Method, system and computer readable storage medium for registering intraoral measurements - Patents.com
Deep learning techniques are employed to semantically register intraoral measurements by distinguishing between rigid and non-rigid oral structures, addressing registration errors and enhancing the accuracy of dental scanning.
Patent Information
- Application Number
- JP2022518298
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2019-09-24
- Filing Date
- 2020-09-22
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2040-09-22
AI Technical Summary
Existing dental scanning technologies face challenges in accurately registering intraoral measurements due to soft tissue deformation, which introduces registration errors and distortions, particularly when scanning deformable regions.
Utilizing deep learning methods to semantically register intraoral measurements by identifying and weighting rigid and non-rigid body parts, and employing neural networks to segment and correct registration errors in real-time, thereby forming accurate 3D images.
The method significantly reduces registration errors by accurately distinguishing between rigid and non-rigid oral cavity structures, resulting in precise and reliable 3D models for dental applications.
Smart Images

Figure 0007581337000001 
Figure 0007581337000002 
Figure 0007581337000003
Abstract
Description
[Technical field]
[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This patent application claims the benefit of and priority to U.S. patent application Ser. No. 16 / 580,084, filed Sep. 24, 2019, which is incorporated herein by reference for all purposes.
[0002] The present application relates generally to methods, systems, and computer-readable storage media for registering intraoral measurements, and more particularly to methods, systems, and computer-readable storage media for semantically registering intraoral measurements utilizing deep learning methods. [Background technology]
[0003] Dental practitioners may be trained to produce satisfactory acquisition results during scanning by using appropriate scanning techniques, such as keeping soft tissue out of the dental camera's field of view, which may cause errors and / or disturbances during registration as the soft tissue deforms during scanning causing the same area to take on multiple shapes.
[0004] Feature-based techniques such as Fast Point Feature Histogram (FPFH) can now be used to compute transformations that allow scans / three-dimensional (3D) measurements to be registered without prior knowledge of the relative orientation of the scans, although the success of these techniques may require avoiding scans / 3D measurements of deformable regions.
[0005] US Patent No. 9,456,754 B2 discloses a method for recording multiple three-dimensional images of a dental object, each of which may contain 3D measurement data of the object and color data of the measured surfaces, and combining the individual images into an overall image using a computer-aided recording algorithm, which is incorporated herein by reference for all purposes as if fully set forth herein.
[0006] US Patent No. 7,698,068 B2 discloses a method for providing data useful for a procedure relating to the oral cavity by providing at least one numerical entity representing the three-dimensional surface geometry and color of at least a portion of the intraoral cavity and manipulating the entity to provide desired data therefrom. Typically the numerical entity includes surface geometry and color data associated with said portion of the intraoral cavity, the color data including actual or perceived visual characteristics including hue, saturation, value, translucency and reflectance.
[0007] WO2018219800A1 discloses a method and apparatus for generating and displaying a 3D representation of a portion of an intraoral scene, comprising determining 3D point cloud data representative of the portion of the intraoral scene in a point cloud coordinate space, acquiring a color image of the same portion of the intraoral scene in a camera coordinate space, and labeling color image elements that are within regions of the image that are representative of a surface of the intraoral scene.
[0008] US9436868B2 discloses a method for enabling fast automated object classification of measured three-dimensional (3D) object scenes: the object scene is illuminated with a light pattern and a series of images of the object scene illuminated by the pattern at different spatial phases are acquired.
[0009] US Patent No. 9,788,917 B2 discloses a method for using artificial intelligence in automated orthodontic diagnosis and treatment planning. The method may include providing an intraoral imager configured to be operated by a patient, receiving patient data relating to an orthodontic condition, accessing a database containing or having access to information obtained from the orthodontic treatment, generating an electronic model of the orthodontic condition, and instructing at least one computer program to analyze the patient data and identify at least one diagnosis and treatment plan of the orthodontic condition based on the information obtained from the orthodontic treatment.
[0010] U.S. Patent Application Publication No. 20190026893A1 discloses a method for evaluating the shape of an orthodontic aligner, comprising submitting an analysis image to a deep learning device to determine values of dental attributes for teeth represented on the analysis image and / or values of at least one image attribute for the analysis image.
[0011] PCT application PCT / EP2018 / 055145 discloses a method for constructing restorations, where the condition of the teeth is measured by a dental camera and a three-dimensional (3D) model of the tooth condition is generated. A computer-aided detection algorithm may then be applied to the 3D model of the tooth condition, and the type of restoration, the number of teeth or the position of the restoration is automatically determined.
[0012] US Patent Application Publication No. 20180028294A1 discloses a method for dental CAD automation using deep learning. The method may include receiving patient scan data representing at least a portion of the patient's dentition data set, and identifying one or more dental features in the patient scan using a trained deep neural network. Here, design automation may occur after the complete scan is generated. However, the method does not improve the actual scanning process.
[0013] WO2018158411A1 discloses a method for constructing restorations, in which the condition of the teeth is measured by a dental camera and a 3D model of the tooth condition is generated, where a computer-aided detection algorithm is applied to the 3D model of the tooth condition, in which the type of restoration and / or at least the number of teeth and / or the position of the restoration to be inserted is automatically determined. Summary of the Invention
[0014] The existing limitations associated with the above, as well as other limitations, may be overcome by methods, systems, and computer-readable storage media for semantically registering intra-oral measurements utilizing deep learning methods.
[0015] In one aspect herein, the invention may provide a computer-implemented method for three-dimensional (3D) registration comprising receiving, by one or more computing devices, individual images of a patient's dentition; and automatically identifying sources of registration errors in the individual images using one or more output labels, such as output probability values, of a trained deep neural network, where the output labels / probability values are obtained by segmenting the individual images into regions corresponding to one or more object categories, and the individual images are depth images and / or corresponding color images, and further comprising registering the individual images together based on the one or more output labels, such as the probability values, to form a registered 3D image having no or substantially no registration errors.
[0016] In another aspect herein, the computer-implemented method may further include a combination of one or more of the following: (i) the registration is achieved by generating a point cloud from the depth image by projecting pixels of the depth image into space, assigning a color value and a label / probability value to each point in the point cloud using a corresponding color image and an output label / probability value of the trained deep neural network, respectively, and discarding or partially including points in the point cloud based on the assigned label / probability value using a predefined weight to eliminate or reduce the contribution of the discarded or partially included points to the registration; (ii) the individual images are individual three-dimensional optical images; (iii) the individual images are received as a temporal sequence of images; (iv) the individual images are received as a pair of color and depth images; (v) the one or more object categories include hard gingiva, soft tissue gingiva, teeth and tooth-like objects; (vi) the indication of relevance of the identified causes of registration error is based on the geometry of its surroundings; (vii) the deep neural network is a convolutional neural network ( (vii) using one or more computing devices and a plurality of individual training images, training the deep neural network to map one or more textures in at least a portion of each training image to one or more labels / probability values, the training being performed at a pixel level by classifying the individual training images, pixels of the individual training images, or superpixels of the individual training images into one or more classes corresponding to a type of semantic data and / or a type of error data; (viii) the training images include a 3D mesh and a registered pair of depth and color images; (ix) the 3D mesh is labeled, and the labels are transferred to the registered pair of 3D and color images using a transformation function.
[0017] In yet another aspect of the present invention, there may be provided a non-transitory computer readable storage medium having stored thereon a program that, when executed by a computer system, causes the computer system to perform a procedure comprising: receiving, by one or more computing devices, individual images of a patient's dentition; and automatically identifying sources of registration errors in the individual images using one or more output probability values of a trained deep neural network, the output probability values being obtained by segmenting the individual images into regions corresponding to one or more object categories, wherein the individual images are depth images and / or corresponding color images, and further comprising registering the individual images together based on the one or more output probability values to form a registered 3D image having no or substantially no registration errors.
[0018] Further, a system for three-dimensional (3D) registration can be provided, comprising a processor configured to receive, by one or more computing devices, individual images of a patient's dentition and automatically identify sources of registration errors in the individual images using one or more output probability values of a trained deep neural network, the output probability values being obtained by segmenting the individual images into regions corresponding to one or more object categories, wherein the individual images are depth images and / or corresponding color images, and wherein the processor is configured to register the individual images together based on the one or more output probability values to form a registered 3D image having no or substantially no registration errors.
[0019] In a further aspect of the invention, the system comprises a deep neural network selected from the group consisting of a convolutional neural network (CNN), a fully convolutional neural network (FCN), a recurrent neural network (RNN) and a recurrent convolutional neural network (recurrent CNN).
[0020] The example embodiments will be more fully understood from the detailed description given herein below and the accompanying drawings, in which like elements are represented by like numerals, and which are provided by way of example only and therefore not by way of limitation of the example embodiments herein. [Brief description of the drawings]
[0021] [Figure 1] 1 is a sketch of a cross section of the oral cavity showing different surrounding geometries due to soft tissue deformation. [Figure 2A] FIG. 2 is a sketch of a top view of the oral cavity showing scans / recordings of individual images of the patient's dentition. [Figure 2B] 1 is a sketch illustrating an example registration according to an embodiment of the present invention; [Figure 3A] 1 is a high level block diagram of a system in accordance with one embodiment of the present invention; [Figure 3B] 1 illustrates exemplary training images according to one embodiment of the present invention. [Figure 4] FIG. 13 is a perspective view of a global 3D image of a dentition formed from individual images with soft tissue. [Diagram 5] FIG. 2 is a perspective view of a corrected global 3D image of a dentition formed from individual images with soft tissue contribution removed or weighted less according to an embodiment of the present invention. [Figure 6] FIG. 1 is a high-level block diagram illustrating the structure of a deep neural network according to one embodiment. [Figure 7A] 2 is a flow chart illustrating a method according to one embodiment of the present invention. [Figure 7B] 2 is a flow chart illustrating a method according to one embodiment of the present invention. [Figure 8] FIG. 1 is a block diagram illustrating a training method according to one embodiment of the present invention. [Figure 9] FIG. 1 is a block diagram illustrating a computer system according to an exemplary embodiment of the present invention.
[0022] Different figures within the drawings may have at least some reference numbers that may be the same to identify the same components, and detailed descriptions of each of such components may not be provided below with respect to each figure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0023] In accordance with example aspects described herein, methods, systems, and computer-readable storage media can be provided for utilizing deep learning methods to semantically segment individual intraoral measurements and to register the individual intraoral measurements.
[0024] System for registering intraoral measurements - Patents.com
[0025] Accurate 3D measurements of a patient's oral cavity can be hindered by incorrect registration. Intraoral measurements of the jaws use a camera that produces single scans that capture only a subset of the entire jaw, which can be registered together to form a complete model. The camera may be handheld, and the exact positions where the single scans are obtained are generally not known. Based on information from these single scans (3D data, color data, etc.), a transformation is determined to bring the single scans into a common frame of reference (common 3D coordinate system). However, most registration processes work under the assumption that the bodies are rigid, so parts of the oral cavity that deform / change shape while the camera is frequently acquiring many single scans can distort the registration. Hence, only rigid parts are considered for registration.
[0026] As the scans are taken at different time points, the geometry of certain tissues (especially the soft tissues of the oral cavity) may change between different scans due to soft tissue deformation or the presence of moving foreign bodies, which may hinder registration, which relies on matching 3D data (see, for example, Figure 1, which shows typical errors introduced by techniques based on minimizing the sum-of-squares error).
[0027] Improvements in these techniques may be achieved by considering only the rigid parts for registration and discarding irrelevant (i.e. non-rigid) parts or by weighting their contribution to the registration less, i.e. if the rigid parts / hard tissues 12 such as teeth are considered for registration, the registration is robust (iii) as shown in FIG. 1, and the surrounding geometries 13a, 13b of the rigid parts / hard tissues 12 remain aligned and vice versa (iv). Thus, the term "rigid body" may be used hereafter to describe an anatomical structure or part of an anatomical structure that is unlikely to deform during the scanning procedure. Although the gums near the teeth may deform if a sufficiently high force is applied thereto, this is not usually the case during scanning with an intraoral scanner. It would be a reasonable assumption to consider it as rigid. On the other hand, the inside of the cheek may deform between scans and therefore may be considered as a soft tissue / soft part 15 (FIG. 2A).
[0028] The system described herein is preferably capable of obtaining images such as individual three-dimensional optical images 2 (FIG. 2A), where each three-dimensional optical image 2 preferably contains 3D measurement data and color data of a measured surface of a tooth and is preferably recorded consecutively in the oral cavity by direct intraoral scanning, which may for example be performed in a dental office or clinic, or by a dentist or dental technician, or these images may be obtained indirectly from a series of stored images.
[0029] Using these images, preferably acquired as a temporal sequence of images, a computer-implemented system may automatically identify regions in the images that can be considered suitable for registration. This may be done in real time. It is also understood that the images may be individual two-dimensional (2D) images, RGB images, distance images (2.5 dimensions, 2.5D), four-channel images (RGB-D), where the depth and color may not be perfectly aligned, i.e. the depth and color images may be acquired at different time periods.
[0030] The scanning process may generate multiple individual images, which may then be combined to form an overall / global 3D image 10 (FIG. 4). More specifically, the individual three-dimensional optical images 2, shown in the form of a rectangle as shown in FIG. 2A, may be obtained by a scanner / dental camera 3 that may be moved relative to the object 1 along a measurement path 4 during measurement. In some embodiments, the measurement path may be any arbitrary path, i.e. measurements may be taken from different directions. The dental camera 3 may be a handheld camera that measures the object 1, for example, using a fringe projection method. Other methods of 3D measurement may be understood by those skilled in the art. The first overlapping area 5 between the first image 6 and the second image 7, shown in dashed lines, may be checked using a computer to determine whether the recording condition is met, and if so, the three-dimensional optical images 2 may be combined / registered together to form a global 3D image 10.
[0031] The recording conditions include the appropriate size, appropriate waviness, appropriate roughness and / or the appropriate number and arrangement of characteristic geometries. However, it may be difficult to program a conventional computer to determine the causes of registration errors and how to prevent them. Manual programming features used for registration or segmentation methods that cover all possible scenarios may be tedious to do, especially considering the high frequency of measurements. This is especially true when the context of the entire image must be considered. Machine learning methods, especially neural networks and proper training data, can be used to more effectively solve this problem. On the other hand, neural networks may be trained to recognize the causes of registration errors and to semantically segment data from a single scan / single 3D measurement and determine whether these oral cavity regions can be considered for registration. For this purpose, labels for different objects / object categories for segmentation may be defined to include, but are not limited to, (i) hard tissues (such as teeth, crowns, bridges, hard gums near teeth and other tooth-like objects), (ii) soft tissues (such as tongue, cheeks, soft gums) and (iii) instruments / disposable instruments applied in the oral cavity (such as mirrors, scanbodies, cotton rolls, brackets). Of course, other definitions such as glare areas 21 (FIG. 2B, caused by bright light) may be added accordingly. The segmentation may be based on color, which may be interpreted in a context-aware manner, i.e. an indication of the relevance of a potential source of registration error may be based on its surrounding geometry 13a, 13b. Furthermore, the segmentation may be based on single pixels of color data or on larger regions of the individual three-dimensional optical images 2.
[0032] Since the crown, tooth or hard gums near the tooth are rigid bodies, registration errors may be eliminated or substantially reduced by a registration algorithm that takes into account correct segmentation. Additionally, clutter introduced by accessories such as cotton rolls may be removed to generate a cleaned 3D model, which includes only data relevant for dental treatment.
[0033] The system may therefore train a neural network, such as a deep neural network, with multiple training data sets to automatically recognise sources of registration errors in the 3D optical image 2 and prevent them from contributing to the registration, preferably in real time. Thus, the misregistrations (Fig. 1, iv) propagated to the global 3D image 10 (Fig. 4) may be reduced or eliminated as shown in the corrected global 3D image 9 of Fig. 5, and / or the scan flow may be improved by reducing or eliminating the disturbances caused by the misregistrations.
[0034] The system may also identify and label the data semantically (context-aware, i.e., context may be important for selecting the appropriate correction method; for example, gums close to the teeth may be considered as hard tissue 12, while gums farther from the teeth may be considered as soft tissue 15).
[0035] Furthermore, the system may determine corrective measures and / or apply said corrective measures upon detecting a cause of registration error. For example, if there is a high proportion of hard tissue 12 relative to soft tissue 15 in the individual three-dimensional optical images 2, it may be advantageous to weight the hard tissue 12 more than the soft tissue 15, since deformations or movements of the patient's cheeks or lips may cause deformations of the soft tissues and thus a defective registration as shown in FIG. 1(iv). However, if the proportion of hard tissue 15 in the image is low, the soft tissue 15 may be weighted more heavily to improve the quality of the registration. As an example, FIG. 2B shows an image in which the first region with hard tissue 12 is about 10% of the total area of the individual three-dimensional optical images 2 and the second region with soft tissue 15 is about 50% of the total area of the image, due to the missing two middle teeth. The remaining third region, which cannot be assigned to any of the tissues, is about 40%, and the proportion of the first region with hard tissue 12 falls below the predefined threshold of 30%. Thus, the first region having hard tissue 12 is weighted by a first weighting factor and the second region having soft tissue is weighted by a second weighting factor such that the second weighting factor increases with the decrease of the first region. For example, the first weighting factor for the first region having hard tissue 12 may be 1.0 and the second variable weighting factor for the second region having soft tissue 15 may be 0.5 above a predefined threshold and may increase to 1.0 with the decrease of the first region. The dependence of the second weighting factor on the first region may be determined according to any function, such as an exponential or linear function, and the hard and soft tissues may be segmented using a neural network as described herein.
[0036] FIG. 3A shows a block diagram of a system 200 for recognizing dental information from an individual three-dimensional optical image 2 of a patient's dentition according to an embodiment. The system 200 may include a dental camera 3, a training module 204, an image registration module 206, a computer system 100, and a database 202. In another embodiment, the database 202, the image registration module 206, and / or the training module 204 may be part of the computer system 100 and / or may directly and / or indirectly adjust the parameters of the dental camera 3 based on the corrective measures. The computer system 100 may also include at least one computer processor 122, a user interface 126, and an input device 130. The computer processor may receive various requests and may load appropriate instructions stored in the storage device into the memory and then execute the loaded instructions. The computer system 100 may also include a communication interface 146 that allows software and data to be transferred between the computer system 100 and external devices.
[0037] The computer system 100 may receive a registration request from an external device, such as a dental camera 3, or a user (not shown), and may be loaded with appropriate instructions for semantic registration. Preferably, the computer system may independently register individual three-dimensional optical images 2 as they are received without waiting for a request.
[0038] In one embodiment, the computer system 100 may use multiple training data sets from the database 202 (which may include, for example, multiple individual 3D optical images 2) to train one or more deep neural networks, which may be part of the training module 204. FIG. 3B(i-iv) shows example images used for training, including a color image, a depth image, a color image mapped to a depth image, and a depth image mapped to a color image, respectively. Since convolutional neural networks (CNNs) operate in local neighborhoods, it may be good for the CNN to map these images. Thus, regions of the RGB image and the depth image may be represented by the same 2D pixel coordinates. For example, mapping a depth image to an RGB image or vice versa means generating an image such that pixels with the same 2D pixel coordinates in the RGB image and the generated image correspond to the same point in space. Typically this involves the application of a pinhole camera model and a transformation to accommodate the movement (determined by the registration algorithm). Steps to compensate for the movement may be omitted, and the net can be expected to be able to handle the resulting offsets, which are expected to be small. In some embodiments, the system 200 may include a neural network module (not shown) that includes various deep learning neural networks, such as a convolutional neural network (CNN), a fully convolutional neural network (FCN), a recurrent neural network (RNN), and a recurrent convolutional neural network (recurrent CNN). An exemplary fully convolutional neural network is described in a publication entitled "Fully Convolutional Networks for Semantic Segmentation" by Jonathan Long et al., published March 8, 2015, which is incorporated by reference in its entirety as if fully set forth herein.Thus, a fully convolutional neural network (an efficient convolutional network architecture used for pixel-by-pixel segmentation) may be trained to segment an RGB(D) image or a sequence of RGB(D) images by using a recurrent model. A recurrent model may be used in contrast to a simple forward propagation network. The network may thus receive the output of a layer as input for the next forward calculation, such that its current activation can be considered as a state depending on all previous inputs, thus allowing the processing of sequences. Further exemplary recurrent CNN models are described in a publication entitled "Recurrent Convolutional Neural Networks: A Better Model of Biological Object Recognition" by Courtney J. Spoerer et al., Front. Psychol, September 12, 2017, which is incorporated by reference in its entirety as if fully set forth herein.
[0039] The training data set and / or input to the neural network may be pre-processed. For example, calibration (such as determining parameters of a camera model) may be applied to align color images with 3D surfaces in order to process color data together with the 3D measurements. Furthermore, standard data augmentation procedures such as compound rotation, scaling, etc. may be applied to the training data set and / or input.
[0040] The training module 204 may supervise the learning process of the deep neural network using a training data set with labels. The labels may be used to weight the data points. Conversely, the training module 204 may train the generative deep neural network using an unlabeled training data set.
[0041] In an exemplary embodiment, a plurality of real individual three-dimensional optical image data sets having the tissue types and object categories described above may be used to train the deep neural network to detect sources of registration errors. In another example, a plurality of other training data sets from real dental patients having one or more hard gingival regions near one or more teeth and one or more soft gingival regions away from one or more teeth may be selected to form a training data set group to train the deep neural network to recognize semantic data (e.g., hard gingiva near teeth). Thus, the database 202 may include different training data sets, e.g., one group for each object category and / or one group for each semantic data type.
[0042] In some embodiments, the training module 204 may pre-train one or more deep neural networks with a training data set from the database 204, so that the computer system 100 can easily use one or more pre-trained deep neural networks to detect sources of registration errors. It may then transmit information about the detected sources and / or the individual three-dimensional optical images 2, preferably automatically and in real time, to the image registration module 206, where the sources of registration errors are taken into account prior to registration.
[0043] The database 204 may also store data related to the deep neural networks and the identified causes along with the corresponding individual three-dimensional optical images 2. Additionally, the computer system 100 may have a display device 126 and an input device 130, which may be used by a user to perform functions such as submitting requests and receiving and reviewing causes of registration errors identified during training.
[0044] In an exemplary embodiment of the training process (S600) as shown in FIG. 8, labels may be generated by collecting images (step S602) that represent real cases in the field. These cases may include meshes (such as 3D triangular meshes) and single images (registered pairs of depth and color images). These meshes may be segmented by an expert who can crop teeth in them. The cropped meshes may then be labeled (step S604). They may be labeled as teeth, crowns, implants or other tooth-like objects (step S606) or hard and soft gums (step S608). Additionally, outliers such as 3D points in the optical path of the camera 3 that are not part of the oral cavity may be labeled (step S610). The labeling of the meshes is transferred to the pixels of the single image, thus reducing the amount of work done in the training process. In step S612, all final labels may be determined by combining the information from steps S606, S608 and S610 (step S612). Furthermore, since we know the transformation that aligns the single images together (because they are registered), we can transfer these final labels from the cropped mesh to the single image. In this way, we can quickly label many images by cutting / slicing the mesh.
[0045] Other embodiments of system 200 may include different and / or additional components, and their functionality may be distributed among the components differently than described herein.
[0046] FIG. 6 shows a block diagram illustrating the structure of a deep neural network 300 according to an embodiment of the present invention. It may have several layers, including an input layer 302, one or more hidden layers 304, and an output layer 306. Each layer may consist of one or more nodes 308, represented by small circles. Information may flow from the input layer 302 to the output layer 306, i.e. from left to right, but in other embodiments it may flow from right to left or both. For example, a recurrent network may take into account previously observed data in a sequence of images 8 when processing new data (e.g., segmenting a current image taking into account previous images), while a non-recurrent network may process new data individually.
[0047] Nodes 308 may have inputs and outputs, and nodes in the input layer 308 may be passive, meaning that they may not modify the data. For example, nodes 308 in the input layer 302 may each receive a single value (e.g., a pixel value) for their input and replicate the value to their multiple outputs. Conversely, nodes in the hidden layer 304 and output layer 306 may be active and therefore may modify the data. In an exemplary structure, each value from the input layer 302 may be replicated and sent to all of the hidden nodes. Values arriving at the hidden nodes may be multiplied by a weight, which may be a set of predefined numbers associated with each of the hidden nodes. The weighted inputs may then be summed to generate a single number.
[0048] In one embodiment according to the present invention, the deep neural network 300 may use pixels of the individual 3D optical images 2 as inputs in detecting object categories. The individual 3D optical images 2 may be color and / or depth images. Herein, the number of nodes in the input layer 302 may be equal to the number of pixels in the individual 3D optical images 2.
[0049] In an exemplary embodiment, one neural network may be used for all object categories, while in another embodiment, different networks may be used for different object categories. In another example, the deep neural network 300 may classify / label individual 3D optical images 2 instead of individual pixels upon detecting an object category, such as one caused by ambient light. In a further embodiment, the images may be subsampled inputs, such as every fourth pixel.
[0050] In yet another embodiment, the deep neural network 300 may have as input a number of data acquired by the dental camera 3, such as color images, depth measurements, accelerations, as well as device parameters, such as exposure time, aperture, etc. The deep neural network may output a label, which may for example be a probability vector including one or more probability values for each pixel input belonging to a particular object category. For example, the output may include a probability vector including probability values, where the highest probability value may define the location of the hard tissue 12. The deep neural network may also output a map of label values that does not include any probabilities. A deep neural network can be created for each classification, although this may not be necessary.
[0051] Method for registration in intraoral measurements - Patents.com
[0052] Having now described the system 200 of FIG. 3A, reference is now made to FIG. 7A, which illustrates a process S400 in accordance with at least some of the example embodiments herein.
[0053] Process S400 may begin by obtaining and labeling regions of interest in a training dataset with predefined labels (step S402). For example, sample soft tissue 415 on sample image 413 shown in FIG. 3B(i) may be labeled as soft tissue. Sample hard tissue 412 on sample image 413 shown in FIG. 3B(i) may be labeled as hard tissue. Labeling of training images may be done digitally, for example, by placing dots on the image that correspond to points of interest.
[0054] The training data may be labeled in order to assign semantics to the individual 3D optical images 2. This may be done at a pixel-by-pixel level for color or depth information. Alternatively, for a single image, a mesh of the complete 3D model may be cut and the corresponding pixel-by-pixel labels calculated. Said mesh may be segmented so that the labeling process can be further automated. These labels may distinguish between teeth, cheeks, lips, tongue, gums, fillings and ceramics, while nothing else is assigned a label. Irrelevant for the registration may be cheeks, lips, tongue, glare and unlabeled data.
[0055] The training data may also be labeled in order to assign registration error cause labels to the individual 3D optical images 2. Again, this may be done at pixel-by-pixel level, e.g. for image or depth information. The training data may also be labeled at pixel-by-pixel level, e.g. for hard tissue 12 and soft tissue 15 and / or instruments / disposable instruments applied in the oral cavity, etc.
[0056] Semantic labels may overlap with markers for sources of registration error, e.g., labels such as "hard tissue + glare", "soft tissue close to hard tissue", and "tongue + hard tissue", and these labels may be distinguished from other labels such as "cheek + glare".
[0057] With this set of labeled or classified images, a deep neural network 300 may be constructed and fed labeled images, thereby allowing the network to “learn” from them, so that the network can wire up to segment new images on its own.
[0058] As an alternative option to segmentation involving classification on a pixel-by-pixel basis, segmentation may involve classifying and training at a level slightly higher than the pixel-by-pixel level (i.e., at the "superpixel-by-superpixel" level, where a "superpixel" is a portion of an image that is larger than a regular pixel in the image).
[0059] The instructions and algorithms of process S400 may be stored in a memory of computer system 100 and may be loaded and executed by processor 122 to train one or more deep neural networks with a training data set to detect one or more defects 15 based on one or more output labels / probability values (step S404). For example, if one of the probability values of the probability vector corresponding to glare is 90%, then the neural network may detect glare portion 21 as one of the causes of registration error in each three-dimensional optical image 2.
[0060] The training step may be performed once, multiple times or intermittently. The training step may also be "semi-supervised" or "self-supervised". For example, after a first training step, the deep neural network may receive or obtain images and outputs that it has not seen before, and may provide corresponding feedback so that the network can preferably operate on its own and ultimately classify the images without human help. Thus, the deep neural network 300 may be trained such that when a sequence of images 8 of the individual three-dimensional optical images 2 are input to the deep neural network 300, the deep neural network can return a resulting label / probability vector for each image indicating the category to which parts of those images belong.
[0061] After the training step, the deep neural network may acquire or receive a sequence of images 8 of the individual 3D optical images from the dental camera 3 for real-time segmentation (step S406) and detect sources of registration error in those images (step 408). Upon detection of said sources, the image registration module 206 may register the images together based on predefined weights for the segments by ensuring that the detected sources of registration error do not contribute to the registration process (step S410). Steps S406-S410 of FIG. 7A are also included in the flow chart of FIG. 7B, discussed later in this specification.
[0062] FIG. 7B shows process S500, which may be a subset of process S400. Process S500 may begin with step S502, where a color image and / or depth image is obtained from dental camera 3. In step S504, the color and depth images are used as inputs to a trained deep neural network 300 to obtain corresponding outputs of labeled images indicating the probability of the segments belonging to the object category. Using both images may make it easier to discriminate between different labels. One would expect that it would be harder to train a net to distinguish between soft and hard gums based on color than with depth and color, for example. These images may be "matched" and labeled, so it may not make much difference. In one embodiment, one may be labeled / segmented and the other may simply provide additional features to determine segmentation. Depending on the embodiment, there may be a one-to-one correspondence either between the resulting labeled images and depth images or between the resulting labeled images and color images. The labeled image may have the same lateral resolution as the depth / color image and the channel for that label. In step S506, a point cloud may be generated from the depth image by projecting each pixel of the depth image into space. Each point in these point clouds may be assigned a color value from the color image and a probability vector from the labeled image. In one embodiment, the labeled image, point cloud and resulting mesh may all have a label, a probability of the label or a probability vector assigned to them. In step S508, more images and corresponding output labels are obtained so that each input point cloud is registered to the already registered point cloud. Herein, points with a high probability of being soft tissue (e.g., above a predefined threshold or weighted as a function of probability in a predefined manner) are discarded or weighted less than other points with a high probability of being hard tissue. In step S510, each point in the input point cloud is added to the corresponding grid cell until it has an average position, color and probability.A transformation may then be optimized to align the single images with each other by using predefined weights for soft tissue 15, hard tissue 12 and / or any other object category. If the transformation changes, the input in the grid sampler may be updated accordingly. Of course, other embodiments different from FIG. 7B may be achieved taking into account this specification.
[0063] Computer system for registering intraoral measurements - Patents.com
[0064] Having now described the processes of Figures 7A and 7B, reference is now made to Figure 9, which illustrates a block diagram of a computer system 100 that may be used in accordance with at least some of the exemplary embodiments herein. Although various embodiments may be described herein with respect to this exemplary computer system 100, it will become apparent to one of ordinary skill in the art after reading this specification how to implement the invention using other computer systems and / or architectures.
[0065] The computer system 100 may include or be separate from the training module 204, the database 202, and / or the image registration module 206. These modules may be implemented in hardware, firmware, and / or software. The computer system may also include at least one computer processor 122, a user interface 126, and an input device 130. The input device 130 may be used by a dental practitioner in an exemplary embodiment in conjunction with a display device 128, such as a monitor, to send commands or requests during the training process. In another exemplary embodiment herein, the input device 130 is a finger or a stylus used on a touch screen interface (not shown). The input device 130 may alternatively be a gesture / voice recognition device, a trackball, a mouse, or other input device such as a keyboard or a stylus. In one example, the display device 128, the input device 130, and the computer processor 122 may collectively form a user interface 126.
[0066] The computer processor 122 may include, for example, a central processing unit, a multi-processing unit, an application specific integrated circuit ("ASIC"), or a field programmable gate array ("FPGA"). The processor 122 may be connected to a communication infrastructure 124 (e.g., a communication bus or network). In one embodiment herein, the processor 122 may receive a request for 3D measurements, automatically detect sources of registration errors in the images, and automatically register the images based on the detected sources of registration errors using these image registration modules 206. The processor 122 may accomplish this by loading corresponding instructions stored in a non-transitory storage device in the form of computer readable program instructions, and executing the loaded instructions.
[0067] Computer system 100 may further include a primary memory 132, which may be random access memory ("RAM"), and may also include a secondary memory 134. The secondary memory 134 may include, for example, a hard disk drive 136 and / or a removable storage drive 138. The removable storage drive 138 may read from and / or write to the removable storage device 140 in a well-known manner. The removable storage device 140 may be, for example, a floppy disk, magnetic tape, optical disk, flash memory device, etc. that can be written to and read by the removable storage drive 138. The removable storage device 140 may include a non-transitory computer-readable storage medium having computer-executable software instructions and / or data stored thereon.
[0068] In still other embodiments, the secondary storage device 134 may include other computer-readable media that store computer-executable programs or other instructions that are loaded into the computer system 100. Such devices may include removable storage devices 144 and interfaces 142 (e.g., program cartridges and cartridge interfaces), removable memory chips (e.g., erasable programmable read-only memories ("EPROMs") or programmable read-only memories ("PROMs")) and associated memory sockets, other removable storage devices 144 and interfaces 142 that allow software and data to be transferred from the removable storage devices 144 to other portions of the computer system 100.
[0069] Computer system 100 may also include a communications interface 146 that allows software and data to be transferred between computer system 100 and external devices. Such interfaces include a modem, a network interface (e.g., an Ethernet card, a wireless interface, a cloud hosting service over the Internet, etc.), a communications port (e.g., a Universal Serial Bus ("USB") port or a Firewire port), a Personal Computer Memory Card International Association ("PCMCIA") interface, Bluetooth, etc. The software and data transferred through communications interface 146 may be in the form of a signal, which may be an electronic, electromagnetic, optical, or another type of signal that may be capable of being transmitted and / or received by communications interface 146. The signal may be provided to communications interface 146 via a communications path 148 (e.g., a channel). Communications path 148 may carry signals and may be implemented using wire or cable, fiber optics, telephone lines, cellular links, or radio frequency ("RF") links, etc. Communications interface 146 may be used to transfer software, data or other information between computer system 100 and a remote server or cloud-based storage.
[0070] One or more computer programs or computer control logic may be stored in primary memory 132 and / or secondary memory 134. Computer programs may also be received via communications interface 146. The computer programs may include computer-executable instructions that, when executed by computer processor 122, cause computer system 100 to perform the methods described herein.
[0071] In another embodiment, the software may be stored on a non-transitory computer readable storage medium and loaded into the primary memory 132 and / or secondary memory 134 of the computer system 100 using the removable storage drive 138, the hard disk drive 136 and / or the communication interface 146. The control logic (software), when executed by the processor 122, causes the computer system 100, and more generally the system for detecting scanning interference, to perform all or some of the methods described herein.
[0072] Other hardware and software configurations for carrying out the functions described herein will be apparent to those of ordinary skill in the art in view of this specification.
Claims
1. receiving, by one or more computing devices, respective images configured as a depth image and a corresponding color image of the patient's dentition, wherein the depth image and the corresponding color image are associated together; providing pixels of said individual images as input to a trained deep neural network; automatically identifying sources of registration error in the individual images prior to registration using one or more output label values of the trained deep neural network, the output label values being obtained by segmenting the individual images into regions corresponding to one or more object categories; registering the individual images together based on the one or more output label values to form a registered 3D image having no or substantially no registration error; Including, the individual images are depth images and corresponding color images that are associated together by associating the depth image with the corresponding color image or by associating the corresponding color image with the depth image; generating a point cloud from the depth image by projecting pixels of the depth image into space; assigning a color value and a label value to each point in the point cloud using the corresponding color image and the output label values of the trained deep neural network, respectively; discarding or partially including one or more points in the point cloud based on the assigned label values using a predetermined weight to eliminate or reduce the contribution of the discarded or partially included one or more points to registration; further comprising: A computer-implemented method for three-dimensional (3D) registration.
2. The method of claim 1 , wherein the individual images are individual three-dimensional optical images.
3. The method of claim 1 , wherein the individual images are received as a temporal sequence of images.
4. The method of claim 1 , wherein the one or more object categories include hard gums, soft tissue gums, tongue, cheek, teeth, and tooth-like objects.
5. 2. The method of claim 1, wherein the deep neural network is a network selected from the group consisting of a convolutional neural network (CNN), a fully convolutional neural network (FCN), a recurrent neural network (RNN), and a recurrent convolutional neural network (recurrent CNN).
6. training the deep neural network using the one or more computing devices and a plurality of individual training images to map one or more textures in at least one portion of each training image to one or more label values; further comprising the training step is performed at the pixel level by classifying the individual training images, pixels of the individual training images, or superpixels of the individual training images into one or more classes corresponding to types of semantic data and / or types of registration error; The method of claim 1.
7. The method of claim 6 , wherein the training images include a 3D mesh and a registered pair of depth and color images.
8. The method of claim 7 , wherein the 3D mesh is labeled and the labels are transferred to the registered pair of 3D and color images using a transformation function.
9. When executed by a computer system, receiving, by one or more computing devices, respective images configured as a depth image and a corresponding color image of the patient's dentition, wherein the depth image and the corresponding color image are associated together; providing pixels of said individual images as input to a trained deep neural network; automatically identifying sources of registration error in the individual images prior to registration using one or more output label values of the trained deep neural network, the output label values being obtained by segmenting the individual images into regions corresponding to one or more object categories; Execute a procedure that includes the individual images are depth images and corresponding color images that are associated together by associating the depth image with the corresponding color image or by associating the corresponding color image with the depth image; the procedure includes registering the individual images together based on the one or more output label values to form a registered 3D image having no or substantially no registration error; The method further comprises: generating a point cloud from the depth image by projecting pixels of the depth image into space; assigning a color value and a label value to each point in the point cloud using the corresponding color image and the output label values of the trained deep neural network, respectively; discarding or partially including one or more points in the point cloud based on the assigned label values using a predetermined weight to eliminate or reduce the contribution of the discarded or partially included one or more points to registration; Including, A non-transitory computer-readable storage medium storing a program.
10. receiving, by one or more computing devices, respective images configured as a depth image and a corresponding color image of the patient's dentition, and correlating the depth image and the corresponding color image together; providing pixels of said individual images as input to a trained deep neural network; and automatically identifying sources of registration error in the individual images prior to registration using one or more output label values of the trained deep neural network, the output label values being obtained by segmenting the individual images into regions corresponding to one or more object categories. a processor configured to: the individual images are depth images and corresponding color images that are associated together by associating the depth image with the corresponding color image or by associating the corresponding color image with the depth image; the processor is configured to register the individual images together based on the one or more output label values to form a registered 3D image having no or substantially no registration error; The processor further generating a point cloud from the depth image by projecting pixels of the depth image into space; assigning a color value and a label value to each point in the point cloud using the corresponding color image and the output label values of the trained deep neural network, respectively; configured to discard or partially include one or more points in the point cloud based on the assigned label values and using predetermined weights to eliminate or reduce the contribution of the discarded or partially included one or more points to registration. A system for three-dimensional (3D) registration.
11. 11. The system of claim 10, wherein the deep neural network is a network selected from the group consisting of a convolutional neural network (CNN), a fully convolutional neural network (FCN), a recurrent neural network (RNN), and a recurrent convolutional neural network (recurrent CNN).
Citation Information
Patent Citations
Image registration
JP2010515502A
A method for recording multiple 3D images of dental objects.
JP2014527861A
Tooth wear detection using intraoral 3D scanning
JP2017526495A
Intraoral scanning with surface differentiation
WO2019158442A1
