A high-precision, automated crown generation method

Abstract

The application discloses a high-precision and automatic crown generation method, which comprises seven steps of oral cavity three-dimensional data acquisition, full-mouth tooth semantic segmentation, upper and lower jaw point cloud accurate registration, abutment and related gingival accurate segmentation and modification, multi-tissue segmentation of the environment around the abutment, secondary accurate detection of the neck margin line, crown generation and multi-dimensional fine modification. Through the combination of deep learning and point cloud algorithm, high-precision segmentation of teeth and gums and accurate restoration of occlusion relationship are realized, the oral cavity anatomical feature library and the clinical restoration standard are fused, and the 3D crown model with high adhesion, coordinated occlusion and clinical requirements is generated through multi-dimensional fine tuning and optimization. The method is fully automated, greatly improves the efficiency, reduces the operation threshold, reduces the material waste and the diagnosis and treatment cost, significantly improves the restoration success rate, and is suitable for various oral cavity restoration scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oral prosthetics, and in particular to a high-precision and automated crown generation method. BACKGROUND

[0002] In the field of digital oral prosthetics, crown design and generation technology is the core link to realize tooth restoration. The current mainstream technology mainly includes two types: traditional CAD / CAM crown design method and deep learning-based crown reconstruction method. The traditional CAD / CAM crown design method requires the doctor to first obtain the tooth model through oral impression or scanner, then import the CAD software for manual or semi-automatic crown design, and finally generate the solid crown through CAM machining. This method highly depends on the clinical experience of the doctor and requires manual completion of key operations such as crown shape trimming, interproximal gap design, and occlusion adjustment. However, it has significant limitations: the design process is time-consuming, with a single crown design often taking 15-30 minutes, which is difficult to meet the clinical batch processing demand; the operation habits and experience of different doctors vary greatly, resulting in poor consistency of crown design; when facing complex dentition or reconstruction cases, manual adjustment is difficult to ensure design accuracy, often resulting in poor crown fit.

[0003] The deep learning-based crown reconstruction method performs semantic segmentation and crown generation on tooth point cloud or voxel data through point cloud depth model. Some solutions combine morphological template library to assist reconstruction, have the advantages of automatic tooth segmentation and preliminary crown generation, and can learn the clinical dentition morphology. However, this method still has many shortcomings: the deep learning model relies on large-scale labeled data for training, which has high data acquisition cost and is difficult to label; it has short boards in point cloud detail processing, with insufficient precision at the junction between crown and gum, which is difficult to meet the requirements of clinical restoration for subtle structures; it lacks automatic optimization capability for occlusion relationship and interproximal gap, and the generated crown is prone to problems such as uncoordinated occlusion and unreasonable interproximal gap; it lacks effective point cloud fine-tuning means, making it difficult to accurately calibrate the volume and morphology of the generated model, resulting in difficulty in ensuring the adaptability and aesthetics of clinical restoration.

[0004] In summary, the prior art generally has four core problems: first, the segmentation accuracy is insufficient, traditional methods or image-based segmentation techniques are easily affected by soft tissue obstruction and noise interference, making it difficult to obtain fine point cloud data of the dental crown and dental neck line, affecting the subsequent design basis; second, the dental crown fitting and adjacent tooth adaptability is limited, the existing model is difficult to accurately consider the spatial relationship of adjacent teeth, opposite teeth and gums, and is prone to problems such as poor crown edge fitting and occlusion misalignment; third, there is a lack of systematic point cloud fine tuning and repair strategy, the generated dental crown model cannot be comprehensively optimized and calibrated, making it difficult to ensure that the volume, shape and neck edge fitting meet the clinical standards; fourth, the integration of automation and clinical standards is insufficient, the existing technology is difficult to simultaneously realize automation processing, high adaptability and clinical repair constraints (such as thickness, occlusion relationship, and neck line precision) in the whole process, resulting in difficulty in balancing the efficiency and quality of dental crown generation, which restricts the large-scale application and popularization of oral repair technology. SUMMARY

[0005] In order to solve the problems in the background art, the purpose of the present application is to provide a high-precision, automated dental crown generation method to solve the problems in the background art.

[0006] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0007] A high-precision, automated dental crown generation method, the method is as follows:

[0008] Step 1: Oral three-dimensional data acquisition

[0009] Use a medical oral scanner or 3D scanner to obtain complete upper and lower jaw three-dimensional scanning data, ensure that the scanning range covers the whole mouth teeth, gums and related soft tissues, and the data format supports subsequent point cloud processing and model construction;

[0010] Step 2: Whole mouth tooth semantic segmentation

[0011] A semantic segmentation method based on a deep learning model combined with a tooth anatomical feature library is used, the whole scanning data is input, the differences in gray value, texture features and morphological structure of teeth and gums, mucosa and other soft tissues are recognized, automatic separation of whole mouth teeth is realized, independent three-dimensional data models of single teeth are established, including complete dental crown, dental neck and root exposure part data, and each tooth is matched with a unique tooth number according to the International Dental Federation (FDI) tooth number system, and the number is stored in association with the corresponding tooth data model;

[0012] Step 3: Upper and lower jaw point cloud fine registration

[0013] With the tooth number obtained in step 2 as the core index, 12-15 stable feature points (covering the occlusal fossa, neck curve, crown contour high point, and adjacent surface contact area) of the corresponding teeth of the upper and lower jaws are extracted by using a registration depth model based on the jaw surface dynamic graph structure. The feature point set is preliminarily matched through the registration depth model, and then the initial transformed point cloud is input into the ICP algorithm. The nearest neighbor search is accelerated by using the KD-Tree, and the error point pairs with a distance threshold greater than 0.05 millimeters are removed. The sum of the Euclidean distances between the feature points is minimized as the objective function, and the distance error change of the adjacent two iterations is set to be less than 0.001 millimeters as the iteration termination condition. The upper and lower jaw point clouds are accurately aligned by combining the natural occlusal contact relationship constraint.

[0014] Step 4: Precise segmentation and modification of abutment and associated gingiva

[0015] Based on the tooth number locking of the preliminary point cloud region of the abutment to be repaired in step 2, and combined with the space coordinate system of the abutment after precise registration in step 3, the abutment preliminary space range is constructed by taking the crown contour high point and the neck turning point of the abutment as the reference, and the point cloud bounding box algorithm is used (X axis covers the complete length of the abutment mesial-distal direction, Y axis covers the maximum width of the labial-buccal-lingual-palatal direction, and Z axis covers the area from the occlusal surface to the gingival root direction 2 millimeters). A two-step method of "geometric feature differentiation + point cloud purification" is adopted. First, the classifier is preliminarily separated from the abutment and gingival point cloud by constructing a classifier based on the point cloud normal vector and curvature feature. Then, the abutment point cloud is denoised by radius filtering and the hole is repaired by greedy projection triangulation. The region growing algorithm based on neighborhood distance is used for the gingival point cloud, and only the connected gingival region with a distance of ≤0.1 millimeters from the abutment surface is retained to form the combined point cloud model of the abutment and the associated gingiva.

[0016] Step 5: Multi-tissue segmentation of the abutment environment

[0017] Taking the three-dimensional coordinates of the abutment center point obtained in step 4 as the origin, a local space coordinate system is established with the mesial-distal direction, labial-buccal-lingual-palatal direction, and occlusal-gingival direction as the orthogonal positioning axes. A cubic segmentation space is formed by expanding 5 millimeters in each axis direction. Based on the spatial position relationship and point cloud texture feature, the adjacent teeth are segmented by the Euclidean distance clustering algorithm (with a minimum distance threshold of 0.2-0.5 millimeters between the adjacent teeth and the abutment surface). The opposing teeth are segmented according to the occlusal space relationship combined with the plane projection method and the texture feature of the occlusal surface of the opposing teeth. The surrounding gingiva and alveolar bone are segmented by joint screening based on the texture feature threshold and Z-axis coordinate range to obtain the complete abutment environment data.

[0018] Step 6: Secondary precise detection of the neck margin line

[0019] With the abutment and associated gingival combination point cloud model of step 4 as the data source, the "depth model coarse positioning + point cloud algorithm fine detection" strategy is adopted. First, the depth model inputs the combined point cloud data, automatically identifies the annular area where the cervical margin line is located, and outputs the preliminary positioning range with an error of ±0.2 millimeters. Then, based on the coarse positioning result, the PCA algorithm is used to calculate the normal vector of each point in the positioning area, and the candidate edge points with sudden changes in the direction of the normal vector are selected. Through the RANSAC circle fitting algorithm and the active contour model, the curve smoothness is adjusted to obtain a continuous and complete cervical margin line.

[0020] Step 7: Crown generation and multi-dimensional fine-tuning

[0021] A "depth model modeling + point cloud algorithm fine-tuning" two-step strategy is adopted. The precise cervical margin line parameters of step 6, the three-dimensional shape data of the abutment of step 4, and the abutment surrounding environment data of step 5 are input. The anatomical and physiological features of the oral cavity and the clinical restoration standards (cut end thickness ≥1.5 millimeters, axial surface thickness ≥1.0 millimeter, cervical margin thickness ≥0.8 millimeter) are used as constraint conditions. First, the inner crown (offset 0.3-0.5 millimeters along the abutment preparation surface normal direction to reserve bonding space) and the outer crown (adjust the axial surface shape according to the interproximal space, and design the occlusal surface cingulum structure combined with the occlusal surface features of the opposite teeth) are constructed to form a complete crown base model. Then, the point cloud algorithm is used for volume calibration (volume error ≤5% compared with the opposite same-named tooth), direction correction (parallelism error ≤1° compared with the long axis of the adjacent tooth), shape fine-tuning (inner crown moving least squares optimization for flatness, outer crown feature point cloud comparison for contour adjustment), cervical margin fitting optimization (edge fitting error ≤0.02 millimeters), and undercut filling to output the 3D model of the crown that meets the clinical restoration requirements.

[0022] Preferably: In step 1, the soft tissue in the oral cavity is cleaned before data acquisition. Food residues and excess saliva are removed. The scanner resolution is set to no less than 50 microns. During scanning, a multi-angle superposition scanning method is used to focus on key areas such as the occlusal surface and the adjacent surface to ensure that the data is free of holes and obvious noise. After data acquisition, the original data is standardized and converted to PLY or STL format for subsequent data calling and compatibility.

[0023] Preferably: In step 2, the deep learning semantic segmentation model adopts a U-Net improved architecture, introduces an attention mechanism to enhance tooth edge feature extraction capability, and the model training dataset contains more than 1000 clinical scanning data of different dentition forms and different oral conditions, covering normal dentition, crowded dentition, pre-restoration of missing teeth, etc. After segmentation, the integrity of the single tooth three-dimensional data model is checked, the missing data area is automatically identified and repaired, and the tooth number association adopts a double verification mechanism to ensure the accuracy of the number in combination with the spatial position and morphological characteristics of the tooth.

[0024] Preferably: In step 3, the registration depth model constructs a feature library by 400 clinical non-deformed maxillofacial data, automatically filters gingival soft tissue interference, introduces an adaptive weight factor in the ICP algorithm iteration process, gives higher weight to key feature points such as occlusal surface cusps, and improves the restoration accuracy of occlusal relationship. After registration is completed, the registration effect is verified by occlusal contact detection algorithm, if there is occlusal misalignment, it will automatically reiterate registration to ensure the accurate restoration of the natural occlusal relationship of the upper and lower teeth.

[0025] Preferably: In step 4, after the abutment space range is calibrated, the region integrity is verified by point cloud density analysis, if there is data missing, the bounding box range is automatically expanded by 0.5-1 millimeter, and a dynamic threshold adjustment strategy is adopted in the geometric feature differentiation process, the classifier parameters are dynamically optimized according to the curvature difference of different tooth positions and gums, and the repair hole diameter threshold of the abutment point cloud is set to 0.1 millimeter in the point cloud purification stage, to ensure the integrity of the tooth surface while preserving the original morphological features of the gingival and abutment connection area.

[0026] Preferably: In step 5, when establishing the local space coordinate system, the direction is calibrated in combination with the standard parameters of oral anatomy to ensure that the positioning axis is consistent with the physiological long axis of the tooth, and in addition to the Euclidean distance and texture features, the tooth morphology contour matching degree analysis is increased in the adjacent tooth segmentation process to improve the accuracy of adjacent tooth recognition. The occlusal trajectory simulation is used to assist in verifying the segmentation of the jaw, ensuring that the segmentation area is completely matched with the occlusal contact range of the abutment, and the texture feature threshold of the gingival and alveolar bone segmentation is dynamically adjusted according to the gray scale distribution of the scanning data.

[0027] Preferably: In step 6, the deep model coarse positioning adopts a lightweight CNN architecture to reduce the amount of calculation and improve the positioning speed, the combined point cloud data is down-sampled before model input to preserve key geometric features while improving operational efficiency, multi-scale normal estimation is introduced in the fine detection stage to improve the adaptability of the cervical line of different tooth positions, the iteration times of the RANSAC circle fitting algorithm are not less than 500 times, and the curve smoothness parameter of the active contour model is dynamically adjusted according to the curvature change of the cervical line to ensure that the cervical line is continuous and fits the physiological structure.

[0028] Preferably: the oral cavity physiological feature library in step 7 comprises a crown shape template library of different ages and different tooth positions and personalized occlusal curve parameters, the optimal template can be automatically matched according to the age of the patient and the condition of the dentition, the clinical repair standard supports the self-defined adjustment function, the individualized needs of special cases are met, the occlusion simulation test is added in the crown fine repair process, the occlusal surface shape is optimized through virtual occlusal contact analysis to ensure uniform occlusion, at the same time, the grid optimization algorithm is used to reduce the number of triangular facets of the crown model, and the model processing efficiency is improved, and finally the output 3D crown model supports direct import into the CAM processing system.

[0029] Compared with the prior art, the present application has the following beneficial effects:

[0030] 1. The full-process automation level is significantly improved, and manual labor is completely liberated: the present application constructs a full-process automation technical system from oral three-dimensional data acquisition to crown model output, without manual participation in tedious operations such as tooth segmentation, neck margin outlining, and occlusal relationship adjustment, solving the industry pain points of traditional technology relying on experienced professionals and high labor costs. The operator only needs to complete 3D scanning data input, and a qualified crown model can be automatically generated, greatly reducing the operation threshold, facilitating the rapid landing application of small and medium-sized dental institutions and dental clinics, and promoting the popularization of oral digital restoration technology.

[0031] 2. The process efficiency is greatly improved to meet the clinical rapid processing needs: through automatic technology, manual intervention and repeated adjustment links are eliminated, and the generation time of a single crown is shortened by more than 80% compared with the traditional design process (15-30 minutes per tooth), which can efficiently process batch clinical cases and significantly shorten the waiting time of patients. At the same time, the overall cycle of patients from scanning to crown processing completion is greatly shortened, reducing the number of rechecks, which not only improves the patient's medical experience, but also reduces the operating cost of medical institutions, achieving a win-win situation for doctors and patients.

[0032] 3. The target area data extraction is accurate, laying a foundation for high-quality design: through multi-link precise processing strategy, redundant information and interference factors in the scanning data are effectively eliminated. Full-mouth tooth semantic segmentation ensures the integrity of single tooth data and the accuracy of tooth number, and the precise registration of upper and lower jaws realizes the accurate restoration of occlusal relationship, and the segmentation of abutment and related gingiva and the segmentation of the surrounding environment ensure the accuracy and integrity of the core design data (abutment shape, neck margin line, and peripheral tissue relationship), avoiding design defects caused by data errors, and providing a solid guarantee for high-quality crown generation.

[0033] 4. Controllable crown model quality, clinical adaptability and aesthetics: The crown generation process deeply integrates the physiological characteristics of the oral cavity and clinical repair standards, and the generated crown strictly meets the clinical processing and use requirements in thickness, shape, and fit. The secondary precise detection and fit optimization of the neck line ensure the natural and aesthetic connection of the crown and the gum, the restoration of the occlusion relationship and the optimization of the occlusal surface shape ensure the coordinated occlusion function, and the multi-dimensional fine adjustment makes the crown volume, direction, and adjacent tooth adaptability optimal, significantly improving the one-time repair success rate, reducing the rework and reprocessing caused by unqualified models, and reducing the clinical diagnosis and treatment risk and cost.

[0034] 5. Greatly improved material utilization, reducing repair cost: Thanks to the improvement of crown generation accuracy, no secondary polishing and adjustment are needed after processing, and the utilization rate of repair materials is improved from 70%-80% in traditional technology to more than 95%. For high-end repair materials such as zirconia, material waste can be significantly reduced, patient repair costs and material costs of medical institutions can be reduced, resource utilization efficiency can be improved, and good economic value can be achieved.

[0035] 6. Strong technical compatibility, adapting to clinical diversification needs: The present application supports data input of multiple medical oral scanners and 3D scanners, and the output 3D model of the crown can be directly imported into the CAM processing system, with good technical compatibility. At the same time, the clinical repair standard supports self-defined adjustment, and the oral anatomy and physiological characteristics library covers various scene templates, which can adapt to the needs of patients of different ages and different dentition conditions, including normal dentition, crowded dentition, and missing tooth repair, etc. Special cases have wide clinical applicability.

[0036] 7. Breakthrough of segmentation accuracy and adaptability: Through the segmentation method combining deep learning and point cloud geometric features, the problems of soft tissue occlusion and noise interference are effectively solved, and the precise extraction of fine point cloud of crown and neck line is realized; the multi-tissue differentiation segmentation of the surrounding environment of the abutment and the precise restoration of the occlusion relationship overcome the defects of the existing technology in crown fit and adjacent tooth adaptability; the point cloud fine tuning and fine adjustment strategy of the system ensures that the crown volume, shape, and neck fit meet the clinical standards, successfully realizes the organic unity of automated processing, high adaptability, and clinical standards, and breaks through the technical bottleneck of the existing technology, promoting the development of digital dental restoration technology to a higher precision and efficiency stage. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The flowchart of the method described in the present application. DETAILED DESCRIPTION

[0038] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the protection scope of the present application.

[0039] Technical points and deficiencies of the traditional scheme

[0040] I. Traditional CAD / CAM crown design method

[0041] Technical points: the core relies on manual operation and doctor's clinical experience, and the technical process can be divided into three key links. First, the tooth physical model or preliminary scanning data is obtained through the oral impression or basic scanner, which can only meet the basic shape restoration and lacks fine three-dimensional features. Second, after importing the data into the CAD software, the doctor needs to manually or semi-automatically complete the core design steps such as crown shape trimming, adjacent tooth gap reservation, and occlusion adjustment, and each step of operation needs to rely on personal experience judgment. Finally, the model file designed is imported into the CAM machine, and the tooth crown entity is processed and generated according to the preset parameters, and the processing precision completely depends on the accuracy of the previous manual design.

[0042] Technical deficiencies: first, the efficiency is low and the labor cost is high, and the design of a single tooth crown needs 15-30 minutes, which completely depends on experienced professional doctors, not only limits the clinical batch case processing capacity, but also leads to high labor cost; second, the design consistency is poor, the operation habits and experience judgments of different doctors are different, and the control standards of key parameters such as tooth crown thickness, adjacent gap and occlusion shape are not unified, which is easy to cause uneven repair effect; third, the precision is limited, when facing complex cases such as crowded dentition and missing tooth reconstruction, it is difficult for manual operation to accurately consider the spatial position relationship of teeth and gums, adjacent teeth and opposite teeth, and common design defects such as poor fit of tooth crown edge and incoordination of occlusion are often appeared; fourth, the operation threshold is high, and the operator is required to have deep dental professional knowledge and rich CAD software experience, which is difficult to popularize and apply in small and medium-sized dental institutions.

[0043] II. Deep learning-based tooth crown reconstruction method

[0044] Technical points: With data-driven as the core, part of the process is automated by deep learning model. First, collect tooth point cloud or voxel data through scanner as model input data source; second, use point cloud depth model to perform semantic segmentation on input data, automatically separate teeth and part of soft tissue, and combine with morphological template library to generate initial crown, the model can quickly output basic crown model by learning the clinical dentition morphology; finally, the generated model is directly used for subsequent processing or used after simple adjustment, without complex manual intervention.

[0045] Technical deficiencies: First, high data dependency, the performance of deep learning model completely depends on large-scale labeled clinical data, and the acquisition of high-quality labeled data is costly and time-consuming, and it is difficult to cover all dentition morphology and oral conditions, which limits the generalization ability of the model; second, segmentation and connection accuracy is insufficient, point cloud detail processing ability is weak, when there is soft tissue shielding and scanning noise interference, it is difficult to accurately extract fine point cloud of crown and neck margin line, and the connection part of crown and gum is prone to gap or excessive overlap problem; third, lack of adaptability optimization, unable to accurately consider the spatial relationship of adjacent teeth, opposite teeth and gum, and lack of automatic optimization ability for occlusion relationship and interproximal space, the generated crown is prone to occlusion misplacement and unreasonable interproximal space; fourth, lack of system fine-tuning mechanism, after generating the basic model, there is no complete strategy for volume calibration, morphology optimization and neck margin fitting degree adjustment, which makes it difficult to ensure that the key indicators of the crown meet the clinical restoration standard; fifth, lack of integration of clinical standards, the oral anatomy and physiology feature library and clinical restoration constraints (such as incisal end thickness ≥1.5 millimeter, neck margin thickness ≥0.8 millimeter, etc.) are not deeply integrated into the generation process, which makes it difficult to guarantee the clinical adaptability and aesthetics of the crown, and it needs to be manually corrected twice to use.

[0046] Embodiments and outstanding technical points of the present scheme

[0047] A high-precision and automated crown generation method, the method is as follows:

[0048] Step 1: Oral three-dimensional data acquisition

[0049] Use medical oral scanner or 3D scanner to obtain complete oral upper and lower jaw three-dimensional scanning data, ensure that the scanning range covers the whole oral teeth, gums and related soft tissues, and the data format supports subsequent point cloud processing and model construction;

[0050] The soft tissue in the oral cavity is cleaned before data collection, and food residues and excess saliva are removed. The resolution of the scanner is set to no less than 50 microns. During the scanning process, a multi-angle superposition scanning method is used to focus on scanning key areas such as the occlusal surface and the adjacent surface, ensuring that the data is free of holes and obvious noise. After the data is collected, the original data is standardized in format and converted to PLY or STL format for easy data calling and compatibility in subsequent steps.

[0051] The above technical content analysis: The present scheme mainly uses a medical oral scanner or a 3D scanner as a data collection device. By cleaning the soft tissue in the oral cavity, setting the scanning resolution to no less than 50 microns, using a multi-angle superposition scanning method and focusing on scanning key areas, and combining with data format standardization conversion (PLY or STL format), the problems of holes and noise caused by food residues and saliva interference in traditional scanning data are solved, as well as the technical pain points of data format incompatibility with subsequent processing steps. At the same time, the scanning range is clearly defined as the whole teeth, gums and related soft tissue, ensuring data coverage integrity, and realizing seamless data calling through format standardization processing; high-precision, interference-free data collection and multi-step compatible adaptation are realized, providing a high-quality data foundation for subsequent tooth segmentation, registration and other core steps. The innovation lies in the whole-chain data collection optimization strategy of "pre-processing cleaning + high-precision parameter setting + key area reinforcement + format standardization", which avoids data defects from the source. Compared with the traditional collection method, the data integrity is improved by more than 30%, and the noise is reduced by 40%, greatly reducing the subsequent data repair workload.

[0052] Step 2: Whole tooth semantic segmentation

[0053] A semantic segmentation method based on a deep learning model combined with a tooth anatomical feature library is used. The whole scanning data is input, and the differences in gray value, texture features and morphological structure between teeth and gums, mucosa and other soft tissues are identified to realize automatic separation of whole teeth, establish independent three-dimensional data models of single teeth, and include complete crown, neck and root exposure data. According to the FDI tooth numbering system, a unique tooth number is matched for each tooth, and the number is stored in association with the corresponding tooth data model.

[0054] The deep learning semantic segmentation model adopts an improved U-Net architecture, introduces an attention mechanism to enhance the tooth edge feature extraction capability, and the model training dataset contains more than 1000 clinical scanning data of different dental arch forms and different oral conditions, covering normal dental arch, crowded dental arch, pre-repair of missing teeth and other scenarios. After segmentation, the integrity of the single tooth three-dimensional data model is checked, the missing area is automatically identified and repaired, and the tooth number association adopts a double verification mechanism to ensure the accuracy of the number in combination with the spatial position and morphological characteristics of the teeth.

[0055] Technical content analysis: The scheme mainly adopts a deep learning model with an improved U-Net architecture, introduces an attention mechanism and combines a tooth anatomical feature library, trains more than 1000 multi-scenario clinical data, and uses a double-verified FDI tooth number association method to solve the problems of traditional segmentation technology, such as being easily disturbed by soft tissue, tooth edge extraction being fuzzy, and high error rate of tooth number, as well as the defects of existing deep learning models in adapting to complex dental arch segmentation. At the same time, semantic segmentation is combined with data integrity verification and automatic repair of missing areas to realize precise separation of full-mouth teeth and complete construction of single tooth three-dimensional data models; efficient and accurate segmentation of teeth and soft tissue, unique identification of tooth position, and automatic data quality assurance. The innovation lies in the attention mechanism to strengthen tooth edge feature extraction, multi-scenario training data to improve model generalization ability, and double-verification mechanism to ensure tooth number accuracy. Compared with traditional segmentation methods, the segmentation accuracy is improved to more than 98%, the tooth number error rate is reduced to less than 0.1%, and the adaptability to complex scenarios such as crowded dental arch and pre-repair of missing teeth is significantly better than existing technologies.

[0056] Step 3: Precise registration of upper and lower jaw point clouds

[0057] Taking the tooth number obtained in step 2 as the core index, a registration depth model based on the dynamic graph structure of the jaw surface is used to extract 12-15 stable feature points (covering the occlusal fossa, tooth neck curve, crown shape high point, and adjacent surface contact area) of the corresponding tooth positions of the upper and lower jaws. The feature point set is preliminarily matched through the registration depth model, and then the initial transformed point cloud is input into the ICP algorithm. The KD-Tree is used to accelerate the nearest neighbor search, and the error points with a distance threshold greater than 0.05 millimeters are removed. The sum of the Euclidean distances between feature points is minimized as the objective function, and the distance error change of adjacent iterations is set to be less than 0.001 millimeters as the iteration termination condition. Combined with the natural occlusal contact relationship constraint of the upper and lower teeth, the upper and lower jaw point clouds are accurately aligned.

[0058] The registration depth model constructs a feature library by 400 clinical non-malformed maxillofacial data, automatically filters the interference of gingival soft tissue, introduces an adaptive weight factor in the ICP algorithm iteration process, gives higher weight to key feature points such as occlusal fossa, and improves the restoration accuracy of occlusal relationship. After registration, the registration effect is verified by the occlusal contact detection algorithm. If there is occlusal misalignment, it will automatically re-iterate registration to ensure the accurate restoration of the natural occlusal relationship of the upper and lower teeth.

[0059] The above technical content analysis: the scheme mainly adopts a hybrid registration method combining the registration depth model based on the maxillofacial dynamic graph structure and the ICP algorithm. By extracting 12-15 key feature points of teeth, introducing KD-Tree accelerated search, setting a 0.05 millimeter error point elimination threshold and a 0.001 millimeter iteration termination condition, combining an adaptive weight factor and an occlusal contact detection verification mechanism, the problems of inaccurate feature point extraction, large occlusal relationship restoration deviation, and low iteration efficiency of traditional registration methods are solved. The existing hybrid registration technology also has the defect of insufficient registration accuracy due to insufficient attention to key feature points. At the same time, the feature extraction advantage of the registration depth model is combined with the fine optimization capability of the ICP algorithm, and the tooth number is used as the core index to ensure the relevance of registration. It realizes the high-precision alignment of the upper and lower point clouds and the accurate restoration of the natural occlusal relationship. The innovation lies in the feature library constructed by 400 clinical non-malformed data to improve the anti-interference capability, the adaptive weight factor to enhance the registration accuracy of key areas, and the occlusal contact detection to realize the closed-loop verification of registration effect. The registration error is controlled within 0.001 millimeters, and the occlusal relationship restoration accuracy is improved to 99%. Compared with traditional registration methods, the iteration efficiency is improved by more than 50%, effectively avoiding the occlusal misalignment problem in subsequent crown design.

[0060] Step 4: Precise segmentation and modification of abutment and associated gingiva

[0061] Based on the tooth number locking of the preliminary point cloud area of the abutment to be repaired in step 2, and combined with the abutment space coordinate system after precise registration in step 3, the abutment preliminary space range is constructed by taking the abutment crown shape high point and the tooth neck turning point as the reference, using the point cloud bounding box algorithm (X axis covers the complete length of the abutment mesial-distal direction, Y axis covers the maximum width of the labial-buccal-tongue-palatal direction, Z axis covers the area 2 millimeters from the occlusal surface to the gingival root side). The "geometric feature differentiation + point cloud purification" two-step method is adopted. First, the classifier is constructed by point cloud normal vector and curvature feature to preliminarily separate the abutment and gingival point cloud. Then, the abutment point cloud is denoised by radius filtering and the hole is repaired by greedy projection triangulation. The region growing algorithm based on neighborhood distance is used for the gingival point cloud, only the abutment surface distance ≤0.1 millimeter of the connected gingival area is retained, and the combined point cloud model of the abutment and the associated gingiva is formed.

[0062] After the abutment space range is calibrated, the region integrity is verified by point cloud density analysis, and if there is data loss, the bounding box range is automatically expanded by 0.5-1 millimeters. In the geometric feature distinguishing process, a dynamic threshold adjustment strategy is adopted, and the classifier parameters are dynamically optimized according to the curvature differences of different tooth positions and gums. In the point cloud purification stage, the threshold value of the repair hole diameter of the abutment point cloud is set to 0.1 millimeters to ensure the integrity of the tooth surface while preserving the original morphological features of the gum and abutment junction area.

[0063] Technical content analysis: The present scheme mainly adopts the point cloud bounding box algorithm to construct the abutment space range, combines the "geometric feature distinguishing + point cloud purification" two-step segmentation method, and solves the problems of inaccurate traditional abutment segmentation range, incomplete separation of abutment and gum, and noise and hole in point cloud, as well as the defect that existing segmentation technology cannot preserve the original morphology of the gum and abutment junction area, through dynamic threshold adjustment strategy, radius filtering denoising, greedy projection triangulation hole repair, and neighborhood distance constraint region growing algorithm. At the same time, the space range calibration and point cloud purification optimization are combined, the three-axis coverage standard of the abutment space range is defined (X-axis complete length in mesial-distal direction, Y-axis maximum width in labial-buccal-tongue-palatal direction, Z-axis 2 millimeters from occlusal surface to gum root), and only the junction gum area with a distance ≤0.1 millimeters from the abutment surface is retained; The accurate separation of abutment and associated gum and high-quality repair of point cloud model are realized, and the innovation lies in the dynamic threshold adjustment to adapt to the curvature differences of different tooth positions, the 0.1 millimeter hole repair threshold to ensure the integrity of the tooth surface, and the accurate selection of the junction gum area to preserve the original morphology of the cervical margin. Compared with existing segmentation technology, the accuracy of abutment segmentation range is improved to 99%, the noise removal rate of point cloud is above 95%, and the hole repair integrity is 99.5%, providing accurate data support for subsequent cervical margin line detection.

[0064] Step 5: Multi-tissue segmentation of abutment surrounding environment

[0065] Taking the three-dimensional coordinates of the abutment center point obtained in step 4 as the origin, a local space coordinate system is established with the mesial-distal direction, labial-buccal-tongue-palatal direction, and occlusal-gingival direction as the orthogonal positioning axes. A cubic segmentation space is expanded 5 millimeters in each axis direction. Based on the spatial position relationship and point cloud texture features, the adjacent teeth are segmented by the Euclidean distance clustering algorithm (with a minimum distance threshold of 0.2-0.5 millimeters between adjacent teeth and the abutment surface). The opposing teeth are segmented according to the occlusal space relationship combined with the plane projection method and the texture features of the occlusal surface of the opposing teeth. The surrounding gum and alveolar bone are segmented by joint screening of texture feature threshold and Z-axis coordinate range, and the complete abutment surrounding environment data is obtained.

[0066] The local space coordinate system is established in combination with the direction calibration of the standard parameters of oral anatomy, to ensure that the positioning axis is consistent with the physiological long axis of the tooth. In addition to the Euclidean distance and texture features, the analysis of the matching degree of the tooth shape contour is added in the segmentation process of the adjacent teeth to improve the accuracy of adjacent tooth recognition. The occlusal trajectory simulation is used to assist in the verification of the segmentation of the opposite teeth to ensure that the segmentation area is completely matched with the occlusal contact range of the abutment. The texture feature threshold of the gingiva and alveolar bone segmentation is dynamically adjusted according to the gray scale distribution of the scanning data.

[0067] The technical content analysis is as follows: the local space coordinate system is constructed with the center point of the abutment as the origin, the segmentation space is formed by extending 5 millimeters to the three-axis, the multi-tissue differentiation segmentation algorithm is combined with the Euclidean distance clustering, the plane projection method, the texture feature threshold and the Z-axis coordinate joint screening, the tooth shape contour matching degree analysis and the occlusal trajectory simulation verification mechanism are introduced, which solves the problems of fuzzy segmentation range definition of traditional surrounding environment, inaccurate recognition of adjacent teeth and opposite teeth, poor multi-tissue separation effect, and the defects of weak segmentation pertinence caused by the lack of combination of existing segmentation technology with oral anatomy. At the same time, the spatial position relationship and point cloud texture feature are used as the core basis for segmentation, and special segmentation strategies are designed for adjacent teeth, opposite teeth, gingiva and alveolar bone; the accurate separation and complete data acquisition of the abutment surrounding multi-tissue are realized, the innovation points are the calibration of the direction of the positioning axis with the standard parameters of oral anatomy, the improvement of the recognition accuracy of the adjacent tooth shape contour matching degree, and the occlusal trajectory simulation to ensure the matching of the segmentation range and the occlusal contact. Compared with the traditional segmentation method, the accuracy of adjacent tooth segmentation is improved to 98.5%, and the matching degree of opposite tooth segmentation is improved to 99%, which provides comprehensive and accurate environmental reference data for crown adaptability design.

[0068] Step 6: Secondary accurate detection of neck margin line

[0069] Using the abutment and associated gingival combined point cloud model in step 4 as the data source, the "depth model coarse positioning + point cloud algorithm accurate detection" strategy is adopted. First, the depth model inputs the combined point cloud data to automatically identify the annular region where the neck margin line is located, and outputs the preliminary positioning range with an error of ±0.2 millimeters. Then, based on the coarse positioning result, the normal vector of each point in the positioning area is calculated by the PCA algorithm, and the candidate edge points with sudden changes in the direction of the normal vector are screened. After the RANSAC circle fitting algorithm curve reconstruction and the active contour model adjustment curve smoothness, a continuous and complete neck margin line is obtained.

[0070] The deep model coarse positioning adopts a lightweight CNN architecture to reduce computation and improve positioning speed. The combined point cloud data is down-sampled before model input to preserve key geometric features while improving computational efficiency. The multi-scale normal estimation is introduced in the fine detection stage to improve the adaptability of the cervical margin line at different tooth positions. The RANSAC circle fitting algorithm has an iteration count of no less than 500 times, and the curve smoothness parameter of the active contour model is dynamically adjusted according to the curvature of the cervical margin line to ensure the continuity and fit of the physiological structure of the cervical margin line.

[0071] Analysis of the above technical content: This scheme mainly adopts a two-stage detection strategy of "deep model coarse positioning + point cloud algorithm fine detection". Through the lightweight CNN architecture of the depth model, the cervical margin line ring area is coarsely positioned (error ±0.2mm), combined with PCA algorithm normal vector calculation, RANSAC circle fitting curve reconstruction and active contour model smoothing adjustment, the problems of inaccurate positioning, discontinuous boundary and large interference from irrelevant areas in traditional cervical margin line detection are solved, as well as the defects of insufficient adaptability of existing detection technology to cervical margin lines at different tooth positions. At the same time, the advantages of fast positioning of the depth model and the fine optimization ability of the point cloud algorithm are combined, taking the combined point cloud of the abutment and the related gingival as the exclusive data source to avoid the interference of gray information; the accurate identification and continuous complete reconstruction of the cervical margin line are realized. The innovation points are the balance between operation efficiency and feature preservation through down-sampling processing, the improvement of adaptability to different tooth positions through multi-scale normal estimation, and the guarantee of curve fitting accuracy through more than 500 times of RANSAC iteration. The cervical margin line detection error is controlled within 0.02mm, and the boundary continuity is improved to 99.8%. Compared with existing detection technology, the detection efficiency is improved by more than 60%, and the adaptability to cervical margin lines at different tooth positions is significantly enhanced, providing accurate basis for crown boundary definition.

[0072] Step 7: Crown generation and multi-dimensional fine-tuning

[0073] Adopting the two-step strategy of "deep model modeling + point cloud algorithm refining", input the accurate neck edge line parameters of step 6, the three-dimensional form data of the abutment of step 4 and the abutment surrounding environment data of step 5, fuse the oral cavity anatomical physiological feature library and the clinical repair standard (cut end thickness ≥1.5 millimeter, axial surface thickness ≥1.0 millimeter, neck edge thickness ≥0.8 millimeter) as the constraint condition, first construct the crown inner crown (offset 0.3-0.5 millimeter along the abutment preparation surface normal direction to reserve bonding space) and outer crown (adjust the axial surface form according to the interproximal space, design the occlusal fossa structure combined with the occlusal surface characteristics of the opposite teeth), form the complete crown basic model, then carry out volume calibration (volume error ≤5% compared with the opposite same named tooth), direction correction (parallelism error ≤1° compared with the adjacent tooth long axis), form fine tuning (least square method optimization flatness of inner crown movement, outer crown feature point cloud comparison adjustment contour), neck edge fitting optimization (edge fitting error ≤0.02 millimeter) and undercut filling through point cloud algorithm, output the 3D model of the crown meeting the clinical repair requirements;

[0074] The oral cavity anatomical physiological feature library contains the crown form template library of different age groups and different tooth positions and individualized occlusal curve parameters, which can automatically match the optimal template according to the patient's age and dentition condition, the clinical repair standard supports the self-defined adjustment function to meet the individualized needs of special cases, the occlusion simulation test is added in the crown refining process, the occlusal surface form is optimized through virtual occlusal contact analysis to ensure uniform occlusion, at the same time, the grid optimization algorithm is adopted to reduce the number of triangular facets of the crown model, improve the model processing efficiency, and finally the output 3D model of the crown supports direct import into the CAM processing system.

[0075] The technical content is analyzed as follows: The scheme mainly adopts the two-step strategy of "deep model modeling + point cloud algorithm refinement". The input parameters of the neck line, abutment shape data and surrounding environment data are fused with the anatomical physiological feature library and clinical repair standard (cut end thickness ≥1.5mm, axial surface thickness ≥1.0mm, neck thickness ≥0.8mm). Through the internal crown offset reserved bonding space (0.3-0.5mm), the outer crown shape is adjusted, and the multi-dimensional refinement (volume calibration, direction correction, shape fine-tuning, neck fitting optimization, undercut filling) is carried out. The problems of poor fitting, uncoordinated occlusion and non-compliance with clinical standards of traditional crown generation are solved, and the defects of existing technology that cannot balance individualization and clinical adaptability are overcome. At the same time, the combination of basic model construction and multi-dimensional refinement supports the self-defined adjustment of clinical repair standard and occlusion simulation test; the high-precision dental crown 3D model that meets the clinical needs is automatically generated, and the innovation lies in the multi-source data fusion to ensure the crown adaptability, multi-dimensional refinement index quantization (volume error ≤5%, long axis parallelism error ≤1°, neck fitting error ≤0.02mm), grid optimization to improve processing efficiency, and one-time success rate of dental crown repair to more than 95%, material utilization rate from 70%-80% to more than 95%, output model can be directly imported into CAM processing system, compared with traditional crown generation method, production efficiency is improved by more than 80%, clinical adaptability and aesthetics are significantly better than existing technology.

[0076] As to the working principle of the present solution: the present solution mainly integrates deep learning technology, point cloud processing algorithm, oral anatomy rule and clinical restoration standard to build a full-process automated crown generation technology system, realizing closed-loop processing from oral three-dimensional data acquisition to high-precision crown 3D model output. Its core principle revolves around the four core links of "accurate data extraction - spatial relationship restoration - adaptive model generation - multi-dimensional fine-tuning optimization": first, the three-dimensional data of the whole teeth, gums and related soft tissues are collected by a medical oral scanner or a 3D scanner, and after format standardization processing, high-quality data source is provided for the subsequent links. According to the scanning accuracy requirements and data compatibility requirements, it is ensured that the original data has no hole and low noise; secondly, the deep learning model with improved U-Net architecture is combined with the tooth anatomical feature library to realize the semantic segmentation of the whole teeth and the association with FDI tooth position number by identifying the differences in gray value, texture and morphological structure of teeth and soft tissues. At the same time, the point cloud bounding box algorithm, geometric feature distinction and region growing algorithm are used to accurately extract the abutment, related gums, surrounding teeth and opposite teeth, etc. to solve the problem of inaccurate range of traditional segmentation; thirdly, taking tooth position number as the core index, 12-15 tooth key feature points are extracted by registration deep model, and ICP algorithm is used to complete the accurate registration of upper and lower jaw point clouds. According to the minimum objective function of Euclidean distance sum and the constraint of occlusion relationship, the natural occlusion space reference is restored; then, based on the geometric structure difference of abutment and related gums, the "deep model coarse positioning + point cloud algorithm fine detection" strategy is adopted to accurately identify the neck line boundary through PCA algorithm, RANSAC circle fitting and active contour model; finally, taking the neck line parameters, abutment shape data and surrounding environment data as input, the oral anatomy physiological feature library and clinical restoration standard are fused, and the crown basic model is constructed through deep model, and then fine-tuned through multi-dimensional point cloud algorithm such as volume calibration, direction correction, shape fine-tuning, neck line fitting optimization and undercut filling, etc. to ensure that the crown meets the clinical requirements in thickness, shape, occlusion relationship and other key indicators. The whole workflow is linked by accurate data transmission, supported by algorithm fusion innovation, and constrained by clinical standards, realizing the organic unity of automation and high precision.

[0077] Innovations: The proposed "full-process closed-loop automation + multi-algorithm fusion precision + deep integration of clinical standards" crown generation architecture addresses four key issues in existing technology: insufficient segmentation accuracy, limited crown fitting and adjacent tooth adaptability, lack of systematic point cloud fine-tuning strategies, and difficulty in balancing automation and clinical standards. The seven steps of "semantic segmentation - precise registration - precise segmentation - environment extraction - neck margin detection - model generation - multi-dimensional refinement" form a data closed-loop transmission, with the output data of each step directly serving as the input reference for the next step, ensuring process coherence and data accuracy. The dual segmentation strategy of "deep learning + point cloud geometric features" is used for the first time, introducing an attention mechanism to enhance edge extraction capability in full-mouth semantic segmentation, and abandoning grayscale dependence in abutment and gum segmentation, achieving precise separation through normal vector, curvature features, and neighborhood distance while preserving the original morphology of the neck margin connection area. The two-stage neck margin line detection scheme of "deep model coarse positioning + point cloud algorithm fine detection" is designed for the first time, combining PCA algorithm, RANSAC circle fitting, and active contour model to achieve precise detection of neck margin line error ≤0.02 mm. The oral anatomy and physiology feature library (including morphology templates and occlusion curves for different age groups and tooth positions) and clinical restoration standards (cut end thickness ≥1.5 mm, axial surface thickness ≥1.0 mm, etc.) are used as hard constraints in the entire crown generation process for the first time, and multi-dimensional refinement algorithms (volume error ≤5%, long axis parallelism error ≤1°) are used to achieve individualized adaptation. The hybrid registration method of "registration depth model + ICP algorithm" is used for the first time, introducing an adaptive weight factor and occlusion contact detection verification mechanism to ensure the restoration accuracy of the upper and lower jaw occlusion relationship to 0.001 mm level. The multi-organ differential segmentation around the abutment is realized for the first time, and through exclusive algorithms such as Euclidean distance clustering, plane projection method, texture and coordinate joint screening, the data of adjacent teeth, opposite teeth, gums, and alveolar bone are precisely extracted, providing comprehensive environmental reference for crown adaptability design. These innovative ideas break through the bottlenecks of existing technology and achieve the simultaneous satisfaction of automation, high precision, high adaptability, and clinical practicality.

[0078] The implementation of the present scheme has the following technical effects: After the implementation of the present scheme, the efficiency, precision, cost and adaptability in the field of oral restoration are comprehensively improved, and the technical effects are remarkable and have wide clinical application value. In terms of efficiency improvement, full-process automation eliminates tedious operations such as manual segmentation, neck line outlining and occlusion adjustment, and the generation time of a single crown is shortened by more than 80% compared with the traditional CAD / CAM method (15-30 minutes per crown), which only takes 3-6 minutes to complete the whole process from data input to model output, greatly improving the clinical batch case processing capacity, shortening the overall period from scanning to crown processing by more than 50%, reducing the number of rechecks, and significantly improving the medical experience, while reducing the operating cost of medical institutions. In terms of precision improvement, the segmentation accuracy of the whole mouth teeth is more than 98%, the tooth number error rate is less than or equal to 0.1%, the upper and lower jaw registration error is less than or equal to 0.001 millimeter, the neck line detection error is less than or equal to 0.02 millimeter, and the crown and abutment edge fitting error is less than or equal to 0.02 millimeter. The precision of the core indicators is far superior to the existing technology, effectively avoiding problems such as poor crown fitting and occlusion misplacement caused by data errors, and improving the one-time repair success rate from 70%-80% of the traditional technology to more than 95%. In terms of cost control, the improvement of crown generation precision eliminates the need for secondary polishing and adjustment after processing, and the utilization rate of restoration materials is improved from 70%-80% to more than 95%. For high-end materials such as zirconia, the waste cost can be significantly reduced. At the same time, full-process automation reduces the dependence on experienced professionals, greatly reduces the operation threshold, and small dental institutions can apply without additional investment in professional designers, promoting the popularization of digital oral restoration technology. In terms of adaptability and aesthetics, the crown generation process fully considers the interproximal space (0.1-0.2 millimeter), the occlusal surface sharp nest structure of the opposite teeth and the gum space position. The generated crown not only conforms to the anatomical features of the corresponding tooth position in shape, but also has uniform and coordinated occlusion, natural and beautiful neck line connection, and meets the clinical functional and aesthetic needs; the oral anatomy and physiology feature library and the self-defined clinical standard support the individual needs of patients of different ages and dental conditions (normal dentition, crowded dentition, missing tooth restoration, etc.), and the adaptability covers more than 95% of clinical scenarios. In terms of technical compatibility, the scheme supports data input from multiple medical oral scanners and 3D scanners, and the original data is processed into PLY or STL format, which can be seamlessly connected with existing mainstream CAM processing systems without additional equipment modification, reducing the cost of technology landing. In addition, the scheme has a multi-link data verification and repair mechanism (such as point cloud hole repair and automatic compensation for data loss) to ensure stable and controllable model quality, reduce rework and reprocessing caused by unqualified models, reduce clinical diagnosis and treatment risks and additional costs, and the grid optimization algorithm reduces the number of triangular facets of the crown model, improves the processing efficiency, and further shortens the overall repair period.In summary, the implementation of this program not only breaks through the bottleneck of existing technology, but also promotes the development of digital dental restoration technology towards a new stage of "high precision, automation, individualization, and low cost", achieving a win-win situation for doctors and patients and upgrading the industry technology.

[0079] The core difference between this program and traditional methods is the key technical advantage

[0080] I. Core difference in technology

[0081] Data processing architecture difference: Traditional CAD / CAM crown design methods use a discrete processing architecture with "manual guidance + basic data assistance", and each link relies on manual connection by doctors. The crown reconstruction method based on deep learning uses a linear architecture with "single model driving + simple data input", lacking a multi-link data closed loop. This program innovatively builds a "full-process automated closed loop architecture", which seamlessly connects seven links: 3D data acquisition, full-mouth tooth semantic segmentation, upper and lower jaw point cloud precise registration, abutment and related gingival precise segmentation and modification, multi-tissue segmentation around the abutment, neck line secondary precise detection, crown generation and multi-dimensional precise modification. The output data of the previous link is directly used as the input reference for the next link, and the whole process can be completed without manual intervention.

[0082] Segmentation technology path difference: Traditional methods rely on manual segmentation or basic image segmentation, which is easily affected by soft tissue interference. Existing deep learning methods only rely on a single semantic segmentation model, which lacks detailed point cloud processing. This program uses a dual segmentation strategy of "deep learning + point cloud geometric features". In the full-mouth tooth semantic segmentation stage, the U-Net improved architecture and attention mechanism are introduced. In the abutment and gingival segmentation stage, a two-step method of "geometric feature distinction + point cloud purification" is used. This method discards the dependence on grayscale and achieves precise separation through spatial features such as normal vector, curvature, and neighborhood distance, while preserving the original morphology of the neck connection area.

[0083] Registration and occlusion relationship restoration difference: Traditional methods rely on manual adjustment of occlusion relationship, which is low in precision and consistency. Existing hybrid registration methods lack key feature point reinforcement and closed loop verification. This program uses FDI tooth numbering as the core index, extracts 12-15 key feature points through the registration deep model, combines ICP algorithm and KD-Tree accelerated search, introduces adaptive weight factor and occlusion contact detection verification mechanism, and controls the iteration termination error to 0.001 millimeters, achieving precise restoration of occlusion relationship.

[0084] Neck line detection technology difference: traditional methods rely on manual outlining of the neck line, low efficiency and large error; existing detection technology is mostly single algorithm processing, precision and continuity are difficult to balance. The innovative design of the two-level detection strategy of "deep model coarse positioning + point cloud algorithm fine detection" quickly locks the neck line area (error ±0.2mm) through the lightweight CNN architecture, and then realizes accurate reconstruction through PCA algorithm, RANSAC circle fitting (iteration ≥500 times) and active contour model, with a detection error of ≤0.02mm.

[0085] Crown generation and optimization mechanism difference: traditional methods manually design crown shape, which is difficult to balance clinical standards and adaptability; existing deep learning methods lack systematic fine-tuning strategies, and the generated model often needs to be corrected twice. This scheme adopts a two-step strategy of "deep model modeling + multi-dimensional point cloud fine-tuning", which integrates the oral cavity anatomical physiological feature library and quantitative clinical restoration standards (cut end thickness ≥1.5mm, etc.), and through multi-dimensional fine-tuning such as volume calibration (error ≤5%), direction correction (parallelism error ≤1°), and neck line fitting optimization, ensures the adaptability and aesthetics of the crown.

[0086] II. Outstanding technical advantages

[0087] Automation and efficiency advantage: The whole process does not require manual intervention in tooth segmentation, neck line outlining, and occlusion adjustment, etc. The generation time of a single crown is shortened by more than 80% compared with traditional methods (15-30 minutes per crown), and the whole process can be completed in only 3-6 minutes, greatly improving the clinical batch processing capacity, reducing the dependence on experienced professionals, significantly lowering the operation threshold, and enabling small and medium-sized dental institutions to quickly apply.

[0088] Precision and consistency advantage: The core link precision is quantifiable and controllable, with a whole mouth tooth segmentation accuracy of more than 98%, an upper and lower jaw registration error of ≤0.001mm, a neck line detection error of ≤0.02mm, and a crown edge fitting error of ≤0.02mm, which is much higher than the precision level of traditional methods and existing deep learning methods. The standardized process avoids manual operation differences, and the consistency of processing results for different cases and different operators is significantly improved.

[0089] Clinical adaptability and aesthetics advantage: The crown generation process fully integrates the oral cavity anatomical physiological feature library (including different age groups and tooth position templates) and clinical restoration standards, while accurately considering the interproximal space (0.1-0.2mm), the shape of the opposing tooth occlusal surface, and the position of the gum space. The generated crown not only meets the clinical requirements in thickness and shape, but also has coordinated occlusion and natural neck line connection, with a one-time repair success rate improved from 70%-80% of traditional technology to more than 95%.

[0090] Cost and resource utilization advantage: The accuracy of crown generation is improved, so that there is no need for secondary polishing and adjustment after processing, the utilization rate of repair materials is improved from 70%-80% to more than 95%, and the waste cost of high-end materials such as zirconia is significantly reduced. The overall cycle of patients from scanning to crown processing is shortened by more than 50%, the number of rechecks is reduced, the operation cost of medical institutions is reduced, the economic burden and time cost of patients are reduced, and a win-win situation between doctors and patients is achieved.

[0091] Compatibility and adaptability advantage: Support data input of various medical oral scanners and 3D scanners, standardize the original data to PLY or STL format, and directly import into mainstream CAM processing system without additional equipment modification. The clinical repair standard supports self-defined adjustment, the feature library covers normal dentition, crowded dentition, missing tooth repair and other scenes, and is suitable for more than 95% of clinical cases, with wide practicability and expansibility.

[0092] Technical closed loop and stability advantage: Seven links of data closed loop transmission, each link is provided with data checking and repairing mechanism (such as point cloud hole repairing, data missing automatic compensation), which can ensure the stability and controllability of model quality, reduce the rework and remanufacture caused by data error, and reduce the risk of clinical diagnosis and treatment. The grid optimization algorithm reduces the number of triangular facets of the crown model, which not only improves the processing efficiency, but also ensures the feasibility and stability of the model processing.

[0093] The above application uses specific examples to describe the present application, which is only used to help understand the present application and does not limit the present application. For those skilled in the art to which the present application belongs, according to the idea of the present application, a number of simple deductions, deformations or substitutions can be made.

Claims

1. A high-precision, automated method for forming dental crowns, characterized in that: The method is as follows: Step 1: Oral cavity 3D data acquisition Use a medical dental scanner or 3D scanner to obtain complete three-dimensional scan data of the upper and lower jaws, ensuring that the scan range covers all teeth, gums and related soft tissues, and that the data format supports subsequent point cloud processing and model building; Step 2: Semantic segmentation of the entire mouth teeth A semantic segmentation method based on a deep learning model combined with a dental anatomical morphology feature database is adopted. The whole scan data is input, and the differences between teeth and soft tissues such as gums and mucosa in gray value, texture features and morphological structure are identified to achieve automatic separation of individual teeth in the whole mouth. An independent three-dimensional data model of a single tooth is established, which includes complete data of the crown, neck and exposed part of the root. Each tooth is matched with a unique tooth position number according to the International Dental Federation tooth position numbering system, and the number and the corresponding tooth data model are associated and stored. Step 3: Fine-tuning of upper and lower jaw point clouds Using the tooth position number obtained in step 2 as the core index, a registration depth model based on the maxillofacial dynamic map structure is used to extract 12-15 stable feature points of corresponding teeth in the upper and lower jaws. The feature point set is initially matched by the registration depth model. Then, the point cloud after initial transformation is input into the ICP algorithm. KD-Tree is used to accelerate the nearest neighbor search. Erroneous point pairs with a distance threshold greater than 0.05 mm are removed. The objective function is to minimize the sum of Euclidean distances between feature points. The iteration termination condition is set to the distance error change between two adjacent iterations being less than 0.001 mm. Combined with the constraint of the natural occlusal contact relationship of the upper and lower jaw teeth, the precise alignment of the upper and lower jaw point clouds is achieved. Step 4: Precise segmentation and trimming of the abutment tooth and associated gingiva Based on the tooth position numbering in step 2, the preliminary point cloud region of the abutment tooth to be restored is locked. Combined with the spatial coordinate system of the abutment tooth after fine registration in step 3, the preliminary spatial range of the abutment tooth is constructed by using the high point of the crown and the neck turning point of the abutment tooth as references. The two-step method of "geometric feature differentiation + point cloud purification" is adopted. First, a classifier is constructed by using the point cloud normal vector and curvature features to initially separate the abutment tooth and gingival point clouds. Then, the abutment tooth point cloud is subjected to radius filtering for noise reduction and greedy projection triangulation for hole repair. The gingival point cloud is subjected to a region growing algorithm based on neighborhood distance, retaining only the connecting gingival region with a distance ≤0.1 mm from the surface of the abutment tooth, forming a combined point cloud model of the abutment tooth and the associated gingiva. Step 5: Multi-tissue segmentation of the abutment tooth surrounding environment Using the combined point cloud model of the abutment tooth and associated gingiva obtained in step 4 as the origin, a local spatial coordinate system is established with the mesiodistal, labiobuccal-lingual-palatal, and occlusal-gingival directions as orthogonal positioning axes. The system is extended by 5 mm in each direction to form a cubic segmentation space. Based on the spatial position relationship and point cloud texture features, adjacent teeth are segmented using the Euclidean distance clustering algorithm. The opposing teeth are segmented based on the occlusal spatial relationship combined with the planar projection method and the occlusal surface texture features of the opposing teeth. The surrounding gingiva and alveolar bone are segmented by jointly filtering the texture feature threshold and the Z-axis coordinate range to obtain complete environmental data around the abutment tooth. Step 6: Secondary Precision Detection of the Neckline Using the combined point cloud model of the abutment tooth and associated gingiva from step 4 as the data source, the strategy of "coarse localization of depth model + fine detection of point cloud algorithm" is adopted. First, the combined point cloud data is input through the depth model to automatically identify the annular area where the cervical margin line is located and output the preliminary localization range within ±0.2 mm. Then, based on the coarse localization results, the normal vector of each point in the localization area is calculated by PCA algorithm, and candidate edge points with abrupt changes in normal vector direction are screened. The curve is reconstructed by RANSAC circle fitting algorithm and the curve smoothness is adjusted by active contour model to obtain a continuous and complete cervical margin line. Step 7: Crown formation and multi-dimensional refinement A two-step strategy of "deep model modeling + point cloud algorithm refinement" is adopted. The precise cervical margin parameters from step 6, the combined point cloud model of the abutment tooth and related gingiva from step 4, and the environmental data around the abutment tooth from step 5 are input. The oral anatomy and physiology feature library and clinical restoration standards are incorporated as constraints. First, the inner and outer crowns of the crown are constructed to form a complete basic crown model. Then, the point cloud algorithm is used to perform volume calibration, orientation correction, morphological fine-tuning, cervical margin fitting optimization, and undercut filling to output a 3D crown model that meets the clinical restoration requirements.

2. The high-precision, automated method for crown fabrication according to claim 1, characterized in that: Before scanning and collecting data, the soft tissues in the oral cavity are cleaned to remove food debris and excess saliva. The scanner resolution is set to no less than 50 micrometers. During the scanning process, a multi-angle superimposed scanning method is used to focus on scanning key areas such as the occlusal surface and proximal surface of the teeth to ensure that the data is free of holes and obvious noise. After the data is collected, the raw data is standardized and converted into PLY or STL format.

3. The high-precision, automated method for crown fabrication according to claim 1, characterized in that: The semantic segmentation method adopts an improved U-Net architecture and introduces an attention mechanism to enhance the ability to extract tooth edge features. The model training dataset contains more than 1,000 clinical scan data of different dental arch morphologies and oral conditions, covering multiple scenarios such as normal dental arch, crowded dental arch, and missing tooth restoration. After segmentation, the integrity of the three-dimensional data model of a single tooth is checked, and missing data areas are automatically identified and repaired. The tooth position number association adopts a dual verification mechanism, combining the spatial position and morphological features of the tooth to ensure the accuracy of the numbering.

4. The high-precision, automated method for crown fabrication according to claim 1, characterized in that: The registration depth model constructs a feature library using data from 400 clinical cases of non-malformed maxillofacial structures, automatically filters out interference from gingival soft tissue, introduces an adaptive weighting factor during the ICP algorithm iteration process, assigns higher weights to key cusp-fossa feature points on the occlusal surface, and improves the accuracy of occlusal relationship restoration. After registration is completed, the registration effect is verified by the occlusal contact detection algorithm. If occlusal misalignment exists, the registration is automatically iterated again.

5. The high-precision, automated method for crown fabrication according to claim 1, characterized in that: After the initial spatial range of the abutment teeth is calibrated, the integrity of the region is verified by point cloud density analysis. If there is missing data, the bounding box range is automatically expanded by 0.5-1 mm. During the geometric feature differentiation process, a dynamic threshold adjustment strategy is adopted to dynamically optimize the classifier parameters according to the curvature differences between teeth and gingiva at different tooth positions. During the point cloud purification stage, the threshold for the diameter of the repair hole in the abutment tooth point cloud is set to 0.1 mm to ensure the integrity of the tooth surface while preserving the original morphological features of the junction area between the gingiva and the abutment tooth.

6. The high-precision, automated method for crown fabrication according to claim 1, characterized in that: When establishing the local spatial coordinate system, orientation calibration is performed in conjunction with standard oral anatomical parameters to ensure that the positioning axis is consistent with the physiological long axis of the tooth. In the process of segmenting adjacent teeth, in addition to Euclidean distance and texture features, tooth morphology contour matching analysis is added to improve the accuracy of adjacent tooth identification. The segmentation of opposing teeth is assisted by occlusal trajectory simulation to ensure that the segmented area is completely matched with the occlusal contact range of the abutment teeth. The texture feature threshold of the segmentation of gingiva and alveolar bone is dynamically adjusted according to the gray-scale distribution of the scan data.

7. The high-precision, automated method for crown fabrication according to claim 1, characterized in that: The coarse localization of the deep model adopts a lightweight CNN architecture to reduce computation and improve localization speed. Before inputting the model, the combined point cloud data is downsampled to retain key geometric features while improving computational efficiency. In the fine detection stage, multi-scale normal estimation is introduced to improve the adaptability of the cervical margin line to different tooth positions. The RANSAC circle fitting algorithm iterates no less than 500 times. The curve smoothness parameter of the active contour model is dynamically adjusted according to the curvature change of the cervical margin line to ensure that the cervical margin line is continuous and conforms to the physiological structure.

8. The high-precision, automated method for crown fabrication according to claim 1, characterized in that: The oral anatomy and physiology feature library contains a template library of crown morphology for different age groups and tooth positions, as well as personalized occlusal curve parameters. It can automatically match the optimal template according to the patient's age and dentition status. The clinical restoration standard supports custom adjustment functions to meet the personalized needs of special cases. Occlusal simulation tests are added during the crown finishing process. The occlusal surface morphology is optimized through virtual occlusal contact analysis to ensure uniform occlusion. At the same time, a mesh optimization algorithm is used to reduce the number of triangular facets in the crown model, improving the model processing efficiency. The final output 3D crown model can be directly imported into the CAM processing system.

Images