Dention reconstruction and orthodontic simulation method and system and storage medium

Abstract

The invention provides a dentition reconstruction and orthodontic simulation method and system and a storage medium, and relates to the technical field of medical image processing, and the method comprises the steps: obtaining a plurality of intraoral photos shot by a user, and carrying out the image preprocessing to obtain standardized intraoral photos; performing tooth segmentation processing on the standardized intraoral picture by adopting a deep neural network to obtain a tooth region mask; performing three-dimensional reconstruction on the plurality of standardized intraoral photos to obtain a three-dimensional dental crown model; performing registration fusion on the three-dimensional dental crown model and a tooth root model in the CBCT image to obtain a dental crown-tooth root fusion model; and constructing a tooth movement mechanical model based on the dental crown-tooth root fusion model for orthodontic effect simulation and outputting an orthodontic prediction result. According to the method, the three-dimensional dentition model is directly reconstructed from the two-dimensional picture, and professional equipment such as an oral scanner is not needed; a complete anatomical relation is established through registration fusion of the dental crown and the tooth root; mechanical simulation is carried out based on the periodontal membrane elastic model, and the prediction result better conforms to the actual treatment effect.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method, system, and storage medium for dental reconstruction and orthodontic simulation. Background Technology

[0002] Dental scanners, as the mainstream equipment for obtaining three-dimensional models in orthodontics, play a vital role in clinical diagnosis and treatment. Doctors can clearly observe the shape and position of teeth through these three-dimensional models, providing a basis for disease diagnosis and treatment planning. However, existing dental scanners suffer from high equipment costs and significant dependence on clinical settings. Patients must visit the clinic in person to complete the scanning procedure, making it difficult for small and medium-sized clinics to afford the investment. Furthermore, patients in remote areas or with mobility issues face difficulties in accessing the clinic. These factors severely restrict the widespread adoption of three-dimensional orthodontic technology.

[0003] With the continuous improvement of smartphone camera performance, mobile devices now possess high-resolution image acquisition capabilities, making it possible to convert two-dimensional intraoral photographs into three-dimensional models using algorithms. Existing technologies mostly employ a three-dimensional-to-three-dimensional data processing approach, acquiring three-dimensional data from oral scanners or CBCT devices before subsequent analysis and processing. This lack of effective methods for directly constructing diagnostically usable three-dimensional models from ordinary two-dimensional images results in patients being unable to complete data acquisition independently, requiring them to make a special trip to the medical institution for each follow-up visit, increasing the frequency and time costs of medical visits.

[0004] In predicting orthodontic treatment outcomes, existing systems typically only consider the surface morphology of the crown for analysis and planning, failing to incorporate deeper anatomical structures such as the root and alveolar bone into the biomechanical analysis framework. While CBCT can provide information on the root and bone structure, it cannot accurately reflect the direction and magnitude of forces acting on the root during treatment. How to effectively integrate the crown surface model with the root bone structure without increasing reliance on advanced equipment has become a key issue in improving the accuracy of orthodontic prediction.

[0005] Furthermore, existing orthodontic prediction schemes mostly employ geometric path optimization algorithms, failing to simulate the actual response of teeth under orthodontic forces from a biomechanical perspective. The periodontal ligament, as a crucial tissue connecting the tooth root and alveolar bone, directly influences the direction and amplitude of tooth movement through its elastic properties. Ignoring this factor will lead to discrepancies between predicted results and actual treatment outcomes, making it difficult to provide reliable clinical decision support for physicians.

[0006] Chinese patent document CN119672047A discloses a method and system for three-dimensional segmentation of dental CBCT images. It discloses a technical solution that combines CBCT scanning with dynamic MRI examination, uses generative adversarial networks to generate virtual time-point data, and employs an attention mechanism to extract multi-time-point features and perform three-dimensional convolutional neural network segmentation. This solution achieves the technical effects of integrating multimodal data, grasping dynamic changes during treatment, and accurately focusing on key structural features. However, it still has the problems of requiring the support of two professional imaging devices, CBCT and MRI, requiring patients to undergo examinations at medical institutions, and being unable to directly reconstruct three-dimensional models from ordinary two-dimensional photographs.

[0007] Chinese patent document CN119925008A discloses a soft and hard tissue collaborative orthodontic auxiliary treatment system based on dual-modal imaging of CBCT and pseudo-MRI. It discloses a technical solution that generates pseudo-MRI images from CBCT images through deep learning, generates soft and hard tissue collaborative images using pixel-level and feature-level fusion algorithms, and combines multi-objective optimization algorithms for orthodontic path planning. It achieves the technical effects of obtaining high-contrast information of soft tissue without real MRI equipment, improving diagnostic accuracy, and realizing personalized path planning. However, it still has problems such as relying on CBCT equipment as the only data input source, being unable to reconstruct three-dimensional models from two-dimensional images, and orthodontic prediction being based only on geometric optimization without introducing biomechanical simulation. Summary of the Invention

[0008] The purpose of this invention is to provide a method, system, and storage medium for dentition reconstruction and orthodontic simulation, which can reconstruct a three-dimensional dentition model based on intraoral photos taken with a mobile phone without the need for an oral scanner, and combine CBCT fusion and biomechanical models to achieve accurate prediction of orthodontic effects.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A method for dentition reconstruction and orthodontic simulation includes the following steps: S1: Acquire multiple intraoral photos taken by the user, perform image preprocessing on the intraoral photos, and obtain standardized intraoral photos; S2: Perform tooth segmentation processing on the standardized intraoral photograph to obtain a tooth region mask; S3: Based on the multi-view geometric method, three-dimensional reconstruction is performed on multiple standardized intraoral photographs to obtain a three-dimensional crown model; S4: Register and fuse the three-dimensional crown model with the root model in the CBCT image to obtain a crown-root fusion model; S5: Construct a tooth movement biomechanics model based on the crown-root fusion model, simulate orthodontic effects, and output orthodontic prediction results.

[0011] Further: The image preprocessing in S1 includes: performing white balance correction, local contrast enhancement, highlight area repair and lens distortion correction on the intraoral photograph in sequence; wherein the local contrast enhancement includes converting the image to the LAB color space, applying limited contrast adaptive histogram equalization to the luminance channel and then converting it back to the RGB color space.

[0012] Furthermore, the deep neural network in S2 adopts an encoder-decoder structure. The encoding path extracts multi-scale features through convolutional and pooling layers, the decoding path restores spatial resolution through upsampling and skip connections, and the output layer outputs a pixel-level segmentation probability map.

[0013] Further: S3 includes: S31: Perform feature point detection and description on multiple standardized intraoral photographs to obtain a feature point set for each image; S32: Perform feature matching on feature points between different images in the feature point set, and eliminate erroneous matches through geometric consistency verification to obtain valid matching point pairs; S33: Recover the camera pose parameters of each image based on the effective matching point pairs; S34: Based on the effective matching point pairs and the camera pose parameters, the matching point pairs are back-projected to the three-dimensional space using the triangulation method to generate a sparse point cloud on the dental surface, and global optimization is performed. S35: Based on the camera pose parameters, perform dense matching on the standardized intraoral photograph to generate a dense point cloud on the dental surface; S36: Perform meshing and texture mapping on the dense point cloud to obtain the three-dimensional tooth crown model.

[0014] Further: When the number of intraoral images is less than a preset threshold or feature matching fails, a view augmentation method is used to generate a synthetic view. The view augmentation method includes: extracting visual features of existing views, using the visual features and the target camera pose as input conditions, and generating a synthetic intraoral image under the target camera pose through a pre-trained deep learning image generation model.

[0015] Further: S4 includes: S41: Identify anatomical landmarks in the three-dimensional crown model and the root model from the CBCT image, respectively; S42: Perform coarse registration based on the anatomical landmarks to obtain the initial transformation matrix; S43: Using the initial transformation matrix as the initial value for iteration, fine registration is performed using the iterative nearest point algorithm, where different distance weights are set for different anatomical sites to obtain the final transformation matrix; S44: Calculate the average registration error based on the final transformation matrix. If the average registration error exceeds a preset threshold, it is marked as a registration failure. Otherwise, the three-dimensional crown model and the root model are fused according to the final transformation matrix to obtain the crown-root fusion model.

[0016] Furthermore, the construction of the tooth movement mechanics model in S5 includes: establishing a periodontal ligament elastic model for each tooth, calculating the local elastic coefficient based on the root surface area and bone parameters, calculating the rigid body displacement and elastic deformation correction amount according to the applied orthodontic force vector, and using a regression model trained based on historical cases to predict and correct the residuals of the rigid body displacement and elastic deformation correction amount.

[0017] A dental arch reconstruction and orthodontic simulation system includes: The image acquisition module is used to acquire multiple intraoral photos taken by the user and perform sharpness and exposure detection; The image preprocessing module is used to preprocess the intraoral photographs to obtain standardized intraoral photographs; The tooth segmentation module is used to perform tooth segmentation processing on the standardized intraoral photograph using a deep neural network to obtain a tooth region mask. The 3D reconstruction module is used to perform 3D reconstruction of multiple standardized intraoral photographs based on multi-view geometric methods to obtain a 3D tooth crown model. The registration and fusion module is used to register and fuse the three-dimensional crown model with the root model in the CBCT image to obtain a crown-root fused model; The orthodontic simulation module is used to construct a tooth movement biomechanics model based on the crown-root fusion model to simulate orthodontic effects and output orthodontic prediction results.

[0018] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0019] Compared with the prior art, the present invention has the following advantages: I. This invention uses photos taken within a mobile device as the input source and directly reconstructs a three-dimensional dental crown model from a two-dimensional photo using a multi-view geometric method. This achieves a breakthrough in the technical path from two-dimensional to three-dimensional. Patients can obtain three-dimensional dental information by taking photos themselves with their smartphones, without relying on professional equipment such as dental scanners, which greatly reduces the equipment threshold and usage cost of obtaining three-dimensional oral data.

[0020] Second, this invention registers and fuses the optically reconstructed three-dimensional crown model with the root model in CBCT images, establishing a complete crown-root-bone anatomical relationship. This overcomes the limitations of existing technologies that only analyze the surface morphology of the crown, providing a complete data foundation including deep anatomical structures for orthodontic biomechanical analysis.

[0021] Third, this invention constructs a tooth movement mechanics model based on the periodontal ligament elasticity model and combines it with a regression model trained on historical cases for residual correction. It simulates the real response of teeth under orthodontic forces from a biomechanical perspective. Compared with the path planning method of pure geometric optimization, the prediction results are more consistent with the actual treatment effect and can provide a more reliable reference for clinical decision-making. Attached Figure Description

[0022] Figure 1 A flowchart of a dental arch reconstruction and orthodontic simulation method provided by the present invention.

[0023] Figure 2 This is a schematic diagram of the structure of a dental arch reconstruction and orthodontic simulation system provided by the present invention. Detailed Implementation

[0024] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0025] Example 1 like Figure 1 The present invention provides a method for dentition reconstruction and orthodontic simulation, comprising the following steps: acquiring multiple intraoral photographs taken by the user; preprocessing the intraoral photographs to obtain standardized intraoral photographs; performing tooth segmentation processing on the standardized intraoral photographs to obtain tooth region masks; performing three-dimensional reconstruction of the multiple standardized intraoral photographs based on a multi-view geometric method to obtain a three-dimensional crown model; registering and fusing the three-dimensional crown model with a root model from CBCT images to obtain a crown-root fusion model; constructing a tooth movement biomechanics model based on the crown-root fusion model, simulating orthodontic effects, and outputting orthodontic prediction results. This method realizes the reconstruction from two-dimensional intraoral photographs to a three-dimensional dentition model and combines CBCT data to complete orthodontic effect prediction, allowing patients to obtain three-dimensional dentition information without an oral scanner.

[0026] In one specific embodiment of this example, image preprocessing includes sequentially performing white balance correction, local contrast enhancement, highlight area repair, and lens distortion correction on the intraocular photograph. Local contrast enhancement includes converting the image to the LAB color space, applying limited contrast adaptive histogram equalization to the luminance channel, and then converting it back to the RGB color space. White balance correction eliminates color temperature deviations caused by different light sources, contrast enhancement improves local details, highlight repair reduces interference from glaze reflections, and distortion correction eliminates lens geometric distortion, ensuring the image meets the requirements for subsequent processing.

[0027] In one specific embodiment of this example, tooth segmentation employs a deep neural network with an encoder-decoder structure. The encoding path extracts multi-scale features through convolutional and pooling layers, while the decoding path restores spatial resolution through upsampling and skip connections. The output layer outputs a pixel-level segmentation probability map. Skip connections transfer high-resolution features from the encoding stage to the decoding stage, enhancing the ability to recover boundary details and achieving accurate segmentation of the tooth region.

[0028] In one specific implementation of this embodiment, the three-dimensional reconstruction includes: performing feature point detection and description on multiple standardized intraoral photographs to obtain a feature point set for each image; Feature matching is performed on feature points of different images in the feature point set, and erroneous matches are eliminated through geometric consistency verification to obtain valid matching point pairs; Camera pose parameters of each image are recovered based on effective matching point pairs; Based on the effective matching point pairs and camera pose parameters, the matching point pairs are back-projected into 3D space using the triangulation method to generate a sparse point cloud on the dental surface, and then global optimization is performed. Dense matching is performed on standardized intraoral photographs based on camera pose parameters to generate dense point clouds on the dental surface. Dense point clouds are meshed and textured to obtain a 3D dental crown model. Global optimization employs bundle adjustment to minimize reprojection error, improving the overall accuracy of the point cloud and pose.

[0029] In one specific implementation of this embodiment, when the number of intraoral images is less than a preset threshold or feature matching fails, a view augmentation method is used to generate a synthetic view. The view augmentation method extracts the visual features of existing views, uses the visual features and the target camera pose as input conditions, and generates a synthetic intraoral image under the target camera pose through a pre-trained deep learning image generation model to compensate for the reconstruction failure problem caused by sparse viewpoints.

[0030] In one specific embodiment of this example, registration fusion includes: identifying anatomical landmarks in the three-dimensional crown model and the root model of the CBCT image, respectively; Coarse registration is performed based on the anatomical landmarks to obtain the initial transformation matrix; Using the initial transformation matrix as the initial value for iteration, the iterative nearest point algorithm is used for fine registration, in which different distance weights are set for different anatomical sites to obtain the final transformation matrix; The average registration error is calculated based on the final transformation matrix. When the average registration error exceeds a preset threshold, it is marked as a registration failure. Otherwise, the three-dimensional crown model and the root model are fused according to the final transformation matrix to obtain the crown-root fusion model. Anatomical landmarks include cusps, occlusal plane feature points, and incisal edge feature points. A weighted strategy enables higher registration accuracy for key areas such as the incisal edge of the anterior teeth.

[0031] In one specific embodiment of this example, constructing a tooth movement mechanics model includes: establishing a periodontal ligament elastic model for each tooth; calculating the local elastic coefficient based on the root surface area and bone parameters; calculating the rigid body displacement and elastic deformation correction amount based on the applied orthodontic force vector; and using a regression model trained on historical cases to predict and correct the residuals of the rigid body displacement and elastic deformation correction amount. The periodontal ligament elastic model simulates the real biomechanical response, and the regression model corrects for prediction biases caused by complex biological factors.

[0032] In one specific embodiment of this example, prior to image preprocessing, the method further includes: guiding the user via a mobile application to take intraoral photographs from preset perspectives and performing sharpness and exposure checks on the captured images. The preset perspectives include the upper anterior teeth perspective, the lower anterior teeth perspective, the left lateral view perspective, the right lateral view perspective, and the closed occlusion perspective. The shooting guidance ensures that the user obtains multi-view images that meet the reconstruction requirements, and quality checks eliminate unqualified images.

[0033] Example 2 like Figure 2As shown, this invention also provides a dentition reconstruction and orthodontic simulation system, including an image acquisition module, an image preprocessing module, a tooth segmentation module, a 3D reconstruction module, a registration and fusion module, and an orthodontic simulation module. The image acquisition module acquires multiple intraoral photographs taken by the user and performs sharpness and exposure checks. The image preprocessing module preprocesses the intraoral photographs to obtain standardized intraoral photographs. The tooth segmentation module uses a deep neural network to segment the standardized intraoral photographs to obtain tooth region masks. The 3D reconstruction module performs 3D reconstruction of multiple standardized intraoral photographs based on multi-view geometric methods to obtain a 3D crown model. The registration and fusion module registers and fuses the 3D crown model with a root model from CBCT images to obtain a crown-root fusion model. The orthodontic simulation module constructs a tooth movement biomechanics model based on the crown-root fusion model to simulate orthodontic effects and outputs orthodontic prediction results. All modules work collaboratively to achieve a complete processing flow from intraoral photograph acquisition to orthodontic prediction.

[0034] Example 3 The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for three-dimensional dental reconstruction and orthodontic simulation based on intraoral photographs.

[0035] Example 4 This embodiment further explains and illustrates the algorithm and process used in this invention. The specific process of algorithm implementation is as follows: I. Intraoral Image Enhancement and Distortion Correction Module.

[0036] Module objective: To enhance and geometrically correct intraoral photographs taken by the user, so that they achieve the image quality and geometric accuracy required for subsequent segmentation, registration, and 3D reconstruction.

[0037] Detailed steps: Step 1: Reading and Preliminary Verification: Receive the original image uploaded by the user and check the file integrity and basic metadata (resolution, file format). If the image resolution is lower than the application's set threshold (e.g., minimum width 1024 pixels), the system marks it as unqualified and prompts for a retake.

[0038] Step 2: Automatic White Balance (Color Constantization): Perform grayscale world or grayscale-based white balance correction on the image. Specifically, calculate the average or median of the three channels of the image and adjust the channel gain proportionally to make the average brightness more consistent. This step aims to reduce color temperature deviations caused by different light sources.

[0039] Step 3: Local Contrast Enhancement: Convert the image from BGR color space to LAB color space, and apply Adaptive Histogram Equalization or CLAHE (Constrained Contrast Adaptive Histogram Equalization) to the L channel to enhance local details and avoid over-amplifying noise. Then convert LAB back to RGB.

[0040] Step 4: Specular Highlight and Reflection Suppression: Detect overexposed areas (e.g., pixel brightness exceeding a set threshold, such as 245 / 255) and mark them as highlights. Replace the highlight pixels in the highlight areas using neighborhood-based inpainting or a reflection-diffuse reflection separation algorithm to reduce the interference of specular reflection on texture matching.

[0041] Step 5: Lens Distortion Correction: Perform anti-distortion processing on the image using the device's camera intrinsic parameters (if available) or a pre-calibrated standard camera matrix and distortion coefficients. If the device intrinsic parameters are unknown, the system can guide the user to complete the calibration on the first use using a self-calibration process based on a checkerboard pattern or a known calibration image, or use approximate standard parameters and fine-tune them through registration correction in subsequent steps.

[0042] Step 6: Noise Reduction and Edge Preservation: Apply edge-preserving filtering (such as bilateral filtering) or non-local mean denoising to the image. The purpose is to remove sensor noise while preserving tooth edge information, which will directly affect the accuracy of key point detection and segmentation.

[0043] Step 7: Output normalized image and mask: Adjust the processed image to a standard size and color space, generate and save the processing log (including whether highlight restoration was applied, camera parameters used, etc.) for traceability and quality control.

[0044] II. Tooth Semantic / Instance Segmentation Module.

[0045] Module objective: To identify and label teeth, gingiva, lips, tongue, and other instrument / retractor areas on each enhanced intraoral photograph, providing pixel-level masks and tooth contours for 3D reconstruction and morphological analysis.

[0046] Implementation principle and steps: Input preparation: Scale or crop the enhanced image to the model input size (e.g., 512×512 or 1024×1024) and normalize the pixel values ​​(e.g., map the pixel values ​​to [0,1]).

[0047] Feature encoding (downsampling): Local and mesoscale features are extracted through several layers of convolution + standardization + activation function (e.g., ReLU), while spatial downsampling is achieved by using stride or pooling operations to obtain multi-scale feature representations.

[0048] Feature Decoding (Upsampling): Bilinear interpolation or deconvolution is used in the upsampling stage to restore spatial resolution, and high-resolution features from the encoding stage are passed to the decoding stage through skip connections to enhance the recovery of boundary details.

[0049] Output layer and activation: Finally, the feature map is compressed to the number of classes (e.g., 1 foreground / background or multiple classes) through 1×1 convolution. For each pixel, a class probability map is output, using Sigmoid (single class) or Softmax (multi-class).

[0050] Post-processing: Thresholding is performed on the probability map to obtain a binary mask. Then, morphological operations (opening / closing operations) are used to remove isolated noise points and fill small holes. Connectivity analysis is used to identify and label independent tooth instances (instance segmentation can be achieved by combining boundary detection and the watershed algorithm).

[0051] Training Description: The model is trained using a combination of pixel cross-entropy loss and Dice loss. Training data should include labeled masks from multiple devices, multiple light sources, and patients of different age groups. An initial training set of at least 1000 images is recommended; to achieve clinical-grade accuracy, it is suggested to expand to 5,000-10,000 images or more, and use data augmentation (color jitter, random cropping, rotation / flipping).

[0052] III. Multi-view geometric reconstruction module.

[0053] Module objective: Estimate camera pose based on multiple registered intraoral images and triangulate it to obtain a 3D sparse point cloud. Then, obtain a high-quality point cloud and mesh through dense matching and completion.

[0054] 1. Feature detection and description: For each image, a robust local feature detector (such as SIFT, SURF or ORB) is used to detect key points, and descriptors are calculated to express the local texture information of the key points; at the same time, deep learning descriptors (such as SuperPoint) can be combined to improve repeatability.

[0055] 2. Feature matching: Nearest neighbor matching is used between image pairs and a ratio test (Loweratiotest) is applied to eliminate low-quality matching pairs, resulting in a set of candidate matching point pairs.

[0056] 3. Initial Geometric Verification (RANSAC): Use RANSAC to estimate the fundamental matrix or homography matrix (depending on whether the scene is planar) for the matched point pairs and filter out outliers to ensure geometric consistency of the matching.

[0057] 4. Camera pose recovery: Recover the rotation and translation parameters (i.e. extrinsic parameters) of each image relative to the reference coordinate system from the essential matrix / fundamental matrix or directly through PnP (if there is a 2D-3D correspondence).

[0058] 5. Triangulation to generate sparse point cloud: Using the triangulation method of multi-view geometry, the matching point pairs are transformed into three-dimensional points to form a sparse point cloud representation of the coarse geometry of the tooth row.

[0059] 6. Global Optimization (BundleAdjustment): Joint nonlinear optimization of all camera parameters and 3D points is performed to minimize reprojection error, thereby improving the overall consistency and accuracy of sparse point cloud and camera pose.

[0060] 7. Dense Matching (MVS) and Point Cloud Completion: Based on the known camera pose, a dense depth map is generated using parallax or stereo matching methods. Multiple depth maps are fused to obtain a high-density point cloud. Hole filling and noise filtering are performed when necessary.

[0061] 8. Meshing and Texture Mapping: Mesh triangulation of high-density point clouds (e.g., Poisson reconstruction or Delaunay triangulation) and mapping of the original image texture onto the mesh surface to reconstruct a 3D crown model with color information.

[0062] IV. Visual augmentation and neural implicit reconstruction.

[0063] Note: For cases where traditional SfM fails due to sparse shooting perspectives (e.g., only 5 or fewer images) or repetitive tooth surface textures, the system employs learning-based view augmentation and neural implicit field reconstruction as enhancement methods.

[0064] View augmentation (concept): Based on existing real-world views and corresponding camera poses, a conditional image generation model is trained to synthesize realistic intraoral images under any target camera pose. The model uses the visual features of the input view and the target pose as conditions to output a synthesized view image.

[0065] View augmentation (training and loss): During training, real multi-view data pairs are used, and the loss function consists of pixel reconstruction loss, edge consistency loss and mask consistency loss; to ensure geometric consistency, beam consistency or parallax consistency constraints are also introduced.

[0066] Neural implicit field reconstruction (concept): After obtaining more synthetic views, a continuous 3D field is learned from the multi-view information using implicit representations (such as SDF or differentiable volume rendering functions), where the zero isosurface of the field represents the target surface.

[0067] Neural implicit field (training loss): The training loss includes photometric consistency loss (the difference between the pixels of the rendered image and the real image), Eikonal loss (constraining the gradient to 1 to ensure the SDF property), and mask consistency loss.

[0068] V. CBCT and optical surface fusion algorithm.

[0069] Objective and Prerequisites: The objective is to align the reconstructed crown surface from photographs with the root / bone structures in CBCT. Prerequisites include: CBCT has completed DICOM import and segmentation of the tooth / bone regions; the optical model is in mesh or point cloud format.

[0070] Coarse registration (feature matching): First, common anatomical landmarks (such as cusps, occlusal planes, or incisal edge features) are identified in both modalities. Based on these landmarks, rigid registration is performed to obtain an initial transformation matrix.

[0071] Fine registration (weighted ICP): In the fine registration stage, the Iterative Closest Point (ICP) algorithm is used, but weights based on site importance are introduced into the distance metric (e.g., the incisal edge of the anterior tooth has a higher weight than the lingual root), and the range of non-rigid local adjustments for each tooth is limited to avoid overall mismatch.

[0072] Registration quality control: Calculate the statistical distribution of the average point-to-surface distance, maximum error, and registration residual. If the average error exceeds a set threshold (e.g., 1.0 mm), the system marks the registration as failed and prompts for manual review or re-collection of data.

[0073] Fusion Output: After registration is successful, the system outputs an anatomical model in a unified coordinate system. The model includes a crown mesh, a root voxel model, and an alveolar bone voxel, and generates a rendered view that visualizes the root-crown relationship.

[0074] VI. Orthodontic tooth movement prediction algorithm.

[0075] Overall approach: A hybrid approach combining physical approximation (rigidity + elasticity) and data-driven residual correction is adopted to ensure interpretability while correcting complex biological responses by learning from historical cases.

[0076] Input items include the current three-dimensional position of the tooth, root morphology (if CBCT is available), alveolar bone mineral density (which can be estimated by CBCT surrogate values), contact / crowding of adjacent teeth, and the setting of the orthodontic force vector and time step to be applied.

[0077] Rigid propulsion: For each tooth, the approximate rigid body displacement under a given force is first calculated, and the preliminary displacement direction and magnitude are obtained based on a simple torque balance and resistance model.

[0078] Elastic correction (PDL simulation): A linear elastic model is introduced for each tooth to represent the periodontal ligament (PDL). The local elastic coefficient is calculated based on the root surface area and bone surrogate value, and the elastic deformation correction amount under small displacement is solved.

[0079] ML Residual Corrector: A trained regression model (such as time series regression or Transformer) is used to predict and correct the residuals of the rigidity + elasticity calculation. The model training data is the mapping between actual displacement and virtual force in historical cases.

[0080] Time iteration and convergence determination: The displacement of each time step is superimposed, and the local stress and displacement changes are evaluated after each step. If the convergence reaches the set threshold or the maximum number of iterations is reached, the iteration stops.

[0081] Uncertainty quantification and risk warning: Output the prediction uncertainty (e.g., standard deviation) for each tooth and indicate the potential risk of root resorption or bone damage based on root morphology, bone weakness, or high torque.

[0082] The specific implementation process used by users is as follows: 1. User-side shooting and uploading.

[0083] 1) Users take photos according to the application's shooting prompts (standard angles include upper front teeth, lower front teeth, left and right sides, and closed bite position).

[0084] 2) The client performs real-time sharpness and exposure checks. If the check fails, it prompts the user to retake the photo and provides improvement suggestions (such as adding a light source or adjusting the phone angle).

[0085] 3) Qualified photos are compressed to the standard size and uploaded to the server. During the upload process, EXIF ​​and device information are recorded for subsequent correction.

[0086] 2. Image enhancement and preprocessing.

[0087] 1) After receiving the image, the server performs automatic white balance, local contrast enhancement, reflection repair, and distortion correction.

[0088] 2) The system generates a processing log for each image (including the specific parameters of the application and whether highlight restoration was performed) and archives it for quality auditing.

[0089] 3. Tooth segmentation and key point detection.

[0090] 1) Run a semantic / instance segmentation model on each enhanced image to obtain tooth masks and tooth contours.

[0091] 2) Further extract key points of the teeth (incisal edge, cusp, cervical edge) and tooth boundaries on the mask for subsequent registration and tooth instance segmentation.

[0092] 4. Cross-view registration and camera pose estimation.

[0093] 1) Perform viewpoint registration using local feature descriptors and their matching pairs, and use RANSAC to remove incorrect matches.

[0094] 2) If there is a pre-existing pose estimate (e.g., an approximate pose provided by AR guidance), use it as initialization to improve the SfM convergence rate.

[0095] 3) Perform BundleAdjustment to optimize all camera extrinsic parameters.

[0096] 5. Sparse / dense point cloud construction and meshing.

[0097] 1) Use triangulation to generate sparse point clouds from matching points and then perform multi-view stereo (MVS) densification.

[0098] 2) Filter out noise points and segment the point cloud to extract the dental arch region.

[0099] 3) Generate a closed mesh from the point cloud using Poisson or other reconstruction methods and then smooth it.

[0100] 6. CBCT import and registration.

[0101] 1) Import the CBCTDICOM file uploaded by the user or provided by the clinic into the system, and perform threshold / deep learning segmentation to extract tooth and bone voxels.

[0102] 2) Calculate the initial rigid registration between the optical model and the CBCT model and perform weighted ICP fine registration.

[0103] 3) When the registration is successful, generate a fusion model and update the tooth number and the correspondence between the root and crown.

[0104] 7. Tooth movement simulation and report generation.

[0105] 1) Displacement simulation is performed based on the fusion model and the user-defined treatment plan (e.g., force direction, appliance type).

[0106] 2) Record the displacement and stress values ​​at each time step during the simulation and calculate the uncertainty.

[0107] 3) Generate a diagnostic report that includes a 3D view, displacement curve, risk warnings and suggestions, and provide it for clinical staff to review.

[0108] The selection of key parameters, thresholds, and output terms is as follows: Image resolution: It is recommended to upload an original image with a width of ≥1024 pixels; the model input can be scaled to 512 or 1024 resolution to balance speed and accuracy.

[0109] Reconstruction error alarm threshold: When comparing with the IOS gold standard, a quality warning is triggered when the average point-to-point error is >1.0mm; if the average registration error is >1.2mm, manual review is recommended.

[0110] Shooting pose tolerance: AR-guided recommended pose tolerance: ±10° in each direction of pitch / yaw / roll, and ±20mm tolerance in distance from the camera center.

[0111] Segmentation threshold and training details: The probability threshold for segmentation models is commonly 0.5; the training loss uses a combination of BCE and Dice loss; the recommended initial learning rate is 1e-4, and the batch size is 8–16.

[0112] Predictive model training samples: It is recommended that the initial number of longitudinal cases used for training be no less than 300, and each case should include at least 3 time points: preoperative, midoperative, and postoperative, for time-series learning.

[0113] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent transformations or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method of dentition reconstruction and orthodontic simulation, characterized by, The method comprises the following steps: S1: obtaining multiple intraoral photos taken by a user, performing image preprocessing on the intraoral photos to obtain standardized intraoral photos; S2: performing tooth segmentation processing on the standardized intraoral photos to obtain a tooth region mask; S3: performing three-dimensional reconstruction on the multiple standardized intraoral photos based on a multi-view geometry method to obtain a three-dimensional crown model; S4: registering and fusing the three-dimensional crown model with a tooth root model in a CBCT image to obtain a crown-tooth root fusion model; S5: constructing a tooth movement mechanics model based on the crown-tooth root fusion model, performing orthodontic effect simulation, and outputting an orthodontic prediction result.

2. The method of claim 1, wherein: The image preprocessing in S1 comprises: sequentially performing white balance correction, local contrast enhancement, highlight area repair, and lens distortion correction on the intraoral photos; wherein the local contrast enhancement comprises converting the image to an LAB color space, applying a limited contrast adaptive histogram equalization processing to the luminance channel, and then converting back to an RGB color space.

3. The method of claim 1, wherein: The deep neural network in S2 adopts an encoder-decoder structure, the encoding path extracts multi-scale features through convolution layers and pooling layers, the decoding path restores the spatial resolution through upsampling and skip connection, and the output layer outputs a pixel-level segmentation probability map.

4. The method of claim 1, wherein: S3 comprises: S31: detecting and describing feature points of the multiple standardized intraoral photos to obtain a feature point set of each image; S32: performing feature matching on the feature points in the feature point set between different images, and eliminating false matches through geometric consistency verification to obtain valid matching point pairs; S33: restoring camera pose parameters of each image based on the valid matching point pairs; S34: based on the valid matching point pairs and the camera pose parameters, using a triangulation method to back-project the matching point pairs to a three-dimensional space, generating a sparse point cloud of a tooth surface, and performing global optimization; S35: performing dense matching on the standardized intraoral photos based on the camera pose parameters to generate a dense point cloud of the tooth surface; S36: performing meshing processing and texture mapping on the dense point cloud to obtain the three-dimensional crown model.

5. The method of claim 4, wherein: When the number of intraoral photos is less than a preset threshold or feature matching fails, a view expansion method is used to generate a synthetic view, the view expansion method comprising: extracting visual features of existing views, taking the visual features and a target camera pose as input conditions, and generating a synthetic intraoral image under the target camera pose through a pre-trained deep learning image generation model.

6. The method of claim 1, wherein: S4 comprises: S41: identifying anatomical landmark points in the three-dimensional crown model and the tooth root model of the CBCT image, respectively; S42: performing coarse registration based on the anatomical landmark points to obtain an initial transformation matrix; S43: taking the initial transformation matrix as an initial value for iteration, performing fine registration using an iterative closest point algorithm, wherein different distance weights are set for different anatomical parts to obtain a final transformation matrix; S44: calculating an average registration error based on the final transformation matrix, marking as registration failure when the average registration error exceeds a preset threshold, or otherwise fusing the three-dimensional crown model and the tooth root model according to the final transformation matrix to obtain a crown-tooth root fusion model.

7. The method of claim 1, wherein: The S5 includes: establishing a periodontal membrane elastic model for each tooth, calculating a local elastic coefficient based on a tooth root surface area and bone quality parameters, and calculating a rigid body displacement and an elastic deformation correction amount according to an applied correction force vector; and using a regression model trained based on historical cases to predict and correct the residual error of the rigid body displacement and the elastic deformation correction amount.

8. The method of claim 1, wherein: Before the S1, the method further includes: guiding a user to shoot intraoral photos from preset visual angles through a mobile terminal application, and detecting the clarity and exposure of the shot images, wherein the preset visual angles include an upper anterior tooth visual angle, a lower anterior tooth visual angle, a left lateral visual angle, a right lateral visual angle, and a closed bite visual angle.

9. A dentition reconstruction and orthodontic simulation system implementing the dentition reconstruction and orthodontic simulation method of any one of claims 1-8, characterized in that, The method comprises: an image acquisition module configured to acquire a plurality of intraoral photos shot by a user and detect the clarity and exposure of the intraoral photos; an image preprocessing module configured to preprocess the intraoral photos to obtain standardized intraoral photos; a tooth segmentation module configured to perform tooth segmentation processing on the standardized intraoral photos using a deep neural network to obtain tooth region masks; a three-dimensional reconstruction module configured to perform three-dimensional reconstruction on a plurality of the standardized intraoral photos based on a multi-view geometry method to obtain a three-dimensional crown model; a registration and fusion module configured to register and fuse the three-dimensional crown model and a tooth root model in a CBCT image to obtain a crown-tooth root fusion model; an orthodontic simulation module configured to construct a tooth movement mechanics model based on the crown-tooth root fusion model, simulate orthodontic effects, and output orthodontic prediction results.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, and the computer program is executed by the processor to implement the dental arch reconstruction and orthodontic simulation method in any one of claims 1-8.

Images