A Deep Learning-Based Method and System for Monitoring Radicular Relationship in Orthodontic Treatment Follow-up

Abstract

This invention discloses a deep learning-based method and system for monitoring and diagnosing root-bone relationships in orthodontic treatment. It innovatively integrates jawbone anatomical information into the assessment system, which can replace the traditional method of repeated CBCT scans for follow-up visits. By fusing data from the initial CBCT scan and intraoral scans (IOS) during treatment, it automatically and accurately predicts the position of the tooth root and directly quantifies the distance between the tooth root and the buccal and lingual cortical bone. This achieves a complete transformation from geometric accuracy verification to clinical risk quantification, providing radiation-free, high-precision, and automated risk assessment and visual early warning for bone fenestration and bone fracture. It provides orthodontists with a follow-up monitoring tool that can be truly used for clinical decision-making.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical treatment, in particular to an orthodontic treatment root bone relationship follow-up visit monitoring method and system based on deep learning. BACKGROUND

[0002] At present, the core of orthodontic treatment is the accurate control and dynamic evaluation of teeth, especially the roots in three-dimensional space, which is directly related to the long-term stability of treatment effect, periodontal health and prevention of complications. During the orthodontic treatment process, the doctor needs to monitor whether the tooth movement causes the roots to excessively approach the buccal and lingual cortical bone, so as to prevent the occurrence of serious complications such as bone fenestration and bone cracking. Cone beam computed tomography (CBCT) is the "gold standard" for obtaining three-dimensional information of tooth roots and their surrounding bone tissue. The traditional follow-up visit monitoring method mainly relies on regular CBCT scanning to evaluate the root-cortical bone distance.

[0003] In recent years, with the development of digital oral technology, researchers have begun to explore root position prediction technology based on the fusion of intraoral scanning (IOS) and CBCT. The existing related technology mainly fuses the initial diagnosis CBCT root information and the oral scanning crown information through registration, constructs a virtual root model, and verifies the prediction accuracy through surface shell difference analysis of the real tooth root. These technologies have made certain progress in tooth root geometry reconstruction, but are mainly limited to the verification level of geometric accuracy.

[0004] However, the existing technology has significant limitations: on the one hand, the frequent scanning in the traditional CBCT follow-up visit method brings the risk of cumulative ionizing radiation, especially for adolescent patients who need close monitoring; on the other hand, although the existing fusion technology can predict the geometry of the tooth root, it does not include the jaw bone anatomy in the evaluation system, and cannot directly quantify the core risk indicator of the root-cortical bone distance that clinicians are most concerned about, making it difficult to be transformed into a clinical tool that truly serves the risk assessment of bone fenestration and bone cracking.

[0005] In the traditional orthodontic follow-up visit monitoring method, orthodontic doctors need to take CBCT regularly to evaluate whether there is a risk of bone fenestration and bone cracking during tooth movement, mainly by analyzing the accurate distance between the tooth root and the buccal and lingual cortical bone to determine the risk level. In recent years, some similar technologies based on the fusion of oral scanning data and CBCT have emerged. These methods can use oral crown data to construct a virtual tooth root and perform shell difference analysis with the real CBCT segmented tooth root, to a certain extent, verifying the geometric accuracy of tooth root position prediction. However, the existing technology still has the following key defects:

[0006] Firstly, the previous similar technology research focuses mainly on the reconstruction accuracy of the root geometry, and the evaluation dimension is limited to the surface shell difference between the virtual root and the real root, the root spacing, the tooth long axis angle and other geometric accuracy indexes, and the jaw bone anatomy is not included in the evaluation system, which cannot directly quantify the core risk index of the tooth root-bone cortex distance which is most concerned in the clinic.

[0007] Secondly, the traditional CBCT recheck method still needs frequent scanning, which leads to high cumulative radiation risk, especially for adolescent patients and women of childbearing age, which poses a threat to health.

[0008] Thirdly, the process of manually extracting key clinical risk indicators such as the distance between the tooth root and the buccal and lingual bone cortex from the CBCT image is tedious, time-consuming and highly dependent on the operator's experience, which is highly subjective, affecting the objectivity and consistency of the evaluation.

[0009] These factors together make it difficult for existing technology to transform from simple geometric accuracy verification to a practical tool that truly serves clinical risk assessment, limiting the timely warning and early intervention of treatment risks. SUMMARY

[0010] To solve these problems, the present application proposes a deep learning-based orthodontic treatment root-bone relationship recheck monitoring method and system.

[0011] According to an aspect of the present application, a deep learning-based orthodontic treatment root-bone relationship recheck monitoring method is proposed, comprising the following steps:

[0012] S1, collecting CBCT data and IOS data of a patient and preprocessing the CBCT data and IOS data;

[0013] S2, performing segmentation processing on the CBCT data to generate CBCT tooth root grid data, and performing segmentation processing on the IOS data to generate IOS tooth crown grid data;

[0014] S3, registering the CBCT tooth root grid data and the IOS tooth crown grid data based on the ICP algorithm to obtain a baseline model;

[0015] S4, optimizing the ICP algorithm based on the movement characteristics of the teeth, the stability requirements and the individual differences, obtaining an optimized ICP algorithm, acquiring the patient's recheck IOS data, and predicting the current tooth root position based on the baseline model, the recheck IOS data and the optimized ICP algorithm.

[0016] The application provides an orthodontic follow-up monitoring method and system based on deep learning, which realizes radiation-free follow-up monitoring through a technical route of constructing an initial diagnosis baseline model, predicting tooth root positions in real time, and quantifying risk indicators. The overall architecture of the system follows the design concept of "one CBCT modeling, multiple IOS predictions", which converts the complex three-dimensional tooth root-cortical relationship into a visual risk assessment tool.

[0017] Specifically, the preprocessing of the CBCT data in S1 specifically includes: threshold segmentation and morphological opening operation are performed on the CBCT data to remove metal artifacts, then isotropic resampling is performed to a uniform voxel spacing and linear transformation normalization mapping is performed, and finally threshold method is used to extract the jaw cortical region in the CBCT data as the region of interest;

[0018] The preprocessing of the IOS data in S1 specifically includes: converting the IOS data into a voxel grid form, wherein the voxel density value of the IOS data is calculated by nearest neighbor interpolation, and the local contrast is enhanced based on an adaptive histogram equalization algorithm, then a principal component matrix of the IOS data is calculated and the orientation consistency of the IOS data in the three-dimensional space is realized by limiting the principal component matrix of the IOS data through a rotation matrix.

[0019] Specifically, S3 specifically includes:

[0020] The centroids and principal direction matrices of the CBCT tooth root grid data and the IOS tooth crown grid data are calculated based on a principal component analysis algorithm, the corresponding initial transformation matrix is calculated based on the centroids and principal direction matrices of the CBCT tooth root grid data and the IOS tooth crown grid data, coarse registration alignment is performed based on the initial transformation matrix, then accurate registration is performed based on the ICP algorithm, the overlapping region of the CBCT tooth root grid data and the IOS tooth crown grid data is smoothly transitioned using a thin-plate spline deformation algorithm, then the fusion weight function is set to fuse the overlapping region, and finally the Breguet reconstruction algorithm is used for optimization to obtain the baseline model.

[0021] Specifically, S4 specifically includes: setting a tooth movement amount decay function based on the movement characteristics of teeth, an anatomical stability function based on stability requirements, and an individual adjustment factor based on individual differences requirements, optimizing the weight items of the objective function of the ICP algorithm based on the tooth movement amount decay function, the anatomical stability function and the individual adjustment factor, adding a regularization term to the objective function of the ICP algorithm, obtaining the optimized ICP algorithm, registering the baseline model and the follow-up IOS data, calculating the spatial transformation parameters based on the optimized ICP algorithm, and applying the spatial transformation parameters to the baseline model to predict the current tooth root position.

[0022] Specifically, the tooth movement amount attenuation function is an exponential function that attenuates based on an increase in the movement distance of the tooth root position, the anatomical stability function is graded based on different tooth positions of the tooth, and the individual adjustment factor is obtained based on individual differences to obtain different values, with a value range of 0.8-1.2.

[0023] Specifically, it further comprises: S5, segmenting the CBCT data, generating CBCT jaw bone mesh data, obtaining a jaw bone model based on the CBCT jaw bone mesh data, calculating the shortest distance from each point on the current predicted tooth root surface to the cortical surface of the jaw bone based on the predicted current tooth root position and the jaw bone model, setting a risk distance threshold, and performing risk level assessment.

[0024] Specifically, S2 specifically comprises: segmenting the tooth region and the cortical region of the jaw bone in the CBCT data using the nnUNet framework, converting the mesh data using the Marching Cubes algorithm, optimizing the data through edge folding algorithm and performing Laplacian smoothing processing, generating the CBCT tooth root mesh data and the CBCT jaw bone mesh data.

[0025] In S2, the IOS data is segmented to generate IOS crown mesh data, specifically comprising: using TSegNet algorithm to segment the IOS data to generate the IOS crown mesh data.

[0026] Specifically, in S5, the shortest distance from each point on the current predicted tooth root surface to the cortical surface of the jaw bone is calculated, specifically comprising:

[0027] The cortical surface of the jaw bone model is segmented into a plurality of triangular facets using octree;

[0028] The plane distance between each point on the current predicted tooth root surface and the plurality of triangular facets is calculated;

[0029] Based on the plane distance, the projection point of each point on the current predicted tooth root surface on the plane corresponding to the plurality of triangular facets is determined;

[0030] Based on the barycentric coordinate method, it is judged whether the projection point is on the corresponding triangular facet, if so, it is an effective projection point;

[0031] The distance between each point on the current predicted tooth root surface and the effective projection point of the plurality of triangular facets is calculated and compared to obtain the shortest distance from each point on the current predicted tooth root surface to the cortical surface of the jaw bone.

[0032] Specifically, S5 further comprises pre-processing the jawbone model to remove irrelevant risk areas, the irrelevant risk areas including alveolar crest top area and maxillary sinus bottom boundary area in the jawbone model, identifying the alveolar crest top area in the jawbone model through curvature analysis, and locating the maxillary sinus bottom boundary area in the jawbone model through anatomical landmark points, and removing the irrelevant risk areas using morphological operations and region growing algorithm.

[0033] Specifically, setting the risk distance threshold in S5 specifically comprises: grading and assigning values based on the shortest distance values from each point on the root surface to the cortical surface of the jawbone, and adjusting the threshold of the grading and assigning values based on the age of the patient, periodontal phenotype, and type of treatment performed to obtain the risk distance threshold.

[0034] According to an aspect of the present application, a deep learning-based orthodontic treatment root bone relationship follow-up monitoring system is provided, according to the method of any one of the first aspect, comprising the following modules:

[0035] A data processing module configured to collect CBCT data and IOS data of a patient and pre-process the CBCT data and IOS data;

[0036] A data segmentation module configured to segment the CBCT data to generate CBCT root mesh data, and segment the IOS data to generate IOS crown mesh data;

[0037] A model construction module configured to register the CBCT root mesh data and IOS crown mesh data based on an ICP algorithm to obtain a baseline model;

[0038] A tooth root prediction module configured to optimize the ICP algorithm based on the movement characteristics of the teeth, stability requirements, and individualized difference requirements, obtain an optimized ICP algorithm, acquire follow-up IOS data of the patient, and predict the current tooth root position based on the baseline model, the follow-up IOS data, and the optimized ICP algorithm.

[0039] The present application has the following advantages:

[0040] (1) The purpose of the present application is to achieve a key breakthrough on the basis of previous geometric registration technology, and to provide a deep learning-based automatic orthodontic follow-up monitoring method and system, which innovatively integrates jawbone anatomical information into the evaluation system.

[0041] (2) By fusing the initial diagnosis CBCT and the intraoral scanning (IOS) data during the initial diagnosis process to establish a baseline model, the baseline model and the IOS data during the follow-up process are registered based on a dynamic weight registration algorithm oriented by clinical targets to automatically and accurately predict the position of the dental root, predict the movement of the tooth, and realize the prediction of the position of the dental root through one initial diagnosis CBCT combined with subsequent IOS data, completely avoiding the exposure of ionizing radiation in the follow-up stage, and being able to replace the traditional repeated CBCT scanning follow-up mode.

[0042] (3) Directly quantifying the distance between the dental root and the buccal and lingual cortical bone, and predicting the risk based on the distance between the dental root and the buccal and lingual cortical bone, so as to realize the complete transformation from geometric accuracy verification to clinical risk quantification, provide radiation-free, high-precision, automated bone windowing, bone cracking risk assessment and visual warning, and provide orthodontists with a follow-up monitoring tool that can be truly used for clinical decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0043] The accompanying drawings are included to provide a further understanding of embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of the application. Other embodiments and many of the intended advantages of the present application will be readily appreciated as the same becomes better understood by reference to the following detailed description when considered in connection with the accompanying drawings. The elements of the drawings are not necessarily to scale relative to each other. Like reference numerals designate corresponding similar parts.

[0044] Figure 1 A flowchart of an orthodontic treatment root bone relationship follow-up monitoring method based on deep learning according to the present application is shown;

[0045] Figure 2 A flowchart of an orthodontic treatment root bone relationship follow-up monitoring method based on deep learning according to the present application is shown;

[0046] Figure 3 A structural diagram of the nnUNet CBCT segmentation network model is shown;

[0047] Figure 4 A visual result diagram of risk assessment according to the method of the present application is shown;

[0048] Figure 5 A structural diagram of an orthodontic treatment root bone relationship follow-up monitoring system based on deep learning according to the present application is shown. DETAILED DESCRIPTION

[0049] The application will be described in further detail below with reference to the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the related application, and not to limit the application. In addition, it should be noted that only parts related to the application are shown in the drawings for ease of description.

[0050] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the drawings and embodiments.

[0051] Figure 1 An orthodontic treatment root bone relationship follow-up monitoring method based on deep learning is shown, as shown in Figure 1 and shown, comprising the following steps:

[0052] S1, collecting CBCT data and IOS data of a patient and pre-processing the CBCT data and IOS data;

[0053] S2, segmenting the CBCT data to generate CBCT tooth root grid data, and segmenting the IOS data to generate IOS tooth crown grid data;

[0054] S3, registering the CBCT tooth root grid data and IOS tooth crown grid data based on an ICP algorithm to obtain a baseline model;

[0055] S4, optimizing the ICP algorithm based on the movement characteristics, stability requirements and individual difference requirements of teeth, obtaining an optimized ICP algorithm, acquiring follow-up IOS data of the patient, and predicting the current tooth root position based on the baseline model, the follow-up IOS data and the optimized ICP algorithm.

[0056] Specifically, the pre-processing of the CBCT data in S1 specifically includes: threshold segmentation and morphological opening operation to remove metal artifacts are performed on the CBCT data, then isotropic resampling to a unified voxel spacing and linear transformation normalization mapping are performed, and finally threshold method is used to extract the jaw bone cortical region in the CBCT data as the region of interest;

[0057] The pre-processing of the IOS data in S1 specifically includes: converting the IOS data into a voxel grid form, wherein the voxel density value of the IOS data is calculated by nearest neighbor interpolation, and the local contrast is enhanced based on an adaptive histogram equalization algorithm, then the principal component matrix of the IOS data is calculated and the orientation consistency of the IOS data in the three-dimensional space is realized by limiting the principal component matrix of the IOS data through a rotation matrix.

[0058] Firstly, the initial CBCT data and intraoral scan (IOS) data of the patient are acquired, and the original data is standardized and preprocessed to ensure the robustness and accuracy of the subsequent algorithm. In the CBCT data preprocessing process, first, the metal artifacts are removed based on threshold segmentation (Hounsfield unit > 1,500) and morphological opening operation, then isotropic resampling is performed to a uniform voxel spacing (0.15-0.25 cubic millimeters, preferably 0.2 cubic millimeters), the CT value range is limited to [-1000, 2000], and is normalized to the [0,1] interval through linear transformation: wherein is the original CT value, =-1000, =2000. The threshold method is used to extract the jaw bone region as the region of interest (ROI) to reduce computational redundancy.

[0059] In the IOS data preprocessing, the point cloud data is converted to a 256³ voxel grid, and the voxel density value is calculated by nearest neighbor interpolation: wherein is a weight factor, is a kernel function, . Adaptive histogram equalization (CLAHE) is applied to enhance local contrast, and the contrast is limited to 2.0-4.0. To ensure the orientation consistency of all input data in three-dimensional space, the principal component matrix is calculated for each data, wherein is the first principal component. The axial uniformity is realized by the rotation matrix R: for maxillary data, ensure , for mandibular data, ensure , and the rotation angle θ is calculated as . represents the coordinate value of the data point in the Z-axis direction, represents the target Z-axis coordinate.

[0060] Specifically, S3 specifically comprises:

[0061] calculating the centroids and principal direction matrices of the CBCT tooth root grid data and the IOS tooth crown grid data based on the principal component analysis algorithm, calculating the corresponding initial transformation matrix based on the centroids and principal direction matrices of the CBCT tooth root grid data and the IOS tooth crown grid data, performing coarse registration alignment based on the initial transformation matrix, then performing accurate registration based on the ICP algorithm, then using a thin plate spline deformation algorithm to realize smooth transition in the overlapping region of the CBCT tooth root grid data and the IOS tooth crown grid data, then setting a fusion weight function to fuse the overlapping region, and finally using the B-spline reconstruction algorithm for optimization to obtain the baseline model.

[0062] AsFigure 2 The core innovation of the present application is to fuse the tooth root morphology information provided by CBCT and the high-precision tooth crown surface information provided by IOS. Through the above process, a baseline three-dimensional model containing accurate tooth crown surface and complete tooth root morphology is constructed. Through combination with IOS data for tooth displacement prediction, and in combination with CBCT jaw bone grid data composed of jaw bone cortical structure, it can be applied to risk assessment. That is, after completing the preprocessing of the data, the following steps are performed by the segmentation algorithm: ① segmentation of the jaw bone cortical data in the CBCT data, ② segmentation of the tooth data in the CBCT data, and ③ segmentation of the tooth data in the IOS data, then based on the segmentation results of ② and ③, ④ coarse registration is performed, then ⑤ fine registration is performed, and ⑥ baseline model is obtained, and ⑦ risk assessment is performed in combination with the segmentation results of ① jaw bone data CBCT data.

[0063] The stages ④ and ⑤ adopt a dual-source registration algorithm. First, in the ④ coarse registration stage, initial alignment is performed based on principal component analysis. The centroids of the CBCT tooth root grid data and the IOS tooth crown grid data are calculated and , and the principal direction matrices and are calculated. The initial transformation matrix is calculated by the following steps: translation alignment centroid ; rotation alignment principal direction ; scale alignment , where σ is the standard deviation of the principal component.

[0064] The fine registration stage uses the ICP algorithm for accurate registration, and the objective function is , where and are corresponding point pairs, is a weight term, and N is the number of point pairs participating in registration. The convergence condition is set as , where is the convergence threshold, and the maximum number of iterations . R represents the rotation matrix, which describes the rotation transformation of the tooth in the three-dimensional space; t represents the translation vector, which describes the translation transformation of the tooth in the three-dimensional space.

[0065] After registration is completed, thin plate spline deformation is used in the overlapping area to achieve smooth transition, and the deformation function is , where is a radial basis function, P(x) is a polynomial term, is a control parameter. x represents the coordinates of the overlapping position data points.

[0066] The fusion weight function of the crown and the tooth root connection area is where d is the shortest Euclidean distance from each vertex on the predicted root surface to the nearest cortical surface of the jawbone, g = 5.0 is the steepness parameter, is the cutoff distance.

[0067] The mesh optimization stage generates a smooth surface using a Poisson reconstruction algorithm that solves the Poisson equation where V is the normalized gradient field. The Laplacian smoothing iteration formula is where is the set of neighboring vertices of vertex i, and λ = 0.3-0.7 is the smoothing factor.

[0068] Specifically, S4 specifically comprises: setting a tooth movement amount attenuation function based on the movement characteristics of teeth, an anatomical stability function based on the stability requirement, and an individualized adjustment factor based on the individualized difference requirement, optimizing the weight term of the objective function of the ICP algorithm based on the tooth movement amount attenuation function, the anatomical stability function, and the individualized adjustment factor, adding a regularization term to the objective function of the ICP algorithm, obtaining the optimized ICP algorithm, registering the baseline model and the follow-up IOS data, and calculating the spatial transformation parameter based on the optimized ICP algorithm, and then applying the spatial transformation parameter to the baseline model to predict the current root position.

[0069] Specifically, the tooth movement amount attenuation function is an exponential function that attenuates based on the increase of the movement distance of the tooth root position, the anatomical stability function is graded and assigned based on the different tooth positions of the teeth, and the individualized adjustment factor is assigned different values based on individual differences, with a value range of 0.8-1.2.

[0070] In the follow-up stage, after obtaining the intraoral scanning data in the treatment process, a dynamic weight optimization ICP registration mechanism guided by clinical targets is used to predict the current root position.

[0071] The core innovation of this algorithm is to assign different weights to the registration process according to the movement characteristics and stability requirements of different teeth in orthodontic treatment, significantly improving the prediction accuracy.

[0072] In the dynamic weight calculation, the anatomical stability function S(τ) is determined according to the biomechanical characteristics of the tooth position τ: for the anterior tooth anchorage area a high weight of 0.9-0.98 is given, for the premolar transition area a medium weight of 0.4-0.6 is given, and for the molar movement area a low weight of 0.1-0.3 is given.

[0073] The movement amount attenuation function M(d) is calculated based on the expected movement distance d (unit: mm): , where α=2.5 is the attenuation coefficient, to ensure a stronger weighted inhibition effect on teeth that have moved significantly.

[0074] Overall weight ,in The individualized adjustment factor is (0.8-1.2). The optimization objective function is: ,in The regularization term is β = 0.01, which is the regularization coefficient. By obtaining the optimized spatial transformation parameters R and t, the current tooth root position can be predicted.

[0075] Solve using the Levenberg-Marquardt algorithm, and update the formula as follows: ,in Let μ be the Jacobian matrix, and μ be the damping factor (initial value). ), For parameter increments, Let be the residual vector. The algorithm converges if it satisfies any of the following conditions: the objective function changes relatively... Parameter changes The maximum number of iterations was reached: 100.

[0076] Specifically, it also includes: S5, segmenting the CBCT data and generating CBCT jawbone mesh data, obtaining a jawbone model based on the CBCT jawbone mesh data, calculating the shortest distance from each point on the currently predicted tooth root surface to the jawbone cortex surface based on the predicted current tooth root position and the jawbone model, setting a risk distance threshold, and performing a risk level assessment.

[0077] Specifically, S2 includes: using the nnUNet framework to segment the tooth region and jawbone cortex region in the CBCT data, then using the Marching Cubes algorithm to transform the mesh data, optimizing the data through the edge folding algorithm and performing Laplacian smoothing to generate the CBCT tooth root mesh data and CBCT jawbone mesh data;

[0078] In step S2, the IOS data is segmented to generate IOS crown mesh data. Specifically, this includes using the TSegNet algorithm to segment the IOS data and generate the IOS crown mesh data.

[0079] To accurately extract the three-dimensional geometric information of teeth and jawbones from multimodal data, two specially optimized deep learning segmentation algorithms were used for segmentation.

[0080] For IOS crown segmentation, the TSegNet algorithm was employed. This is a highly efficient and accurate tooth segmentation network specifically designed for 3D dental models. The TSegNet algorithm can robustly handle challenging clinical cases, such as pre-orthodontic abnormalities like missing teeth, crowding, or misalignment.

[0081] The algorithm employs a two-stage network architecture: the first stage detects the center points of all teeth through a distance-aware tooth center point voting mechanism. The problem of tooth detection is transformed into the problem of tooth center point prediction, and the voting function is: Where G is the Gaussian kernel function, The standard deviation is given. This method is based on observations that regardless of tooth shape, position, and orientation, the center point is a stable feature point within the tooth shape, making it more reliable than bounding boxes. The second stage employs a cascaded segmentation module based on confidence levels, calculated using the following formula: ,in The distance to the center point. To adjust parameters, this stage first uses the predicted tooth center point cropping point cloud and corresponding features as tooth proposals. Then, all tooth proposals are sent to the segmentation module to generate individual tooth labels. A point-level confidence map based on an attention mechanism is introduced to enhance label learning, particularly improving segmentation accuracy for teeth with blurred boundary signals. TSegNet demonstrates good robustness in handling complex clinical scenarios such as missing teeth, additional braces, and crowded teeth, achieving a segmentation accuracy of 98.0±1.2% Dice coefficient and an F1 score of 94.2%.

[0082] like Figure 3 As shown, the nnUNet CBCT segmentation algorithm is based on a 3D U-Net architecture and uses residual connections and attention mechanisms to process tooth and jawbone segmentation from CBCT data. After ROI detection, the CBCT data generates segments with dimensions of [missing information]. The single-channel image is downsampled and divided into blocks. The image blocks are converted into embedded vector memories through linear projection, and semantic segmentation is performed to initially extract the target structure and generate attention intent and binary mesh.

[0083] The image is processed through two branches. Branch ① involves two convolutional layers with 3×3×3 kernels and ReLU activation function, followed by normalization, resulting in an image with dimensions of [missing information]. Features Figure 1 .

[0084] Branch ② is the Transformer encoder module, comprising a multi-head self-attention section and a multilayer perceptron section. The input to Branch ②, after processing by the multi-head self-attention section, is processed by a first convolutional module consisting of three stacked 2×2×2 deconvolutional layers and 3×3×3 standard convolutional layers; a second convolutional module consisting of two stacked 2×2×2 deconvolutional layers and 3×3×3 standard convolutional layers; and finally, the multilayer perceptron module. The output of the multilayer perceptron module is processed by a third convolutional module consisting of a 2×2×2 deconvolutional layer and 3×3×3 standard convolutional layers, and a single 2×2×2 deconvolutional layer. A fully connected layer then connects the outputs of the third convolutional module and the deconvolutional layer, and finally, the output is processed by two stacked 3×3×3 standard convolutional layers to obtain the final output. Features Figure 2 This feature Figure 2 After further processing by a 2×2×2 deconvolution layer and concatenation with the output of the second convolutional module, we obtain... Features Figure 3 .feature Figure 3 After processing by a 2×2×2 deconvolutional layer and concatenating with the output of the first convolutional module, the output dimension is... Features Figure 4 .feature Figure 4 Processed by two convolutional layers with 3×3×3 kernels and one deconvolutional layer with 2×2×2 kernels, and then combined with a dimension of... Features Figure 1 The fusion yields a dimension of Features Figure 5 The data is then processed through two deconvolutional layers with 2×2×2 kernels and one convolutional layer with 3×3×3 kernels to obtain the segmentation results of the jawbone cortex and teeth. The first to third convolutional modules also include ReLU activation functions and normalization layers. Upsampling is performed using 2×2×2 transposed convolutions (deconvolutions) to double the resolution of the feature maps, followed by a standard 3×3×3 convolutional layer to refine and optimize the upsampled features. , , Representing features respectively Figure 1 Height, width and depth , , , These represent different feature channel numbers.

[0085] The network loss function is ,in for loss, The cross-entropy loss is given by λ = 0.5, which is the weight balancing factor. The Dice loss calculation formula is as follows: Where P is the predicted result and G is the true label. Training parameter settings: learning rate Batch size The iteration count was N=1000. The Dice coefficient for tooth segmentation reached 96.2±2.1%, and the segmentation accuracy for the jawbone cortex reached 95.8±1.8%. After segmentation, the Marching Cubes algorithm was used to convert the segmented voxel data into 3D mesh data, with an isosurface threshold T=0.5. The generated mesh was optimized using an edge-folding algorithm (target number of facets). ) and Laplace smoothing (iteration number K=5-10, smoothing factor ).

[0086] in, Figure 2 The feature extraction network used in the fourth coarse registration stage is the encoder part of the nnUNet segmentation network. Although nnUNet is described as a segmentation tool, its encoder automatically extracts the geometric features of the jawbone while performing the segmentation task, including anatomical landmarks, spatial location information and shape features. The output of the intermediate layer of the encoder can be used to identify the geometric center and main orientation of the jawbone.

[0087] The encoder processing results are then passed to stage ① for center point registration. Because patient head positions vary during CBCT scans, the orientation of the jawbone in three-dimensional space is inconsistent. Therefore, all jawbone data needs to be unified into a standard anatomical coordinate system, a prerequisite for subsequent CBCT-IOS multimodal fusion. Pre-registration reduces the search space of the nnUNet segmentation algorithm, and standardized input improves the network's accuracy in recognizing jawbone boundaries, especially for the complex-shaped maxillary sinus and mandibular angle regions. Furthermore, this step maintains consistency with the overall system workflow, using principal component analysis to unify the orientation of IOS data. Jawbone center point registration achieves the same standardization of CBCT data, ensuring that both modalities are processed in the same coordinate system.

[0088] The ultimate goal of this invention is to transform predicted tooth root locations into clinically applicable risk assessment information. The system calculates the shortest distance from each point on the predicted tooth root surface to the jawbone cortex surface and classifies the risk level based on clinical thresholds. In the distance calculation algorithm, for each vertex on the tooth root surface... Calculate the shortest distance to the surface of the jawbone cortex. ,in For each vertex of the root surface In the triangular facet The projection point on the surface.

[0089] Specifically, S5 calculates the shortest distance from each point on the currently predicted tooth root surface to the surface of the jawbone cortex, including:

[0090] An octree is constructed to segment the cortical surface of the jawbone model into multiple triangular patches;

[0091] Calculate the planar distance between each point on the currently predicted root surface and the plurality of triangular facets;

[0092] Based on planar distance, the projection point of each point on the currently predicted root surface on the plane corresponding to the plurality of triangular facets is determined;

[0093] The projection point is determined based on the centroid coordinate method to determine whether it is on the corresponding triangular facet. If it is, it is a valid projection point.

[0094] The distance between each point on the currently predicted root surface and the effective projection point of the plurality of triangular facets is calculated and compared to obtain the shortest distance from each point on the currently predicted root surface to the surface of the jawbone cortex.

[0095] The calculation of projection points involves three steps:

[0096] Calculate the planar distance from each point on the currently predicted root surface to the triangular facet. ,in Let n be the vertex of the triangle on the triangular facet, and n be the normal vector.

[0097] Calculate projection points Use the centroid coordinate method to determine whether the projection point is inside the triangular facet.

[0098] To improve computational efficiency, an octree is constructed to segment the jawbone surface, with each leaf node containing no more than 10 triangular faces. Simultaneously, a KD-Tree spatial index structure is employed to accelerate nearest neighbor search, reducing the search complexity from... Down to .

[0099] Specifically, S5 also includes preprocessing the jawbone model to remove irrelevant risk areas, which include the alveolar crest region and the maxillary sinus floor boundary region in the jawbone model. The alveolar crest region in the jawbone model is identified by curvature analysis, and the maxillary sinus floor boundary region in the jawbone model is located by anatomical landmarks. Morphological operations and region growing algorithms are used to remove the irrelevant risk areas.

[0100] Considering that root-cortical bone distance analysis primarily focuses on the relationship between the tooth root and the buccal and lingual cortical bone, which is crucial for assessing biomechanical boundaries in orthodontic treatment, this invention first preprocesses the jawbone model to improve the clinical relevance of the analysis results. Since the alveolar ridge crest and maxillary sinus floor regions may result in clinically insignificant shortest distance values ​​in the distance calculation, affecting the accurate assessment of truly risky areas, a three-dimensional mesh processing algorithm is used to selectively remove the alveolar ridge crest and maxillary sinus floor regions, allowing distance measurements to focus more on the buccal and lingual cortical bone boundaries that clinicians are concerned with.

[0101] The specific preprocessing steps include: identifying the alveolar ridge crest region based on curvature analysis and setting a curvature threshold. The boundaries of the maxillary sinus floor were located using anatomical landmarks; morphological manipulations and region growing algorithms were used to accurately segment and remove these interfering regions.

[0102] Specifically, setting the risk distance threshold in S5 includes: assigning a graded value based on the different shortest distance values ​​from each point on the root surface to the surface of the jawbone, and adjusting the graded value results in a personalized threshold based on the patient's age, periodontal phenotype, and type of treatment to obtain the risk distance threshold.

[0103] Risk level classification is based on a four-level system using the distance value d: High-risk areas Assignment Medium risk Assignment Relatively safe Assignment ; safe zone Assignment .

[0104] The threshold is dynamically adjusted based on patient characteristics; the personalized threshold calculation formula is as follows: ,in The age-adjusted factor is 0.8-1.0 for adolescents and 1.0-1.2 for adults. The periodontal phenotypic factors (0.8-1.0 for thick gingiva, 1.0-1.3 for thin gingiva) are... Treatment type factor (extraction orthodontics 0.9-1.1, non-extraction orthodontics 1.0-1.2). and And so on.

[0105] like Figure 4 As shown, visualization rendering is performed using a real-time rendering engine based on OpenGL, which supports lighting models, shadow effects, and transparency adjustment. Color interpolation uses linear interpolation in the HSV color space to ensure visual continuity, and the frame rate is maintained at ≥30fps to ensure a smooth interactive experience.

[0106] The orthodontic follow-up monitoring system of this invention can achieve excellent high-precision prediction performance. The average surface shell difference is 0.0565 mm (95% confidence interval: [0.0501, 0.0630] mm), the average root-cortical bone distance prediction error is 0.1004 mm, and the intraclass correlation coefficient (ICC) reaches 0.985, indicating that there is excellent consistency between the predicted value and the true value, and the prediction accuracy reaches or even exceeds the accuracy of traditional CBCT measurement.

[0107] This invention significantly reduces radiation risks. Zero radiation exposure is achieved during follow-up visits, completely resolving the cumulative ionizing radiation problem caused by frequent CBCT scans in traditional methods. It is particularly suitable for adolescent patients requiring long-term monitoring, and also addresses radiation safety issues for women of childbearing age during orthodontic treatment, providing technical assurance for safer orthodontic treatment.

[0108] This invention significantly improves clinical work efficiency. The automated analysis process significantly reduces manual intervention, generates real-time visual reports to optimize the diagnosis and treatment process, and supports high-frequency on-demand monitoring. Compared to the large amount of time and effort required for traditional manual analysis of CBCT data, the automated prediction process based on IOS data creates the conditions for achieving higher-frequency on-demand monitoring.

[0109] This invention enhances risk warning capabilities. The 0.17mm warning threshold provides a 1.7 times safety margin relative to the system's prediction error, enabling reliable early warnings before bone fenestration or fracture occurs. The four-level risk classification (red-yellow-green-blue) and 360° rotation viewing function achieve comprehensive risk assessment. Clinicians can adjust treatment strategies promptly based on the warning chart, effectively preventing complications such as alveolar bone fenestration and fracture.

[0110] According to one aspect of the present invention, a deep learning-based orthodontic treatment pedicle relationship follow-up monitoring system is proposed, comprising the method described in any one of the first aspects, such as... Figure 5 As shown, it includes the following modules:

[0111] Data processing module 51 is configured to acquire patient CBCT data and IOS data and preprocess the CBCT data and IOS data;

[0112] The data segmentation module 52 is configured to segment the CBCT data to generate CBCT root mesh data and segment the IOS data to generate IOS crown mesh data.

[0113] The model building module 53 is configured to register the CBCT root mesh data and IOS crown mesh data based on the ICP algorithm to obtain a baseline model.

[0114] The tooth root prediction module 54 is configured to optimize the ICP algorithm based on the tooth's movement characteristics, stability requirements, and individual differences, obtain an optimized ICP algorithm, acquire patient follow-up IOS data, and predict the current tooth root position based on the baseline model, follow-up IOS data, and the optimized ICP algorithm.

[0115] The key improvement of this invention lies in the realization of radiation-free follow-up monitoring technology. Traditional methods require frequent CBCT scans, resulting in a high risk of cumulative radiation. This invention uses a single initial CBCT scan combined with subsequent IOS data to predict tooth root position, completely avoiding ionizing radiation exposure during follow-up visits. It is particularly suitable for adolescent patients and women of childbearing age, solving the problem of cumulative radiation safety in orthodontic treatment.

[0116] This invention innovatively proposes a clinical goal-oriented dynamic weight registration algorithm. Traditional ICP algorithms use homogeneous weights for all teeth, leading to accumulated registration biases in the moving tooth regions. This invention, based on a differentiated weight allocation strategy considering the expected tooth movement and stability requirements, significantly improves prediction accuracy, resulting in a prediction difference of only 0.0041 mm across different tooth positions. This solves the registration bias problem caused by the spatiotemporal heterogeneity of tooth movement in orthodontic treatment.

[0117] This invention achieves accurate prediction of key clinical risk indicators. Existing methods mainly focus on geometric accuracy and lack quantitative assessment of clinical risk indicators. This invention specifically optimizes the root-cortical bone distance prediction algorithm to achieve sub-millimeter accuracy, with a root-bone distance prediction error of only 0.1004 mm, reaching the accuracy of CBCT measurement, and can provide reliable quantitative risk indicators for clinical decision-making.

[0118] This invention constructs an intelligent visual early warning system. Traditional analysis results are abstract and difficult to understand intuitively regarding risk levels. This invention uses a three-dimensional heat map to display the root-bone distance in real time, supports personalized threshold adjustment and 360° rotation viewing, transforming complex medical concepts into an intuitive interface, significantly improving clinical decision-making efficiency and patient treatment compliance.

[0119] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A deep learning-based orthodontic treatment root bone relationship follow-up monitoring method, characterized in that, The method comprises the following steps: S1, collecting CBCT data and IOS data of a patient and preprocessing the CBCT data and the IOS data; S2, performing segmentation processing on the CBCT data to generate CBCT tooth root grid data, and performing segmentation processing on the IOS data to generate IOS tooth crown grid data; S3, registering the CBCT tooth root grid data and the IOS tooth crown grid data based on an ICP algorithm to obtain a baseline model; S4, setting a tooth movement amount attenuation function based on movement characteristics of teeth , an anatomical stability function based on stability requirements , and an individualized adjustment factor based on individualized difference requirements , wherein the tooth movement amount attenuation function is an exponential function that attenuates based on an increase in movement distance of a tooth root position , wherein represents an attenuation coefficient, represents an expected movement distance, the anatomical stability function is graded based on different tooth positions of the teeth, the individualized adjustment factor is assigned different values based on individual differences, and a comprehensive weight is obtained by optimizing a weight term of an objective function of the ICP algorithm based on the tooth movement amount attenuation function, the anatomical stability function, and the individualized adjustment factor , a regularization term is added to the objective function of the ICP algorithm to obtain an optimized ICP algorithm, IOS data of a patient at a follow-up visit is obtained, and current tooth root positions are predicted based on the baseline model, the IOS data at the follow-up visit, and the optimized ICP algorithm.

2. The deep learning-based orthodontic treatment root bone relationship follow-up monitoring method of claim 1, wherein, The preprocessing of the CBCT data in S1 specifically comprises: performing threshold segmentation and morphological opening operation on the CBCT data to remove metal artifacts, then performing isotropic resampling to a uniform voxel spacing and linear transformation normalization mapping, and finally using a threshold method to extract the cortical region of the jaw bone in the CBCT data as a region of interest; The preprocessing of the IOS data in S1 specifically comprises: converting the IOS data into a voxel grid form, wherein the voxel density value of the IOS data is calculated by nearest neighbor interpolation, and the local contrast is enhanced based on an adaptive histogram equalization algorithm, then a principal component matrix of the IOS data is calculated and the orientation consistency of the IOS data in a three-dimensional space is realized by limiting the principal component matrix of the IOS data through a rotation matrix.

3. The deep learning-based orthodontic treatment root bone relationship follow-up monitoring method of claim 1, wherein, S3 specifically comprises: The mass center and the principal direction matrix of the CBCT tooth root grid data and the IOS tooth crown grid data are calculated based on a principal component analysis algorithm, the corresponding initial transformation matrix is calculated based on the mass center and the principal direction matrix of the CBCT tooth root grid data and the IOS tooth crown grid data, coarse registration alignment is performed based on the initial transformation matrix, then accurate registration is performed based on the ICP algorithm, the overlapping region of the CBCT tooth root grid data and the IOS tooth crown grid data is smoothly transitioned using a thin-plate spline deformation algorithm, then the fusion weight function is set to fuse the overlapping region, and finally the baseline model is obtained by using the Persephone reconstruction algorithm for optimization.

4. The deep learning-based orthodontic treatment root bone relationship follow-up monitoring method of claim 1, wherein, S4 specifically comprises: registering the baseline model and the follow-up IOS data, and calculating the spatial transformation parameters based on the optimized ICP algorithm, and then applying the spatial transformation parameters to the baseline model to predict the current tooth root position.

5. The deep learning-based orthodontic treatment root bone relationship recheck monitoring method of claim 4, wherein, The assignment range of the individual adjustment factor is 0.8-1.

2.

6. The deep learning-based orthodontic treatment root bone relationship follow-up monitoring method according to claim 1, characterized in that, Further comprising: S5, performing segmentation processing on the CBCT data to also generate CBCT jaw bone grid data, obtaining a jaw bone model based on the CBCT jaw bone grid data, calculating the shortest distance from each point on the current predicted tooth root surface to the cortical surface of the jaw bone based on the predicted current tooth root position and the jaw bone model, setting a risk distance threshold, and performing risk level evaluation.

7. The deep learning-based orthodontic treatment root bone relationship recheck monitoring method of claim 6, wherein, S2 specifically comprises: using the nnUNet framework to segment the tooth region and the cortical region of the jaw bone in the CBCT data, then using the Marching Cubes algorithm to perform grid data conversion, performing data optimization through edge folding algorithm and performing Laplacian smoothing processing to generate the CBCT tooth root grid data and the CBCT jaw bone grid data; In S2, the IOS data is segmented to generate IOS crown mesh data, specifically including: using TSegNet algorithm to segment the IOS data to generate the IOS crown mesh data.

8. The deep learning-based orthodontic treatment root bone relationship recheck monitoring method of claim 6, wherein, In S5, the shortest distance from each point on the current predicted root surface to the cortical surface of the jawbone is calculated, specifically including: The cortical surface of the jawbone model is segmented into a plurality of triangular facets by constructing an octree; The plane distance between each point on the current predicted root surface and the plurality of triangular facets is calculated; Based on the plane distance, the projection point of each point on the current predicted root surface on the plane corresponding to the plurality of triangular facets is determined; Based on the barycentric coordinate method, it is determined whether the projection point is on the corresponding triangular facet, and if so, it is an effective projection point; The distance between each point on the current predicted root surface and the effective projection point of the plurality of triangular facets is calculated and compared to obtain the shortest distance from each point on the current predicted root surface to the cortical surface of the jawbone.

9. The deep learning-based orthodontic treatment root bone relationship recheck monitoring method of claim 6, wherein, S5 also includes preprocessing the jawbone model to remove irrelevant risk areas, including the alveolar crest region and the maxillary sinus bottom boundary region in the jawbone model, identifying the alveolar crest region in the jawbone model through curvature analysis, and positioning the maxillary sinus bottom boundary region in the jawbone model through anatomical landmark points, using morphological operations and region growing algorithm to remove the irrelevant risk areas.

10. The deep learning-based orthodontic treatment root bone relationship recheck monitoring method of claim 6, wherein, In S5, the risk distance threshold is set, specifically including: based on the different values of the shortest distance from each point on the root surface to the cortical surface of the jawbone, the results of the hierarchical assignment are adjusted based on the patient's age, periodontal phenotype, and the type of treatment performed, to obtain the risk distance threshold.

11. A deep learning-based orthodontic treatment root bone relationship follow-up monitoring system, characterized in that, The method according to any one of claims 1 to 10, comprising the following modules: A data processing module configured to collect patient CBCT data and IOS data and preprocess the CBCT data and IOS data; A data segmentation module configured to segment the CBCT data to generate CBCT root mesh data, and segment the IOS data to generate IOS crown mesh data; A model construction module configured to register the CBCT root mesh data and the IOS crown mesh data based on ICP algorithm to obtain a baseline model; a tooth movement amount attenuation function based on movement characteristics of the tooth an anatomical stability function based on stability requirements and an individualized adjustment factor based on individualized difference requirements wherein the tooth movement amount attenuation function is an exponential function that attenuates based on an increase in a movement distance of a tooth position wherein denotes an attenuation coefficient, denotes an expected movement distance, the anatomical stability function is assigned different values based on different tooth positions of the tooth, the individualized adjustment factor is assigned different values based on individual differences, and a comprehensive weight is obtained by optimizing a weight term of an objective function of the ICP algorithm based on the tooth movement amount attenuation function, the anatomical stability function, and the individualized adjustment factor a regularization term is added to the objective function of the ICP algorithm to obtain an optimized ICP algorithm, IOS data of a patient at a follow-up visit is obtained, and a current tooth position is predicted based on the baseline model, the IOS data at the follow-up visit, and the optimized ICP algorithm.

Images