A method for evaluating the effective bonding area of ​​proximal surfaces of bridge abutment teeth in Maryland

Abstract

This invention discloses a method for evaluating the effective bonding area of ​​proximal surfaces of bridge abutments in Maryland. The method involves acquiring a three-dimensional digital model of the oral cavity containing the target abutment tooth, standardizing the model to construct a standardized triangular mesh, extracting sampling point clouds, and establishing a correspondence between sampling points and mesh patches. Tooth position segmentation is performed on the sampling point cloud to determine the target abutment tooth and its adjacent teeth. Integrity detection is performed on the proximal surface region of the target abutment tooth, and data completion is performed when missing or distorted data is detected. Interactive three-dimensional segmentation is performed based on the surface data of the target abutment tooth to obtain the proximal surface segmentation results. Constraints are then applied using enamel boundaries, gingival margin safety distances, incisal or occlusal boundaries, and undercut exclusion conditions to obtain the effective bonding area. The effective bonding area is mapped to the standardized triangular mesh, and the three-dimensional surface areas of the corresponding mesh patches are accumulated to output the effective bonding area evaluation result.

Description

Technical Field

[0001] This invention relates to the field of digital dental restoration and computer-aided design technology, specifically to a method for evaluating the effective bonding area of ​​proximal surfaces of bridge abutment teeth in Maryland. Background Technology

[0002] With the development of digital dental restoration technology, data analysis methods based on intraoral scanning, cone-beam computed tomography (CBCT), and 3D modeling have been gradually applied to tooth morphology observation, restoration design, and preoperative assessment. Compared to traditional methods relying on plaster models, clinical visual inspection, or two-dimensional images, 3D digital models can more completely reflect the spatial morphology of the tooth surface, providing conditions for the identification and quantitative analysis of local areas. In applications involving adhesive restoration design, the usable area and size of local regions on the abutment tooth surface directly affect preoperative assessment and subsequent restoration plan development. Therefore, how to objectively evaluate relevant areas based on 3D digital models has become a practical requirement in clinical digital analysis.

[0003] However, current assessment methods for localized areas of the tooth still largely rely on physician experience, manual selection, or auxiliary measurements using general-purpose 3D software. Because the tooth surface is a continuously changing and complex curved surface, there is usually no clear, fixed geometric boundary between the localized area and the surrounding area. Operators' understanding of boundary locations, range selection, and measurement paths can easily vary, leading to inconsistent assessment results across different personnel and at different times. At the same time, existing general segmentation or modeling methods often focus on overall contour extraction, general crown reconstruction, or routine surface treatment, lacking specific constraints corresponding to actual restorative conditions. This makes it difficult to simultaneously consider anatomical morphological relationships, adjacent spatial relationships, and clinical application requirements during the identification of localized areas, thus resulting in discrepancies between assessment results and actual restorative needs.

[0004] Furthermore, oral cavity 3D data is easily affected by factors such as occlusion, reflection, scanning angle, and local defects during acquisition, leading to sparse sampling, missing pores, surface artifacts, or local distortion in the target area. When the raw data has the above problems, existing methods usually lack effective processing mechanisms for local area evaluation, which can easily amplify boundary identification errors and area measurement errors, affecting the reliability of subsequent analysis results.

[0005] In view of this, the present invention proposes a method for evaluating the effective bonding area of ​​the proximal surfaces of bridge abutment teeth in Maryland. Summary of the Invention

[0006] The purpose of this invention is to provide a method for evaluating the effective bonding area of ​​the proximal surface of abutment teeth in Maryland bridges. This method aims to address the problem that in existing preoperative evaluations for Maryland bridge restorations, the area available for bonding on the proximal surface of abutment teeth often relies on manual experience or measurement using general three-dimensional software. This makes it difficult to simultaneously consider the integrity of proximal surface data, clinical bonding constraints, and objective calculation of three-dimensional surface area, resulting in insufficient consistency in the evaluation results of the effective bonding area.

[0007] In a first aspect, the present invention provides a method for evaluating the effective bonding area of ​​the proximal surfaces of abutment teeth for Maryland bridges, comprising the following steps:

[0008] A three-dimensional digital model of the oral cavity containing the target abutment tooth is obtained. The three-dimensional digital model of the oral cavity is standardized to construct a standardized triangular mesh. Sampling point cloud is extracted from the surface of the standardized triangular mesh, and the correspondence between sampling points and mesh patches is established.

[0009] Tooth position segmentation is performed on the sampling point cloud. The target abutment tooth and its adjacent teeth are determined according to clinical input. The surface data of the target abutment tooth are extracted, and the integrity of the candidate area of ​​the proximal surface of the target abutment tooth is checked. When there are missing or distorted areas in the candidate area of ​​the proximal surface, the proximal surface data is completed based on the proximal surface data of the target abutment tooth, the spatial relationship of the adjacent teeth, and the tooth position information of the target abutment tooth.

[0010] Interactive three-dimensional segmentation is performed based on the surface data of the target abutment tooth to obtain the segmentation results of the proximal surface of the target abutment tooth. The segmentation results are then constrained and filtered according to the enamel boundary, gingival margin safety distance, incisal boundary or occlusal boundary and the undercut exclusion condition of the path of insertion to obtain the effective bonding area.

[0011] The effective bonding area is mapped to the standardized triangular mesh according to the correspondence, and the three-dimensional surface area of ​​the mesh patches belonging to the effective bonding area is calculated to obtain the effective bonding area evaluation result of the target abutment tooth proximal surface.

[0012] As a preferred embodiment of the present invention, the standardization process includes:

[0013] The oral cavity 3D digital model was subjected to denoising, mesh repair and coordinate normalization in sequence to obtain a topologically continuous standardized triangular mesh.

[0014] A uniformly distributed cloud of sampling points is extracted from the standardized triangular mesh, wherein each sampling point records its corresponding mesh patch index.

[0015] The adjacent face segmentation results obtained from the sampled point cloud can be reflected back to the triangular mesh to perform area calculation.

[0016] As a preferred technical solution of the present invention, each sampling point in the sampling point cloud is configured with point-by-point geometric features, which include three-dimensional coordinates, normal vectors, curvature features and corresponding mesh patch indices.

[0017] The curvature features are obtained by fitting the local surfaces around the sampling points in different scale neighborhoods, and the curvature estimation results at different scales are weighted and fused to generate a point-by-point geometric feature table containing multi-scale fused curvature values.

[0018] As a preferred embodiment of the present invention, the adjacent candidate region is determined by the following method:

[0019] Determine the mesial or distal direction of the target abutment tooth based on its position information;

[0020] Based on the long axis direction of the target abutment tooth, the contact relationship between adjacent teeth, or the location of the edentulous space, candidate surface points facing adjacent teeth or edentulous spaces are extracted from the surface data of the target abutment tooth.

[0021] By combining the normal vector changes and multi-scale fused curvature values ​​of the candidate surface points, surface points corresponding to the labial, lingual, incisal, or occlusal surfaces are eliminated to obtain the candidate region of the proximal surface of the target abutment tooth.

[0022] The integrity detection is performed on the adjacent candidate region, which includes sampling point distribution detection, mesh hole detection and local surface continuity detection. When there is insufficient sampling point distribution, hole boundary loop located inside the adjacent candidate region, or the normal vector and curvature change do not satisfy the local surface continuity relationship, it is determined that the adjacent candidate region is missing or distorted.

[0023] As a preferred embodiment of the present invention, the adjacent face data completion includes:

[0024] The incomplete set of proximal points of the target abutment tooth, the set of points of adjacent teeth physically adjacent to the target abutment tooth, and the tooth position identifier of the target abutment tooth are jointly input into the morphological completion model.

[0025] The incomplete set of proximal points of the target abutment tooth is used to provide local geometric trends, and the set of proximal tooth points is used to provide spatial adjacency constraints.

[0026] The target abutment tooth position identifier is used to provide the morphological prior of the corresponding tooth position, and the completed proximal surface data is generated.

[0027] The completion results are then subjected to boundary continuity verification and adjacent junction rationality verification. After the verification is passed, the completed data is fused with the original adjacent face data.

[0028] As a preferred embodiment of the present invention, interactive three-dimensional segmentation includes:

[0029] Extract multi-scale surface geometric features from the target abutment tooth surface data;

[0030] Receive interactive prompts input by the physician and convert the interactive prompts into prompt features;

[0031] The multi-scale surface geometric features and cue features are used together to determine the proximal surface region of the target abutment tooth, and the proximal surface segmentation result is output.

[0032] The adjacent face segmentation results include adjacent face deterministic segmentation results and adjacent face probabilistic segmentation results. The adjacent face probabilistic segmentation results are used for subsequent boundary transition calculations of the effective bonding area.

[0033] As a preferred embodiment of the present invention, the interactive prompt information includes at least one of positive prompt points, negative prompt points, boundary key points, and prior masks;

[0034] For positive and negative prompts, extract the local geometric features of the sampling points in the neighborhood of the prompt location, and combine them with the prompt location coordinates, prompt location normal vector and prompt type to generate point prompt features;

[0035] For boundary key points, establish a continuity relationship according to the arrangement order of multiple boundary key points on the target tooth surface, and generate boundary cue features to characterize the orientation of adjacent boundary.

[0036] For the prior mask, based on the region markings of each sampling point in the prior mask, a point-by-point region hint feature corresponding to the sampling point on the target abutment tooth surface is generated;

[0037] At least one of the point hint features, boundary hint features, and point-by-point region hint features is combined to obtain interactive hint features for adjacent area determination.

[0038] As a preferred embodiment of the present invention, the interactive three-dimensional segmentation further includes:

[0039] The segmentation mode is adaptively determined based on the segmentation confidence of the interactive 3D segmentation output;

[0040] When the segmentation confidence level is higher than the first threshold, an automatic segmentation mode is adopted;

[0041] When the segmentation confidence level is between the first threshold and the second threshold, the confirmation segmentation mode is adopted;

[0042] When the segmentation confidence level is lower than the second threshold, a fine-grained interactive segmentation mode is adopted so that standard cases and complex cases correspond to different interaction intensities.

[0043] As a preferred embodiment of the present invention, the constraint screening includes:

[0044] Glaze coverage constraints are generated based on glaze boundaries to preserve adjacent face regions within the glaze coverage area;

[0045] Gingival margin safety constraints are generated based on the gingival margin line and its coronal safety distance, and are used to exclude adjacent areas close to the gingival margin.

[0046] Generate incisal boundary constraints or occlusal boundary constraints based on the target abutment tooth type to limit the effective bonding area from extending to the incisal or occlusal side;

[0047] Undercut exclusion constraints are generated based on the preset placement path direction and the surface normal vector of the target abutment tooth, which are used to exclude undercut areas that cannot be covered by the restoration along the preset placement path direction.

[0048] The effective bonding area is obtained by intersecting the enamel coverage constraint, gingival margin safety constraint, incisal boundary constraint or occlusal boundary constraint and undercut exclusion constraint with the adjacent surface segmentation results.

[0049] As a preferred embodiment of the present invention, the area calculation includes:

[0050] Multiple candidate placement path directions are set within a preset range near the long axis of the target abutment tooth. The effective bonding area corresponding to each candidate placement path direction is calculated. An optimization target is constructed based on the effective bonding area and the perimeter of the effective bonding area boundary to determine the target placement path direction.

[0051] The effective bonding area obtained under the direction of the target placement path is then mapped onto a standardized triangular mesh;

[0052] Calculate the weight value of each mesh surface that belongs to the effective bonding area, and multiply the weight value by the three-dimensional surface area of ​​the corresponding mesh surface and sum them to obtain the effective bonding area evaluation result of the target abutment tooth proximal surface.

[0053] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0054] This invention standardizes the three-dimensional digital model of the oral cavity and establishes a correspondence between sampled point clouds and triangular meshes, enabling subsequent region identification and area measurement to be based on a unified data foundation. On the one hand, the point cloud structure facilitates tooth segmentation and local surface feature extraction; on the other hand, the mesh structure facilitates the calculation of surface accumulation and spatial relationships. This avoids the region judgment bias caused by relying solely on manual observation, two-dimensional projection, or a single model structure in existing technologies, thereby improving the consistency and reusability of the local usable area assessment process.

[0055] This invention performs integrity checks on the proximal surface region after determining the target abutment tooth and its adjacent teeth. When missing or distorted areas are detected, the system incorporates the geometric trend of the proximal surface itself, the spatial relationship between adjacent teeth, and prior knowledge of tooth position morphology to complete the data. This ensures the continuity and rationality of the surface data input into subsequent segmentation processes. Consequently, it reduces boundary drift, region fragmentation, or measurement distortion caused by occlusion, noise, or local defects in the original scan data, allowing local region definition to be established on a more stable geometric basis.

[0056] This invention performs interactive three-dimensional segmentation on the proximal surfaces of the target abutment teeth and combines the enamel boundary, gingival margin safety distance, incisal or occlusal boundary, and undercut exclusion conditions of the path of insertion to constrain and screen the segmentation results. This further transforms the anatomically significant proximal surface region into an effective bonding region that meets the conditions for restorative application. At the same time, the evaluation results are output by using a three-dimensional surface accumulation method mapped to a standardized triangular mesh. This ensures that the results are no longer limited to empirical judgment or general regional measurements, but can be connected with the actual restorative design conditions, providing a quantitative basis for clinical plan selection and restoration design. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0058] Figure 1 This is a schematic diagram of the overall architecture of the three-dimensional tooth proximal segmentation network of the present invention;

[0059] Figure 2 This is a schematic diagram of the process and algorithm for encoding interactive operations into semantic vectors in this invention;

[0060] Figure 3 This is a schematic diagram of the stepwise constraint solution algorithm for the effective bonding area of ​​the proximal surface of the abutment tooth in this invention;

[0061] Figure 4 This is a triangular mesh model of the maxillary full dental arch obtained from the original scan;

[0062] Figure 5 This is a user interface diagram for assessing the available adhesive area on adjacent surfaces. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings.

[0064] Throughout the accompanying drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. The described embodiments are only a part of the embodiments of this application, not all of them. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0065] Example 1

[0066] This embodiment provides a method for evaluating the effective bonding area of ​​proximal surfaces of abutment teeth for Maryland bridges. On a three-dimensional digital model of the abutment tooth, the effective area of ​​the proximal surface that meets the bonding conditions is precisely defined, and its three-dimensional surface area is repeatedly quantified. In most standard cases, proximal surface segmentation and area calculation can be completed fully automatically, with minimal interactive prompts for rapid correction, achieving an AI-driven process with manual fine-tuning. The method includes the following steps:

[0067] S101: Obtain a three-dimensional digital model of the oral cavity containing the target abutment tooth, perform standardization processing on the three-dimensional digital model of the oral cavity, construct a standardized triangular mesh, extract sampling point cloud based on the standardized triangular mesh, and establish the correspondence between sampling points and mesh patches.

[0068] In this embodiment, the data sources of the oral cavity three-dimensional digital model include: triangular mesh files output by an intraoral optical scanner, the format of which includes STL, OBJ, or PLY; and a three-dimensional model generated by cone-beam CT after threshold segmentation and surface reconstruction. The neural network recognizes the geometric features of the tooth surface using point clouds as the computational carrier, while the three-dimensional surface accumulation, geodesic distance calculation, and boundary loop detection depend on the mesh topology connection relationship. The triangular mesh structure and point cloud structure are preserved, and the one-to-one correspondence between the two is maintained through index mapping.

[0069] First, the original triangular mesh is standardized. This standardization process includes bilateral filtering for noise reduction, mesh restoration, and coordinate normalization. Bilateral filtering smooths high-frequency noise while preserving geometric features of the tooth surface, such as ridges, grooves, and corners. Mesh restoration fills small cavities and eliminates non-manifold edges and degenerate triangles to obtain a topologically continuous mesh structure. Coordinate normalization establishes a unified coordinate system with the occlusal plane as a reference, thereby reducing the impact of different cases and scanning postures on subsequent geometric analysis and segmentation results. After these processes, a standardized triangular mesh is obtained. ,in: For vertex set, For edge set, This is a set of triangular facets. This mesh is retained throughout all five steps and used for the final area and distance calculations.

[0070] After obtaining the standardized triangular mesh Then, uniformly distributed sampling points are extracted from the surface of the standardized triangular mesh to form a point cloud of the entire dental arch. The Poisson disk sampling method is preferred for sampling to ensure that the sampling points are evenly distributed on the tooth surface and to avoid localized over-dense or under-dense sampling. For each sampling point... The index of the nearest triangular facet is recorded to form a "sampling point-facet index mapping table", so that the subsequent point-by-point segmentation results obtained based on the sampling point cloud can be accurately mapped to the corresponding facet of the standardized triangular mesh.

[0071] Furthermore, geometric features are extracted for each sampling point and used as input to the subsequent segmentation network. For any sampling point... Its input feature vector is denoted as:

[0072] ;

[0073] in: In three-dimensional coordinates, It is the normal vector. , These are the maximum principal curvature and the minimum principal curvature, respectively. For the mean curvature, The curvature is Gaussian. Curvature features are used to characterize the local bending changes in the transition area between the adjacent surfaces and the labial and lingual surfaces. The corner lines, clinically referred to as corners, are geometrically represented as the local maxima of the maximum principal curvature. Therefore, curvature features are an important basis for subsequent determination of adjacent surface boundaries.

[0074] The 3D Dental-SAM module of this invention relies on these curvature patterns to determine the spatial extension range of adjacent surfaces. Curvature is extremely sensitive to high-frequency noise in intraoral scan data caused by saliva reflection or scanner precision limitations. Even small noise disturbances can cause drastic fluctuations in curvature values, thus interfering with subsequent boundary determination. To address this, a multi-scale neighborhood weighted fitting scheme is employed on the mesh after bilateral filtering and denoising: quadratic surface fitting is performed on local patches at multiple different neighborhood radii, curvature estimates at each scale are calculated, and then a Gaussian weighted average is taken as the final curvature. Quadratic surface fitting is performed on local patches at neighborhood radii of 0.5mm, 1.0mm, and 1.5mm, and the Gaussian weighted average of the estimates at each scale is taken as the final curvature. Small-scale neighborhoods capture fine local changes but are sensitive to residual noise, while large-scale neighborhoods provide good smoothing effects but may blur feature details. Multi-scale fusion achieves a balance between fidelity and noise reduction.

[0075] Local neighborhood weighted fitting: at each scale Below, collect with Center, radius All mesh vertices within. Established using... Let the origin be the point and the normal vector of that point be... for The local coordinate system of the axis, for neighboring vertices Local coordinates Perform quadratic surface fitting:

[0076] ;

[0077] in: The coefficients are the fitting coefficients for the quadratic surface, and the coefficient vector is... The solution is obtained using the weighted least squares (WLS) method. For ease of uniform representation, the above equation can be written in linear parametric form. ,in , Point The height component along the normal axis in the local coordinate system.

[0078] Within the neighborhood Fitting weights for each point The accuracy decreases with spatial distance to enhance the fit to the center point:

[0079] ;

[0080] in: and Represent the center point and the th neighboring point, respectively. The three-dimensional coordinates of each point; Represents the Euclidean norm; Indicates the first The neighborhood radius is a scale; Let be the vector of coefficients to be estimated; For the first The first scale The fitting weights for each neighborhood point. The objective function of WLS can be written as:

[0081] ;

[0082] in: Indicates the first Under each scale, Centered and satisfying The set of neighborhood indexes.

[0083] Analytical curvature calculation: based on fitting coefficients At the origin The first and second derivatives of the local surface can be obtained as follows:

[0084] , , , , .

[0085] Based on this, the mean curvature and Gaussian curvature at this scale can be calculated:

[0086] ;

[0087] ;

[0088] The principal curvature is derived from the mean curvature and Gaussian curvature:

[0089] ;

[0090] in: This is usually used as a convention for sorting principal curvatures. If the local slope is small, i.e. and If the result is approximately zero, then the above result can be approximated as a Hessian matrix. The eigenvalues ​​are then consistent with the original abbreviation of "eigenvalue decomposition".

[0091] Multi-scale Gaussian fusion: The estimates from each scale are fused into the final curvature feature using a Gaussian weighted average. Taking the maximum principal curvature as an example:

[0092] ;

[0093] ;

[0094] in: The total number of scales; For the first One scale radius; For reference scale; For scale bandwidth; For the first Fusion weights at each scale; For point In the The maximum principal curvature is estimated at each scale.

[0095] S102: Perform tooth position semantic segmentation on the preprocessed full arch sampling point cloud P, determine the target abutment tooth and its adjacent teeth according to clinical input, extract the surface data corresponding to the target abutment tooth, and perform integrity detection on the proximal area of ​​the target abutment tooth. When missing or distorted data is detected, the proximal data is completed to obtain complete proximal data that meets the requirements of subsequent proximal segmentation and area calculation.

[0096] In this embodiment, the tooth segmentation preferably utilizes a pre-trained three-dimensional tooth segmentation network to perform semantic segmentation on the point cloud of the entire dental arch, assigns an FDI tooth position label to each sampling point, and automatically decomposes the entire dental arch into independent single teeth and gingival regions.

[0097] Specifically, the backbone architecture of the segmentation network is based on PointNet++ hierarchical feature extraction or DGCNN dynamic graph convolution, containing four Set Abstraction (SA) layers for progressive downsampling and expanding the receptive field, and four symmetrical Feature Propagation (FP) layers for restoring pointwise resolution through skip connections. Based on the clinically input abutment tooth location, the system automatically determines the target abutment tooth. and its related adjacent teeth And extract the point set on the surface of the abutment tooth. and the corresponding grid subset.

[0098] In a preferred embodiment, in the first Within each SA layer, based on sampling points Centered on, within radius Local features are aggregated within a spherical neighborhood (layers are 2mm, 4mm, 8mm, and 16mm respectively): The relative spatial coordinates of the neighboring points and the center point (providing local translation invariance) are concatenated with the geometric features of the previous layer at that point. Local information is extracted through a shared multilayer perceptron, and then aggregated by max pooling to form a new feature vector for that sampling point.

[0099] ;

[0100] in: Indicates the first Sampling points in each SA layer eigenvectors; Indicates the neighboring points of the previous level. eigenvectors; Indicates the initial input features; Indicates the first The neighborhood radius of each SA layer; Indicates feature splicing; Indicates the first Layer-sharing multilayer perceptron; This represents the max pooling operation; Represents a point in three-dimensional space; Represents the Euclidean norm; Representation layer index.

[0101] Preferably, when the point cloud maintains the true intra-orbit coordinate system and the unit is mm, the neighborhood radius can be directly taken as 2 mm, 4 mm, 8 mm and 16 mm; if the input point cloud is normalized and scaled, the neighborhood radius should be converted proportionally according to the scaling factor to maintain physical scale consistency.

[0102] The final feature propagation layer outputs a logit vector for each sampling point corresponding to 33 categories. The class probabilities are obtained by performing softmax normalization on the logit vector:

[0103] ;

[0104] Then select the category with the highest probability as the predicted label:

[0105] ;

[0106] in: Indicates sampling point Belongs to the The class's logit value; This represents the corresponding normalized probability; Indicates sampling point Predicted tooth position labels; Indicates a category index; The operator that retrieves the index corresponding to the maximum value; It is an exponential function; This indicates summing over all categories.

[0107] After completing the point-by-point tooth position prediction, connected component analysis is performed based on the mesh topology connectivity preserved in the preprocessing stage. This process suppresses noise in isolated connected components with too few points of the same type. Specifically, connected component analysis is performed on points of the same type using mesh topology or point adjacency graphs to obtain connected components. If the number of points in a connected component satisfies If it is a segmentation noise point, then its label is reset to the category that appears most frequently in its adjacent region, so as to obtain a set of single tooth points with smooth boundaries and no fragmentation.

[0108] in: Indicates the first One connected component; This indicates the number of points in the connected component; The minimum connected component threshold is preferably 50, and this threshold can be configured based on the statistical results of the training set or the clinical calibration results.

[0109] Based on the clinically input abutment tooth location (FDI number) and restoration design type, the system automatically determines the target abutment tooth according to physical adjacency rules. and its adjacent teeth set The physical adjacency rule is as follows: Only includes with Teeth that are directly in contact with or adjacent to each other on the dental arch, and teeth on the opposite side that cross the edentulous space, are not included in the adjacent tooth assembly. For example, in the case of missing tooth 14 and abutment tooth 15... Only the outer adjacent tooth 16 is included, while 13 and 15 are separated by a gap in the edentulous space and do not constitute a physical adjacency relationship. This definition is used to ensure that the spatial constraint of "the restoration must not penetrate the adjacent tooth" can be applied during the subsequent proximal restoration process, so that the restoration result has physical rationality.

[0110] After determining the target abutment tooth, extract the set of abutment tooth surface points from the tooth segmentation results:

[0111] ;

[0112] The mesh subset 〖M^'〗_T⊂M^' corresponding to the target base tooth is extracted through the sampling point-patch index mapping table, and together with the point set P_(N(T)) corresponding to the adjacent tooth set, it is sent to subsequent processing.

[0113] Where: P represents the point cloud of the entire dental arch; P_T represents the surface point set of the target abutment tooth T; ŷ_i represents the predicted tooth position label of the sampling point p_i; FDI(T) represents the FDI number of the target abutment tooth T; M^' represents the normalized triangular mesh; 〖M^'〗_T represents the mesh subset corresponding to the target abutment tooth; P_(N(T)) represents the point set corresponding to the set of adjacent teeth N(T).

[0114] Since the proximal surfaces of target abutment teeth are often located in the interproximal zone, the edge of edentulous spaces, or scanning blind spots, they are easily affected by occlusion, reflection, or angular limitations during intraoral scanning, leading to sparse sampling, missing cavities, or local artifacts in the proximal surface region. Therefore, before proceeding to precise segmentation of the proximal surface boundary, an integrity check is performed on the proximal surface region to determine whether the proximal surface data meets the requirements for subsequent precise segmentation and area calculation. To avoid definition loops before the "proximal surface" is precisely segmented, this embodiment preferably first obtains the candidate proximal surface region using a coarse localization method. Integrity checks are then performed on the candidate region.

[0115] In this embodiment, the integrity detection includes at least three criteria: sampling density, hole area ratio, and morphological anomaly. For sampling density, each sampling point within the adjacent candidate region... Calculate its radius The number of points in the neighborhood is normalized to the areal density, and the average value of the region is taken.

[0116] ;

[0117] in: , ; This represents the set of candidate adjacent face points to be detected; Point In radius The set of neighboring points within; This represents the number of points in the adjacent candidate point set; This represents the number of points within that neighborhood; This represents the average point density of the adjacent area; This represents the density detection radius, preferably 1.0 mm; This represents an approximate value for a local area. Here, it is preferred to use... As the projected area under the local tangent plane approximation, it is used to perform a uniform comparison of different local neighborhoods; when When the curvature is sufficiently small and the local curvature changes are not drastic, this approximation meets the requirements for engineering applications.

[0118] when Below the preset threshold When the minimum density threshold is reached, it indicates that there is sparse sampling or local missing sampling in the adjacent area. Preferred setting is This value can be configured based on the scanner's nominal sampling density, training set statistics, or clinical calibration results. When the sampling density is too low, subsequent point cloud-based boundary segmentation results are prone to boundary instability or local misjudgments, and the area calculation results after mesh mapping will also be affected. Therefore, density detection is used to screen out obviously incomplete areas first.

[0119] To determine the percentage of void area, we detect boundary loops (closed loops formed by non-shared edges) in the candidate mesh subsets of adjacent faces, and calculate the sum of the missing areas enclosed by all boundary loops. Estimated total area of ​​candidate regions with adjacent faces The ratio:

[0120] ;

[0121] in: This represents the total area of ​​all the holes; This represents the estimated total area of ​​the adjacent candidate regions; This represents the proportion of the hole area and is a dimensionless quantity. When Exceeding the allowable threshold When the maximum hole area is too large, the missing area may significantly affect subsequent area calculations. The maximum hole percentage threshold is... The preferred setting is 15%, meaning that completion is triggered when the cumulative area of ​​the holes exceeds 15% of the estimated total area of ​​the adjacent surfaces. This threshold can be configured based on the target prosthesis's tolerance for area errors, training set statistics, or clinical calibration results. Even if the overall sampling density is acceptable, if continuous holes form locally, the geometric topology of the adjacent surfaces will be incomplete, and subsequent area calculations will be significantly underestimated due to missing hole boundaries. Therefore, completion should be prioritized before segmentation.

[0122] For morphological anomalies, the reconstruction error of a pre-trained autoencoder is used to determine the presence of scanning artifacts. The candidate point set of adjacent surfaces is fed into the pre-trained autoencoder to reconstruct the adjacent surfaces, and the average reconstruction distance between the input and the reconstruction output is calculated.

[0123] ;

[0124] in: This represents the point set reconstructed by the autoencoder; Indicates the input point; Points representing the set of reconstructed points; Represents the Euclidean norm; This indicates that the point in the set of reconstructed points is the closest to the input point. This represents the average reconstruction error, expressed in mm. If the autoencoder output is an unordered point set, the error can be defined using the nearest neighbor correspondence method, thus avoiding ambiguity in the one-to-one correspondence during point set reconstruction.

[0125] when Exceeding the threshold When this occurs, it indicates that the proximal morphology deviates from the normal geometric distribution of that tooth position, potentially indicating scanning artifacts or severe distortion. The aforementioned morphological abnormality threshold... The preferred setting is 0.15 mm. This value can be configured based on the statistical distribution of normal reconstruction errors or the clinical tolerance range, and is approximately three times the upper limit of normal reconstruction errors. This ensures detection sensitivity while avoiding misclassifying normal anatomical variations as artifacts. Some scanning artifacts or local distortions may not necessarily manifest as obvious holes or insufficient density, but their surface morphology deviates from the statistical distribution of normal tooth positions, which can still interfere with boundary segmentation and effective area calculation. Such geometric abnormalities can be detected by using an autoencoder to reconstruct the error.

[0126] In summary, if any one of the data exceeds the permissible range, it is determined that the proximal surface data is missing or distorted, and supplementation is required. The goal of the supplementation process is not to simply fill in the geometric gaps, but to reconstruct a reasonable proximal surface that is consistent with the anatomical structure of the target tooth and does not penetrate the adjacent tooth, based on the existing geometric trend of the target abutment tooth, the spatial constraints of the adjacent tooth, and the prior morphology of the target tooth position.

[0127] Therefore, before entering the subsequent adjacent surface completion network, the system will remove the incomplete target tooth adjacent surface point set. The set of adjacent tooth points physically adjacent to the target abutment tooth and the target abutment tooth position marker As input, the morphological completion network is automatically invoked. The missing region is reconstructed based on the three complementary information paths, resulting in the completed set of adjacent points. :

[0128] ;

[0129] in: This is the set of proximal points of the target abutment tooth after completion; To complete the network for morphological purposes; For network parameters; For an incomplete set of proximal points of the target tooth base; For adjacent tooth point set; Number the tooth position of the target abutment tooth; For the target abutment tooth.

[0130] The incomplete set of proximal points provides the geometry and curvature trend of the hole edge, which is the most direct local basis for inferring the missing area; the set of adjacent tooth points provides spatial adjacency constraints, which restrict the completed proximal surface from penetrating the adjacent tooth surface and should maintain a reasonable adjacency distance with the adjacent tooth; the tooth position identifier provides the anatomical morphology prior corresponding to the tooth position, which guides the completion result to converge to the reasonable tooth position morphology when local geometric information is insufficient.

[0131] To effectively fuse the three heterogeneous information streams, features are first extracted using a multi-source information fusion encoder; then, multi-scale features are extracted from the set of proximal points of the abutment teeth using a 4-layer PointNet++ algorithm. Adjacent-tooth spatial constraints are injected at each scale through cross-attention:

[0132] ;

[0133] in: Indicates the incomplete adjacent face at the th Characteristics of the layer; Indicates the adjacent tooth in the first position Characteristics of the layer; This indicates the fusion features after injecting adjacent tooth constraints; Represents the cross-attention operator; , , These represent Query, Key, and Value, respectively. The representation layer index. Tooth position priors are injected via adaptive instance normalization (AdaIN). Mapped to embedding vector Generate channel-wise scaling and offset parameters, and perform affine transformation on the fused features:

[0134] ;

[0135] in: Indicates the first Layer fusion characteristics; and These represent the channel-wise scaling and offset parameters generated by tooth position embedding, respectively; Represents the characteristic mean; Indicates the characteristic standard deviation; This indicates element-wise multiplication. The conditional parameters allow the same network to produce different reconstruction tendencies at different tooth positions without requiring separate model training for each tooth position.

[0136] The fused features output by the encoder are then fed into a hierarchical decoder, which maps the high-dimensional features back to three-dimensional space, employing a two-stage generation strategy from coarse to fine. The first stage starts from the deepest fused features of the encoder and generates features through fully connected layers. Seed points, preferred The first stage is used to outline the overall contour of the missing area. The second stage, using the features of each seed point as conditions, simultaneously generates a 3D spatial offset and a surface normal vector through a local multilayer perceptron, expanding the seed points into a cluster of refined oriented points:

[0137] ;

[0138] , , ;

[0139] in: Indicates the first Seed points; Represents seed point The corresponding feature vector; This represents a random vector sampled from a three-dimensional standard Gaussian distribution, used to introduce local surface diversity; Indicates the first The seed point at the th th Spatial offset predicted under each random branch; This represents the corresponding normal vector; Indicates the coordinates of the generated point; Represents the normalized normal vector; This represents the index of the generated branch under the same seed point. Two-level expansion output. A reconstruction point with normal information is used to construct a fine-grained geometric representation of the adjacent faces.

[0140] To ensure the generated point cloud is realistic and reasonable, the network training uses a multi-objective constrained loss function, defined as:

[0141] ;

[0142] in: Indicates the total loss; Indicates the first One loss item; Indicates the first The weighting coefficients of each loss term; This indicates the index of the loss term. The geometric coincidence loss uses a bidirectional chamfer distance (ChamferDistance) to constrain overall shape fit.

[0143] ;

[0144] in: Represents the reconstructed point set; Represents the real point set; and Let each represent the number of points in the two point sets; This represents the L2 norm. The normal vector consistency loss uses the normal vectors predicted by the decoder to constrain the reconstructed surface orientation to be consistent with the real surface.

[0145] ;

[0146] in: Indicates the generation point The predicted normal vector; Indicates the concentration of real points and The normal vector of the nearest neighbor; Indicates the nearest neighbor index; Represents the vector dot product. Adversarial loss. The discriminator helps the reconstructed results approximate the real data in terms of morphological distribution. Adjacency distance constraint. Using signed projection distance to severely punish physical clipping or gap anomalies:

[0147] ;

[0148] ;

[0149] in: Point Signed distance to the surface of the adjacent tooth; The distance from the adjacent tooth The nearest point; This is the normal vector of the adjacent tooth point facing outwards; This indicates the maximum permissible normal clearance, preferably 0.5mm; when If a clipping collision occurs, the network will impose a penalty. Finally, there is the shape regularization loss. The local curvature distribution of the constrained reconstructed surface is consistent with the statistical distribution of the tooth position to prevent sharp protrusions or abnormal angles.

[0150] The completed adjacent faces must pass two verifications. The first is a transition continuity verification, which checks whether the surface normal change at the boundary between the completed region and the original adjacent face is smooth, avoiding abrupt changes in normal, sharp angles, or geometric steps at the splicing boundary. Preferably, the normal deflection angle of adjacent point pairs in the boundary region is calculated:

[0151] ;

[0152] in: This represents a pair of adjacent points in the boundary region; Representing points respectively With point The normal vector; This represents the normal deflection angle between two points. Further, the maximum normal deflection angle of the boundary region is defined as:

[0153] ;

[0154] in: Represents the set of adjacent point pairs on the boundary; This represents the maximum normal deflection angle of the boundary region. When The continuity verification is passed at the specified time. .

[0155] The second aspect is the verification of interproximal junction rationality. This verification is used to check whether the closest distance between the repaired proximal surface and the adjacent tooth is within a reasonable range, ensuring that the repair neither intrudes into the surface of the adjacent tooth nor creates an abnormally large gap between them. Preferably, the signed distance from the repaired proximal surface to the adjacent tooth surface is calculated:

[0156] ;

[0157] in: This indicates that sampling points on adjacent faces are completed; Indicates and The nearest adjacent tooth point; Point The external normal vector; This represents the signed distance. Further, the minimum signed distance and the average signed distance are defined as follows:

[0158] ;

[0159] ;

[0160] in: Represents the minimum signed distance; Indicates the average signed distance; This represents the set of adjacent points after completion. When... and The time determination is passed through the adjacent join rationality verification, where mm.

[0161] When the completion result satisfies both the transition continuity verification and the adjacency joining rationality verification, the completed data is fused with the original adjacent surface data to obtain the complete adjacent surface data. During the fusion process, the unmissing parts of the original scan data remain unchanged, and only the missing areas in the original proximal surfaces are filled with the completion results. This preserves the true tooth geometry information reflected in the original scan while also supplementing the proximal surfaces that were missing due to scanning limitations.

[0162] S103: Perform interactive three-dimensional segmentation based on the target abutment tooth surface data to obtain the proximal segmentation results of the target abutment tooth. Combine the enamel boundary, gingival margin safety distance, incisal or occlusal boundary and undercut exclusion conditions of the path of insertion to constrain and filter the proximal segmentation results to obtain the effective bonding area.

[0163] It should be noted that the transitional surface between the proximal surface and the labial / lingual surface is a continuous curved transition area, a conceptual transition zone rather than a sharp geometric boundary. Different surgeons may differ by several millimeters in their judgment of the extent of the proximal surface extension, and a boundary offset of 1mm can lead to an error of 5% to 8% in the adhesive area. The accurate and repeatable definition of the proximal surface boundary is the most significant factor affecting the final area accuracy.

[0164] The interactive three-dimensional segmentation model is preferably the 3DDental-SAM model, whose input is the target tooth surface point set. and a collection of prompts The output is a pointwise probability mask of the adjacent surfaces of the target tooth. The process can be represented as follows:

[0165] ;

[0166] in: Represents the segmentation model. Indicates model parameters; Represents the set of surface points of the target abutment tooth; This represents the prompt set. Physician interaction operations are encoded to form prompt signals. In fully automatic mode, the prompt set is empty. Represents a point-by-point probability mask; This indicates the total number of sampling points on the surface of the target abutment tooth; This indicates that each sampling point corresponds to a probability value between 0 and 1.

[0167] The probability value obtained for each surface point represents the likelihood that the point belongs to a neighboring surface. After thresholding, this mask yields the deterministic boundary of the neighboring surfaces. Preferably, the hard mask can be written as:

[0168] ;

[0169] in: Indicates the first Binary neighbor label for each sampling point; Indicates the first The probability of adjacent faces of each sampling point; This represents the threshold value, preferably 0.5. This indicates an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise. Indicates the sampling point index.

[0170] The 3D Dental-SAM model consists of three parts: a surface feature encoder, an interactive cue encoder, and a mask decoder. The surface feature encoder is used to extract multi-scale geometric features of the target abutment tooth surface; the interactive cue encoder is used to convert the physician's input interactive operations into cue embeddings; and the mask decoder is used to fuse surface geometric features and cue embeddings, and output the segmentation probability mask of the target abutment tooth proximal surface and the corresponding confidence score.

[0171] Details are as attached Figure 1 As shown, component A is a surface feature encoder. The task of the surface feature encoder is to encode the geometric features of each location on the tooth surface into a high-dimensional feature vector, which is then called by the mask decoder when determining the boundary.

[0172] 1) Site Alignment: Before encoding, principal component analysis (PCA) is used to determine the three principal axis directions of the abutment tooth point cloud. The tooth long axis is aligned to the Z-axis, the mesiodistal direction is aligned to the X-axis, and the labiolingual direction is aligned to the Y-axis. This operation ensures that the same tooth position in different patients and different scanning postures has a consistent spatial orientation when entering the network.

[0173] 2) Encoder Architecture: The encoder adopts an improved Point Transformer V3 architecture, containing four encoding layers. The point cloud is downsampled progressively between layers at a 4x ratio. The core computational method is sequential attention: mapping the 3D point cloud along a space-filling curve (Z-order or Hilbert curve) into a one-dimensional sequence, within a fixed-size local window (window size...). Internal computation of self-attention:

[0174] ;

[0175] in: Represents the Query matrix; Represents the Key matrix; Represents the Value matrix; The feature dimensions representing the query / key; Indicates matrix transpose; Represents the normalization function; Indicates the size of a local window.

[0176] Local windows reduce computational complexity to that of global attention. Reduced to ,in This represents the number of input points, enabling efficient processing of tens of thousands of sampling points. Each encoding layer simultaneously uses multiple space-filling curves with different configurations to serialize the point cloud, and each path independently calculates attention before fusing the results. Since different curves have different spatial traversal paths, multi-path fusion can keep the feature extraction stable against small pose deviations and sampling randomness of the point cloud.

[0177] 3) Multi-path serialization fusion: Each coding layer uses multiple spatial filling curves with different configurations to serialize the point cloud simultaneously, and the attention of each path is calculated independently before fusion. Different curves have different spatial traversal paths, and multi-path fusion keeps the feature extraction stable against small pose deviations and sampling randomness of the point cloud.

[0178] The encoder outputs a multi-scale feature pyramid. The resolutions are as follows: , , , The feature dimension per scale is (Default value: 256). The highest resolution feature is denoted as... High-resolution layers preserve fine-grained curvature variations at adjacent surface boundaries, while low-resolution layers capture the overall morphological information of adjacent surfaces.

[0179] 4) Two-stage pre-training. The first stage is self-supervised learning: Masking autoencoders are performed on over 500 unlabeled full dental arch scans, randomly covering 60% of the surface area. The encoder is then trained to predict the coordinates and normal vectors of the covered portion.

[0180] ;

[0181] in: This represents the mask autoencoding loss; The index set representing the occluded points; Indicates the number of points that are covered; Indicates the first The coordinates of a real point; Indicates the first Reconstructed coordinates of each point; Represents the Euclidean norm; Indicates the first Predicted normal vectors for each point; Indicates the first The true normal vector of each point; The normal vector loss weights are represented.

[0182] This stage enables the encoder to learn general geometric patterns of the tooth surface from a large amount of data, and the second stage performs supervised fine-tuning on a dataset with tooth surface partition annotations.

[0183] Component B is an interactive cue encoder. This encoder transforms click interactions into semantic vectors of the same dimension as surface features, enabling the mask decoder to utilize both geometric features and the physician's intent. This encoder supports four types of cues, which are then concatenated into a cue embedding matrix. As attached Figure 2 As shown, taking the intraoral scan point cloud of the mesial surface of the maxillary second premolar as an example, three types of prompts, including point prompts, boundary key point prompts, and prior mask prompts, are transformed into embedding vectors of a unified dimension through independent encoding paths. These are then combined into a prompt embedding matrix, which is input to the segmentation decoder. (a) Point prompt encoding path: joint encoding based on local geometric context extraction of spherical neighborhood and intrinsic features of the click location; (b) Boundary key point prompt encoding path: spatial continuity modeled by 1D Transformer for key point sequences sampled along the corner line; (c) Prior mask prompt encoding path: point-by-point MLP encoding of coarse segmentation binary masks. The right-hand convergence area shows the process of combining the three outputs into a unified prompt embedding matrix.

[0184] The dentist clicks a point on the surface of the abutment tooth. The system obtains the coordinates, normal vector, and curvature of the point. The key design element is the introduction of a local geometric context for each click: automatically extracting the coordinates, normal vector, and curvature of the point centered on the click location. Features of all sampled points within a spherical neighborhood (default 2mm) are aggregated into a context vector using a lightweight PointNet network:

[0185] ;

[0186] in: Indicates the first The local context vector corresponding to each click; Represents the shared pointwise feature extraction function; Indicates the number of neighbors within the neighborhood. Input features of each sampling point; Indicates the first The coordinates of each sampling point; Indicates the first Coordinates of each click point; The radius of the local neighborhood is represented, preferably 2 mm; This represents the max pooling operation. The dot hint embedding is:

[0187] ;

[0188] in: Indicates the first Embedded vectors for each point suggestion; A dedicated multilayer perceptron for indicating point cues; Indicates feature splicing; This represents the normal vector of the click point; Indicates the curvature at the click point; The learnable type embedding is used to distinguish between positive prompts "this belongs to the neighboring face" and negative prompts "this does not belong to the neighboring face". The role of local context is that, since the geometric features of the 2 mm neighborhood near the center of the neighboring face and the corner line are very different, the contextual information enables the network to distinguish whether the click is located inside the neighboring face or near the boundary, thus inferring the range of the neighboring face more accurately.

[0189] The physician clicks on several key points (usually 2 to 5) sequentially on the adjacent boundary line. Each point is independently coded and then fed into a one-dimensional Transformer to model spatial continuity.

[0190] ;

[0191] in: The encoding result representing the boundary key point hints; Indicates the first Embedded vectors of key boundary points; Indicates the total number of boundary critical points; This represents a one-dimensional Transformer. This cue allows physicians to directly specify the orientation of adjacent face boundaries, reducing positioning errors in corner areas from millimeters to sub-millimeter levels. The prior mask cue inputs a coarsely segmented binary mask into point-by-point MLP encoding to obtain the corresponding cue embedding, denoted as . ,in This indicates the embedding of prior mask hints. This represents a coarsely segmented binary mask. This represents a masked encoding multilayer perceptron. The three cue embeddings are concatenated along the token dimension to form a unified cue embedding matrix. .

[0192] Component C is a mask decoder, which fuses surface geometric features with fine-grained segmentation operations, outputting an adjacent-face segmentation mask. Its input is surface features. , prompt embedding And 3 learnable output query tokens The decoding process consists of two iterations, each of which sequentially performs self-attention fusion, query-to-surface cross-attention, surface-to-query reverse cross-attention, and point-by-point feedforward network.

[0193] (1) Perform self-attention fusion on the prompt embedding and output query so that the output query can understand the interactive intent of the physician input;

[0194] (2) Perform cross-attention on the query surface to retrieve geometric information related to the adjacent surfaces from the surface features of the target abutment tooth:

[0195] ;

[0196] in: Represents the cross-attention operator; Represents the query matrix; Represents the key matrix; Represents a value matrix; This indicates that the query token matrix will be output. This represents the surface feature matrix.

[0197] (3) Perform reverse cross-attention from surface to query so that surface features can be updated according to the current query state:

[0198] ;

[0199] in: The surface features before the update are denoted as ; the surface features after the update are denoted as . .

[0200] (4) The updated surface segmentation features are output through a pointwise feedforward network. After two iterations, each output query is multiplied by the updated surface features and normalized by sigmoid to output the segmentation probability mask and confidence score:

[0201] ;

[0202] ;

[0203] in: Indicates the first Each output granularity corresponds to a segmentation probability mask; This represents the sigmoid function; This represents the updated surface feature matrix; Indicates the first One output query vector; Indicates the first The confidence level corresponding to each output query; This indicates a confidence-based prediction multilayer perceptron. This indicates the output granularity index, which can be 1, 2, or 3. The three outputs correspond to the complete adjacent surface, the functional sub-region such as the gingival 1 / 3, and the adjacent contact area, respectively. The default granularity for adhesive area evaluation is 1.

[0204] Furthermore, to account for the differences in interaction intensity between standard and complex cases, the interactive 3D segmentation adaptively determines the segmentation mode based on the segmentation confidence output by the mask decoder. Specifically, the system uses the confidence level of the output granularity corresponding to the complete adjacent face as the current adjacent face segmentation confidence level, denoted as:

[0205]

[0206] in: This indicates the segmentation confidence level corresponding to the current adjacent face segmentation result; This represents the confidence level corresponding to the first output granularity; the first output granularity corresponds to the complete adjacent face segmentation result. and The value ranges from 0 to 1. The larger the value, the higher the reliability of the current adjacent surface segmentation probability mask being consistent with the actual adjacent surface boundary of the target tooth.

[0207] The system presets a first threshold. Second threshold .in: This represents the first confidence threshold used to determine whether to enter automatic segmentation mode; This represents the second confidence threshold used to determine whether to enter the fine-grained interactive segmentation mode; The value range is from 0.83 to 0.95; The value range is from 0.69 to 0.82; and In a preferred embodiment, , .

[0208] (1) If the segmentation confidence satisfies:

[0209]

[0210] The system identifies the current case as a standard case with regular adjacent surface morphology and clear boundaries, and adopts automatic segmentation mode. In automatic segmentation mode, a set of suggestions is displayed. If the input is empty or no further prompts are required from the physician, the system directly uses the adjacent probability mask output by the mask decoder as the adjacent segmentation result.

[0211] (2) When the segmentation confidence level satisfies:

[0212]

[0213] The system employs a confirmation-based segmentation mode. In this mode, the physician only needs to input a confirmation point in the target proximal area or confirm the current automatic segmentation result. The system will then encode this confirmation operation as a point prompt and update the proximal segmentation mask. If the updated boundary still deviates from the physician's judgment, a few additional positive or negative prompt points can be added for fine-tuning. This mode is suitable for cases where the dentition is generally regular but there is some uncertainty regarding local corner lines, proximal areas, or edentulous space boundaries.

[0214] (3) When the segmentation confidence level satisfies:

[0215]

[0216] The system identifies the current case as complex and employs a refined interactive segmentation mode. In this mode, physicians use a combination of positive and negative prompts, boundary key point prompts, and / or prior mask prompts to precisely define adjacent surface boundaries. Preferably, the physician inputs 2 to 5 boundary key points along the corner line between the adjacent surface and the lip or tongue surface, and inputs negative prompt points in missegmented areas, allowing the mask decoder to regenerate the adjacent surface probability mask based on the physician's interactive intent.

[0217] Specifically: positive dot hints indicate that the physician marks the location as belonging to an adjacent region; negative dot hints indicate that the physician marks the location as not belonging to an adjacent region; boundary key point hints indicate the sequence of key points entered by the physician along the adjacent boundary or corner line; prior mask hints indicate binary mask hints formed by coarse segmentation results or historical segmentation results. Therefore, standard cases, general cases, and complex cases correspond to automatic segmentation mode, confirmation segmentation mode, and fine interactive segmentation mode, respectively, allowing different cases to obtain different levels of interaction.

[0218] After the doctor confirms, the system outputs two forms of adjacent face masks: hard mask. The deterministic boundary of adjacent faces is defined in binary form, where The default value is 0.5; probability mask It is used to retain continuous probability information of the boundary transition region and to be used for subsequent effective bonding region screening and probability-weighted area calculation.

[0219] S104: Map the effective bonding area to the standardized triangular mesh, and accumulate the three-dimensional surface area of ​​the corresponding mesh facets to obtain the effective bonding area evaluation result of the target abutment tooth proximal surface.

[0220] It should be noted that not all anatomically complete proximal areas are suitable for bonding with a laminar flow plate. The edge of the laminar flow plate near the gingival margin can irritate periodontal tissues, undercuts along the path of insertion cannot be covered by the laminar flow plate, and laminar flow plates near the incisal or occlusal ends may affect aesthetics or interfere with occlusion. This step translates the above clinical requirements into spatial constraints one by one, finds their intersection with the proximal areas, and retains the effective bonding area (EBA) that meets all the conditions.

[0221] As shown in Figure 3, taking the mesial surface of the maxillary second premolar as an example, five clinical constraints are sequentially applied to the three-dimensional proximal surface segmentation results in the form of a probability mask. The candidate regions are progressively narrowed down through point-by-point product operations, ultimately yielding an effective bonding region that meets all clinical requirements. This process can be represented as:

[0222] ;

[0223] in: This represents the probability mask of the final effective bonding area; Represents the probability mask for adjacent face segmentation; Represents the glaze boundary constraint mask; Indicates the gingival margin safety distance constraint mask; Represents the boundary constraint mask for the cutting edge or the biting side; This indicates the indentation mask to exclude constraints in the in-situ path. This represents element-wise multiplication. The advantage of using probability multiplication instead of hard-threshold Boolean operations is that the transition regions near each constraint boundary reflect uncertainty with a gradual change in probability. For example, the enamel probability near CEJ can continuously transition from 1.0 to 0.0, avoiding area jumps caused by hard boundaries and making the result more stable for small offsets at the boundary positions.

[0224] Glaze boundary constraints In this study, the resin bonding strength of enamel is much higher than that of dentin and cementum; therefore, the effective bonding area should be limited to the enamel region. When CBCT data is available, the enamel volume is extracted through grayscale thresholding, and the cementoenamel junction (CEJ) is mapped to the abutment tooth surface model via iterative nearest-neighbor registration. When only IOS data is available, a pre-trained graph convolutional regression network predicts the signed distance from each surface point to the CEJ, and the zero contour line represents the CEJ location. The coronal region of the CEJ is marked as the enamel area. The value approaches 1; CEJ marks the gingival region as a non-enamel area. The value approaches 0; the probability of CEJ changes gradually to reflect the uncertainty of positioning.

[0225] Gingival margin safety distance constraint In this process, the gingival margin position is extracted through curvature analysis or a pre-trained network, and then offset coronally by a safe distance. The default value is 0.5 mm, and the offset line's outermost marker is the safety zone. The safety distance... It is used to prevent the edge of the wing plate from getting too close to the gingival margin, which can cause plaque buildup and gingivitis.

[0226] Boundary constraints between the cutting edge and the biting side In the middle, the effective zone is marked by the surface contour line 2 to 3 mm below the incisal edge for anterior teeth and 1 to 2 mm below the marginal ridge for posterior teeth. This prevents the wing plate from extending excessively towards the incisal or occlusal surface.

[0227] In-place path undercut to eliminate constraints In the middle, the direction of the positioning path is given. A region where the angle between the surface normal and the placement direction is greater than 90° constitutes an undercut, and the restoration cannot be positioned and covered in this direction, can be written as:

[0228] ;

[0229] in: Indicates the first Does each sampling point satisfy the no-inverted-concave constraint? Indicates the first The external normal vector of each sampling point; The unit vector representing the direction of the in-place path; This represents the indicator function. This formula only retains the region where the normal vector is at an acute angle to the placement direction.

[0230] Since the orientation of the placement path directly affects the usable area, an optimal orientation is automatically searched to maximize the bonding area. However, if only area is considered as the objective, an orientation may result in the EBA appearing as a thin, forked, or fragmented shape; although the area of ​​this region is large, the stress concentration in this region makes it prone to wing plate detachment in clinical settings. Therefore, a morphological compactness penalty term is added to the optimization objective, written as:

[0231] ;

[0232] in: Indicates the optimal placement path direction; Indicates the direction of the candidate placement path; Represents a unit sphere; Indicates direction The effective bonding area obtained below; Indicates direction The perimeter of the EBA boundary below; and For weighting, the perimeter penalty makes the algorithm tend to choose directions that produce a rounded and compact EBA, rather than directions with a slightly larger area but an irregular shape. To maintain strict dimensional consistency, the area and perimeter can be normalized, or... It has the dimension of length.

[0233] Subsequently, the system maps the adjacent face segmentation results determined by the automatic segmentation mode, the confirmed segmentation mode, or the fine interactive segmentation mode in S103 to the face of the triangular mesh M′ through the sampling point-face index mapping table, and inputs them into the area calculation module.

[0234] Then, the point-by-point mask is mapped to the triangular mesh using a sampling point-patch index mapping table. On the facet, input the next area calculation module. After EBA definition is complete, the system calculates the area on the triangular mesh. The system calculates the three-dimensional surface area and outputs multi-dimensional quality evaluation indicators. To ensure that the area values ​​smoothly reflect the continuous probability of boundary transition regions in the EBA mask, the system calculates the probability weight of each triangular facet belonging to the EBA, multiplies it by the facet's physical area, and then sums the results.

[0235] ;

[0236] ;

[0237] in: Indicates the first The probability weights of each facet; , , These represent the probability values ​​of the three vertices of the face in the EBA mask; Indicates the first A triangular facet; This represents the set of faces involved in the area calculation; , , Represents the three vertices of the face; Represents a piece of dough The three-dimensional physical area is calculated using the following formula:

[0238] ;

[0239] in: Represents the cross product of vectors; This represents the Euclidean norm of the vector. From this, the effective bonding area assessment result of the target abutment tooth proximal surface can be obtained.

[0240] In a preferred embodiment, in addition to outputting the effective bonding area value, the system can further calculate quality evaluation indicators related to regional morphology, such as regional boundary perimeter, regional continuity, regional compactness, and segmentation confidence level, and organize the effective bonding area and quality evaluation indicators into a clinical evaluation report. This evaluation report can be used to assist restorative dentists in determining whether the proximal surface of the target abutment tooth has sufficient usable bonding area, and accordingly adjust the abutment tooth selection, apron coverage, and placement direction in the Maryland bridge restoration plan.

[0241] In summary, this invention, based on the anatomical segmentation results of the proximal surfaces, incorporates clinical conditions such as enamel boundaries, gingival margin safety distances, incisal or occlusal boundaries, and exclusion of undercuts along the path of insertion into the determination process of the effective bonding area. This ensures that the final calculated effective bonding area is no longer simply the anatomical proximal surface area, but rather a usable bonding area consistent with the actual wing plate coverage conditions of the Maryland bridge. Furthermore, through candidate placement orientation optimization and probability-weighted area calculation, the obtained area results have higher reliability in both clinical significance and geometric stability.

[0242] Example 2

[0243] This embodiment, based on Embodiment 1, analyzes data from one subject. This patient had a congenital absence of the left maxillary lateral incisor (International Dental Federation tooth number FDI 22), and the clinical plan was to use a resin-bonded fixed bridge for restoration. The maxillary left canine (FDI 23) was used as the abutment tooth, and the wing plate was bonded to the mesial proximal surface of tooth 23. This embodiment fully demonstrates the entire process from intraoral scan data, through model standardization, abutment tooth positioning and proximal surface completion, precise proximal surface segmentation, effective bonding area definition, to the output of bonding area and quality evaluation.

[0244] (I) Acquisition and Standardization of 3D Models

[0245] An intraoral optical scanner (IOS) was used to scan the patient's entire maxillary dental arch, outputting an STL format triangular mesh file containing approximately 743,120 triangular faces, covering 13 teeth (11-17, 21, 23-27) and the surrounding gingiva. A gap in the edentulous space was visible at tooth position 22. Figure 4 As shown, the patient also underwent cone-beam computed tomography (CBCT) before the operation, which was used to determine the enamel boundary. There was an edentulous gap at position 22, and the mesial surface of tooth 23 faced the gap.

[0246] (1) Construction of the standardized grid M′

[0247] Three preprocessing steps were performed on the original mesh sequentially: ① Bilateral filtering denoising, with the spatial domain standard deviation set to 0.3 mm and the signal domain standard deviation set to 0.2 mm, iterated 3 times. After processing, the surface noise amplitude decreased from ±0.025 mm to ±0.008 mm, while the curvature peak at the corner of tooth 23 remained clear. ② Mesh repair, filling 3 micro-holes, eliminating 7 non-manifold edges and 12 degenerate triangles, making the mesh satisfy the second manifold condition; ③ Coordinate normalization, establishing a unified coordinate system with the interlocking plane as a reference. The resulting normalized mesh M′, containing approximately 516,230 faces, served as the geometric reference for all subsequent area and distance calculations.

[0248] (2) Construction of uniform sampling point cloud P

[0249] 80,000 uniformly distributed points are extracted from M′ using Poisson disk sampling, with a minimum point spacing of approximately 0.12 mm. Each sampling point records its nearest patch index and centroid coordinates, forming a sampling point-patch index mapping table, which is used to convert the point-by-point segmentation results into the physical area on the mesh patch.

[0250] (3) Point-by-point feature calculation

[0251] For each sampling point, a 10-dimensional feature vector is extracted, including three-dimensional coordinates (3D), normal vector (3D), maximum principal curvature κ1 (1D), minimum principal curvature κ2 (1D), mean curvature H (1D), and Gaussian curvature K (1D). Curvature is calculated using a multi-scale neighborhood weighted fitting scheme: weighted least-squares quadratic surface fitting is performed at three neighborhood radii of 0.5mm, 1.0mm, and 1.5mm, respectively. Gaussian weighted averages are used to fuse the estimates at each scale. The reference scale is 1.0mm, the bandwidth is 0.5mm, and the fusion weights for the three scales are approximately 0.607, 1.000, and 0.607, respectively. Taking a sampling point at the mesial labial corner of tooth 23 as an example, the estimated maximum principal curvature values ​​at the three scales are 1.6, 1.2, and 0.9 mm⁻¹, respectively, and the fused value is 1.24 mm⁻¹, which is significantly higher than the typical value of 0.3-0.6 mm⁻¹ at the center of the adjacent surface. This difference is the key geometric clue for the subsequent network to determine the boundary of the adjacent surface.

[0252] (ii) Proximal restoration of abutment teeth

[0253] Automatic segmentation of the dental arch

[0254] The PointNet++ segmentation network performed 33-class semantic segmentation on 80,000 points in the entire dental arch. The network consists of 4 ensemble abstraction layers (neighborhood radii of 2, 4, 8, and 16 mm respectively) and 4 feature propagation layers. After softmax to select the class with the highest probability, isolated connected components with fewer than 50 points are merged into the surrounding majority class. Approximately 4200 sampling points were obtained for tooth 23, and its corresponding mesh subset M′_T contains approximately 6800 triangular faces.

[0255] Integrity assessment of the proximal surface of the abutment tooth: Based on clinical experience, the dentist determined that tooth 23 was the abutment tooth. According to the physical adjacency rules, tooth 24, which is distal to tooth 23, is included in the adjacent tooth set N(23)={24}. Due to the absence of tooth 22, the mesial proximal surface of tooth 23 faces the open edentulous space, and data loss is likely to occur in this area during intraoral scanning due to the limited scanning angle. The system performed three automatic tests on the mesial proximal surface area (approximately 850 sampling points):

[0256] Sampling density detection: ① The average surface density of adjacent areas ρ̄≈6.8 points / mm², which is higher than the minimum threshold of 5 points / mm², so it passes.

[0257] Hole area percentage detection. A hole was detected on the labial side of the gingival margin, with an area A_hole≈0.2mm². The estimated total area of ​​the adjacent surfaces A_prox≈14.5mm² is less than the allowable threshold of 15%, so it passes the test.

[0258] Morphological anomaly detection. The reconstruction error of the pre-trained autoencoder, e_recon≈0.09mm, is below the threshold of 0.15mm, and no scanning artifacts were detected, so the test passed.

[0259] (iii) Precise division of the proximal surfaces of the abutment teeth

[0260] The 3D Dental-SAM module of this invention solves this problem through the collaborative work of three components. Component A, the surface feature encoder, first aligns the tooth long axis, mesiodistal direction, and labiolingual direction of the 23rd tooth point cloud to the Z, X, and Y axes respectively through principal component analysis, eliminating differences in scanning position. The encoder adopts an improved Point Transformer V3 architecture, containing four encoding layers. The point cloud is downsampled layer by layer at a 4x ratio. Each layer simultaneously uses four spatial filling curves with different configurations (two Z-order curves and two Hilbert curves, each with a different starting orientation) to independently complete the local sequential attention calculation and fusion results with a window size W=64. The feature dimension is 256. The output is features at four scales.

[0261] The pyramids are interconnected at different scales through jumps.

[0262] The system initially runs in fully automatic mode (without physician prompts). The mask decoder (component C) processes the surface feature pyramid and three learnable output query tokens (corresponding to three granularities: complete neighboring face, neighboring face core region, and neighboring face boundary zone, respectively) through two iterative layers. Each layer sequentially includes self-attention fusion, cross-attention from query to surface, reverse cross-attention from surface to query, and a point-by-point feedforward network—outputting near-neighboring face probability masks and corresponding confidence scores for the three granularities.

[0263] The system first performs interactive 3D segmentation without physician input. The mask decoder processes the surface feature pyramid, cue embedding matrix, and three learnable output query tokens through two iterative layers, outputting near-neighbor probability masks and corresponding confidence scores at three granularities. Granularity 1 corresponds to the complete neighbor, granularity 2 to the neighbor core region, and granularity 3 to the neighbor boundary zone. This embodiment uses the confidence score corresponding to granularity 1. Used as the confidence level for the current adjacent segmentation.

[0264] in: This represents the confidence level corresponding to granularity 1, i.e., the complete adjacent face segmentation result, with a value ranging from 0 to 1; The larger the value, the higher the system's assessment of the reliability of the current complete adjacent face segmentation result.

[0265] In this example, the system presets a first threshold. It falls within the range of 0.83 to 0.95; a second preset threshold is used. It falls within the range of 0.69 to 0.82, and .in: This represents the first confidence threshold corresponding to the automatic segmentation mode; This represents the second confidence threshold corresponding to the fine-grained interactive segmentation mode.

[0266] The segmentation confidence level of granularity 1 in this example ,satisfy:

[0267]

[0268] Therefore, the system did not directly use the automatic segmentation results, nor did it require physicians to perform multi-point fine-tuning interactions. Instead, it automatically entered the segmentation confirmation mode. In the segmentation confirmation mode, the physician observed the initial adjacent surface segmentation results on the 3D visualization interface and found that the lip-side boundary was offset from the expected position by approximately 0.8 mm towards the inner lip surface, and did not yet completely encompass the adjacent surface area near the lip-side corner line. The physician entered a positive prompt point near the mesial lip-side corner line, and the system automatically extracted the radius of that clicked location. Sampling points within the spherical neighborhood are processed by a shared MLP to extract features point-by-point, then aggregated into a local context vector using MaxPool. This vector is then concatenated with the coordinates, normal vector, and curvature of the clicked point and encoded as a point cue embedding. The mask decoder receives the updated cue embedding matrix and re-outputs the adjacent surface probability mask. The lip boundary is then extended to the corner position, consistent with the physician's expectations, and the physician confirms the adjacent surface segmentation result.

[0269] (iv) Constraints and delineation of the effective bonding area

[0270] The system immediately and automatically performs five clinical constraint calculations on the same interface.

[0271] Constraint 1: Adjacent-face segmentation mask The nearest neighbor probability mask output from the previous 3D Dental-SAM step. (Middle of the neighboring face) Approaching 1.0, outside the corner line Approaching 0, with a smooth transition near the corner line.

[0272] Constraint 2: Glaze Boundary The bonding strength of resin adhesives to acid-etched enamel (approximately 20 to 30 MPa) is significantly higher than that to dentin (approximately 8 to 15 MPa), and effective bonding should be limited to the enamel coverage area. In this example, CBCT data was used to determine the cementoenamel junction (CEJ) location: the enamel volume was segmented using a grayscale threshold (approximately 1200 grayscale threshold on the KaVo OP 3D device used in this example), and the gingival boundary surface of the enamel volume, i.e., the CEJ surface, was extracted and mapped to the abutment tooth surface using an iterative nearest-point algorithm (registration error 0.12 mm). CEJ coronal gingival , The internal sigmoid pathway is gradual. In this case, the CEJ is located approximately 1.5 mm below the gingival margin, with an exclusion area of ​​approximately 1.2 mm².

[0273] Constraint 3: Safe distance from the gingival margin The edge of the wing plate should not be close to the free gingival margin, otherwise it will hinder gingival sulcus cleaning and promote plaque accumulation. The system automatically detects the gingival margin line through the curvature sign reversal feature and offsets it coronally along the tooth surface. .default Offset line crown The sigmoid line gradually changes from the offset line to the gingival margin, and the gingival side... The area excluded is approximately 3.8 mm².

[0274] Constraint 4: Tangential Boundary Tooth #23 is a maxillary canine, and the wing plate should not extend excessively to the vicinity of the cusp. This area bears functional contact forces during protrusive occlusion and may be exposed during smiling on the labial proximal incisal region. Measurements are taken gingivally along the long axis of the tooth, using the cusp apex as a reference. The intersection of this horizontal section and the adjacent plane forms a tangential constraint boundary. (Boundary gingival side) The coronal sigmoid gradient is set to 0. The excluded area is approximately 4.1 mm². The texture is displayed in the interface with blue diagonal lines filling the area. It can be adjusted through the parameter panel.

[0275] Constraint 5: Orientation of the path of insertion. Near the long axis of the tooth. Within the cone angle range Step-size sampling yields approximately 592 candidate directions, based on the objective function.

[0276]

[0277] The optimal search direction is as follows: For effective bonding area (mm²), The perimeter of the boundary (mm) Perimeter penalty coefficient (make (with dimensions in mm²). In this example, the optimal direction is approximately labial along the tooth's long axis. The traversal calculation took approximately 0.15 seconds.

[0278] The five constraints are found by intersecting the element-wise probabilities:

[0279] ;

[0280] Taking a point at the center of the adjacent surface as an example, the values ​​are 0.97, 1.00, 1.00, 1.00, and 1.00, with a product of 0.97; taking a point near the gingival margin as an example, the values ​​are 0.92, 0.85, 0.40, 1.00, and 1.00, with a product of 0.31, and this point contributes to the area with a weight of 0.31.

[0281] (v) Real-time interactive adjustment by physicians

[0282] like Figure 5 As shown, the physician observed the fully automatic segmentation results on the 3D visualization interface and found that the lip-side boundary was offset towards the inner side of the lip surface by approximately 0.8 mm, and the segmentation did not completely encompass the adjacent surface region near the lip-side corner line. The physician clicked a positive indicator point near the mesial lip-side corner line, and the system automatically extracted the radius of that clicked location using component B. The sampling points (approximately 45 points) within the spherical neighborhood are used to extract features point-by-point through a shared MLP, and then aggregated into a 256-dimensional context vector using MaxPool. After being concatenated with the coordinates, normal vector, and curvature of the click point, it is processed by MLP point Encoded as 256-dimensional cue embeddings, overlaid with learnable positive cue type embeddings. The mask decoder receives the updated cue embedding matrix and completes forward inference in approximately 85 milliseconds. The lip-side boundary extends outward to the corner line position, consistent with the physician's expectation, and the physician confirms the result.

[0283] (vi) Validation of area calculation results and methods

[0284] After the above full-process processing, the system outputs the effective bonding area A_EBA = 12.918 mm², boundary perimeter L_perim = 15.1 mm, region compactness (4π·A / L²) = 0.712, and segmentation confidence level = 0.94 for the mesial proximal surface of tooth 23. The entire calculation process takes approximately 2.1 seconds (including 0.15 seconds for path optimization).

[0285] To verify the repeatability and accuracy of the method, the following comparative experiments were conducted:

[0286] (1) Intra-operator repeatability: The same physician performed the same operation on the same case 5 times at different times. In the fully automatic mode, the output area of ​​the 5 times was 12.92 mm² (standard deviation 0.00 mm², coefficient of variation CV < 0.1%). In the single-click confirmation mode, the outputs of the 5 times were 12.92, 13.15, 12.76, 12.89 and 12.83 mm² (standard deviation 0.15 mm², CV = 1.1%).

[0287] (2) Comparison with manual measurement: Three dentists with more than 5 years of restorative experience were invited to manually select and calculate the area of ​​the proximal surface of the same abutment tooth using the general three-dimensional measurement software (Geomagic Wrap 2021). The results were 14.52, 12.83, and 13.76 mm², respectively, with an inter-operator mean of 13.70 mm², a standard deviation of 0.85 mm², and a coefficient of variation (CV) of 6.2%. The output value of this method (12.92 mm²) was about 5.7% lower than the manual mean. The reason for this was that this method strictly applied the gingival margin safety distance (0.5 mm) and enamel boundary constraint, while manual selection tended to include some areas near the gingival margin and the cementoenamel junction. This systematic bias is consistent with the clinical conservative principle.

[0288] (3) Verification of multiple tooth positions: Ten additional cases with different tooth positions (2 cases of maxillary lateral incisors and 2 cases of mandibular lateral incisors) were selected, with effective bonding areas ranging from 8.34 to 16.71 mm². In a case-by-case comparison between the output area of ​​this method and the mean of manual measurements by three physicians, the mean absolute deviation was 0.82 mm², and the systematic bias of this method was low (mean deviation -0.61 mm²). The largest single-case deviation occurred in a case of mandibular premolar (9.47 mm² in this method, 11.23 mm² in manual measurements, deviation -1.76 mm²). Analysis showed that this case had a large lingual undercut area that was excluded by this method but not by manual measurements. The intragroup correlation coefficient (ICC) was 0.91 (95% confidence interval 0.74-0.97). The results indicate that this method is generally consistent with the judgment of clinical experts, and the direction of systematic bias conforms to the principle of conservative clinical design. The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for evaluating the effective bonding area of ​​proximal surfaces of abutment teeth in Maryland bridges, characterized in that, Includes the following steps: A three-dimensional digital model of the oral cavity containing the target abutment tooth is obtained. The three-dimensional digital model of the oral cavity is standardized to construct a standardized triangular mesh. Sampling point cloud is extracted from the surface of the standardized triangular mesh, and the correspondence between sampling points and mesh patches is established. Tooth position segmentation is performed on the sampled point cloud. The target abutment tooth and its adjacent teeth are determined according to clinical input. The surface data of the target abutment tooth are extracted, and the integrity of the candidate area of ​​the proximal surface of the target abutment tooth is detected. The integrity detection includes sampling density detection, cavity area ratio detection, and morphological abnormality detection. When there are missing or distorted areas in the candidate area of ​​the proximal surface, the proximal surface data is completed based on the proximal surface data of the target abutment tooth, the spatial relationship of the adjacent teeth, and the tooth position information of the target abutment tooth. Based on the surface data of the target abutment tooth, interactive three-dimensional segmentation is performed using a segmentation model that includes a surface feature encoder, an interactive prompt encoder, and a mask decoder to obtain the proximal surface segmentation results of the target abutment tooth. The proximal surface segmentation results are constrained and filtered according to the enamel boundary, gingival margin safety distance, incisal boundary or occlusal boundary, and undercut exclusion conditions of the path of insertion to obtain the effective bonding area. The adjacent candidate regions are determined in the following way: Determine the mesial or distal direction of the target abutment tooth based on its position information; Based on the long axis direction of the target abutment tooth, the contact relationship between adjacent teeth, or the location of the edentulous space, candidate surface points facing adjacent teeth or edentulous spaces are extracted from the surface data of the target abutment tooth. By combining the normal vector changes and multi-scale fused curvature values ​​of the candidate surface points, surface points corresponding to the labial, lingual, incisal, or occlusal surfaces are eliminated to obtain the candidate region of the proximal surface of the target abutment tooth. The integrity detection is performed on the adjacent candidate region, which includes sampling point distribution detection, mesh hole detection, and local surface continuity detection. When there is insufficient sampling point distribution, hole boundary loop located inside the adjacent candidate region, or the normal vector and curvature change do not satisfy the local surface continuity relationship, it is determined that the adjacent candidate region is missing or distorted. The effective bonding area is mapped to the standardized triangular mesh according to the correspondence, and the three-dimensional surface area of ​​the mesh patches belonging to the effective bonding area is calculated to obtain the effective bonding area evaluation result of the target abutment tooth proximal surface.

2. The method for evaluating the effective bonding area of ​​the proximal surfaces of abutment teeth in a Maryland bridge according to claim 1, characterized in that, The standardization process includes: The oral cavity 3D digital model was subjected to denoising, mesh repair and coordinate normalization in sequence to obtain a topologically continuous standardized triangular mesh; A uniformly distributed cloud of sampling points is extracted from the standardized triangular mesh, wherein each sampling point records its corresponding mesh patch index. The adjacent face segmentation results obtained from the sampled point cloud can be reflected back to the triangular mesh to perform area calculation.

3. The method for evaluating the effective bonding area of ​​the proximal surfaces of abutment teeth in a Maryland bridge according to claim 2, characterized in that, Each sampling point in the sampling point cloud is configured with point-by-point geometric features, which include three-dimensional coordinates, normal vectors, curvature features, and corresponding mesh patch indices. The curvature features are obtained by fitting the local surfaces around the sampling points in different scale neighborhoods, and the curvature estimation results at different scales are weighted and fused to generate a point-by-point geometric feature table containing multi-scale fused curvature values.

4. The method for evaluating the effective bonding area of ​​the proximal surfaces of abutment teeth for Maryland bridges according to claim 1, characterized in that, The adjacent data completion includes: The incomplete set of proximal points of the target abutment tooth, the set of points of adjacent teeth physically adjacent to the target abutment tooth, and the tooth position identifier of the target abutment tooth are jointly input into the morphological completion model. The incomplete set of proximal points of the target abutment tooth is used to provide local geometric trends, and the set of proximal tooth points is used to provide spatial adjacency constraints. The target abutment tooth position identifier is used to provide the morphological prior of the corresponding tooth position, and the completed proximal surface data is generated. The completion results are then subjected to boundary continuity verification and adjacent junction rationality verification. After the verification is passed, the completed data is fused with the original adjacent surface data.

5. The method for evaluating the effective bonding area of ​​the proximal surfaces of abutment teeth in a Maryland bridge according to claim 1, characterized in that, Interactive 3D segmentation includes: Extract multi-scale surface geometric features from the target abutment tooth surface data; Receive interactive prompts input by the physician and convert the interactive prompts into prompt features; The multi-scale surface geometric features and cue features are used together to determine the proximal surface region of the target abutment tooth, and the proximal surface segmentation result is output. The adjacent face segmentation results include adjacent face deterministic segmentation results and adjacent face probabilistic segmentation results. The adjacent face probabilistic segmentation results are used for subsequent boundary transition calculations of the effective bonding area.

6. The method for evaluating the effective bonding area of ​​the proximal surfaces of abutment teeth in a Maryland bridge according to claim 5, characterized in that, The interactive prompt information includes at least one of positive prompt points, negative prompt points, boundary key points, and prior masks; For positive and negative prompts, extract the local geometric features of the sampling points in the neighborhood of the prompt location, and combine them with the prompt location coordinates, prompt location normal vector and prompt type to generate point prompt features; For boundary key points, establish a continuity relationship according to the arrangement order of multiple boundary key points on the target tooth surface, and generate boundary cue features to characterize the orientation of adjacent boundary. For the prior mask, based on the region markings of each sampling point in the prior mask, a point-by-point region hint feature corresponding to the sampling point on the target abutment tooth surface is generated; At least one of the point hint features, boundary hint features, and point-by-point region hint features is combined to obtain interactive hint features for adjacent area determination.

7. The method for evaluating the effective bonding area of ​​the proximal surfaces of abutment teeth in a Maryland bridge according to claim 1, characterized in that, The interactive 3D segmentation also includes: The segmentation mode is adaptively determined based on the segmentation confidence of the interactive 3D segmentation output; When the segmentation confidence level is higher than the first threshold, an automatic segmentation mode is adopted; When the segmentation confidence level is between the first threshold and the second threshold, the confirmation segmentation mode is adopted; When the segmentation confidence level is lower than the second threshold, a fine-grained interactive segmentation mode is adopted so that standard cases and complex cases correspond to different interaction intensities.

8. The method for evaluating the effective bonding area of ​​the proximal surfaces of abutment teeth for Maryland bridges according to claim 1, characterized in that, The constraint filtering includes: Glaze coverage constraints are generated based on glaze boundaries to preserve adjacent face regions within the glaze coverage area; Gingival margin safety constraints are generated based on the gingival margin line and its coronal safety distance, and are used to exclude adjacent areas close to the gingival margin. Generate incisal boundary constraints or occlusal boundary constraints based on the target abutment tooth type to limit the effective bonding area from extending to the incisal or occlusal side; Undercut exclusion constraints are generated based on the preset placement path direction and the surface normal vector of the target abutment tooth, which are used to exclude undercut areas that cannot be covered by the restoration along the preset placement path direction. The effective bonding area is obtained by intersecting the enamel coverage constraint, gingival margin safety constraint, incisal boundary constraint or occlusal boundary constraint and undercut exclusion constraint with the adjacent surface segmentation results.

9. The method for evaluating the effective bonding area of ​​the proximal surfaces of abutment teeth for Maryland bridges according to claim 8, characterized in that, Area calculation includes: Multiple candidate placement path directions are set within a preset range near the long axis of the target abutment tooth. The effective bonding area corresponding to each candidate placement path direction is calculated. An optimization target is constructed based on the effective bonding area and the perimeter of the effective bonding area boundary to determine the target placement path direction. The effective bonding area obtained under the target placement path direction is then mapped onto a standardized triangular mesh; Calculate the weight value of each mesh surface that belongs to the effective bonding area, and multiply the weight value by the three-dimensional surface area of ​​the corresponding mesh surface and sum them to obtain the effective bonding area evaluation result of the target abutment tooth proximal surface.

Images