A skull surgery navigation method based on curvature guidance and online pose maintenance

Abstract

This invention discloses a method for intraoperative navigation of craniostomy based on curvature guidance and online pose maintenance, belonging to the field of intraoperative navigation of craniostomy. This invention constrains the curvature prior of the natural human structure: curvature estimation uses the auricle (along the arcuate ridge line of the helix / antihelix) and the nasal ala (along the arcuate ridge line of the nasal ala rim / nasal groove) as anatomical references, preferentially preserving these curvature features and suppressing their tangential "sawtooth" jitter; it provides a continuous and stable 6DoF initial pose on the time axis, and gating freeze / backoff is used to avoid initial value jumps in case of occlusion or mismatch; spatially, curvature-guided filtering is used, combined with anisotropic diffusion of prior constraints to complete edges and holes, improving effective overlap and connectivity. The correspondence between point clouds is more reliable, and the initial values ​​are closer to the true solution, thus enabling robust kernel ICP to converge stably under normal / distance gating, meeting the requirements of high-precision rigid registration and continuous navigation of intraoperative-preoperative 3D models.

Description

Technical Field

[0001] This invention relates to the field of intraoperative navigation in craniectomy, and more specifically to an intraoperative navigation method for craniectomy based on curvature guidance and online pose maintenance. Background Technology

[0002] Intraoperative navigation in cranial surgery is a specific application of image-guided neurosurgery (IGS) for cranial surgery. Its core is to use precise registration of preoperative images with the patient during surgery, combined with spatial positioning technology, to display the relative positions of surgical instruments, cranial defects / lesions, and intracranial blood vessels, nerves, and / or brain tissue in real time during surgery. This achieves sub-millimeter-level precision in cranial manipulation, solving the pain point of "judging anatomical positions based on experience" in traditional cranial surgery. It is applicable to almost all cranial-related surgical procedures, such as cranioplasty, tumor resection, deformity correction, and drill biopsy.

[0003] However, in landmark-free navigation in narrow surgical areas such as the skull base / temporal bone, the intraoperative RGB-D point cloud is often "local and with low overlap, uneven density and many holes, and large initial pose jitter between frames", which directly leads to the classic ICP being extremely sensitive to initial values ​​and outliers, with a narrow convergence domain that is prone to getting trapped in erroneous minima or divergence. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, this invention provides a craniostomy navigation method based on curvature guidance and online pose maintenance. This method solves the problems of inaccurate navigation caused by the extreme sensitivity of classical ICP to initial values ​​and outliers, narrow convergence domain that easily falls into erroneous minima or diverges, resulting in classical navigation without markers.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A method for intraoperative navigation of craniectomy based on curvature guidance and online pose maintenance is provided, which includes the following steps: A reconstructed 3D model of the natural structure is obtained from a preoperative 3D CT model containing the natural structure, and the reconstructed 3D model of the natural structure is placed in the intraoperative field of view; the natural structure includes the auricle and the nasal ala. Using the reconstructed 3D model of the natural structure as the tracking target, the RGB-D image frames containing the natural structure captured in real time during the operation are back-projected. The point cloud obtained by back-projection is selected and reweighted according to the feature curvature, so as to form a point cloud with natural structure preservation without sacrificing edge details. Using the visible region of the preoperative 3D CT model containing natural structures and the distance from the preoperative 3D CT model containing natural structures as prior boundary constraints, a diffusion coefficient field coupled with local feature curvature and observation uncertainty is constructed on the edge-preserving point cloud of natural structures. Under the control of the diffusion coefficient field, points in the neighborhood of deep holes, occlusions or sparse regions are taken as points to be completed and anisotropic diffusion interpolation is performed to update the position of the points to be completed to the weighted average of the coordinates of its neighborhood points, thereby generating a natural structure complete point cloud that takes into account both shape continuity and edge sharpness. Using the most recent stable pose of the reconstructed 3D model of the natural structure as the initial value, the point cloud of the natural structure is precisely registered with the preoperative 3D CT model containing the natural structure. The results of the superimposed display and error monitoring are then output to complete intraoperative navigation of the skull.

[0006] The beneficial effects of this invention are as follows: This invention constrains the curvature prior of the natural human structure: curvature estimation uses the auricle (along the arcuate ridge line of the helix / helix-antihelix) and the nasal ala (along the arcuate ridge line of the nasal ala rim / nasal ala groove) as anatomical references, preferentially preserving these curvature features and suppressing their tangential "sawtooth" jitter; it provides a continuous and stable 6DoF initial pose on the time axis, and gating freeze / backoff is used to avoid initial value jumps when occlusion or mismatch occurs; spatially, curvature-guided filtering is used, combined with anisotropic diffusion of prior constraints to complete edges and holes, improving effective overlap and connectivity. The correspondence between point clouds is more reliable, and the initial values ​​are closer to the true solution, thus enabling robust kernel ICP to converge stably under normal / distance gating, meeting the requirements of high-precision rigid registration and continuous navigation of intraoperative-preoperative 3D models. Attached Figure Description

[0007] Figure 1 This is a flowchart illustrating the method. Figure 2 This is a schematic diagram illustrating the weighted ICP effect combined with characteristic curvature in the embodiment; Figure 3 This is a schematic diagram illustrating the use of natural structures such as the auricle / nose wing as landmark structures for heterogeneous diffusion interpolation in the embodiment. Detailed Implementation

[0008] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0009] like Figure 1 As shown, this intraoperative cranioscopy navigation method based on curvature guidance and online pose maintenance includes the following steps: S1. Obtain a reconstructed 3D model of the natural structure based on the preoperative 3D CT model containing the natural structure, and place the reconstructed 3D model of the natural structure within the intraoperative field of view; wherein the natural structure includes the auricle and the nasal ala. S2. Using the reconstructed 3D model of the natural structure as the tracking target, the RGB-D image frames containing the natural structure captured in real time during the operation are back-projected. The point cloud obtained by back-projection is selected and reweighted according to the feature curvature, so as to form a point cloud with natural structure preservation without sacrificing edge details. S3. Using the visible area of ​​the preoperative 3D CT model containing natural structures and the distance from the preoperative 3D CT model containing natural structures as prior boundary constraints, construct a diffusion coefficient field coupled with local feature curvature and observation uncertainty on the edge-preserving point cloud of natural structures. S4. Under the control of the diffusion coefficient field, points in the neighborhood of deep holes, occlusions or sparse regions are taken as points to be completed and anisotropic diffusion interpolation is performed to update the position of the points to be completed to the weighted average of the coordinates of its neighborhood points, thereby generating a natural structure completion point cloud that takes into account both shape continuity and edge sharpness. S5. Using the most recent stable pose of the reconstructed 3D model of the natural structure as the initial value, perform fine registration between the natural structure complete point cloud and the preoperative 3D CT model containing the natural structure, output the superimposed display and error monitoring results, and complete the intraoperative navigation of the skull.

[0010] In some embodiments, a specific method for selecting and reweighting the point cloud obtained by back projection according to its feature curvature to form an edge-preserving point cloud with a natural structure without sacrificing edge details includes the following steps: A1. Calculate the maximum principal curvature of the point in the point cloud obtained by back projection, the angle or difference between the point normal and the average normal of its neighborhood, and the reference threshold for depth uncertainty. A2. Based on the curvature template of the pre-constructed natural structure, for points belonging to the natural structure region, according to the difference between the actual curvature sampled along the edge or arc of the natural structure and the curvature in the curvature template, a bias is applied to the reference threshold to obtain the final maximum principal curvature constraint threshold, the angle or difference constraint threshold between the point normal and its neighborhood average normal, and the depth uncertainty constraint threshold. A3. Using the maximum principal curvature of a point, the angle or difference between the point's normal and the average normal of its neighborhood, the depth uncertainty, and the corresponding constraint threshold as constraints for the edge-preserving point cloud of the natural structure, the edge-preserving point cloud of the natural structure is obtained. Its expression is:

[0011] in Preservation of edge points in natural cloud structures; Refers to points in the point cloud obtained by back projection; , and Points The maximum principal curvature, the angle or difference between the point normal and the average normal of its neighborhood, and the depth uncertainty; , and These are the final maximum principal curvature constraint threshold, the angle or difference constraint threshold between the point normal and its neighborhood average normal, and the depth uncertainty constraint threshold, respectively.

[0012] For example, the benchmark threshold is a preset quantile, or the sum of the median and the mean absolute deviation. The bias is applied by applying an additive bias or a multiplicative bias.

[0013] In some embodiments, the expression for the diffusion coefficient field is:

[0014] in Indicates the anisotropic diffusion coefficient; This is the edge-stopping function; This is a priori exponential decay function used to adjust the diffusion intensity based on depth uncertainty; The prior boundary constraint function can be the Perona–Malik thermal conductivity function, the Sigmoid gate function, and / or the Heaviside step gate function, used to allow diffusion near the CT prior boundary to naturally decay or regress, thereby balancing hole crossing and edge sharpness maintenance. The distance field of a preoperative 3D CT model containing natural structures represents the point. The Euclidean distance to the preoperative 3D CT model containing natural structures is used as a distance prior. This indicates calculating the gradient.

[0015] For example, such as Figure 3 As shown, the color / depth frames superimposed on the left are back-projected to obtain the intraoperative point cloud. However, due to occlusion, reflection, or sensor defects, the point cloud has large areas of missing data (holes). The red box marks natural structural regions (such as the auricle / nasal ala), whose local geometry has stable and identifiable curvature features; the small image in the lower right corner shows a local detail of this region, used to emphasize that "high-curvature edges / ridges" are important and reliable information sources for subsequent completion and registration. Using natural structures such as the auricle / nasal ala as landmark structures... Figure 3 The right side shows the point cloud result sequence after diffusion iteration: as the iteration progresses, the void area is gradually filled and the surface continuity is enhanced; at the same time, due to the "curvature-uncertainty-CT" coupling constraint of the diffusion coefficient, high curvature edge structures such as the auricle are preserved, and the overall shape is consistent with the CT prior, avoiding boundary blurring or shape drift caused by unconstrained smoothing.

[0016] For example, heterogeneous diffusion interpolation is performed on the point cloud structure in the form of a weighted graph Laplacian, and CT prior regression is superimposed to complete the unobserved areas. The corresponding expression is:

[0017] in Iteration The position of the point to be completed in three-dimensional space; Iteration The position of the point to be completed in three-dimensional space; The diffusion step size; To use the anisotropic diffusion coefficient The Laplacian term, calculated on the adjacency graph of the point cloud, is used to spread the geometric information in the neighborhood to the point to be completed to achieve smooth interpolation. These are the prior regression coefficients; express exist The gradient at the point; the adjacency graph of the point cloud is constructed from edge-preserving point clouds with natural structures; In some embodiments, during the heterogeneous diffusion interpolation process for the points to be filled, the condition for stopping the iteration is one of the following: Condition 1: The local principal curvature growth rate of the points to be filled exceeds a preset threshold; Condition 2: The rate of decrease in distance from the point to be completed to the preoperative 3D CT model of the natural structure is lower than the preset threshold; Condition 3: The increase in the crossover ratio of the CT projection contours of the natural structure before and after iteration is less than a set threshold.

[0018] In some embodiments, after obtaining the natural structure-completed point cloud, the following operations are also performed: The observation contour of the natural structure is obtained by completing the point cloud based on the natural structure. If the overlap between the template projection contour of the natural structure and the observation contour, the visibility of the natural structure, or the depth hole ratio is lower than the preset threshold, the pose update of the current frame is frozen. If the pose update is frozen for consecutive frames, the tracking is restarted from the most recent key frame. The key frame selection must simultaneously meet the following conditions: ICP convergence, stability of residual and attitude increment, and the overlap of contours meets the standard.

[0019] In some embodiments, fine registration employs points with robust kernel functions. The ICP algorithm, also known as ICP fine registration, is used in the process of ICP fine registration. If the distance between corresponding point pairs is greater than or equal to the distance threshold, the normal angle is greater than or equal to the angle threshold, the depth uncertainty of any point is higher than the preset upper limit, or the corresponding point pair falls in the edge region of the depth hole, then it is determined to be a suspicious corresponding point pair and given a preset low weight in the robust kernel or directly removed.

[0020] For example, the points to be completed and heterogeneous diffusion interpolation is performed to generate a completed point cloud and fine registration are run in a cascaded or alternating manner: In the current pose, several steps are performed to generate the point cloud to be completed and heterogeneous diffusion interpolation is executed. When the effective overlap between the completed point cloud and the preoperative 3D CT model of the natural structure exceeds a threshold, a local ICP fine registration is triggered. That is, based on the results of the previous round of ICP fine registration, the pose obtained at the end of the previous round of ICP fine registration is used as the initial value to perform several iterations of optimization to minimize the point-to-surface residual, thereby further improving the registration accuracy. If the residual increases or the overlap between the template projection contour and the observed contour decreases below the threshold during the ICP fine registration process, the generation of the completed point cloud is paused and the model is reverted to the pose obtained at the end of the previous round of ICP fine registration to ensure overall convergence and robustness.

[0021] The expression for minimizing the point-to-surface residual during the ICP fine registration process, involving several iterations of optimization, is as follows:

[0022] in To complete the point cloud with natural structure One point; To complete the point cloud structure The joint weight of each point , , and The first point cloud is the one with natural structure completion. The maximum principal curvature of a point, the angle or difference between the point normal and the average normal of its neighborhood, and the depth uncertainty; It is a function that monotonically increases with respect to curvature, used to enhance the weight of points on high-curvature ridges; and These are functions that monotonically decrease with respect to normal consistency and depth uncertainty, respectively, and are used to suppress points with normal instability or unreliable depth. , and The shape and parameters are calibrated during pre-training; For preoperative 3D CT models containing natural structures in Unit normal vector at the location transpose; For the preoperative 3D CT model of natural structure and Corresponding points; Represents the rotation matrix; This represents the translation vector.

[0023] The aforementioned joint weights can enhance the weight of points on high-curvature ridges while suppressing points with unstable normals or unreliable depths, thereby increasing the registration weight of "feature points with stable curvature in natural structures such as auricles / nasal alae" and obtaining more accurate registration results.

[0024] The purpose of ICP (Intraoperative Point Cloud) is to align the intraoperative point cloud onto the CT model. Each iteration is equivalent to having many points find their corresponding positions on the CT scan. Since not every correspondence is reliable, we use joint weights to enhance reliable correspondences while suppressing suspicious ones. A commonly used residual in ICP is the distance along the target surface normal. This joint weight serves as the coefficient for each corresponding residual term and is directly fed into the objective function of ICP, forming a weighted ICP. Figure 2 As shown, weighted ICP allows points with "high discrimination and low uncertainty" to contribute more, thereby achieving more accurate registration.

[0025] In the specific implementation process, to achieve directional suppression, in the estimated curvature ridge local coordinate system, the ridge direction is defined as tangential, and the normal is defined as the cross-ridge direction. A one-dimensional low-pass filter is applied to the tangential coordinate sequence to smooth high-frequency "sawtooth noise," while the smoothing of the normal direction is weakened to preserve cross-ridge details. Furthermore, the normal and curvature at each scale of the pyramid are calculated separately. When the angle between the normals of the same point in two adjacent scales exceeds a preset angle threshold or the difference in principal curvature exceeds a preset amplitude threshold, it is determined to be a cross-scale inconsistency. Attenuate or remove it.

[0026] In summary, this invention designs point cloud completion as a three-stage process: "edge preservation, diffusion prevention, and prior regression." First, curvature analysis is performed on the point cloud, with weighted / filtered data to prioritize the preservation of high-curvature edges and ridge details as the input base for diffusion completion. Then, iterative diffusion is performed, with the curvature growth rate, the distance reduction rate from the point cloud to the preoperative 3D CT model of the natural structure, and the improvement of the intersection-union ratio with the CT projection contour as stopping criteria to avoid over-diffusion and invalid iteration. Finally, a CT prior regression term is introduced to constrain the missing region, ensuring that the completion result reasonably regresses to the preoperative 3D CT model of the natural structure, preventing morphological drift caused by unconstrained smoothing.

Claims

1. A method for intraoperative navigation of a craniotomy based on curvature guidance and online pose maintenance, characterized in that, Includes the following steps: A reconstructed 3D model of the natural structure is obtained from a preoperative 3D CT model containing the natural structure, and the reconstructed 3D model of the natural structure is placed in the intraoperative field of view; the natural structure includes the auricle and the nasal ala. Using the reconstructed 3D model of the natural structure as the tracking target, the RGB-D image frames containing the natural structure captured in real time during the operation are back-projected. The point cloud obtained by back-projection is selected and reweighted according to the feature curvature, so as to form a point cloud with natural structure preservation without sacrificing edge details. Using the visible region of the preoperative 3D CT model containing natural structures and the distance from the preoperative 3D CT model containing natural structures as prior boundary constraints, a diffusion coefficient field coupled with local feature curvature and observation uncertainty is constructed on the edge-preserving point cloud of natural structures. Under the control of the diffusion coefficient field, points in the neighborhood of deep holes, occlusions or sparse regions are taken as points to be completed and anisotropic diffusion interpolation is performed to update the position of the points to be completed to the weighted average of the coordinates of its neighborhood points, thereby generating a natural structure complete point cloud that takes into account both shape continuity and edge sharpness. Using the most recent stable pose of the reconstructed 3D model of the natural structure as the initial value, the point cloud of the natural structure is precisely registered with the preoperative 3D CT model containing the natural structure. The results of the superimposed display and error monitoring are then output to complete intraoperative navigation of the skull.

2. The intraoperative navigation method for craniectomy based on curvature guidance and online pose maintenance according to claim 1, characterized in that, The specific method for selecting and reweighting point clouds obtained by back projection according to feature curvature to form edge-preserving point clouds with natural structure without sacrificing edge details includes the following steps: Calculate the maximum principal curvature of the point in the point cloud obtained by back projection, the angle or difference between the point normal and the average normal of its neighborhood, and the benchmark threshold for depth uncertainty. Based on the curvature template of the pre-constructed natural structure, for points belonging to the natural structure region, the difference between the actual curvature sampled along the edge or arc of the natural structure and the curvature in the curvature template is applied to the reference threshold to obtain the final maximum principal curvature constraint threshold, the angle or difference constraint threshold between the point normal and its neighborhood average normal, and the depth uncertainty constraint threshold. The constraints formed by the maximum principal curvature of a point, the angle or difference between the point's normal and the average normal of its neighborhood, the depth uncertainty, and the corresponding constraint threshold are used as constraints for the edge-preserving point cloud of the natural structure. The expression for this edge-preserving point cloud of the natural structure is as follows: in Preservation of edge points in natural cloud structures; Refers to points in the point cloud obtained by back projection; , and Points The maximum principal curvature, the angle or difference between the point normal and the average normal of its neighborhood, and the depth uncertainty; , and These are the final maximum principal curvature constraint threshold, the angle or difference constraint threshold between the point normal and its neighborhood average normal, and the depth uncertainty constraint threshold, respectively.

3. The intraoperative navigation method for craniostomy based on curvature guidance and online pose maintenance according to claim 2, characterized in that, The baseline threshold is a preset quantile, or the sum of the median and the mean absolute deviation.

4. The intraoperative navigation method for craniectomy based on curvature guidance and online pose maintenance according to claim 2, characterized in that, The method for applying the bias is as follows: Apply additive or multiplicative biases.

5. The intraoperative navigation method for craniostomy based on curvature guidance and online pose maintenance according to claim 2, characterized in that, The expression for the diffusion coefficient field is: in Indicates the anisotropic diffusion coefficient; This is the edge-stopping function; This is a priori exponential decay function used to adjust the diffusion intensity based on depth uncertainty; The prior boundary constraint function is used to allow diffusion near the CT prior boundary to naturally decay or regress, thereby balancing hole crossing and edge sharpness maintenance. The distance field of a preoperative 3D CT model containing natural structures represents the point. The Euclidean distance to the preoperative 3D CT model containing natural structures is used as a distance prior. This indicates calculating the gradient.

6. The intraoperative navigation method for craniostomy based on curvature guidance and online pose maintenance according to claim 5, characterized in that, Heterogeneous diffusion interpolation is performed on the point cloud structure in the form of a weighted graph Laplacian, and CT prior regression is superimposed to complete the unobserved areas. The corresponding expression is: in Iteration The position of the point to be completed in three-dimensional space; Iteration The position of the point to be completed in three-dimensional space; The diffusion step size; To use the anisotropic diffusion coefficient The Laplacian term, calculated on the adjacency graph of the point cloud, is used to spread the geometric information in the neighborhood to the point to be completed to achieve smooth interpolation. These are the prior regression coefficients; express exist The gradient at the point; the adjacency graph of the point cloud is constructed from edge-preserving point clouds with natural structures; During the heterogeneous diffusion interpolation process for the points to be filled, the iteration stops under one of the following conditions: Condition 1: The local principal curvature growth rate of the points to be filled exceeds a preset threshold; Condition 2: The rate of decrease in distance from the point to be completed to the preoperative 3D CT model of the natural structure is lower than the preset threshold; Condition 3: The increase in the crossover ratio of the CT projection contours of the natural structure before and after iteration is less than a set threshold.

7. The intraoperative navigation method for craniostomy based on curvature guidance and online pose maintenance according to claim 6, characterized in that, After obtaining the point cloud with natural structure completion, the following operations are performed: The observation contour of the natural structure is obtained by completing the point cloud based on the natural structure. If the overlap between the template projection contour of the natural structure and the observation contour, the visibility of the natural structure, or the depth hole ratio is lower than the preset threshold, the pose update of the current frame is frozen. If the pose update is frozen for consecutive frames, the tracking is restarted from the most recent key frame. The keyframe selection must simultaneously meet the following requirements: ICP convergence, stability of residuals and attitude increments, and compliance of contour overlap.

8. The intraoperative navigation method for craniectomy based on curvature guidance and online pose maintenance according to claim 1, characterized in that, Fine registration uses points with robust kernel functions The ICP algorithm, also known as ICP fine registration, is used in the process of ICP fine registration. If the distance between corresponding point pairs is greater than or equal to the distance threshold, the normal angle is greater than or equal to the angle threshold, the depth uncertainty of any point is higher than the preset upper limit, or the corresponding point pair falls in the edge region of the depth hole, then it is determined to be a suspicious corresponding point pair and given a preset low weight in the robust kernel or directly removed.

9. The intraoperative navigation method for craniostomy based on curvature guidance and online pose maintenance according to claim 8, characterized in that, For the points to be filled in, heterogeneous diffusion interpolation is performed to generate a filled point cloud, and fine registration is carried out in a cascaded or alternating manner: In the current pose, several steps are performed to generate the point cloud to be completed and heterogeneous diffusion interpolation is executed. When the effective overlap between the completed point cloud and the preoperative 3D CT model of the natural structure exceeds a threshold, a local ICP fine registration is triggered. That is, based on the results of the previous round of ICP fine registration, the pose obtained at the end of the previous round of ICP fine registration is used as the initial value to perform several iterations of optimization to minimize the point-to-surface residual, thereby further improving the registration accuracy. If the residual increases or the overlap between the template projection contour and the observed contour decreases below the threshold during the ICP fine registration process, the generation of the completed point cloud is paused and the model is reverted to the pose obtained at the end of the previous round of ICP fine registration to ensure overall convergence and robustness.

10. The intraoperative navigation method for craniectomy based on curvature guidance and online pose maintenance according to claim 9, characterized in that, The expression for minimizing the point-to-surface residual during the ICP fine registration process, involving several iterations of optimization, is as follows: in To complete the point cloud with natural structure One point; To complete the point cloud structure The joint weight of each point , , and The first point cloud is the one with natural structure completion. The maximum principal curvature of a point, the angle or difference between the point normal and the average normal of its neighborhood, and the depth uncertainty; It is a function that monotonically increases with respect to curvature, used to enhance the weight of points on high-curvature ridges; and These are functions that monotonically decrease with respect to normal consistency and depth uncertainty, respectively, and are used to suppress points with normal instability or unreliable depth. , and The shape and parameters are calibrated during pre-training; For preoperative 3D CT models containing natural structures in Unit normal vector at the location transpose; For the preoperative 3D CT model of natural structure and Corresponding points; Represents the rotation matrix; This represents the translation vector.

Images