CBCT data and oral scanning data automatic registration method and system based on deep learning
Through deep learning and improved registration algorithms, efficient and low-cost automatic registration of CBCT data and oral scan data is achieved, solving the problems of complex manual point selection and large errors in traditional methods. It is suitable for virtual implantation guide design and occlusion analysis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-09
- Publication Date
- 2026-05-15
AI Technical Summary
Traditional methods for registering CBCT data with intraoral scan data rely on manual point selection, which is costly and difficult to automate efficiently. It also makes feature point matching difficult and results in large errors.
A deep learning segmentation network was used for single-tooth segmentation. An improved 3D YOLOv12 network and point cloud segmentation algorithm were combined to obtain the cusp feature point cloud. Coarse registration was performed using the sample consistency initial registration algorithm, and fine registration was performed using the improved ICP registration algorithm to generate a registered ligament model that is spatially aligned with the CBCT data.
It achieves efficient and accurate automatic registration of CBCT data and oral scan data, reduces operating costs, and improves registration accuracy and consistency. It is suitable for virtual implant guide design and occlusion analysis.
Smart Images
Figure CN122049003A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of automatic data registration technology, specifically to a deep learning-based method for automatic registration of CBCT (Conebeam Computer Tomography) data and oral scan data, and a deep learning-based system for automatic registration of CBCT data and oral scan data. Background Technology
[0002] In dental implant surgery, traditional implant guides, which rely on plaster casts, are prone to risks such as bone fractures and nerve canal perforation due to the inability to accurately capture the jawbone structure, posing significant safety defects. With the development of medical imaging technology, virtual implant guides, by integrating CBCT-reconstructed oral models with dental model data obtained from intraoral scanning, can effectively improve surgical precision and safety.
[0003] In related technologies, registration methods often rely on manual point selection or reference marking, which not only increases operational costs but also makes it difficult to achieve efficient automation. In particular, due to the significant differences in information composition between CBCT data and intraoral scan data—such as the lack of jawbone features in dental model data, the coarse tooth features in CBCT data, and the potential inclusion of interfering structures like the base in intraoral scan data—feature point correspondence becomes difficult in automatic registration, resulting in large errors. Summary of the Invention
[0004] To address the aforementioned technical problems, this invention provides a deep learning-based automatic registration method for CBCT data and oral scan data, which can efficiently and accurately perform automatic registration of CBCT data and oral scan data, and has low operating costs.
[0005] The technical solution adopted in this invention is as follows: An automatic registration method for CBCT data and intraoral scan data based on deep learning includes the following steps: using a deep learning segmentation network to perform single-tooth segmentation on the CBCT data to obtain first-modality single-tooth segmentation data; using a point cloud segmentation algorithm to perform single-tooth segmentation on the intraoral scan data to obtain second-modality single-tooth segmentation data; using an improved 3D YOLOv12 network to obtain the first cusp feature point cloud corresponding to the first-modality single-tooth segmentation data and the second cusp feature point cloud corresponding to the second-modality single-tooth segmentation data; using a sample consistency initial registration algorithm to perform coarse registration on the first cusp feature point cloud and the second cusp feature point cloud and obtain the corresponding initial transformation matrix; using an improved ICP registration algorithm to perform fine registration on the first-modality single-tooth segmentation data and the second-modality single-tooth segmentation data based on the initial transformation matrix and obtain the corresponding target rigid body transformation matrix; and mapping the coordinates corresponding to all vertices of the second-modality single-tooth segmentation data to the CBCT coordinate system according to the target rigid body transformation matrix to generate a registered ligament model aligned with the CBCT data space.
[0006] In one embodiment of the present invention, the target loss function of the improved three-dimensional YOLOv12 network is generated by the following formula: , in, Denotes the target loss function. This represents the localization loss function. This represents the first weighting factor corresponding to the localization loss function. Represents the confidence loss function. This represents the second weighting factor corresponding to the confidence loss function. Represents the category loss function, This represents the third weighting factor corresponding to the category loss function.
[0007] In one embodiment of the present invention, a sample consistency initial registration algorithm is used to coarsely register the first cusp feature point cloud and the second cusp feature point cloud and obtain the corresponding initial transformation matrix. Specifically, this includes: calculating a local geometric feature descriptor for each point in the first cusp feature point cloud and the second cusp feature point cloud; randomly selecting multiple key points in the second cusp feature point cloud and finding corresponding points in the first cusp feature point cloud based on the similarity of the local geometric feature descriptors corresponding to each key point; generating candidate corresponding point pairs based on the key points and the corresponding corresponding points; for each candidate corresponding point pair, solving the rigid estimation transformation matrix from the second cusp feature point cloud to the first cusp feature point cloud using the least squares method; calculating the consistency score of each rigid estimation transformation matrix and selecting the rigid estimation transformation matrix with the highest consistency score as the initial transformation matrix from the second cusp feature point cloud to the first cusp feature point cloud.
[0008] In one embodiment of the present invention, an improved ICP registration algorithm is used to perform fine registration of the first modal single-tooth segmentation data and the second modal single-tooth segmentation data based on the initial transformation matrix, and to obtain the corresponding target rigid body transformation matrix. This includes the following steps: performing an initial transformation on the second tooth cusp feature point cloud using the initial transformation matrix to obtain the initial spatial position of the second tooth cusp feature point cloud initially aligned to the first tooth cusp feature point cloud; searching for the point in the first tooth cusp feature point cloud with the closest Euclidean distance for each point in the second tooth cusp feature point cloud as a matching point; obtaining corresponding weights for each pair of matching points, and using the weighted least squares method to solve for the candidate rigid body transformation matrix from the second tooth cusp feature point cloud to the first tooth cusp feature point cloud based on the weights; iteratively updating the spatial position of the second tooth cusp feature point cloud based on the candidate rigid body transformation matrix, and obtaining the corresponding target rigid body transformation matrix when the iteration converges.
[0009] An automatic registration system for CBCT data and intraoral scan data based on deep learning includes: a first acquisition module, which is used to perform single-tooth segmentation on the CBCT data using a deep learning segmentation network to obtain first modality single-tooth segmentation data, and to perform single-tooth segmentation on the intraoral scan data using a point cloud segmentation algorithm to obtain second modality single-tooth segmentation data; a second acquisition module, which is used to use an improved 3D YOLOv12 network to obtain first cusp feature point clouds corresponding to the first modality single-tooth segmentation data and second cusp feature point clouds corresponding to the second modality single-tooth segmentation data; and a third acquisition module. The system is configured to: 1) perform coarse registration of the first cusp feature point cloud and the second cusp feature point cloud using a sample consistency initial registration algorithm, and obtain the corresponding initial transformation matrix; 2) perform fine registration of the first modal single-tooth segmentation data and the second modal single-tooth segmentation data using an improved ICP registration algorithm based on the initial transformation matrix, and obtain the corresponding target rigid body transformation matrix; and 3) generate a registration ligament model aligned with the CBCT data space, based on the target rigid body transformation matrix, using the target rigid body transformation matrix to map the coordinates corresponding to all vertices of the second modal single-tooth segmentation data to the CBCT coordinate system.
[0010] A computer device includes 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 computer program, it implements the above-mentioned method for automatic registration of CBCT data and oral scan data based on deep learning.
[0011] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned deep learning-based automatic registration method for CBCT data and oral scan data.
[0012] The beneficial effects of this invention are: This invention can efficiently and accurately perform automatic registration of CBCT data and oral scan data, and has low operating costs. Attached Figure Description
[0013] Figure 1 This is a flowchart of an automatic registration method for CBCT data and oral scan data based on deep learning, according to an embodiment of the present invention. Figure 2 This is a block diagram of an automatic registration system for CBCT data and oral scan data based on deep learning, according to an embodiment of the present invention. Detailed Implementation
[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0015] Figure 1 This is a flowchart of an embodiment of the deep learning-based automatic registration method for CBCT data and oral scan data.
[0016] like Figure 1 As shown, the deep learning-based automatic registration method for CBCT data and oral scan data in this embodiment of the invention may include the following steps: S1. A deep learning segmentation network is used to segment CBCT data into individual teeth to obtain the first modality of individual tooth segmentation data. A point cloud segmentation algorithm is used to segment oral scan data into individual teeth to obtain the second modality of individual tooth segmentation data.
[0017] Among them, the deep learning segmentation network can be the nnU-Net network, and the point cloud segmentation algorithm can be the PointNet algorithm.
[0018] S2, using an improved 3D YOLOv12 network, the first cusp feature point cloud corresponding to the first modal single tooth segmentation data and the second cusp feature point cloud corresponding to the second modal single tooth segmentation data are obtained respectively.
[0019] Specifically, firstly, a large number of CBCT and intraoral scan data samples can be collected and labeled, covering different tooth arrangements, missing teeth, metal artifacts, restorations, etc., to ensure the diversity and representativeness of the samples. Three dental experts used 3D annotation tools to annotate the cusps of each tooth in the CBCT volume data and intraoral scan mesh data, ensuring the accurate location of key points. The cusps of each tooth are saved in 3D coordinates and assigned a unique tooth ID for label matching during model training.
[0020] To unify the coordinate system, the oral scan mesh was voxelized. Considering that the YOLOv12 feature point detection algorithm is better suited to voxel data, the input oral scan data (STL or PLY format) was imported into the VTK environment. The vtkPolyDataToImageStencil and vtkImageStencil modules were used to convert the surface mesh data into 3D voxel data. During the voxelization process, the voxel spacing was set to 0.3 mm, and the coordinate system of the origin was unified with that of the CBCT volumetric data, ensuring consistency between the two in physical space coordinates, thus achieving a unified representation of multimodal data. After the conversion, 3D grayscale voxel data with the same size as CBCT voxels was generated, providing an input format for subsequent deep learning models.
[0021] Sample preprocessing and data augmentation were performed on the CBCT and voxelized intraoral scan data, normalizing the grayscale values to the [0, 1] range. Data augmentation techniques were used to expand the samples, including random translation (ranging from -5 to 5 voxels to simulate different head positions in patients), random rotation (ranging from -10° to 10° to enhance the model's robustness to posture changes), adding Gaussian noise to simulate a noisy acquisition environment, and randomly cropping and scaling the data to 128×128×128 voxel blocks to facilitate model training. These preprocessing steps help improve the model's generalization performance and reduce overfitting.
[0022] An improved 3D YOLOv12 network was used to train a tooth cusp detection model, employing the improved YOLOv12 network structure as the 3D feature point detection model. The network consists of three parts: Backbone, Neck, and Head. The Backbone uses a CSPNet structure combined with the RepStage module (RepStageBlock), achieving multi-scale feature extraction and reuse through re-parameterized convolutions. The Neck uses a Bidirectional Feature Pyramid (BiFPN) fusion mechanism to achieve contextual information interaction between features at different scales. The Head introduces a Task-Aligned Assigner dynamic anchor point allocation strategy and a Transformer attention mechanism (CoT Attention) to improve the localization accuracy of keypoint detection.
[0023] For each voxel prediction unit, the network output is defined as follows: , in, Indicates the prediction confidence level. This represents the predicted offset coordinates of the tooth cusp.
[0024] In one embodiment of the present invention, the target loss function of the improved three-dimensional YOLOv12 network is generated by the following formula: , in, Represents the target loss function. This represents the localization loss function. This represents the first weight factor corresponding to the localization loss function. Represents the confidence loss function. This represents the second weighting factor corresponding to the confidence loss function. Represents the category loss function, This represents the third weighting factor corresponding to the categorical loss function.
[0025] In one embodiment of the present invention, the localization loss function can be generated by the following formula. : , in, This represents the square of the Euclidean distance between the predicted and actual locations. This represents the length of the diagonal of the circumscribed cube. and These represent the weights for controlling aspect ratio and consistency, respectively. This indicates intersection, union, and ratio.
[0026] In one embodiment of the present invention, the confidence loss function can be generated by the following formula. : , in, Indicates the number of samples. and Let represent the true label and predicted probability of the i-th sample, respectively.
[0027] In one embodiment of the present invention, the category loss function can be generated by the following formula. : , in, Indicates the predicted probability. and These represent the corresponding adjustment coefficients.
[0028] For model training and validation, the dataset was divided into training, validation, and test sets in a 7:2:1 ratio. During validation, the Mean Euclidean Distance (MED) was used to measure the accuracy of cusp prediction. , Among them, when the validation set MED no longer decreases or the loss converges, the optimal model weights are saved.
[0029] Feature point extraction and initial registration point cloud generation: The trained YOLOv12 feature point detection model was applied to CBCT and intraoral scan data respectively, outputting the 3D coordinates of the cusps for each. Non-maximum suppression (NMS) was used to select the prediction results with the highest confidence.
[0030] in, Point The set of neighborhood voxels.
[0031] Finally, the cusp feature point cloud (i.e., the first modal single tooth segmentation data and the second modal single tooth segmentation data) corresponding to the CBCT data and the oral scan data is formed, which serves as the initial matching input for the subsequent coarse registration algorithm.
[0032] S3. The sampling consistency initial registration algorithm is used to coarsely register the feature point clouds of the first cusp and the feature point clouds of the second cusp, and obtain the corresponding initial transformation matrix.
[0033] In one embodiment of the present invention, a sample consistency initial registration algorithm is used to coarsely register the feature point clouds of the first cusp and the feature point clouds of the second cusp, and to obtain the corresponding initial transformation matrix. Specifically, the method includes the following steps: S31, calculate the local geometric feature descriptor for each point in the first cusp feature point cloud and the second cusp feature point cloud respectively.
[0034] Specifically, a local geometric feature descriptor is calculated for each point in the feature point cloud of the first cusp and the feature point cloud of the second cusp. For example, a fast point feature histogram. From this, the feature set of the first tooth cusp feature point cloud is obtained. and the feature set of the second cusp feature point cloud .
[0035] S32, randomly select multiple key points in the second tooth cusp feature point cloud, and find corresponding points in the first tooth cusp feature point cloud based on the similarity of the local geometric feature descriptors corresponding to each key point, and generate candidate corresponding point pairs based on the key points and the corresponding points.
[0036] Specifically, random sampling matching is used to randomly select several key points from the feature point cloud of the second tooth cusp. Based on the similarity of the feature descriptors, corresponding points are found in the feature point cloud of the first cusp. ,Right now .
[0037] S33. For each candidate corresponding point pair, the least squares method is used to solve the rigid estimation transformation matrix from the feature point cloud of the second cusp to the feature point cloud of the first cusp.
[0038] Specifically, for each candidate corresponding point pair, the least squares method is used to solve for the rigid estimation transformation matrix from the feature point cloud of the second cusp to the feature point cloud of the first cusp, i.e. , in, Represents the rigid estimation transformation matrix. Represents the rotation matrix. This represents the translation vector.
[0039] S34. Calculate the consistency score of each rigid estimation transformation matrix, and select the rigid estimation transformation matrix with the highest consistency score as the initial transformation matrix from the second cusp feature cloud to the first cusp feature cloud.
[0040] Specifically, the consistency score for each rigid estimation transformation matrix is calculated using the following formula: , in, This represents the point after distance transformation in the feature point cloud of the first tooth cusp. The nearest point, This indicates the preset distance threshold. This indicates an indicator function.
[0041] The rigid estimation transformation matrix with the highest consistency score is selected as the initial transformation matrix from the second cusp feature cloud to the first cusp feature cloud, i.e. , in, This represents the initial transformation matrix.
[0042] S4. An improved ICP registration algorithm is used to perform fine registration of the first mode single tooth segmentation data and the second mode single tooth segmentation data based on the initial transformation matrix, and the corresponding target rigid body transformation matrix is obtained.
[0043] Specifically, an improved ICP registration algorithm is used to perform fine registration of the first-mode single-tooth segmentation data and the second-mode single-tooth segmentation data based on the initial transformation matrix, and to obtain the corresponding target rigid body transformation matrix, including the following steps: S41, the initial transformation matrix is used to perform an initial transformation on the feature point cloud of the second cusp, so as to obtain the initial spatial position of the feature point cloud of the second cusp to the feature point cloud of the first cusp.
[0044] Specifically, an initial transformation is performed on the feature point cloud of the second cusp, and the initial transformation matrix is applied to the feature point cloud of the second cusp. For each point, the feature point cloud of the second cusp is initially aligned with the feature point cloud of the first cusp. Initial spatial position: , in, Represents the rotation matrix. This step, which represents the translation vector, ensures that the subsequent fine registration algorithm starts from a better initial position, thus avoiding getting trapped in a local optimum.
[0045] S42, for each point in the feature point cloud of the second cusp, search for the point with the closest Euclidean distance in the feature point cloud of the first cusp as the matching point.
[0046] Specifically, the nearest point correspondence search involves searching for each point in the feature point cloud of the second cusp. Search for the point with the closest Euclidean distance in the feature point cloud of the first tooth cusp as the matching point. .
[0047] S43. For each pair of matching points, obtain the corresponding weights, and use the weighted least squares method to solve the candidate rigid body transformation matrix from the feature point cloud of the second tooth cusp to the feature point cloud of the first tooth cusp based on the weights.
[0048] Specifically, the robustness of the algorithm to noise in CBCT data and missing points in oral scan data is improved by outlier removal and weighting strategies. Specifically, a corresponding weight is introduced for each pair of matching points. ,Right now , in, Indicates the distance threshold. This represents the normal vector consistency threshold. This step effectively removes outliers and noise points, improving the stability and accuracy of point cloud registration.
[0049] Furthermore, the weighted least squares rigid transformation is used to solve for the candidate rigid body transformation matrix from the feature cloud of the second cusp point to the feature cloud of the first cusp point for each matching point pair in each iteration. : , in, Represents the rotation matrix. This represents the translation vector. By using the weighted least squares method, reliable matching points can be fully utilized, while reducing the impact of outliers on the results.
[0050] S44, the spatial position of the feature point cloud of the second tooth tip is iteratively updated according to the candidate rigid body transformation matrix, and the corresponding target rigid body transformation matrix is obtained when the iteration converges.
[0051] Specifically, the spatial position of the feature point cloud of the second tooth cusp is iteratively updated based on the candidate rigid body transformation matrix, i.e. , As the iteration proceeds, the feature point cloud of the second cusp gradually overlaps with the feature point cloud of the first cusp, and the registration error continues to decrease.
[0052] Iterative convergence determination, i.e., calculating the weighted mean square error of the current iteration: , Wherein, when the error change of two consecutive iterations satisfies The algorithm stops when the iteration count reaches a preset maximum value, or when the iteration count reaches a preset maximum value.
[0053] The optimal transformation output, the final iteration yields the target rigid body transformation matrix: , This transformation serves as a precise registration result from the feature point cloud of the second cusp to the feature point cloud of the first cusp. It can be used for subsequent applications such as tooth morphology analysis, virtual crown placement, and occlusal simulation.
[0054] S5. Based on the target rigid transformation matrix, the coordinates corresponding to all vertices of the second modality single-tooth segmentation data are mapped to the CBCT coordinate system to generate a registration ligament model aligned with the CBCT data space.
[0055] Specifically, the process involves acquiring the registered diaphragm data, including rotating and translating all vertices of the intraoral scan data based on the target rigid body transformation matrix obtained in the fine registration stage, mapping their coordinates to the CBCT coordinate system, and generating a registered diaphragm model aligned with the CBCT data space.
[0056] Therefore, this invention, based on the single-tooth segmentation model of nnUNet and PointNet and the YOLOv12 3D feature point detection network, can automatically identify teeth and their key point locations, achieving full automation from feature extraction to registration, significantly reducing manual point selection and calibration operations, and improving efficiency and consistency. By voxelizing the intraoral scan mesh data and strictly aligning the Origin, Spacing, and Direction parameters of the CBCT data, a unified representation of the two modal data in the same physical coordinate system is achieved, significantly reducing registration errors caused by feature differences. A two-stage strategy of "Sampling Consistency Initial Registration (SAC-IA) + Improved ICP Fine Registration" is adopted. SAC-IA achieves global coarse alignment, avoiding getting trapped in local optima; the improved weighted ICP introduces distance and normal vector constraints, effectively eliminating outliers and improving accuracy and stability; through the finally calculated rigid body transformation matrix, the intraoral scan dental membrane model is accurately mapped to the CBCT coordinate system, achieving high-precision spatial alignment of the tooth geometry. The integrated model can be directly used for virtual implant guide design, automatic restoration generation, and occlusal analysis, improving the intelligence level of digital oral diagnosis and treatment.
[0057] In summary, the deep learning-based automatic registration method for CBCT and intraoral scan data according to embodiments of the present invention employs a deep learning segmentation network to segment CBCT data into single teeth to obtain first-modality single-tooth segmentation data, employs a point cloud segmentation algorithm to segment intraoral scan data into single teeth to obtain second-modality single-tooth segmentation data, employs an improved 3D YOLOv12 network to obtain the first cusp feature point cloud corresponding to the first-modality single-tooth segmentation data and the second cusp feature point cloud corresponding to the second-modality single-tooth segmentation data, employs a sample consistency initial registration algorithm to perform coarse registration of the first and second cusp feature point clouds and obtain the corresponding initial transformation matrix, employs an improved ICP registration algorithm to perform fine registration of the first and second-modality single-tooth segmentation data based on the initial transformation matrix and obtain the corresponding target rigid body transformation matrix, and maps the coordinates corresponding to all vertices of the second-modality single-tooth segmentation data to the CBCT coordinate system according to the target rigid body transformation matrix to generate a registered ligament model aligned with the CBCT data space. Therefore, it can efficiently and accurately perform automatic registration of CBCT data and oral scan data, and the operating cost is low.
[0058] Corresponding to the above embodiments of the automatic registration method for CBCT data and oral scan data based on deep learning, the present invention also proposes an automatic registration system for CBCT data and oral scan data based on deep learning.
[0059] like Figure 2As shown, the CBCT data and oral scan data automatic registration system based on deep learning in this embodiment of the invention may include: a first acquisition module 100, a second acquisition module 200, a third acquisition module 300, a fourth acquisition module 400, and a generation module 500.
[0060] The first acquisition module 100 is used to perform single-tooth segmentation on CBCT data using a deep learning segmentation network to obtain first-modality single-tooth segmentation data, and to perform single-tooth segmentation on intraoral scan data using a point cloud segmentation algorithm to obtain second-modality single-tooth segmentation data; the second acquisition module 200 is used to use an improved 3D YOLOv12 network to obtain the first cusp feature point cloud corresponding to the first-modality single-tooth segmentation data and the second cusp feature point cloud corresponding to the second-modality single-tooth segmentation data, respectively; the third acquisition module 300 is used to perform coarse registration on the first cusp feature point cloud and the second cusp feature point cloud using a sample consistency initial registration algorithm, and obtain the corresponding initial transformation matrix; the fourth acquisition module 400 is used to perform fine registration on the first-modality single-tooth segmentation data and the second-modality single-tooth segmentation data based on the initial transformation matrix using an improved ICP registration algorithm, and obtain the corresponding target rigid body transformation matrix; the generation module 500 is used to map the coordinates corresponding to all vertices of the second-modality single-tooth segmentation data to the CBCT coordinate system according to the target rigid body transformation matrix, so as to generate a registered ligament model aligned with the CBCT data space.
[0061] In one embodiment of the present invention, the target loss function of the improved three-dimensional YOLOv12 network is generated by the following formula: , in, Represents the target loss function. This represents the localization loss function. This represents the first weight factor corresponding to the localization loss function. Represents the confidence loss function. This represents the second weighting factor corresponding to the confidence loss function. Represents the category loss function, This represents the third weighting factor corresponding to the categorical loss function.
[0062] In one embodiment of the present invention, the third acquisition module 300 is specifically configured to: calculate local geometric feature descriptors for each point in the first cusp feature point cloud and the second cusp feature point cloud; randomly select multiple key points in the second cusp feature point cloud, and find corresponding points in the first cusp feature point cloud based on the similarity of the local geometric feature descriptors corresponding to each key point; generate candidate corresponding point pairs based on the key points and the corresponding corresponding points; for each candidate corresponding point pair, solve the rigid estimation transformation matrix from the second cusp feature point cloud to the first cusp feature point cloud using the least squares method; calculate the consistency score of each rigid estimation transformation matrix, and select the rigid estimation transformation matrix with the highest consistency score as the initial transformation matrix from the second cusp feature point cloud to the first cusp feature point cloud.
[0063] In one embodiment of the present invention, the fourth acquisition module 400 is specifically used to: perform an initial transformation on the second cusp feature point cloud using an initial transformation matrix to obtain the initial spatial position of the second cusp feature point cloud initially aligned to the first cusp feature point cloud; search for the point in the first cusp feature point cloud with the closest Euclidean distance for each point in the second cusp feature point cloud as a matching point; obtain the corresponding weight for each pair of matching points, and solve the candidate rigid body transformation matrix from the second cusp feature point cloud to the first cusp feature point cloud using the weighted least squares method according to the weight; iteratively update the spatial position of the second cusp feature point cloud according to the candidate rigid body transformation matrix, and obtain the corresponding target rigid body transformation matrix when the iteration converges.
[0064] It should be noted that more specific embodiments of the deep learning-based automatic registration system for CBCT data and oral scan data of the present invention can be found in the embodiments of the deep learning-based automatic registration method for CBCT data and oral scan data described above. To avoid redundancy, they will not be described in detail here.
[0065] According to an embodiment of the present invention, an automatic registration system for CBCT data and intraoral scan data based on deep learning comprises the following steps: a first acquisition module uses a deep learning segmentation network to segment CBCT data into single teeth to obtain first modal single-tooth segmentation data, and a point cloud segmentation algorithm to segment intraoral scan data into single teeth to obtain second modal single-tooth segmentation data; a second acquisition module uses an improved 3D YOLOv12 network to obtain first cusp feature point clouds corresponding to the first modal single-tooth segmentation data and second cusp feature point clouds corresponding to the second modal single-tooth segmentation data; a third acquisition module uses a sample consistency initial registration algorithm to perform coarse registration of the first and second cusp feature point clouds and obtain the corresponding initial transformation matrix; a fourth acquisition module uses an improved ICP registration algorithm based on the initial transformation matrix to perform fine registration of the first and second modal single-tooth segmentation data and obtain the corresponding target rigid body transformation matrix; and a generation module maps the coordinates of all vertices of the second modal single-tooth segmentation data to the CBCT coordinate system according to the target rigid body transformation matrix to generate a registered dental lamina model aligned with the CBCT data space. Therefore, it can efficiently and accurately perform automatic registration of CBCT data and oral scan data, and the operating cost is low.
[0066] Corresponding to the above embodiments, the present invention also proposes a computer device.
[0067] The computer device of this invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the deep learning-based automatic registration method for CBCT data and oral scan data described in the above embodiments.
[0068] The computer device according to embodiments of the present invention can efficiently and accurately perform automatic registration of CBCT data and oral scan data, and has low operating costs.
[0069] Corresponding to the above embodiments, the present invention also proposes a non-transitory computer-readable storage medium.
[0070] The non-transitory computer-readable storage medium of this invention stores a computer program that, when executed by a processor, implements the above-described method for automatic registration of CBCT data and oral scan data based on deep learning.
[0071] The non-transitory computer-readable storage medium according to embodiments of the present invention can efficiently and accurately perform automatic registration of CBCT data and oral scan data, and has low operating costs.
[0072] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. "A plurality of" means two or more, unless otherwise explicitly specified.
[0073] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0074] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.
[0075] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0076] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.
[0077] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1. A method for automatic registration of CBCT data and intraoral scan data based on deep learning, characterized in that, Includes the following steps: A deep learning segmentation network is used to perform single-tooth segmentation on the CBCT data to obtain first-modality single-tooth segmentation data, and a point cloud segmentation algorithm is used to perform single-tooth segmentation on the intraoral scan data to obtain second-modality single-tooth segmentation data. An improved 3D YOLOv12 network was used to obtain the first cusp feature point cloud corresponding to the first modal single tooth segmentation data and the second cusp feature point cloud corresponding to the second modal single tooth segmentation data. The sample consistency initial registration algorithm is used to coarsely register the feature point clouds of the first cusp and the second cusp, and the corresponding initial transformation matrix is obtained. An improved ICP registration algorithm is used to perform fine registration of the first modal single-tooth segmentation data and the second modal single-tooth segmentation data based on the initial transformation matrix, and the corresponding target rigid body transformation matrix is obtained. Based on the target rigid transformation matrix, the coordinates corresponding to all vertices of the second modality single-tooth segmentation data are mapped to the CBCT coordinate system to generate a registered ligament model aligned with the CBCT data space.
2. The method for automatic registration of CBCT data and intraoral scan data based on deep learning according to claim 1, characterized in that, The target loss function of the improved 3D YOLOv12 network is generated using the following formula: ， in, Denotes the target loss function. This represents the localization loss function. This represents the first weighting factor corresponding to the localization loss function. Represents the confidence loss function. This represents the second weighting factor corresponding to the confidence loss function. Represents the category loss function, This represents the third weighting factor corresponding to the category loss function.
3. The method for automatic registration of CBCT data and intraoral scan data based on deep learning according to claim 2, characterized in that, A sample consistency initial registration algorithm is used to coarsely register the feature point clouds of the first and second cusps, and to obtain the corresponding initial transformation matrix, specifically including: For each point in the first cusp feature point cloud and the second cusp feature point cloud, calculate the local geometric feature descriptor; Multiple key points are randomly selected in the second tooth cusp feature point cloud, and corresponding points are found in the first tooth cusp feature point cloud based on the similarity of the local geometric feature descriptors corresponding to each key point. Candidate corresponding point pairs are generated based on the key points and the corresponding points. For each candidate corresponding point pair, the least squares method is used to solve the rigid estimation transformation matrix from the feature point cloud of the second cusp to the feature point cloud of the first cusp. Calculate the consistency score for each rigid estimation transformation matrix, and select the rigid estimation transformation matrix with the highest consistency score as the initial transformation matrix from the second cusp feature point cloud to the first cusp feature point cloud.
4. The method for automatic registration of CBCT data and intraoral scan data based on deep learning according to claim 3, characterized in that, An improved ICP registration algorithm is used to perform fine registration of the first modal single-tooth segmentation data and the second modal single-tooth segmentation data based on the initial transformation matrix, and the corresponding target rigid body transformation matrix is obtained. This includes the following steps: An initial transformation matrix is used to perform an initial transformation on the second cusp feature point cloud to obtain the initial spatial position of the second cusp feature point cloud initially aligned with the first cusp feature point cloud; For each point in the second cusp feature point cloud, the point with the closest Euclidean distance in the first cusp feature point cloud is searched as the matching point; For each pair of matching points, obtain the corresponding weights, and use the weighted least squares method to solve the candidate rigid body transformation matrix from the feature point cloud of the second tooth cusp to the feature point cloud of the first tooth cusp based on the weights. The spatial position of the second tooth cusp feature point cloud is iteratively updated based on the candidate rigid body transformation matrix, and the corresponding target rigid body transformation matrix is obtained when the iteration converges.
5. A deep learning-based automatic registration system for CBCT data and intraoral scan data, characterized in that, include: The first acquisition module is used to perform single-tooth segmentation on the CBCT data using a deep learning segmentation network to obtain first-modality single-tooth segmentation data, and to perform single-tooth segmentation on the oral scan data using a point cloud segmentation algorithm to obtain second-modality single-tooth segmentation data. The second acquisition module is used to acquire the first cusp feature point cloud corresponding to the first modal single tooth segmentation data and the second cusp feature point cloud corresponding to the second modal single tooth segmentation data using an improved three-dimensional YOLOv12 network. The third acquisition module is used to perform coarse registration of the first cusp feature point cloud and the second cusp feature point cloud using the sample consistency initial registration algorithm, and to acquire the corresponding initial transformation matrix. The fourth acquisition module is used to perform fine registration of the first modal single tooth segmentation data and the second modal single tooth segmentation data based on the initial transformation matrix using an improved ICP registration algorithm, and to acquire the corresponding target rigid body transformation matrix. The generation module is used to map the coordinates corresponding to all vertices of the second modal single-tooth segmentation data to the CBCT coordinate system according to the target rigid transformation matrix, so as to generate a registered ligament model aligned with the CBCT data space.
6. 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 computer program, it implements the method for automatic registration of CBCT data and oral scan data based on deep learning according to any one of claims 1-4.
7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the deep learning-based automatic registration method for CBCT data and oral scan data according to any one of claims 1-4.