to optimize the optimal transformation matrix; The overlap score vector is calculated based on the source point cloud features and the target point cloud features, including: The cosine similarity is used to calculate the similarity between the source point cloud features and the target point cloud features, and a point-by-point feature similarity matrix is ​​constructed; The point-by-point feature similarity matrix is ​​processed by the first multi-layer perceptron network, and the point-by-point feature similarity vector is obtained by using the Sigmoid function; The source point cloud features are passed through the global maximum pooling layer to obtain the source point cloud global feature descriptor, which is combined with the target point cloud features to form a global feature matrix; the global feature matrix is ​​passed through the second multi-layer perceptron network to obtain a global similarity vector; Multiply the point-wise feature similarity vector with the global similarity vector to obtain the final overlap score vector.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in a feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery as described in any one of claims 1 to 6 are implemented.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the feature-guided rigid point cloud registration method for computer-assisted orthopedic surgery as described in any one of claims 1-6 are implemented.

Images