Method and system for tooth wear detection

Abstract

A computer-implemented method for detecting tooth wear on a virtual 3D model (101) of a tooth condition is disclosed. The method includes receiving, by a processor, a virtual 3D model of a tooth condition (101). The method further comprises generating an input (301) of a trained neural network (300) for tooth wear detection from the virtual 3D model (101), where the input (301) comprises at least one input element corresponding to at least a portion of the virtual 3D model (101), the at least one input element comprising geometric information of the at least a portion of the virtual 3D model (101), wherein the geometric information of at least a portion of the virtual 3D model (101) comprises curvature information. The method further comprises providing an input (301) to the trained neural network (300) and obtaining an output (305) from the trained neural network (300) based on the provided input (301), where the output (305) comprises at least one probability value, where the at least one probability value represents a likelihood that the at least one input element corresponds to a tooth wear category. Further, the method comprises assigning a tooth wear category to the at least one input element, and marking tooth wear on at least a portion of the virtual 3D model (101) based on the tooth wear category assigned to the at least one input element. At least a portion of the virtual 3D model (101) is displayed along with the tagged tooth wear.

Description

Technical Field

[0001] The present disclosure relates to a computer-implemented method and system for detecting tooth wear on a virtual 3D model of a dental condition. Background Art

[0002] Tooth wear is a dental condition characterized by loss of tooth substance. It is often painful and impairs tooth function. Three types of tooth wear occur: abrasion, attrition, and erosion. Abrasion is the physical wear of teeth caused by factors other than tooth-to-tooth contact, such as improper brushing. Attrition is the loss of tooth substance caused by tooth-to-tooth contact. Erosion is the dissolution of tooth enamel due to the presence of acids in the mouth.

[0003] The damage caused by tooth wear is irreversible and can be difficult to repair. Therefore, timely detection and monitoring of tooth wear by your GP is crucial to protecting tooth structure.

[0004] Clinical detection and diagnosis of tooth wear is currently based on direct visual inspection of the teeth, which is highly subjective. When performed by general practitioners, visual inspection is characterized by low sensitivity and low reproducibility. Furthermore, tooth material loss is visually detectable only when a significant amount of hard tooth tissue has been lost.

[0005] Digital dentistry and the use of intraoral scanners have enabled the development of more accurate methods for detecting dental conditions, reducing or completely eliminating physician subjectivity. For example, it is now possible to obtain two intraoral scans of a patient's dental condition at different time periods. Tooth wear can then be detected by comparing the geometry of the two scans and identifying quantifiable geometric differences. However, the disadvantage of this method is that tooth wear cannot be determined from a single scan alone; instead, two scans of the dental condition must be taken at different points in time.

[0006] There is a clear need to develop methods and systems that will aid in tooth wear detection from a single intraoral scan of a patient's dental condition. Summary of the Invention

[0007] Disclosed herein is a computer-implemented method for detecting tooth wear on a virtual 3D model of a dental condition, the computer-implemented method comprising:

[0008] - receiving, by a processor, said virtual 3D model of said dental condition,

[0009] - generating an input for a trained neural network for tooth wear detection from the virtual 3D model, wherein the input comprises at least one input element corresponding to at least a portion of the virtual 3D model, the at least one input element comprising geometric information of the at least a portion of the virtual 3D model,

[0010] - providing said input to said trained neural network,

[0011] - obtaining an output from the trained neural network based on the provided input, wherein the output comprises at least one probability value, wherein the at least one probability value represents a likelihood that the at least one input element corresponds to a tooth wear category,

[0012] - assigning said tooth wear category to said at least one input element,

[0013] - marking tooth wear on said at least one portion of said virtual 3D model based on said tooth wear category assigned to said at least one input element,

[0014] - displaying said at least a portion of said virtual 3D model together with said annotated tooth wear.

[0015] A method according to an embodiment may include receiving, by a processor, a virtual 3D model of the dental condition.

[0016] Throughout this disclosure the expression "3D" refers to the term "three-dimensional". Similarly, the term "2D" refers to the term "two-dimensional". The term "virtual 3D model of a dental case" refers to a virtual, three-dimensional, computer-generated representation of a patient's dental case. Such a virtual 3D model may be constructed by a processor based on scan data collected during an intraoral scanning process, wherein an intraoral scanner is used to scan the patient's dental case including teeth and gums. The virtual 3D model may alternatively be generated by scanning a plaster model of the patient's dental case using a conventional 3D scanner, also known as a lab scanner or desktop scanner. The virtual 3D model may be stored in a memory of a computer system, for example, in a Standard Triangulation Language (STL) format.

[0017] The virtual 3D model may be received or accessed by the processor. The virtual 3D model may typically be displayed on a display screen in the form of a 3D mesh, a point cloud, a volume representation, or any other suitable 3D representation.

[0018] A virtual 3D model of a dental case can be obtained from scan data from a single scan of the dental case and used in the methods according to the present disclosure. This means that, using a trained neural network, tooth wear can be detected from just a single scan of the patient's dental case. In particular, there is no need to obtain a second scan at a later point in time and compare the geometries of the two scans. A single scan of the dental case can include one or more correction scans to accurately construct the virtual 3D model.

[0019] Furthermore, the method may include generating an input from the virtual 3D model for a trained neural network for tooth wear detection, wherein the input may include at least one input element corresponding to at least a portion of the virtual 3D model, and the at least one input element may include geometric information of the at least a portion of the virtual 3D model.

[0020] The method can utilize a trained neural network to detect the presence of tooth wear on a virtual 3D model. To this end, input to the trained neural network can be generated from the virtual 3D model. Depending on the type of trained neural network, the input can be in 3D or 2D format. For example, one type of trained neural network may be capable of processing 2D images of the virtual 3D model, while another type of trained neural network may be capable of processing the virtual 3D model in the form of a point cloud. Thus, the generating step results in a virtual 3D model suitable for processing by the trained neural network.

[0021] In one embodiment, the input may be in a 3D format. Examples of input in a 3D format may be a point cloud, a 3D mesh, a volumetric representation, or a graphical representation of the virtual 3D model. It may be possible to transform the input from one 3D representation type (e.g., a mesh representation) to another 3D representation type (e.g., a point cloud representation). The type of trained neural network capable of processing input in a 3D format may be a trained PointNet network or a PointNet-type network. These neural networks may be capable of processing the virtual 3D model or at least a portion of the virtual 3D model in the form of a point cloud.

[0022] At least one input element can be a point from a point cloud representation of the virtual 3D model. Similarly, at least one input element can be a facet from a mesh representation of the virtual 3D model or a voxel from a volume representation of the virtual 3D model. At least one of these input elements can be included in the input of the trained neural network. All input elements corresponding to the virtual 3D model can also be provided as input to the trained neural network. For example, all points from a point cloud representation of the virtual 3D model can be provided to the trained neural network.

[0023] Generating the input may include sampling at least one input element or converting at least one input element into a digital form. For example, one or more points from a point cloud representation of a virtual 3D model may be converted into a digital form, such as a matrix of input values. The matrix may then form the input to the trained neural network.

[0024] If the virtual 3D model is represented by a 3D mesh, it can be converted to a point cloud representation by collapsing all vertices from the 3D mesh representation so that only points remain, thus forming a point cloud. The retained point cloud can then be processed by a PointNet network or a PointNet-type network.

[0025] In another embodiment, the input may be in a 2D format, such as a 2D image or multiple 2D images of at least one tooth surface of at least one tooth in the virtual 3D model.

[0026] Generating an input for a trained neural network for tooth wear detection from a virtual 3D model may include generating one or more 2D images of at least one tooth surface of at least one tooth in the virtual 3D model. The one or more 2D images may then form at least one input element. Alternatively, the at least one input element may be a pixel belonging to the one or more 2D images. The one or more 2D images may be generated with the aid of a virtual camera, as will be explained later.

[0027] The geometric information included in the at least one input element, also referred to as topological information, may be depth information related to the camera position, angles between facet normals and the camera direction, facet normals, and / or curvature information. The term "camera" in this context relates to a virtual camera used to view the virtual 3D model. The camera may be characterized by virtual camera properties, such as field of view (FOV) and / or focal length.

[0028] The depth information may include distance information of at least a portion of the virtual 3D model from the camera. The depth information may be in the form of a depth image. The depth image may display the at least a portion of the virtual 3D model according to its distance from the camera. For example, a portion of the virtual 3D model that is farther from the camera may be shaded darker than a portion that is closer to the camera. A distance threshold may be defined such that when the distance between the camera and a portion of the virtual 3D model is greater than the distance threshold, the portion of the virtual 3D model is not shown in the depth information. The distance of at least a portion of the virtual 3D model from a given camera position indirectly describes the topology of at least a portion of the virtual 3D model, and therefore describes the smoothness of the surface of at least a portion of the virtual 3D model. The difference in surface smoothness may represent a relevant parameter when detecting tooth wear in at least a portion of the virtual 3D model.

[0029] The facet normal can be a vector that is perpendicular to the facet of the virtual 3D model. The facet can be regarded as a "building block" of the virtual 3D model in the format of a 3D mesh, and can be, for example, a triangle in the case where the virtual 3D model is in the format of a triangular mesh. A triangular facet can be defined by three points (vertices) interconnected by three edges. Two adjacent facets with small differences in their respective facet normals can indicate a smooth local area. However, if a significant change in the facet normal can be observed, the adjacent facets define a non-smooth surface that is relevant for tooth wear detection. The change in two surface normals can be expressed by the angle between the two surface normals.

[0030] In an embodiment, the geometric information of the at least a portion of the virtual 3D model may include a change in facet normals of at least two adjacent facets.

[0031] The curvature information can be in the form of a value for each facet and can be obtained by measuring the change in adjacent facets of the virtual 3D model along the facet normal. This change can give an idea of whether the associated section of the virtual 3D model is flat or curved. Thus, the curvature information can describe the topology of the tooth by revealing smooth and non-smooth surfaces. Tooth wear can be characterized by sudden changes in relative curvature between areas in the tooth surface. Tooth wear can also be associated with local smoothness (low curvature areas) in certain areas. Therefore, curvature information is an important parameter when detecting tooth wear, detecting the location of tooth wear lesions, tooth wear type and / or tooth wear severity.

[0032] Once generated, the input can be fed into the trained neural network in order to obtain the output from the trained neural network. All the information that forms the input can be concatenated into the final input tensor.

[0033] In an embodiment, generating input from the virtual 3D model for the trained neural network may include generating one or more two-dimensional (2D) images of the at least one tooth surface of the at least one tooth in the virtual 3D model.

[0034] Thus, the trained neural network can be adapted to process 2D input formats. One or more of the 2D images can be generated by a virtual camera, also referred to as a camera. The virtual camera can take snapshots of individual teeth of the virtual 3D model. For example, the teeth in the segmented 3D model can be positioned so that the desired tooth surfaces of the teeth face the camera. The teeth can be rotated and translated via the local coordinate system so that one or more of the tooth surfaces are photographed. This can be an automatic process, where predetermined movements of the teeth have been set in order to capture one or more of the tooth surfaces via the virtual camera. Alternatively, the virtual camera can be moved to a plurality of predetermined positions to capture snapshots of the desired tooth surfaces. Snapshots of individual teeth can alternatively be generated by manually moving the virtual camera to the desired position.

[0035] At least one tooth in the virtual 3D model can be obtained by dividing the virtual 3D model into multiple teeth. Each tooth may include multiple tooth surfaces. The multiple tooth surfaces of each tooth may include buccal surface, lingual surface, distal surface, mesial surface and / or occlusal surface. The multiple tooth surfaces may also include a cervical surface and / or an incisor surface of the tooth. The lingual surface of at least one tooth may be a surface facing the tongue and may be closest to the tongue. The buccal surface may face the cheek. The occlusal surface may be a chewing surface. The occlusal surface may also be an incisor surface, which refers to the biting edge of the incisor. The cervical surface may constitute the boundary between the enamel covering the crown of the tooth and the cementum covering the root of the tooth.

[0036] Segmenting the virtual 3D model can be performed via a segmentation process that allows for identification of different dental objects in the virtual 3D model, such as individual teeth and / or surrounding gums. Individual teeth can be assigned tooth identifiers, for example, based on the Universal Notation (UNN) method, which assigns numbers 1 to 32 to human teeth. The segmentation process can include the use of algorithms such as principal component analysis (PCA) or harmonic fields. Alternatively or additionally, the segmentation process can include the use of a machine learning model.

[0037] Optionally, providing input to the trained neural network may include providing semantic information of at least a portion of the virtual 3D model to the trained neural network. Thus, the input may additionally include the semantic information of the at least a portion of the virtual 3D model.

[0038] The semantic information of at least a portion of the virtual 3D model may include a tooth identifier, a tooth surface identifier, and / or a jaw identifier. The tooth surface identifier may provide information about the surface type (e.g., buccal, lingual, or occlusal) in the input, while the jaw identifier may provide information about which jaw (upper or lower) the input or at least one input element represents.

[0039] By incorporating semantic information into the input of a trained neural network, semantic context can be imparted to the data present in the input, thereby assisting the trained neural network in processing geometric information. The semantic information can be converted into a digital format, i.e., another data format, before being processed by the trained neural network. Better performance of the trained neural network can be achieved by additionally incorporating this semantic information into the input. For example, tooth wear lesions may occur in a common manner between teeth of the same or similar types. To more efficiently detect these tooth wear lesions, the trained neural network can use the provided additional tooth identifiers as semantic information.

[0040] Because different types of teeth exhibit different patterns of tooth wear lesions, semantic information in the form of tooth identifiers can be particularly useful. For example, the difference in tooth wear behavior between molars and incisors is striking. While tooth wear lesions on incisors typically occupy the entire incisal edge, molars often have multiple, pore-like lesions on their occlusal surfaces. Incorporating tooth identifiers as additional features into the input improves the performance of trained neural networks.

[0041] In one embodiment, instead of receiving tooth identifiers as part of the input, the trained neural network can predict tooth shape by classifying the input according to predetermined tooth shape categories, such as "incisor," "molar," or "canine." Thus, the trained neural network, including a classification head for predicting tooth shape, can have a multi-task architecture.

[0042] Optionally, providing input to the trained neural network may include providing color information, such as red, green, and blue (RGB) intensity values for at least one input element. Thus, the color information of at least one input element may be part of the input to the trained neural network. The presence of color information may be beneficial because it facilitates the identification of tooth wear lesions that may expose dentin because dentin is more yellow compared to enamel. This is particularly beneficial in cases of severe tooth wear characterized by loss of tooth substance enamel. Similarly, for example, once the enamel is etched, the portion of the remaining enamel that becomes visible has a different color due to the lack of any marking or staining that may occur from tea or coffee stains. The color difference may be indicative of a tooth wear lesion. Therefore, color information may be highly relevant to the trained neural network in the detection of tooth wear lesions, particularly those that expose dentin.

[0043] Further, the method may include obtaining an output from the trained neural network based on the provided input, wherein the output includes at least one probability value, wherein the at least one probability value represents a likelihood that the at least one input element corresponds to a tooth wear category.

[0044] The tooth wear category may indicate whether tooth wear exists on the at least one input element. For example, the tooth wear category may be one of “yes” or “no” for indicating whether tooth wear exists.

[0045] In another example, the tooth wear category may be a tooth wear severity value indicating how severe the detected tooth wear lesion is.

[0046] In yet another example, the tooth wear category may be a tooth wear type indicating a specific type of tooth wear lesion detected. As previously mentioned, the tooth wear type may be one of abrasion, erosion, and tooth erosion.

[0047] Based on the at least one obtained probability value, a tooth wear category can be assigned to the at least one input element. For example, if the at least one input element is a specific pixel, and the at least one probability value representing the likelihood that the specific pixel corresponds to the tooth wear category "yes" is 95%, the tooth wear category "yes" is assigned to that specific pixel. Thus, the tooth wear category with the highest probability value can be assigned to the at least one input element. As mentioned, the tooth wear category "yes" indicates that tooth wear is present at the specific pixel.

[0048] Based on the tooth wear category assigned to the at least one input element, the tooth wear can be annotated on at least a portion of the virtual 3D model. As mentioned, the at least one input element corresponds to at least a portion of the virtual 3D model. In this way, the tooth wear is associated with the at least one portion of the virtual 3D model. The correspondence of the at least one input element to the at least one portion of the virtual 3D model can be established with the aid of an identifier. For example, the at least one input element can be a point with an identifier that defines the position of the point in the virtual 3D model. The same principle can be applied to other types of input elements, such as vertices, facets, voxels, pixels.

[0049] Once the tooth wear has been marked, the at least portion of the virtual 3D model can be displayed. In this way, information regarding the presence of tooth wear is conveyed to the user. For example, the at least portion of the virtual 3D model can be highlighted to visually distinguish it from the rest of the virtual 3D model. Additionally or alternatively, the at least portion of the virtual 3D model can be colored a different color or circled to visually distinguish the at least portion of the virtual 3D model from the rest of the virtual 3D model. Alternatively, a line defining the boundary between the at least portion of the virtual 3D model and the rest of the virtual 3D model can be highlighted.

[0050] According to an embodiment of the present disclosure, the tooth wear category can be a tooth wear severity value. In this embodiment, a trained neural network can perform a classification task and can classify the at least one input element, such as one or more 2D images of the tooth surface of the at least one tooth, into a corresponding tooth wear severity value. Thus, a tooth wear severity value can be assigned to each tooth surface of the at least one tooth. As previously mentioned, the tooth surface of the at least one tooth in the virtual 3D model can be the lingual surface, buccal surface, mesial surface, distal surface and / or occlusal surface of the tooth.

[0051] Multiple tooth wear severity values can be defined. In one example, the tooth wear severity value can correspond to the value of a clinically valid tooth wear index such as the tooth wear index (TWI). Generally, TWI includes a five-point ordinal scale for grading various tooth surfaces, with a score of zero to four (0 to 4). For example, a TWI score of four indicates the most severe tooth wear condition, with complete loss of enamel or pulp exposure. Instead of TWI, the Lussi index, the Basic Acid Etch Wear Examination (BEWE) index can also be used to define the tooth wear severity value. The advantage of using a clinically valid tooth wear index is that the user can easily interpret the tooth wear indication.

[0052] The trained neural network can perform a classification task in the manner described above to classify at least one input element, such as one or more 2D images of at least one tooth surface of at least one tooth in a virtual 3D model, into an associated tooth wear severity value according to clinically accepted criteria for tooth wear.

[0053] The output from the trained neural network may be in a fixed-size vector format that may be derived from a compact representation of the input. The values in the output may include the at least one probability value representing the likelihood that the at least one input element corresponds to a tooth wear category.

[0054] In one embodiment, the at least one input element with the assigned tooth wear category can indicate the location of a tooth wear lesion on the virtual 3D model. Thus, the output from the trained neural network can indicate an area on the virtual 3D model corresponding to a portion of a tooth condition that is missing due to tooth wear. In this case, the trained neural network can perform a semantic segmentation task, wherein a tooth wear category can be assigned to the at least one input element, which is a pixel, facet, or point. Thus, the effect of performing the semantic segmentation task is to identify the exact location and size of the tooth wear lesion on the virtual 3D model.

[0055] In one embodiment, obtaining an output from the trained neural network may include obtaining a segmentation map.

[0056] The segmentation map can be used to assign a tooth wear class to the at least one input element. For example, a tooth wear class can be assigned to pixels of an input 2D image. This assignment can be achieved by superimposing the input, such as a 2D image, with a segmentation mask.

[0057] The tooth wear category may be a predefined numerical value. For example, a value of "one" (1) may indicate tooth wear, a value of "two" (2) may indicate tooth surface, and a value of "three" (3) may indicate gum.

[0058] The segmentation mask may indicate the location of the structure of interest within the input, e.g. the location of the structure of interest within the input 2D image. The segmentation mask may be a 2D vector of the same size as the input. Thus, the segmentation mask has the same number of spatial elements as the input and represents a reconstruction of the input. Each value of the segmentation mask may be a tooth wear class. For example, for a given pixel or facet of the input, the trained neural network may assign a tooth wear class. For example, the tooth wear class may have a value of "one" to indicate the presence of tooth wear. The tooth wear class may have a value of "two" to indicate a tooth surface or a value of "three" to indicate a gum. In this way, the at least one input element, e.g. a pixel, point, voxel or facet of the input, may be assigned a corresponding tooth wear class.

[0059] In one example, the output from a trained neural network can be reconstructed into a 2D format, such as a 2D image, that can be overlaid or projected onto a virtual 3D model.

[0060] In one example, the displaying step may include displaying a tabular overview showing each tooth surface of the at least one tooth in the virtual 3D model and the corresponding tooth wear severity value. The tabular overview may be in the form of a dental chart. Alternatively or additionally, the tooth wear severity value may be displayed directly on the virtual 3D model, for example, adjacent to the associated tooth surface.

[0061] According to one embodiment, the trained neural network may be a trained convolutional neural network (CNN). The trained CNN may include multiple convolutional layers capable of capturing low-level features of the input, i.e., obtaining convolutional features. In addition, one or more pooling layers may be present to reduce the spatial size of the convolutional features. Convolutional neural networks may be particularly suitable for processing matrix information, such as 2D images.

[0062] Another example of a trained neural network that is particularly suitable for performing classification tasks is a trained residual network (ResNet) or a variant thereof.

[0063] An example of a trained neural network suitable for performing semantic segmentation tasks is a trained U-Net or its variants, such as U-Net++.

[0064] In an alternative embodiment, the trained neural network may be a trained transformer network. The transformer network may be suitable for processing matrix information such as 2D images and may have similar performance to a convolutional neural network.

[0065] In the case of a three-dimensional (3D) format of input (such as a point cloud), a suitable neural network architecture can be a PointNet network or its variants. The PointNet network is capable of processing input in a 3D format, such as a point cloud.

[0066] Alternatively or additionally, the tooth wear category may be a tooth wear type.

[0067] Three different tooth wear type categories may be defined, such as abrasion, wear, and tooth erosion. The trained neural network may classify at least one input element according to the defined tooth wear types. In this case, the at least one input element may be a 2D image or multiple 2D images of at least one tooth surface of at least one tooth in the virtual 3D model.

[0068] In one embodiment, the tooth wear category may include a tooth wear severity value and a tooth wear type. In this case, the trained neural network may perform two simultaneous classification tasks. In this case, the trained neural network may be a multi-task learning architecture. In this type of architecture, different branches may be employed to determine the tooth wear severity value and the tooth wear type in the input. Alternatively, the trained neural network may include two different trained neural networks. The tooth wear severity value may be assigned to the at least one input element based on the output from the first trained neural network. The tooth wear type may be assigned to the at least one input element based on the output from the second trained neural network.

[0069] In an embodiment, the tooth wear category may include the location of the tooth wear lesion. In this way, the clinical utility of the method according to the present disclosure is increased, as the tooth wear lesion may be pointed out in the virtual 3D model.

[0070] Tooth wear lesions can be areas of tooth loss due to tooth wear. A trained neural network can be capable of performing semantic segmentation tasks to indicate areas of tooth loss due to tooth wear. This can be achieved by generating a segmentation mask that indicates the location of specific structures of interest within an input, such as a 2D image. In this way, the trained neural network can detect where tooth wear is present within the input.

[0071] The tooth wear category may be a single category parameter, for example including only a tooth wear severity value, or a multi-category parameter, for example including a tooth wear severity value and a tooth wear type.

[0072] In one embodiment, displaying at least a portion of the virtual 3D model with the annotated tooth wear can include displaying tooth wear lesions. In this manner, the user is visually presented with the tooth wear indications visible on the virtual 3D model. This is particularly beneficial to the user because it allows tooth wear to be detected, quantified, and presented, thereby avoiding subjective visual inspection.

[0073] In an embodiment, obtaining an output from the trained neural network may include obtaining a probability matrix, wherein the probability matrix allows assigning a tooth wear category to the at least one input element of the virtual 3D model.

[0074] In one embodiment, the tooth wear classification can include a tooth wear severity and a location of the tooth wear lesion. In this case, the trained neural network can perform both a classification task and a semantic segmentation task. To this end, the trained neural network can include two different trained neural networks, wherein the tooth wear severity can be assigned to the at least one input element based on the output of a first trained neural network that performs the classification task. Additionally, the location of the tooth wear lesion can be determined based on the output of a third trained neural network that performs the semantic segmentation task.

[0075] According to one embodiment of the present disclosure, the method may further include determining a distribution of tooth wear categories in the input and selecting outliers in the determined distribution. Based on the selected outliers, a tooth wear indication may be identified and displayed. For example, the distribution of tooth wear categories in the input tooth surfaces may include surfaces classified as having a tooth severity value of 4.

[0076] In this way, a custom reference point regarding the state of tooth wear can be determined.The custom reference point regarding the state of tooth wear may also include characteristics of the patient, such as age, sex, race and / or region of residence.

[0077] The selection of outliers can be based on Linear Discriminant Analysis (LDA), Principal Component Analysis (PCA) or similar methods.

[0078] Furthermore, a data processing device is disclosed, comprising means for performing the method as set forth in any one or more of the presented embodiments.

[0079] The data processing device may be configured to perform a method for detecting tooth wear on a virtual 3D model of a dental case, wherein the method comprises the steps of:

[0080] - receiving, by the processor, a virtual 3D model of the dental condition;

[0081] - generating an input for the trained neural network for tooth wear detection from the virtual 3D model, wherein the input comprises the at least one input element corresponding to the at least one portion of the virtual 3D model, the input element comprising geometric information of the at least one portion of the virtual 3D model;

[0082] - Provide input to the trained neural network;

[0083] - obtaining an output from the trained neural network based on the provided input, wherein the output comprises at least one probability value, wherein the at least one probability value represents a likelihood that the at least one input element corresponds to a tooth wear category;

[0084] - assigning a tooth wear category to the at least one input element;

[0085] - marking the tooth wear on the at least one portion of the virtual 3D model based on the tooth wear category assigned to the at least one input element;

[0086] - displaying the at least part of the virtual 3D model together with the annotated tooth wear.

[0087] A computer program product is further disclosed. According to an embodiment, the computer program product may include instructions which, when executed by a computer, cause the computer to perform the method as described in any one or more of the embodiments set forth.

[0088] The computer program product may comprise instructions which, when executed by a computer, cause the computer to perform a method for detecting tooth wear on a virtual 3D model of a dental condition, wherein the method comprises the steps of:

[0089] - receiving, by the processor, a virtual 3D model of the dental condition;

[0090] - generating an input from the virtual 3D model for a trained neural network for tooth wear detection, wherein the input comprises the at least one input element corresponding to the at least one portion of the virtual 3D model, the input element comprising geometric information of the at least one portion of the virtual 3D model;

[0091] - providing said input to said trained neural network;

[0092] - obtaining said output from said trained neural network based on the provided input, wherein said output comprises said at least one probability value, wherein said at least one probability value represents said likelihood that said at least one input element corresponds to said tooth wear category;

[0093] - assigning said tooth wear category to said at least one input element;

[0094] - marking the tooth wear on the at least one portion of the virtual 3D model based on the tooth wear category assigned to the at least one input element;

[0095] - displaying the at least part of the virtual 3D model together with the annotated tooth wear.

[0096] Furthermore, the present disclosure includes a non-transitory computer-readable medium. The non-transitory computer-readable medium may include instructions, which, when executed by a computer, cause the computer to perform the method as described in any one or more of the embodiments.

[0097] The non-transitory computer readable medium may include instructions that, when executed by a computer, cause the computer to perform a method for detecting tooth wear on a virtual 3D model of a dental condition, wherein the method comprises the steps of:

[0098] - receiving, by the processor, the virtual 3D model of the dental condition;

[0099] - generating an input from the virtual 3D model, the input being used for a trained neural network for tooth wear detection, wherein the input comprises the at least one input element corresponding to the at least one portion of the virtual 3D model, the input element comprising geometric information of the at least one portion of the virtual 3D model;

[0100] - providing said input to said trained neural network;

[0101] - obtaining said output from said trained neural network based on the provided input, wherein said output comprises said at least one probability value, wherein said at least one probability value represents said likelihood that said at least one input element corresponds to said tooth wear category;

[0102] - assigning said tooth wear category to said at least one input element;

[0103] - marking tooth wear on the at least one portion of the virtual 3D model based on the tooth wear category assigned to the at least one input element;

[0104] - displaying said at least part of said virtual 3D model together with the annotated tooth wear.

[0105] Furthermore, the present disclosure includes a computer-readable medium. The computer-readable medium may include instructions, which, when executed by a computer, cause the computer to perform the method as described in any one or more of the embodiments.

[0106] In one embodiment of the present disclosure, a computer-implemented method for training a neural network for tooth wear detection is disclosed, wherein the method comprises:

[0107] a) receiving, by a processor, a training dataset comprising a plurality of virtual 3D models of dental conditions with associated labels of tooth wear lesions, wherein the labels represent true tooth wear class values;

[0108] b) generating an input to the neural network for tooth wear detection from the training dataset, wherein the input comprises an input element of a virtual 3D model from the plurality of virtual 3D models, geometric information in the form of curvature information associated with the input element, and a real tooth wear category value associated with the input element;

[0109] c) providing the input to the neural network;

[0110] d) obtaining an output of the neural network, the output comprising a predicted tooth wear category value for the input element;

[0111] e) comparing the predicted tooth wear category value for the input element with the actual tooth wear category value for the input element to obtain a loss function value;

[0112] f) updating the weights of the neural network based on the loss function value;

[0113] g) Repeat steps c) to f) until the stopping criterion is met.

[0114] Tooth wear lesions in a training dataset may be manually annotated with labels representing true tooth wear class values, also referred to as "ground truth."

[0115] Training a neural network to recognize tooth wear may be based on geometric information associated with input elements of the virtual 3D model of the training dataset. The geometric information may include curvature information.

[0116] In an example, obtaining an output can further include applying a smoothing factor to the input elements as a function of the position of the input elements within the tooth wear lesion. The smoothing factor can refer to a probability value assigned to the input elements to account for an imperfect manual labeling process. Thus, input elements, such as facets on the boundary of a tooth wear lesion, can be assigned a different probability than non-boundary facets. Consequently, the neural network's confidence in predicting tooth wear at boundary elements (e.g., facets) can be slightly reduced.

[0117] Applying the smoothing factor may include assigning a probability within the range [0.35, 0.65] to input elements that are located on a boundary of a tooth wear lesion. The probability range [0.35, 0.65] represents a probability of 35% to 65%.

[0118] Applying a smoothing factor may include assigning probabilities in the range [0.2, 0.8] to input elements that are non-boundary input elements (closer to the center) of the tooth wear lesion. The probability range [0.2, 0.8] represents a probability of 20% to 80%.

[0119] The training dataset can include both virtual 3D models with color information and virtual 3D models without color information. The virtual 3D models without color information can comprise 30% to 40% of the training dataset. Therefore, the neural network will not rely solely on color information for predicting tooth wear, but will use it as an important auxiliary tool while prioritizing geometric features that describe tooth wear lesions.

[0120] According to an embodiment of the training method, the training method may include providing a virtual 3D model without color information to the neural network in one training instance, and providing a virtual 3D model with color information to the neural network in another training instance.

[0121] The training dataset may include virtual 3D models with dental fillings. The neural network may assign a "none" class to dental fillings identified in the output, thereby ignoring dental fillings in the process of identifying tooth wear.

[0122] The training dataset can include a plurality of virtual 3D models of dental conditions that do not have labels associated with tooth wear lesions. In one example, a maximum percentage of the plurality of virtual 3D models of dental conditions that do not have labels associated with tooth wear lesions ranges from 5% to 40% of the training dataset. In this manner, a balance can be achieved during training based on how often the neural network processes samples from the training dataset that include labels (target annotations) and how often samples do not include any labels (target annotations).

[0123] The neural network can be pre-trained to recognize individual teeth and gums of the training dataset. This has the advantage of improving segmentation results for the task of tooth wear lesion segmentation, as the pre-trained neural network is a better starting point for identifying tooth wear than a neural network with randomly initialized weights.

[0124] In one example, the loss function is a combination of a distribution-based loss and a region-based loss. Using either the distribution-based loss or the region-based loss in isolation may not be sufficient to guide the training process.

[0125] The stopping criterion may be the number of training epochs before a performance metric stops increasing. A performance metric may be a loss function value. BRIEF DESCRIPTION OF THE DRAWINGS

[0126] The embodiments of the present disclosure can be best understood from the following detailed description taken in conjunction with the accompanying drawings. The drawings are schematic and simplified for clarity, and they only show details that are useful for improving understanding of the claims, with other details omitted. The individual features of each embodiment can each be combined with any or all features of the other embodiments. These and other embodiments, features, and / or technical effects will become apparent from and will be elucidated with reference to the description below, in which:

[0127] Figure 1 A virtual 3D model of the teeth is shown.

[0128] Figure 2 is a flowchart illustrating a method according to an embodiment.

[0129] Figure 3 A trained neural network used in a method according to an embodiment is shown.

[0130] Figure 4 A displayed portion of a virtual 3D model is shown with a plurality of teeth highlighted as being affected by tooth wear according to an embodiment of the invention.

[0131] Figure 5 A displayed portion of a virtual 3D model is shown with highlighted tooth wear on two tooth surfaces of a single tooth according to an embodiment of the invention.

[0132] Figure 6 A displayed portion of a virtual 3D model is shown with highlighted tooth wear lesions located on a plurality of teeth according to an embodiment of the invention.

[0133] Figure 7 A dental scanning system was demonstrated. DETAILED DESCRIPTION

[0134] In the following description, reference is made to the accompanying drawings that show, by way of illustration, how the invention may be practiced.

[0135] Figure 1 A user interface 100 is shown having a displayed virtual 3D model 101 of a patient's dental condition. Virtual 3D model 101 can be displayed on a display screen in the form of a 3D mesh, a point cloud, a 3D graph, a volumetric representation, or any other suitable 3D representation. 3D model 101 can represent the patient's dental condition, i.e., 3D model 101 can include a combined teeth and gum representation of the patient's dental condition. User interface 100 can include a button 104 that, upon user activation, initiates a method for analyzing virtual 3D model 101 to identify the presence of tooth wear. The method can also be automatically initiated without user activation.

[0136] To identify and separate individual teeth 102 and gums 103 within virtual 3D model 101, virtual 3D model 101 can be segmented. This means that facets belonging to individual teeth 102 according to the Universal Notation (UNN) can be determined, such as facets of a 3D mesh representation of the virtual 3D model. In this way, individual teeth 102 or portions of individual teeth 102 can be analyzed to determine the presence of tooth wear. Segmentation of virtual 3D model 101 is an optional step, and the entire virtual 3D model 101 can be analyzed to determine the presence of tooth wear instead.

[0137] Figure 2 A flow chart illustrating a method 200 according to an embodiment is shown.

[0138] First, in step 201, a processor may receive a virtual 3D model 101. The virtual 3D model 101 may be stored in a computer system memory, for example, in Standard Triangulation Language (STL) format. The processor may receive the virtual 3D model 101, for example, when a user touches a button 104 in the user interface 100. The virtual 3D model 101 may be displayed on a display screen in the form of a 3D mesh, a point cloud, a 3D graph, a volume representation, or any other suitable 3D representation.

[0139] Step 202 of method 200 illustrates generating an input from the virtual 3D model 101 for use in a trained neural network for tooth wear detection. The input may include at least one input element corresponding to at least a portion of the virtual 3D model 101. The at least one input element may include geometric information of at least a portion of the virtual 3D model 101.

[0140] The input to the trained neural network can be in 2D or 3D format. For example, the input can include one or more 2D images of at least one tooth surface of at least one tooth 102 in the virtual 3D model 101. In one example, the input can include one or more pixels of one or more 2D images. In another example, the input can include one or more points representing a point cloud of the virtual 3D model 101, or one or more facets of a 3D mesh representation of the virtual 3D model 101.

[0141] Any of the above items may be referred to as input elements. Each input element may correspond to a portion of the virtual 3D model 101, and this correspondence may be defined by an input element identifier. For example, the one or more points representing the point cloud of the virtual 3D model 101 may have an identifier that defines the location of the one or more points in the point cloud.

[0142] One or more of the 2D images may be generated by a virtual camera, also referred to as a camera, that takes snapshots of individual teeth 102 of the virtual 3D model 101. For example, the teeth 102 in the virtual 3D model 101 may be positioned so that the desired tooth surface of the teeth 102 faces the camera. The teeth 102 may be rotated and translated via a local coordinate system so that one or more of the tooth surfaces can be photographed. This may be an automated process that has already set predetermined movements of the teeth 102 to capture one or more of the tooth surfaces via the virtual camera.

[0143] The individual teeth 102 may be identified by segmenting the virtual 3D model 101 into a plurality of individual teeth 102 and gums 103. Each tooth 102 from the plurality of individual teeth may include a plurality of tooth surfaces.

[0144] Segmentation of the virtual 3D model 101 can be accomplished in several ways. According to an example, segmentation can include using surface curvature to identify the boundaries of the tooth representation. A curvature threshold can be selected to distinguish the tooth boundary region from the rest of the surface.

[0145] In another example, segmentation of the virtual 3D model 101 can include identifying tooth boundaries using a harmonic field. On the virtual 3D model 101, the harmonic field can be a scalar attached to each mesh vertex that satisfies the condition: ΔΦ = 0, where Δ is the Laplacian operator subject to Dirichlet boundary constraints. The above equation can be solved, for example, using the least squares method to calculate the harmonic field. The segmented teeth 102 can then be extracted by selecting the best contour line connecting data points with the same value as the tooth representation boundary.

[0146] In yet another example, segmenting the virtual 3D model 101 can include using a segmentation machine learning model. Specifically, the virtual 3D model 101 can be converted into a series of 2D virtual images. The segmentation machine learning model can be applied to the series of 2D virtual images. For each 2D virtual image, segmentation can be performed to distinguish different tooth classes and gums. After the classification of each element of each 2D virtual image, a back projection onto the virtual 3D model 101 can be performed. This method for segmenting the virtual 3D model 101 can be advantageous because the segmentation machine learning model utilizes the series of 2D virtual images to produce a fast and accurate facet classification overall.

[0147] The geometric information included in the at least one input element may be depth information related to a camera position, an angle between a facet normal and a camera direction, the facet normal and / or curvature information.

[0148] Step 203 represents providing input to the trained neural network 300 for tooth wear detection. The trained neural network 300 can be adapted to process the input, whether the input is in 2D or 3D format. By applying the trained neural network 300 to the generated input, the presence of tooth wear in the input is identified, which represents at least a portion of the dental condition.

[0149] Step 204 of method 200 represents obtaining an output from the trained neural network 300 based on the provided input, wherein the output may include at least one probability value, wherein the at least one probability value represents a likelihood that at least one input element corresponds to a tooth wear category.

[0150] In one example, the tooth wear category may be a tooth wear severity value. In another example, the tooth wear category may be an indication of the presence of tooth wear on at least one input element. For example, the indication of the presence of tooth wear may be "yes" to indicate the presence of tooth wear, or "no" to indicate the absence of tooth wear, also referred to as the absence of teeth on at least one input element. In yet another example, the tooth wear category may be a tooth wear type. The tooth wear category may be any combination of an indication of the presence of tooth wear, a tooth wear severity value, and / or a tooth wear type.

[0151] In the case of the obtained output, it is possible to assign a tooth wear category to at least one input element. This is shown in step 205 of method 200. Thus, the at least one input element can be assigned the tooth wear category having the highest probability value in the output.

[0152] Step 206 of method 200 illustrates marking tooth wear on at least a portion of the virtual 3D model 101 based on the tooth wear category assigned to the at least one input element. In this way, tooth wear can be associated with at least a portion of the virtual 3D model 101. The correspondence between the at least one input element and the at least one portion of the virtual 3D model 101 can be established with the aid of an identifier. For example, the at least one input element can be a point with an identifier that defines the position of the point in the virtual 3D model 101. The same principle can be applied to other types of input elements. The identifier associated with the at least one input element allows the tooth wear to be assigned to the at least one portion of the virtual 3D model 101, the at least one portion corresponding to the portion of the dental condition affected by the tooth wear.

[0153] Step 207 of the method 200 represents displaying the at least a portion of the virtual 3D model 101 together with the annotated tooth wear.

[0154] The at least one portion of the virtual 3D model 101 can be displayed on a display screen. Additionally or alternatively, the display can include playing an audio signal when tooth wear is detected. The display step can indicate to the user the at least one portion of the virtual 3D model 101 with the annotated tooth wear. The user can see on the virtual 3D model 101 that tooth wear exists on the at least one tooth 102 in the virtual 3D model 101, the severity of the tooth wear, and / or the exact location of the tooth wear lesion.

[0155] There may be various ways to display the at least one portion of the virtual 3D model 101 with the annotated tooth wear. For example, the color, transparency, zoom level, and / or shading of the at least one portion of the virtual 3D model 101 may vary relative to the rest of the virtual 3D model 101.

[0156] Figure 3 A trained neural network 300 used in a method according to an embodiment is shown. In this case, the trained neural network 300 may be a convolutional neural network and may include a plurality of convolutional layers 302 capable of capturing low-level features of the input, i.e., obtaining convolutional features. In addition, one or more pooling layers 303 may be present to reduce the spatial size of the convolutional features. Convolutional neural networks may be particularly suitable for processing matrix information, such as 2D images.

[0157] The input 301 may be in 2D or 3D format and may include at least one input element. Figure 3An input in a 2D format is shown. The at least one input element may be a 2D image of a tooth 102 whose occlusal surface is affected by tooth wear. The at least one input element may alternatively include one or more of the 2D images of the occlusal surface of the tooth 102 or at least one pixel of the 2D image of the tooth 102.

[0158] The output 305 may include the at least one probability value representing a likelihood that the at least one input element corresponds to a tooth wear category.

[0159] exist Figure 3 In the case of , the tooth wear category is an indication of the presence of tooth wear on the at least one input element of the input 301. In this case, the trained neural network 300 can perform a classification task to classify the input 301 into a "yes" or "no" output category. Alternatively or additionally, the tooth wear category can include a tooth wear severity value associated with the tooth wear on the tooth 102. For example, the tooth wear severity value associated with the tooth wear on the tooth 102 can indicate "severe", where the "severe" value can correspond to a tooth wear index (TWI) value of 4.

[0160] The training process for a classification task for a trained neural network 300 can be described as follows. The overall goal of an untrained neural network can be to develop a trained neural network 300 that is capable of classifying inputs according to a desired tooth wear category. The tooth wear category can be an indication of the presence of tooth wear, a tooth wear severity value, and / or a tooth wear type.

[0161] The training process may be referred to as "supervised training" due to the use of an already labeled training dataset. The training dataset may be in a 2D format, i.e. it may comprise a plurality of 2D training images of teeth or parts of teeth. Additionally, the training dataset may comprise associated labels, wherein the labels correspond to tooth wear categories. The labeling may be performed, for example, by a qualified dentist. The training dataset may be labeled according to an indication of the presence of tooth wear, for example wherein a "yes" or "no" label indicates the presence or absence of tooth wear, respectively. Additionally or alternatively, the training dataset may be labeled with a tooth wear severity label and / or a tooth wear type label associated with the corresponding tooth wear state.

[0162] Labels can be thought of as targets or true values for a given training dataset.

[0163] The training dataset may include at least 1,000 3D scans that may yield over 17,000 individual teeth.

[0164] A training dataset can be fed into an untrained neural network. The output of the untrained neural network can be a probability value indicating which output category an input or input element belongs to. The error between the true value and the output can be determined. This error can also be considered a loss function. An example of such a loss function is the mean squared error function. To train the untrained neural network, this loss function should be minimized. Minimizing the loss function can be achieved, for example, using a stochastic gradient descent algorithm. By minimizing the loss function, the weights of the untrained neural network can be updated.

[0165] The training process can be an iterative process where the same training dataset can be iteratively fed into an untrained neural network until a stopping criterion is reached. The stopping criterion can be a specific threshold of the loss function. The stopping criterion can be the number of epochs (training cycles) before the performance metric stops improving.

[0166] Due to the nature of tooth wear lesions, they typically represent a small fraction of the total number of input samples (e.g., teeth) of the training dataset. This means that the choice of loss function or loss functions to guide the training process should cater to this characteristic. The loss function can be used as a performance metric.

[0167] Due to the imbalance in the number of elements between the background and target classes, a single-shot distribution-based loss, such as the cross-entropy loss function, may not be suitable for tooth wear recognition. Even with weighting techniques, neural networks may perform poorly with this type of method. Given that the distribution of elements across classes does not directly impact the computation, region-based losses, such as the Dice loss or the Tversky loss, are more appropriate.

[0168] In one embodiment, the loss function may be a region-based loss function. In another embodiment, the loss function may be a combination of a distribution-based loss and a region-based loss, which may produce optimal results for the tooth wear lesion segmentation task. Using either the distribution-based loss or the region-based loss in isolation may not be sufficient to guide the training process.

[0169] The training dataset may alternatively be in a 3D format comprising a plurality of teeth represented as a 3D mesh, a 3D graph or a point cloud.The tooth wear category may be an indication of the presence of tooth wear, a tooth wear severity value and / or a tooth wear type.

[0170] The labeling (annotation) of the training dataset is typically done by annotators using a brush-like tool. This imposes a limit on how precise the annotations can be due to the nature of the tool and the granularity of the samples from the training dataset (e.g., a sample 3D mesh can include hundreds of thousands of facets). To account for this loss in accuracy, it is possible to introduce a higher smoothing factor at the borders of the tooth wear lesions compared to the center of the tooth wear lesions. For example, facets on the borders of the tooth wear lesions have a higher smoothing factor than facets at the center of the tooth wear lesions. This means that the uncertainty / impossibility of millimeter precision during the labeling process is transferred to the neural network training process, as elements of the tooth wear lesion that are closer to its border and therefore less reliable from an annotation perspective are modeled as more uncertain. This concept can be applied to the neural network training process, where the probabilistic labels of all elements (e.g., facets) at the borders of the tooth wear lesions are highly smoothed, e.g., assigned probabilities of [0.35, 0.65], in contrast to the rest of the elements closer to the center of the tooth wear lesion (which are less smoothed, e.g., assigned probabilities of [0.2, 0.8]). Therefore, the confidence of the neural network in predicting tooth wear at boundary elements (such as facets) can be slightly reduced.

[0171] Regardless of the format of the training dataset, the training dataset can include geometric information associated with the tooth or parts of the tooth. Similar to the input elements of the trained neural network input, the geometric information of the training dataset can include curvature information related to the tooth or parts of the tooth, facet normal information, depth information related to the virtual camera position, and / or angles between the facet normals and the virtual camera direction.

[0172] The curvature information of the training dataset can be correlated with the successful training process of an untrained neural network. For example, a sudden change in relative curvature between regions in a tooth surface in the training dataset can indicate tooth wear.

[0173] The depth information of the training dataset may be relevant to the training process, as information about the tooth surface topology of the teeth may be included from the training dataset.

[0174] The facet normal information of the training dataset may be relevant to the training process. Adjacent facets of teeth from the training dataset may have significant differences between their facet normals, indicating tooth wear caused by sudden topological changes.

[0175] The training dataset can include color information for teeth or parts of teeth. Severe stages of tooth wear can be characterized by the loss of tooth material called dentin, leaving it exposed. The color difference between the white hue of enamel and the yellow color of dentin is a strong indicator of severe tooth wear. Therefore, this information can be highly relevant during the training of an untrained neural network for detecting tooth wear.

[0176] In one embodiment, the training dataset may include samples with color information and samples without color information. For example, only some samples (2D images, 3D meshes, 3D graphics, or point clouds) of the training dataset may include color information, while other samples of the training dataset may have the color information values of all their elements set to zero. In one example, the samples with color information set to zero may comprise 30% to 40% of the total number of samples in the training dataset.

[0177] The partial presence of color information in the training dataset reflects the behavior of human annotators, who typically disable the color of 3D scans when assigning labels (annotations) for tooth wear in order to focus solely on the geometric features of the 3D scans. Therefore, color is a visual aid that is typically turned off during the annotation process, even though tooth wear lesions are still visible and recognizable with color information turned on. This behavior can be mimicked during the training of the neural network, thus forcing the neural network to detect tooth wear based solely on geometric information without color.

[0178] Furthermore, due to the cyclical nature of the training process, a neural network may often receive samples from the training dataset where the color information for that sample is set to zero during one training instance and the color information for that sample is present during another training instance. Therefore, the neural network will not rely solely on color information for predicting tooth wear, but will use it as an important auxiliary tool while prioritizing geometric features that describe tooth wear lesions.

[0179] Optionally, the training dataset can include semantic information associated with the tooth or parts of the tooth. This type of information aids the training process, as an untrained neural network can learn how to process visual geometric information in an improved way by adding semantic context to the geometric information.

[0180] Alternatively, the training dataset can include samples with dental fillings. Given that the neural network relies on geometric information about the samples, fillings can be mistaken for tooth wear lesions due to their smoothness. Therefore, samples containing fillings can be incorporated into the training dataset. The neural network is not designed to predict fillings, and it may learn that fillings are not part of the tooth wear lesion being predicted. Fillings can be set to the "None" class and thus ignored.

[0181] The training process for the segmentation task can be the same as the training process for the classification task.

[0182] In order to achieve a balance in the training process based on the frequency with which the neural network processes samples from the training dataset that include a label (target annotation) and the frequency with which samples do not contain any label (target annotation), a threshold for the maximum percentage of samples from the training dataset that do not contain any label (target annotation) can be set. In one example, the maximum percentage of samples from the training dataset that do not have any label can be in the range of 5% to 40% of the total number of samples.

[0183] For example, in the case where the input samples are teeth, this threshold can reflect the average fraction of teeth that do not have wear lesions. The goal of setting this threshold is to prevent the neural network from wasting too many training iterations processing samples from the training dataset that do not contribute to the detection of tooth wear lesions, while having a distribution of samples with and without tooth wear that is similar to the distribution of samples naturally found in clinical scenarios, with a ratio of approximately 70% of samples containing wear to 30% of samples without wear. Thus, improving the neural network's ability to minimize "false positive" detections.

[0184] In one embodiment, a neural network can be pre-trained to recognize individual teeth and gums for a training dataset prior to training for tooth wear lesion recognition. In this case, the training dataset can additionally include samples, such as a 3D mesh or point cloud with labeled teeth and gums. Thus, a pre-trained neural network for teeth and gum recognition can be used as a starting point for training for the tooth wear recognition task. This pre-training has been shown to improve segmentation results for the tooth wear lesion segmentation task, as the pre-trained neural network is a better starting point than a neural network with randomly initialized weights.

[0185] Figure 4 A displayed portion 400 of a virtual 3D model 101 is shown with a plurality of teeth 401-405 highlighted due to the effects of tooth wear according to an embodiment of the invention. The portion 400 of the virtual 3D model 101 may be a portion of the virtual 3D model 101, such as a portion of the mandible of the virtual 3D model 101. Figure 4In the case of tooth wear, the tooth wear category may be an indication of the presence of tooth wear. The presence of tooth wear may be in the form of a visual notification 407. Additionally, the presence of tooth wear may be indicated by displaying teeth 401 to 405 affected by tooth wear in a different color compared to teeth 406 not affected by tooth wear. The indication of the presence of tooth wear may additionally include an audio signal.

[0186] Figure 5 A displayed portion of a virtual 3D model is shown with highlighted tooth wear on two tooth surfaces of a single tooth according to an embodiment of the invention.

[0187] Figure 5 The displayed portion 500 of the virtual 3D model 101 is shown with tooth wear emphasized on two tooth surfaces 502, 503 of a tooth 501 according to an embodiment of the invention. Figure 5 In the case of , the tooth wear category may be an indication of the presence of tooth wear on the tooth surfaces 502, 503 of the teeth 501 that are part 500 of the virtual 3D model 101. Figure 5 As can be seen in FIG, the neck surface 503 and the occlusal surface 502 are affected by tooth wear. These surfaces can be displayed in a different color than teeth that are not affected by tooth wear. The indication of whether tooth wear is present can additionally include a visual notification 507.

[0188] Figure 6 A displayed portion 600 of a virtual 3D model 101 is shown with highlighted tooth wear lesions 602, 603 located on a plurality of teeth according to an embodiment of the invention. Figure 6 In the case of tooth wear, specific tooth wear lesions 602, 603 on the portion 600 of the virtual 3D model 101 can be shown. In contrast, Figure 4 and Figure 5 Shown are highlighted entire teeth or tooth surfaces affected by tooth wear. Figure 6 In the case of tooth wear, the tooth wear category may additionally include a tooth wear severity value 607. The three tooth wear severity values may be "no tooth wear," "moderate tooth wear," and "severe tooth wear" that may be associated with a tooth or a tooth wear lesion on the tooth. A tooth or tooth wear lesion associated with one of the tooth wear severity values may be a different color than a tooth or tooth wear lesion associated with another tooth wear severity value.

[0189] Figure 7A dental scanning system 700 is shown, which may include a computer 710 capable of performing a method according to any one or more embodiments of the present invention, such as method 200. The computer may include a wired or wireless interface with a server 715, a cloud server 720, and an intraoral scanner 725. The intraoral scanner 725 may be capable of recording scan data of a patient's dentition.

[0190] The dental scanning system 700 may include a processor configured to perform the method according to one or more embodiments of the present disclosure. The processor may be part of the computer 710 , the server 715 , or the cloud server 720 .

[0191] A non-transitory computer-readable storage medium may be included in the dental scanning system 700. The non-transitory computer-readable medium may carry instructions that, when executed by a computer, cause the computer to perform a method according to one or more embodiments of the present disclosure.

[0192] The computer program product may be embodied in a non-transitory computer-readable storage medium. The computer program product may include instructions that, when executed by a computer, cause the computer to perform a method according to any of the embodiments presented herein.

[0193] Although some embodiments have been described and shown in detail, the present disclosure is not limited to such details, but may also be embodied in other ways within the scope of the subject matter defined in the following claims. In particular, it should be understood that other embodiments may be utilized and structural and functional modifications may be made without departing from the scope of the present invention.

[0194] It should be understood that references throughout this specification to "one embodiment" or "an embodiment" or "aspect" or features that "may" include mean that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present disclosure. Furthermore, particular features, structures, or characteristics may be combined where appropriate in one or more embodiments of the present disclosure. The foregoing description is provided to enable any person skilled in the art to practice the various aspects described herein. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects.

Claims

1. A computer-implemented method for detecting tooth wear on a virtual 3D model of a dental condition, the method comprising: - receiving, by a processor, said virtual 3D model of said dental condition; - generating an input from the virtual 3D model for a trained neural network for tooth wear detection, wherein the input comprises at least one input element corresponding to at least a portion of the virtual 3D model, the at least one input element comprising geometric information of the at least a portion of the virtual 3D model, wherein the geometric information of the at least a portion of the virtual 3D model is curvature information; - providing said input to said trained neural network; - obtaining an output from the trained neural network based on the provided input, wherein the output comprises at least one probability value, wherein the at least one probability value represents a likelihood that the at least one input element corresponds to a tooth wear category; - assigning said tooth wear category to said at least one input element; - marking tooth wear on the at least one portion of the virtual 3D model based on the tooth wear category assigned to the at least one input element; - displaying said at least a portion of said virtual 3D model together with said annotated tooth wear.

2. The method according to claim 1, wherein The at least one input element comprises an identifier defining a position of the input element in the virtual 3D model.

3. A method according to any preceding claim, wherein: The geometric information of the at least a portion of the virtual 3D model further includes facet normal information.

4. A method according to any preceding claim, wherein: The geometric information of the at least a portion of the virtual 3D model further includes a difference in facet normals of at least two adjacent facets.

5. The method according to claim 4, wherein: The difference in the facet normals of at least two adjacent facets is the angle between the facet normals of the at least two adjacent facets.

6. A method according to any preceding claim, wherein: The tooth wear category indicates the presence of tooth wear on the at least one input element.

7. A method according to any preceding claim, wherein: The tooth wear category is a tooth wear severity value.

8. A method according to any preceding claim, wherein: The at least one input element comprises a 2D image of a tooth surface of the virtual 3D model.

9. A method according to any preceding claim 7 or 8, wherein The tooth wear severity value corresponds to the tooth wear index (TWI).

10. The method according to any preceding claim 7 to 9, further comprising displaying the tooth wear severity value directly on the virtual 3D model.

11. A method according to any preceding claim 7 to 10, wherein: The tooth wear category further includes a location of a tooth wear lesion on the virtual 3D model.

12. The method according to claim 11, wherein The tooth wear severity value is assigned to the at least one input element based on an output from a first trained neural network performing a classification task, and wherein the location of the tooth wear lesion is determined based on an output from a third trained neural network performing a semantic segmentation task.

13. The method according to any one of the preceding claims 1 to 6, wherein The tooth wear category is a tooth wear type.

14. The method according to claim 13, wherein: The tooth wear category further includes the tooth wear severity value.

15. The method according to claim 14, wherein The trained neural network includes different branches for determining the tooth wear severity value and the tooth wear type in the input.

16. The method according to claim 14, wherein The tooth wear severity value is assigned to the at least one input element based on the output from the first trained neural network, and wherein the tooth wear type is assigned to the at least one input element based on the output from the second trained neural network.

17. The method according to claim 16, wherein The trained neural network includes the first trained neural network and the second trained neural network.

18. A method according to any preceding claim, further comprising determining a distribution of the tooth wear categories in the input and selecting outliers to determine a custom reference point for tooth wear status.

19. The method according to claim 18, wherein The custom reference points regarding the tooth wear status include the patient's age, gender, race and / or region of residence.

20. The method of any preceding claim 18 or 19, further comprising identifying tooth wear indications based on the selected outliers and displaying the tooth wear indications.

21. A method according to any preceding claim, wherein: The input further includes semantic information of the virtual 3D model, wherein the semantic information includes a tooth identifier.

22. A method according to any preceding claim, wherein: The input further includes semantic information of the virtual 3D model, wherein the semantic information includes a tooth surface identifier.

23. A method according to any preceding claim, further comprising assigning a tooth profile class to the at least one input element.

24. A method according to any preceding claim, wherein: The input further includes color information of the at least one input element.

25. A method according to any preceding claim, wherein: Obtaining the output from the trained neural network includes obtaining a segmentation map for assigning the tooth wear class to the at least one input element.

26. A computer program product comprising instructions which, when the program is executed by a computer, cause the computer to perform the method according to any one of claims 1 to 25.

27. A non-transitory computer-readable medium comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 25.

28. A data processing apparatus comprising means for performing the method of any one of the preceding claims 1 to 25.

29. A computer-implemented method for training a neural network for tooth wear detection, the method comprising: a) receiving, by a processor, a training dataset comprising a plurality of virtual 3D models of dental conditions with associated labels of tooth wear lesions, wherein the labels represent real tooth wear class values; b) generating an input for the neural network for tooth wear detection from the training dataset, wherein the input comprises an input element of a virtual 3D model from the plurality of virtual 3D models, geometric information in the form of curvature information associated with the input element, and a real tooth wear category value associated with the input element; c) providing the input to the neural network; d) obtaining an output of the neural network, the output comprising a predicted tooth wear category value for the input element; e) comparing the predicted tooth wear category value of the input element with the actual tooth wear category value of the input element to obtain a loss function value; f) updating the weights of the neural network based on the loss function value; g) Repeat steps c) to f) until the stopping criterion is met.

30. The method according to claim 29, wherein Obtaining the output further comprises applying a smoothing factor to the input element as a function of the position of the input element within the tooth wear lesion.

31. The method according to claim 30, wherein Applying a smoothing factor involves assigning a probability in the range [0.35, 0.65] to input elements that are located on the boundary of said tooth wear lesion.

32. A method according to any preceding claim 30 or 31, wherein Applying a smoothing factor includes assigning probabilities within the range [0.2, 0.8] to input elements that are non-boundary input elements of the tooth wear lesion.

33. A method according to any one of claims 29 to 32, wherein The training data set includes a virtual 3D model with color information and a virtual 3D model without color information.

34. The method according to claim 33, wherein The virtual 3D model without color information comprises 30% to 40% of the training dataset.

35. The method of any one of claims 29 to 34, further comprising providing the neural network with a virtual 3D model without color information in one training instance, and providing the neural network with the virtual 3D model with color information in another training instance.

36. A method according to any one of claims 29 to 35, wherein The training dataset includes virtual 3D models with dental fillings.

37. The method of claim 36, further comprising assigning a "None" class to the dental filling identified in the output of the neural network.

38. A method according to any one of claims 29 to 37, wherein The training dataset includes a plurality of virtual 3D models of dental conditions without associated labels of tooth wear lesions.

39. The method according to claim 38, wherein A maximum percentage of the plurality of virtual 3D models of dental conditions having no associated labels of tooth wear lesions is in the range of 5% to 40% of the training dataset.

40. The method according to any one of the preceding claims 29 to 39, wherein The neural network is pre-trained to recognize individual teeth and gums of the training dataset.

41. The method according to any one of the preceding claims 29 to 40, wherein The loss function is a combination of distribution-based loss and region-based loss.

42. The method according to any one of the preceding claims 29 to 41, wherein The stopping criterion is the number of training epochs before the performance metric stops increasing.

43. The method according to claim 42, wherein The performance indicator is the loss function value.

Images