Deep learning-based root canal therapy guide plate design method and device

Abstract

The invention discloses a method for designing a root canal therapy guide plate based on deep learning. The method comprises the following steps: acquiring first data with consistent size through preprocessing such as normalization based on CBCT (Cone Beam Computed Tomography) data; based on the first data and treatment data corresponding to the CBCT data, obtaining a training data set through mapping; network parameters are updated through an nnUNet module based on the training data set; inputting CBCT data of a patient, and obtaining oral cavity marking data through an nnUNet module, wherein the marking data are data for identifying and marking dentitions, single teeth, medullary cavities and root canals of the CBCT data of the patient; updating the oral marking data through manual auditing and revising on the basis of the oral marking data, and updating the network parameters through the nnUNet module on the basis of the oral marking data; acquiring assembly data based on the oral marker data, wherein the assembly data is an STL guide plate file and comprises a metal inner ring and a plastic shell; and generating a guide plate of a 3D structure through 3D printing based on the assembly body data.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical auxiliary device design, and particularly relates to a root canal treatment guide plate design method and device based on deep learning. BACKGROUND

[0002] Root canal treatment is a key means of oral medicine, which can remove infected dental pulp and eliminate root canal inflammation, thereby preserving natural teeth, avoiding tooth extraction, restoring molar function, relieving pain, maintaining oral health and function, and is a core way to treat dental pulp disease and periapical disease. However, there are many problems in the preoperative pulp opening path design of root canal treatment, which seriously affect the precision, efficiency and standardization of treatment: Dependence on doctor's experience: the pulp opening scheme design has not established a standardized system taking the individualized root canal anatomic morphology of patients as the absolute basis, resulting in a lack of objective and unified evaluation standard for the treatment scheme. Especially in the clinical treatment of multi-rooted teeth such as molars, since the spatial anatomic position, direction and entrance of multiple root canals need to be comprehensively analyzed and judged, the operation process is complex, which further magnifies the limitations of the experience-dominated mode in precision and consistency.

[0003] Software tool defects: the existing software needs to be manually analyzed, which is low in efficiency and large in error; only single-rooted canal independent paths can be processed, and the overall modeling and intelligent reasoning ability of multi-rooted canal collaborative structure is lacking, which cannot meet the complex treatment needs and is difficult to realize the automation and precision of multi-rooted canal path planning. Preoperative and postoperative fragmentation: it is difficult to output the preoperative design results, which cannot directly connect with the guide plate generation or enter the navigation system, and it is difficult to standardize the output as a guide plate 3D model or navigation system recognizable data, resulting in a disconnection between preoperative and intraoperative, data conversion relying on manual work, and affecting the treatment precision and efficiency. SUMMARY

[0004] To solve the above problems, the application provides a root canal treatment guide plate design method and device based on deep learning. Through a deep learning driven automatic segmentation mechanism, the target tooth's multi-root canal and pulp cavity anatomical structure can be accurately extracted from the ROI area CBCT image, which gets rid of the inefficiency and error of traditional manual segmentation, provides accurate and personalized anatomical basis for multi-root canal collaborative path planning, and solves the limitations of relying on experience to judge anatomical morphology. Based on the segmentation result, a 3D printed guide plate model is output through multi-root canal overall path planning, which ensures the fit of the guide plate and the tooth structure, and can guide the pulp opening and root canal entrance positioning. Through the split type easy-to-assemble feature, the clinical operation convenience is improved, the preoperative planning and intraoperative operation are seamlessly connected, and the accuracy, efficiency and standardization level of root canal treatment are greatly improved. This method provides a completely adaptive and one-key production root canal treatment guide plate design scheme for any root canal morphology by means of voxel domain geometry construction, parameterized sleeve synthesis body modeling and automatic process closely coupled with deep learning segmentation, which not only takes into account the clinical feasibility, but also realizes a number of original technical breakthroughs at the algorithm level.

[0005] The first aspect of the application provides a root canal treatment guide plate design method based on deep learning, comprising: obtaining first data with consistent size based on CBCT data through preprocessing; obtaining training data set through mapping based on the first data and treatment data corresponding to the CBCT data; updating network parameters based on the training data set through an nnUNet module; obtaining oral marking data by inputting patient CBCT data through an nnUNet module, wherein the marking data is data for identifying and marking the dentition, single tooth, pulp cavity and root canal of the patient CBCT data; updating the oral marking data based on artificial review and revision, and updating the network parameters based on the oral marking data through the nnUNet module; obtaining assembly data based on the oral marking data, wherein the assembly data is an STL guide plate file and includes a metal inner ring and a plastic shell; generating a 3D structure guide plate based on the assembly data through 3D printing, wherein the 3D structure guide plate is used for guiding pulp opening and root canal entrance positioning.

[0006] Preferably, the step of obtaining first data with consistent size based on CBCT data through preprocessing further comprises: obtaining short side scaling factor and long side scaling factor based on the CBCT data through gray scale normalization and scaling ratio in sequence, wherein the range of gray scale normalization is [0, 1]; scaling the first pre-processing data based on the CBCT data, the short side scaling factor and the long side scaling factor by bilinear interpolation; obtaining first data based on the first pre-processing data by random cropping and random flipping in sequence.

[0007] Preferably, the step of obtaining training data set based on the first data and the treatment data corresponding to the CBCT data by mapping specifically comprises: constructing the training data set based on the first data and the treatment data by density peak clustering, the training data set comprising tooth labels and corresponding index data of each panoramic data in the CBCT data; the tooth labels at least comprising three-dimensional structure data of each tooth, four-level label data, ROI data corresponding to the four-level label data, the four-level label data comprising labels of dentition, single tooth, pulp cavity and root canal.

[0008] Preferably, the step of updating network parameters based on the training data set by nnUNet module further comprises: obtaining pseudo-label data based on the training data set by 3DU-Net encoder, the pseudo-label data comprising predicted labels of dentition, single tooth, pulp cavity and root canal; constructing a total loss function based on the training data and the pseudo-label data and updating network parameters, the calculation expression being: L total =L seg +L topo +L conf ; wherein L seg is a supervised segmentation loss, L topo is a topological consistency loss, and L conf is a confidence weighted loss.

[0009] Preferably, the calculation expression of the supervised segmentation loss is: wherein, is a voxel of pseudo-label data, λ dice and λ ce are preset parameters, and c in Pc is probability data of dentition, single tooth, pulp cavity and root canal respectively; the calculation expression of the topological consistency loss is: wherein, is a hyperparameter, is a penalty amount of root canal data of label data; the calculation expression of the confidence weighted loss is: wherein, is a hyperparameter, .

[0010] Preferably, the step of obtaining assembly data based on the oral marker data further comprises: checking and revising the 3 landmarks of the oral marker data by artificial, the 3 landmarks are located on the crown side of each tooth to the near root side in turn, including the crown end starting point of the most convex point of the root canal orifice, the middle segment inflection point of the vertex of the first significant bend of the root canal, and the root apex stop point 0.5-1 mm behind the root foramen; obtaining path data based on the landmarks, the path data is a smooth path fitted from the crown end starting point, the middle segment inflection point, and the root apex stop point; judging the number of path data of each tooth, if the number is less than or equal to 1, prompting artificial checking and revising, otherwise, checking through and updating the oral marker data based on the revised landmarks and the path data.

[0011] Preferably, the step of obtaining assembly data based on the oral marker data further comprises: obtaining shell data adapted to the marker data; updating the shell data by removing islands based on the shell data, the specific process is: marking all connected domains of the shell data, calculating the voxel volume of each connected domain, and retaining the connected domain with the largest voxel volume, and removing the remaining connected domains and updating the shell data; obtaining local base plate data corresponding to the specified region by selecting the specified region of the shell data and through coordinate mapping and voxel-level cropping, updating the local base plate data based on the local base plate data by correcting the root canal curve, identifying the end point of the calibration curve, and updating the local base plate data based on the root canal vector; obtaining a plurality of metal sleeves based on the end point of the calibration curve and the root canal vector, and obtaining assembly data by fusion based on the metal sleeves and the local base plate data; obtaining assembly data by morphological smoothing processing and interference checking based on the assembly data, and exporting; updating the assembly data by parameter reconstruction based on the assembly data.

[0012] Preferably, the step of obtaining shell data adapted to the marker data further comprises: obtaining tooth mask data based on the marker data, the tooth mask data includes maxilla, mandible, upper teeth, lower teeth, and mandibular canal; Binary data is obtained through binarization processing based on the dental mask data, and the binary data at least includes data type, geometric feature, and voxel spacing. Based on the binary data, a hollow shell is generated through morphological expansion, and shell data is obtained through a script, wherein the shell data includes data with a preset thickness and a closed structure.

[0013] Preferably, the step of obtaining a plurality of metal sleeves based on the end point of the calibration curve and the root canal vector further comprises: A plurality of metal sleeves are obtained through coordinate conversion, main cylindrical voxels, auxiliary cylindrical voxels, and array merging based on the end point of the calibration curve and the root canal vector, wherein the metal sleeves include voxel data. The metal sleeve is updated through drilling processing, thickening sleeve wall processing, and hole filling processing based on the metal sleeve data.

[0014] The second aspect of the present application provides a root canal treatment guide plate design device based on deep learning, comprising: A preprocessing module obtains first data with consistent sizes based on CBCT data through preprocessing; A data generation module is configured to obtain a training data set through mapping based on the first data and corresponding treatment data of the CBCT data; A training module is configured to update network parameters through an nnUNet module based on the training data set; A scheme generation module is configured to obtain oral marking data through an nnUNet module by inputting patient CBCT data, wherein the marking data is data for identifying and marking the dentition, single tooth, pulp cavity, and root canal of the patient CBCT data; A correction marking module is configured to update the oral marking data through artificial review and revision based on the oral marking data, and to update the network parameters through the nnUNet module based on the oral marking data; A guide plate data generation module is configured to obtain assembly data based on the oral marking data, wherein the assembly data is an STL guide plate file and includes a metal inner ring and a plastic shell; A printing module is configured to generate a 3D structure guide plate through 3D printing based on the assembly data, wherein the 3D structure guide plate is used to guide the pulp opening and root canal entrance positioning.

[0015] Compared with the prior art, the present application has the following advantages and positive effects: The present application file provides high-quality input for model training by gray scale normalization, scaling, random cropping and flipping processing of data, and reduces the interference of data difference on model learning. Further refining the training data set for the first data enables the model to more comprehensively learn the anatomical features of the dentition, single tooth, pulp cavity and root canal, and improves the relevance of feature learning. Updating the network parameters based on the training data set can ensure the segmentation accuracy while strengthening the learning of the root canal topology, reducing the prediction bias caused by complex anatomical structures, significantly improving the recognition and identification accuracy of the dental marker data, and enhancing the generalization ability of the model. After inputting the patient CBCT data, the dental marker data is obtained through the nnUNet module, and the data is further optimized through artificial review and revision, and the network parameters are updated, forming a closed loop of automatic recognition, artificial verification and model iteration, which can not only reduce the subjectivity and error of artificial analysis, but also improve the recognition ability of the model for individualized anatomical structure through continuous iteration, ensuring the accuracy of the dental marker data. On this basis, the three key identification points (coronal starting point, middle inflection point and apical ending point) are revised and a smooth path is fitted through artificial verification, and the number of multi-root canal paths is verified, which can effectively avoid unreasonable path planning and provide accurate path basis for subsequent guide plate design. In the process of obtaining the assembly data, the hollow shell is generated by the tooth mask data, and the isolated islands are removed to ensure the integrity of the shell structure; the local base guide plate data is obtained by cutting the specified area of the shell, combined with root canal curve correction and metal sleeve design, and then smoothed by morphological smoothing and interference checking to ensure the precise fitting of the guide plate and the tooth structure. The final STL guide plate file contains a metal inner ring and a plastic shell. The guide plate made by 3D printing not only has high-precision physical structure and can accurately guide the pulp opening and root canal entrance positioning, but also takes into account stability and easy assembly due to its split design, which can reduce the installation time during the operation and improve the operation efficiency, realizing seamless connection between preoperative planning and intraoperative operation. The overall technical scheme realizes the automatic process from CBCT data processing, anatomical structure recognition, path planning to guide plate design through deep learning, greatly reduces the dependence on personal experience of doctors and reduces the complexity and error rate of manual operation. At the same time, the high precision and adaptability of the guide plate can improve the accuracy of pulp opening and root canal treatment, reduce surgical trauma and shorten treatment time, promote the standardization, precision and efficiency of root canal treatment, and significantly improve the clinical treatment quality.

[0016] A guide plate of a 3D structure is generated based on the assembly data by 3D printing, and the guide plate of the 3D structure is used to guide the opening of the pulp and the positioning of the root canal entrance. After all the voxel operations are completed, the STL files are output respectively by the local bottom guide plate data of the assembly data and the several metal sleeves; since the two are in a unified local coordinate system and have geometric tolerances, they can be assembled after printing, and the external CAD secondary alignment process commonly used in the traditional digital workflow is omitted. Through the Slicer extension, this process is scripted from the loading of the image to the export of the printable model, and the doctor only needs to complete the ROI and root canal curve interaction once in the interface, which greatly reduces the cross-software operation cost.

[0017] Image domain Boolean and morphological fusion throughout the process. Unlike subtractive / additive Boolean on STL meshes, this method always completes Hollow, cropping and solid construction geometry operations in voxel space, naturally avoiding mesh gap and self-intersection problems, ensuring uniform thickness and no broken surface, and having a higher printing success rate. Parameter acquisition sleeve array for any root canal. By reading the direction of each root canal endpoint, a primary and auxiliary multi-level cylinder is automatically generated, and the height, diameter and other parameters are controlled by adjustable parameters; regardless of the number of root canals or the diversity of arrangement, the algorithm can complete the layout of the sleeve combination body at one time. Two-way interference constraint and real-time abnormal feedback. During the generation of the guide plate, the system performs voxel intersection detection with the tooth body for the metal and plastic guide plates respectively, and outputs and prompts immediately if a conflict occurs. Coupling with CBCT segmentation depth. Because the guide plate shape is completely derived from the nnU-Net dental arch segmentation result, the internal fit and external contour are highly consistent with the real tooth surface, and stable bonding can be obtained without manual grinding. Script is a module, plug and play. All geometric steps are connected through the 3DSlicer Python API, and the extended module is published on the clinical end, so that the doctor can complete the closed-loop operation from CBCT to STL without CAD background, which significantly improves the guide plate design efficiency and standardization. BRIEF DESCRIPTION OF DRAWINGS

[0018] The specific embodiments of the application will be further described in detail below with reference to the accompanying drawings, in which: Figure 1 The main flowchart of the root canal treatment guide plate design method based on deep learning of the application; Figure 2 The schematic diagram of the oral marker data of the root canal treatment guide plate design method based on deep learning of the application. DETAILED DESCRIPTION

[0019] The application will be described in further detail below with reference to the drawings and specific embodiments. The advantages and features of the present application will become apparent from the following description and claims. It should be noted that the drawings are in extremely simplified form and are not drawn to precise scale, and are merely intended to facilitate the understanding of embodiments of the present application.

[0020] It should be noted that all directional indications, such as upper, lower, left, right, front, back, etc., are only used to explain the relative position relationship, movement condition, etc. between components in a certain posture (as shown in the drawings), and if the certain posture changes, the directional indications will also change accordingly.

[0021] First embodiment Referring to Figure 1 and Figure 2 , the first aspect of the present application provides a deep learning-based root canal treatment guide plate design method, comprising: S100: obtaining first data with consistent size based on CBCT data through preprocessing; S200: obtaining training data set through mapping based on first data and treatment data corresponding to CBCT data; S300: updating network parameters based on the training data set through the nnUNet module; S400: inputting patient CBCT data to obtain oral marking data through the nnUNet module, the marking data being data for identifying the dentition, single tooth, pulp cavity and root canal of the patient CBCT data and marking; S500: updating the oral marking data through artificial review and revision based on the oral marking data, and updating the network parameters through the nnUNet module based on the oral marking data; S600: obtaining assembly data based on the oral marking data, the assembly data being an STL guide plate file and including a metal inner ring and a plastic shell; S700: generating a 3D structure guide plate through 3D printing based on the assembly data, the 3D structure guide plate being used for guiding the pulp opening and root canal entrance positioning.

[0022] S100: obtaining first data with consistent size based on CBCT data through preprocessing. The CBCT data is a multi-center database facing multi-source heterogeneous CBCT data, the CBCT data is generated from CBCT data from different manufacturers' equipment, the first data includes image data with different imaging parameters and different resolutions, and the CBCT data includes a variety of tooth type variations and image artifacts. Through this step, the differences caused by equipment and image parameters are eliminated, and the data is standardized. By correcting the spatial deformation, the size of each data of the first data is aligned.

[0023] S200: Obtain a training data set by mapping based on the first data, the treatment data corresponding to the CBCT data; obtain the training data set by spatial mapping mechanism based on the first data containing the CBCT data and the treatment data, realize multi-modal data fusion integration, and the training data set includes etiology, intervention and scheme containing a treatment guide plate.

[0024] S300: Update the network parameters based on the training data set through the nnUNet module; the system introduces a self-supervised learning strategy to pre-train the visual encoder of the nnUNet module, and the training process includes mechanisms such as occlusion recovery, projection consistency constraint and structure pseudo-label generation, and the above processing enhances the expression ability of the model to complex dentition structure.

[0025] S400: Obtain oral marker data by inputting patient CBCT data through the nnUNet module, and the marker data is data for identifying the dentition, single tooth, pulp cavity and root canal of the patient CBCT data and marking; automatic segmentation by the nnUNet greatly shortens the annotation work of multiple teeth, can quickly provide a three-dimensional visual report, shortens the preoperative preparation time, and the obtained oral marker data is used to connect a surgical navigation system to realize dynamic path planning guide plate generation. This scheme makes the oral diagnosis and treatment from experience-driven to model-driven and data-driven, while ensuring medical quality, significantly reduces the technical threshold and operation cost.

[0026] S500: Update the oral marker data through artificial review and revision based on the oral marker data, and update the network parameters based on the oral marker data through the nnUNet module; the nnUNet automatically adjusts the sensitivity of the neural network to key features through back propagation, so that the response to these areas is stronger when predicting next time. Adjust the Dice coefficient dynamically in each iteration, and realize model convergence and reach expert level through multiple iterations.

[0027] S600: Obtain assembly data based on the oral marker data, and the assembly data is an STL guide plate file and includes a metal inner ring and a plastic shell; parameterized modeling logic: automatically calculate the metal ring wall thickness and the plastic layer thickness according to the oral marker data, and balance between mechanical strength, lightweight balance and the fit degree of the treatment scheme of the diseased tooth root canal.

[0028] S700: generating a guide plate of a 3D structure based on the assembly data, the guide plate of the 3D structure being used to guide the opening of the pulp and the positioning of the root canal entrance. After the completion of all voxel operations, the local bottom guide plate data based on the assembly data and the plurality of metal sleeve data are fused by calling vtkSlicerSegmentationsModuleLogic to output an STL file; since the local bottom guide plate data and the plurality of metal sleeve data are in a unified local coordinate system and have geometric tolerances, they can be assembled after printing, which eliminates the external CAD secondary alignment process commonly used in traditional digital workflow. Through Slicer extension, this process is fully scripted from image loading to printable model export, and only one ROI and root canal curve interaction is required by the doctor on the interface, which greatly reduces the cross-software operation cost.

[0029] Preferably, the step of obtaining the first data with consistent sizes based on the CBCT data through preprocessing further comprises: Based on the CBCT data, the short side scaling factor and the long side scaling factor are obtained by gray scale normalization and scaling ratio in sequence, and the range of gray scale normalization is [0, 1]; Based on the CBCT data, the short side scaling factor and the long side scaling factor, the first preprocessing data is obtained by bilinear interpolation scaling; Based on the first preprocessing data, the first data is obtained by random cropping and random flipping in sequence.

[0030] The gray scale dynamic range difference under different devices and different scanning protocols is unified through gray scale normalization ([0, 1]), so that the gray scale distribution of the structures such as tissues in the oral cavity falls within a unified interval, avoiding model deviation caused by original gray scale deviation. Through bilinear interpolation and independent scaling of long and short sides, the inherent anisotropic distortion in CBCT data is solved. Further, by independently calculating the short side / long side scaling factor, accurate correction of non-uniform deformation is realized. Bilinear interpolation balances the weighted fusion of neighborhood information while maintaining computational efficiency. Through random cropping and random flipping, the model focuses on local features rather than absolute position. Single data can generate multiple effective sample data through multiple augmentations, which alleviates the problem of medical data scarcity.

[0031] Preferably, the step of obtaining the training data set by mapping based on the first data and the treatment data corresponding to the CBCT data comprises: Based on the first data and the treatment data, a training data set is constructed by density peak clustering, and the training data set includes tooth marks and corresponding index data of each panoramic data in the CBCT data; the tooth marks at least include three-dimensional structure data of each tooth, four-level label data, and ROI data corresponding to the four-level label data, and the four-level label data includes labels of dentition, single tooth, pulp cavity and root canal.

[0032] The label generation is based on the open source DentalSegmentator framework and combines the seed region growing and topological boundary restriction method to realize the labeling of the boundaries of the teeth and soft tissues. The process is a nnUNet model framework which can adapt to any sample type of data set and independently configure the hyperparameters of the segmentation whole process, thereby improving the deployability while maintaining high performance. The above process includes: performing dental arch segmentation on the CBCT image in the first data, and automatically identifying the centroid position and number of each tooth through the density peak clustering method. The dental arch label can be used to position the teeth in the oral cavity as a whole, ensuring the accuracy of the spatial topological relationship of the whole mouth teeth. The single tooth label solves the boundary ambiguity problem of the overlapping area of adjacent teeth; the pulp cavity label is used to describe the spatial features of the tooth internal pulp cavity, providing anatomical basis for guide plate design; the root canal label is used to identify the running path of the main and auxiliary root canals, providing real-time three-dimensional landmarks for microapical surgery navigation and providing basic data for constructing guide plates. Each four-level label automatically generates a corresponding ROI mask to establish the mapping relationship between the tooth and the tooth opening path. The four-level label provides a data basis for measuring tooth volume, pulp cavity volume, root canal curvature and other parameters. The two-dimensional oral scan and treatment data are registered through nonlinear mapping to solve the spatiotemporal consistency problem of multi-source data of the same patient.

[0033] Preferably, the step of updating the network parameters based on the training data set through the nnUNet module further comprises: obtaining pseudo-label data based on the training data set through a certain specific 3D U-Net encoder, the pseudo-label data including predicted labels of dental arch, single tooth, pulp cavity and root canal; constructing a total loss function based on the training data and the pseudo-label data and updating the network parameters, the calculation expression being: L total =L seg +L topo +L conf ; In the formula, L seg is a supervised segmentation loss, L topo is a topological consistency loss, and L conf is a confidence weighted loss.

[0034] Preferably, the calculation expression of the supervised segmentation loss is: In the formula, is a voxel of the pseudo-label data, λ dice and λ ce are preset parameters, and c in Pc is 1-4, which are probability data of dental arch, single tooth, pulp cavity and root canal, respectively. The calculation expression of the topological consistency loss is: wherein, is a hyper-parameter, is a penalty quantity of root canal data of label data; The calculation expression of the confidence weighted loss is: wherein, is a hyper-parameter, .

[0035] Optionally, the network parameters are further updated by the four-level label data and the pseudo-label data of the training data through the discriminator. The potential feature space distribution difference between the real label and the generated label is learned. The gradient direction provided by the discriminator guides the main network to escape from the local minimum value; the spectral normalization constraint is applied to the adversarial loss, which suppresses overfitting.

[0036] The pseudo-label data including the predicted labels of the dentition, single tooth, pulp cavity and root canal are obtained based on the training data set through the 3D U-Net encoder. This step is to obtain the predicted pseudo-label data of the label data in the training data set through the encoder. The specific steps are as follows: the three-dimensional ROI region of the target tooth of each data in the training data set is obtained, and the pulp cavity and root canal data are obtained based on each three-dimensional ROI region through the mapping relationship of the four-level label and are segmented by the system. Due to the large individual difference of the root canal structure form according to the patient and the tooth type, and the problem of insufficient contrast and imaging blur in the root tip region of the CBCT image of the target tooth, considering the effective solution of the above problems, the geometric boundary representation of the pulp cavity and the tooth is inferred through the region growing and topological restriction based on the predicted label generated by the DentalSegmentator, and the network parameter training update under the label constraint condition based on the treatment data is realized. The segmentation model adopts the multi-task segmentation network structure configured by the nnUNet framework, and the backbone network simultaneously outputs the mask of the pulp cavity and the root canal, the positioning regression result of the apical foramen and the root canal path vector field when processing the ROI body data. Lseg adopts dynamic weighting considering λdice+λce, which solves the problem of missed detection in the label data, Ltopo introduces a neighborhood relationship constraint term to maintain the consistency of the tooth arrangement order and the anatomical spatial topology; Lconf realizes adaptive sampling of samples, and higher gradient is allocated to the low confidence area to realize adaptive update of the weight.

[0037] Preferably, the step of obtaining the assembly data based on the oral marker data further comprises: The three identification points of the oral marker data are checked and revised by artificial, and the identification points are sequentially located on the crown side of each affected tooth extending to the near root side, including the crown end starting point of the most convex point of the root canal, the middle segment inflection point of the vertex of the first significant bending place, and the root tip stop point of the root tip hole retreating 0.5-1mm; The path data is used to fit a smooth path based on the identification points, the path data being a coronal starting point, a middle segment inflection point, and an apical ending point. The number of path data of each tooth is determined, and if the number is less than or equal to 1, manual checking and revision are prompted, otherwise, the checking is passed, and the oral marker data is updated based on the revised identification points and path data.

[0038] The assembly data is obtained based on the oral marker data, and the tooth position number corresponding to one or more dental bodies in the oral marker data can be input. The target dental body is located by a segmentation mask, and the ROI region around the target dental body is extracted, thereby providing a structural basis for subsequent root canal identification and guide plate generation. Manual review of the three anatomical landmark points of each dental body avoids fatal errors caused by AI misidentification. The number of paths is checked to ensure that each root canal is independently modeled.

[0039] Preferably, the step of obtaining assembly data based on oral marker data further comprises: Obtaining shell data compatible with the marker data; Updating the shell data by removing islands based on the shell data, and the specific process is as follows: all connected domains of the shell data are marked, the voxel volume of each connected domain is calculated, and the connected domain with the largest voxel volume in the connected domain is retained, and the remaining connected domains are removed and the shell data is updated; The specified region of the shell data is selected, and the local base guide plate data corresponding to the specified region is obtained through coordinate mapping and voxel-level cropping. The local base guide plate data is updated based on the local base guide plate data by correcting the root canal curve, identifying the end point of the calibration curve, and updating the root canal vector; A plurality of metal sleeves are obtained based on the end point of the calibration curve and the root canal vector, and the assembly data is obtained by fusion based on the metal sleeve and the local base guide plate data; The assembly data is obtained by morphological smoothing processing and interference checking based on the assembly data, and is exported; The assembly data is updated by parameter reconstruction based on the assembly data.

[0040] Optionally, it further comprises detecting the availability of the assembly data based on two-level interference, and the specific logic is as follows: based on whether the voxels of the metal sleeve of the assembly data overlap with the voxels of the dental body, if they overlap, an exception is thrown and the assembly data is prompted to be regenerated, otherwise, it is further determined whether the base guide plate of the assembly data invades the dental body, and if so, the reconstruction parameter is prompted and the assembly data is regenerated.

[0041] First, based on the oral marker data, different types of teeth in the upper and lower jaws are identified and corresponding voxel-level labels are generated. This replaces the traditional manual rough cutting and painting link, laying the foundation for subsequent geometric operations. After segmentation, the generation of assembly data is entered. First, the shell data compatible with the oral marker data is generated. Based on the index number corresponding to one or more teeth, local base plate data is generated. The local base plate data is used for the plastic base plate of root canal treatment navigation, which realizes the coverage of the local base plate data in the operation area specified by the doctor. Specifically, the doctor extracts the partial area around the target tooth by boxing the ROI in Slicer. The RAS mapping to the IJK matrix is used to realize the world coordinate to index coordinate transformation. After the geometric boundary is mapped to the voxel space, the local base plate data is accurately cut through logical "and" operation. Based on the local base plate data, the local base plate data is updated by modifying the root canal curve, identifying the end point of the calibration curve and the root canal vector. Based on the end point of the calibration curve and the root canal vector, a group of main cylinders and auxiliary cylinders are generated in the ROI. A plurality of metal sleeve holes are generated in the local base plate data through Boolean "subtract, add, and close" operations. The metal sleeve hole is collinear with the root canal access of the tooth specified by the doctor. The metal sleeve hole is fused with the local base plate data to generate assembly data. Based on the assembly data, the assembly data is updated through morphological smoothing processing and interference checking. Specifically, the outer part of the plastic base plate shell of the assembly data and the outer part of the metal sleeve are respectively marked with assembly marks and enhanced structural strength. In this process, the closed operation including morphological dilation and erosion is used to realize the smooth construction of the assembly surface of the geometric entity and avoid the broken surface caused by traditional triangular net level Boolean. Since the inner diameter, height and wall thickness of the metal sleeve data are explicit parameters, based on the explicit parameters, tooth position, different drill bit specifications or special-shaped root canals, the corresponding parameters can be adjusted to quickly recalculate and update the assembly data. Regardless of the number and spatial arrangement of the root canals, the algorithm can automatically generate and position multiple sleeve arrays by reading the curve nodes drawn by the doctor, realizing true "arbitrary form adaptation".

[0042] Preferably, the step of obtaining shell data compatible with the marker data further comprises: obtaining tooth mask data based on the marker data, the tooth mask data including maxilla, mandible, upper teeth, lower teeth, and mandibular canal; obtaining binary volume data through binary processing based on the tooth mask data, the binary volume data including at least data type, geometric feature, and voxel spacing; generating a hollow shell through morphological outer expansion based on the binary volume data, and obtaining shell data through a script, the shell data including data with a preset thickness and a closed structure.

[0043] Different label data in the dental mask data is converted into binary data, and based on the binary data, Hollow is automatically called in the SegmentEditor environment of 3DSlicer by using a script to offset the dental surface by 1.5 mm to obtain a shell data with a constant wall thickness, and based on the shell data, discrete isolated noise foreground points are eliminated and shell data closely bonded to the dentition without cavities are obtained through island analysis and maximum connected domain reservation operation. The above process is automatically completed through a scripted interface, and parameters such as Hollow thickness and display transparency can be modified with one key at the time of calling, without manual outlining in CAD software.

[0044] Preferably, the step of obtaining a plurality of metal sleeves based on the end point of the calibration curve and the root canal vector further comprises: A plurality of metal sleeves are obtained based on the end point of the calibration curve and the root canal vector through coordinate conversion, main cylindrical voxels, auxiliary cylindrical voxelization and array merging, and the metal sleeve includes voxel data. The metal sleeve is updated based on the metal sleeve data through drilling processing, thickening sleeve wall processing and hole filling processing.

[0045] The root canal vector is used as a reference axis, and real-time calibration is combined with the end point of the calibration curve to eliminate the positioning deviation of the traditional static template; the main cylinder defines the main body direction of the root canal, the auxiliary cylinder compensates for the morphology of the lateral root canal / isthmus, and the array density of the cylinder is automatically adjusted according to the curvature of the root canal, thereby adapting to complex root canal morphology.

[0046] Second embodiment The second aspect of the present application provides a root canal treatment guide plate design device based on deep learning, comprising: A preprocessing module obtains first data with consistent sizes based on CBCT data through preprocessing; A data generation module is configured to obtain a training data set by mapping based on the first data and treatment data corresponding to the CBCT data; A training module is configured to update network parameters based on the training data set through an nnUNet module; A scheme generation module is configured to obtain oral marking data by inputting patient CBCT data through an nnUNet module, wherein the marking data is data for identifying the dentition, single tooth, pulp cavity and root canal of the patient CBCT data and marking the same; A correction marking module is configured to update the oral marking data based on the oral marking data through artificial review and revision, and update the network parameters based on the oral marking data through the nnUNet module; A guide plate data generation module is configured to obtain assembly body data based on the oral marking data, wherein the assembly body data is an STL guide plate file and includes a metal inner ring and a plastic shell; A printing module is configured to generate a guide plate of a 3D structure based on the assembly data through 3D printing, and the guide plate of the 3D structure is used to guide the opening of the pulp and the positioning of the root canal entrance.

[0047] The preprocessing module completes CBCT data standardization in one key, replacing the traditional time-consuming manual adjustment. The data generation module realizes four-level accurate identification of dentition, single tooth, pulp cavity and root canal through nnUNet, solves the problem of strong subjectivity and easy missed detection of traditional manual labeling, and automatically calibrates the root canal vector and the end point of the calibration curve based on the CBCT image, eliminating the scanning angle deviation. The training module enhances the model's ability to distinguish complex anatomical regions through a confidence weighting mechanism, effectively reducing the influence of pseudo-label noise on training, and provides an intelligent diagnosis and treatment method in the vertical field for root canal treatment. The scheme generation module uses nnUNet to adaptively extract different case features, without the need for individual parameter adjustment for each case; the correction marking module reversely transmits artificial review opinions to the training network to realize continuous detection, correction and learning. The guide plate data generation module generates STL files through oral marking data, reads the curve nodes to automatically complete the generation and positioning of multiple sleeve arrays, and realizes true adaptive shape. The printing module seamlessly connects 3D printing equipment based on STL files, converts the doctor's experience into a repeatable standardized process, and improves the efficiency of guide plate production.

[0048] In the description of the present application, it should be noted that the terms "in", "out" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship commonly used when the product of the present application is used, and are only for the convenience of describing the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the indicated device or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application. In addition, the terms "first", "second" and the like are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0049] It should also be noted that unless otherwise explicitly specified and limited, the terms "set", "connected" should be broadly understood, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be directly connected, or indirectly connected through an intermediate medium, or it can be connected inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0050] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific execution of the recognition content of the above-described system and device can refer to the corresponding process in the foregoing method embodiment.

[0051] The embodiments of the present application are described in detail above with reference to the accompanying drawings, but the present application is not limited to the above-described embodiments. Even if various changes are made to the present application, provided that the changes fall within the scope of the present claims and their equivalents, they still fall within the protective scope of the present application.

Claims

1. A deep learning-based method for designing root canal treatment guides, characterized in that, include: First data with consistent size is obtained by preprocessing based on CBCT data; A training dataset is obtained by mapping the treatment data corresponding to the first data and the CBCT data. The network parameters are updated using the nnUNet module based on the training dataset. The input patient's CBCT data is used to obtain oral cavity marker data through the nnUNet module. The marker data is data that identifies and marks the dentition, single tooth, pulp chamber, and root canal of the patient's CBCT data. The oral cavity marker data is updated through manual review and revision based on the oral cavity marker data, and the network parameters are updated through the nnUNet module based on the oral cavity marker data. Based on the oral cavity marking data, the assembly data is obtained as an STL guide plate file and includes a metal inner ring and a plastic outer shell. Based on the assembly data, a 3D structure guide plate is generated by 3D printing. The 3D structure guide plate is used to guide the opening of the medullary canal and the positioning of the root canal entrance.

2. The deep learning-based root canal treatment guide design method according to claim 1, characterized in that, The step of obtaining first data of consistent size based on CBCT data through preprocessing further includes: Based on the CBCT data, the short side scaling factor and long side scaling factor are obtained sequentially through grayscale normalization and scaling ratio, and the range of grayscale normalization is [0, 1]. The first preprocessed data is obtained by scaling the CBCT data, the short-side scaling factor, and the long-side scaling factor through bilinear interpolation. Based on the first preprocessed data, the first data is obtained by randomly cropping and randomly flipping the data.

3. The deep learning-based root canal treatment guide design method according to claim 1, characterized in that, The steps of obtaining a training dataset by mapping based on the first data and the treatment data corresponding to the CBCT data specifically include: The training dataset is constructed based on the first data and the treatment data through density peak clustering. The training dataset includes tooth markers and corresponding index data for each panoramic data in the CBCT data. The tooth markers include at least the three-dimensional structural data of each tooth, the four-level label data, and the ROI data corresponding to the four-level label data. The four-level label data includes labels for dentition, single tooth, pulp chamber, and root canal.

4. The deep learning-based root canal treatment guide design method according to claim 1, characterized in that, The step of updating network parameters using the nnUNet module based on the training dataset further includes: Based on the training dataset, pseudo-label data is obtained through a 3DU-Net encoder. The pseudo-label data includes predicted labels for dentition, single tooth, pulp chamber, and root canal. Based on the training data and the pseudo-label data, a total loss function is constructed and the network parameters are updated. The calculation expression is as follows: L total =L seg +L topo +L conf ; In the formula L seg To monitor the loss segmentation, L topo For topological consistency loss, L conf The loss is weighted by confidence level.

5. The deep learning-based root canal treatment guide design method according to claim 4, characterized in that, The expression for calculating the supervised segmentation loss is as follows: In the formula, For voxels of pseudo-labeled data, λ dice , λ ce These are preset parameters, where c in Pc ranges from 1 to 4, representing the probability data for the dentition, single tooth, pulp chamber, and root canal, respectively. The expression for calculating the topology consistency loss is as follows: In the formula, For hyperparameters, The penalty amount for root canal data in the labeled data; The expression for calculating the confidence-weighted loss is as follows: In the formula, For hyperparameters, .

6. The deep learning-based root canal treatment guide design method according to claim 1, characterized in that, The step of obtaining assembly data based on the oral cavity marker data further includes: The three marker points of the oral marking data are manually verified and revised. These marker points are located sequentially on the crown side of each affected tooth and extend to the root side, including the coronal starting point of the most convex point of the root canal orifice, the mid-section inflection point of the apex at the first significant bend of the root canal, and the apical termination point 0.5-1mm behind the apical foramen. Based on the marker points, path data is obtained, and the path data is a smooth path fitted from the coronal starting point, the middle inflection point, and the root apex ending point. Determine the number of path data for each affected tooth. If the number is less than or equal to 1, prompt for manual verification and revision. Otherwise, the verification passes and the oral cavity marker data is updated based on the revised marker points and the path data.

7. The deep learning-based root canal treatment guide design method according to claim 1, characterized in that, The step of obtaining assembly data based on the oral cavity marker data further includes: Obtain shell data that is compatible with the marked data; The shell data is updated by removing isolated islands based on the shell data. The specific process is as follows: all connected components of the shell data are marked, the voxel volume of each connected component is calculated, and the connected component with the largest voxel volume is retained. The remaining connected components are removed and the shell data is updated. By selecting a specified area of ​​the shell data and obtaining the local base plate data corresponding to the specified area through coordinate mapping and voxel-level clipping, the local base plate data is updated by correcting the root canal curve, identifying the end point of the calibration curve and the root canal vector based on the local base plate data. Several metal sleeves are obtained based on the end point of the calibration curve and the root canal vector. Assembly data is obtained by fusing the metal sleeves and the local bottom guide plate data. Based on the assembly data, morphological smoothing and interference verification are used to obtain and export the assembly data; The assembly data is updated by reconstructing parameters based on the assembly data.

8. The deep learning-based root canal treatment guide design method according to claim 7, characterized in that, The step of obtaining shell data adapted to the marked data further includes: Based on the labeled data, tooth mask data is obtained, which includes maxilla, mandible, upper teeth, lower teeth, and mandibular canal; Binary data is obtained by binarization based on the tooth mask data. The binary data includes at least data type, geometric features, and voxel spacing. Based on the binary data, a hollow shell is generated by morphological expansion, and the shell data is obtained by script. The shell data includes data with a preset thickness and a closed structure.

9. The deep learning-based root canal treatment guide design method according to claim 7, characterized in that, The step of obtaining several metal sleeves based on the end point of the calibration curve and the root canal vector further includes: Based on the endpoint of the calibration curve and the root canal vector, several metal sleeves are obtained through coordinate transformation, principal cylinder voxelization, auxiliary cylinder voxelization, and array merging. The metal sleeves include voxel data. Based on the data of the metal sleeve, the metal sleeve is updated through drilling, thickening of the sleeve wall, and patching.

10. A deep learning-based root canal treatment guide design device, characterized in that, include: The preprocessing module obtains first data of consistent size based on CBCT data through preprocessing. The data generation module is used to obtain a training dataset by mapping the treatment data corresponding to the first data and the CBCT data. The training module is used to update network parameters through the nnUNet module based on the training dataset; The scheme generation module is used to input patient CBCT data and obtain oral cavity marker data through the nnUNet module. The marker data is data that identifies and marks the dentition, single tooth, pulp chamber, and root canal of the patient's CBCT data. The correction tagging module is used to update the oral cavity tagging data through manual review and revision based on the oral cavity tagging data, and to update the network parameters through the nnUNet module based on the oral cavity tagging data; The guide plate data generation module is used to obtain assembly data based on the oral cavity marking data. The assembly data is an STL guide plate file and includes a metal inner ring and a plastic outer shell. The printing module is used to generate a 3D structure guide plate based on the assembly data through 3D printing. The 3D structure guide plate is used to guide the opening of the medullary canal and the positioning of the root canal entrance.

Images