WO2016122890A1 - Estimating and predicting tooth wear using intra-oral 3d scans - Google Patents

Estimating and predicting tooth wear using intra-oral 3d scans Download PDF

Info

Publication number
WO2016122890A1
WO2016122890A1 PCT/US2016/013329 US2016013329W WO2016122890A1 WO 2016122890 A1 WO2016122890 A1 WO 2016122890A1 US 2016013329 W US2016013329 W US 2016013329W WO 2016122890 A1 WO2016122890 A1 WO 2016122890A1
Authority
WO
WIPO (PCT)
Prior art keywords
teeth
digital model
tooth
module
wear
Prior art date
Application number
PCT/US2016/013329
Other languages
French (fr)
Inventor
Evan J. Ribnick
Guruprasad Somasundaram
Brian J. Stankiewicz
Aya EID
Ravishankar Sivalingam
Shannon D. SCOTT
Anthony J. SABELLI
Robert D. Lorentz
Original Assignee
3M Innovative Properties Company
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 3M Innovative Properties Company filed Critical 3M Innovative Properties Company
Priority to AU2016212031A priority Critical patent/AU2016212031B2/en
Priority to DK16743844.9T priority patent/DK3250111T3/en
Priority to EP20166109.7A priority patent/EP3708069A1/en
Priority to EP16743844.9A priority patent/EP3250111B1/en
Publication of WO2016122890A1 publication Critical patent/WO2016122890A1/en

Links

Classifications

    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/45For evaluating or diagnosing the musculoskeletal system or teeth
    • A61B5/4538Evaluating a particular part of the muscoloskeletal system or a particular medical condition
    • A61B5/4542Evaluating the mouth, e.g. the jaw
    • A61B5/4557Evaluating bruxism
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/0059Measuring for diagnostic purposes; Identification of persons using light, e.g. diagnosis by transillumination, diascopy, fluorescence
    • A61B5/0062Arrangements for scanning
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/0059Measuring for diagnostic purposes; Identification of persons using light, e.g. diagnosis by transillumination, diascopy, fluorescence
    • A61B5/0082Measuring for diagnostic purposes; Identification of persons using light, e.g. diagnosis by transillumination, diascopy, fluorescence adapted for particular medical purposes
    • A61B5/0088Measuring for diagnostic purposes; Identification of persons using light, e.g. diagnosis by transillumination, diascopy, fluorescence adapted for particular medical purposes for oral or dental tissue
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/45For evaluating or diagnosing the musculoskeletal system or teeth
    • A61B5/4538Evaluating a particular part of the muscoloskeletal system or a particular medical condition
    • A61B5/4542Evaluating the mouth, e.g. the jaw
    • A61B5/4547Evaluating teeth
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/72Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235Details of waveform analysis
    • A61B5/7264Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/72Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7271Specific aspects of physiological measurement analysis
    • A61B5/7275Determining trends in physiological measurement data; Predicting development of a medical condition based on physiological measurements, e.g. determining a risk factor
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/72Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7271Specific aspects of physiological measurement analysis
    • A61B5/7278Artificial waveform generation or derivation, e.g. synthesising signals from measured signals
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/74Details of notification to user or communication with user or patient ; user input means
    • A61B5/742Details of notification to user or communication with user or patient ; user input means using visual displays
    • A61B5/743Displaying an image simultaneously with additional graphical information, e.g. symbols, charts, function plots
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/0002Inspection of images, e.g. flaw detection
    • G06T7/0012Biomedical image inspection
    • G06T7/0014Biomedical image inspection using an image reference approach
    • GPHYSICS
    • G16INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16HHEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H30/00ICT specially adapted for the handling or processing of medical images
    • G16H30/40ICT specially adapted for the handling or processing of medical images for processing medical images, e.g. editing
    • GPHYSICS
    • G16INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16HHEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/50ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for simulation or modelling of medical disorders
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B2576/00Medical imaging apparatus involving image processing or analysis
    • A61B2576/02Medical imaging apparatus involving image processing or analysis specially adapted for a particular organ or body part
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61CDENTISTRY; APPARATUS OR METHODS FOR ORAL OR DENTAL HYGIENE
    • A61C9/00Impression cups, i.e. impression trays; Impression methods
    • A61C9/004Means or methods for taking digitized impressions
    • A61C9/0046Data acquisition means or methods
    • A61C9/0053Optical means or methods, e.g. scanning the teeth by a laser or light beam
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/10Image acquisition modality
    • G06T2207/10028Range image; Depth image; 3D point clouds
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/30Subject of image; Context of image processing
    • G06T2207/30004Biomedical image processing
    • G06T2207/30036Dental; Teeth

Definitions

  • Tooth wear associated with Bruxism
  • gingival recession are both conditions that, if not treated in a timely manner by dental professionals, can have serious medical consequences.
  • lateral movements and tooth grinding can cause significant tooth wear and lead to muscle pain, temporomandibular joint issues, and headaches. In some cases, this may lead to the dentin being exposed, dental decay, and even tooth fracture.
  • the tools available to dental professionals for diagnosing and assessing the severity of tooth wear and gingival recession are limited.
  • these tools include patient questionnaires, clinical examination by a dentist, and bite force measurements. Clinical examinations may be performed using the Individual Tooth-Wear Index, which provides a rating between 0 and 3 based on visual assessment by a dentist. Accordingly, a need exists for additional tools to assess tooth wear, particularly using intra-oral 3D scans.
  • a first method for estimating teeth wear includes receiving a 3D digital model of teeth and segmenting the 3D digital model of teeth to identify individual teeth within the 3D digital model of teeth.
  • a digital model of a tooth is selected from the segmented 3D digital model of teeth, and an original shape of the selected tooth is predicted to obtain a digital model of a predicted original shape.
  • the digital model of the tooth is compared with the digital model of the predicted original shape to estimate wear areas in the tooth.
  • a second method for predicting teeth wear includes receiving a 3D digital model of teeth.
  • a mapping function is applied to the digital model of the teeth based upon values relating to tooth wear. Wear areas in the teeth are predicted based upon the applying step.
  • FIG. 1 is a diagram of a system for detecting tooth wear using a 3D digital model based upon intra-oral 3D scans
  • FIG. 2 illustrates a 3D model of teeth from intra-oral scans
  • FIG. 3 illustrates a 3D model of teeth segmented to digitally separate each tooth
  • FIG. 4 is a flow chart of a method for segmenting teeth in a 3D digital model
  • FIG. 5 illustrates over-segmentation of teeth by a geometric hill-climbing method
  • FIG. 6 illustrates detection of boundary vertices between teeth in a 3D digital model
  • FIGS. 7A and 7B illustrate morphological clean up to fix boundaries between teeth in a 3D digital model
  • FIGS. 8 A and 8B illustrate removal of non-aligned boundaries in 3D digital model
  • FIGS. 9A-9H illustrate merging of the results of segmentation methods to segment teeth in a 3D digital model
  • FIG. 10 is a flow chart of a method of estimating tooth wear using 3D scans
  • FIG. 11 is a diagram of a user interface for displaying estimated tooth wear
  • FIG. 12 is a flow chart of a method of predicting tooth wear using 3D scans.
  • FIG. 13 is a diagram of a user interface for displaying predicted tooth wear.
  • Embodiments includes analyzing tooth wear from a single 3D scan of the patient's dentition.
  • One approach is based on estimating the original shape of the surface through use of a database (or collection) of known tooth shapes and learned mathematical model and for reconstructing shape, and then comparing this estimated original shape with the current shape of the surface.
  • Another approach is based upon comparing the current shape of the surface with annotated scans where the annotation indicates an amount of tooth wear.
  • FIG. 1 is a diagram of a system 10 for detecting tooth wear using a digital 3D model based upon intra-oral 3D scans.
  • System 10 includes a processor 20 receiving digital 3D models of teeth (12) from intra-oral 3D scans or scans of impressions of teeth.
  • System 10 can also include an electronic display device 16, such as a liquid crystal display (LCD) device, for displaying indications of tooth wear and an input device 18 for receiving user commands or other information.
  • LCD liquid crystal display
  • FIG. 2 An example of digital 3D model of a patient's teeth from a scan is shown in FIG. 2.
  • Systems to generate digital 3D images or models based upon image sets from multiple views are disclosed in U.S. Patent Nos.
  • System 10 can receive the 3D scans locally or remotely via a network.
  • the individual teeth in the model need to be segmented from one another before the desired analysis or manipulation can be performed.
  • a software interface may be presented in order for a user to perform this
  • FIG. 3 An example of teeth that have been segmented in a digital model is shown in FIG. 3.
  • the segmentation provides for separating individual teeth in the digital 3D model, as represented by the shading in FIG. 3, and each tooth in the model can essentially be digitally separated from the other teeth for further processing to detect tooth wear.
  • Using a segmented digital 3D model for comparing or analyzing individual teeth is more accurate than comparing whole or partial arches within the model.
  • Described herein are techniques for tooth segmentation within a digital 3D model.
  • the technique is a combination of two separate algorithms and combines the strengths of both of them.
  • the first algorithm is a geometric hill-climbing approach which takes into account topological structures such as height and curvature.
  • the second algorithm is a machine learning approach which classifies each point on the surface as belonging to either a boundary or a non-boundary.
  • the second algorithm is interstice detection which classifies a set of planes (or points) that approximate the intersticial spaces between teeth.
  • the second algorithm can be complementary to the first algorithm (geometric hill-climbing) and combined with the first algorithm to produce a resulting segmentation.
  • the first algorithm can be combined with user input estimating centroids of teeth in the digital 3D model. Instead of merging the results of two algorithms, only one algorithm can be used to segment the digital 3D model such as any one of the algorithms described herein.
  • the 3D scans addressed herein are represented as triangular meshes.
  • the triangular mesh is common representation of 3D surfaces and has two components.
  • the first component referred to as the vertices of the mesh, are simply the coordinates of the 3D points that have been reconstructed on the surface - i.e., a point cloud.
  • the second component the mesh faces, encodes the connections between points on the object and is an efficient way of interpolating between the discrete sample points on the continuous surface.
  • Each face is a triangle defined by three vertices, resulting in a surface that can be represented as a set of small triangular planar patches.
  • FIG. 4 is a flow chart of a method 22 for segmenting teeth in a digital 3D model.
  • Method 22 can be implemented in software or firmware modules, for example, for execution by processor 20.
  • Method 22 can alternatively be implemented in hardware modules or a combination of software and hardware.
  • Method 22 includes receiving a digital 3D model of a patient's teeth (step 24) and optionally aligning the model (step 25). Method 22 then involving segmenting the model by geometric hill-climbing (step 26) and point classification (step 28). Optionally, post processing on boundaries of the segmentation by point classification is performed (step 32). As an alternative to point classification, the model can be segmented by interstice detection (step 29). As another alternative to point classification, method 22 can receive user input identifying centroids of each tooth in the model (step 31).
  • the results of the segmentation methods are iteratively merged (step 30).
  • the results of segmentation by hill-climbing are merged with the results of segmentation by point classification or interstice detection or user input identifying the centroids.
  • the merged segmentation can optionally be refined based upon manual, for example user-entered, input (step 34).
  • the results of the segmentation are stored (step 38).
  • the segmentation results in a separate mesh for each tooth from the digital 3D model, as illustrated in FIG. 3.
  • the optional alignment step 25 can be implemented using a Support Vector Regression (SVR) method to find the occlusal plane fitted to a mesh of the teeth in the digital 3D model.
  • the alignment can be used to have the teeth in the digital 3D model essentially aligned with the Y axis.
  • the alignment can use the LIBSVM toolbox and £ ⁇ ⁇ method.
  • the training is based on the assumption that teeth are roughly pointing up along the Y axis.
  • the output is sample points from the occlusal plane which is given to a simple principal component analysis (PCA) method to find the normal direction.
  • PCA principal component analysis
  • SVR uses a linear loss function with a zero part within the margins which performs better for teeth dataset than the quadratic loss function in regular least square regression methods. It helps to decrease the effect of gingiva cut-lines which can be very jagged and bumpy in mesh scans.
  • Table 1 provides exemplary pseudocode for implementing the alignment step.
  • Input a 3D mesh with a set of vertices V specified in 3D coordinate system X,Y and Z.
  • Y represents the rough direction of vertical axis in which the teeth point upwards.
  • Output the normal vector perpendicular to occlusal plane which represents the correct upward direction of teeth.
  • Teeth are roughly pointing up along the Y axis. The mesh has been truncated below the gum line.
  • One of the algorithms for segmentation is based upon geometric operations on the mesh. Specifically, the main idea behind this approach is that, if one starts from any point on the surface and moves upwards through a series of points, one will converge to a high point that is a local maximum. In most cases it would be expected all points on a tooth (or on the same cusp of a tooth) will converge to the same local maximum. This type of segmentation can produce very accurate boundaries between teeth, but it typically results in an over- segmentation in which a single tooth may be divided into multiple segments.
  • the mesh is preprocessed using Laplacian smoothing. This preprocessing is an effective way of removing high-frequency noise in the surface reconstruction.
  • the energy function at each vertex is composed of two terms, where for the z ' -th vertex:
  • yi is the y-coordinate (height) of the z ' -th vertex
  • is its angular divergence
  • ⁇ > 0 is a weighting parameter.
  • the parameter ⁇ can be any value greater than zero or, alternatively, ⁇ can be equal to zero.
  • Angular divergence is a measure of overall curvature around a point.
  • a face F comprised of vertices v ; , vy, and vk, with normal vectors , tij, and m, respectively.
  • the angular divergence of the z ' -th vertex v is the mean of the angular divergences of the faces of which Vi is a part.
  • segmentation is performed according to a hill-climbing procedure.
  • the algorithm can be understood as follows. For each vertex on the surface, the algorithm initializes a hill-climb, in which at each iteration it moves to the connected neighbor (as defined by the faces) that has the highest energy function value. The algorithm continues climbing until it reaches a local maximum that has higher energy than all of its neighbors. All vertices that were passed through along this route are assigned to this local maximum, and all such paths that converge to this local maximum define a segment. This process is repeated until all vertices on the mesh have been traversed.
  • This segmentation assigns vertices to segments defined by local energy maxima that can be reached through a monotonically-increasing path through the energy function.
  • the energy function is defined such that each iteration of hill-climbing moves upwards in height, but is discouraged from crossing an area with high curvature by the angular divergence term. This helps ensure that the boundaries between teeth are not crossed.
  • FIG. 5 An example of a segmentation produced by this algorithm is shown in FIG. 5.
  • the algorithm over-segments the teeth by separating each cusp of a tooth into its own segment - this can be understood intuitively as a result of the hill-climbing procedure, since each cusp will have its own unique local maximum.
  • the digital model of tooth 40 is segmented into five sections.
  • the boundaries produced by this approach are quite precise and accurately separate teeth from one another.
  • Table 2 provides exemplary pseudocode for implementing the geometric hill- climbing algorithm.
  • Segmentation by Point Classification is a data-driven approach. Unlike the geometric hill-climbing approach, this approach relies on manually provided groundtruth segmentation. Groundtruth can be obtained from a user providing nearly accurate segmentation manually using mesh manipulation tools such as the MeshLab system. A selection of an individual tooth can be made using a face selection tool. Individual teeth are selected in this manner and saved as individual mesh files. Using the original mesh and the individual teeth files, a labeling of the vertices in the original mesh can then be inferred. Once groundtruth for a full scan is completed, the inferred labels of all the segments can be visualized.
  • mesh manipulation tools such as the MeshLab system.
  • a selection of an individual tooth can be made using a face selection tool. Individual teeth are selected in this manner and saved as individual mesh files. Using the original mesh and the individual teeth files, a labeling of the vertices in the original mesh can then be inferred. Once groundtruth for a full scan is completed, the inferred labels of all the segments can
  • the boundary vertices between segments can be determined. For each vertex the distribution of vertex labels around that vertex is examined. If the distribution is not unimodal (i.e., the vertex labels are predominantly the same), then that vertex is considered an interior vertex. If not, the vertex is considered a boundary vertex. This data can be manually entered one time, for example, as training data and then used repeatedly in the point classification algorithm.
  • the algorithm Given the groundtruth boundary vertices labels from multiple training meshes, the algorithm provides for a function that is capable of predicting whether a vertex on a mesh lies in the interior of a tooth or on the boundary between teeth.
  • the algorithm can classify or label points in the mesh as being on a tooth or on a boundary between teeth. This process involves two tasks: feature extraction and classification.
  • FIG. 6 illustrates detection of boundary vertices 42 between teeth in a digital 3D model.
  • Table 3 provides exemplary pseudocode for implementing the point classification (machine learning) training data algorithm.
  • Input Multiple 3D meshes with a sets of vertices V specified in 3D coordinate system X,Y and Z.
  • Y represents the vertical axis or the general direction in which the teeth point upwards.
  • the mesh also has a set of triangulations or faces F based on the vertices. Also the groundtruth segmentation in the form of the vertices corresponding to boundaries and those in the interior as indicated by manual annotation.
  • Output A predictive model that is capable of generating the boundary vertex prediction labels for a query set of vertices.
  • Teeth are roughly pointing up along the Y axis. The mesh has been truncated below the gum line.
  • the point classification algorithm extracts many characteristic features for every vertex in the mesh. It is often difficult to determine which features are useful in a segmentation algorithm.
  • features which can be used for segmentation in this framework including but not limited to multi-scale surface curvature, singular values extracted from PCA of local shape, shape diameter, distances from medial surface points, average geodesic distances, shape contexts, and spin images.
  • the algorithm implements the following features: absolute and mean curvature, direction of normal at vertex, local covariance of the mesh around the vertex and its principal Eigen values, spin images, Fourier features, shape contexts, and PCA features. Classification
  • the function f Given the feature set for a vertex X, the function f is defined as follows: f: X -> ⁇ 1,0 ⁇ , that is the function f maps the set of features X to either a 1 or 0. A value 1 indicates that vertex is a boundary vertex and the value 0 indicates otherwise.
  • This function can be one or a combination of many classification methods such as support vector machines, decision trees, conditional random fields, and the like. Additionally, in the segmentation as a classification problem, there is a class imbalance. The number of interior vertices is much greater than the number of boundary vertices. The ratio of interior vertices to boundary vertices is typically 100: 1. In such extreme class imbalance situations, regular classifiers are not optimal.
  • one option involves using classifier ensembles such as boosting.
  • the classification algorithm uses RUSBoosting on decision stumps as a classifier.
  • RUSBoost stands for random undersampling boosting and is known to handle the class imbalance very well. Additionally RUSBoost is already implemented in the MATLAB program "fitensemble" function. Based on preliminary analysis, RUSBoost was performed on 700 decision stumps. This number was chosen using cross-validation on the training set with the resubstitution loss as the metric. For our experiments, we used a "leave-scan-out" cross-validation scheme. Our dataset consisted of 39 scans, and for every test scan the remaining 38 scans were used for training. The resulting predictions were compared to the groundtruth boundary labels of the test scan. A confusion matrix can then be obtained by comparing the groundtruth labels with the predicted labels. From this we obtained the false alarm rate and the hit rate. With cross-validation testing on 39 scans we obtained an 80% hit rate and 1.7% false alarm rate on average.
  • Table 4 provides exemplary pseudocode for implementing the point classification (machine learning) algorithm.
  • Input a 3D mesh with a set of vertices V specified in 3D coordinate system X,Y and Z.
  • Y represents the vertical axis or the general direction in which the teeth point upwards.
  • the mesh also has a set of triangulations or faces F based on the vertices.
  • Output Binarized mesh where for each vertex Vi in the mesh, a label li corresponding to whether the vertex belongs to a boundary or not.
  • Teeth are roughly pointing up along the Y axis. The mesh has been truncated below the gum line.
  • the second algorithm for segmentation can use interstice detection (step 29 in method 22).
  • Table 5 provides exemplary pseudocode for implementing the interstice detection algorithm.
  • Input a 3D mesh with a set of vertices V specified in 3D coordinate system X,Y and Z.
  • Y represents the vertical axis or the general direction in which the teeth point upwards.
  • the mesh also has a set of triangulations or faces F based on the vertices.
  • Output a set of planes that approximate the intersticial spaces between each pair of teeth.
  • Teeth are roughly pointing up along the Y axis.
  • Intersticial spaces are identified as sample points that are local minima in the sum of heights computed in step 4.
  • the orientation of the intersticial space is given by the direction of the normal to the one-dimensional parameterization curve at the corresponding sample point.
  • Morphological operations such as mesh erosion and dilation can be done in tandem, resulting in an operation known morphological opening. Unlike images, mesh erosion and dilation are non-trivial since there are no sliding windows. Instead to perform mesh erosion, one can use the connected v-ring of every vertex as its neighborhood.
  • FIGS. 7 A and 7B illustrate morphological clean up to fix boundaries between teeth in a digital 3D model with FIG. 7B illustrating clean up of the boundaries shown in FIG. 7A. This morphological clean up can be used to for the optional step 32 in method 22 after the segmentation by point classification. Complementary Approaches to Segmentation
  • the hill-climbing approach captures the general geometry of cusp and has a tendency to form good boundaries around teeth, but it over-segments and thus creates more false boundaries.
  • the classification approach on the contrary has a somewhat less than desired hit rate on boundaries but has a very low false alarm rate.
  • a method to merge the results helps reduce the demerits of both approaches and boost the merits of both.
  • a hierarchical merging algorithm is used, which merges the segments in the hill-climbing approach using the boundary predictions of the classification approach. Every boundary predicted by the hill-climbing approach is given a score based on the predicted boundary vertices from the classification approach.
  • a hierarchical merging is performed. All the boundaries with a score less than a threshold are discarded and the corresponding segments are merged and the boundary scores are corrected accordingly.
  • This threshold is gradually increased. For example, all boundaries that have score less than 5 are discarded first. The corresponding segments are merged, and then this process is repeated by increasing the threshold step-by-step to, for example, 50. This heuristic provides correct segmentation of the teeth in one of the merge steps in most cases. Elimination of Non-Aligned Boundaries
  • FIGS. 8A and 8B illustrates removal of a boundary from the model of FIG. 8 A. This can be achieved by determining the principal direction of orientation of each boundary segment using PCA. The principal components (PCs) of each consecutive tooth boundary should be aligned, thus resulting in eliminating the boundaries which have misaligned PCs. This process is applied after merging the hill climbing result with the machine learning result.
  • FIGS. 9A-9H Sample results of the classification or machine learning (ML), hill-climbing (HC), and the merging steps are shown in FIGS. 9A-9H.
  • the machine learning output (FIG. 9 A) shows the mesh labeling for the boundary vertices and the interior vertices.
  • the second mesh (FIG. 9B) is the result of the hill climbing. As shown in FIG. 9B, the hill-climbing over-segments each tooth but in general there is a reduced chance of a segment being shared across teeth. This is also a behavior associated with the choice of the parameter ⁇ .
  • the meshes displayed in FIGS. 9C-9H indicate iteratively the result of each merge step.
  • Merge 1 corresponds to discarding boundaries with a score less than 5 and merge 2 corresponds to scores less than 10 and so on.
  • the correct segmentation was achieved at step 6.
  • Successive merge steps indicate how aggressively nearby segments are merged and, therefore, in some cases changes are only noticeable at later merge steps.
  • the score used for merging can represent, for example, the number of points classified as a boundary from the point classification algorithm within a particular vicinity of a boundary determined from the hill-climbing algorithm.
  • An exemplary score of 5 means at least 5 points classified as a boundary are within a particular vicinity of a boundary determined by the hill-climbing algorithm.
  • the particular vicinity used can be based upon, for example, empirical evidence, the typical width or size of a true boundary, or other factors.
  • the best result would be achieved earlier than the 6th merging step and it is possible to get an over-merged result at step 6.
  • an under-merged or over-segmented result can occur even after step 6.
  • a cursor control device and user interface a user could manually select ("click on") and merge the segments that require merging to extract the teeth correctly, for example.
  • the final segmented digital 3D model can then be stored in an electronic storage device for later processing.
  • Table 6 provides exemplary pseudocode for implementing the algorithm for merging hill-climbing segmentation with point classification (machine learning) segmentation.
  • Table 7 provides exemplary pseudocode for implementing the algorithm for merging hill-climbing segmentation with interstice detection segmentation.
  • Input a 3D mesh with a set of vertices V specified in 3D coordinate system X,Y and Z.
  • Y represents the vertical axis or the general direction in which the teeth point upwards.
  • the mesh also has a set of triangulations or faces F based on the vertices. Label assignments from hill climbing and boundary vertex labels predicted by the machine learning are also provided.
  • Output Segmented mesh, where for each vertex Vi in the mesh, a label li corresponding to the segment to which that vertex belongs is assigned.
  • Teeth are roughly pointing up along the Y axis. The mesh has been truncated below the gum line.
  • Table 7 - Pseudocode for Merging Hill-Climbing Segments using Interstice Detection Input a 3D mesh with a set of vertices V specified in 3D coordinate system X,Y and Z.
  • Y represents the vertical axis or the general direction in which the teeth point upwards.
  • the mesh also has a set of triangulations or faces F based on the vertices. Label assignments from hill climbing, as well as detected intersticial spaces, are also provided.
  • Output Segmented mesh, where for each vertex Vi in the mesh, a label li corresponding to the segment to which that vertex belongs is assigned.
  • Teeth are roughly pointing up along the Y axis.
  • the algorithm can merge the hill-climbing segmentation with user input identifying centroids of teeth (step 31 in method 22).
  • This segmentation method requires input from a user at the beginning of the process.
  • the user identifies the centroid of each tooth in the digital 3D model of teeth.
  • input device 18 such as a cursor control device
  • the centroid can include the actual centroid or an estimation of the centroid as perceived by the user.
  • This user entered information is used as the initialization for the step of the segmentation which merges the hill-climbing segments using the Kmeans method.
  • These user-identified centroids need to be close to actual centroids of the teeth in order for the segmentation process to work well and not require post-processing by the user.
  • the only parameter required for this method to be trained is ⁇ in SVR for normal direction extraction described above for the alignment process.
  • the user-entered information to identify the centroids of each tooth is then merged with the results of the hill-climbing segmentation using the Kmeans clustering method.
  • the vertices should first be replaced by the corresponding local maximum from the hill- climbing step.
  • Kmeans method is applied on the new set of vertices to cluster them in k segments, where k is equal to the number of inputs ("clicks") of the user at the beginning of the process.
  • the user's inputs (estimation of teeth centroids) are used as the centroid starting locations of the Kmeans method.
  • This merging method can result in successful segmentation as follows: clustering is applied on the local maxima (mostly located on the teeth cusps) and not the full mesh, yielding accuracy and speed benefits.
  • the local maxima of larger clusters find higher weights in Kmeans method, and the centroid starting locations entered by the user avoid converging to other possible local optima of Kmeans methods.
  • Table 8 provides exemplary pseudocode for implementing the algorithm for merging hill-climbing segmentation with user-entered estimations of teeth centroids.
  • a single scan of a patient's dentition for a person presumed to have tooth wear is compared to an estimated reconstruction of the original "virgin" tooth shape for each tooth or particular teeth (i.e., the shape before any tooth wear occurred), as predicted by a mathematical model as applied to the current patient's teeth.
  • This approach can be summarized according to the following steps: a 3D scan of the patient's teeth is acquired; the individual teeth are segmented from one another using a segmentation algorithm; for each tooth comprising the patient's dentition, the original "virgin" shape of the tooth is predicted using a mathematical model of tooth shape learned from a large database of teeth; and the current scan of each tooth is compared to the predicted virgin tooth shape in order to assess the degree of wear that has occurred.
  • FIG. 10 is a flow chart of a method 50 to implement this approach of estimating tooth wear using 3D scans.
  • Method 50 can be implemented in software or firmware modules, for example, for execution by processor 20.
  • Method 50 can alternatively be implemented in hardware modules or a combination of software and hardware.
  • Method 50 includes receiving a segmented model (step 52), for example from the segmentation methodology described above.
  • a tooth is selected from the model (step 54), and the original "virgin" shape of the tooth is predicted based upon a database of tooth scans (step 56).
  • a large database of 3D tooth models can be accessed and used, where the degree of wear in each of the teeth in the database is known.
  • an aggregate generic mathematical model can be formed of the canonical original "virgin” shape of a tooth location. This can be accomplished for each tooth.
  • multiple "virgin” models can be formed, which can include clustering of the space of tooth shape for that particular tooth location.
  • the original "virgin" shape of a tooth from the current scan can be predicted.
  • One approach is as follows. First, the appropriate model from the database for the current tooth is determined, since multiple clustered models may exist for each tooth location, depending on the variability of tooth shape at this bite location. If multiple models exist for this tooth, the appropriate model is determined by computing the similarity of this tooth to each model. Then, a mapping is computed from the current tooth to the model tooth shape. This mapping can be accomplished through use of a non-rigid registration algorithm. Then, once the new tooth is mapped to the model space, its original "virgin" shape is associated with that of the model. Using the inverse of the mapping estimated previously, this model is mapped back to the space of the current tooth, resulting in a prediction of the original shape of this tooth.
  • the original "virgin” tooth shape can be compared with the actual current shape in order to assess the amount of wear exhibited (step 60).
  • these two models may need to be registered (step 58), using a 3D registration algorithm, so that they are aligned with one another as closely as possible.
  • An example of a registration algorithm is disclosed in the application referenced above.
  • the areas in which the actual and predicted "virgin" models are in disagreement must be located and compared (step 62). These represent the areas of the tooth that have been worn.
  • Wear areas can optionally be estimated using local smooth contours on the tooth (step 63), and this estimation can be used to supplement the estimated wear areas from step 62.
  • local discontinuities in the tooth surface can be detected by analysis of the smooth contours in the digital model localized on or near the top surface of the tooth. A discontinuity in the model of that surface satisfying particular criteria can tend to indicate a worn area.
  • the results of the estimated tooth wear, such as the heights or volumes of the worn areas, can be computed and displayed (step 64).
  • FIG. 11 is a diagram of a user interface 68 for displaying estimated tooth wear, for example on display device 16.
  • User interface 68 includes a section 70 for displaying a predicted original shape of the selected tooth, a section 72 for displaying the actual shape as shown in a 3D scan, and a section 74 for displaying a comparison of the shapes to indicate tooth wear.
  • Section 74 can display, for example, the model of the actual shape superimposed on the model of the predicted original shape.
  • Another approach involves predicting tooth wear by determining from a single 3D scan of a person's dentition a score or a rating an amount of tooth wear for that person.
  • This approach takes advantage of a large number of annotated 3D scan data of dentitions that have been given scores related to tooth wear. For example, the Smith and Knight tooth wear index could be used for annotating scans. Given these annotations this approach learns a mapping function that uses low-level mesh features such as curvature, spin images, and the features to predict the Smith and Knight tooth wear index for a particular tooth.
  • Such an approach has a benefit for predicting the onset of conditions such as Bruxism at an early stage without the requirement of multiple scans spread out over a longer period of time.
  • FIG. 12 is a flow chart of a method 80 to implement this approach of predicting tooth wear using 3D scans.
  • Method 80 can be implemented in software or firmware modules, for example, for execution by processor 20.
  • Method 80 can alternatively be implemented in hardware modules or a combination of software and hardware.
  • Method 80 includes optionally receiving a segmented model (step 82), for example from the segmentation methodology described above. Segmentation is optional in that mapping can be applied to the full mesh or a portion of it, representing a full or partial arch of teeth in the digital model of the teeth.
  • a tooth or arch is selected from the model (step 84), and a mapping function is applied to the selected tooth or arch to predict tooth wear (step 86).
  • the results of the predicted tooth wear can be displayed (step 88).
  • Step 86 can be implemented as follows.
  • Many low level mesh features can be computed using well known computer vision and geometric methods. Some examples are features such as multi-scale surface curvature, singular values extracted from Principal Component Analysis of local shape, shape diameter, distances from medial surface points, average geodesic distances, shape contexts, and spin images.
  • a function f Given the ensemble feature set for a mesh X, a function f is defined as follows: f: X -> ⁇ 0, 1,2,3 ⁇ , that is the function f maps the set of features X to a Smith and Knight tooth wear index which takes one of 4 values. In terms of classification, this is a 4-class classification problem. Many different types of classifiers are possible to model this function f.
  • This function can be one or a combination of many classification methods such as support vector machines, decision trees, conditional random fields, or other methods.
  • Support vector machines are known to provide a high a degree of separation between classes with the use of kernels by comparing the features in the appropriate kernel space.
  • Decision trees and ensemble versions of decision trees/stumps such as boosting, bootstrapping, and other versions can be used to combine multiple weak classifiers into strong classifiers with very high performance.
  • conditional random fields provide great performance by taking neighborhood and group labeling into account. This can be used for localized
  • classification such as classifying the top portion of teeth, i.e. the cusps. Bag-of-features are also possible for use in classifying objects globally by the weighted assimilation of local mesh features computed all over the mesh.
  • FIG. 13 is a diagram of a user interface 90 displaying predicted tooth wear, for example on display device 16.
  • User interface 90 includes a section 92 for displaying predicted tooth wear by showing, for example, an annotated scan of the selected tooth.

Abstract

Methods for estimating and predicting tooth wear based upon a single 3D digital model of teeth. The 3D digital model is segmented to identify individual teeth within the model. A digital model of a tooth is selected from the segmented model, and its original shape is predicted. The digital model is compared with the predicted original shape to estimate wear areas. A mapping function based upon values relating to tooth wear can also be applied to the selected digital model to predict wear of the tooth.

Description

ESTIMATING AND PREDICTING TOOTH
WEAR USING INTRA-ORAL 3D SCANS
BACKGROUND
Tooth wear (associated with Bruxism) and gingival recession are both conditions that, if not treated in a timely manner by dental professionals, can have serious medical consequences. In the case of Bruxism, lateral movements and tooth grinding can cause significant tooth wear and lead to muscle pain, temporomandibular joint issues, and headaches. In some cases, this may lead to the dentin being exposed, dental decay, and even tooth fracture. Despite the potential severity of these consequences, the tools available to dental professionals for diagnosing and assessing the severity of tooth wear and gingival recession are limited. In the case of tooth wear, these tools include patient questionnaires, clinical examination by a dentist, and bite force measurements. Clinical examinations may be performed using the Individual Tooth-Wear Index, which provides a rating between 0 and 3 based on visual assessment by a dentist. Accordingly, a need exists for additional tools to assess tooth wear, particularly using intra-oral 3D scans.
SUMMARY
A first method for estimating teeth wear, consistent with the present invention, includes receiving a 3D digital model of teeth and segmenting the 3D digital model of teeth to identify individual teeth within the 3D digital model of teeth. A digital model of a tooth is selected from the segmented 3D digital model of teeth, and an original shape of the selected tooth is predicted to obtain a digital model of a predicted original shape. The digital model of the tooth is compared with the digital model of the predicted original shape to estimate wear areas in the tooth.
A second method for predicting teeth wear, consistent with present invention, includes receiving a 3D digital model of teeth. A mapping function is applied to the digital model of the teeth based upon values relating to tooth wear. Wear areas in the teeth are predicted based upon the applying step. BRIEF DESCRIPTION OF THE DRAWINGS
The accompanying drawings are incorporated in and constitute a part of this specification and, together with the description, explain the advantages and principles of the invention. In the drawings,
FIG. 1 is a diagram of a system for detecting tooth wear using a 3D digital model based upon intra-oral 3D scans;
FIG. 2 illustrates a 3D model of teeth from intra-oral scans;
FIG. 3 illustrates a 3D model of teeth segmented to digitally separate each tooth;
FIG. 4 is a flow chart of a method for segmenting teeth in a 3D digital model; FIG. 5 illustrates over-segmentation of teeth by a geometric hill-climbing method;
FIG. 6 illustrates detection of boundary vertices between teeth in a 3D digital model;
FIGS. 7A and 7B illustrate morphological clean up to fix boundaries between teeth in a 3D digital model;
FIGS. 8 A and 8B illustrate removal of non-aligned boundaries in 3D digital model;
FIGS. 9A-9H illustrate merging of the results of segmentation methods to segment teeth in a 3D digital model;
FIG. 10 is a flow chart of a method of estimating tooth wear using 3D scans;
FIG. 11 is a diagram of a user interface for displaying estimated tooth wear;
FIG. 12 is a flow chart of a method of predicting tooth wear using 3D scans; and
FIG. 13 is a diagram of a user interface for displaying predicted tooth wear.
DETAILED DESCRIPTION
Embodiments includes analyzing tooth wear from a single 3D scan of the patient's dentition. One approach is based on estimating the original shape of the surface through use of a database (or collection) of known tooth shapes and learned mathematical model and for reconstructing shape, and then comparing this estimated original shape with the current shape of the surface. Another approach is based upon comparing the current shape of the surface with annotated scans where the annotation indicates an amount of tooth wear.
A method for determining tooth wear based upon comparison of multiple 3D scans of a patient is described in U.S. Patent Application Serial No. 14/321318, entitled "Detecting Tooth Wear Using Intra-Oral 3D Scans," and filed July 1, 2014, which is incorporated herein by reference as if fully set forth.
3D Scan Acquisition and Segmentation
FIG. 1 is a diagram of a system 10 for detecting tooth wear using a digital 3D model based upon intra-oral 3D scans. System 10 includes a processor 20 receiving digital 3D models of teeth (12) from intra-oral 3D scans or scans of impressions of teeth. System 10 can also include an electronic display device 16, such as a liquid crystal display (LCD) device, for displaying indications of tooth wear and an input device 18 for receiving user commands or other information. An example of digital 3D model of a patient's teeth from a scan is shown in FIG. 2. Systems to generate digital 3D images or models based upon image sets from multiple views are disclosed in U.S. Patent Nos. 7,956,862 and 7,605,817, both of which are incorporated herein by reference as if fully set forth. These systems can use an intra-oral scanner to obtain digital images from multiple views of teeth or other intra-oral structures, and those digital images are processed to generate a digital 3D model representing the scanned teeth. System 10 can be
implemented with, for example, a desktop, notebook, or tablet computer. System 10 can receive the 3D scans locally or remotely via a network.
For certain diagnostic tasks, the individual teeth in the model need to be segmented from one another before the desired analysis or manipulation can be performed. In some cases, a software interface may be presented in order for a user to perform this
segmentation, or some parts of it, manually. However, this process can be quite labor intensive and tedious. As such, the automation of this task is desirable. An example of teeth that have been segmented in a digital model is shown in FIG. 3. The segmentation provides for separating individual teeth in the digital 3D model, as represented by the shading in FIG. 3, and each tooth in the model can essentially be digitally separated from the other teeth for further processing to detect tooth wear. Using a segmented digital 3D model for comparing or analyzing individual teeth is more accurate than comparing whole or partial arches within the model.
Described herein are techniques for tooth segmentation within a digital 3D model.
The technique is a combination of two separate algorithms and combines the strengths of both of them. The first algorithm is a geometric hill-climbing approach which takes into account topological structures such as height and curvature. The second algorithm is a machine learning approach which classifies each point on the surface as belonging to either a boundary or a non-boundary. Alternatively, the second algorithm is interstice detection which classifies a set of planes (or points) that approximate the intersticial spaces between teeth. The second algorithm can be complementary to the first algorithm (geometric hill-climbing) and combined with the first algorithm to produce a resulting segmentation. As another alternative to the second algorithm, the first algorithm can be combined with user input estimating centroids of teeth in the digital 3D model. Instead of merging the results of two algorithms, only one algorithm can be used to segment the digital 3D model such as any one of the algorithms described herein.
The 3D scans addressed herein are represented as triangular meshes. The triangular mesh is common representation of 3D surfaces and has two components. The first component, referred to as the vertices of the mesh, are simply the coordinates of the 3D points that have been reconstructed on the surface - i.e., a point cloud. The second component, the mesh faces, encodes the connections between points on the object and is an efficient way of interpolating between the discrete sample points on the continuous surface. Each face is a triangle defined by three vertices, resulting in a surface that can be represented as a set of small triangular planar patches.
FIG. 4 is a flow chart of a method 22 for segmenting teeth in a digital 3D model. Method 22 can be implemented in software or firmware modules, for example, for execution by processor 20. Method 22 can alternatively be implemented in hardware modules or a combination of software and hardware.
Method 22 includes receiving a digital 3D model of a patient's teeth (step 24) and optionally aligning the model (step 25). Method 22 then involving segmenting the model by geometric hill-climbing (step 26) and point classification (step 28). Optionally, post processing on boundaries of the segmentation by point classification is performed (step 32). As an alternative to point classification, the model can be segmented by interstice detection (step 29). As another alternative to point classification, method 22 can receive user input identifying centroids of each tooth in the model (step 31).
The results of the segmentation methods are iteratively merged (step 30). In particular, the results of segmentation by hill-climbing are merged with the results of segmentation by point classification or interstice detection or user input identifying the centroids. The merged segmentation can optionally be refined based upon manual, for example user-entered, input (step 34). The results of the segmentation are stored (step 38). The segmentation results in a separate mesh for each tooth from the digital 3D model, as illustrated in FIG. 3. These steps are described in more detail below.
The optional alignment step 25 can be implemented using a Support Vector Regression (SVR) method to find the occlusal plane fitted to a mesh of the teeth in the digital 3D model. The alignment can be used to have the teeth in the digital 3D model essentially aligned with the Y axis.
The alignment can use the LIBSVM toolbox and £ ~ ΕΨΜ method. The kernel is chosen to be linear and€ = 5. The training is based on the assumption that teeth are roughly pointing up along the Y axis. The output is sample points from the occlusal plane which is given to a simple principal component analysis (PCA) method to find the normal direction. SVR uses a linear loss function with a zero part within the margins which performs better for teeth dataset than the quadratic loss function in regular least square regression methods. It helps to decrease the effect of gingiva cut-lines which can be very jagged and bumpy in mesh scans. It also tries to rule out the vertical points on the teeth (buccal part) and give more weight of importance to the horizontal points on teeth (cuspal part) in determining the occusal plane orientation. The RANSAC method and Robust PCA method can alternatively be used for the alignment.
Table 1 provides exemplary pseudocode for implementing the alignment step.
Table 1 - Pseudocode for Normal Direction Extraction
Input: a 3D mesh with a set of vertices V specified in 3D coordinate system X,Y and Z. Y represents the rough direction of vertical axis in which the teeth point upwards.
Output: the normal vector perpendicular to occlusal plane which represents the correct upward direction of teeth.
Assumptions: Teeth are roughly pointing up along the Y axis. The mesh has been truncated below the gum line.
Method steps:
Subtract the mean of data points to centralize the data points around (0,0,0).
2 Apply the Support Vector Regression with linear kernel and margin value £ to find the occlusal plane.
3 Find the normal direction of the occlusal plane by geometrical methods or applying a simple PCA.
Segmentation by Geometric Hill-Climbing
One of the algorithms for segmentation is based upon geometric operations on the mesh. Specifically, the main idea behind this approach is that, if one starts from any point on the surface and moves upwards through a series of points, one will converge to a high point that is a local maximum. In most cases it would be expected all points on a tooth (or on the same cusp of a tooth) will converge to the same local maximum. This type of segmentation can produce very accurate boundaries between teeth, but it typically results in an over- segmentation in which a single tooth may be divided into multiple segments.
Before performing the segmentation, the mesh is preprocessed using Laplacian smoothing. This preprocessing is an effective way of removing high-frequency noise in the surface reconstruction.
An energy function is then computed for each vertex on the mesh, on which the algorithm will attempt to find local maxima later in the hill-climbing process. The energy function at each vertex is composed of two terms, where for the z'-th vertex:
Figure imgf000008_0001
where yi is the y-coordinate (height) of the z'-th vertex, ώ is its angular divergence, and λ > 0 is a weighting parameter. The parameter λ can be any value greater than zero or, alternatively, λ can be equal to zero.
Angular divergence is a measure of overall curvature around a point. For a face F comprised of vertices v;, vy, and vk, with normal vectors , tij, and m, respectively, the angular divergence is given by
Figure imgf000008_0002
If the area around a face is completely flat, then all the normal vectors of all three of its vertices will point in the same direction, and the DF will be zero. Then the angular divergence of the z'-th vertex v; is the mean of the angular divergences of the faces of which Vi is a part.
Once the energy is computed for each vertex, segmentation is performed according to a hill-climbing procedure. Conceptually, the algorithm can be understood as follows. For each vertex on the surface, the algorithm initializes a hill-climb, in which at each iteration it moves to the connected neighbor (as defined by the faces) that has the highest energy function value. The algorithm continues climbing until it reaches a local maximum that has higher energy than all of its neighbors. All vertices that were passed through along this route are assigned to this local maximum, and all such paths that converge to this local maximum define a segment. This process is repeated until all vertices on the mesh have been traversed. This segmentation assigns vertices to segments defined by local energy maxima that can be reached through a monotonically-increasing path through the energy function. The energy function is defined such that each iteration of hill-climbing moves upwards in height, but is discouraged from crossing an area with high curvature by the angular divergence term. This helps ensure that the boundaries between teeth are not crossed.
An example of a segmentation produced by this algorithm is shown in FIG. 5. As can be seen, the algorithm over-segments the teeth by separating each cusp of a tooth into its own segment - this can be understood intuitively as a result of the hill-climbing procedure, since each cusp will have its own unique local maximum. For example, the digital model of tooth 40 is segmented into five sections. However, the boundaries produced by this approach are quite precise and accurately separate teeth from one another.
Table 2 provides exemplary pseudocode for implementing the geometric hill- climbing algorithm.
Figure imgf000009_0001
Segmentation by Point Classification The segmentation by point classification is a data-driven approach. Unlike the geometric hill-climbing approach, this approach relies on manually provided groundtruth segmentation. Groundtruth can be obtained from a user providing nearly accurate segmentation manually using mesh manipulation tools such as the MeshLab system. A selection of an individual tooth can be made using a face selection tool. Individual teeth are selected in this manner and saved as individual mesh files. Using the original mesh and the individual teeth files, a labeling of the vertices in the original mesh can then be inferred. Once groundtruth for a full scan is completed, the inferred labels of all the segments can be visualized.
From this groundtruth labeling, the boundary vertices between segments can be determined. For each vertex the distribution of vertex labels around that vertex is examined. If the distribution is not unimodal (i.e., the vertex labels are predominantly the same), then that vertex is considered an interior vertex. If not, the vertex is considered a boundary vertex. This data can be manually entered one time, for example, as training data and then used repeatedly in the point classification algorithm.
Given the groundtruth boundary vertices labels from multiple training meshes, the algorithm provides for a function that is capable of predicting whether a vertex on a mesh lies in the interior of a tooth or on the boundary between teeth. In particular, the algorithm can classify or label points in the mesh as being on a tooth or on a boundary between teeth. This process involves two tasks: feature extraction and classification. FIG. 6 illustrates detection of boundary vertices 42 between teeth in a digital 3D model.
Table 3 provides exemplary pseudocode for implementing the point classification (machine learning) training data algorithm.
Table 3 - Pseudocode for Machine Learning Training
Input: Multiple 3D meshes with a sets of vertices V specified in 3D coordinate system X,Y and Z. Y represents the vertical axis or the general direction in which the teeth point upwards. The mesh also has a set of triangulations or faces F based on the vertices. Also the groundtruth segmentation in the form of the vertices corresponding to boundaries and those in the interior as indicated by manual annotation.
Output: A predictive model that is capable of generating the boundary vertex prediction labels for a query set of vertices.
Assumptions: Teeth are roughly pointing up along the Y axis. The mesh has been truncated below the gum line.
Method steps:
1 For each vertex in every mesh in the training set of data, compute the following features:
a. Normal direction
b. Absolute, mean and Gaussian curvature
c. Shape context
d. Mesh fourier
e. Spin image
f. Mesh local covariance
2 Construct a data matrix X which is M X N where M is the total number of vertices in all the meshes and N is the total number of feature dimensions when all features in step 1 are concatenated
3 Train a RUSBoosted decision tree classifier that can predict the labels
corresponding to whether a vertex lies on the boundary or not. (An alternate classifier can be used.)
Feature Extraction
In order to perform this task, the point classification algorithm extracts many characteristic features for every vertex in the mesh. It is often difficult to determine which features are useful in a segmentation algorithm. There are many features which can be used for segmentation in this framework, including but not limited to multi-scale surface curvature, singular values extracted from PCA of local shape, shape diameter, distances from medial surface points, average geodesic distances, shape contexts, and spin images. Of these, the algorithm implements the following features: absolute and mean curvature, direction of normal at vertex, local covariance of the mesh around the vertex and its principal Eigen values, spin images, Fourier features, shape contexts, and PCA features. Classification
Given the feature set for a vertex X, the function f is defined as follows: f: X -> { 1,0}, that is the function f maps the set of features X to either a 1 or 0. A value 1 indicates that vertex is a boundary vertex and the value 0 indicates otherwise. This function can be one or a combination of many classification methods such as support vector machines, decision trees, conditional random fields, and the like. Additionally, in the segmentation as a classification problem, there is a class imbalance. The number of interior vertices is much greater than the number of boundary vertices. The ratio of interior vertices to boundary vertices is typically 100: 1. In such extreme class imbalance situations, regular classifiers are not optimal. This is because it is possible to obtain very high accuracy by always predicting that a vertex is in the interior, and that would be practically useless since no vertices would be classified as being on a boundary. To remedy this issue, one option involves using classifier ensembles such as boosting.
The classification algorithm uses RUSBoosting on decision stumps as a classifier. RUSBoost stands for random undersampling boosting and is known to handle the class imbalance very well. Additionally RUSBoost is already implemented in the MATLAB program "fitensemble" function. Based on preliminary analysis, RUSBoost was performed on 700 decision stumps. This number was chosen using cross-validation on the training set with the resubstitution loss as the metric. For our experiments, we used a "leave-scan-out" cross-validation scheme. Our dataset consisted of 39 scans, and for every test scan the remaining 38 scans were used for training. The resulting predictions were compared to the groundtruth boundary labels of the test scan. A confusion matrix can then be obtained by comparing the groundtruth labels with the predicted labels. From this we obtained the false alarm rate and the hit rate. With cross-validation testing on 39 scans we obtained an 80% hit rate and 1.7% false alarm rate on average.
Table 4 provides exemplary pseudocode for implementing the point classification (machine learning) algorithm. Table 4 - Pseudocode for Machine Learning Prediction
Input: a 3D mesh with a set of vertices V specified in 3D coordinate system X,Y and Z. Y represents the vertical axis or the general direction in which the teeth point upwards. The mesh also has a set of triangulations or faces F based on the vertices.
Output: Binarized mesh where for each vertex Vi in the mesh, a label li corresponding to whether the vertex belongs to a boundary or not.
Assumptions: Teeth are roughly pointing up along the Y axis. The mesh has been truncated below the gum line.
Method steps:
1 For each vertex Viin V, compute the following features:
a. Normal direction
b. Absolute, mean and Gaussian curvature
c. Shape context
d. Mesh fourier
e. Spin image
f. Mesh local covariance
2 Construct a data matrix X which is M X N where M is the number of vertices in the mesh and N is the total number of feature dimensions when all features in step 1 are concatenated
3 Predict using the learned decision tree RUSBoost classifier the labels corresponding to whether a vertex lies on the boundary or not
Segmentation by Interstice Detection
As an alternative to point classification, the second algorithm for segmentation can use interstice detection (step 29 in method 22). Table 5 provides exemplary pseudocode for implementing the interstice detection algorithm.
Table 5 - Pseudocode for Interstice Detection
Input: a 3D mesh with a set of vertices V specified in 3D coordinate system X,Y and Z. Y represents the vertical axis or the general direction in which the teeth point upwards. The mesh also has a set of triangulations or faces F based on the vertices.
Output: a set of planes that approximate the intersticial spaces between each pair of teeth.
Assumptions: Teeth are roughly pointing up along the Y axis.
Method steps:
1 Form a plan-view range image of the mesh. That is, a range image from the top view, where each pixel represents the height of the surface at the corresponding point.
2 Estimate a one-dimensional parameterization of the dental arch using the Locally- Linear Embedding (LLE) algorithm, which results in a curve that represents the general shape of the arch and passes roughly through the centers of the teeth.
3 Compute a set of evenly-spaced sample points along the one-dimensional
parameterization.
4 For each sample point along the curve, compute the sum of heights in the range image along a line normal to the curve at that point.
5 Intersticial spaces are identified as sample points that are local minima in the sum of heights computed in step 4. The orientation of the intersticial space is given by the direction of the normal to the one-dimensional parameterization curve at the corresponding sample point.
6 Detected intersticial spaces, and their orientations, are mapped back to the three- dimensional coordinates of the original mesh.
Morphological Clean Up
Morphological operations such as mesh erosion and dilation can be done in tandem, resulting in an operation known morphological opening. Unlike images, mesh erosion and dilation are non-trivial since there are no sliding windows. Instead to perform mesh erosion, one can use the connected v-ring of every vertex as its neighborhood.
Performing morphological opening removes islands and small streaks which can interfere with the merging algorithm mentioned later. FIGS. 7 A and 7B illustrate morphological clean up to fix boundaries between teeth in a digital 3D model with FIG. 7B illustrating clean up of the boundaries shown in FIG. 7A. This morphological clean up can be used to for the optional step 32 in method 22 after the segmentation by point classification. Complementary Approaches to Segmentation
Based on the results of the hill-climbing approach and the classification approach, it was observed that the hill-climbing captures the general geometry of cusp and has a tendency to form good boundaries around teeth, but it over-segments and thus creates more false boundaries. The classification approach on the contrary has a somewhat less than desired hit rate on boundaries but has a very low false alarm rate. From this complementary result, a method to merge the results helps reduce the demerits of both approaches and boost the merits of both. In order to accomplish this, a hierarchical merging algorithm is used, which merges the segments in the hill-climbing approach using the boundary predictions of the classification approach. Every boundary predicted by the hill-climbing approach is given a score based on the predicted boundary vertices from the classification approach. Then a hierarchical merging is performed. All the boundaries with a score less than a threshold are discarded and the corresponding segments are merged and the boundary scores are corrected accordingly. This threshold is gradually increased. For example, all boundaries that have score less than 5 are discarded first. The corresponding segments are merged, and then this process is repeated by increasing the threshold step-by-step to, for example, 50. This heuristic provides correct segmentation of the teeth in one of the merge steps in most cases. Elimination of Non-Aligned Boundaries
Even after the merging process, there are some strong false boundaries predicted by the machine learning classifier which are not eliminated completely. These boundaries can be removed using a hypothesis of boundary direction alignment. Since each consecutive tooth boundary is roughly parallel, there cannot be any stark changes in the boundary direction between consecutive teeth. In FIGS. 8A and 8B, a misaligned boundary is removed using such a hypothesis where FIG. 8B illustrates removal of a boundary from the model of FIG. 8 A. This can be achieved by determining the principal direction of orientation of each boundary segment using PCA. The principal components (PCs) of each consecutive tooth boundary should be aligned, thus resulting in eliminating the boundaries which have misaligned PCs. This process is applied after merging the hill climbing result with the machine learning result.
Segmentation Results Sample results of the classification or machine learning (ML), hill-climbing (HC), and the merging steps are shown in FIGS. 9A-9H. The machine learning output (FIG. 9 A) shows the mesh labeling for the boundary vertices and the interior vertices. The second mesh (FIG. 9B) is the result of the hill climbing. As shown in FIG. 9B, the hill-climbing over-segments each tooth but in general there is a reduced chance of a segment being shared across teeth. This is also a behavior associated with the choice of the parameter λ. The meshes displayed in FIGS. 9C-9H indicate iteratively the result of each merge step. Merge 1 corresponds to discarding boundaries with a score less than 5 and merge 2 corresponds to scores less than 10 and so on. In this example, the correct segmentation was achieved at step 6. As shown in the example of FIGS. 9C-9H, it is possible there are no changes between some of the successive (iterative) merge steps. Successive merge steps indicate how aggressively nearby segments are merged and, therefore, in some cases changes are only noticeable at later merge steps.
The score used for merging can represent, for example, the number of points classified as a boundary from the point classification algorithm within a particular vicinity of a boundary determined from the hill-climbing algorithm. An exemplary score of 5 means at least 5 points classified as a boundary are within a particular vicinity of a boundary determined by the hill-climbing algorithm. The particular vicinity used can be based upon, for example, empirical evidence, the typical width or size of a true boundary, or other factors.
In some cases, the best result would be achieved earlier than the 6th merging step and it is possible to get an over-merged result at step 6. In this case one could use the result at step 5 manually or attempt to separate manually just the teeth that are over- merged. Sometimes, an under-merged or over-segmented result can occur even after step 6. In this scenario, by using a cursor control device and user interface a user could manually select ("click on") and merge the segments that require merging to extract the teeth correctly, for example. The final segmented digital 3D model can then be stored in an electronic storage device for later processing.
Table 6 provides exemplary pseudocode for implementing the algorithm for merging hill-climbing segmentation with point classification (machine learning) segmentation. For the alternative intestice detection segmentation, Table 7 provides exemplary pseudocode for implementing the algorithm for merging hill-climbing segmentation with interstice detection segmentation. Table 6 - Pseudocode for Merging Hill-Climbing and Machine Learning Prediction
Input: a 3D mesh with a set of vertices V specified in 3D coordinate system X,Y and Z. Y represents the vertical axis or the general direction in which the teeth point upwards. The mesh also has a set of triangulations or faces F based on the vertices. Label assignments from hill climbing and boundary vertex labels predicted by the machine learning are also provided.
Output: Segmented mesh, where for each vertex Vi in the mesh, a label li corresponding to the segment to which that vertex belongs is assigned.
Assumptions: Teeth are roughly pointing up along the Y axis. The mesh has been truncated below the gum line.
Method ste s:
Figure imgf000017_0001
Table 7 - Pseudocode for Merging Hill-Climbing Segments using Interstice Detection Input: a 3D mesh with a set of vertices V specified in 3D coordinate system X,Y and Z. Y represents the vertical axis or the general direction in which the teeth point upwards. The mesh also has a set of triangulations or faces F based on the vertices. Label assignments from hill climbing, as well as detected intersticial spaces, are also provided.
Output: Segmented mesh, where for each vertex Vi in the mesh, a label li corresponding to the segment to which that vertex belongs is assigned.
Assumptions: Teeth are roughly pointing up along the Y axis.
Method steps:
Figure imgf000017_0002
As an alternative to point classification and interstice detection, the algorithm can merge the hill-climbing segmentation with user input identifying centroids of teeth (step 31 in method 22). This segmentation method requires input from a user at the beginning of the process. In particular, the user identifies the centroid of each tooth in the digital 3D model of teeth. For example, when viewing the digital 3D model of teeth, such as viewing the model in FIG. 2, on display device 16, the user can use input device 18, such as a cursor control device, to select ("click on") the centroid of each tooth in the model or otherwise identify the centroid. The centroid can include the actual centroid or an estimation of the centroid as perceived by the user. This user entered information is used as the initialization for the step of the segmentation which merges the hill-climbing segments using the Kmeans method. These user-identified centroids need to be close to actual centroids of the teeth in order for the segmentation process to work well and not require post-processing by the user. The only parameter required for this method to be trained is ^ in SVR for normal direction extraction described above for the alignment process.
The user-entered information to identify the centroids of each tooth is then merged with the results of the hill-climbing segmentation using the Kmeans clustering method. The vertices should first be replaced by the corresponding local maximum from the hill- climbing step. Then Kmeans method is applied on the new set of vertices to cluster them in k segments, where k is equal to the number of inputs ("clicks") of the user at the beginning of the process. The user's inputs (estimation of teeth centroids) are used as the centroid starting locations of the Kmeans method.
This merging method can result in successful segmentation as follows: clustering is applied on the local maxima (mostly located on the teeth cusps) and not the full mesh, yielding accuracy and speed benefits. The local maxima of larger clusters find higher weights in Kmeans method, and the centroid starting locations entered by the user avoid converging to other possible local optima of Kmeans methods.
Table 8 provides exemplary pseudocode for implementing the algorithm for merging hill-climbing segmentation with user-entered estimations of teeth centroids.
Figure imgf000019_0001
The exemplary pseudocode in Tables 1-8 is provided for illustrative purposes of particular implementations of the described algorithms, and other implementations are possible.
Tooth Wear Estimation and Prediction
In one approach, a single scan of a patient's dentition for a person presumed to have tooth wear is compared to an estimated reconstruction of the original "virgin" tooth shape for each tooth or particular teeth (i.e., the shape before any tooth wear occurred), as predicted by a mathematical model as applied to the current patient's teeth. This approach can be summarized according to the following steps: a 3D scan of the patient's teeth is acquired; the individual teeth are segmented from one another using a segmentation algorithm; for each tooth comprising the patient's dentition, the original "virgin" shape of the tooth is predicted using a mathematical model of tooth shape learned from a large database of teeth; and the current scan of each tooth is compared to the predicted virgin tooth shape in order to assess the degree of wear that has occurred.
FIG. 10 is a flow chart of a method 50 to implement this approach of estimating tooth wear using 3D scans. Method 50 can be implemented in software or firmware modules, for example, for execution by processor 20. Method 50 can alternatively be implemented in hardware modules or a combination of software and hardware. Method 50 includes receiving a segmented model (step 52), for example from the segmentation methodology described above.
A tooth is selected from the model (step 54), and the original "virgin" shape of the tooth is predicted based upon a database of tooth scans (step 56). In order to accomplish this, a large database of 3D tooth models can be accessed and used, where the degree of wear in each of the teeth in the database is known. Given this database, an aggregate generic mathematical model can be formed of the canonical original "virgin" shape of a tooth location. This can be accomplished for each tooth. For teeth that have a large variability in shape from person-to-person, multiple "virgin" models can be formed, which can include clustering of the space of tooth shape for that particular tooth location.
Given this database and the mathematical shape models learned from it, the original "virgin" shape of a tooth from the current scan can be predicted. Several approaches exist for performing this step. One approach is as follows. First, the appropriate model from the database for the current tooth is determined, since multiple clustered models may exist for each tooth location, depending on the variability of tooth shape at this bite location. If multiple models exist for this tooth, the appropriate model is determined by computing the similarity of this tooth to each model. Then, a mapping is computed from the current tooth to the model tooth shape. This mapping can be accomplished through use of a non-rigid registration algorithm. Then, once the new tooth is mapped to the model space, its original "virgin" shape is associated with that of the model. Using the inverse of the mapping estimated previously, this model is mapped back to the space of the current tooth, resulting in a prediction of the original shape of this tooth.
Once the original "virgin" tooth shape has been estimated, it can be compared with the actual current shape in order to assess the amount of wear exhibited (step 60). First, these two models may need to be registered (step 58), using a 3D registration algorithm, so that they are aligned with one another as closely as possible. An example of a registration algorithm is disclosed in the application referenced above. Then, the areas in which the actual and predicted "virgin" models are in disagreement must be located and compared (step 62). These represent the areas of the tooth that have been worn.
Wear areas can optionally be estimated using local smooth contours on the tooth (step 63), and this estimation can be used to supplement the estimated wear areas from step 62. In particular, local discontinuities in the tooth surface can be detected by analysis of the smooth contours in the digital model localized on or near the top surface of the tooth. A discontinuity in the model of that surface satisfying particular criteria can tend to indicate a worn area. The results of the estimated tooth wear, such as the heights or volumes of the worn areas, can be computed and displayed (step 64).
FIG. 11 is a diagram of a user interface 68 for displaying estimated tooth wear, for example on display device 16. User interface 68 includes a section 70 for displaying a predicted original shape of the selected tooth, a section 72 for displaying the actual shape as shown in a 3D scan, and a section 74 for displaying a comparison of the shapes to indicate tooth wear. Section 74 can display, for example, the model of the actual shape superimposed on the model of the predicted original shape.
Another approach involves predicting tooth wear by determining from a single 3D scan of a person's dentition a score or a rating an amount of tooth wear for that person. This approach takes advantage of a large number of annotated 3D scan data of dentitions that have been given scores related to tooth wear. For example, the Smith and Knight tooth wear index could be used for annotating scans. Given these annotations this approach learns a mapping function that uses low-level mesh features such as curvature, spin images, and the features to predict the Smith and Knight tooth wear index for a particular tooth. Such an approach has a benefit for predicting the onset of conditions such as Bruxism at an early stage without the requirement of multiple scans spread out over a longer period of time.
FIG. 12 is a flow chart of a method 80 to implement this approach of predicting tooth wear using 3D scans. Method 80 can be implemented in software or firmware modules, for example, for execution by processor 20. Method 80 can alternatively be implemented in hardware modules or a combination of software and hardware. Method 80 includes optionally receiving a segmented model (step 82), for example from the segmentation methodology described above. Segmentation is optional in that mapping can be applied to the full mesh or a portion of it, representing a full or partial arch of teeth in the digital model of the teeth. A tooth or arch is selected from the model (step 84), and a mapping function is applied to the selected tooth or arch to predict tooth wear (step 86). The results of the predicted tooth wear can be displayed (step 88).
Step 86 can be implemented as follows. Many low level mesh features can be computed using well known computer vision and geometric methods. Some examples are features such as multi-scale surface curvature, singular values extracted from Principal Component Analysis of local shape, shape diameter, distances from medial surface points, average geodesic distances, shape contexts, and spin images. Given the ensemble feature set for a mesh X, a function f is defined as follows: f: X -> {0, 1,2,3 }, that is the function f maps the set of features X to a Smith and Knight tooth wear index which takes one of 4 values. In terms of classification, this is a 4-class classification problem. Many different types of classifiers are possible to model this function f. This function can be one or a combination of many classification methods such as support vector machines, decision trees, conditional random fields, or other methods. Support vector machines are known to provide a high a degree of separation between classes with the use of kernels by comparing the features in the appropriate kernel space. Decision trees and ensemble versions of decision trees/stumps such as boosting, bootstrapping, and other versions can be used to combine multiple weak classifiers into strong classifiers with very high performance. Finally, conditional random fields provide great performance by taking neighborhood and group labeling into account. This can be used for localized
classification such as classifying the top portion of teeth, i.e. the cusps. Bag-of-features are also possible for use in classifying objects globally by the weighted assimilation of local mesh features computed all over the mesh.
FIG. 13 is a diagram of a user interface 90 displaying predicted tooth wear, for example on display device 16. User interface 90 includes a section 92 for displaying predicted tooth wear by showing, for example, an annotated scan of the selected tooth.

Claims

1. A method for estimating teeth wear, comprising steps of:
receiving a 3D digital model of teeth;
segmenting the 3D digital model of teeth to identify individual teeth within the 3D digital model of teeth;
selecting a digital model of a tooth from the segmented 3D digital model of teeth; predicting an original shape of the selected tooth to obtain a digital model of a predicted original shape; and
comparing the digital model of the tooth with the digital model of the predicted original shape to estimate wear areas in the selected tooth.
2. The method of claim 1, wherein the segmenting step comprises:
performing a first segmentation method that over segments at least some of the teeth within the 3D digital model of teeth;
performing a second segmentation method that classifies points within the 3D digital model of teeth as being either on an interior of a tooth or on a boundary between teeth in the 3D digital model of teeth; and
combining results of the first and second segmentation methods to generate a segmented 3D digital model of teeth.
3. The method of claim 1, wherein the predicting step comprises selecting a digital model of a tooth from known tooth shapes based upon characteristics of a patient corresponding with the 3D digital model of teeth.
4. The method of claim 1, further comprising estimating wear areas in the selected tooth by detecting discontinuities in the digital model of the tooth that satisfy particular criteria.
5. The method of claim 1, wherein the comparing step comprises detecting differences in surface area between the digital model of the tooth and the digital model of the predicted original shape.
6. The method of claim 1, further comprising displaying an indication of the estimated wear areas.
7. A method for predicting teeth wear, comprising steps of:
receiving a 3D digital model of teeth;
applying a mapping function to the digital model of the teeth based upon values relating to tooth wear; and
predicting wear areas in the teeth based upon the applying step.
8. The method of claim 7, further comprising, prior to the applying step, segmenting the 3D digital model of teeth to identify individual teeth within the 3D digital model of teeth, comprising:
performing a first segmentation method that over segments at least some of the teeth within the 3D digital model of teeth;
performing a second segmentation method that classifies points within the 3D digital model of teeth as being either on an interior of a tooth or on a boundary between teeth in the 3D digital model of teeth; and
combining results of the first and second segmentation methods to generate a segmented 3D digital model of teeth.
9. The method of claim 7, further comprising annotating the digital model of the tooth based upon the predicted wear areas.
10. The method of claim 9, further comprising displaying the annotated digital model of the tooth.
11. A system for estimating teeth wear, comprising:
a module for receiving a 3D digital model of teeth;
a module for segmenting the 3D digital model of teeth to identify individual teeth within the 3D digital model of teeth;
a module for selecting a digital model of a tooth from the segmented 3D digital model of teeth; a module for predicting an original shape of the selected tooth to obtain a digital model of a predicted original shape; and
a module for comparing the digital model of the tooth with the digital model of the predicted original shape to estimate wear areas in the tooth.
12. The system of claim 11, wherein the module for segmenting comprises:
a module for performing a first segmentation method that over segments at least some of the teeth within the 3D digital model of teeth;
a module for performing a second segmentation method that classifies points within the 3D digital model of teeth as being either on an interior of a tooth or on a boundary between teeth in the 3D digital model of teeth; and
a module for combining results of the first and second segmentation methods to generate a segmented 3D digital model of teeth.
13. The system of claim 11, wherein the module for predicting comprises selecting a digital model of a tooth from known tooth shapes based upon characteristics of a patient corresponding with the 3D digital model of teeth.
14. The system of claim 11, further comprising a module for estimating wear areas in the selected tooth by detecting discontinuities in the digital model of the tooth that satisfy particular criteria.
15. The system of claim 11, wherein the module for comparing comprises a module for detecting differences in surface area between the digital model of the tooth and the digital model of the predicted original shape.
16. The system of claim 11, further comprising a module for displaying an indication of the estimated wear areas.
17. A system for predicting teeth wear, comprising:
a module for receiving a 3D digital model of teeth;
a module for applying a mapping function to the digital model of the teeth based upon values relating to tooth wear; and a module for predicting wear areas in the digital model of the teeth based upon the applying step.
18. The system of claim 17, further comprising a module for segmenting the 3D digital model of teeth to identify individual teeth within the 3D digital model of teeth, comprising:
a module for performing a first segmentation method that over segments at least some of the teeth within the 3D digital model of teeth;
a module for performing a second segmentation method that classifies points within the 3D digital model of teeth as being either on an interior of a tooth or on a boundary between teeth in the 3D digital model of teeth; and
a module for combining results of the first and second segmentation methods to generate a segmented 3D digital model of teeth.
19. The system of claim 17, further comprising a module for annotating the digital model of the tooth based upon the predicted wear areas.
20. The system of claim 19, further comprising a module for displaying the annotated digital model of the tooth.
PCT/US2016/013329 2015-01-30 2016-01-14 Estimating and predicting tooth wear using intra-oral 3d scans WO2016122890A1 (en)

Priority Applications (4)

Application Number Priority Date Filing Date Title
AU2016212031A AU2016212031B2 (en) 2015-01-30 2016-01-14 Estimating and predicting tooth wear using intra-oral 3D scans
DK16743844.9T DK3250111T3 (en) 2015-01-30 2016-01-14 ESTIMATION AND PREDICTION OF TEETH WEAR USING INTRA-ORAL 3D SCANNINGS
EP20166109.7A EP3708069A1 (en) 2015-01-30 2016-01-14 Estimating and predicting tooth wear using intra-oral 3d scans
EP16743844.9A EP3250111B1 (en) 2015-01-30 2016-01-14 Estimating and predicting tooth wear using intra-oral 3d scans

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US14/609,529 2015-01-30
US14/609,529 US9737257B2 (en) 2015-01-30 2015-01-30 Estimating and predicting tooth wear using intra-oral 3D scans

Publications (1)

Publication Number Publication Date
WO2016122890A1 true WO2016122890A1 (en) 2016-08-04

Family

ID=56544172

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2016/013329 WO2016122890A1 (en) 2015-01-30 2016-01-14 Estimating and predicting tooth wear using intra-oral 3d scans

Country Status (5)

Country Link
US (2) US9737257B2 (en)
EP (2) EP3250111B1 (en)
AU (1) AU2016212031B2 (en)
DK (1) DK3250111T3 (en)
WO (1) WO2016122890A1 (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170311873A1 (en) * 2015-01-30 2017-11-02 3M Innovative Properties Company Estimating and predicting tooth wear using intra-oral 3d scans

Families Citing this family (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8108189B2 (en) * 2008-03-25 2012-01-31 Align Technologies, Inc. Reconstruction of non-visible part of tooth
BR112013021378B1 (en) 2011-02-23 2021-06-15 3Shape A/S METHOD FOR MODIFYING THE GINGIVAL PART OF A VIRTUAL MODEL OF A SET OF TEETH
US9814549B2 (en) * 2015-09-14 2017-11-14 DENTSPLY SIRONA, Inc. Method for creating flexible arch model of teeth for use in restorative dentistry
US11207161B2 (en) * 2016-05-30 2021-12-28 3Shape A/S Predicting the development of a dental condition
ES2808210T3 (en) * 2016-08-15 2021-02-25 Trophy Dynamic dental arch map
US10828130B2 (en) * 2017-03-20 2020-11-10 Align Technology, Inc. Automated 2D/3D integration and lip spline autoplacement
CN108986123A (en) * 2017-06-01 2018-12-11 无锡时代天使医疗器械科技有限公司 The dividing method of tooth jaw three-dimensional digital model
US10327693B2 (en) 2017-07-07 2019-06-25 3M Innovative Properties Company Tools for tracking the gum line and displaying periodontal measurements using intra-oral 3D scans
TWI644655B (en) 2017-11-23 2018-12-21 勘德股份有限公司 Digital dental mesh segmentation method and digital dental mesh segmentation device
US10916053B1 (en) * 2019-11-26 2021-02-09 Sdc U.S. Smilepay Spv Systems and methods for constructing a three-dimensional model from two-dimensional images
US11403813B2 (en) 2019-11-26 2022-08-02 Sdc U.S. Smilepay Spv Systems and methods for constructing a three-dimensional model from two-dimensional images
US11270523B2 (en) * 2017-11-29 2022-03-08 Sdc U.S. Smilepay Spv Systems and methods for constructing a three-dimensional model from two-dimensional images
US11553988B2 (en) 2018-06-29 2023-01-17 Align Technology, Inc. Photo of a patient with new simulated smile in an orthodontic treatment review software
EP3899984A1 (en) 2018-12-21 2021-10-27 The Procter & Gamble Company Apparatus and method for operating a personal grooming appliance or household cleaning appliance
US11488062B1 (en) * 2018-12-30 2022-11-01 Perimetrics, Inc. Determination of structural characteristics of an object
CN109864829A (en) * 2019-01-28 2019-06-11 苏州佳世达光电有限公司 Scanning system and scan method
US11030801B2 (en) 2019-05-17 2021-06-08 Standard Cyborg, Inc. Three-dimensional modeling toolkit
JP6800358B1 (en) * 2020-02-03 2020-12-16 株式会社松風 Dental prosthodontic device design method and design device
CN112991273B (en) * 2021-02-18 2022-12-16 山东大学 Orthodontic feature automatic detection method and system of three-dimensional tooth model

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6766217B1 (en) * 1999-08-02 2004-07-20 Kabushiki Kaisya Advance Method of manufacturing dental prosthesis, method of placing object for measurement and measuring device
US20050095562A1 (en) * 1999-11-30 2005-05-05 Peer Sporbert Three-dimensional occlusal and interproximal contact detection and display using virtual tooth models
US20090034811A1 (en) * 2007-08-02 2009-02-05 Kuo Eric E Mapping abnormal dental references
US7605817B2 (en) 2005-11-09 2009-10-20 3M Innovative Properties Company Determining camera motion
WO2010033404A1 (en) 2008-09-19 2010-03-25 3M Innovative Properties Company Methods and systems for determining the positions of orthodontic appliances
EP2258303A1 (en) 2000-04-19 2010-12-08 OraMetrix, Inc. System for creating an individual three-dimensional virtual tooth model
US20130054190A1 (en) * 2011-08-23 2013-02-28 Yusei Kadobayashi Occlusal wear evaluation apparatus and occlusal wear evaluation method
US8422751B2 (en) * 2008-08-28 2013-04-16 E-Woo Technology Co., Ltd. Method and apparatus for generating virtual teeth, and recording media storing program performing the method
EP2604220A1 (en) 2010-08-10 2013-06-19 Hidefumi Ito Information processing device, information processing method and program
US20130308843A1 (en) 2011-02-10 2013-11-21 Straumann Holding Ag Method and analysis system for the geometrical analysis of scan data from oral structures

Family Cites Families (33)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP0840574B1 (en) * 1995-07-21 2003-02-19 Cadent Ltd. Method and system for acquiring three-dimensional teeth image
US6152731A (en) * 1997-09-22 2000-11-28 3M Innovative Properties Company Methods for use in dental articulation
US6227850B1 (en) * 1999-05-13 2001-05-08 Align Technology, Inc. Teeth viewing system
US6406292B1 (en) * 1999-05-13 2002-06-18 Align Technology, Inc. System for determining final position of teeth
AU2164100A (en) * 1998-12-04 2000-06-26 Align Technology, Inc. Reconfigurable dental model system for fabrication of dental appliances
US6648640B2 (en) * 1999-11-30 2003-11-18 Ora Metrix, Inc. Interactive orthodontic care system based on intra-oral scanning of teeth
US6371761B1 (en) * 2000-03-30 2002-04-16 Align Technology, Inc. Flexible plane for separating teeth models
US7471821B2 (en) 2000-04-28 2008-12-30 Orametrix, Inc. Method and apparatus for registering a known digital object to scanned 3-D model
US7027642B2 (en) * 2000-04-28 2006-04-11 Orametrix, Inc. Methods for registration of three-dimensional frames to create three-dimensional virtual models of objects
US7040896B2 (en) 2000-08-16 2006-05-09 Align Technology, Inc. Systems and methods for removing gingiva from computer tooth models
US7080979B2 (en) * 2001-04-13 2006-07-25 Orametrix, Inc. Method and workstation for generating virtual tooth models from three-dimensional tooth data
US7362890B2 (en) * 2001-05-24 2008-04-22 Astra Tech Inc. Registration of 3-D imaging of 3-D objects
US20030040009A1 (en) * 2001-08-14 2003-02-27 University Of Southern California Saliva-based methods for preventing and assessing the risk of diseases
US7156661B2 (en) * 2002-08-22 2007-01-02 Align Technology, Inc. Systems and methods for treatment analysis by teeth matching
US7077647B2 (en) * 2002-08-22 2006-07-18 Align Technology, Inc. Systems and methods for treatment analysis by teeth matching
US7029279B2 (en) * 2002-10-07 2006-04-18 Mark Schomann Prosthodontia system
AU2004227999B2 (en) * 2003-04-01 2009-09-24 Proactive Oral Solutions, Inc. Caries risk test for predicting and assessing the risk of disease
US7695278B2 (en) 2005-05-20 2010-04-13 Orametrix, Inc. Method and system for finding tooth features on a virtual three-dimensional model
US20070024611A1 (en) 2005-07-27 2007-02-01 General Electric Company System and method for aligning three-dimensional models
US7813591B2 (en) * 2006-01-20 2010-10-12 3M Innovative Properties Company Visual feedback of 3D scan parameters
US20070207441A1 (en) * 2006-03-03 2007-09-06 Lauren Mark D Four dimensional modeling of jaw and tooth dynamics
US8099268B2 (en) * 2007-05-25 2012-01-17 Align Technology, Inc. Tooth modeling
US8075306B2 (en) 2007-06-08 2011-12-13 Align Technology, Inc. System and method for detecting deviations during the course of an orthodontic treatment to gradually reposition teeth
CN101689309A (en) * 2007-06-29 2010-03-31 3M创新有限公司 The synchronized views of video data and three-dimensional modeling data
US8244028B2 (en) 2010-04-30 2012-08-14 Align Technology, Inc. Virtual cephalometric imaging
JP5959539B2 (en) * 2011-02-18 2016-08-02 スリーエム イノベイティブ プロパティズ カンパニー Orthodontic digital setup
EP3417830B1 (en) * 2012-02-10 2020-07-01 3Shape A/S Virtually designing a post and core restoration using a digital 3d shape
US9626462B2 (en) * 2014-07-01 2017-04-18 3M Innovative Properties Company Detecting tooth wear using intra-oral 3D scans
US10192003B2 (en) * 2014-09-08 2019-01-29 3M Innovative Properties Company Method of aligning intra-oral digital 3D models
US11147652B2 (en) * 2014-11-13 2021-10-19 Align Technology, Inc. Method for tracking, predicting, and proactively correcting malocclusion and related issues
US9737257B2 (en) * 2015-01-30 2017-08-22 3M Innovative Properties Company Estimating and predicting tooth wear using intra-oral 3D scans
US9770217B2 (en) * 2015-01-30 2017-09-26 Dental Imaging Technologies Corporation Dental variation tracking and prediction
US10032271B2 (en) * 2015-12-10 2018-07-24 3M Innovative Properties Company Method for automatic tooth type recognition from 3D scans

Patent Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6766217B1 (en) * 1999-08-02 2004-07-20 Kabushiki Kaisya Advance Method of manufacturing dental prosthesis, method of placing object for measurement and measuring device
US20050095562A1 (en) * 1999-11-30 2005-05-05 Peer Sporbert Three-dimensional occlusal and interproximal contact detection and display using virtual tooth models
EP2258303A1 (en) 2000-04-19 2010-12-08 OraMetrix, Inc. System for creating an individual three-dimensional virtual tooth model
US7605817B2 (en) 2005-11-09 2009-10-20 3M Innovative Properties Company Determining camera motion
US7956862B2 (en) 2005-11-09 2011-06-07 3M Innovative Properties Company Determining camera motion
US20090034811A1 (en) * 2007-08-02 2009-02-05 Kuo Eric E Mapping abnormal dental references
US8422751B2 (en) * 2008-08-28 2013-04-16 E-Woo Technology Co., Ltd. Method and apparatus for generating virtual teeth, and recording media storing program performing the method
WO2010033404A1 (en) 2008-09-19 2010-03-25 3M Innovative Properties Company Methods and systems for determining the positions of orthodontic appliances
EP2604220A1 (en) 2010-08-10 2013-06-19 Hidefumi Ito Information processing device, information processing method and program
US20130308843A1 (en) 2011-02-10 2013-11-21 Straumann Holding Ag Method and analysis system for the geometrical analysis of scan data from oral structures
US20130054190A1 (en) * 2011-08-23 2013-02-28 Yusei Kadobayashi Occlusal wear evaluation apparatus and occlusal wear evaluation method

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
DETECTING TOOTH WEAR USING INTRA-ORAL 3D SCANS, 1 July 2014 (2014-07-01)
See also references of EP3250111A4

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170311873A1 (en) * 2015-01-30 2017-11-02 3M Innovative Properties Company Estimating and predicting tooth wear using intra-oral 3d scans
US10405796B2 (en) * 2015-01-30 2019-09-10 3M Innovative Properties Company Estimating and predicting tooth wear using intra-oral 3D scans

Also Published As

Publication number Publication date
US20170311873A1 (en) 2017-11-02
AU2016212031A1 (en) 2017-08-10
US20160220173A1 (en) 2016-08-04
AU2016212031B2 (en) 2018-05-10
DK3250111T3 (en) 2020-09-21
EP3708069A1 (en) 2020-09-16
US10405796B2 (en) 2019-09-10
EP3250111B1 (en) 2020-06-17
US9737257B2 (en) 2017-08-22
EP3250111A1 (en) 2017-12-06
EP3250111A4 (en) 2018-07-25

Similar Documents

Publication Publication Date Title
US10405796B2 (en) Estimating and predicting tooth wear using intra-oral 3D scans
US10410346B2 (en) Detecting tooth wear using intra-oral 3D scans
Maitra et al. Technique for preprocessing of digital mammogram
US20180300877A1 (en) Method for automatic tooth type recognition from 3d scans
Hogeweg et al. Clavicle segmentation in chest radiographs
US20060241412A1 (en) Method for visualizing damage in the myocardium
Maitra et al. Accurate breast contour detection algorithms in digital mammogram
CN113782184A (en) Cerebral apoplexy auxiliary evaluation system based on facial key point and feature pre-learning
Dharmalingham et al. A model based segmentation approach for lung segmentation from chest computer tomography images
Davis et al. Automated bone age assessment using feature extraction
Sundararajan et al. A multiresolution support vector machine based algorithm for pneumoconiosis detection from chest radiographs
Lee et al. Hybrid airway segmentation using multi-scale tubular structure filters and texture analysis on 3D chest CT scans
JP5640280B2 (en) Osteoporosis diagnosis support device and osteoporosis diagnosis support program
Mortaheb et al. Metal artifact reduction and segmentation of dental computerized tomography images using least square support vector machine and mean shift algorithm
Viitaniemi et al. Detecting hand-head occlusions in sign language video
del Fresno et al. Application of color image segmentation to estrus detection
Atuhaire Reconstruction of three-dimensional facial geometric features related to fetal alcohol syndrome using adult surrogates
CN114078105A (en) Blood vessel detection method, ultrasound apparatus, and computer-readable storage medium
Zhihong et al. Interactive feature extraction on 3D meshes
Uriondo et al. Computed Tomography CAD system for monitoring and modeling the evolution of lung cancer nodule
Samanu et al. Semi-automatic spine extraction for disc space narrowing diagnosis
Kim et al. Advanced medical image visualization and analysis systems for diagnosis and treatment planning
Dandu et al. Scale focusing of statistical shape models for breast region segmentation and pectoral muscle suppression
Su et al. A Knowledge-Based Lung Nodule Detection System for Helical CT Images
Mohd Adib et al. Detection of cardiac geometry via difference intensity of echocardiogram images

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 16743844

Country of ref document: EP

Kind code of ref document: A1

REEP Request for entry into the european phase

Ref document number: 2016743844

Country of ref document: EP

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2016212031

Country of ref document: AU

Date of ref document: 20160114

Kind code of ref document: A