Method and system for detecting dental plaque

Abstract

According to one embodiment, a computer-implemented method for detecting plaque on a digitized 3D model (101) of a tooth condition is disclosed. The method comprises: receiving, by a processor, a digitized 3D model of a tooth condition (101); at least a portion of the digitized 3D model (101) is provided as input (601) to the trained neural network (600), wherein the at least a portion of the digitized 3D model (101) contains natural color information and / or geometric information associated with at least one face of the tooth condition. The geometric information includes curvature information, facet normal information, depth information related to the position of the virtual camera, and / or an angle between the facet normal and the direction of the virtual camera. The method further includes obtaining an output (605) of the trained neural network (600) based on the provided input (610), wherein the output (605) includes a plaque parameter associated with the at least one dental face; assigning plaque parameters to at least a portion of the digitized 3D model; and displaying the digitized 3D model, wherein the plaque parameters are assigned to at least a portion of the digitized 3D model (101).

Description

Technical Field

[0001] This disclosure relates to a computer-implemented method and system for detecting dental plaque in a digital 3D dental model of a patient's dentition. Background Technology

[0002] Dental plaque, or simply plaque, is a dental disease that forms a biofilm on human teeth. Plaque biofilm is a sticky, acidic biofilm formed by the breakdown of food debris by bacteria present in the mouth. It can be removed through regular professional cleanings, daily brushing, and flossing. However, if left untreated, it can lead to cavities, tartar, gingivitis, or other oral diseases. Therefore, timely plaque detection and monitoring by a dentist is crucial for protecting a patient's tooth structure and overall oral health.

[0003] Currently, the clinical detection and diagnosis of plaque primarily relies on direct visual and tactile examination of teeth by dentists. Developing agents (e.g., in tablet form) can be used to stain plaque with a bright color, typically red or blue. Some developing agents contain several dyes that may stain plaque at different stages of its development differently. For example, new plaque may stain red, while older, pathogenic plaque may stain blue.

[0004] The development of intraoral scanning technology has played a crucial role in the transition to modern digital dentistry. Using intraoral 3D scanners allows dentists to quickly and accurately capture the condition of a patient's oral cavity, which can then be visualized on a monitor as a digital 3D dental model. The resulting digital 3D dental model of the patient's teeth can thus serve as a digital impression, offering numerous advantages over traditional physical impressions. The overall efficiency of dental procedures is thus improved, for example, because the created digital impressions can be applied to many applications such as the design of dental restorations and orthodontic treatments. Furthermore, digital 3D dental models can be used for diagnostic purposes to detect and visualize the presence of dental diseases, such as plaque, caries, gingivitis, gingival recession, and / or tooth wear, to users (e.g., dentists).

[0005] There is a clear need to develop methods and systems capable of detecting plaque in teeth from a single digital 3D dental model. Dentists also need a solution to visualize plaque on digital 3D dental models for patients in a manner similar to the use of a developer. This would then allow for the design of timely dental hygiene improvement recommendations to protect patients' oral health.

[0006] In addition to identifying plaque in dental conditions, it is also necessary to digitally determine plaque thickness and plaque type, such as whether the plaque is newly formed low-risk plaque or old, pathogenic plaque. In particular, it is desirable to identify and visualize areas with high-risk pathogenic plaque. Summary of the Invention

[0007] In one embodiment, a computer-implemented method is disclosed for detecting dental plaque on a digital 3D model of dental conditions, wherein the method includes: - A digital 3D model of the teeth's condition is received by the processor; - At least a portion of a digital 3D model is provided as input to a trained neural network, wherein the at least a portion of the digital 3D model contains natural color information and / or geometric information associated with at least one tooth surface of a tooth condition; - The output is obtained from the trained neural network based on the provided input, wherein the output contains plaque parameters associated with the at least one tooth surface; - Assigning plaque parameters to at least a portion of the digital 3D model; and - Display a digital 3D model in which dental plaque parameters are assigned to at least a portion of the digital 3D model.

[0008] In one embodiment, a computer-implemented method for detecting dental plaque on a digital 3D model of tooth condition includes: - A digital 3D model of the teeth's condition is received by the processor; - At least a portion of a digital 3D model is provided as input to a trained neural network, wherein at least a portion of the digital 3D model contains natural color information and / or geometric information associated with at least one tooth surface of the tooth condition, wherein the geometric information includes curvature information, facet normal information, depth information related to the virtual camera position, and / or the angle between the facet normal and the virtual camera orientation. - The output is obtained from the trained neural network based on the provided input, wherein the output contains plaque parameters associated with at least one tooth surface; - Assign plaque parameters to at least a portion of the digital 3D model; - Display a digital 3D model in which dental plaque parameters are assigned to at least a portion of the digital 3D model.

[0009] Throughout this disclosure, the term "3D" refers to the term "three-dimensional". Similarly, the term "2D" refers to the term "two-dimensional". The term "digital 3D model of dental condition" refers to a digital, three-dimensional, computer-generated representation of a patient's dental condition.

[0010] This digital 3D model can accurately correspond to the actual condition of teeth, so that dental objects such as teeth, tooth surfaces, restorations and gums on the digital 3D model correspond to the dental objects of the actual condition of teeth.

[0011] Digital 3D models can be built by a processor based on scan data collected during intraoral scanning, which uses an intraoral scanner to scan the condition of a patient's teeth and gums. The digital 3D model can be stored in the memory of a computer system, for example, in Standard Triangle Language (STL) format.

[0012] The processor can receive or access digital 3D models. Digital 3D models are typically displayed on a screen as 3D meshes to represent the surface of teeth and gingival tissue in relation to dental conditions. A 3D mesh can consist of facets (e.g., triangular facets). Alternatively, digital 3D models can be displayed as point clouds containing points, graphs containing nodes and edges, volume representations containing voxels, or any other suitable 3D representation.

[0013] The method may include providing at least a portion of a digitized 3D model as input to a trained neural network. The at least a portion of the digitized 3D model may correspond to at least one tooth surface or a portion of at least one tooth surface of a tooth condition. The trained neural network may be adapted to process at least a portion of the digitized 3D model directly in its three-dimensional format, or may be adapted to process a two-dimensional representation of at least a portion of the digitized 3D model, such as a 2D image or multiple 2D images generated by taking a snapshot of the digitized 3D model, or a 2D image generated by an intraoral scanner. The 2D image generated by the scanner may include white light images, fluorescence images, and / or infrared images.

[0014] Three-dimensional or two-dimensional data structures that serve as input to a trained neural network can first be converted into numerical representations before being input into the trained neural network.

[0015] In one embodiment, the input to the trained neural network can be in a 3D format. Examples of 3D format inputs include point clouds, 3D meshes, volumetric representations, or graph representations of at least a portion of a digitized 3D model. The input can be converted from one 3D representation type, such as a mesh representation, to another 3D representation type, such as a point cloud representation. The type of trained neural network capable of processing 3D format inputs can be a trained PointNet network or a PointNet-type network. These neural networks can be capable of processing digitized 3D models or at least a portion of digitized 3D models in point cloud form.

[0016] Providing at least a portion of a digitized 3D model as input may include sampling at least a portion of the digitized 3D model or converting at least a portion of the digitized 3D model into a numerical form. For example, one or more points in a point cloud or mesh representation of the digitized 3D model may be converted into a numerical form, such as an input value matrix describing a feature set. This matrix can then form the input to a trained neural network.

[0017] If the digitized 3D model is in the form of a 3D mesh, it can be converted into a point cloud by folding all the vertices in the 3D mesh format, leaving only the points. The remaining point cloud can then be processed directly by a PointNet neural network or a PointNet-type neural network.

[0018] 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 a digitized 3D model. Alternatively, the input may be in a 2D format, such as a 2D image or multiple 2D images of at least one tooth surface of an actual tooth condition captured by an intraoral scanner.

[0019] Therefore, a trained neural network can be adapted to process 2D input formats. One or more 2D images can be generated from a digitized 3D model using a virtual camera (also called a camera). The virtual camera can take snapshots of individual teeth in the digitized 3D model. For example, teeth in the digitized 3D model can be positioned such that the desired tooth surfaces face the camera. Positioning the teeth can include rotation to align the virtual camera with the target tooth surface. Positioning the teeth can also include translation to bring the entire target tooth surface within the virtual camera's field of view, such as capturing all facets of the tooth surface. Both rotation and translation operations can be performed via the tooth's local coordinate system. This process can be automated, where predetermined movements of the teeth are set to capture one or more tooth surfaces via the virtual camera. Alternatively, the virtual camera can be moved to multiple predetermined positions to capture snapshots of the desired tooth surfaces. Alternatively, snapshots of individual teeth can be generated by manually moving the virtual camera to the desired positions.

[0020] A method according to one embodiment may include generating input for a trained neural network, wherein the input is one or more 2D images of at least one tooth surface in a digitized 3D model of a tooth, the method comprising: - Rotate the tooth using its local coordinate system so that the surface normal of at least one tooth surface coincides with the direction of the virtual camera. - Translate the tooth using the local coordinate system of the tooth so that at least one tooth surface is within the field of view of the virtual camera; - Create one or more 2D images by taking snapshots of at least one tooth surface using a virtual camera.

[0021] At least one tooth surface normal can coincide with the axis of the tooth's local coordinate system that protrudes through the labial tooth surface. The orientation of the virtual camera indicates the camera's aiming position.

[0022] In addition to generating images by taking snapshots of a digital 3D model using a virtual camera, multiple 2D images of at least one tooth surface of a digital 3D model can also be generated directly by an intraoral scanner that records the condition of the teeth.

[0023] Providing at least a portion of a digitized 3D model as input to a trained neural network may include generating one or more 2D images of at least one tooth surface.

[0024] At least one tooth surface in a digital 3D model may have a direct corresponding tooth surface in an actual tooth condition. Therefore, the process of generating one or more 2D images of at least one tooth surface in a digital 3D model using a virtual camera can correspond, for example, to taking a photograph of at least one tooth surface in an actual tooth condition using an intraoral scanner. The one or more 2D images generated in this way can be provided to a trained neural network.

[0025] The generated one or more 2D images can form the input to a trained neural network. Alternatively or additionally, the input can be at least one pixel of one or more 2D images.

[0026] In one embodiment, providing at least a portion of a digitized 3D model as input to a trained neural network may include providing a 3D mesh of at least a portion of the digitized 3D model to the trained neural network.

[0027] In another embodiment, providing at least a portion of the digitized 3D model as input to a trained neural network may include providing the trained neural network with a point cloud of at least a portion of the digitized 3D model.

[0028] In yet another embodiment, providing at least a portion of the digitized 3D model as input to a trained neural network may include providing a voxel representation of at least a portion of the digitized 3D model to the trained neural network.

[0029] At least a portion of the digital 3D model may contain natural color information and / or geometric information associated with at least one tooth surface of the tooth condition. The tooth surface may be any one of the occlusal surface, labial surface, buccal surface, mesial surface, distal surface, cervical surface, and / or incisal surface.

[0030] Natural color information and / or geometric information can be included in the scan data obtained via an intraoral scanner used to scan the condition of teeth.

[0031] Natural color information can relate to the surface color of the scanned teeth and / or gums and can be obtained by scanning the patient's intraoral situation using an intraoral scanner. This information can be expressed using red-green-blue (RGB) color intensity values. The presence of natural color information can be advantageous because its color difference compared to healthy, clean enamel facilitates the identification of plaque. Color differences on the tooth surface can indicate the presence of plaque because plaque is more yellow than healthy enamel. Therefore, natural color information can be highly correlated with trained neural networks during plaque detection. In addition to plaque, natural color information is particularly effective in detecting stains, tartar (also known as calcified plaque), and / or food debris.

[0032] Geometric information (also known as topological information) is included in at least a portion of the digitized 3D model and can be depth information related to the camera position, the angle between the facet normal and the camera orientation, facet normal and / or curvature information. The term "camera" as used herein refers to a virtual camera used to observe the digitized 3D model. This camera can be characterized by virtual camera properties such as field of view (FOV) and / or focal length. Examples of geometric information mentioned herein can serve as descriptors of geometric properties used as input to a trained neural network.

[0033] Depth information may include distance information of at least a portion of the digitized 3D model relative to a camera. The depth information may be in the form of a depth image. The depth image can display the at least a portion of the digitized 3D model based on its distance from the camera. For example, portions of the digitized 3D model farther from the camera may appear as darker shadows than portions closer to the camera. A distance threshold may be defined such that when the distance between the camera and the portion of the digitized 3D model is greater than the distance threshold, that portion of the digitized 3D model is not displayed in the depth information. The distance of at least a portion of the digitized 3D model from a given camera position indirectly describes the topology of that at least a portion of the digitized 3D model, and therefore describes the surface smoothness of that at least a portion of the digitized 3D model. Variations in surface smoothness can represent relevant parameters when detecting plaque in at least a portion of the digitized 3D model, since plaque has different surface waviness compared to a clean tooth surface.

[0034] Facet normals can be vectors perpendicular to the facets of a digitized 3D model. Facets can be considered "building units" of a digitized 3D model in a 3D mesh format, and if the digitized 3D model is in a triangular mesh format, then the facets can be, for example, triangles. A triangular facet can be defined by three points (vertices) interconnected by three edges. A small difference in the facet normals of two adjacent facets can indicate a smooth local area. However, a significant difference in the facet normals of adjacent facets may indicate the presence of plaque. Variations in two surface normals can be expressed by the angle between the two surface normals.

[0035] Curvature information can be presented as facet values ​​and can be obtained by measuring the change in facet normals along adjacent facets of a digital 3D model. This change provides insight into whether the associated portion of the digital 3D model is flat or curved. Thus, curvature information can describe the topology of teeth by revealing smooth and non-smooth surfaces. Plaque causes variations at the surface level of teeth, which can then be directly reflected in changes in surface curvature. Plaque can also be associated with localized smoothness (low-curvature areas) in certain regions. Therefore, curvature information can be an important parameter for detecting plaque presence, plaque location, plaque type, and / or plaque severity.

[0036] Both geometric and natural color information can be represented as numerical values ​​for each unit element (e.g., a vertex) of at least a portion of a digitized 3D model. These values ​​(e.g., values ​​between 0 and 255 in the red, green, and blue (RGB) color space) can form a matrix that can be processed by a trained neural network.

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

[0038] At least a portion of the semantic information of a digital 3D model can be tooth identifiers, tooth surface identifiers, and / or jawbone identifiers. Tooth identifiers can provide information about which tooth is included in the input (e.g., incisor, canine, molar, or premolar). Tooth surface identifiers can provide information about which tooth surface is included in the input (e.g., buccal, lingual, mesial, distal, or occlusal surface), while jawbone identifiers can provide information about at least a portion of the associated jawbone (maxilla or mandible) in the digital 3D model.

[0039] By incorporating semantic information into the input of a trained neural network, semantic context can be given to the data present in the input, thus providing semantic background to the trained neural network. The trained neural network can then adapt its behavior based on the provided semantic information. The semantic information can be converted into a numerical format, such as other data formats, before being processed by the trained neural network. By additionally incorporating this semantic information into the input, better performance of the trained neural network can be achieved. For example, premolars are highly prone to plaque buildup. By incorporating this information into the input, the behavior of the trained neural network can be adapted. For instance, the numerical representation of a tooth as input to the trained neural network can be modified to reflect the fact that the tooth is a premolar. The exact manner of this numerical modification can be learned by the trained neural network.

[0040] Advantageously, by inputting natural color and / or geometric information associated with at least one tooth surface of the tooth condition into a trained neural network, various plaque parameters on a digital 3D model can be detected, such as the presence, type, and / or thickness of plaque. Additionally, tartar, stains, or food debris can also be detected due to their unique color.

[0041] Therefore, a method according to one embodiment may include generating input for a trained neural network from a received digitized 3D model. Depending on the type of trained neural network, the input may be in 3D or 2D format. For example, one type of trained neural network may be able to process numerical representations of 2D images generated from the digitized 3D model, while another type of trained neural network may be able to process numerical representations of the digitized 3D model in the form of point clouds or graph representations. Thus, the received digitized 3D model can be adapted for processing by the trained neural network.

[0042] Once generated, it can be fed into a trained neural network to obtain the network's output. All the information forming the input can be concatenated into the final input tensor.

[0043] In addition, the method may include obtaining an output from a trained neural network based on the provided input, wherein the output contains plaque parameters associated with at least one tooth surface.

[0044] The output obtained from the trained neural network based on the provided input can include detecting plaque parameters in the provided input.

[0045] The output can be in the form of a probability value representing the likelihood that at least a portion of the digitized 3D model contains the plaque parameter. This probability value can be associated with at least one vertex, at least one facet, at least one point, at least one node, at least one pixel, or at least one voxel contained in at least a portion of the digitized 3D model, depending on the format of the provided input. Typically, the output can be the probability of the presence of the plaque parameter for each unit element of the input. The unit element can be a vertex, pixel, voxel, node, point, and / or facet.

[0046] By obtaining the output in this way, information about plaque in a given dental condition can be obtained from a single digital 3D model of that condition.

[0047] The output obtained from the trained neural network based on the provided input may include detecting plaque parameters in the input that can be associated with at least one tooth surface of the tooth condition.

[0048] The method according to this disclosure may include assigning plaque parameters to at least a portion of a digital 3D model. In this manner, detected plaque parameters can be associated with corresponding regions in the digital 3D model, and thereby with corresponding regions representing the actual tooth condition as shown by the digital 3D model. The correspondence between at least a portion of the digital 3D model and the detected plaque parameters can be established using identifiers, which may be in numerical form. For example, plaque parameters may be detected for vertices, facets, points, voxels, nodes, pixels, or generally for any unit element in the input. The identifiers can define the exact location of the facet, point, voxel, node, and / or pixel within at least a portion of the digital 3D model. The assignment step allows for the correct labeling and display of plaque parameters on the digital 3D model.

[0049] The trained neural network can perform a classification task for each unit element of the input in the manner described above, and thereby classify the input according to the presence of plaque, for example, dividing the input into tooth surfaces containing plaque and tooth surfaces not containing plaque.

[0050] The output of a trained neural network can be in a fixed-size vector format, which can be derived from the processed representation of the input. The values ​​in the output can contain at least one probability value representing the likelihood that the input contains plaque.

[0051] According to one embodiment, the trained neural network can be a trained convolutional neural network (CNN). This trained CNN may contain multiple convolutional layers that capture low-level features in the input, i.e., obtain convolutional features. Additionally, one or more pooling layers may be present, responsible for reducing the spatial size of the convolutional features. Convolutional neural networks are particularly well-suited for processing matrix information representing 2D images.

[0052] Another example of a trained neural network that is particularly well-suited for performing general classification tasks is the trained residual network (ResNet) or its variants.

[0053] Another example of a trained neural network suitable for performing a classification task (also known as a semantic segmentation task) on each unit element of the input is a trained U-Net or its variants, such as U-Net++.

[0054] In alternative embodiments, the trained neural network can use a self-attention mechanism to process the data, as in transformer networks. Convolution, traditionally derived from convolutional networks, can also be combined with self-attention, traditionally derived from transformer networks.

[0055] When the input is in a three-dimensional (3D) format (such as point clouds), a suitable neural network architecture is the PointNet network or its variants. The PointNet network is capable of handling inputs in 3D format, such as point clouds.

[0056] In one embodiment, obtaining the output of a trained neural network may include obtaining a probability matrix, wherein the probability matrix allows plaque parameters to be assigned to at least a portion of a digitized 3D model. The probability matrix may contain probability values ​​associated with each element of the input. The probability values ​​may specify the probability of a plaque parameter present in the corresponding input element.

[0057] Dental plaque parameters can be plaque presence values, which can simply indicate the presence of plaque on at least one facet of the tooth surface in binary form. Therefore, a trained neural network can classify this at least one facet of the tooth surface as containing plaque or not containing plaque based on previously learned data. In addition to this at least one facet, other unit elements of the input, such as vertices, points, nodes, voxels, and pixels, can be classified according to the input format.

[0058] For example, the label "plaque" can be used to mark the presence of plaque, or the label "plaque-free" can be used to indicate the absence of plaque. Advantageously, users can obtain information about the presence or absence of dental plaque in a single digital 3D dental model of the dental condition.

[0059] Alternatively or additionally, plaque parameters can be plaque type. For example, plaque detected on at least one tooth surface can be classified as pathogenic plaque, non-pathogenic plaque, or tartar. Pathogenic plaque is acidic and, if left untreated, can lead to other oral diseases such as caries or periodontal disease.

[0060] Pathogenic plaques, also known as old or mature plaques, have a high concentration of bacterial metabolites, which makes them acidic.

[0061] Therefore, information about the types of plaque present in a dental condition can be obtained from a single digital 3D model of that condition. Information about high-risk pathogenic plaque accumulation has significant clinical implications in order to develop appropriate treatment plans for patients.

[0062] Alternatively or additionally, plaque parameters can also be plaque thickness values. The detected plaque can be classified into one of predefined plaque thickness categories, such as "no plaque," "thin plaque," "medium plaque," and "thick plaque." The "no plaque" category can cover plaque with a thickness of less than 20 micrometers. "Thin plaque" can refer to plaque with a thickness between 20 and 100 micrometers. "Medium plaque" can refer to plaque with a thickness between 100 and 200 micrometers. "Thick plaque" can refer to plaque with a thickness exceeding 200 micrometers. Information on plaque thickness is clinically important because plaque of different thicknesses requires different treatment approaches. For example, thin plaque layers smaller than 100 micrometers can be removed through consistent brushing. However, thick plaque layers may require dentist intervention. Knowing plaque thickness information allows dentists to develop appropriate treatment plans for patients.

[0063] As part of the method according to this disclosure, treatment recommendations can be automatically provided to the user based on plaque thickness values.

[0064] In one embodiment, dental plaque parameters can be multiple types of parameters, including plaque presence value, plaque thickness value, and / or plaque type.

[0065] In one embodiment, the method according to this disclosure may include determining a plaque thickness value associated with at least one tooth surface by means of the following steps: - At least a portion of the digitized 3D model is fed to a first autoencoder to obtain the output of the first autoencoder, wherein the first autoencoder is trained to detect plaque thickness exceeding a first threshold; - Compare the output of the first autoencoder with at least a portion of the digitized 3D model to obtain a first difference; - At least a portion of the digitized 3D model is provided to a second autoencoder to obtain the output of the second autoencoder, wherein the second autoencoder is trained to detect plaque thickness exceeding a second threshold and reaching a first threshold; - The output of the second autoencoder is compared with at least a portion of the digitized 3D model to obtain a second difference; - At least a portion of the digitized 3D model is fed to a third autoencoder to obtain the output of the third autoencoder, wherein the third autoencoder is trained to detect plaque thickness reaching a second threshold; - The output of the third autoencoder is compared with at least a portion of the digitized 3D model to obtain a third difference; - Identify the minimum value among the first, second, and third differences; - The minimum value identified is used to assign plaque thickness to at least one tooth surface.

[0066] According to the mentioned embodiments, the first autoencoder can be trained solely on data consisting of cases with thick plaques, the second autoencoder can be trained solely on data consisting of cases with medium plaques, and the third autoencoder can be trained solely on data consisting of cases with thin plaques. This, in turn, ensures that each autoencoder creates only the plaque thickness used for its training. Therefore, the autoencoder that produces the output with the least difference from the input can most accurately determine the plaque thickness level present in the input. The first threshold can be set in the range of 150 to 250 micrometers, the second threshold can be set in the range of 50 to 150 micrometers, and the first threshold can be higher than the second threshold.

[0067] Each autoencoder can be trained using a sufficient number of samples with the corresponding plaque thickness level to ensure that each autoencoder can adequately represent the plaque thickness level. In one example, the number of samples can be at least 1,000, preferably at least 10,000.

[0068] Instead of autoencoder networks, variational autoencoders can be used. Additionally, instead of plaque thickness, autoencoders or variational autoencoders can be trained to detect plaque type or plaque severity in the input data structure.

[0069] In another embodiment, determining a plaque thickness value associated with at least one tooth surface may include comparing the profile of the at least one tooth surface with a predicted profile of the at least one tooth surface. For example, the profile of the buccal surface of a tooth containing plaque may be compared with a predicted profile of the buccal surface without plaque. Based on this comparison, the plaque thickness can be determined.

[0070] Typically, predicted surface profiles can be obtained through statistical analysis of multiple teeth. For example, principal component analysis (PCA) or neural networks can be used to obtain a model for predicting the buccal profile of a tooth. This model can predict the buccal profile of a tooth given a local profile. For example, a local profile of a tooth can be provided to the model, and as output, a prediction for the buccal profile can be obtained. The local profile can be the top of the buccal tooth surface, for example, reaching one-third.

[0071] Therefore, the predicted profile of at least one tooth surface can be obtained based on the local profile of that at least one tooth surface.

[0072] The profile of at least one tooth surface can be obtained by converting the tooth containing the at least one tooth surface into a two-dimensional representation; and sampling points in the two-dimensional representation that represent the profile of the at least one tooth surface.

[0073] The method may include, for example, displaying a digital 3D model on a display screen, wherein plaque parameters are assigned to at least a portion of the digital 3D model. The display may include coloring at least one facet of the tooth surface containing plaque. In this manner, areas (such as facets) on the digital 3D model corresponding to areas with plaque in an actual tooth condition are colored to facilitate differentiation from healthy areas in the digital 3D model.

[0074] By displaying the plaque, information about the dental condition can be conveyed to the user. For example, at least a portion of the plaque-containing digital 3D model can be highlighted to visually distinguish it from the rest of the model. Alternatively, at least a portion of the digital 3D model can be colored with a different color, or it can be circled, to visually differentiate the plaque-containing portion from the rest of the model. Alternatively, the line defining the boundary between the plaque-containing portion and the rest of the model can be highlighted.

[0075] In one example, the display may include a table overview showing each tooth surface and corresponding plaque parameters of at least a portion of the digital 3D model. This table overview may be in the form of a dental chart. Alternatively or additionally, plaque parameters may be displayed directly on the digital 3D model, for example, near the corresponding tooth surface.

[0076] One particularly advantageous way to represent plaque on a digital 3D model is to stain at least one facet of the tooth surface containing the plaque with a color corresponding to the color of the developer. This can be advantageous for users such as dentists, as the resulting appearance is similar to that when using a developer.

[0077] The method according to one embodiment may further include determining the plaque density of at least a portion of a digital 3D model. This plaque density can serve as another clinically relevant plaque parameter. Typically, high plaque density reflects a high abundance of plaque-causing bacteria, which can subsequently increase the risk of dental caries and periodontal disease. Plaque density values ​​provide valuable input regarding the density of plaque on the tooth surface.

[0078] In one example, determining plaque density may include, for each plaque-bearing facet, counting the number of adjacent facets containing plaque. The counting may be performed within a sphere of a predetermined radius, such as at least five millimeters. Alternatively, the counting may be performed within a sphere of an adjustable radius, where the radius value depends on the anatomy of the tooth.

[0079] In another example, determining plaque density may include calculating the distance from each facet with plaque to its nearest neighboring facet with plaque.

[0080] The method may also include coloring each facet of at least a portion of the digitized 3D model according to a density-based gradient color scheme based on the determined plaque density. For example, areas with higher plaque density may be colored with darker colors, while areas with lower plaque density may be colored with lighter colors.

[0081] To color appropriate unit elements (e.g., facets) identified as containing plaque, a digitized 3D model can be segmented. In one embodiment, the method may include obtaining a segmented 3D model by segmenting the digitized 3D model into multiple teeth and gingiva. The segmented 3D model may contain at least one facet of at least a portion of the digitized 3D model. This segmentation process of the digitized 3D model should not be confused with general semantic segmentation tasks that can be performed by trained neural networks, which are also referred to herein as classifying each unit element of the input.

[0082] Segmenting a digital 3D model can be performed via a segmentation process that allows the identification of different dental objects (such as individual teeth and / or surrounding gingiva) within the virtual 3D model. Individual teeth can be assigned tooth identifiers, such as those based on the Universal Numbering Notation (UNN), where the numbers 1 through 32 are assigned to human teeth. The segmentation process may include the use of algorithms such as Principal Component Analysis (PCA) or harmonic fields. The segmentation process may alternatively or additionally include the use of machine learning models.

[0083] In one embodiment, the method may further include, within at least a portion of a digital 3D model containing plaque: determining a first surface area of ​​the tooth containing plaque; determining a second surface area of ​​the tooth corresponding to a plaque-free surface; and determining a ratio of the first surface area to the sum of the first and second surface areas. This ratio may be referred to as the planar plaque index (PPI) and provides insight into the percentage of the tooth surface containing plaque. This ratio may also be determined for the entire digital 3D model, thereby providing insight into the percentage of the entire tooth surface that may be covered by plaque.

[0084] According to one embodiment, the method may further include: determining a first number of tooth surfaces containing plaque; determining a second number of tooth surfaces without plaque; and determining a ratio of the first number of tooth surfaces to the sum of the first number of tooth surfaces and the second number of tooth surfaces. This ratio may be referred to as the plaque index and provides insight into the percentage of the entire tooth surface that may be covered by plaque. Determining the plaque index is generally a desirable indicator for dentists. The plaque index determined according to the method of this disclosure is more accurate and reliable than that determined by a dentist manually counting tooth surfaces.

[0085] According to one embodiment, the plaque index can be calculated by assigning a score to each tooth surface and obtaining a total score by summing all scores and dividing by the number of tooth surfaces. The score assigned to each tooth surface can be a value between 0 and 3. Alternatively, the score assigned to each tooth surface can be a value between 0 and 5, where the total score corresponds to a clinical measurement known as the Turesky-modified Quigley Hein Plaque Index (TQHPI). The score assigned to each tooth surface can correspond to the plaque density of each tooth surface. Alternatively, the score assigned to each tooth surface can correspond to the plaque thickness value of each tooth surface.

[0086] Generally, all examples of plaque indices determined according to this disclosure are more accurate and reliable than those determined by dentists through manual counting of tooth surfaces and assignment of relevant scores.

[0087] The trained neural network can be used to additionally process a separate digitized 3D model of the dental condition. This additional digitized 3D model can represent the dental condition captured by an intraoral scanner at a later time point, compared to the previously described digitized 3D model. In this way, the results of the trained neural network processing the digitized 3D model and the additional digitized 3D model can be compared. This comparison can be helpful in tracking the progression of plaque development in the dental condition.

[0088] Therefore, the method according to one embodiment may further include: - The processor receives an additional digital 3D model of the teeth's condition; - At least a portion of the additional digital 3D model is provided as additional input to the trained neural network, wherein at least a portion of the additional digital 3D model contains natural color information and / or geometric information associated with at least one tooth surface of the tooth condition. - Based on the additional input provided, obtain an additional output of the trained neural network, wherein the additional output contains additional plaque parameters associated with at least one tooth surface; - Assign additional plaque parameters to at least a portion of an additional digital 3D model; and - Display an additional digital 3D model, wherein additional plaque parameters are assigned to at least a portion of the additional digital 3D model.

[0089] The additional plaque parameters can be a single type of parameter or multiple types of parameters, such as those described above.

[0090] A method according to one embodiment may include obtaining a further segmented 3D model by segmenting a further digital 3D model into a further plurality of teeth and gums. The further segmented 3D model may contain at least one unit element, such as a facet, of at least a portion of the further digital 3D model. Segmentation may be performed according to the segmentation process described above for the digital 3D model.

[0091] Displaying an additional digital 3D model having additional plaque parameters assigned to at least a portion of an additional digital 3D model may include simultaneously displaying the plaque parameters and the additional plaque parameters. This can be done by displaying an overlaid digital 3D model, wherein the overlaid digital 3D model can be obtained by geometrically aligning the digital 3D model of the tooth condition with the additional digital 3D model.

[0092] A controller can be configured to enable adjustable display of the plaque parameters and / or other plaque parameters on an overlaid digital 3D model. Through user operation, the controller can be positioned between a first position and a second position. The first position of the controller may correspond to displaying essentially only the plaque parameters on the overlaid digital 3D model. The second position of the controller may correspond to displaying essentially only other plaque parameters on the overlaid digital 3D model. In this way, a visual representation of plaque progression in the dental condition can be obtained. The controller may also have other positions between the first and second positions, which may correspond to displaying at least a portion of both the plaque parameters and other plaque parameters. This controller may be referred to as a slider or a 3D slider.

[0093] To develop the ability to accurately identify plaques on a digital 3D model, a neural network can first be trained for this specific task. Training may involve setting appropriate weights within the neural network. Once the neural network meets the training criteria, it can be called a trained neural network.

[0094] In one embodiment of this disclosure, a computer-implemented method for training a neural network to detect dental plaque is disclosed, the method comprising: - Obtain training data for the neural network, which includes: a first training digitized 3D model of the condition of teeth with plaque; and a second training digitized 3D model of the condition of teeth after plaque removal; - Target data is generated by geometrically subtracting a second training digitized 3D model of the dental condition from a first training digitized 3D model of the dental condition to obtain a 3D layer of plaque; - Input the first training digital 3D model of the dental condition into the neural network to obtain the output of the neural network; - Compare the output of the neural network with the generated target data to obtain the loss value; - Adjust the weights of the neural network based on the obtained loss value; - Repeat the input steps until the loss value meets the stopping condition.

[0095] Training data can contain multiple digital 3D models of different dental conditions, also known as training digital 3D models. This data can be multiple intraoral scans obtained by scanning different patients. For each patient's dental condition, at least two digital 3D models can be obtained. For example, the training data can contain a first training digital 3D model of a first dental condition with plaque. The presence of plaque can be determined, for example, by using a common developing agent. Then, a second training digital 3D model of the first dental condition after plaque removal can be obtained. For example, plaque in the dental condition can be removed by simple brushing to obtain the second training digital 3D model.

[0096] Therefore, for multiple different dental conditions, two training digital 3D models can be obtained, one of which includes plaque and the other does not.

[0097] Next, target data can be generated, which can be multiple isolated 3D plaque layers. Generating the target data may involve geometrically subtracting a second trained virtual 3D model of the dental condition from a first trained virtual 3D model of the dental condition to obtain 3D plaque layers. This layer represents a portion of the target data. The same process can be repeated for other trained digitized 3D models.

[0098] For cases with thick plaque, the 3D plaque layer can be up to 2 millimeters thick. Considering that modern intraoral scanners can observe differences in tooth condition as small as 10 micrometers, the training method mentioned above is particularly suitable for detecting the presence of plaque.

[0099] Next, a first training digitized 3D model of the teeth's condition can be input into the neural network to obtain its output. This output can be compared with the generated target data to obtain a loss value. Based on the obtained loss value, the weights of the neural network can be adjusted, and the input steps can be repeated until the loss value meets the stopping condition.

[0100] The particular advantage of the mentioned training method lies in its ability to autonomously generate target data through geometric subtraction of trained 3D models with and without plaques. Specifically, no manual data labeling is required.

[0101] In addition to being used to determine the presence of plaque, or alternatively to determine the presence of plaque, the above training methods can also be used to train neural networks to determine plaque thickness and / or plaque type.

[0102] Another example of a computer-implemented method for training a neural network to detect dental plaque can be described below. The overall goal of the training process for the neural network can be to develop a trained neural network that can classify its inputs according to desired plaque parameters, such as the presence of plaque, plaque type, and / or plaque thickness.

[0103] The training process may include using a manually labeled training dataset. This training dataset may be in 2D format, meaning it may contain multiple 2D training images of teeth or areas of teeth. Additionally, the training dataset may contain associated labels corresponding to dental plaque parameters. Labeling can be performed by, for example, a qualified dentist. The training dataset can be labeled based on indicators of plaque presence, such as using "yes" or "no" labels to indicate the presence or absence of plaque, respectively. Additionally or alternatively, the training dataset may also be labeled with plaque type labels and / or plaque thickness labels.

[0104] In one example, the training dataset can also be in 3D format, with labels directly embedded within it.

[0105] Labels can be viewed as targets or true results for a given training dataset. Additionally or alternatively, true results can be information about plaque presence, plaque thickness, and / or plaque type obtained by applying a developer to multiple different dental conditions. The advantage of such true results lies in their clinical accuracy and acceptance, as the information originates from the use of the developer.

[0106] In another embodiment of this disclosure, a computer-implemented method is disclosed for training a neural network to detect dental plaque on a digital 3D model of a dental condition, the method comprising: - Obtain training data for the neural network, the training data comprising: a first training digital 3D model of the dental condition having plaque colored by a developer; and a second training digital 3D model of the dental condition not having colored plaque. - Target data is generated by identifying plaques that are colored on a first training digitized 3D model and transferring the identified plaques to a second training digitized 3D model; - Input a second training digitized 3D model of the dental condition into the neural network to obtain the output of the neural network; - Compare the output of the neural network with the generated target data to obtain the loss value; - Adjust the weights of the neural network based on the obtained loss value; and - Repeat the input steps until the loss value meets the stopping condition.

[0107] The advantage of this training scheme is that the target data can be generated automatically without the need for manual data labeling.

[0108] Identifying the colored plaques on the first trained digitized 3D model can include: - The color of the facet of the first trained digitized 3D model is compared with a reference color representing the developer to calculate the Euclidean distance between the facet color and the reference color. The reference color can be determined experimentally.

[0109] In this way, the degree of deviation between the facet's color and the reference color can be determined, for example, in the Lab color space. This comparison can be performed on all facets belonging to the first training digitized 3D model, or on all facets corresponding to teeth in the first training digitized 3D model.

[0110] Identifying the colored plaques on the first training digitized 3D model can also include determining the green channel difference between the green channel value of the facet color and the green channel value of the reference color. In this way, different shades of the reference color can be considered and processed. The green channel value can be chosen because the reference color has a pinkish hue, and in the red-green-blue (RGB) color space, different shades of red are controlled by the green channel value.

[0111] Furthermore, facets can be assigned to the developer mask if the Euclidean distance is below the color distance threshold and if the green channel difference is below the green channel threshold. By using these two conditions, the resulting developer mask accurately represents the stained plaque and avoids classifying cavities or other stains as developer.

[0112] The above process can be repeated for all facets belonging to the first training digital 3D model or for all facets corresponding to teeth in the first training digital 3D model. Therefore, the resulting developer mask can contain multiple facets. To exclude individual facets or a small number of connected facets from being classified as developer masks, a minimum number of connected facets can be defined for the developer mask.

[0113] Next, the developer mask can be transferred from the first training digitized 3D model to the corresponding portion of the second training digitized 3D model. This transfer of the developer mask can be performed by first aligning the first and second digitized 3D models, and then identifying the nearest facet of the second training digitized 3D model from the developer mask, after which the developer mask can be transferred. Any of the aforementioned neural network training methods can be used to train the neural network configured to detect dental plaque according to any method of this disclosure.

[0114] A training dataset can be fed into a neural network. The output of the neural network can be a probability value indicating the output category to which the input belongs. A loss function determines the error between the true result and the output. An example of such a loss function is the mean squared error function. To train the neural network, this loss function should be minimized. This can be achieved, for example, by using the stochastic gradient descent algorithm. Through this process of minimizing the loss function, the weights of the neural network can be updated.

[0115] The training process can be an iterative process, in which the same training dataset can be repeatedly fed into the neural network until the stopping criterion is met.

[0116] Regardless of the training process used, the stopping criterion can be a specific threshold for the loss value. Alternatively, the stopping criterion can be the number of training epochs.

[0117] In addition to having an optical system to capture geometric and color information about the condition of teeth, a 3D intraoral scanner can also be equipped with a device for capturing fluorescence emitted from the teeth in the dental condition. This can be achieved by integrating the following components into the intraoral scanner: a light source capable of emitting light to excite fluorescence in a portion of the dental condition; a sensor capable of measuring the emitted fluorescence; and a filter that blocks the light excitation from the sensor while transmitting fluorescence, thereby enabling fluorescence observation. By scanning the dental condition using such a 3D intraoral scanner, fluorescence information can be obtained. This can be significant for plaque detection because the metabolites of bacteria that cause plaque emit red fluorescence when illuminated by a light source with an emission spectrum, such as below 500 nanometers. For each unit element in at least a portion of the 3D digital model, the obtained fluorescence information can be presented as red (R) fluorescence signal values ​​and green (G) fluorescence signal values ​​in the red-green-blue (RGB) color space. Thus, fluorescence information can be stored in each point, facet, voxel, and / or node of at least a portion of the digital 3D model.

[0118] At least a portion of the digital 3D model may additionally include fluorescence information associated with at least one tooth surface of the tooth condition, wherein the fluorescence information includes red (R) fluorescence signal values ​​and green (G) fluorescence signal values.

[0119] Fluorescence information may include red (R) fluorescence signal values ​​and green (G) fluorescence signal values ​​per unit element of at least a portion of the digitized 3D model.

[0120] This fluorescence information can be used as additional input to a trained neural network. Alternatively, it can be used to filter the output of a trained neural network to determine the presence of pathogenic plaque. In another example, the fluorescence information can be used independently to determine the presence of plaque in dental conditions.

[0121] In one embodiment, a computer-implemented method for detecting dental plaque on a digital 3D model of tooth condition is disclosed, the method comprising: - A digital 3D model of the teeth's condition is received by the processor; - At least a portion of a digital 3D model is provided as input to a trained neural network, wherein at least a portion of the digital 3D model contains natural color information and / or geometric information associated with at least one tooth surface of a tooth condition; - Based on the provided input, obtain the output of the trained neural network, wherein the output contains dental plaque parameters associated with the at least one tooth surface, and wherein the dental plaque parameter is a plaque presence value; - Based on a function of red (R) fluorescence signal values ​​and green (G) fluorescence signal values ​​associated with at least a portion of a digital 3D model, the plaque present on at least one tooth surface is determined to be pathogenic plaque.

[0122] The digital 3D model can then be displayed, and the presence of plaques can be visualized, such as by using a trained neural network, and pathogenic plaques can be identified by using fluorescence information.

[0123] In the above method, the trained neural network can be used to detect the presence of plaques, and fluorescence information can also be used to assess whether the identified plaques are pathogenic. Fluorescence information can also be specifically used to establish the presence of pathogenic plaques on at least a portion of the digitized 3D model.

[0124] According to one example, the method may include using fluorescence information to filter the output of a trained neural network to identify the presence of pathogenic plaque on at least one tooth surface. That is, within the facets determined to contain plaque, fluorescence information can then be used to determine which facets contain pathogenic plaque.

[0125] According to another example, the method may include using fluorescence information to filter the output of a trained neural network to identify plaque types associated with at least one tooth surface.

[0126] By filtering the output of a trained neural network using fluorescence information, further information about the state of plaque in dental information can be obtained beyond simply identifying the presence of plaque. Analyzing the fluorescence information of plaque detected by the trained neural network provides insight into the plaque type. By using additional fluorescence filtering, it can be determined whether the detected plaque is, for example, pathogenic (acidic or cariogenic) plaque that may lead to further dental diseases such as tooth decay.

[0127] The function of the red (R) fluorescence signal value to the green (G) fluorescence signal value can be the ratio R / G. Alternatively, the function can be the ratio (RG) / (R+G). In another example, the function can be the difference RG. Different functions of the red (R) fluorescence signal value to the green (G) fluorescence signal value can be used to detect the increase in the red (R) fluorescence signal value as the plaque level increases. The difference (RG) has been observed to be particularly advantageous due to its lowest noise contribution, thus effectively determining the red fluorescence signal.

[0128] The difference between the red (R) fluorescence signal value and the green (G) fluorescence signal value, which can also be called the differential signal, can be a particularly advantageous function because the differential signal has a low noise contribution and can be calibrated simply and directly with the desired red fluorescence signal.

[0129] Detecting the presence of pathogenic plaques can include determining whether the amplitude of the differential signal is greater than the noise component of that differential signal. The noise component of the differential signal can refer to the noise contribution from both the red (R) fluorescence signal value and the green (G) fluorescence signal value.

[0130] In another example, the function of red (R) fluorescence signal value and green (G) fluorescence signal value can be expressed as: RG - (Rh - Gh), where Rh is the health reference value for red fluorescence and Gh is the health reference value for green fluorescence. The health reference value (Rh) for red fluorescence and the health reference value (Gh) for green fluorescence can be calculated for each tooth in the digital 3D model.

[0131] For a digital 3D model of a tooth, a health reference value (Gh) for green fluorescence can be obtained by sampling all vertices belonging to that particular tooth and averaging all values ​​of green (G) fluorescence.

[0132] Alternatively or additionally, for teeth in a digital 3D model, a health reference value (Gh) for green fluorescence can be obtained by averaging all values ​​of green (G) fluorescence belonging to all vertices of the tooth, where the green (G) fluorescence value is higher than a threshold Gt. This additional condition can be advantageous in excluding areas with caries and pathogenic plaque from the calculation of the health reference value. That is, caries can also fluoresce upon irradiation. However, the fluorescence curve of caries is characterized by a loss of green (G) fluorescence value, while pathogenic plaque is characterized by an increase in red (R) fluorescence value. By including only green (G) fluorescence values ​​above the threshold Gt in the calculation of the health reference value (Gh) for green fluorescence, plaque parameters can be detected and false positives in the form of caries presence detection can be excluded. The threshold Gt can also act as a filter to exclude non-dental materials, such as fillings, that do not fluoresce upon excitation.

[0133] The health reference value (Gh) for green fluorescence can be obtained experimentally and set as a predefined value.

[0134] The calculation method for the health reference value (Rh) of red fluorescence is similar to that used for the green fluorescence health reference value (Gh). That is, the health reference value (Rh) of red fluorescence can be calculated for each tooth in the digital 3D model. For a single tooth in the digital 3D model, the health reference value (Rh) of red fluorescence can be obtained by sampling all values ​​of red (R) fluorescence from all vertices belonging to that single tooth and averaging all these values.

[0135] The health reference value (Rh) for red fluorescence can be obtained experimentally and set as a predefined value.

[0136] Using the function RG-(Rh-Gh) is advantageous because it accurately reflects fluorescence behavior. Due to its sensitivity, this function can display plaque parameters over large areas of the tooth.

[0137] In another example, for at least one unit element in at least a portion of the digitized 3D model, the function of the red (R) fluorescence signal value to the green (G) fluorescence signal value can be the difference between the ratio of the red (R) fluorescence signal value to the green (G) fluorescence signal value R / G and the ratio of the healthy reference value Rh of red fluorescence to the healthy reference value Rg of green fluorescence, Rh / Gh. This difference in the two ratios can be expressed as: R / G – Rh / Gh. The characteristic of this difference function is that it can more realistically reflect the formation of pathogenic plaque because sharp edges are visible on the digitized 3D model 101. For dentists, the results obtained and visualized in this way provide a universal and familiar perspective on the formation of pathogenic plaque.

[0138] The examples above using functions based on health reference values ​​(RG-(Rh-Gh) and R / G–Rh / Gh) can be advantageous because they provide insight into the relative increase in red fluorescence compared to green fluorescence, a characteristic of pathogenic plaque fluorescence curves. Due to this property, these functions can be used on multiple different 3D models for various dental conditions, yielding standardized results. That is, these functions are robust to the variable absolute values ​​of green fluorescence in different individuals. Additionally, the fact that different 3D models are affected by different ambient light conditions during scanning is not reflected in the obtained results. Generally, these functions have low sensitivity to factors influencing the variability of red (R) versus green (G) fluorescence signal values ​​across different 3D models. The contribution of including health reference values ​​is reflected in their ability to standardize for different environmental and / or patient characteristics.

[0139] Fluorescence information can be used to filter training data and thereby improve its quality. That is, a dataset intended for training a neural network can contain tooth surfaces with and without fluorescence information. The dataset can then be filtered to exclude tooth surfaces without fluorescence information, thus obtaining training data containing tooth surfaces with fluorescence information.

[0140] After scanning teeth using an intraoral scanner, a digital 3D model can be generated. This digital 3D model may include surfaces artificially generated by a reconstruction algorithm, rather than based on the scan data. This may occur for tooth surfaces that were not accurately captured by the intraoral scanner. The reconstruction algorithm can then create artificial surfaces to compensate for the missing scan data. These artificially generated surfaces can be identified by their lack of associated fluorescence data. That is, the scan data corresponding to the scanned tooth surface may contain fluorescence data, while the artificially generated surface does not. Therefore, filters can be limited based on the presence of fluorescence data.

[0141] In one aspect, the method may include filtering training data to exclude tooth surfaces that lack fluorescence information. Thus, training data containing only tooth surfaces with the corresponding fluorescence information can be obtained. In this way, the training process of the neural network can be improved because the neural network does not consider artificially generated tooth surfaces, thereby reducing the probability of learning incorrect patterns.

[0142] In another aspect, the method may include filtering training data based on fluorescence information to exclude tooth surfaces in the training data that lack fluorescence information.

[0143] A computer program product is also disclosed. According to one embodiment, the computer program product may include instructions that, when executed by a computer, cause the computer to perform the methods of any or more of the embodiments presented above.

[0144] Furthermore, this disclosure includes a non-volatile computer-readable medium. This non-volatile computer-readable medium may contain instructions that, when executed by a computer, cause the computer to perform the methods of any or more of the proposed embodiments. Attached Figure Description

[0145] Various aspects of this disclosure are best understood from the following detailed description taken in conjunction with the accompanying drawings. For clarity, these drawings are schematic and simplified, showing only details to enhance understanding of the claims, while other details are omitted. Each feature of each aspect can be combined with any or all features of the other aspects. These and other aspects, features, and / or technical effects will be apparent from the examples described below and illustrated with reference to these examples in the drawings: Figure 1 The user interface of the displayed digital 3D model with dental features is shown.

[0146] Figure 2 This is a flowchart illustrating a method according to one embodiment.

[0147] Figure 3AThis demonstrates how plaque can be visualized on tooth condition when a developer is used.

[0148] Figure 3B This demonstrates how plaque can be visualized on a digital 3D model of a tooth condition, in a manner similar to that using a developer.

[0149] Figure 4 A user interface is shown on a digital 3D model of plaque used to detect and visualize the condition of teeth.

[0150] Figure 5A A digital 3D model with the displayed plaques and plaque index is shown.

[0151] Figure 5B A digital 3D model with the displayed plaques and plaque index is shown, along with another digital 3D model.

[0152] Figure 6 A trained neural network is shown in a method according to one embodiment.

[0153] Figure 7 This is a flowchart illustrating a method for training a neural network for plaque detection.

[0154] Figure 8 A dental scanning system is shown.

[0155] Figure 9 An exemplary workflow for a patient visiting a dentist is shown. Detailed Implementation

[0156] In the following description, these descriptions are made with reference to the accompanying drawings, which schematically illustrate embodiments of the invention.

[0157] Figure 1 A user interface 100 is shown for a digital 3D model 101 displaying a patient's dental condition. The digital 3D model 101 can be displayed on a screen as a 3D mesh, point cloud, 3D graphics, volumetric representation, or any other suitable 3D representation. The digital 3D model 101 includes a mandible and / or maxilla with multiple teeth. The digital 3D model 101 accurately represents the patient's dental condition and can be constructed based on scan data collected during a scan, such as using an intraoral scanner 825 to scan the patient's teeth and gums. The digital 3D model 101 can be constructed using the captured scan data and presented on a screen. The scan data or the digital 3D model 101 can be stored in and accessed from a storage device of a computer system. The collected scan data may include natural color information, geometric information, infrared information, and / or fluorescence information related to the patient's dental condition.

[0158] User interface 100 may include button 104, which, once triggered by a user, initiates a method for identifying dental plaque on digital 3D model 101. Throughout this disclosure, dental plaque is also referred to as plaque. Identifying plaque may include recognizing the presence of plaque, as well as recognizing any other plaque parameters, such as plaque thickness and / or plaque type. The method may also be initiated automatically without user triggering.

[0159] To identify and separate individual teeth 102 and gingiva 103 within the digital 3D model 101, the digital 3D model 101 can be segmented. This means that facets belonging to each tooth 102 according to a Universal Numbering System / Symbol (UNN) can be determined, such as facets represented by a 3D mesh of the digital 3D model. In this way, individual teeth 102 or local areas of individual teeth 102 can be analyzed to identify plaque. Segmentation of the digital 3D model 101 is an optional step; instead, plaque presence analysis can be performed on the entire digital 3D model 101.

[0160] Figure 2 A flowchart illustrating a method 200 according to one embodiment is shown.

[0161] First, in step 201, the processor may receive the digitized 3D model 101. The digitized 3D model 101 may be stored in the memory of a computer system, for example, in a Standard Trigonometric Language (STL) format. For example, the processor may receive the digitized 3D model 101 when a user triggers button 104 in the user interface 100. The digitized 3D model 101 may typically be displayed on a screen in the form of a 3D mesh, point cloud, 3D graph, volumetric representation, or any other suitable 3D representation.

[0162] Step 202 of method 200 illustrates providing at least a portion of a digital 3D model 101 as input to a trained neural network 600, wherein the at least a portion of the digital 3D model 101 may contain natural color information and / or geometric information associated with at least one tooth surface of a tooth condition. It can be noted that the tooth surface of an actual tooth condition is precisely represented by the tooth surface on the digital 3D model, thereby at least one tooth surface of a tooth condition may have a corresponding tooth surface on the digital 3D dental model.

[0163] At least a portion of the digitized 3D model 101 can be provided as input 601 in any format suitable for processing by the trained neural network 600. For example, the trained neural network 600 may be able to process 2D input formats (such as 2D images) or 3D input formats (such as point clouds or 3D meshes). Providing at least a portion of the digitized 3D model 101 to the trained neural network 600 may include generating numerical representations, such as matrix representations, of 2D and / or 3D input formats.

[0164] Step 203 illustrates obtaining the output 605 of the trained neural network 600 based on the provided input 601, wherein the output 605 may contain plaque parameters associated with at least one tooth surface. These plaque parameters can be understood as single-class or multi-class parameters characterizing dental plaque. For example, plaque parameters may be plaque presence values, plaque thickness values, and / or plaque type.

[0165] If the input is in the format of a 3D mesh, plaque parameters can be obtained for each facet of the input. If the input is in the format of a point cloud, plaque parameters can be obtained for each point of the input; if the input is in the format of a volume representation, plaque parameters can be obtained for each voxel of the input. If the input is in the format of a 2D image, plaque parameters can be obtained for each pixel of the input. Typically, the output 605 of the trained neural network 600 can contain plaque parameters obtained for unit elements of the input 601. Unit elements can be understood as pixels, voxels, faces, nodes, and / or points.

[0166] Step 204 of method 200 illustrates the assignment of plaque parameters to at least a portion of the digitized 3D model 101. For example, the output 605 of a trained neural network 600 may contain plaque parameters, such as a plaque presence value “yes,” associated with a facet of the digitized 3D model 101 as input to 601. This facet may then be labeled as containing plaque, and may have its own identifier defining its location within the digitized 3D model 101. This assignment step establishes an association between the correct region of the digitized 3D model 101 and plaque presence. This principle can also be applied to other types of unit elements in the input, which may be pixels, voxels, points, and / or nodes. All types of unit elements in the input may have their own specified identifiers that associate the unit elements of input 601 with the correct regions on the digitized 3D model 101. For example, one or more points representing a point cloud of the digitized 3D model 101 may have identifiers defining the location of those points within the point cloud.

[0167] Step 205 of method 200 illustrates the process of displaying a digital 3D model 101, wherein dental plaque parameters are assigned to at least one tooth surface. The display may therefore include displaying a digital 3D model, as determined by method 200, having highlighted areas containing plaque. (Refer to Figure 3.) Figure 4 Figure 5 and the display of the digital 3D model 101 can be further understood.

[0168] Figure 3A The image shows the condition of teeth with a developing agent applied to stain the plaque layer, typically red or blue, to make the plaque layer visible to the dentist. This effect can be reproduced once plaque is detected on the digital 3D model 101. Figure 3B As shown.

[0169] Figure 3B A digital 3D model 101 is shown, in which at least a portion of a tooth or the surface of a tooth can be stained to visualize the presence of plaque. The color used to stained the plaque area can be the same as or similar to the color of a developer. In this way, a familiar visualization of plaque on the digital 3D model 101 can be presented to the dentist.

[0170] Figure 4 A user interface 100 is shown for detecting and visualizing plaque on a digital 3D model 101 for detecting and visualizing dental conditions. The digital 3D model 101 can be colored based on plaque parameter values ​​obtained in a method 200 for determining plaque. Coloring can be performed on each unit element of the digital 3D model 101, such as each face, voxel, node, and / or point.

[0171] Dental plaque parameters can include plaque thickness values. Based on the detected plaque thickness values, areas of the digitized 3D model 101 can be colored with different shades of the color used. This visualizes thin, medium, and thick plaque. A gradient bar 401 can be displayed, which can serve as a diagram to explain the thickness of specific points on the plaque. The plaque thickness value can also be displayed when the user hovers a pointer (e.g., a mouse pointer) over a point containing plaque on the digitized 3D model 101.

[0172] In some embodiments of this disclosure, the presence of pathogenic plaque on at least one tooth surface associated with a tooth condition can be determined using fluorescence information. This information can be used as part of the input 601 of a trained neural network 600. Alternatively, the fluorescence information can be used to filter the output 605 of the trained neural network 600 and identify pathogenic plaque. Figure 4In the digital 3D model 101, an indicator 402 can be displayed to show the color of pathogenic plaques as displayed. The color used to visualize pathogenic plaques is typically darker than the color used to visualize non-pathogenic plaques.

[0173] Figure 5A A digital 3D model 101 is shown, in which plaque areas have been colored according to detected plaque thickness values. A gradient bar 401 illustrates the correlation between the plaque thickness values ​​and the colors used to color the areas of the digital 3D model 101. Additionally or alternatively, an indicator 402 may indicate the color used to visualize pathogenic plaques on the digital 3D model 101. Additionally, the value of the plaque index 501 may be displayed. Additionally or alternatively, a plaque presence indicator 502 may be displayed to indicate to the user that plaques have been detected.

[0174] from Figure 5A As can be seen, a single digital 3D model 101 can detect multiple types of dental plaque parameters. This single digital 3D model 101 can also be referred to as a baseline scan.

[0175] To track plaque progression in dental conditions, an additional digitized 3D model 503 may be received. A neural network trained according to this disclosure can be additionally applied to this additional 3D model 503 to detect plaque. The additional digitized 3D model 503 can be obtained from additional scan data, which is obtained by scanning the dental condition at a slightly later time point compared to an acquired baseline. This slightly later time point could be, for example, one year later. The additional digitized 3D model 503 may also be referred to as a follow-up scan.

[0176] The baseline scan 101 and subsequent scan 503 can then be compared to identify changes in plaque parameters. The two 3D models 101 and 503 can be displayed simultaneously, for example, adjacent to each other. Alternatively, the 3D models 101 and 503 can be aligned and overlaid. By aligning the baseline scan 101 and subsequent scan 503, an overlaid 3D model 504 can be obtained. To correctly align the two 3D models 101 and 503, it may be necessary to segment each 3D model and then align corresponding points between the two 3D models 101 and 503.

[0177] The segmentation process can be performed in various ways, such as based on the identification of individual facets or groups of facets belonging to a tooth representation. This allows for the identification of objects such as individual teeth and / or surrounding gingiva in digital 3D models 101, 503 representing a patient's dentition. Therefore, the output of the segmentation process can be an individual tooth representation, typically displayed as a solid tooth object, a tooth mesh, or a tooth point cloud.

[0178] According to one example, segmentation may include using surface curvature to identify the boundaries of the tooth representation. Surface properties can be quantitatively measured using minimum principal curvature and mean principal curvature, and then cutting planes for separating the gingival and tooth portions can be generated based on a principal component analysis (PCA) algorithm. A curvature threshold (e.g., a constant value) can be selected to distinguish the tooth boundary region from the rest of the surface.

[0179] In another example, segmenting digitized 3D models 101 and 503 may include using harmonic fields to identify tooth boundaries. In the digitized 3D models, the harmonic field is a scalar attached to each mesh vertex, satisfying the condition ΔΦ = 0, where Δ is the Laplace operator and is subject to Dirichlet boundary constraints. The above equation can be solved, for example, using the least squares method to compute the harmonic field. The segmented tooth representation can then be extracted by selecting the optimal contour lines connecting data points with the same values ​​as the tooth representation boundaries.

[0180] In yet another example, segmenting the digitized 3D model 101 and / or another digitized 3D model 503 may include using a segmentation machine learning model. Specifically, the digitized 3D model can be converted into a series of 2D digital images taken from different perspectives. The segmentation machine learning model can be applied to this series of 2D digital images. For each 2D digital image, a classification operation can be performed to distinguish different tooth categories from gums. After classifying each 2D image, a back projection onto the digitized 3D model can be performed. This method for segmenting the digitized 3D model can be advantageous because the segmentation machine learning model utilizes the series of 2D images, achieving fast and accurate classification overall. In this example, instead of using 2D digital images, the segmentation machine learning model can use point clouds as the data structure.

[0181] A controller 505 can be configured to allow a user to alternate between displaying a digital 3D model 101 and displaying another digital 3D model 503. The controller 505 allows the user to easily switch between displaying different 3D models. The superimposed digital 3D model 504 may include a boundary, shown by dashed line 506, which can be translated based on user manipulation of the controller 505. The boundary 506 can define which parts of the superimposed digital 3D model 504 can be displayed as digital 3D model 101, and which parts of the superimposed digital 3D model 504 should be displayed as another digital 3D model 503.

[0182] Figure 6A trained neural network 600 used in a method according to one embodiment is shown. This trained neural network 600 can in this case be a convolutional neural network and may contain multiple convolutional layers 602 capable of capturing low-level features of the input, i.e., obtaining convolutional features. Furthermore, one or more pooling layers 603 may be present to reduce the spatial size of the convolutional features. Convolutional neural networks are particularly well-suited for processing matrix information representing 2D images.

[0183] Input 601 can be in 2D or 3D format and can contain at least a portion of a digitized 3D model 101. Figure 6 An input 601 in 2D format is shown. At least a portion of the digitized 3D model 101 may be a 2D image of a tooth 102, the labial surface of which is affected by plaque. Alternatively, at least a portion of the digitized 3D model 101 may include one or more 2D images of the labial surface of the tooth 102, or at least one pixel of a 2D image of the tooth 102. It should be understood that providing input in 2D or 3D format may involve converting these data structures into input tensors that can be processed by a trained neural network.

[0184] Output 605 may contain at least one probability value that represents the likelihood that at least a portion of the digitized 3D model 101 contains dental plaque parameters.

[0185] exist Figure 6 In this case, the plaque parameters are plaque presence values ​​on at least a portion of the digitized 3D model 101. In this case, the trained neural network 600 can perform a classification task on each unit element of the input 601, thereby classifying each unit element of the input 601 into a "yes" or "no" output category. Alternatively or additionally, the plaque parameters may include plaque thickness values ​​and / or the type of plaque present on the tooth 102.

[0186] Figure 7 A flowchart of a computer-implemented method 700 for training a neural network to detect dental plaque is shown.

[0187] Step 701 of method 700 illustrates obtaining training data for a neural network, wherein the training data may include: a first training digitized 3D model of a tooth condition with plaque present; and a second training digitized 3D model of the tooth condition after plaque removal.

[0188] Training data can contain multiple different digital 3D models, referred to as training digital 3D models. This data can be multiple intraoral scans obtained by scanning different patients. For example, the training data could contain a first training digital model of a tooth condition with plaque. The presence of plaque can be determined, for example, by using a common developing agent. Then, a second training digital 3D model of the same tooth condition after plaque removal can be obtained. For example, plaque in the tooth condition can be removed simply by brushing, resulting in the second training digital 3D model.

[0189] Step 702 illustrates generating target data by geometrically subtracting a second training digitized 3D model of the dental condition from a first training digitized 3D model of the dental condition to obtain an isolated 3D plaque layer. This 3D layer can be used as target data to "teach" a neural network to recognize this plaque layer in the input.

[0190] The same process can be repeated for other trained digital 3D models associated with other dental conditions to generate multiple different isolated plaque layers. These isolated plaque layers serve as the "real result" in this case.

[0191] Step 703 of method 700 shows inputting a first training digitized 3D model of the dental condition into a neural network to obtain the output of the neural network.

[0192] Step 704 shows the comparison of the neural network output with the generated target data to obtain a loss value.

[0193] Step 705 illustrates adjusting the weights of the neural network based on a comparison of the output with the target data. For example, by using the stochastic gradient descent (SGD) algorithm, the resulting loss value should be minimized.

[0194] Step 706 illustrates repeating the input steps until the loss value meets the stopping criterion. Therefore, the training process is iterative. The training process can stop when the loss value reaches the stopping criterion. For example, the stopping criterion could be a threshold value for the loss. The stopping criterion could also be multiple iterations (training cycles) after which the performance metric stops increasing.

[0195] The special advantage of the training method mentioned is that it can autonomously generate target data by performing geometric subtraction on training 3D models with and without plaques.

[0196] In addition to determining the presence of plaque, or as a substitute for determining the presence of plaque, the above training method can also be used to train neural networks to determine other types of plaque parameters, such as plaque thickness or plaque type.

[0197] Another example of obtaining true-to-life data could be using 3D scan data with manually annotated labels associated with the plaque, instead of generating a 3D plaque layer through geometric subtraction. Yet another example could be using a chromogenic agent to obtain information about the presence of plaque, which is likely the most accurate from a clinical perspective. Once the plaque has been stained with the chromogenic agent, the corresponding tooth condition can be scanned to obtain a digital representation of the tooth condition. The resulting digital scan data can be segmented, and segments with specific characteristics, such as those with red, green, and blue (RGB) values ​​corresponding to the color of the stained plaque, can be selected as true-to-life data. True-to-life data can refer to the general presence of plaque, or to plaque thickness and / or plaque type, such as pathogenic plaque.

[0198] Real-world results relating to the presence of pathogenic plaque in a patient's dental condition can be obtained in the same manner as described above regarding plaque presence. Digital scan data can be obtained by scanning the patient's teeth with a two- or three-color developer to highlight pathogenic plaque using an intraoral scanner 825. This digital scan data can be segmented to isolate areas corresponding to pathogenic plaque. These isolated areas can then be used as target data to train a neural network to detect pathogenic plaque in the input.

[0199] Figure 8 A dental scanning system 800 is shown, which may include a computer 810 capable of performing any of the methods disclosed herein. The computer may include a wired or wireless interface for connecting to a server 815, a cloud server 820, and an intraoral scanner 825. The intraoral scanner 825 may be capable of recording scan data containing geometric information, natural color information, and / or fluorescence information associated with the patient's dental condition. The intraoral scanner 825 may be equipped with various modules such as fluorescence modules, natural color modules, and / or infrared modules, thereby enabling the capture of information for diagnosing dental diseases such as caries, cracked teeth, gingivitis, gingival recession, and / or plaque.

[0200] The dental scanning system 800 may include a data processing device configured to perform methods according to one or more embodiments of the present disclosure. The data processing device may be part of a computer 810, a server 815, a cloud server 820, or a handheld device not shown in the figures.

[0201] The dental scanning system 800 may include a non-volatile computer-readable storage medium. This non-volatile computer-readable medium may carry instructions that, when executed by a computer, cause the computer to perform methods according to one or more embodiments of this disclosure.

[0202] A computer program product may be embodied in a non-volatile computer-readable storage medium. The computer program product may contain instructions that, when executed by a computer, cause the computer to perform a method according to any of the embodiments presented herein.

[0203] Figure 9 An exemplary workflow 900 of a patient visiting a dentist is shown. In step 1, the patient's oral cavity may be scanned using an intraoral scanner 825. Figure 9 In step 1 of the workflow 900, during the scan, the dentist can view the scan image on the display unit of the computer 810.

[0204] exist Figure 9 In step 2, one or more software applications can be used to further analyze the scan data to detect plaque parameters in the patient's oral cavity. Step 2 of workflow 900 can utilize a software application (i.e., an application module) configured to detect, classify, monitor, predict, prevent, visualize, and / or record plaque that may be present in the patient's dental condition.

[0205] Step 3 of workflow 900 illustrates the use of information obtained in step 2 to populate dental chart 910, such as plaque presence, plaque index, plaque thickness value, and / or plaque type. This information can be recorded separately for each tooth surface. Dental chart 910 can be included as part of a broader patient management system.

[0206] In addition, such as Figure 9 As shown in step 4, at least a portion of the dental scanning system can be connected to the handheld device 920, thereby enabling the transmission of relevant information to the patient. This relevant information may include any plaque parameters associated with the patient's dental condition. In this way, contact with the patient or any other entity using the analyzed scan data can be achieved outside of a dental clinic.

[0207] It should be understood that, without departing from the scope of the present invention, embodiments other than those mentioned above may be adopted, and modifications may be made to the structure and function.

[0208] It should be understood that references to "an embodiment," "an aspect," or "a feature" throughout this specification refer to a specific feature, structure, or characteristic described in connection with an embodiment that is included in at least one embodiment of this disclosure. Furthermore, specific features, structures, or characteristics may be appropriately combined in one or more embodiments of this disclosure. The foregoing description is provided to enable those skilled in the art to practice the various aspects described herein. Various modifications to these aspects will be apparent to those skilled in the art, and the general principles defined herein may also be applied to other aspects.

Claims

1. A computer-implemented method for detecting dental plaque on a digital 3D model (101) of a tooth condition, the method comprising: - The processor receives the digital 3D model of the tooth condition (101). - At least a portion of the digital 3D model (101) is provided as input (601) to a trained neural network (600), wherein the at least a portion of the digital 3D model (101) contains natural color information and / or geometric information associated with at least one tooth surface of the tooth condition, wherein the geometric information includes curvature information, facet normal information, depth information related to the virtual camera position, and / or the angle between the facet normal and the virtual camera direction; - Wherein, the at least portion of the digital 3D model (101) further includes fluorescence information associated with at least one tooth surface of the tooth condition, wherein the fluorescence information includes red (R) fluorescence signal values ​​and green (G) fluorescence signal values; - Based on the provided input (601), the output (605) of the trained neural network (600) is obtained, wherein the output (605) contains a plaque presence value associated with the at least one tooth surface; - The output (605) of the trained neural network (600) is filtered by using the fluorescence information to identify the plaque type associated with the at least one tooth surface; - Assign the plaque presence value and the plaque type to at least a portion of the digital 3D model (101); - Display the digital 3D model (101), wherein the plaque presence value and the plaque type are assigned to at least a portion of the digital 3D model (101).

2. The method according to claim 1, wherein, Filtering the output (605) of the trained neural network (600) includes calculating a function of the red (R) fluorescence signal value and the green (G) fluorescence signal value, wherein the function is the difference between the red (R) fluorescence signal value and the green (G) fluorescence signal value.

3. The method according to any of the preceding claims further includes determining a plaque thickness value associated with the at least one tooth surface.

4. The method according to claim 3, wherein, Determining the plaque thickness value includes: - Provide at least a portion of the digitized 3D model (101) to a first autoencoder to obtain the output of the first autoencoder, wherein the first autoencoder is trained to detect plaque thickness exceeding a first threshold; - The output of the first autoencoder is compared with at least a portion of the digitized 3D model (101) to obtain a first difference; - Provide at least a portion of the digitized 3D model (101) to a second autoencoder to obtain the output of the second autoencoder, wherein the second autoencoder is trained to detect plaque thickness exceeding a second threshold and reaching the first threshold; - The output of the second autoencoder is compared with at least a portion of the digitized 3D model (101) to obtain a second difference; - At least a portion of the digitized 3D model (101) is provided to a third autoencoder to obtain the output of the third autoencoder, wherein the third autoencoder is trained to detect plaque thickness up to the second threshold; - The output of the third autoencoder is compared with at least a portion of the digitized 3D model (101) to obtain a third difference; - Identify the minimum value among the first difference, the second difference, and the third difference; - The minimum value identified is used to assign plaque thickness value to at least one tooth surface.

5. The method according to claim 4, wherein, The first threshold is in the range of 150 to 250 micrometers.

6. The method according to claim 4 or 5, wherein, The second threshold is in the range of 50 to 150 micrometers.

7. The method according to claim 3, wherein, Determining the plaque thickness value associated with the at least one tooth surface includes comparing the profile of the at least one tooth surface with a predicted profile of the at least one tooth surface.

8. The method according to claim 7, wherein, The at least one tooth surface is a buccal surface.

9. The method according to claim 7 or 8, wherein, The predicted profile is obtained based on the local profile of the at least one tooth surface.

10. The method according to any one of claims 7 to 9, wherein, The contour of the at least one tooth surface is obtained by converting a tooth including the at least one tooth surface into a two-dimensional representation and sampling points in the two-dimensional representation that represent the contour of the at least one tooth surface.

11. The method according to any one of claims 3 to 10, further comprising making a treatment recommendation based on the plaque thickness value associated with the at least one tooth surface.

12. The method according to any of the preceding claims further comprises determining a plaque density value for the at least portion of the digital 3D model (101).

13. The method according to claim 12, wherein, Determining the plaque density value involves counting the number of adjacent faces containing the plaque for each facet with plaque.

14. The method according to claim 13, wherein, Counting is performed within a sphere of a predetermined radius.

15. The method according to claim 13 of the preceding paragraph, wherein, Counting is performed within a sphere of adjustable radius, depending on the tooth anatomy.

16. The method according to any of the preceding claims, further comprising: Determine the first number of tooth surfaces containing plaque in at least a portion of the digital 3D model (101); Determine a second number of tooth surfaces in at least a portion of the digital 3D model (101) that do not contain plaque; and determine the ratio of the first number of tooth surfaces to the sum of the first number of tooth surfaces and the second number of tooth surfaces.

17. The method according to any of the preceding claims, wherein, The input (601) of the trained neural network (600) also includes the semantic information of at least a portion of the digitized 3D model (101).

18. The method according to claim 17, wherein, The semantic information includes tooth surface identifiers.

19. The method according to any one of claims 1 to 18, wherein, The trained neural network (600) was trained in the following manner: - Obtain training data for the neural network, the training data comprising: a first training digital 3D model of the condition of teeth with plaque; and a second training digital 3D model of the condition of teeth after plaque removal; - Target data is generated by geometrically subtracting a second training digital 3D model of the dental condition from a first training digital 3D model of the dental condition to obtain a 3D layer of plaque; - Input the first trained digital 3D model of the tooth condition into the neural network to obtain the output of the neural network; - The output of the neural network is compared with the generated target data to obtain a loss value; - Adjust the weights of the neural network based on the obtained loss value; and - Repeat the input steps until the loss value meets the stopping condition.

20. The method according to claim 19 further includes filtering the training data based on the fluorescence information to exclude tooth surfaces that do not possess the fluorescence information from the training data.

21. The method according to any one of claims 1 to 18, wherein, The trained neural network (600) was trained in the following manner: - Obtain training data for the neural network, the training data including: a first training digital 3D model of the tooth condition having plaque colored by a developer, and a second training digital 3D model of the tooth condition without the colored plaque. - Target data is generated by identifying plaques that are colored on the first training digitized 3D model and transferring the identified plaques to the second training digitized 3D model; - Input the second training digital 3D model of the tooth condition into the neural network to obtain the output of the neural network; - The output of the neural network is compared with the generated target data to obtain the loss value; - Adjust the weights of the neural network based on the obtained loss value; and - Repeat the input steps until the loss value meets the stopping condition.

22. The method according to claim 21, wherein, Identifying the bacterial plaques that appear colored on the first trained digitized 3D model includes: - The color of the facet of the first trained digitized 3D model is compared with a reference color representing the colorant to calculate the Euclidean distance between the facet color and the reference color.

23. The method according to claim 22 further comprises: - Determine the green channel difference between the green channel value of the facet color and the green channel value of the reference color.

24. The method according to claim 22 or 23 further comprises: If the Euclidean distance is below the color distance threshold and if the green channel difference is below the green channel threshold, then the facet is assigned to the developer mask.

25. A data processing apparatus comprising means for performing the method of any one of claims 1 to 24.

26. A computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 24.

27. A 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 24.

Images