Dental three-dimensional identification and model generation method based on CBCT and deep learning

Abstract

The invention relates to the technical field of digital medical treatment, and discloses a dental three-dimensional identification and model generation method based on CBCT and deep learning, and the method comprises the steps: carrying out the preprocessing of obtained CBCT image data; the preprocessed CBCT image data is segmented, and the segmentation step comprises the substeps that a tooth three-dimensional model, an alveolar bone three-dimensional model and a mandibular neural tube three-dimensional model are obtained by adopting a two-stage cascade deep learning network; based on the tooth three-dimensional model and the alveolar bone three-dimensional model, obtaining a periodontal membrane three-dimensional model through geometric and Boolean operation; a tooth root area in the tooth three-dimensional model is positioned, and a self-adaptive threshold method is adopted for CBCT image data corresponding to the tooth root area, and a root canal three-dimensional model is recognized and generated; based on at least one of the obtained three-dimensional models, a personalized clinical application model is generated, high-precision automatic segmentation of the dental complex anatomical structure is achieved, and a reliable digital model basis is provided for personalized precision medical treatment.

Description

Technical Field

[0001] This invention relates to the field of digital medical technology, specifically to a method for dental 3D recognition and model generation based on CBCT and deep learning. Background Technology

[0002] With the widespread adoption of cone-beam computed tomography (CBCT) technology in dental clinics, precise diagnosis, surgical planning, and personalized medical device design based on high-resolution three-dimensional image data have become a trend in modern precision dentistry. The core of this process is to reconstruct discrete two-dimensional CBCT image sequences into three-dimensional digital models that accurately reflect the individual anatomy of the patient, such as teeth, alveolar bone, and mandibular canal.

[0003] However, accurately segmenting various complex dental and jaw anatomy structures from CBCT images is a key bottleneck in realizing digital diagnostic and treatment workflows. Existing technologies have significant limitations in handling this task. On the one hand, due to the small grayscale differences and blurred boundaries between teeth and surrounding alveolar bone, dentin and root canals, and the extremely thin periodontal ligament, traditional image segmentation algorithms (such as thresholding and region growing methods) often fail to achieve ideal accuracy. This typically requires dentists or technicians to invest considerable time in tedious and repetitive manual or semi-automatic sketching and correction, which is not only inefficient but also highly dependent on the operator's experience, lacking consistency and repeatability. Although there have been attempts in recent years to use deep learning methods for automatic segmentation, single network models often struggle to simultaneously achieve the same level of accuracy in recognizing large-scale structures like teeth and alveolar bone as well as small, low-contrast structures like root canals and periodontal ligaments, resulting in segmentation results that are inconsistent and fail to meet clinical requirements.

[0004] Furthermore, existing digital workflows are often fragmented and non-integrated. Clinicians or technicians often need to perform tedious data import, export, and format conversion between multiple independent software systems. For example, preliminary segmentation might be completed in one software, while implant guides or orthodontic archwires are designed in another CAD (computer-aided design) software. This non-integrated workflow not only prolongs the overall design cycle and increases operational complexity, but also easily leads to problems such as file incompatibility or loss of accuracy during data exchange, introducing potential error risks into the entire diagnostic and treatment process.

[0005] On the other hand, the accuracy and safety of downstream clinical application models (such as implant guides, orthodontic archwires, and surgical planning models) depend entirely on the accuracy of upstream anatomical segmentation. If there are deviations in the initial segmentation of key anatomical models such as teeth, alveolar bone, and nerve canals, then any clinical application model generated based on them will be inaccurate. This will not only affect the treatment outcome, but may even lead to serious medical accidents during implant or extraction surgery due to incorrect guide positioning, resulting in damage to important nerves and blood vessels.

[0006] Therefore, there is an urgent need in this field for a method to generate dental 3D models that can achieve automated, high-precision, and integrated processing, in order to overcome the shortcomings of existing technologies and improve the efficiency and quality of digital oral diagnosis and treatment. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method for dental 3D recognition and model generation based on CBCT and deep learning. This method solves the problem that existing technologies struggle to simultaneously, quickly, and accurately segment multiple anatomical structures such as teeth, alveolar bone, mandibular nerve canal, periodontal ligament, and root canals from CBCT images through a single automated processing step.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for dental 3D recognition and model generation based on CBCT and deep learning, comprising: S1. Preprocess the acquired CBCT image data; S2. The preprocessed CBCT image data is segmented. The segmentation step includes: using a two-stage cascaded deep learning network to obtain a three-dimensional model of the teeth, a three-dimensional model of the alveolar bone, and a three-dimensional model of the mandibular nerve canal; and based on the three-dimensional model of the teeth and the three-dimensional model of the alveolar bone, a three-dimensional model of the periodontal ligament is obtained through geometric and Boolean operations. S3. Locate the root region in the three-dimensional tooth model, and use an adaptive thresholding method on the CBCT image data corresponding to the root region to identify and generate a three-dimensional root canal model. S4. Based on at least one of the obtained three-dimensional models of the teeth, alveolar bone, mandibular nerve canal, periodontal ligament, and root canal, a personalized clinical application model is generated, wherein the personalized clinical application model includes at least one of the following: implant guide model, personalized orthodontic archwire model, and impacted tooth extraction planning model.

[0009] Preferably, the preprocessing step includes: sequentially performing Gaussian filtering, artifact correction, and grayscale normalization on the CBCT image data. Gaussian filtering is used to suppress image noise, achieved by convolving the original 3D image data with a 3D Gaussian kernel function.

[0010] Gray-level normalization is used to eliminate gray-level differences caused by different scanning parameters. It linearly maps voxel gray values ​​to the [0,1] interval using the min-max normalization method. The calculation formula is as follows: ; in, The new grayscale values ​​obtained after normalization are located in the interval [0,1]. and These represent the maximum and minimum values ​​of the defined grayscale range, respectively. After this processing step, the output data can be used as the standard input for subsequent deep learning networks. In a specific embodiment, step S2, which involves employing a two-stage cascaded deep learning network, specifically includes: First, the CBCT image data is coarsely segmented using a V-Net network to quickly locate and extract the region of interest containing the complete dentition and the upper and lower jaws from the complete three-dimensional image; Then, using the region of interest as input, the UNETR++ network is used to perform multi-class fine segmentation on the region of interest to separate and generate the three-dimensional model of the tooth and the three-dimensional model of the alveolar bone.

[0011] Preferably, the UNETR++ network integrates an attention gating unit. During feature fusion, this attention gating unit adjusts the weights of the encoder feature map by calculating attention weight coefficients. The purpose is to enhance the learning of features in the boundary region between the teeth and alveolar bone and suppress feature responses from irrelevant background regions.

[0012] In another specific embodiment, step S2, which involves obtaining the three-dimensional model of the periodontal ligament through geometric and Boolean operations, specifically includes: First, a morphological expansion operation is performed on the three-dimensional tooth model to obtain an expanded tooth model. This operation can be represented as: ; in, For expansion operator, For a predefined structuring element or core, For the model of the teeth that are expanded outwards, A three-dimensional model of teeth.

[0013] Secondly, a Boolean intersection operation is performed between the extended tooth model and the alveolar bone 3D model to obtain the alveolar socket model. This operation can be represented as: ; in, This is the Boolean intersection operator. For alveolar socket model, This is a three-dimensional model of the alveolar bone.

[0014] Finally, the tooth 3D model is subtracted from the alveolar socket model using a Boolean subtraction operation. The resulting difference set is the periodontal ligament 3D model. This operation can be represented as: ; in, This is a Boolean subtraction operator. This method avoids the difficulty of directly segmenting low-contrast periodontal ligament tissue, achieving accurate generation of the periodontal ligament model using high-precision tooth and alveolar bone models. This is a three-dimensional model of the periodontal ligament.

[0015] As one implementation method, step S3 specifically includes: The CBCT image data corresponding to the tooth root region is traversed using a sliding window method. For the set of voxels within the window, its local mean and local standard deviation are calculated. Based on the local mean With the local standard deviation Dynamically calculate the adaptive segmentation threshold The calculation formula is: ; in, This is a preset sensitivity coefficient used to adjust the weight of the local standard deviation on the final threshold.

[0016] Voxels with grayscale values ​​higher than the threshold are identified as root canal structures, and the root canal 3D model is generated accordingly.

[0017] In one embodiment, when the personalized clinical application model in step S4 is an implantation guide model, its generation process includes: Based on the preset implant placement and angle information, combined with the three-dimensional tooth model and alveolar bone model, a guide plate body with guide holes is generated through Boolean operations. On the inner wall of the guide hole, a Boolean subtraction operation is performed on a preset thread geometry to form a threaded substitute interface that matches the substitute body; The base surface on the guide plate body for contact with adjacent teeth or mucosa is identified, and geometric morphology matching is performed with the three-dimensional model of the adjacent teeth or mucosa to form the positioning guide structure, and finally the implant guide plate model is generated.

[0018] In another embodiment, when the personalized clinical application model in step S4 is a personalized orthodontic archwire model, its generation process includes: A series of ordered spatial feature points are extracted along the alveolar ridge crest of the three-dimensional alveolar bone model; the spatial feature point set is fitted with a B-spline curve to generate a smooth three-dimensional dental arch morphology curve. The mathematical expression for this B-spline curve is: ; in, These are control points (i.e., extracted feature points). It is a p-th order B-spline basis function. These are parameters in the node vector.

[0019] The three-dimensional curve is projected onto the sagittal plane, and the Spee curve of the dental arch is fitted using a polynomial function; finally, the three-dimensional dental arch morphology curve and the Spee curve are fused to generate the personalized orthodontic archwire model.

[0020] In yet another embodiment, when the personalized clinical application model is an impacted tooth extraction planning model, its generation process includes: Based on the impacted tooth 3D model (i.e., a specific tooth 3D model) and the mandibular nerve canal 3D model, the minimum Euclidean distance between the point sets on the surfaces of the two models is calculated for surgical risk assessment. The calculation formula is as follows: ; in, and These are point vectors on the surfaces of the two models, respectively.

[0021] Based on the risk assessment, an ultrasonic bone removal path is generated to expose the tooth root and maintain a safe distance from the mandibular nerve canal. The bone removal path is then integrated with the three-dimensional model of the impacted tooth and the three-dimensional model of the mandibular nerve canal to form the impacted tooth extraction planning model.

[0022] To ensure the geometric accuracy of the generated model, the method provided by this invention further includes a step of accuracy verification of the personalized clinical application model, which specifically includes: Based on the data from the personalized clinical application model, a physical model is manufactured using additive or subtractive manufacturing techniques. The three-dimensional scanning data of the physical model are obtained using a coordinate measuring machine or an industrial-grade optical scanner; The three-dimensional scan data is then three-dimensionally registered with the original personalized clinical application model, and the geometric deviation between the two is calculated.

[0023] This invention provides a method for dental 3D recognition and model generation based on CBCT and deep learning. It has the following beneficial effects: 1. This invention achieves automated segmentation of major anatomical structures such as teeth and alveolar bone by employing a two-stage cascaded deep learning network composed of V-Net and UNETR++ networks. Furthermore, for complex structures such as periodontal ligament and root canal that are difficult to accurately identify by traditional algorithms, fine processing is performed through geometric and Boolean operations and adaptive thresholding methods. This combination of deep learning and specific algorithms ensures the automation of the overall process while improving the segmentation accuracy of key fine structures.

[0024] 2. The method of the present invention covers the complete process from preprocessing of raw CBCT image data and segmentation of multi-tissue three-dimensional models to generating various types of personalized clinical application models. Users can complete all operations without switching between multiple independent software systems, which simplifies the workflow of clinical digitization, shortens the design cycle, and reduces the risk of errors that may be introduced due to data conversion.

[0025] 3. This invention first obtains a high-precision individual anatomical model containing multiple tissues, including teeth, alveolar bone, and nerve canals. Based on this, by performing specific generation steps, such as downstream Boolean operations, curve fitting, or risk assessment calculations, the structure and function of the implant guide, orthodontic archwire, and surgical planning model are optimized and designed. This ensures that the final generated clinical application model can be highly matched with the patient's individual anatomical structure, thereby improving the accuracy and safety of subsequent clinical treatment. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the overall technical process of the present invention. Figure 2 This is a diagram of the segmentation network architecture of the present invention. Figure 3 This is a system functional block diagram of the present invention; Figure 4 This is a comparison diagram of the tissue segmentation effect of the present invention; Figure 5 This is a CT image processing diagram of the teeth according to the present invention; Figure 6 This is a schematic diagram of the U-Net deep learning model of the present invention; Figure 7 This is a diagram of tooth scanning recognition (distinguishing teeth) according to the present invention; Figure 8 This is a schematic diagram illustrating the generation of a three-dimensional tooth model according to the present invention; Figure 9 This is a schematic diagram of the 3D printing of a three-dimensional tooth model according to the present invention; Figure 10 This is a schematic diagram of the optimized three-dimensional tooth model of the present invention; Figure 11This is a printed display image of the optimized 3D tooth model according to the present invention; Figure 12 This is a schematic diagram of the optimized three-dimensional tooth model printed according to the present invention; Figure 13 This is a schematic diagram of the design of the three-dimensional visualization analysis software plugin for teeth according to the present invention.

[0027] The module includes: 10. Data acquisition module; 20. Data preprocessing module; 30. 3D model segmentation module; and 40. Personalized clinical application model generation module. Detailed Implementation

[0028] 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.

[0029] See attached document Figure 1 -Appendix Figure 13 The method provided in this embodiment of the invention can be run on a computing device, which includes a processor, a graphics processing unit (GPU), a memory, and a data bus for connecting the aforementioned components. The memory stores computer program instructions, and when the processor executes these computer program instructions, it invokes the GPU for parallel computation to implement the steps of the following method.

[0030] In one embodiment, the system may include: a data acquisition module 10, a data preprocessing module 20, a three-dimensional model segmentation module 30, and a personalized clinical application model generation module 40.

[0031] In a specific implementation, the method described in this invention can be implemented through computer programming. For example, Python can be used, and the two-stage cascaded deep learning network can be built and trained using deep learning frameworks such as PyTorch or TensorFlow. The processing of 3D image data and geometric operations can be implemented using open-source libraries such as SimpleITK and VTK (Visualization Toolkit). The computing device can be configured with an NVIDIA RTX series or higher-performance graphics processing unit (GPU) to accelerate the training and inference process of the deep learning network.

[0032] The data acquisition module 10 is used to receive three-dimensional image data stored in DICOM format obtained by a cone-beam computed tomography (CBCT) device.

[0033] The data preprocessing module 20, connected to the data acquisition module 10, is used to perform preprocessing operations on the received three-dimensional image data to suppress image noise, correct artifacts, and normalize grayscale values, and outputs the processed three-dimensional image data.

[0034] The 3D model segmentation module 30, connected to the data preprocessing module 20, receives preprocessed 3D image data and performs segmentation operations on it. This module 30 integrates a one- or two-stage cascaded deep learning network and specific geometric operations and image processing algorithms. Its function is to identify and separate different anatomical tissues such as teeth, alveolar bone, mandibular nerve canal, periodontal ligament, and root canals, and generate corresponding 3D model data for each.

[0035] The personalized clinical application model generation module 40 is connected to the three-dimensional model segmentation module 30. It is used to receive three-dimensional model data of various anatomical tissues and generate personalized models with optimized structure that can be used for specific clinical applications, such as implant guide models, personalized orthodontic archwire models, or impacted tooth extraction planning models.

[0036] See attached document Figure 1 -Appendix Figure 7 The dental 3D recognition and model generation method based on CBCT and deep learning provided by this invention may include the following steps: Step S100: Preprocess the acquired CBCT image data. This step is performed by the data preprocessing module 20, which aims to eliminate interference factors such as noise and artifacts in the original scan data, and to standardize the gray values ​​of voxels to provide consistent data input for subsequent segmentation steps.

[0037] Step S200: Segment the preprocessed CBCT image data. This step is performed by the 3D model segmentation module 30 and is the core of generating a high-precision anatomical structure model. This step first uses a two-stage deep learning network to automatically segment 3D models of teeth, alveolar bone, and mandibular nerve canal from the image data. Subsequently, based on the obtained tooth and alveolar bone models, a 3D model of the periodontal ligament is generated through a series of geometric and Boolean operations.

[0038] Step S300: Identify and generate a 3D model of the root canal. This step is also performed by the 3D model segmentation module 30. Based on the 3D model of the tooth segmented by the deep learning network, the region where the tooth root is located is located, and an adaptive thresholding method is used to process the original CBCT image data corresponding to this region in order to accurately identify and generate the internal 3D model of the root canal.

[0039] Step S400: Generate a personalized clinical application model. This step is performed by the personalized clinical application model generation module 40. This step utilizes one or more anatomical tissue 3D models obtained in the preceding steps as a foundation, and designs and generates clinical application models with specific functional structures by executing a preset algorithm flow. Examples include implant guide models for precise implant placement, personalized orthodontic archwire models for orthodontic treatment, or impacted tooth extraction planning models for surgical risk assessment and pathway planning.

[0040] See attached document Figure 2 Step S100 is the specific implementation process of preprocessing the acquired CBCT image data. The purpose of this step is to eliminate noise and artifacts in the original three-dimensional image data and standardize the gray values ​​of voxels, so as to provide data with consistent characteristics for the subsequent three-dimensional model segmentation step S200.

[0041] In one specific embodiment, the preprocessing step includes three sequential sub-steps: Gaussian filtering, artifact correction, and grayscale normalization.

[0042] First, a Gaussian filtering operation is performed. This operation is used to suppress Gaussian noise and random noise commonly found in the original CBCT 3D image data. Specifically, this is achieved by performing a 3D convolution between the original 3D image data and a 3D Gaussian kernel function. The mathematical expression for the 3D Gaussian kernel function is: ; in, These are the coordinates of the voxel relative to the nuclear center. This is the standard deviation of the Gaussian distribution, and this parameter controls the smoothness of the filter. Through this convolution operation, the gray values ​​of each voxel can be weighted and averaged, thereby achieving the purpose of smoothing the image and suppressing noise.

[0043] Secondly, artifact correction is performed. This operation is primarily used to address high-intensity artifacts caused by metallic implants or prostheses, such as stripe artifacts or beam hardening artifacts. The operation first identifies and segments the metallic regions in the image by setting a high grayscale threshold. Then, the voxel grayscale values ​​in the regions affected by metallic artifacts (typically manifested as low- or high-grayscale stripes around the metallic region) are corrected using a neighborhood interpolation algorithm (e.g., trilinear interpolation) to recover the tissue structure information obscured by the artifacts.

[0044] Finally, grayscale normalization is performed. Since different CBCT devices or scanning parameters can cause differences in the grayscale range of images, this operation linearly maps the grayscale values ​​of all voxels to a unified, preset floating-point range, such as [0,1]. This operation first adjusts or truncates the original Hounsfield Unit (HU) values ​​using window level / window width to preserve the grayscale range of target structures such as teeth and bone tissue, for example, limiting the grayscale values ​​to the range of [-1000, 2000]HU. Subsequently, the truncated grayscale values ​​are calculated using a min-max normalization method, with the following formula: ; in, The new grayscale values ​​obtained after normalization are located in the interval [0,1]. and These represent the maximum and minimum values ​​of the defined grayscale range, respectively. After this processing step, the output data can be used as the standard input for subsequent deep learning networks.

[0045] See attached document Figure 2 Step S200 is the specific implementation process of segmenting the preprocessed CBCT image data to obtain various 3D models of anatomical tissues. This step is executed by the 3D model segmentation module 30 and is the core part of the technical solution of this invention.

[0046] The segmentation process can be divided into two main stages: the first stage is to use a two-stage cascaded deep learning network to automatically segment the main anatomical structures (teeth, alveolar bone and mandibular nerve canal); the second stage is to use geometric and Boolean operations to generate a three-dimensional model of the periodontal ligament based on the output of deep learning.

[0047] Segmentation of key anatomical structures based on a two-stage cascaded deep learning network: This embodiment employs a two-stage cascaded network structure of coarse segmentation and fine segmentation. This structure aims to address the problems of enormous computational resource consumption and background information redundancy encountered when directly performing fine segmentation on the entire high-resolution CBCT image.

[0048] First, a coarse segmentation stage is performed. A V-Net network is used to coarsely segment the preprocessed complete CBCT 3D image data. This V-Net network possesses 3D convolutional properties, enabling efficient processing of 3D volumetric data. The purpose of coarse segmentation is to quickly locate and extract smaller 3D regions of interest (ROIs) containing teeth and alveolar bone, thereby significantly reducing the computational complexity of subsequent fine segmentation.

[0049] Secondly, the fine segmentation stage is performed. The UNETR++ network is used to perform fine segmentation on the region of interest extracted from the coarse segmentation. The UNETR++ network is a U-shaped network based on the Transformer architecture, which has a global receptive field and is suitable for processing complex anatomical structures. This network performs multi-class segmentation on voxels within the region of interest, automatically identifying and generating 3D models of teeth, alveolar bone, and mandibular nerve canal.

[0050] In a preferred embodiment, the UNETR++ network integrates an attention gating unit. This attention gating unit, deployed on a skip connection between the encoder and decoder, dynamically computes attention weight coefficients when fusing multi-scale feature maps. These coefficients enhance the learned weights for the boundary features between teeth and alveolar bone while suppressing the propagation of background noise features irrelevant to the segmentation task, ensuring higher accuracy in the segmentation results, particularly regarding the geometric boundaries between teeth and alveolar bone.

[0051] When converting the segmented voxel data into a 3D model, the Moving Cubes algorithm can be used. This algorithm takes the 3D label map output by the deep learning network as input, extracts the isosurfaces of various tissues (such as teeth and alveolar bone), and converts them into triangular mesh models, for example, storing them in STL (Standard Tessellation Language) or PLY (Polygon File Format) formats for subsequent geometric calculations and personalized clinical application model generation step S400.

[0052] Furthermore, the two-stage cascaded deep learning network requires training before use. The training process includes: A batch of CBCT image data containing expert manual segmentation annotations (Ground Truth) was acquired as the training set; The training set is used to supervise the learning of the V-Net network and the UNETR++ network. The network parameters are iteratively optimized by minimizing the loss function (such as Dice Loss or cross-entropy loss function) between the network prediction results and the expert annotations until the network model converges.

[0053] Generation of a 3D model of the periodontal ligament: Obtaining a 3D model of teeth Three-dimensional model of alveolar bone Subsequently, this embodiment uses geometric and Boolean operations to generate a three-dimensional model of the periodontal ligament that is difficult to identify directly through pure image segmentation. .

[0054] The specific process of this generation step includes: Morphological dilation operation: A minor morphological dilation operation is performed on the three-dimensional tooth model to obtain an expanded tooth model. The purpose of the expansion operation is to uniformly expand the tooth model along the surface normal direction by a predetermined thickness (e.g., 0.15 mm to 0.3 mm), which matches the physiological thickness of the periodontal ligament. This operation can be represented as: ; in, For expansion operator, For a predefined structuring element or kernel.

[0055] In one specific embodiment, the preset thickness value can be set within a reasonable range based on the physiological characteristics of the periodontal ligament. For example, the thickness value can be set to 0.15 mm to 0.30 mm.

[0056] Boolean intersection operation: intersects the extended tooth model with the alveolar bone 3D model. Perform a Boolean intersection operation. Since the expansion of the tooth is uniform outward, the portion where the expanded tooth model intersects with the alveolar bone model can be precisely defined as the alveolar socket model. This operation can be represented as: ; in, This is the Boolean intersection operator.

[0057] Boolean subtraction operation: Subtract the original three-dimensional tooth model from the alveolar socket model using a Boolean subtraction operation. After subtracting the tooth model, the remaining surrounding space is the periodontal ligament region between the tooth and the alveolar socket, with controlled thickness. This operation can be represented as: ; in, This is a Boolean subtraction operator. This method avoids the difficulty of directly segmenting low-contrast periodontal ligament tissue, achieving accurate generation of the periodontal ligament model using high-precision tooth and alveolar bone models.

[0058] See attached document Figure 2 Step S300 is the specific implementation process of identifying and generating the three-dimensional model of the root canal. Given that the three-dimensional model of the tooth has been obtained in step S200, this step aims to accurately segment the root canal system, which has a complex morphology and low contrast with the surrounding dentin, from within the root structure.

[0059] This embodiment does not employ a global thresholding method, but rather an adaptive thresholding method. This method first locates the root region in the three-dimensional tooth model, and then dynamically calculates an independent segmentation threshold for each voxel within the corresponding CBCT image data.

[0060] The specific implementation steps of this method are as follows: First, a three-dimensional sliding window (kernel) is defined, which is a cube of a preset size, such as 5×5×5 voxels. The algorithm will traverse every target voxel within the root region and align the center of the sliding window with the current target voxel. At this point, the set of voxels covered by the sliding window constitutes the local region of the current target voxel.

[0061] Secondly, for each local region, calculate the local mean of all voxels within that region. With local standard deviation The calculation formulas are as follows: ; ; in, It is the first in a local area grayscale value of a voxel It is the total number of voxels in this local area.

[0062] Then, based on the calculated local mean and local standard deviation, an adaptive segmentation threshold is dynamically calculated for the current target voxel. The calculation formula is as follows: ; in, This is a preset sensitivity coefficient used to adjust the weight of the local standard deviation on the final threshold.

[0063] The value of this coefficient affects the sensitivity of segmentation. In a preferred embodiment, the coefficient can be set between -0.5 and 0.5, and the specific value can be determined experimentally on a validation set to balance the recall and precision of root canal identification.

[0064] Finally, the actual grayscale value of the current target voxel is compared with its calculated adaptive segmentation threshold. If the actual grayscale value is greater than the adaptive segmentation threshold, the voxel is determined to belong to the root canal structure. After the algorithm traverses all voxels in the root region and completes the above determination, all voxels determined to belong to the root canal structure are collected to generate the final 3D root canal model.

[0065] See attached document Figure 2Step S400 is the specific implementation process of generating a personalized clinical application model based on one or more 3D anatomical tissue models obtained in steps S200 and S300. This step is executed by the personalized clinical application model generation module 40. The following will use implant guide models, personalized orthodontic archwire models, and impacted tooth extraction planning models as examples to illustrate several specific implementation methods of this step.

[0066] Implementation Method 1: Generation of Planting Guide Plate Model When the personalized clinical application model is an implant guide model, its generation process is based on the preset implant placement position and angle information, and utilizes the obtained three-dimensional tooth model and alveolar bone model.

[0067] The process begins by generating the guide plate body. Based on the three-dimensional tooth model, an initial guide plate geometry is constructed, covering the teeth adjacent to the implant site or the alveolar bone mucosa. Subsequently, according to the preset implant placement position and angle information, a cylindrical model representing the implant drill path is generated. By performing a Boolean subtraction operation on the initial guide plate geometry, the cylindrical model is subtracted to form the guide plate body with precise guide holes.

[0068] Next, a threaded substitute interface is formed on the inner wall of the guide hole. This step is achieved by performing a Boolean subtraction operation on a preset thread geometry that matches the thread of the target substitute. After aligning the thread geometry coaxially with the guide hole, the geometry is subtracted from the guide plate body to generate a precise female thread structure on the inner wall of the guide hole.

[0069] Finally, a base surface for precise positioning is formed. The base surface on the guide body for contact with adjacent teeth or mucosa is identified, and the base surface is trimmed and fitted by performing geometric matching processing with the three-dimensional model of the adjacent teeth or mucosa, such as Boolean intersection or surface projection operations, to ensure that the implant guide has a unique and stable placement path and position in the oral cavity.

[0070] Implementation Method 2: Generation of Personalized Orthodontic Archwire Models When the personalized clinical application model is a personalized orthodontic archwire model, its generation process is mainly based on the alveolar bone three-dimensional model. This process first generates a three-dimensional dental arch morphology curve. Along the alveolar ridge crest of the alveolar bone three-dimensional model, a series of ordered spatial feature points are automatically or interactively extracted. B-spline curves were used to fit the spatial feature point set to generate a smooth three-dimensional curve that accurately reflects the individual dental arch morphology of the patient. The mathematical expression for this B-spline curve is: ; in, These are control points (i.e., extracted feature points). It is a p-th order B-spline basis function. These are parameters in the node vector.

[0071] Next, the Spee curve of the dental arch is fitted. The generated three-dimensional dental arch morphology curve is projected onto the sagittal plane to obtain a series of two-dimensional points. A polynomial function of a preset order (such as a quadratic or cubic polynomial) is used to fit these two-dimensional points using the least squares method to obtain the Spee curve describing the anteroposterior curvature of the dental arch.

[0072] Finally, the three-dimensional dental arch morphology curve and the Spee curve are merged. The fitted Spee curve is used as a height adjustment function to adjust the vertical coordinates (usually the Z coordinate) of each point on the three-dimensional dental arch morphology curve, thereby incorporating the Spee curvature information into the three-dimensional dental arch morphology and ultimately generating the personalized orthodontic archwire model.

[0073] Implementation Method 3: Generation of the Impacted Tooth Extraction Planning Model When the personalized clinical application model is an impacted tooth extraction planning model, its generation process is mainly based on the impacted tooth three-dimensional model (i.e., a specific tooth three-dimensional model) and the mandibular nerve canal three-dimensional model.

[0074] The procedure begins with a surgical risk assessment. This is achieved by calculating the surface point set of the three-dimensional model of the impacted tooth. The point set on the surface of the three-dimensional model of the mandibular nerve canal Minimum Euclidean distance between This is used to quantify the spatial relationship between the tooth root and the nerve canal. The calculation formula is: ; in, and These are point vectors on the surfaces of the two models, respectively.

[0075] Secondly, an ultrasonic osteotome bone removal path is generated based on the risk assessment results. A three-dimensional spatial path for exposing the impacted tooth root is generated according to the calculated minimum Euclidean distance. This path is designed to maintain a preset safe distance from the surface of the mandibular nerve canal three-dimensional model to guide bone tissue removal during surgery.

[0076] See attached document Figure 8 Finally, the impacted tooth extraction planning model is formed. The three-dimensional model of the impacted tooth, the three-dimensional model of the mandibular nerve canal, and the generated bone removal path are integrated into a three-dimensional data set and displayed together in the same three-dimensional coordinate system. This provides the operator with a comprehensive planning model that includes the positional relationships of key anatomical structures, risk assessment data, and suggested surgical paths.

[0077] See attached document Figure 9 -Appendix Figure 13 To further ensure the geometric accuracy of the personalized clinical application model generated by the method of the present invention after physical manufacturing, the present invention also provides an embodiment for accuracy verification of the model. (See attached document) Figure 3 , Figure 3 This is a flowchart of a method for verifying the accuracy of a personalized clinical application model according to an embodiment of the present invention. The verification process is performed after the personalized clinical application model is generated.

[0078] The accuracy verification process specifically includes: First, based on the digital geometric data (e.g., STL files) of the personalized clinical application model generated in step S400, a physical model is manufactured using additive manufacturing technology. Depending on the model type and application requirements, technologies that can be used include stereolithography (SLA) and selective laser sintering (SLS). For example, biocompatible resin is used as a raw material when manufacturing implant guides; metal powder or high-strength polymers can be used as raw materials when manufacturing orthodontic archwire validation models.

[0079] Secondly, high-precision 3D measurement equipment is used to acquire 3D scanning data of the physical model. Acquisition methods include contact and non-contact measurement. Contact measurement can use a coordinate measuring machine (CMM), which uses a probe to contact key surfaces of the physical model point by point to acquire a high-precision set of coordinate points. Non-contact measurement can use industrial-grade optical scanners, such as structured light scanners or laser scanners, which project specific gratings or line lasers onto the surface of the physical model, and a camera captures its deformation, thereby quickly reconstructing a high-density 3D surface point cloud or mesh data containing millions of data points.

[0080] Finally, the acquired 3D scan data is registered in 3D with the original personalized clinical application model digital data, and geometric deviation is calculated. This step first uses the Iterative Closest Point (ICP) algorithm to optimally align the 3D scan data and the original model data in the same coordinate system. After registration, the geometric deviation between the two models is calculated. This calculation is performed by measuring the shortest perpendicular distance from each point on the surface of the scan data model to the surface of the original digital model, and a visual deviation chromatogram is output. For quantitative evaluation, the overall root mean square (RMS) error can also be calculated, using the following formula: ; in, Represents the root mean square error. It is the total number of points measured. It is the first Geometric deviation values ​​at each measurement point.

[0081] By comparing the calculated RMS error value with a preset clinically acceptable tolerance threshold, the objective and quantitative verification of the manufacturing accuracy of the physical model can be completed.

Claims

1. A dental three-dimensional recognition and model generation method based on CBCT and deep learning, characterized in that, The method comprises the following steps: S1, preprocessing the obtained CBCT image data; S2, segmenting the preprocessed CBCT image data, the segmentation step comprising: adopting a two-stage cascaded deep learning network to obtain a tooth three-dimensional model, an alveolar bone three-dimensional model and a mandibular nerve canal three-dimensional model; based on the tooth three-dimensional model and the alveolar bone three-dimensional model, obtaining a periodontal membrane three-dimensional model through geometric and Boolean operations; S3, positioning the tooth root region in the tooth three-dimensional model, and using an adaptive threshold method on the CBCT image data corresponding to the tooth root region to identify and generate a root canal three-dimensional model; S4, based on at least one of the obtained tooth three-dimensional model, alveolar bone three-dimensional model, mandibular nerve canal three-dimensional model, periodontal membrane three-dimensional model and root canal three-dimensional model, generating a personalized clinical application model, wherein the personalized clinical application model at least includes at least one of an implant guide plate model, a personalized orthodontic arch wire model and an impacted tooth extraction planning model. 2.The CBCT and deep learning based dental three-dimensional recognition and model generation method of claim 1, wherein, The preprocessing step comprises: the CBCT image data is sequentially subjected to Gaussian filtering, artifact correction and gray scale normalization processing. 3.The CBCT and deep learning based dental three-dimensional recognition and model generation method of claim 1, wherein, In step S2, the two-stage cascaded deep learning network specifically comprises: using a V-Net network to roughly segment the CBCT image data and extract a region of interest containing teeth and alveolar bone; using a UNETR++ network to finely segment the region of interest. 4.The CBCT and deep learning based dental three-dimensional recognition and model generation method of claim 1, wherein, In step S2, the step of obtaining a periodontal membrane three-dimensional model through geometric and Boolean operations specifically comprises: performing a morphological dilation operation on the tooth three-dimensional model to obtain an expanded tooth model; performing a Boolean intersection operation on the expanded tooth model and the alveolar bone three-dimensional model to obtain an alveolar fossa model; performing a Boolean subtraction of the tooth three-dimensional model from the alveolar fossa model to obtain the periodontal membrane three-dimensional model. 5.The CBCT and deep learning based dental three-dimensional recognition and model generation method of claim 1, wherein, The S3 step specifically comprises: for the local region of the tooth root, calculating the local mean and local standard deviation of the voxels in the local region; based on the local mean and the local standard deviation, dynamically calculating an adaptive segmentation threshold; according to the adaptive segmentation threshold, identifying and generating the root canal three-dimensional model. 6.The CBCT and deep learning based dental three-dimensional recognition and model generation method according to claim 3, characterized in that, The UNETR++ network integrates an attention gate unit to enhance the learning of boundary features between teeth and alveolar bone. 7.The CBCT and deep learning based dental three-dimensional recognition and model generation method of claim 1, wherein, In step S4, when the personalized clinical application model is an implant guide plate model, the generation process comprises: based on the pre-set implant implantation position and angle information, combining the tooth three-dimensional model and the alveolar bone three-dimensional model, and generating a guide plate main body with a guide hole through Boolean operation; on the inner wall of the guide hole, a threaded substitute body interface is formed by performing a Boolean subtraction operation on a pre-set threaded geometry; identifying the basal surface of the guide plate main body for contact with adjacent teeth or mucosa, and forming the implant guide plate model through geometric and morphological matching processing with the three-dimensional model of the adjacent teeth or mucosa. 8.The CBCT and deep learning based dental three-dimensional recognition and model generation method of claim 1, wherein, In step S4, when the personalized clinical application model is a personalized orthodontic arch wire model, the generation process comprises: extracting a series of spatial feature points along the alveolar crest of the alveolar bone three-dimensional model; The spatial feature points are fitted by using a B-spline curve to generate a three-dimensional dental arch shape curve; A polynomial function is used to fit the Spee curve of the dental arch; The three-dimensional dental arch shape curve and the Spee curve are fused to generate the personalized orthodontic archwire model. 9.The CBCT and deep learning based dental three-dimensional recognition and model generation method of claim 1, wherein, When the personalized clinical application model is a supernumerary tooth extraction planning model, the generation process includes: Based on the three-dimensional model of the supernumerary tooth and the three-dimensional model of the mandibular nerve canal, the minimum distance between them is calculated for surgical risk assessment; Based on the risk assessment, an ultrasonic bone knife bone removal path for exposing the tooth root is generated, and the bone removal path is integrated with the three-dimensional model of the supernumerary tooth and the three-dimensional model of the mandibular nerve canal to form the supernumerary tooth extraction planning model. 10.The CBCT and deep learning based dental three-dimensional recognition and model generation method of claim 1, wherein, The dental three-dimensional recognition and model generation method based on CBCT and deep learning further includes a step of precision verification of the personalized clinical application model, which specifically includes: Based on the data of the personalized clinical application model, a physical model is manufactured; A three-coordinate measuring instrument or an optical scanner is used to obtain three-dimensional scanning data of the physical model; The three-dimensional scanning data and the personalized clinical application model are registered, and the geometric deviation between them is calculated.

Images