Tooth edge line detection and undercut evaluation method and device based on generative point cloud network, and storage medium

Abstract

The invention provides a tooth marginal line detection and undercut evaluation method and system based on a generative point cloud network and a storage medium, and relates to the technical field of dental information digitization. The method comprises the following steps: importing a pre-collected single-tooth three-dimensional grid model and a pre-collected preparation body completion line region model, and performing point cloud generation prediction by utilizing a generative point cloud network to generate an edge line; extending a plurality of rays from the edge line, and screening out an optimal ray as an initial insertion axis; evaluating an undercut area formed by the initial insertion shaft, and calculating a corresponding undercut index; and evaluating the undercut index according to a set judgment rule, and outputting the optimal insertion shaft and the undercut index thereof. By screening out the optimal insertion shaft and undercut index, the tooth preparation process can be automated and accurate, the quality and efficiency of preparation body design are improved, personal errors are reduced, better matching of the prosthesis and teeth is ensured, and therefore the repair effect and the patient satisfaction degree are improved.

Description

Technical Field

[0001] This invention mainly relates to the field of dental information digitization technology, specifically to a method, device, and storage medium for tooth edge line detection and undercut evaluation based on generative point cloud networks. Background Technology

[0002] In the field of digital dental diagnosis and treatment, traditional methods rely on manual operation and experience-based judgment for tooth margin detection and undercut assessment, which suffers from problems such as cumbersome operation, strong dependence, and difficulty in standardization. While existing software can provide visual undercut indicators, it lacks quantitative indicators, limiting its objective assessment capabilities. Furthermore, mechanical or laser equipment has limitations in recognizing complex tooth surface details and is difficult to seamlessly integrate with CAD / CAM design processes. The current subjective nature of margin acquisition and insertion axis selection easily leads to recognition errors; undercut assessment results mainly rely on color indicators, lacking unified quantitative indicators, which is detrimental to establishing objective standards and quality control records.

[0003] Therefore, existing technologies have significant shortcomings in terms of accuracy, efficiency, and standardization, and there is an urgent need for a solution that integrates model recognition with traditional geometric algorithms. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method, device and storage medium for tooth edge line detection and undercut evaluation based on generative point cloud network, which addresses the shortcomings of the prior art.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A method for tooth edge line detection and undercut evaluation based on generative point cloud networks, comprising the following steps: Import the pre-acquired single tooth 3D mesh model and the pre-prepared body completion line region model, and use the pre-trained generative point cloud network to perform point cloud generation prediction on the single tooth 3D mesh model and the pre-prepared body completion line region model to obtain the edge line; Multiple rays extend from the points on the edge line, and multiple optimal rays are selected as multiple initial insertion axes, forming an inverted concave region based on the initial insertion axes. The undercut regions formed by each initial insertion axis are evaluated to obtain undercut indices corresponding to multiple initial insertion axes. According to the set judgment rules, the undercut indexes corresponding to multiple initial insertion axes are judged respectively, and multiple optimal insertion axes and their corresponding undercut indexes are obtained and output.

[0006] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: A tooth edge line detection and undercut evaluation device based on generative point cloud network, comprising: The network recognition module is used to import the pre-collected single tooth 3D mesh model and the pre-prepared body completion line region model, and use the pre-trained generative point cloud network to perform point cloud generation prediction on the single tooth 3D mesh model and the pre-prepared body completion line region model to obtain the edge line. The geometry generation module is used to extend multiple rays based on points on the edge line, and select multiple optimal rays as multiple initial insertion axes to form an inverted concave region based on the initial insertion axes. The index calculation module is used to evaluate the inverted concave region formed by each initial insertion axis and obtain the inverted concave index corresponding to multiple initial insertion axes. The index evaluation module is used to judge the inverted indices corresponding to multiple initial insertion axes according to the set judgment rules, and to obtain and output multiple optimal insertion axes and their corresponding inverted indices.

[0007] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: a computer-readable storage medium storing a computer program, which, when executed by a processor, implements a method for detecting tooth edge lines and evaluating undercuts using a generative point cloud network.

[0008] The beneficial effects of this invention are: by enhancing the diversity and quality of training data for generative point cloud networks, using advanced algorithms to optimize the ray selection process, refining the evaluation criteria for undercut areas, dynamically adjusting judgment rules to adapt to different clinical situations, and improving the automation and user interaction experience of the entire process, the accuracy, efficiency, and personalization of tooth preparation are improved, ensuring a better match between restorations and teeth. Attached Figure Description

[0009] Figure 1 A flowchart of the tooth edge line detection and undercut evaluation method provided in the embodiments of the present invention; Figure 2 This is a block diagram of the tooth edge detection and undercut assessment device provided in an embodiment of the present invention. Detailed Implementation

[0010] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0011] In digital dental diagnostic systems, traditional methods for detecting tooth margins and assessing undercuts primarily rely on manual operation and experience-based judgment by clinicians or technicians. Specifically, these methods typically involve the following steps: first, creating a positive dental model through impression taking; second, using a dental observation instrument along a pre-defined insertion path to manually determine the presence of undercuts; or, assessment directly within the patient's mouth or on the model, based on personal experience. However, these methods have significant limitations: the procedures are cumbersome, highly dependent on the operator's experience, and difficult to standardize and document, thus limiting their widespread clinical application and accuracy.

[0012] Furthermore, although existing mainstream restoration design software can provide color mapping or occlusion cues based on insertion paths on 3D models to help identify potential undercut areas, these software programs still require a significant amount of manual work or semi-automatic correction in depicting the finish line. Undercut cues typically remain only at the visualization level, offering very limited support for quantitative indicators (such as area, volume, depth, and continuity) and threshold traceability, which is detrimental to establishing objective standards and quality control records.

[0013] Regarding undercut detection, while existing mechanical or laser-based devices can pinpoint undercut locations under specific conditions, these devices perform poorly in environments with limited intraoral space, complex tooth surface details, and weak correlation with marginal lines. Furthermore, these devices typically cannot be directly quantified and recorded within digital workflows, making seamless integration with subsequent CAD / CAM design processes difficult.

[0014] In the process of obtaining edge lines, most procedures still require manual drawing or extensive interactive corrections. This is not only inefficient and lacks repeatability, but errors in the edge lines can also propagate to the selection of the insertion axis and the determination of undercuts, leading to systemic biases. The selection of the insertion axis is often based on experience or a single heuristic rule (e.g., the default bite direction), lacking joint optimization and constraint evaluation of both global and local factors. This can lead to misjudgments of undercuts or generate false and missed warnings.

[0015] Finally, current undercut assessment results are primarily represented by color-coded indicators, lacking standardized and traceable quantitative metrics (such as undercut area / proportion, maximum depth, longest continuous segment, clinical tolerance grading, etc.). This makes it difficult to establish objective standards and quality control records. These issues highlight the shortcomings of existing technologies in digital dental diagnostic systems, necessitating a more precise, efficient, and standardized solution.

[0016] like Figure 1As shown in the figure, the method for tooth edge line detection and undercut evaluation using generative point cloud networks provided by this invention includes the following steps: S1. Import the pre-collected single-tooth 3D mesh model and the pre-prepared body completion line region model. Use the pre-trained generative point cloud network to generate and predict the point cloud of the single-tooth 3D mesh model and the pre-prepared body completion line region model to obtain the edge line. S2. Extend multiple rays from the points on the edge line, and select multiple optimal rays as multiple initial insertion axes to form an inverted concave region based on the initial insertion axes. S3. Evaluate the undercut regions formed by each initial insertion axis to obtain undercut indices corresponding to multiple initial insertion axes. S4. According to the set judgment rules, the undercut indexes corresponding to multiple initial insertion axes are judged respectively, and multiple optimal insertion axes and their corresponding undercut indexes are obtained and output.

[0017] In the above embodiments, by combining a pre-acquired single-tooth 3D mesh model and a prepared tooth completion line region model, a generative point cloud network is used to generate and predict point clouds, obtaining edge lines and extending multiple rays to select the optimal ray as the initial insertion axis. Then, the undercut regions formed by these initial insertion axes are evaluated to obtain corresponding undercut indices. Finally, these indices are judged based on set judgment rules to identify the optimal insertion axis and its corresponding undercut indices. This can automatically evaluate and select the best prepared tooth design, improving the accuracy and efficiency of tooth restoration, reducing human error, and providing a scientific basis for clinical decision-making. This invention offers batch processing, configurable thresholds, quality control reports, and visual export capabilities, meeting the needs of large-scale clinical, technician, and teaching applications as well as regulatory compliance requirements.

[0018] In the field of dental prosthodontics, preparation refers to the shape of a tooth after preparation. It primarily serves to provide a suitable base for the installation of restorations (such as inlays, veneers, and crowns). The shape and size of the preparation depend on the type of restoration (such as inlays, veneers, and crowns) and the initial state of the tooth. The preparation finish line is the boundary between the prepared portion and the unprepared portion of the tooth, and it is a crucial point for the integration of the restoration with the natural tooth. By preparing the tooth, its morphology and function can be improved, creating favorable conditions for the installation of restorations.

[0019] Preferably, it also includes a training step for the generative point cloud network, comprising: An initial generative point cloud network is constructed, and the initial generative point cloud network is trained using a training point cloud dataset to generate a predicted point cloud dataset. The total loss is calculated by using a constructed total loss function on the predicted point cloud dataset. The process of constructing the total loss function includes: Based on the predicted point cloud dataset and the real point cloud dataset, a Chamfer distance loss function is constructed to quantify the geometric error between the detected tooth edge line and the real edge line. The Chamfer distance loss function is as follows: , in, Let Chamfer distance be the loss function. Let A be the total number of points in the predicted point cloud dataset A corresponding to the tooth edge line to be evaluated. This represents the total number of points in the real tooth edge point cloud corresponding to the real point cloud dataset B. a To predict point cloud data, b For real point cloud data, For predicting point cloud data a Compared with real point cloud data b The Euclidean distance between them For each real point cloud data in the real point cloud dataset B b To each predicted point cloud data in the predicted point cloud dataset A a The minimum Euclidean distance is used to measure the proximity of the real edge point to the detected edge line. For each prediction point cloud data in prediction point cloud dataset A a To each real point cloud data in real point cloud dataset B b The minimum Euclidean distance is used to measure how close the detected edge line point is to the real edge line; An information distance loss function is constructed based on the predicted point cloud dataset and the real point cloud dataset. The information distance loss function is as follows: , in, Let be the information distance loss function, log be the logarithmic function, and exp be the exponential function. d ( a b (For predicting point cloud data) a Compared with real point cloud data b The Euclidean distance between them; A total loss function is constructed based on the Chamfer distance loss function and the information distance loss function. The total loss function is as follows: , in, For the total loss function, and These are the weighting coefficients; The initial generative point cloud network is optimized by using the total loss to obtain the generative point cloud network.

[0020] Specifically, the initial generative point cloud model includes: an FPS layer (Farthest PointSampling) for downsampling, a Dynamic Graph Convolutional Neural Network (DGCNN) for feature extraction, a Transformer encoder for adaptive querying, a Transformer decoder for generating surrogate points, and a Folding Net for reconstructing multi-resolution point sets, which are divided into three levels: coarse, medium, and fine.

[0021] For the training and ensemble of the initial generative point cloud network, a training / test scale of 913 / 134 was adopted on the data side, with five-fold training and ensemble inference. After spline conversion, the median CD≈0.137mm and HD≈0.242mm were obtained. The uncertainty measure was the correlation analysis between the proportion of outliers removed and CD.

[0022] In the above embodiments, by constructing the Chamfer distance loss function and the information distance loss function, the difference between the predicted point cloud dataset and the real point cloud dataset can be measured more accurately, thereby optimizing the performance of the generative point cloud network.

[0023] Preferably, a pre-trained generative point cloud network is used to generate and predict point clouds for the single-tooth 3D mesh model and the pre-body completion line region model, resulting in edge lines, including: The initial point cloud dataset is obtained by predicting the construction of a single-tooth 3D mesh model and a pre-trained body completion line region model through a pre-trained generative point cloud network. The average distance expression is used to calculate the edge line point cloud data in the initial point cloud dataset, resulting in the first... k The average distance of the point cloud data of the edge lines of each tooth sample is expressed as follows: , in, For the first k The average distance between the point cloud data of the edge lines of each tooth sample is used; the smaller the value, the higher the degree of fit between the detection result and the actual edge line. k This is the serial number of the tooth sample. For the first k Predicted edge points in individual tooth samples With real edge points The Euclidean distance between them; The local density of multiple point cloud data is obtained by calculating multiple average distances using a local density expression. The local density expression is as follows: , in, For local density, the closer the value is to 1, the higher the predicted edge point density. The closer the local neighborhood distribution matches the local structure of the real edge line, the better. To predict edge line points of k The set of n nearest neighbors, where exp is an exponential function; Point cloud data with local densities less than or equal to a preset outlier threshold are removed from multiple point cloud data sets to obtain multiple candidate points; Multiple candidate points are calculated using the principal component analysis algorithm to obtain multiple initial edge points; Multiple initial edge points are sorted using a sorting algorithm to obtain multiple edge points; Multiple edge points are fitted into an edge line using a spline fitting algorithm.

[0024] Specifically, principal component analysis (PCA) is used to analyze multiple candidate points to obtain principal component axes, and these candidate points are then grouped together. Mapping these points onto the principal component axes yields the mapping candidate points. And calculate candidate points and mapping candidate points The Euclidean distance between them will be used to eliminate candidate points that are greater than the second threshold (i.e., if...). If the outlier is confirmed, multiple initial edge points are obtained; It should be understood that Principal Component Analysis (PCA) is a commonly used data dimensionality reduction technique that maps the original high-dimensional data to a low-dimensional space through linear transformation, thereby removing noise and redundant information while retaining the main features of the data.

[0025] The Traveling Salesman Problem (TSP) is a classic optimization problem where the goal is to find the shortest path through all cities. Approximate Traveling Salesman Problem (TSP) sorting is a heuristic algorithm used to solve a variant of the TSP, specifically sorting points in point cloud processing.

[0026] B-spline fitting is a technique commonly used in computer graphics, CAD (computer-aided design), CAM (computer-aided manufacturing), and data fitting. B-spline fitting is an iterative process that requires multiple adjustments to control points and node vectors to achieve the best fit.

[0027] In the above embodiments, by calculating the average distance and local density, and removing outliers, noise and inaccurate data points can be effectively removed, thereby improving the overall quality of point cloud data. Principal component analysis (PCA) and ranking algorithms can effectively extract key features from point cloud data, such as edge points, which are crucial for subsequent analysis and processing. Using a spline fitting algorithm to fit edge points into edge lines can generate smooth and accurate edge representations, which is very helpful for the reconstruction and analysis of 3D models.

[0028] Preferably, the pre-body completion line region model includes the pre-body completion line; Multiple rays extend from points on the edge line, including: The completion line of the pre-body is calculated by principal component analysis algorithm to obtain the outward reference axis. A reference hemisphere is generated based on the outward reference axis. The reference hemisphere is sampled by Fibonacci sphere sampling method to obtain multiple reference directions. The edge line is uniformly sampled to obtain multiple new edge points. Based on these new edge points, multiple initial rays are generated according to multiple reference directions. The offset of each initial ray is calculated using an offset calculation expression, which is: , in, To predict edge line points The corresponding offset, 3D model of tooth structure at point The unit of precision for the coordinates at that location. This is the global minimum error tolerance threshold. This is a truncation function that restricts the input value to between 0.25 and 1. For the tooth surface at points Normal deviation at the location; global minimum error tolerance threshold In dental settings, 0.02 is typically used.

[0029] Based on offset By offsetting the corresponding initial ray, multiple rays are obtained, which are represented as follows: , in, For the tooth edge detection point after offset adjustment For normal fine-tuning step size, N i For the 3D mesh model of the tooth at the initial detection point The unit normal vector at that location.

[0030] Specifically, the candidate direction set generation includes: performing principal component analysis (PCA) on each point along the line using a preparatory volume, and providing outward reference axes for the principal directions. On its outward reference axis Sampled using Fibonacci spheres on the hemisphere There are 1 direction, and the reference direction set is 1. Furthermore, multiple reference directions are confined within the cone, with the constraint that the cone angle is +70° or -70°.

[0031] Dynamic ray origin and self-collision exclusion include: uniform resampling of edge lines. From these points, we obtain a new set of edge points. The outward reference axis is expressed by the direction vector expression. and reference direction Calculations are performed to obtain the ray direction vector. According to the ray direction vector Rays are emitted from the outside in (i.e., bidirectional testing with a small number of sampling points, selecting the side with fewer hits as the outside direction (to avoid reverse axis)), resulting in multiple initial rays. The direction vector expression is as follows: , in, sign ( ) is a symbolic function; To suppress grazing (i.e., rays that are almost parallel to the surface of a single-tooth 3D mesh model) and numerical errors, a dynamic offset is used for each ray, and the offset is related to the ULP (offset per unit length) and... (ray direction vector) normal vector of the new edge point The absolute value of the dot product is related to the expression for the ray direction vector calculated using the direction sign. normal vector of the new edge point The dot product is calculated, and the expression for calculating the direction sign is: , in, This is the direction sign parameter.

[0032] It should be understood that the Fibonacci sphere sampling method is a method of uniformly distributing points on the surface of a sphere, which can ensure that the sampling points are evenly distributed on the surface of the sphere.

[0033] In the above embodiments, the outward reference axis is calculated using principal component analysis, which can more accurately determine the direction of the edge line, thereby improving the accuracy of edge detection. A reference hemisphere is generated based on the outward reference axis, and multiple reference directions are obtained by sampling the reference hemisphere using the Fibonacci sphere sampling method, increasing the flexibility and robustness of edge detection. Offset adjustments to the rays ensure more precise and controllable ray generation, optimizing the direction and position of the rays and improving their accuracy and effectiveness.

[0034] Preferably, multiple optimal rays are selected as multiple initial insertion axes, including: Identify the first intersection point of multiple rays with the single-tooth 3D mesh model, calculate the distance between each ray emission point and the corresponding first intersection point, and obtain multiple first-collision distances; The optimal ray is obtained by discarding rays whose first-collision distances are less than the constructed self-collision hit value.

[0035] Specifically, the first collision result is obtained by using triangular mesh rays for intersection, and the line strip region patches are ignored using a pre-set preparatory body; the offset of multiple initial rays is... The product is calculated with the preset local roughness tol_i (i.e. tol_i), to obtain the self-collision hit value, where the preset local roughness tol_i is adaptively selected by the median of the local roughness, and the median of the local roughness is between 1 and 3; multiple rays whose first collision distance is less than the self-collision hit value are eliminated.

[0036] It should be understood that the ray emission point is the new edge point obtained by uniformly sampling the edge line. The culling distance is the distance from the ray emission point (e.g., Pi) to the first collision point where the ray intersects the triangular mesh (i.e., the first collision point is the first intersection point of the ray with the triangular mesh starting from the ray emission point). This distance is used to determine whether the collision point obtained by the ray intersection should be considered a valid collision or should be ignored as a self-collision.

[0037] In the above embodiments, by identifying the first intersection point of multiple rays with the single-tooth 3D mesh model and calculating the distance between the ray emission point and the first intersection point, multiple first-collision distances are obtained. Then, rays with first-collision distances less than the self-collision hit value are eliminated, thereby obtaining the optimal ray. This method can accurately identify and select the best ray intersecting with the 3D model, which helps to improve the accuracy and efficiency of 3D model analysis and processing.

[0038] Preferably, the undercut regions formed by multiple initial insertion axes are evaluated to obtain undercut indices corresponding to the multiple initial insertion axes, including: Multiple effective depths are obtained by calculating multiple initial insertion axes using an effective depth expression, which is: , in, For the first i The effective depth of the ray For the first i The first-collision distance of a ray. d Let be the direction vector of the ray. This is the offset. It is the dot product of the tooth surface normal vector and the ideal direction; Multiple initial insertion axes with effective depths greater than or equal to the depth threshold are marked, and the inverted concave discrete set is obtained based on the index of the marked initial insertion axes. The initial insertion axes of the index in the inverted discrete set are analyzed and processed to obtain the inverted indices corresponding to multiple initial insertion axes.

[0039] Specifically, if the effective depth Greater than or equal to the depth threshold If it is positive, mark it as 1; otherwise, mark it as 0; thus obtaining the inverted concave discrete set. .

[0040] , in, The Boolean label for the i-th ray (indicating whether an indentation exists). The depth threshold (the clinical depth baseline, i.e. the minimum depth threshold at which an indentation is considered meaningful, given as 0.30 mm) In the above embodiments, the effective depth of multiple initial insertion axes is calculated, and these axes are marked based on the comparison results of the effective depth and the depth threshold. Then, the marked axes are subjected to undercut index analysis to obtain the undercut index corresponding to multiple initial insertion axes. This allows for accurate evaluation and selection of the optimal insertion axis, ensuring that the ideal shape and size of the prepared body are achieved during the tooth preparation process, thereby improving the accuracy and success rate of tooth restoration.

[0041] Preferably, the initial insertion axis of the index in the inverted concave discrete set is subjected to index analysis processing to obtain multiple inverted concave indices corresponding to the initial insertion axis, including: A completion line zone is set in the preparatory body completion line region model. The area of ​​the completion line zone is calculated using the completion line zone area calculation expression to obtain the completion line zone area. The completion line zone area calculation expression is as follows: , in, To complete the line band area, f To complete a triangular facet of the line zone domain, To complete the set of all triangular faces in the line zone region, Area( ) is the area function of the graph; The first triangular facet satisfying the inverted concave discrete set in the completed lineband domain is filtered using an image filtering expression to obtain multiple first-collision triangular faces. The image filtering expression is as follows: , in, The set of multiple first-collision triangles (i.e., all first-collision triangles of rays with a Boolean label of 1), unique( ) is the deduplication function, first_tri( ) is the first intersection identification function; The areas of multiple initial-collision triangular faces are calculated using the triangular face area calculation expression to obtain multiple concave areas. The triangular face area calculation expression is as follows: , Among them, A( ) represents the inverted concave area (i.e., the area of ​​the clinically effective first-touch triangle after deduplication and summation); The inverted concave area index is obtained by calculating the area of ​​the completed line zone and the inverted concave area using the inverted concave area index calculation expression: , Among them, Ratio( () represents the area of ​​the concave depression; The effective depth of the initial insertion axis of the index in the inverted concave discrete set is calculated using the maximum depth calculation expression to obtain the maximum effective depth index. The maximum depth calculation expression is as follows: , in, The maximum effective depth indicator It is an inverted concave discrete set; The effective depth of the initial insertion axis of the index in the inverted concave discrete set is calculated using the quantile depth calculation expression to obtain the quantile effective depth index. The quantile depth calculation expression is as follows: , in, The 95th percentile effective depth indicator It is a quantile function; The inverted concave bottleneck index is obtained by calculating the initial insertion axis of the index in the inverted concave discrete set using the bottleneck calculation expression. The bottleneck calculation expression is as follows: , in, The bottleneck indicator is a concave shape. The proportion of the initial insertion axis of the index on the edge line in the inverted concave discrete set is calculated using the inverted concave segment proportion calculation expression to obtain the proportion of the longest continuous inverted concave segment. The inverted concave segment proportion calculation expression is as follows: , Among them, LRF The percentage of the longest continuous concave segment. The longest run length of the initial insertion axis for a continuous index in an inverted discrete set (i.e., the length of the initial insertion axis for a continuous index in an inverted discrete set). Considering it as a closed-loop sequence, let the longest running length of consecutive 1s be . ), M This is the total length; The undercut depth is calculated by using the undercut depth calculation expression to calculate the initial insertion axis and the single-tooth 3D mesh model indexed in the undercut discrete set. The undercut depth calculation expression is as follows: , in, The depth of the concave section. Along the reference direction u The set of vertices in an orthographic projection. Along the reference direction u The set of boundary polygons for orthogonal projection. To obtain the convex hull of the outer envelope, dist(q) e) Point cloud data and point cloud data The Euclidean distance between them.

[0042] Specifically, based on the seed vertices of the completed line projection, the mesh adjacency graph is expanded to obtain a set of strip / ROI patches (used for self-collision exclusion and...). (Statistics), the average step length of the edge lines on both sides of the preparatory body completion line is The region is set to complete the line band domain ROI.

[0043] The concave depth is obtained by calculating the two-dimensional projection along the direction. Orthographic projection: , , in, For along direction u The set of vertices in an orthographic projection. For along direction u The set of boundary polygons for orthogonal projection. To obtain the convex hull of the outer envelope; Since the aperture is defined as the average of the minimum distances from each point in the inner circle to the outer envelope boundary (this invention uses the average minimum distance aperture and takes the maximum within the 30° cone as the final value).

[0044] It should be understood that the first intersection point of multiple optimal rays with the single-tooth 3D mesh model is identified, the distance between the emission point of each optimal ray and the corresponding first intersection point is calculated, and the first-hit distance of multiple optimal rays is obtained. Alternatively, the first-hit distance of the optimal ray can be selected from multiple first-hit distances according to the ray index.

[0045] In the above embodiments, the prepared restoration finish line region model is comprehensively evaluated by calculating multiple dimensions of indicators, including undercut area, maximum effective depth, quantile effective depth, undercut bottleneck, proportion of longest continuous undercut segment, and undercut depth. The calculation of these indicators helps to identify and quantify undercut areas in the model, thereby guiding optimized design, ensuring the model meets specific geometric and functional requirements, improving the processing accuracy and restorative effect of the prepared restoration, and ultimately enhancing the fit between the restoration and the tooth and its chewing function.

[0046] Preferably, the undercut indicators corresponding to multiple initial insertion axes are determined according to the set judgment rules, including: The smallest indented area is selected from multiple indented areas, and the initial insertion axis direction corresponding to the smallest indented area is identified. The feasibility of the ray direction is determined based on the release judgment rule, which is as follows: , in, To minimize the concave area, The absolute area threshold, To complete the line band area, Where S is the relative area threshold, and S is the low-area feasible cone. OR logic symbol; When the clinical hit rate is greater than zero, the rejection decision rule is used to determine whether the radiation direction should be rejected. The rejection decision rule is as follows: , or, The effective depth index is the 95th percentile of the smallest concave area. The maximum depth threshold; , Where E is the volume index. The maximum volume threshold; If the processing result of the release and rejection rules is release and no rejection is triggered, or the clinical hit rate is zero, the corresponding initial insertion axis is set to perfect; otherwise, it is set to allow.

[0047] Specifically, Set to 0.1mm², Set to 0.5%, where S represents the existence of a feasible cone with a base area (adaptive connectivity scale). Set to 0.8mm. Set to 0.05mm³.

[0048] A clinical hit rate of zero indicates a set of multiple first-hit triangles. It is an empty set, meaning that all rays with a Boolean flag of 1 do not collide with any of the triangular faces of the single-tooth 3D mesh model.

[0049] In the above embodiments, by selecting the smallest undercut area from multiple undercut areas and identifying the corresponding initial insertion axis direction, and then evaluating the feasibility of the ray direction based on a series of approval and rejection rules, it is helpful to accurately select the optimal insertion axis direction, ensure the optimal prepared shape in clinical applications, improve the quality and success rate of tooth restoration, and at the same time reduce human error and improve work efficiency through automated evaluation process.

[0050] Preferably, to accommodate the undercut requirements of different restorations, a configurable template is provided. The template version number and threshold group are recorded in the results and can be adjusted on the product side according to the department's requirements; currently, the LRF upper limit is recorded without rejection by default, but can be switched to a hard threshold.

[0051] Table 1 presents experimental data on the differences between the clinical tolerance template and repair type. As shown in Table 1, during the deployment phase, the denoising ratio was used as an uncertainty measure, and its correlation with CD on the test set showed an accuracy of approximately 80–90%, with a sensitivity / precision ratio of approximately 90%. The confidence prediction ratio of the ensemble model was higher than that of the single-fold model (88.81%). This metric can provide self-verification hints for scenarios without true values.

[0052] Table 1 like Figure 2 As shown, an embodiment of the present invention provides a tooth edge line detection and undercut evaluation device based on a generative point cloud network, comprising: The network recognition module is used to import the pre-collected single tooth 3D mesh model and the pre-prepared body completion line region model, and use the pre-trained generative point cloud network to perform point cloud generation prediction on the single tooth 3D mesh model and the pre-prepared body completion line region model to obtain the edge line. The geometry generation module is used to extend multiple rays based on points on the edge line, and select multiple optimal rays as multiple initial insertion axes to form an inverted concave region based on the initial insertion axes. The index calculation module is used to evaluate the inverted concave region formed by each initial insertion axis and obtain the inverted concave index corresponding to multiple initial insertion axes. The index evaluation module is used to judge the inverted indices corresponding to multiple initial insertion axes according to the set judgment rules, and to obtain and output multiple optimal insertion axes and their corresponding inverted indices.

[0053] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for detecting tooth edge lines and evaluating undercuts using a generative point cloud network as described above.

[0054] This invention is applied to the digital restorative design process in dental clinics or hospitals, helping dentists or technicians to quickly and accurately complete the detection of tooth margins, optimization of insertion paths, and assessment of undercuts. It is applicable to the design of various dental restorations, such as inlays, veneers, and full crowns, improving the fit and clinical outcomes of the restorations.

[0055] This invention establishes a closed loop between generative point cloud edge detection and interpretable computational geometry determination: from edge line to insertion axis to inverted concave quantization and grading, it balances accuracy, robustness, and traceability, with the following effects: In terms of accuracy and consistency, the edge lines are directly generated by a generative model and then denoised and splined using local density and principal component analysis (PCA). The median error between the edge lines and the true values ​​in publicly available data is approximately CD ≈ 0.137 mm and HD ≈ 0.242 mm (five-fold and ensemble settings), providing a stable prior for subsequent geometric determination. This quality evidence can be independently verified. Based on stable edge lines, the clinically effective area is used... Candidate direction scanning for the target reduces false alarms / missed detections caused by inaccurate axial selection, making the judgment closer to clinical intuition. To achieve tolerance for real mesh defects, dynamic ray initiation (adaptive magnitude and grazing suppression) and strip self-collision filtering are employed to suppress false hits caused by edge roughness, normal flipping, and micro-holes; the per-ray tolerance coefficient is adaptively determined by local roughness, ranging from [1, 3]. The concave region is calculated using the union area of ​​clinically effective first-collision triangular faces. With completed line band domain The proportion is measured, the caliber is fixed, and it can be recalculated.

[0056] Simultaneously outputting a deep family of values ​​on an interpretable quantization closed loop ( ), the percentage of the longest continuous concave section along the edge line (LRF), and the strict minimum bottleneck thickness. Indicators such as the official caliber depth (the average shortest distance from the outer envelope to the inner circle under 2D projection, taking the maximum within the 30° cone) support clear intraoperative / postoperative review and liability boundaries. Decision-making adopts a three-option release and two-option rejection system, along with a feasible cone connectivity rule, ensuring that the judgment logic is traceable and externally verifiable.

[0057] In terms of templates and engineering implementation, the parameter system is calibrated to the millimeter and supports loading tolerance templates according to repair type (Inlay / Onlay / Full Crown); thresholds and versions are recorded in the output debug_info, which facilitates auditing and consistency control across machines / data domains.

[0058] In terms of human-machine collaboration and incremental recalculation, after slight editing of the edge line or axis, the process supports incremental recalculation and updates each indicator and hit_tris (the set of clinically effective triangular faces), which facilitates the detection of closed-loop operations from correction to review.

[0059] The aforementioned tooth edge detection and undercut evaluation device and storage medium based on generative point cloud networks can be found in the above-described implementation details and beneficial effects of the generative point cloud network-based tooth edge detection and undercut evaluation method, which will not be repeated here.

[0060] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0061] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for tooth edge line detection and undercut evaluation based on generative point cloud networks, characterized in that, Includes the following steps: Import the pre-acquired single tooth 3D mesh model and the pre-prepared body completion line region model, and use the pre-trained generative point cloud network to perform point cloud generation prediction on the single tooth 3D mesh model and the pre-prepared body completion line region model to obtain the edge line; Multiple rays extend from the points on the edge line, and multiple optimal rays are selected as multiple initial insertion axes, forming an inverted concave region based on the initial insertion axes. The undercut regions formed by each initial insertion axis are evaluated to obtain undercut indices corresponding to multiple initial insertion axes. According to the set judgment rules, the undercut indexes corresponding to multiple initial insertion axes are judged respectively, and multiple optimal insertion axes and their corresponding undercut indexes are obtained and output.

2. The method for detecting tooth margins and evaluating undercuts according to claim 1, characterized in that, It also includes the training steps for generative point cloud networks, specifically: An initial generative point cloud network is constructed, and the initial generative point cloud network is trained using a training point cloud dataset to generate a predicted point cloud dataset. The total loss is calculated by using a constructed total loss function on the predicted point cloud dataset. The process of constructing the total loss function includes: Based on the predicted point cloud dataset and the real point cloud dataset, a Chamfer distance loss function is constructed to quantify the geometric error between the detected tooth edge line and the real edge line. The Chamfer distance loss function is as follows: ， in, Let Chamfer distance be the loss function. Let A be the total number of points in the predicted point cloud dataset A corresponding to the tooth edge line to be evaluated. This represents the total number of points in the real tooth edge point cloud corresponding to the real point cloud dataset B. a To predict point cloud data, b For real point cloud data, For predicting point cloud data a Compared with real point cloud data b The Euclidean distance between them For each real point cloud data in the real point cloud dataset B b To each predicted point cloud data in the predicted point cloud dataset A a The minimum Euclidean distance is used to measure the proximity of the real edge point to the detected edge line. For each prediction point cloud data in prediction point cloud dataset A a To each real point cloud data in real point cloud dataset B b The minimum Euclidean distance is used to measure how close the detected edge line point is to the real edge line; An information distance loss function is constructed based on the predicted point cloud dataset and the real point cloud dataset. The information distance loss function is as follows: ， in, Let be the information distance loss function, log be the logarithmic function, and exp be the exponential function. d ( a b (For predicting point cloud data) a Compared with real point cloud data b The Euclidean distance between them; A total loss function is constructed based on the Chamfer distance loss function and the information distance loss function. The total loss function is as follows: ， in, For the total loss function, and These are the weighting coefficients; The initial generative point cloud network is optimized by using the total loss to obtain the generative point cloud network.

3. The method for detecting tooth margins and evaluating undercuts according to claim 1, characterized in that, A pre-trained generative point cloud network is used to generate and predict point clouds for a single-tooth 3D mesh model and a pre-formation body completion line region model, resulting in edge lines, including: The single tooth 3D mesh model and the pre-prepared body completion line region model are input into a pre-trained generative point cloud network. The generative point cloud network then processes the edge line features for recognition and generation to obtain the initial point cloud dataset corresponding to the tooth edge line. The average distance expression is used to calculate the edge line point cloud data in the initial point cloud dataset, resulting in the first... k The average distance of the point cloud data of the edge lines of each tooth sample is expressed as follows: ， in, For the first k The average distance of the point cloud data of the edge lines of each tooth sample is used; the smaller the value, the higher the fit between the detection result and the actual edge line. k This is the serial number of the tooth sample. For the first k Predicted edge points in individual tooth samples With real edge points The Euclidean distance between them; The local density of multiple point cloud data is obtained by calculating multiple average distances using a local density expression. The local density expression is as follows: ， in, For local density, the closer the value is to 1, the higher the predicted edge point density. The closer the local neighborhood distribution matches the local structure of the real edge line, the better. To predict edge line points of k The set of n nearest neighbors, where exp is an exponential function; Point cloud data with local densities less than or equal to a preset outlier threshold are removed from multiple point cloud data sets to obtain multiple candidate points; Multiple candidate points are calculated using the principal component analysis algorithm to obtain multiple initial edge points; Multiple initial edge points are sorted using a sorting algorithm to obtain multiple edge points; Multiple edge points are fitted into an edge line using a spline fitting algorithm.

4. The method for detecting tooth margins and evaluating undercuts according to claim 1, characterized in that, The preparatory body completion line region model includes the preparatory body completion line; Multiple rays extend from points on the edge line, including: The completion line of the pre-body is calculated by principal component analysis algorithm to obtain the outward reference axis. A reference hemisphere is generated based on the outward reference axis. The reference hemisphere is sampled by Fibonacci sphere sampling method to obtain multiple reference directions. The edge line is uniformly sampled to obtain multiple new edge points. Based on these new edge points, multiple initial rays are generated according to multiple reference directions. The offset of each initial ray is calculated using an offset calculation expression, which is: ， in, To predict edge line points The corresponding offset, 3D model of tooth structure at point The unit of precision for the coordinates at that location. This is the global minimum error tolerance threshold. This is a truncation function that restricts the input value to between 0.25 and 1. For the tooth surface at points Normal deviation at the location; Based on offset By offsetting the corresponding initial ray, multiple rays are obtained, which are represented as follows: ， in, These are the detection points for the tooth edge line after offset adjustment. For normal fine-tuning step size, N i For the 3D mesh model of the tooth at the initial detection point The unit normal vector at that location.

5. The method for detecting tooth margins and evaluating undercuts according to claim 1, characterized in that, Multiple optimal rays were selected as multiple initial insertion axes, including: Identify the first intersection point of multiple rays with the single-tooth 3D mesh model, calculate the distance between each ray emission point and the corresponding first intersection point, and obtain multiple first-collision distances; The optimal ray is obtained by discarding rays whose first-collision distances are less than the constructed self-collision hit value.

6. The method for detecting tooth margins and evaluating undercuts according to claim 1, characterized in that, The undercut regions formed by multiple initial insertion axes were evaluated to obtain undercut indices corresponding to the multiple initial insertion axes, including: Multiple effective depths are obtained by calculating multiple initial insertion axes using an effective depth expression, which is: ， in, For the first i The effective depth of the ray For the first i The first-collision distance of a ray. d Let be the direction vector of the ray. This is the offset. It is the dot product of the tooth surface normal vector and the ideal direction; Multiple initial insertion axes with effective depths greater than or equal to the depth threshold are marked, and the inverted concave discrete set is obtained based on the index of the marked initial insertion axes. The initial insertion axes of the index in the inverted discrete set are analyzed and processed to obtain the inverted indices corresponding to multiple initial insertion axes.

7. The method for detecting tooth margins and evaluating undercuts according to claim 6, characterized in that, Indices analysis is performed on the initial insertion axes of the index in the inverted concave discrete set to obtain multiple inverted concave indices corresponding to the initial insertion axes, including: A completion line zone is set in the preparatory body completion line region model. The area of ​​the completion line zone is calculated using the completion line zone area calculation expression to obtain the completion line zone area. The completion line zone area calculation expression is as follows: ， in, To complete the line band area, f To complete a triangular facet of the line zone domain, To complete the set of all triangular faces in the line zone region, Area( ) is the area function of the graph; The first triangular facet satisfying the inverted concave discrete set in the completed lineband domain is filtered using an image filtering expression to obtain multiple first-collision triangular faces. The image filtering expression is as follows: ， in, A set of multiple first-collision triangles, unique( ) is the deduplication function, first_tri( ) is the first intersection identification function; The areas of multiple initial-collision triangular faces are calculated using the triangular face area calculation expression to obtain multiple concave areas. The triangular face area calculation expression is as follows: ， Among them, A( () represents the area of ​​the concave section; The inverted concave area index is obtained by calculating the area of ​​the completed line zone and the inverted concave area using the inverted concave area index calculation expression: ， Among them, Ratio( () represents the area of ​​the concave depression; The effective depth of the initial insertion axis of the index in the inverted concave discrete set is calculated using the maximum depth calculation expression to obtain the maximum effective depth index. The maximum depth calculation expression is as follows: ， in, The maximum effective depth indicator It is an inverted concave discrete set; The effective depth of the initial insertion axis of the index in the inverted concave discrete set is calculated using the quantile depth calculation expression to obtain the quantile effective depth index. The quantile depth calculation expression is as follows: ， in, The 95th percentile effective depth indicator It is a quantile function; The inverted concave bottleneck index is obtained by calculating the initial insertion axis of the index in the inverted concave discrete set using the bottleneck calculation expression. The bottleneck calculation expression is as follows: ， in, The bottleneck indicator is a concave shape. The proportion of the initial insertion axis of the index on the edge line in the inverted concave discrete set is calculated using the inverted concave segment proportion calculation expression to obtain the proportion of the longest continuous inverted concave segment. The inverted concave segment proportion calculation expression is as follows: ， Among them, LRF The percentage of the longest continuous concave segment. The longest run length of the initial insertion axis for a continuous index in an inverted discrete set. M This is the total length; The undercut depth is calculated by using the undercut depth calculation expression to calculate the initial insertion axis and the single-tooth 3D mesh model indexed in the undercut discrete set. The undercut depth calculation expression is as follows: ， in, The depth of the concave section. Along the reference direction u The set of vertices in an orthographic projection. Along the reference direction u The set of boundary polygons for orthogonal projection. To obtain the convex hull of the outer envelope, dist(q) e) Point cloud data and point cloud data The Euclidean distance between them.

8. The method for detecting tooth margins and evaluating undercuts according to claim 7, characterized in that, According to the set judgment rules, the undercut indicators corresponding to multiple initial insertion axes are judged respectively, including: The smallest indented area is selected from multiple indented areas, and the initial insertion axis direction corresponding to the smallest indented area is identified. The feasibility of the ray direction is determined based on the release judgment rule, which is as follows: ， in, To minimize the concave area, The absolute area threshold, To complete the line band area, Where S is the relative area threshold, and S is the low-area feasible cone. OR logic symbol; When the clinical hit rate is greater than zero, the rejection decision rule is used to determine whether the radiation direction should be rejected. The rejection decision rule is as follows: ， or, The effective depth index is the 95th percentile of the smallest concave area. The maximum depth threshold; ， Where E is the volume index. The maximum volume threshold; If the processing result of the approval and rejection rules is approval without triggering rejection, or if the clinical hit rate is zero, the corresponding initial insertion axis is set to perfect; otherwise, it is set to allow. Here, a clinical hit rate of zero represents a set of multiple first-hit triangles. It is an empty set.

9. A device for detecting tooth edge lines and evaluating undercuts based on generative point cloud networks, characterized in that, include: The network recognition module is used to import the pre-collected single tooth 3D mesh model and the pre-prepared body completion line region model, and use the pre-trained generative point cloud network to perform point cloud generation prediction on the single tooth 3D mesh model and the pre-prepared body completion line region model to obtain the edge line. The geometry generation module is used to extend multiple rays based on points on the edge line, and select multiple optimal rays as multiple initial insertion axes to form an inverted concave region based on the initial insertion axes. The index calculation module is used to evaluate the inverted concave region formed by each initial insertion axis and obtain the inverted concave index corresponding to multiple initial insertion axes. The index evaluation module is used to judge the inverted indices corresponding to multiple initial insertion axes according to the set judgment rules, and to obtain and output multiple optimal insertion axes and their corresponding inverted indices.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the tooth edge line detection and undercut evaluation method based on generative point cloud network as described in any one of claims 1 to 8.

Images