Calcaneus fracture area automatic alignment method based on point cloud registration
By constructing a point cloud registration neural network model, the problems of low efficiency and insufficient accuracy in calcaneal fissure region alignment in traditional methods are solved. This achieves highly automated and robust automatic alignment of calcaneal fissure regions, with strong adaptability, accurate capture of geometric relationships, and resistance to noise interference.
Patent Information
- Application Number
- CN202511015461.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-23
- Publication Date
- 2025-11-07
Smart Images
Figure CN120899386A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of medical image processing, and particularly relates to a calcaneus fracture region automatic alignment method based on point cloud registration. BACKGROUND
[0002] With the rapid development of three-dimensional reconstruction technology, point cloud data has become the main data format for representing the three-dimensional world. In medical diagnosis and treatment, accurate modeling of three-dimensional bone structures has important value for preoperative planning and postoperative evaluation. Traditional methods usually perform three-dimensional reconstruction based on CT or MRI images and use image-level or voxel-level methods to detect structural fractures. However, with the advancement of point cloud processing technology, using point clouds to directly perform three-dimensional analysis of bone bodies and fractures has become a new trend. Point cloud registration technology is one of the core tasks in the fields of computer vision and three-dimensional reconstruction, and its goal is to align two or more point cloud data sets through rigid transformation. Early point cloud registration methods mainly relied on manual registration or instrument-dependent registration, and these methods were not only inefficient but also difficult to guarantee the accuracy of registration. With the development of computer technology and algorithms, automatic registration technology has gradually become mainstream. Automatic registration technology uses certain algorithms or statistical laws to calculate the misalignment between two point clouds using a computer, thereby achieving automatic registration of the two point clouds. These algorithms can be roughly divided into two categories: one is a feature matching-based registration method, such as RANSAC and 4PCS, which extracts feature descriptors from point clouds to find corresponding relationships between point clouds; the other is an iterative optimization-based registration method, such as the Iterative Closest Point (ICP) and Normal Distribution Transform (NDT), which iteratively solves the transformation parameters between corresponding point pairs to achieve accurate alignment of point clouds.
[0003] In recent years, the success of deep learning technology has greatly improved the robustness and efficiency of point cloud registration. Deep learning-based point cloud registration methods use deep learning models to extract features from original point clouds to obtain initial configuration values of point clouds, and then further complete fine registration according to the feature values. These methods can learn higher-level features to achieve higher registration accuracy and robustness. In the task of automatic alignment of calcaneus fracture regions, deep learning-based point cloud registration methods can automatically extract features of calcaneus bone bodies and fracture point clouds, and predict rigid transformation parameters through network models, thereby achieving automatic alignment of fracture point clouds and bone body point clouds.
[0004] In the treatment of calcaneal fractures, accurate alignment of the fracture region is crucial for restoring the anatomical structure and stability of the calcaneus. Traditional methods of aligning calcaneal fractures often rely on the experience and manual operation of doctors, which not only is inefficient, but also is difficult to guarantee the accuracy of alignment. Therefore, it is of great application value to develop an automatic alignment method for calcaneal fracture regions based on point cloud registration. By automatically aligning the calcaneal fracture region, doctors can more accurately determine the position, shape and severity of the fracture, so as to develop a more reasonable treatment plan. In addition, the automatic alignment method can also provide an accurate three-dimensional calcaneus model for subsequent treatment such as rehabilitation training, further improving the treatment effect and the quality of patient rehabilitation. SUMMARY
[0005] The application provides an automatic alignment method for calcaneal fracture regions based on point cloud registration to solve the technical problem that there are many uncontrollable factors in the actual scanning process, so that the obtained fracture region point cloud data often has position deviation from the original bone point cloud.
[0006] To solve the above technical problems, one technical solution adopted by the application is: an automatic alignment method for calcaneal fracture regions based on point cloud registration, comprising:
[0007] S1. Based on the CT image of the calcaneus, the bone point cloud and the fracture point cloud of the calcaneal bone region and the fracture region are extracted and preprocessed to obtain preprocessed data;
[0008] S2. Construct a point cloud registration neural network model, and perform deep coding and global aggregation on the preprocessed data to obtain a global feature vector;
[0009] S3. Based on a multi-layer fully connected network, perform nonlinear transformation on the global feature vector to output predicted rigid transformation parameters;
[0010] S4. Based on the rigid transformation parameters, perform confidence analysis to output a confidence interval and improve the credibility of the point cloud registration neural network model.
[0011] Further, the method of step S1 comprises:
[0012] S11. Divide the bone point cloud and the fracture point cloud into three-dimensional grid voxels, and retain the representative points of the voxels;
[0013] S12. Calculate the Euclidean distance mean value between each representative point and a plurality of nearest neighbors, wherein the representative points with a Euclidean distance mean value greater than the global mean value are removed;
[0014] S13. Perform random rigid transformation on the bone point cloud and the fracture point cloud to obtain preprocessed data including perturbed point clouds.
[0015] Further, the method of step S13 comprises:
[0016] Based on formulas (1)-(2), the random rigid body transformation is performed; wherein formulas (1)-(2) are:
[0017] P' frac,aug ={R rand ·q+t rand ∣q∈P' frac} (1);
[0018] Wherein, R∈SO(3) represents a three-dimensional rotation matrix, satisfying And det(R rand )=1, which can be generated by Euler angle rotation:
[0019] R rand =R z (θ z )·R y (θ y )·R x (θ x ) (2);
[0020] Wherein, θ x ,θ y , are the rotation angles around the x, y, z axes respectively. t rand =[t x ,t y ,t z ] T ∈R 3 , wherein t x ,t y , represent three-dimensional translation perturbations.
[0021] Further, the method of step S2 comprises:
[0022] Based on formula (3), the feature mapping is performed on the perturbed bone point cloud and the crack point cloud respectively, to obtain local point features; wherein formula (3) is:
[0023] F bone =φ(P bone )∈R N×d ,F frac =φ(P' frac,aug )∈R M×d (3);
[0024] Wherein, φ represents a point cloud feature encoding function based on PointNet++, and d is the high-dimensional embedding feature dimension of each point.
[0025] Based on formulas (4)-(5) and the farthest point sampling algorithm, local point features are integrated into a global feature vector using a global max-pooling operation; wherein formulas (4)-(5) are:
[0026] f i =MLP(p i ),f j =MLP(q j ) (4);
[0027]
[0028] wherein g bone represents a global feature vector of the bone point cloud, and g frac represents a global feature vector of the crack point cloud.
[0029] Further, the method of step S3 comprises:
[0030] Based on formula (6), a nonlinear transformation is performed to obtain the rigid transformation parameters; wherein formula (6) is:
[0031]
[0032] wherein the translation vector t pred ∈R 3 , q∈P frac,aug is any point in the perturbed point cloud; the three-dimensional rotation matrix R pred ∈SO(3); is the transpose of the rotation matrix, corresponding to its inverse matrix.
[0033] Further, the method of step S4 comprises:
[0034] Based on formulas (7)-(9), confidence analysis is performed to output a confidence interval; wherein formulas (7)-(9) are:
[0035]
[0036] wherein, and are the predicted values of the rigid transformation; and are the mean and covariance estimates; Σ t is the prediction variance.
[0037] The beneficial effects of the present application are: the present application realizes automatic alignment of calcaneal crack regions based on point cloud registration, has high automation degree, greatly reduces manual intervention, has good registration accuracy and robustness, can accurately capture geometric relationships and effectively resist noise interference, has strong adaptability, does not depend on specific device parameters and can adapt to various crack situations, and the data-driven characteristics support continuous optimization. BRIEF DESCRIPTION OF DRAWINGS
[0038] Figure 1 is a flowchart of an embodiment of the automatic alignment method for the calcaneus fracture area based on point cloud registration of the present application;
[0039] Figure 2 is a schematic diagram of the original bone body and fracture mark of an embodiment of the automatic alignment method for the calcaneus fracture area based on point cloud registration of the present application;
[0040] Figure 3 is a flowchart of an embodiment of step S1 in Figure 1
[0041] Figure 4 is the perturbed fracture area and bone body input graph of an embodiment of the automatic alignment method for the calcaneus fracture area based on point cloud registration of the present application;
[0042] Figure 5 is a network structure diagram of an embodiment of the automatic alignment method for the calcaneus fracture area based on point cloud registration of the present application. DETAILED DESCRIPTION
[0043] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with specific embodiments.
[0044] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other ways different from the description, therefore, the present application is not limited to the specific embodiments disclosed in the following description.
[0045] Reference is made to Figure 1 , Figure 1 is a flowchart of an embodiment of the automatic alignment method for the calcaneus fracture area based on point cloud registration of the present application. The method comprises the following steps:
[0046] S1. Based on the calcaneus CT image, extract the bone body point cloud and fracture point cloud of the calcaneus bone body area and fracture area and perform preprocessing to obtain the preprocessed data.
[0047] Specifically, due to the high cost of high-quality three-dimensional medical image labeling, the amount of calcaneus CT data actually available for training is very limited, especially in the fine labeling of the fracture area, the data scarcity problem is particularly prominent. The present application proposes an enhancement strategy for the fracture area point cloud to improve the robustness and adaptability of the model to the spatial distribution changes of the fracture.
[0048] Specifically, the three-dimensional point cloud of the calcaneus bone body area and the fracture area is extracted from the calcaneus CT image, Figure 2 The original bone body and crack indication provided in the present application are in the same spatial coordinate system. The bone body point cloud and the crack point cloud extracted are respectively down-sampled based on a voxel grid to unify the point cloud density, and outliers are removed to filter out abnormal noise points, thereby ensuring data quality. On this basis, in order to simulate the natural change of the position and posture of the crack in the real scene and enhance the diversity of the training data, a random rigid transformation T rand = (R rand ,t rand ) is applied to the crack point cloud, to obtain the perturbed crack point cloud. The random rigid transformation satisfies the mathematical constraints of a three-dimensional rotation matrix R rand ∈ SO(3) and a translation vector t rand ∈ R 3 , ensuring that the transformed point cloud only undergoes translation and rotation, and the shape remains unchanged.
[0049] Referring to Figure 3 , the method of step S1 comprises:
[0050] S11. The bone body point cloud and the crack point cloud are divided into three-dimensional grid voxels, and the representative points of the voxels are retained.
[0051] Specifically, three-dimensional point cloud data of the bone body region and the crack region are extracted from the medical image, and are respectively denoted as the bone body point cloud P and the crack point cloud P . Both are in the same spatial coordinate system and have unified geometric dimensions and coordinate references. To improve the accuracy and efficiency of point cloud registration, the bone body point cloud and the crack point cloud are first preprocessed. The original point cloud is divided into 1mm×1mm×1mm three-dimensional grid voxels. For each voxel, only one representative point in the voxel is retained, thereby realizing point cloud density standardization, and the processed point cloud is denoted as P' bone ,P' frac .
[0052] S12. The Euclidean distance mean between each representative point and a plurality of nearest neighbors is calculated, wherein the representative points with a Euclidean distance mean greater than the global mean are removed.
[0053] Specifically, the Euclidean distance mean between each point and its k nearest neighbors is calculated Points greater than the global mean μ+ασ (where μ is the global mean, σ is the standard deviation, and α is the control factor) are regarded as noise points and removed, to obtain a high-quality point cloud with clear structure and no outliers.
[0054] S13. The bone body point cloud and the crack point cloud are subjected to a random rigid transformation, to obtain preprocessed data including the perturbed point cloud.
[0055] Specifically, to further enrich the data diversity and enhance the generalization ability of the model to the position and attitude changes of the crack, a random rigid transformation T rand =(R rand ,t rand ) is applied to the data augmentation. The random rigid transformation T rand =(R rand ,t rand ) is applied to the crack point cloud P' frac , and a perturbed point cloud is generated:
[0056] P' frac,aug ={R rand ·q+t rand |q∈P' frac} (1).
[0057] wherein R∈SO(3) represents a three-dimensional rotation matrix, satisfies and det(R rand )=1, and can be generated by Euler angle rotation:
[0058] R rand =R z (θ z )·R y (θ y )·R x (θ x ) (2).
[0059] wherein θ x ,θ y , are rotation angles around the x, y and z axes respectively. t rand =[t x ,t y ,t z ] T ∈R 3 , wherein t x ,t y , represent three-dimensional translation perturbations. This constraint ensures that the transformed crack point cloud only undergoes translation and rotation in space, and its shape and geometric features remain unchanged, thereby preserving the essential features of the crack while simulating different positions and directions of the crack in the actual calcaneus, providing more diverse training samples for the model.
[0060] This transformation ensures that the point cloud remains rigidly consistent after perturbation, that is, the Euclidean distance between any two points q1,q2∈P' frac remains unchanged before and after transformation:
[0061] ||q1-q2||=||R rand q1+t rand -(Rrand q2+t rand )||=||R rand (q1-q2)||;
[0062] The application relates to a perturbation process for constructing training paired samples (bone point cloud and perturbed crack point cloud), providing a supervision signal for a subsequent point cloud registration model, and driving the model to learn the rigid transformation parameters required for restoring the real spatial position.
[0063] S2. Constructing a point cloud registration neural network model, deeply encoding and globally aggregating the preprocessed data to obtain a global feature vector.
[0064] Specifically, the perturbed crack point cloud and the corresponding bone point cloud are input into the point cloud registration neural network model.
[0065] Figure 4 The application provides a perturbed crack region and bone input graph. The network architecture combines deep learning theory and point cloud processing technology, mainly including a feature extraction module and a transformation regression module. The application adopts a PointNet++ encoder as the feature extraction module, deeply encodes the point cloud through a multilayer perceptron and a hierarchical feature aggregation operation, and extracts high-dimensional spatial geometric features. Subsequently, the maximum pooling operation is used to globally aggregate the encoded features to generate a representative global feature vector, which provides rich feature information for subsequent transformation regression.
[0066] S21. The perturbed crack point cloud P' frac,aug and the corresponding bone point cloud P bone are input into the point cloud registration neural network model to predict the rigid transformation relationship between the two, and realize accurate registration. The provided network uses the same set of weight encoders to respectively map the features of the two groups of point clouds:
[0067] F bone =φ(P bone )∈R N×d ,F frac =φ(P' frac,aug )∈R M×d (3);
[0068] Wherein, φ represents a point cloud feature encoding function based on PointNet++, and d is the high-dimensional embedding feature dimension of each point.
[0069] S22. Using a PointNet++ feature encoder to extract global features of the calcaneus bone point cloud and the crack point cloud, respectively.
[0070] Figure 5The network structure diagram provided for the present application. In the PointNet++ structure, the feature extraction process mainly includes point-level feature encoding and global feature aggregation. The Farthest Point Sampling (FPS) algorithm and the BallQuery method are used to act on each input point through a Multilayer Perceptron (MLP); the local point features are integrated into a global feature vector by using a Maxpooling operation. The calculation process is as follows:
[0071] f i =MLP(p i ),f j =MLP(q j ) (4);
[0072]
[0073] wherein g bone represents the global feature vector of the bone point cloud, and g frac represents the global feature vector of the crack point cloud.
[0074] S3. Based on a multi-layer fully connected network, the global feature vector is subjected to nonlinear transformation, and the predicted rigid transformation parameters are output.
[0075] Specifically, the transformation regression module of the network subjects the feature vector to nonlinear transformation through a multi-layer fully connected network, and outputs the predicted rigid transformation parameters, including a rotation matrix R pred and a translation vector t pred .
[0076] The transformation regression module outputs the predicted rigid transformation parameters, including a three-dimensional rotation matrix R pred ∈SO(3) and a translation vector t pred ∈R 3 . These two parameters are used for inverse transformation operation on the perturbed crack point cloud, so as to make it regress to the correct position in the original space. For any point q∈P frac,aug in the perturbed point cloud, the coordinate calculation formula after the inverse operation of the predicted transformation is:
[0077]
[0078] wherein, is the transpose of the rotation matrix, corresponding to its inverse matrix.
[0079] S4. Based on the rigid transformation parameters, confidence analysis is performed, and a confidence interval is output, so as to improve the reliability of the point cloud registration neural network model.
[0080] Specifically, the perturbed crack point cloud is applied to the prediction transformation for inverse transformation operation to return to the original spatial position. An uncertainty estimation mechanism is introduced while outputting the rotation matrix and translation vector, and confidence analysis is performed on the prediction results of the transformation parameters through Bayesian regression to output a confidence interval to provide an auxiliary decision basis and improve the credibility of the model.
[0081] To improve the reliability and credibility of model prediction in medical applications, the Monte Carlo Dropout (MC Dropout) method is used to introduce an uncertainty estimation module while outputting the transformation parameters. The Dropout layer is retained during prediction, so that the network presents different sub-network structures during each forward propagation, thereby generating different prediction results.
[0082] The input point cloud (P bone ,P frac,aug ) is repeatedly subjected to K times of forward inference to obtain a group of rigid transformation prediction values:
[0083]
[0084] Mean and covariance estimation is performed on the rotation and translation vectors, respectively:
[0085]
[0086] The corresponding prediction uncertainty can be measured by variance:
[0087]
[0088] Based on the variance results, a 95% confidence interval is further calculated to represent the uncertainty interval range of the current prediction results.
[0089] The present application realizes automatic alignment of calcaneal crack regions based on point cloud registration, has high automation, greatly reduces manual intervention, has good registration accuracy and robustness, can accurately capture geometric relationships and effectively resist noise interference, has strong adaptability and does not depend on specific device parameters and can adapt to various crack conditions, and the data-driven characteristics support continuous optimization.
[0090] The above only describes the embodiments of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.
Claims
1. A method for automatic alignment of calcaneus fracture area based on point cloud registration, characterized in that, The method comprises the following steps: S1. Based on the CT image of the calcaneus, the bone body point cloud and the crack point cloud of the calcaneus bone body region and the crack region are extracted and preprocessed to obtain preprocessed data; S2. Construct a point cloud registration neural network model, deeply encode and globally aggregate the preprocessed data to obtain a global feature vector; S3. Based on a multi-layer fully connected network, the global feature vector is nonlinearly transformed to output the predicted rigid transformation parameters; S4. Based on the rigid transformation parameters, confidence analysis is performed to output the confidence interval, and the credibility of the point cloud registration neural network model is improved.
2. The method of claim 1, wherein, The method of step S1 comprises: S11. The bone body point cloud and the crack point cloud are divided into three-dimensional grid voxels, and the representative points of the voxels are retained; S12. Calculate the Euclidean distance mean value between each representative point and a plurality of nearest neighbors, wherein the representative points with a Euclidean distance mean value greater than the global mean value are removed; S13. Randomly rigidly transform the bone body point cloud and the crack point cloud to obtain the preprocessed data including the perturbed point cloud.
3. The method of claim 2, wherein, The method of step S13 comprises: Based on formulas (1)-(2), the random rigid transformation is performed; wherein the formulas (1)-(2) are: P' frac,aug = {R rand ·q + t rand | q e P' frac} (1) where R ∈ SO(3) represents a 3D rotation matrix satisfying and det(R rand ) = 1, can be generated by Euler angles rotation: R rand = R z (θ z ) · R y (θ y ) · R x (θ x ) (2); where θ x ,θ y , are the rotation angles around the x, y, z axes, respectively; t rand = [t x ,t y ,t z ] T ∈ R 3 , where t x ,t y , denote three-dimensional translational disturbances.
4. The method of claim 3, wherein, The method of step S2 comprises: Based on formula (3), the local point features are obtained by respectively performing feature mapping on the perturbed bone body point cloud and crack point cloud; wherein formula (3) is: F bone = φ(P bone ) ∈ R N×d ,F frac = φ(P' frac,aug ) ∈ R M×d (3) Wherein, φ represents the point cloud feature encoding function based on PointNet++, and d is the high-dimensional embedding feature dimension of each point. Based on formulas (4)-(5) and the farthest point sampling algorithm, the local point features are integrated into the global feature vector using global maximum pooling operation; wherein the formulas (4)-(5) are: f i = MLP(p i ),f j = MLP(q j ) (4); where g bone represents the global feature vector of the bone point cloud, g frac represents the global feature vector of the crack point cloud.
5. The method of claim 4, wherein, The method of step S3 comprises: Based on formula (6), nonlinear transformation is performed to obtain the rigid transformation parameters; wherein the formula (6) is: where t is the translation vector pred ∈R 3 ,q∈P frac,aug is any point in the perturbed point cloud; R pred ∈SO(3) is a three-dimensional rotation matrix. is the transpose of the rotation matrix, corresponding to its inverse matrix.
6. The method of claim 5, wherein, The method of step S4 comprises: Based on formulas (7)-(9), the confidence analysis is performed to output the confidence interval; wherein the formulas (7)-(9) are: where and is the rigid body transformation prediction; and is the mean and covariance estimate;∑ t is the prediction variance.