Surgical navigation point cloud registration system based on global optimization

Abstract

The invention discloses a surgical navigation point cloud registration system based on global optimization, and the system comprises the steps: reconstructing a target point cloud based on a preoperative medical image, collecting an intraoperative anatomical surface source point cloud, and converting the point cloud to a patient reference coordinate system; the method comprises the following steps of: selecting a registration point pair with anatomical representativeness to carry out initial coarse registration; on the basis of the coarse registration, executing fine registration, including establishing a point pair relationship, calculating weighting based on a distance residual error and a neighborhood structure, filtering out abnormal point pairs by adopting a bidirectional pruning mechanism based on statistical analysis, jointly estimating rotation and translation parameters through a robust optimization algorithm, and iteratively updating to convergence; and finally outputting a rigid transformation matrix for surgical navigation. According to the method, stable and accurate point cloud registration can be realized under the conditions of sparse point cloud, noise and low overlapping rate in an operation, reliable space mapping support is provided for operation navigation under image guidance, and the method has relatively high robustness and clinical adaptability.

Description

Technical Field

[0001] This invention relates to the technical field of point cloud registration, and in particular to a surgical navigation point cloud registration system based on global optimization. Background Technology

[0002] Surgical navigation point cloud registration refers to the precise alignment of two spaces by solving a rigid transformation matrix between the preoperative point cloud in the CT image space and the intraoperative point cloud in the patient space. This process provides a spatial reference for intraoperative operations, ensuring that preoperative planning information can be accurately mapped to the patient space. It supports key steps such as CT image-guided surgical navigation, surgical execution, and postoperative evaluation, and is one of the core technologies in computer-assisted surgical systems.

[0003] Existing registration techniques mainly include marker-based registration, interactive registration, and point cloud registration. Marker-based registration achieves registration by acquiring the coordinates of specific markers, offering high accuracy and a simple principle, but it relies on invasive fixation devices or additional attachment devices, potentially leading to risks such as trauma, infection, or displacement. Interactive registration involves the operator clicking on anatomical features point by point to establish correspondences, avoiding additional markers; however, the registration effect is highly dependent on the operator's experience, resulting in significant subjective errors and poor stability. Point cloud registration directly utilizes the overall geometry of the 3D point cloud acquired preoperatively and intraoperatively, achieving spatial alignment through optimizing the rigid transformation matrix, eliminating the need for preoperative marking and minimizing safety risks.

[0004] Despite significant progress in existing point cloud registration techniques, they still face considerable challenges in the medical field. Due to the limited exposure area of ​​the surgical field and the locality of point acquisition methods, intraoperative point clouds often only cover a local area of ​​the preoperative point cloud, which is a typical part-to-whole registration problem. Registration is prone to getting trapped in local optima, thus affecting the overall registration accuracy. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and propose a surgical navigation point cloud registration system based on global optimization, which can improve registration accuracy and robustness and is suitable for point cloud registration scenarios with sparse, noisy and low overlap rates.

[0006] To achieve the above objectives, the technical solution provided by this invention is: a surgical navigation point cloud registration system based on global optimization, comprising:

[0007] The preoperative point cloud acquisition module is used to preprocess preoperative CT images and reconstruct preoperative point clouds. Where, q i Preoperative point cloud representation The i-th point in Let M be a 3x3 matrix of real numbers, where M represents the preoperative point cloud. The number of points in;

[0008] The intraoperative point cloud acquisition module uses an optical positioning system to acquire point clouds of key areas on the patient's bone tissue during surgery, forming an intraoperative point cloud. Where, p i Intraoperative point cloud representation The i-th point in Let N represent a 3xN real matrix space, where N represents the intraoperative point cloud. The number of points in;

[0009] The coarse registration module is used to select corresponding anatomical landmarks of the patient in the preoperative and intraoperative spaces as coarse registration points, and to calculate the initial rigid transformation matrix to complete the coarse registration.

[0010] The fine registration module is used to further iterate and optimize the preoperative point cloud based on the coarse registration. Intraoperative point cloud The rigid transformation matrix between the two is used to output the final rigid transformation matrix, and the preoperative point cloud is precisely aligned to the intraoperative point cloud for surgical navigation.

[0011] Furthermore, the preoperative CT images are preprocessed, including: performing a threshold-based region growth segmentation algorithm on the CT images, and using layer-by-layer hole filling processing to enhance the connectivity and closure of the bone tissue region.

[0012] Furthermore, the intraoperative point cloud is acquired by a point-collecting probe, a structured light scanner, or a laser scanner, and then converted to the patient coordinate system for registration. The patient coordinate system is established by fixing an optical positioning tool to the patient and combining it with an optical positioning system.

[0013] Furthermore, the fine registration module performs the following operations:

[0014] S41: Based on the rigid transformation matrix between the current preoperative point cloud and the intraoperative point cloud, the nearest neighbor of each point in the intraoperative point cloud in the preoperative point cloud is determined according to the least Euclidean distance principle, thus forming a point pair. The formula for calculating the nearest neighbor is as follows:

[0015] c(i) = argmin ||Rp i +tq c(i) || 2

[0016] In the formula, R and t represent the rotation matrix and translation matrix in the rigid transformation matrix under the current iteration, respectively, and q c(i) p represents the i-th point in the intraoperative point cloud. i In the preoperative point cloud, the corresponding nearest neighbor point is represented by c(i), where c(i) represents the nearest neighbor point q. c(i) The subscript index;

[0017] S42: Calculate the distance residual for each pair of points. In the formula q c(i) The normal vector is estimated by fitting points in the neighborhood of the point.

[0018] S43: Calculate the weight w between each pair of points based on the distance residual and its neighborhood structure. i The calculation formula is:

[0019]

[0020] In the formula, e N This represents the maximum value among the distance residuals of all point pairs. The larger the distance residual, the smaller the corresponding weight, so as to adaptively weaken the influence of outlier point pairs and improve the robustness and accuracy of the registration results.

[0021] S44: Design a bidirectional pruning mechanism based on statistical analysis. This mechanism employs a dynamic threshold strategy based on statistics to calculate the mean and standard deviation of the weight distribution for all point pairs. It then identifies and removes outlier point pairs whose weights are below or above a preset deviation threshold, thus obtaining the final weights.

[0022]

[0023] In the formula, m i This represents a pruning mask constructed based on statistical analysis, and its calculation formula is as follows: in, and These represent the mean and standard deviation of the weight distribution for all points, respectively; α represents the pruning factor, used to control the pruning range. This represents the sum of weights after pruning, used to normalize the weights of all point pairs after pruning.

[0024] S45: Jointly optimize the rotation matrix R and translation matrix t in the rigid transformation matrix by adopting the iterative reweighted least squares strategy, i.e., the IRLS strategy, and update the rigid transformation matrix in combination with the Rodriguez formula, and apply the updated rigid transformation matrix to the intraoperative point cloud.

[0025] S46: Determine whether the rigid transformation matrix under the current iteration meets the preset termination condition. If the termination condition is not met, return to step S41 to start the next iteration until the termination condition is met.

[0026] Furthermore, the bidirectional pruning mechanism based on statistical analysis is as follows:

[0027] Two types of outlier point pairs are removed: those with weights lower than the mean minus a set multiple of the standard deviation or higher than the mean plus a set multiple of the standard deviation. Point pairs with weights lower than the mean minus a set multiple of the standard deviation are identified as noise points or outliers, while point pairs with weights higher than the mean plus a set multiple of the standard deviation are identified as point pairs dominated by local geometric features. Both types of outlier point pairs are removed from the current point pair set to prevent them from interfering with or dominating the rigid transformation estimation. This mechanism effectively reduces the dependence of registration on dense local point cloud regions or noise points while ensuring the preservation of effective information, thereby improving the overall registration stability and global optimal solution capability.

[0028] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0029] 1. This invention designs a surgical navigation point cloud registration system based on global optimization, which avoids the problem that existing point cloud registration technology is prone to getting trapped in local optima and improves the registration stability in areas with incomplete anatomical structures or insignificant features.

[0030] 2. The fine registration module designed in this invention adopts a weighted approach that integrates neighborhood structure and residual calculation, and combines it with a bidirectional pruning mechanism based on statistical analysis to effectively suppress outlier interference and overfitting, thereby improving registration accuracy.

[0031] 3. In solving the rigid transformation matrix, the present invention adopts joint optimization, simultaneously optimizing the rotation matrix and translation matrix, reducing the propagation of rotation error to translation error, and improving the accuracy of the overall transformation matrix solution.

[0032] 4. This invention provides a reliable rigid transformation matrix between the preoperative CT image space and the intraoperative patient space, ensuring that the preoperative planning information can be accurately mapped to the patient space, which helps to improve surgical accuracy, reduce operational risks, and provide a high-precision spatial reference for subsequent surgical execution. It can be widely used in surgical navigation systems in various fields. Attached Figure Description

[0033] Figure 1 This is an architecture diagram of the system of the present invention.

[0034] Figure 2 This is a schematic diagram of obtaining preoperative point clouds provided by the present invention.

[0035] Figure 3 This is a schematic diagram of obtaining intraoperative point clouds provided by the present invention.

[0036] Figure 4 This is a schematic diagram of the coarse registration region of the femoral point cloud provided by the present invention.

[0037] Figure 5 This is a schematic diagram of the preoperative and intraoperative point clouds of the femur provided by the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this application clearer, the present invention will be described in further detail below with reference to the embodiments and accompanying drawings, but the implementation of the present invention is not limited thereto.

[0039] This embodiment discloses a surgical navigation point cloud registration system based on global optimization. First, coarse registration is performed on preoperative and intraoperative point clouds to initially estimate the rigid transformation matrix. Then, in the fine registration stage, the importance of point pairs is evaluated using a weighted method based on neighborhood structure and distance residuals, and a bidirectional pruning mechanism based on statistical analysis is introduced to remove outlier point pairs. Next, a joint optimization method is used to simultaneously estimate rotation and translation transformation parameters, continuously iteratively optimizing the point cloud alignment results. Finally, iterative termination conditions are determined, and a stable and accurate registration matrix is ​​output, achieving robust point cloud registration. Figure 1 As shown, it includes the following functional modules:

[0040] 1) Preoperative point cloud acquisition module:

[0041] like Figure 2 As shown, the patient's preoperative three-dimensional CT image data was acquired, grayscale analysis was performed on the preoperative CT images, and the bone tissue region was extracted using a threshold-based region segmentation method to obtain preliminary binarized image data.

[0042] Layer-by-layer hole filling and connectivity processing is performed on binary image data to remove non-bone tissue areas and enhance structural closure.

[0043] In three-dimensional voxel data, voxel points located at the boundary of bone tissue are identified and extracted, i.e., the neighborhood of their location contains at least one background voxel.

[0044] Based on the voxel spacing matrix and origin information of CT image data, the coordinates of the identified boundary voxel points are mapped to CT space to generate a preoperative point cloud. Where q i Preoperative point cloud representation The i-th point in the diagram, M represents the preoperative point cloud. The number of points in the data.

[0045] 2) Intraoperative point cloud acquisition module:

[0046] like Figure 3 As shown, the optical positioning system is activated to calibrate and identify the position and orientation of the optical positioning tool and the sampling probe, thus establishing a dynamic tracking environment. The operator holds the sampling probe and makes point-by-point contact with the target bone surface during the operation.

[0047] An optical positioning tool with reflective markings is fixed to the patient using bone clips or non-invasive methods. A patient coordinate system is defined for the unified representation of intraoperative point cloud data. All subsequent intraoperative point clouds are located in the patient coordinate system.

[0048] The optical positioning system records the three-dimensional coordinates of each probe tip contact point in the patient coordinate system in real time, forming an intraoperative point cloud. Where p i Intraoperative point cloud representation The i-th point in the diagram, N represents the intraoperative point cloud. The number of points in the data.

[0049] The point cloud acquisition device is not limited to a point acquisition probe, but may also include acquisition devices such as structured light scanners and laser scanners. All acquired point cloud data are uniformly converted to the patient coordinate system.

[0050] 3) Coarse registration module:

[0051] Anatomical landmarks corresponding to the patient's structure are selected in the preoperative and intraoperative spaces to construct coarse registration pairs. For example... Figure 4 As shown, taking the femur as an example, anatomical landmarks can be points located at the lateral condyle, lateral condyle process, medial condyle, and the center of the intercondylar fossa.

[0052] Based on the selected coarse registration point pairs, the rigid transformation matrix is ​​estimated using the least squares method or singular value decomposition (SVD) to transform the intraoperative point cloud into the preoperative point cloud coordinate system, generating an intraoperative point cloud for subsequent fine registration.

[0053] A preliminary evaluation of the coarse registration results is performed, such as visualizing the degree of overlap or calculating the coarse registration error, to determine whether the coarse registration is valid as the initial input for subsequent fine registration.

[0054] 4) Fine registration module:

[0055] like Figure 5 As shown, taking femoral point cloud registration in knee replacement surgery as an example, the intraoperative point cloud often only covers a local area of ​​the preoperative point cloud due to limited sampling area, unclear structural features, or noise, resulting in a lack of reliable global correspondence between the two. Traditional fine registration methods, such as ICP, rely on initial point pairs for establishment, are prone to getting trapped in local optima, and struggle to maintain registration accuracy and stability under conditions of partial matching and non-uniform distribution. In addition, noisy points or locally dense regions can easily have a dominant influence on the overall transformation, further weakening the robustness of the registration algorithm in complex scenarios. To solve the above problems, this patent designs a fine registration method based on a weighted strategy and a two-way pruning mechanism. It combines distance residuals and neighborhood structure to calculate point pair weights, and identifies and removes abnormal weight pairs through statistical analysis, significantly improving registration stability and global optimal solution capability in sparse, noisy, and low overlap point cloud registration scenarios.

[0056] The precision registration module performs the following operations:

[0057] 4.1) Based on the rigid transformation matrix between the current preoperative point cloud and the intraoperative point cloud, the nearest neighbor of each point in the intraoperative point cloud in the preoperative point cloud is determined according to the least Euclidean distance principle, thus forming point pairs. The formula for calculating the nearest neighbor is:

[0058] c(i) = argmin ||Rp i +tq c(i) || 2

[0059] In the formula, R and t represent the rotation matrix and translation matrix in the rigid transformation matrix under the current iteration, respectively, and q c(i) p represents the i-th point in the intraoperative point cloud. i In the preoperative point cloud, the corresponding nearest neighbor point is represented by c(i), where c(i) represents the nearest neighbor point q. c(i) The subscript index.

[0060] 4.2) Calculate the distance residual for each pair of points. In the formula q c(i) The normal vector is estimated by fitting points in the neighborhood of the point, for example, using methods such as principal component analysis.

[0061] 4.3) Based on the distance residual and its neighborhood structure, calculate the weight w between each pair of points. i The calculation formula is:

[0062]

[0063] In the formula, e N This represents the maximum value among the distance residuals of all point pairs. A larger distance residual corresponds to a smaller weight, thus adaptively mitigating the influence of outlier point pairs and improving the robustness and accuracy of the registration results.

[0064] 4.4) A bidirectional pruning mechanism based on statistical analysis was designed. A dynamic threshold strategy based on statistics was adopted to calculate the mean and standard deviation of the weight distribution of all point pairs. Outlier point pairs with weights below or above a preset deviation threshold were identified and removed to obtain the final weights.

[0065]

[0066] In the formula, m i This represents a pruning mask constructed based on statistical analysis, and its calculation formula is as follows: in, and These represent the mean and standard deviation of the weight distribution for all points, respectively; α represents the pruning factor, used to control the pruning range. This represents the sum of weights after pruning, used to normalize the weights of all point pairs after pruning.

[0067] 4.5) Joint optimization is performed on the rotation matrix R and translation matrix t in the rigid transformation matrix. The iterative reweighted least squares strategy, i.e. IRLS strategy, is adopted and the rigid transformation matrix is ​​updated in combination with the Rodriguez formula. The updated rigid transformation matrix is ​​then applied to the intraoperative point cloud.

[0068] 4.6) Determine whether the rigid transformation matrix in the current iteration satisfies the preset termination condition. If the termination condition is not met, return to step 4.1) to start the next iteration until the termination condition is met.

[0069] In one embodiment, to improve the robustness and global optimum capability of the fine registration process, after updating the point pair weights in each iteration, a bidirectional pruning mechanism based on statistical analysis is used to optimize the weights of all point pairs. Specifically, the mean and standard deviation of the weights of all current point pairs are first calculated, and pruning upper and lower bound thresholds are constructed based on these. If the weight of a point pair is lower than the mean minus a preset multiple of the standard deviation, the point pair is identified as a noise point or an outlier; if its weight is higher than the mean plus a preset multiple of the standard deviation, it is considered a point pair dominated by local geometric features. Both types of outlier point pairs are removed from the current point pair set to prevent them from interfering with or dominating the rigid transformation estimation. This mechanism effectively reduces the dependence of registration on dense local point cloud regions or noise points while ensuring the preservation of effective information, thereby enhancing the stability and accuracy of the overall transformation estimation.

[0070] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A globally optimized based surgical navigation point cloud registration system, characterized in that, include: A preoperative point cloud acquisition module is configured to pre-process a preoperative CT image and reconstruct a preoperative point cloud wherein q i represents an i-th point in the preoperative point cloud represents a 3-row M-column real matrix space, and M represents a number of points in the preoperative point cloud .​ An intraoperative point cloud acquisition module acquires a key region point cloud on a patient's bone tissue in an operation through an optical positioning system to form an intraoperative point cloud wherein p i represents an i th point in the intraoperative point cloud represents a 3-row N-column real matrix space, and N represents a number of points in the intraoperative point cloud .​ The coarse registration module is used to select corresponding anatomical landmarks of the patient in the preoperative and intraoperative spaces as coarse registration points, and to calculate the initial rigid transformation matrix to complete the coarse registration. A fine registration module is used to further iteratively optimize the preoperative point cloud using fine registration on the basis of coarse registration and the intraoperative point cloud between the intraoperative point cloud and the preoperative point cloud, finally output the rigid transformation matrix, and accurately align the preoperative point cloud to the intraoperative point cloud for surgical navigation.

2. The globally optimized point cloud registration system for surgical navigation of claim 1, wherein, Preprocessing of preoperative CT images includes: performing a threshold-based region growth segmentation algorithm on the CT images and employing layer-by-layer hole filling processing to enhance the connectivity and closure of bone tissue regions.

3. The globally optimized point cloud registration system for surgical navigation of claim 1, wherein, The intraoperative point cloud is acquired by a point-collecting probe, a structured light scanner, or a laser scanner, and then converted to the patient coordinate system for registration. The patient coordinate system is established by fixing an optical positioning tool to the patient and combining it with an optical positioning system.

4. The surgical navigation point cloud registration system based on global optimization according to claim 1, characterized in that, The fine registration module performs the following operations: S41: Based on the rigid transformation matrix between the current preoperative point cloud and the intraoperative point cloud, determine the nearest neighbor point in the preoperative point cloud for each point in the intraoperative point cloud according to the least Euclidean distance principle, thereby forming point pairs. The formula for calculating the nearest neighbor is: c(i) = argmin ||Rp i + t - q c(i) || 2 where R and t represent the rotation matrix and translation matrix in the rigid transformation matrix at the current iteration, respectively, q c(i) denotes the i-th point in the intra-operative point cloud p i its corresponding nearest neighbor in the pre-operative point cloud, c(i) denotes the index of the nearest neighbor q c(i) . S42: Calculate the distance residual for each pair of points. In the formula q c(i) The normal vector is estimated by fitting points in the neighborhood of the point. S43: Calculate the weight w between each pair of point pairs based on the distance residual and its neighborhood structure i The calculation formula is: In the formula, e N represents the maximum value in all point pair distance residuals, the greater the distance residual, the smaller the corresponding weight, to adaptively weaken the influence of abnormal point pairs and improve the robustness and accuracy of the registration result; S44: Design a bidirectional pruning mechanism based on statistical analysis. This mechanism employs a dynamic threshold strategy based on statistics to calculate the mean and standard deviation of the weight distribution for all point pairs. It then identifies and removes outlier point pairs whose weights are below or above a preset deviation threshold, thus obtaining the final weights. In the formula, m i This represents a pruning mask constructed based on statistical analysis, and its calculation formula is as follows: in, and These represent the mean and standard deviation of the weight distribution for all points, respectively; α represents the pruning factor, used to control the pruning range. This represents the sum of weights after pruning, used to normalize the weights of all point pairs after pruning. S45: Jointly optimize the rotation matrix R and translation matrix t in the rigid transformation matrix by adopting the iterative reweighted least squares strategy, i.e., the IRLS strategy, and update the rigid transformation matrix in combination with the Rodriguez formula, and apply the updated rigid transformation matrix to the intraoperative point cloud. S46: Determine whether the rigid transformation matrix under the current iteration meets the preset termination condition. If the termination condition is not met, return to step S41 to start the next iteration until the termination condition is met.

5. A surgical navigation point cloud registration system based on global optimization according to claim 4, characterized in that, The bidirectional pruning mechanism based on statistical analysis is as follows: Two types of outlier point pairs are removed: those with weights lower than the mean minus a set multiple of the standard deviation or higher than the mean plus a set multiple of the standard deviation. Point pairs with weights lower than the mean minus a set multiple of the standard deviation are identified as noise points or outliers, while point pairs with weights higher than the mean plus a set multiple of the standard deviation are identified as point pairs dominated by local geometric features. Both types of outlier point pairs are removed from the current point pair set to prevent them from interfering with or dominating the rigid transformation estimation. This mechanism effectively reduces the dependence of registration on dense local point cloud regions or noise points while ensuring the preservation of effective information, thereby improving the overall registration stability and global optimal solution capability.

Images