Neural network based generation and placement of dental restorative dental appliances

Abstract

Techniques are described for using neural networks to automate the design of dental restorative appliances. An example computing device receives transformation information associated with a current dental anatomy of a dental restoration patient, provides the transformation information associated with the current dental anatomy of the dental restoration patient as input to a neural network trained with transformation information that indicates placement of a dental appliance component relative to one or more teeth of a corresponding dental anatomy for a dental restoration process of the one or more teeth, and executes the neural network using the input to produce placement information for the dental appliance component relative to the current dental anatomy of the dental restoration patient.

Description

Technical Field

[0001] This disclosure relates to dental restorative appliances for reshaping teeth. Background Technology

[0002] Dentists typically use dental instruments to reshape or restore a patient's dental anatomy. Dental instruments are usually constructed based on a model of the patient's dental anatomy and reinforced to the desired structure. This model can be a physical or digital model. The design of dental instruments is often a manual, time-consuming, and imprecise process. For example, dentists often design instrument models through trial and error. They may add, remove, reposition, rearrange features, and / or resize features until they are satisfied with the model. Summary of the Invention

[0003] This disclosure relates to techniques for the design and / or automated placement of dental prosthetic appliances used to restore the dental anatomy of a given patient. A computational system configured according to aspects of this disclosure implements a neural network-based system trained with any of a variety of datasets to generate design aspects (or “geometry”) and / or placement characteristics (or “transformations”) of dental prosthetic appliances for a particular patient. To generate patient-specific customized geometry and / or transformation information for dental prosthetic appliances, the computational system of this disclosure implements neural networks trained with dental anatomy and / or dental appliance information obtainable from a digital library of predefined appliance geometries and / or prefabricated appliance “benchmark truths” (e.g., manually created by a skilled physician or automatically created via a rule-based system as described in WO 2020 / 240351, filed May 20, 2020 (the entire contents of which are incorporated herein by reference).

[0004] The geometry of a dental appliance is represented by a digital three-dimensional (3D) mesh incorporating features of the generated geometry. In various examples, the neural network-based techniques of this disclosure can generate new geometries based on the patient's dental anatomy and a dataset of "benchmark ground truth" appliance geometries used to train the neural network. In some examples, the neural network-based techniques of this disclosure can generate placement data for geometries selected from a digital library of dental appliance geometries, and then place the library geometry relative to one or more of the patient's teeth based on the patient's dental anatomy and a dataset of transformation matrices used to train the neural network.

[0005] In some examples, the neural network-based techniques of this disclosure can generate placement data for dental appliances selected from a digital library of dental appliance geometries. Examples of placement data may relate to one or more of the position, orientation, proportion, or shear mapping of the dental appliance during placement in a patient's dental restorative treatment. The computational system of this disclosure can implement simple neural networks (e.g., with a relatively few hidden layers or no hidden layers), graph convolutional neural networks (GCNN), generative adversarial networks (GAN), conditional generative adversarial networks (cGAN), encoder-decoder-based CNNs, U-Net CNNs, GANs with PatchGAN discriminators, and / or another deep neural network (DNN) to perform the various techniques described herein.

[0006] Thus, the computational system of this disclosure trains and executes a neural network to automate the selection / generation of 3D meshes for dental appliance components and the placement of 3D meshes for dental appliance library components, all of which are later combined and assembled into a complete dental appliance. The computational system of this disclosure can use a variety of data to train the neural network, such as dental anatomical landmarks, patient-to-mesh mappings, two-dimensional (2D) dental anatomical images, etc., to define dental prosthetic appliance components and their placement. The geometry and placement information generated by the neural network trained with a dataset selected according to the techniques of this disclosure automatically generate a mesh (3D digital custom model), which can be used to manufacture patient-specific dental prosthetic appliances, such as 3D printing the prosthetic appliance based on the mesh.

[0007] In one example, a computing device includes an input interface and a neural network engine. The input interface is configured to receive transformation information associated with the current dental anatomy of a dental prosthesis patient. The neural network engine is configured to provide the transformation information associated with the current dental anatomy of the dental prosthesis patient as input to a neural network trained with the transformation information, which indicates the placement of dental appliance components relative to one or more teeth in the corresponding dental anatomy, for dental prosthesis treatment of those teeth. The neural network engine is further configured to use the input to execute the neural network to generate placement information of the dental appliance components relative to the current dental anatomy of the dental prosthesis patient.

[0008] In another example, a method includes receiving transformation information associated with the current dental anatomy of a dental prosthetic patient. The method also includes providing the transformation information associated with the current dental anatomy of the dental prosthetic patient as input to a neural network trained with the transformation information, which indicates the placement of dental appliance components relative to one or more teeth for dental prosthetic treatment of the corresponding dental anatomy. The method further includes using the input to execute the neural network to generate placement information of the dental appliance components relative to the current dental anatomy of the dental prosthetic patient.

[0009] In another example, an apparatus includes: means for receiving transformation information associated with the current dental anatomy of a dental prosthesis patient; means for providing the transformation information associated with the current dental anatomy of the dental prosthesis patient as input to a neural network trained with the transformation information, the transformation information indicating the placement of dental appliance components relative to one or more teeth of the corresponding dental anatomy, the dental appliance being used for dental prosthesis treatment of the one or more teeth; and means for using the input to execute the neural network to generate placement information of the dental appliance components relative to the current dental anatomy of the dental prosthesis patient.

[0010] In another example, a non-transitory computer-readable storage medium is encoded with instructions. When executed, these instructions cause one or more processors of a computing system to: receive transformation information associated with the current dental anatomy of a dental prosthesis patient; provide the transformation information associated with the current dental anatomy of the dental prosthesis patient as input to a neural network trained with the transformation information, which indicates the placement of dental appliance components relative to one or more teeth in the corresponding dental anatomy, the dental appliance being used for dental prosthesis treatment of the one or more teeth; and use the input to execute the neural network to generate placement information of the dental appliance components relative to the current dental anatomy of the dental prosthesis patient.

[0011] In one example, a computing device includes an input interface and a neural network engine. The input interface is configured to receive one or more three-dimensional (3D) tooth meshes associated with the current dental anatomy of a dental prosthetic patient and a 3D part mesh representing the geometry for generating a dental appliance component. The neural network engine is configured to provide the one or more 3D tooth meshes and 3D part meshes received by the input interface as input to a neural network trained with training data including baseline ground truth dental appliance component geometry and 3D tooth meshes for the corresponding dental prosthetic case. The neural network engine is further configured to execute the neural network using the provided input to generate an updated model of the dental appliance component relative to the current dental anatomy of the dental prosthetic patient.

[0012] In another example, a method includes receiving at an input interface one or more three-dimensional (3D) tooth meshes associated with the current dental anatomy of a dental prosthetic patient and a 3D part mesh representing the geometry for generating a dental appliance component. The method also includes a neural network engine communicatively coupled to the input interface providing the one or more 3D tooth meshes and 3D part meshes received by the input interface as input to a neural network trained with training data including baseline ground truth dental appliance component geometry and 3D tooth meshes for the corresponding dental prosthetic case. The method further includes the neural network engine using the provided input to execute the neural network to generate an updated model of the dental appliance component relative to the current dental anatomy of the dental prosthetic patient.

[0013] In another example, an apparatus includes: means for receiving one or more three-dimensional (3D) tooth meshes associated with the current dental anatomy of a dental prosthetic patient and a 3D part mesh representing geometry for generating a dental appliance component; means for providing the one or more 3D tooth meshes and 3D part meshes received by an input interface as input to a neural network trained with training data including baseline ground truth dental appliance component geometry and 3D tooth meshes for the corresponding dental prosthetic case; and means for using the provided input to execute the neural network to generate an updated model of the dental appliance component relative to the current dental anatomy of the dental prosthetic patient.

[0014] In another example, a non-transitory computer-readable storage medium is encoded with instructions. When executed, these instructions cause one or more processors of a computing system to: receive one or more three-dimensional (3D) tooth meshes and a 3D part mesh representing the geometry for generating dental appliance components, associated with the current dental anatomy of a dental prosthetic patient; provide the one or more 3D tooth meshes and 3D part meshes received by an input interface as input to a neural network trained with training data, including baseline ground truth dental appliance component geometry and 3D tooth meshes for the corresponding dental prosthetic case; and execute the neural network using the provided input to generate an updated model of the dental appliance component relative to the current dental anatomy of the dental prosthetic patient.

[0015] In one example, a computing device includes an input interface and a neural network engine. The input interface is configured to receive a two-dimensional (2D) image of the current dental anatomy of a dental patient undergoing restoration. The neural network engine is configured to provide the 2D image of the current dental anatomy of the dental patient as input to a neural network trained with training data including 2D images of pre-restoration and post-restoration dental anatomy from previously performed dental restoration cases. The neural network engine is further configured to use the input to execute the neural network to generate a 2D image of a proposed dental anatomy for the dental patient, which is associated with the post-restoration outcome of the dental restoration plan for the dental patient.

[0016] In another example, a method includes receiving a two-dimensional (2D) image of the current dental anatomy of a dental prosthetic patient. The method also includes feeding the 2D image of the current dental anatomy of the dental prosthetic patient as input to a neural network trained with training data including 2D images of pre-restorative and post-restorative dental anatomy from previously performed dental prosthetic cases. The method further includes using the input to execute the neural network to generate a 2D image of a proposed dental anatomy of the dental prosthetic patient, which is associated with the post-restorative outcome of the dental prosthetic plan for the dental prosthetic patient.

[0017] In another example, an apparatus includes: means for receiving a two-dimensional (2D) image of the current dental anatomy of a dental restoration patient; means for providing the 2D image of the current dental anatomy of the dental restoration patient as input to a neural network trained with training data including 2D images of pre-restoration dental anatomy and corresponding 2D images of post-restoration dental anatomy of previously performed dental restoration cases; and means for using the input to execute the neural network to generate a 2D image of a proposed dental anatomy of the dental restoration patient, which is associated with the post-restoration outcome of a dental restoration plan for the dental restoration patient.

[0018] In another example, a non-transitory computer-readable storage medium is encoded with instructions. When executed, these instructions cause one or more processors of a computing system to: receive a two-dimensional (2D) image of the current dental anatomy of a dental patient undergoing restoration; provide the 2D image of the current dental anatomy of the dental patient as input to a neural network trained with training data including 2D images of pre-restoration and post-restoration dental anatomy of previously performed dental restoration cases; and use the input to execute the neural network to generate a 2D image of a proposed dental anatomy of the dental patient, which is associated with the post-restoration outcome of the dental restoration plan for the dental patient.

[0019] In one example, a computing device includes an input interface and a neural network engine. The input interface is configured to receive one or more three-dimensional (3D) tooth meshes associated with the current dental anatomy of a dental prosthetic patient. The neural network engine is configured to provide the one or more 3D tooth meshes received by the input interface as input to a neural network trained with training data including a baseline ground truth dental appliance component geometry and 3D tooth meshes for the corresponding dental prosthetic case. The neural network engine is further configured to execute the neural network using the provided inputs to generate a customized geometry for the dental appliance component relative to the current dental anatomy of the dental prosthetic patient.

[0020] In another example, a method includes receiving one or more three-dimensional (3D) tooth meshes associated with the current dental anatomy of a dental prosthetic patient. The method also includes feeding the one or more 3D tooth meshes associated with the current dental anatomy of the dental prosthetic patient as input to a neural network trained with training data including baseline ground truth dental appliance component geometry and 3D tooth meshes for the corresponding dental prosthetic case. The method further includes using the provided input to execute the neural network to generate a customized geometry for the dental appliance component relative to the current dental anatomy of the dental prosthetic patient.

[0021] In another example, an apparatus includes: means for receiving one or more three-dimensional (3D) tooth meshes associated with the current dental anatomy of a dental prosthetic patient; means for providing the one or more 3D tooth meshes associated with the current dental anatomy of the dental prosthetic patient as input to a neural network trained with training data including a baseline ground truth dental appliance component geometry and 3D tooth meshes for the corresponding dental prosthetic case; and means for using the provided input to execute the neural network to generate a customized geometry of the dental appliance component relative to the current dental anatomy of the dental prosthetic patient.

[0022] In another example, a non-transitory computer-readable storage medium is encoded with instructions. When executed, these instructions cause one or more processors of a computing system to: receive one or more three-dimensional (3D) tooth meshes associated with the current dental anatomy of a dental prosthesis patient; provide the one or more 3D tooth meshes associated with the current dental anatomy of the dental prosthesis patient as input to a neural network trained with training data including a baseline ground truth dental appliance component geometry and 3D tooth meshes of the corresponding dental prosthesis case; and use the provided input to execute the neural network to generate a customized geometry of the dental appliance component relative to the current dental anatomy of the dental prosthesis patient.

[0023] The techniques and practical applications described herein can offer certain advantages. For example, by automatically determining the geometry and placement of 3D meshes for forming components of dental appliances or for forming an overall model of a dental appliance for a patient's restorative treatment, the computational system of this disclosure can improve data accuracy and save resources. For example, by generating more accurate 3D mesh components for dental appliances, the computational system of this disclosure can improve the function and effectiveness of dental appliances in restorative treatments.

[0024] In cases where the computational system utilizes neural networks with reduced layers to predict placement information for dental appliances, the computational system can reduce computational resource usage by implementing neural networks with fewer hidden layers. In cases where the computational system utilizes GANs, GCNNs, cGANs, encoder-decoder-based CNNs, U-Net CNNs, GANs with PatchGAN discriminators, or other deep neural networks to generate the geometry of 3D meshes, the computational system provides process improvements by reducing iterations caused by defective or suboptimal dental appliances supplied to dentists during the performance of restorative procedures for patients. Thus, the neural network-based dental appliance placement technique of this disclosure improves speed, accuracy, and predictability.

[0025] Faster and / or more accurate restoration of a patient's dental anatomy can improve function (e.g., reduce grinding or interference between teeth), which can improve the patient's quality of life, for example, by reducing pain caused by suboptimal dental morphology, integrity, or functional performance. In some examples, more accurate restoration of a patient's dental anatomy can improve the appearance of the patient's dental anatomy, which can further improve the patient experience and / or quality of life. Furthermore, by creating a precise, rapid, and predictable process for restoring dental anatomy using geometry and / or placement formed by neural networks, the computational system disclosed herein offers increased efficiency for a wider range of dentists and improved affordability for a wider range of patients.

[0026] Details of one or more examples will be shown in the accompanying drawings and the description below. Other features, objects, and advantages of the invention will become apparent from the detailed description, the drawings, and the claims. Attached Figure Description

[0027] Figure 1 A block diagram illustrating an exemplary system for designing and manufacturing dental appliances for restoring a patient's dental anatomy, according to various aspects of this disclosure.

[0028] Figure 2 It is shown Figure 1 The system can be executed by executing a neural network trained according to aspects of this disclosure, and a flowchart of an exemplary process for generating a digital model of a dental appliance is provided.

[0029] Figure 3 This is a flowchart illustrating an exemplary use of a neural network for the placement of library components for dental appliances according to aspects of this disclosure.

[0030] Figure 4 This is a flowchart illustrating an example of neural network-based component geometry generation according to aspects of this disclosure.

[0031] Figure 5 This is a flowchart illustrating a process that a computing device according to aspects of this disclosure can implement to generate component geometry using GAN.

[0032] Figure 6 This is a rendering illustrating an exemplary center fixture placement performed according to the neural network-based placement technique of this disclosure.

[0033] Figure 7 This is a rendering showing an example of an adhesive pad (e.g., a lingual bracket) customized according to the shape of the corresponding tooth.

[0034] Figure 8 This is a rendering showing an example of a set of components that make up the tongue-side bracket.

[0035] Figure 9 This is a flowchart illustrating another example of neural network-based component geometry generation according to aspects of this disclosure.

[0036] Figure 10 This is a conceptual diagram illustrating the co-training process of a generator network and a discriminator network of a cGAN according to aspects of this disclosure, the cGAN being configured to render 2D images of a proposed dental anatomy for a patient.

[0037] Figure 11AThe inputs and outputs of a generator network trained with cGAN are shown. This generator network is configured to generate a 2D image of a proposed dental anatomy using a 2D rendering of the patient’s current dental anatomy.

[0038] Figure 11B A comparison is shown between current dental anatomy images, proposed dental anatomy images of this disclosure, and baseline true-value repaired images.

[0039] Figure 12 It shows Figure 1 A menu that can be displayed on a computing device, which is part of a graphical user interface (GUI) that includes current dental anatomy images and / or proposed dental anatomy images of the present disclosure.

[0040] Figure 13A and Figure 13B This is a conceptual diagram illustrating an exemplary mold parting surface according to various aspects of this disclosure.

[0041] Figure 14 This is a conceptual diagram illustrating an exemplary gingival trimming surface according to various aspects of this disclosure.

[0042] Figure 15 This is a conceptual diagram illustrating an exemplary facet strip according to various aspects of this disclosure.

[0043] Figure 16 This is a conceptual diagram illustrating an exemplary tongue-side shelf according to various aspects of this disclosure.

[0044] Figure 17 This is a conceptual diagram illustrating exemplary dental incisions and dental windows according to various aspects of this disclosure.

[0045] Figure 18 This is a conceptual diagram illustrating an exemplary rear snap clamp according to various aspects of this disclosure.

[0046] Figure 19 This is a conceptual diagram illustrating an exemplary door hinge according to various aspects of this disclosure.

[0047] Figure 20A and Figure 20B This is a conceptual diagram illustrating an exemplary tooth latch according to various aspects of this disclosure.

[0048] Figure 21 This is a conceptual diagram illustrating exemplary ridges according to various aspects of this disclosure.

[0049] Figure 22 This is a conceptual diagram illustrating an exemplary central clamp according to various aspects of this disclosure.

[0050] Figure 23This is a conceptual diagram illustrating an exemplary dental vent according to various aspects of this disclosure.

[0051] Figure 24 This is a conceptual diagram illustrating an exemplary dental incision according to various aspects of this disclosure.

[0052] Figure 25 This is a conceptual diagram illustrating an exemplary interdental matrix according to various aspects of this disclosure.

[0053] Figure 26 This is a conceptual diagram illustrating an exemplary manufacturing shell frame and an exemplary dental appliance according to various aspects of this disclosure.

[0054] Figure 27 This is a conceptual diagram illustrating an exemplary dental appliance including a custom label according to various aspects of this disclosure. Detailed Implementation

[0055] Figure 1 A block diagram illustrating an exemplary system for designing and manufacturing dental appliances for restoring a patient's dental anatomy, according to various aspects of this disclosure. Figure 1 In the example, system 100 includes clinic 104, appliance design facility 108, and manufacturing facility 110.

[0056] Dentist 106 may treat patient 102 at clinic 104. For example, dentist 106 may create a digital model of patient 102's current dental anatomy. Dental anatomy may include any part of the crown or root of one or more teeth of the dental arch, gingiva, periodontal ligament, alveolar bone, cortical bone, bone grafts, implants, pulp fillings, artificial crowns, bridges, dentures, orthodontic appliances, or any structure (natural or synthetic) that may be considered part of patient 102's dental anatomy before, during, or after treatment.

[0057] In one example, the digital model of the current dental anatomy includes a three-dimensional (3D) model of the current (pre-treatment) dental anatomy of patient 102. In various examples, clinic 104 may be equipped with an intraoral scanner, cone-beam computed tomography (CBCT) scanner (e.g., 3D X-ray) device, optical coherence tomography (OCT) device, magnetic resonance imaging (MRI) machine, or any other 3D image capture system that dentist 106 may use to generate a 3D model of the dental anatomy of patient 102.

[0058] exist Figure 1In the example shown, clinic 104 is equipped with computing system 190. Computing system 190 may represent a single device or a group of securely interconnected devices. In these examples, the individual devices of computing system 190 may be securely interconnected by being fully contained within the logical constraints of clinic 104 (e.g., by physical connections within clinic 104, such as using a local area network or "LAN") and / or by securely communicating over a public network such as the Internet via encrypted communication based on a Virtual Private Network (VPN) tunnel. Computing system 190 may include one or more user-facing computing devices, such as personal computers (e.g., desktop computers, laptop computers, netbooks, etc.), mobile devices (e.g., tablets, smartphones, personal digital assistants, etc.), or any other electronic device configured to provide end-user computing capabilities, such as by presenting resources in a human-understandable form (e.g., visual images, such as medical / dental imaging, readable output, symbol / graphic output, audible output, tactile output, etc.).

[0059] The dentist 106 may store a digital model of the patient 102's current dental anatomy in a storage device included in or readable / writable by the computing system 190. In some examples, the computing system 190 may also store a digital model of a proposed dental anatomy for the patient 102. The proposed dental anatomy represents the intended function, integrity, and morphology of the dental anatomy to be achieved by applying dental appliances 112 as part of the dental restorative treatment of the patient 102.

[0060] In one example, dentist 106 may generate a physical model of the proposed dental anatomy and may use an image capture system (e.g., as described above) to generate a digital model of the proposed dental anatomy. In another example, dentist 106 may modify the digital model of the current anatomy of patient 102 (e.g., by adding material to the surface of one or more teeth of the dental anatomy, or otherwise) to generate a digital model of the proposed dental anatomy for patient 102. In yet another example, dentist 106 may use computational system 190 to modify the digital model of the current dental anatomy of patient 102 to generate a model of the proposed dental anatomy for patient 102.

[0061] In one scenario, computing system 190 outputs a digital model representing the current and / or proposed dental anatomy of patient 102 to another computing device, such as computing device 150 and / or computing device 192. Although described herein as performing locally at computing systems 190, 150, and 192, it should be understood that in some examples, one or more of computing systems 190, 150, and 192 may leverage cloud computing capabilities and / or Software as a Service (SaaS) capabilities to perform the underlying processing used for the functions described herein. Figure 1 As shown, in some examples, the computing device 150 of design facility 108, the computing system 190 of clinic 104, and the computing device 192 of manufacturing facility 110 may be communicatively coupled to each other via network 114. In various examples, network 114 may represent or include a dedicated network associated with an association (e.g., a dental service network, etc.), or other entities or groups of entities.

[0062] In other examples, network 114 may represent or include public networks, such as the Internet. Although shown as such for illustrative purposes only. Figure 1 While referring to a single entity, it should be understood that network 114 may include a combination of multiple public and / or private networks. For example, network 114 may represent a private network implemented using public network infrastructure, such as a VPN tunnel implemented over the Internet. Therefore, network 114 may include one or more of a wide area network (WAN) (e.g., the Internet), a LAN, a VPN, and / or another wired or wireless communication network. Network 114 may include networks conforming to one or more standards (such as those implemented via...). Wi-Fi TM , Wired or wireless network components (such as 3G, 4G LTE, 5G, etc.).

[0063] exist Figure 1 In the example, computing device 150 is implemented by or at design facility 108. Computing device 150 is configured to automatically design dental appliances and / or generate placement information for reshaping the dental anatomy of patient 102. Computing device 150 (or components thereof) implements neural network technology to determine the geometry of the dental appliance and / or the placement information of the dental appliance. Figure 1 In the example shown, computing device 150 includes one or more processors 172, one or more user interface (UI) devices 174, one or more communication units 176, and one or more storage devices 178.

[0064] UI device 174 may be configured to receive input data from a user of computing device 150 and / or provide output data to a user of computing device 150. One or more input components of UI device 174 may receive input. Examples of input include, but are only a few examples; haptic, audio, kinetic, and optical input. For example, UI device 174 may include one or more of the following: a mouse, keyboard, voice input system, image capture device (e.g., still camera and / or video camera hardware), physical or logical buttons, control panel, microphone or microphone array, or any other type of device for detecting input from a human user or another machine. In some examples, UI device 174 may include one or more presence-sensitive input components, such as resistive screens, capacitive screens, single-finger or multi-finger touchscreens, stylus-sensitive screens, etc.

[0065] The output components of UI device 174 may output one or more of the following: visual (e.g., symbol / graphic or readable) data, haptic feedback, audio output data, or any other output data that can be understood by a human user or another machine. In various examples, the output components of UI device 174 include one or more of the following: a display device (e.g., a liquid crystal display (LCD) monitor, a touchscreen, a stylus-sensitive screen, a light-emitting diode (LED) display, an optical head-mounted display (HMD), etc.), a speaker or speaker array, headphones or headphone kits, or any other type of device capable of generating output data in a format that can be understood by a human or machine.

[0066] Processor 172 represents one or more types of processing hardware (e.g., processing circuitry), such as a general-purpose microprocessor, a specially designed processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a collection of discrete logic, a fixed-function circuit, a programmable circuit (or a combination of a fixed-function circuit and a programmable circuit), or any type of processing hardware capable of executing instructions for performing the techniques described herein.

[0067] For example, storage device 178 may store program instructions (e.g., software instructions or modules) executed by processor 172 for implementing the techniques described herein. In other examples, these techniques may be executed by specially programmed circuitry of processor 172 (e.g., in the case of fixed-function circuitry or specially programmed programmable circuitry). In these or other ways, processor 172 may be configured to execute the techniques described herein, sometimes by utilizing instructions and other data accessible from storage device 178.

[0068] Storage device 178 may store data for processing by processor 172. Certain portions of storage device 178 represent temporary memory, meaning that the primary purpose of these portions of storage device 178 is not long-term storage. The short-term storage aspect of storage device 178 may include volatile memory that is not configured to retain stored contents upon deactivation and reactivation (e.g., during a power restart). Examples of volatile memory include random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), and other forms of volatile memory known in the art. In some examples, the short-term memory of storage device 178 (e.g., RAM) may include on-chip memory cells that are coupled with portions of processor 172 to form part of an integrated circuit (IC) or a system-on-a-chip (SoC).

[0069] In some examples, storage device 178 may also include one or more computer-readable storage media. Storage device 178 may be configured to store a larger amount of data than volatile memory. Storage device 178 may be further configured to serve as long-term storage of data as non-volatile memory space and to retain data after an activation / deactivation cycle. Examples of non-volatile memory include solid-state drives (SSDs), hard disk drives (HDDs), flash memory, or electrically programmable memory (EPROM) or electrically erasable programmable memory (EEPROM). Storage device 178 may store program instructions and / or data associated with the software components and / or operating system of computing device 150.

[0070] exist Figure 1 In this example, storage device 178 includes an instrument feature library 164, a model library 166, and a physician preference library 168 (collectively, "libraries 164-168"). Libraries 164-168 may include relational databases, multidimensional databases, maps, hash tables, or any other data structures. In one example, model library 166 includes 3D models of a patient's current and / or proposed dental anatomy. In some instances, libraries 164-168 may be stored locally at computing device 150 or accessible via networked file sharing, cloud storage, or other remote data storage accessible using the network interface hardware of communication unit 176.

[0071] The short-term memory of storage device 178 and processor 172 together provide a computing platform for executing operating system 180. Operating system 180 may represent, for example, an embedded real-time multitasking operating system, or any other type of operating system. Operating system 180 provides a multitasking operating environment for executing one or more software components 182-186. In some examples, operating system 180 may execute any of components 182-188 as an instance of a virtual machine or within a virtual machine instance executed on the underlying hardware. Although shown separately from operating system 180 as a non-limiting example, it should be understood that any of components 182-188 may be implemented as part of operating system 180 in other examples.

[0072] According to the technology disclosed herein, computing device 150 automatically or semi-automatically generates a digital model of a dental appliance 112 for repairing the dental anatomy of patient 102 using one or more types of neural networks trained with dental anatomy-related data and / or appliance feature data relating to patient 102 and / or other (actual or hypothetical) patients. Preprocessor 182 is configured to preprocess the proposed dental anatomy of patient 102.

[0073] In one example, preprocessor 182 performs preprocessing to identify one or more teeth in the proposed dental anatomy of patient 102. In some instances, preprocessor 182 identifies the local coordinate system of each individual tooth and may identify the global coordinate system of each tooth in the proposed dental anatomy (e.g., in one or two dental arches of the proposed dental anatomy). As another example, preprocessor 182 may preprocess the digital model of the proposed dental anatomy to identify the root structures of the dental anatomy.

[0074] In another example, preprocessor 182 can identify the gingiva in the proposed dental anatomy, thereby identifying and depicting the gingival portion and the tooth portion of the proposed dental anatomy. As another example, preprocessor 182 can preprocess the digital model of the proposed dental anatomy by extending the tooth roots to identify the top surface of the root of each corresponding tooth. In various use cases, preprocessor 182 can perform one, several, or all of the above-described exemplary functions based on requests provided by dentist 106, based on data availability regarding patient 102 and / or other patients, and possibly based on other factors.

[0075] The computing device 150 (or its hardware / firmware components) may invoke or activate the neural network engine 184 to determine placement information of the dental appliance 112 during dental restorative treatment of the patient 102. In some examples, the neural network engine 184 may be implemented as a two-hidden-layer neural network trained with placement information for the patient 102 and / or other patients with substantially corresponding dental anatomy (current or proposed dental anatomy).

[0076] In these examples, the neural network engine 184 can be implemented as a neural network to accept individual position / orientation information of each of two teeth (e.g., a pair of adjacent teeth) in the current dental anatomy of patient 102 as input, and can output placement information of dental appliance 112 during dental restorative treatment. For example, the placement information may directly or indirectly reflect one or more of the position, orientation, or size of dental appliance 112 when used for dental restorative treatment of patient 102.

[0077] In one example, the neural network engine 184 may use a backpropagation algorithm to train the neural network, which uses a single 4×4 transformation for each of two adjacent teeth, and another 4×4 transformation to identify the “baseline truth” {position, orientation, size} tuple of the dental appliance 112 after placement is completed as part of the dental restoration procedure for patient 102. As used herein, the term “transformation” refers to information about the change (or “increment”) of the {position, orientation, size} tuple, and can therefore also be described as the {translation, rotation, scaling} tuple of the dental appliance 112.

[0078] In some instances, the transformations of this disclosure may also include additional elements, such as a shearing map (or simply "shear") associated with the dental appliance 112. Thus, in various examples, the transformations of this disclosure may represent affine transformations and may include some or all of the transformations possible under automorphism in affine space.

[0079] The neural network engine 184 can extract transformations for specific teeth from 3D mesh data describing the current and / or proposed dental anatomy of patient 102. As used herein, the term "benchmark truth" refers to a verified or otherwise well-founded description of a dental anatomical feature or appliance feature. Therefore, in some examples, benchmark truth transformations may be manually generated by dentist 106 or a technician using CAD tools.

[0080] In other examples, baseline truth transformations can be automatically generated, such as by using the automation techniques described in WO2020 / 240351, filed May 20, 2020, the entire contents of which are incorporated herein by reference. The following non-limiting examples of the positioning, orientation, and size of a central clamp placed in the gap between two adjacent teeth during dental restorations describe various techniques of this disclosure. However, it should be understood that the techniques of this disclosure can also be implemented by the neural network engine 184 to generate geometric and / or placement information for other types of dental appliances.

[0081] As part of the placement information for generating the central fixture, the neural network engine 184 can identify landmarks for the proposed dental anatomy. Exemplary landmarks include slices, midpoints, gingival boundaries, the nearest point between two adjacent teeth (e.g., the contact point or nearest neighbor point between adjacent teeth), convex hulls, centroids, or other landmarks. A slice refers to a cross-section of a dental anatomy. The midpoint of a tooth refers to the geometric center of the tooth within a given slice (also known as the geometric midpoint).

[0082] The gingival boundary refers to the boundary between the gingiva and one or more teeth in dental anatomy. A convex hull is a polygon whose vertices include a subgroup of vertices from a given set of vertices, where the boundary of that subgroup circumscribes the entire set of vertices. The centroid of a tooth refers to its midpoint, center point, centroid, or geometric center. In some instances, the neural network engine 184 can use a local coordinate system for each tooth to determine one or more of these landmarks as expressed.

[0083] In some examples, as part of the landmark identification, the neural network engine 184 determines multiple slices of the patient's proposed dental anatomy. In one example, each slice has the same thickness. In some instances, the thickness of one or more slices differs from the thickness of another slice. The thickness of a given slice can be predefined. In one instance, the neural network engine 184 automatically determines the thickness of each slice using a simplified neural network of this disclosure. In another instance, the thickness of each slice can be user-defined and, for example, obtained as a baseline truth value input to the simplified neural network.

[0084] As part of the landmark identification, in some examples, the neural network engine 184 can determine the midpoint of each tooth in relation to the placement of the dental appliance 112. In one example, the neural network engine 184 uses the midpoint of the particular tooth to identify the landmark by calculating the extreme values ​​of the geometry of the particular tooth based on the entire particular tooth (e.g., without dividing the dental anatomy into slices) and determining the midpoint of the particular tooth based on the extreme values ​​of the tooth geometry.

[0085] In some examples, the neural network engine 184 can determine the midpoint of each tooth in each slice. For example, by calculating the centroids of a series of vertices around the edge of a particular tooth in a particular slice, the neural network engine 184 can determine the midpoint of that particular tooth in that particular slice. In some instances, the midpoint of a particular tooth in a particular slice may be offset toward one edge of the tooth (e.g., in cases where one edge has more points than another).

[0086] In other examples, as part of placing the generated landmark identifier portion, the neural network engine 184 may determine the midpoint of a particular tooth in a particular slice based on the convex hull of that tooth in that particular slice. For example, the neural network engine 184 may determine the convex hull of a set of edge points of a tooth in a given slice. In some instances, as part of the landmark identifier, the neural network engine 184 executes a neural network that determines the geometric center from the convex hull by performing a flood-fill operation on the region circumscribed by the convex hull and calculating the centroid of the flood-filled convex hull.

[0087] In some examples, the neural network executed by neural engine 184 outputs the nearest point between two adjacent teeth. The nearest point between two adjacent teeth can be a contact point or a nearest neighbor point. In one example, neural engine 184 determines the nearest point between two adjacent teeth for each slice. In another example, neural engine 184 determines the nearest point between two adjacent teeth based on the entire adjacent tooth (e.g., without dividing the dental anatomy into slices).

[0088] Using landmarks calculated for the proposed dental anatomy, a neural network executed by neural network engine 184 generates one or more custom appliance features for dental appliance 112, based at least in part on these landmarks. For example, the custom appliance feature generator 184 can generate custom appliance features by determining characteristics such as the size, shape, location, and / or orientation of the custom appliance feature. Examples of custom appliance features include splines, mold parting surfaces, gingival trimming surfaces, shells, facet bands, lingual shelves (also known as “rigid ribs”), incisors, windows, incisal ridges, shell frame struts, interdental matrix wrapping, and so on.

[0089] In some examples, the neural network engine 184 may identify and use features other than those listed above. For example, the neural network engine 184 may identify and use features that are recognizable and processable by the processor 172 within the mathematical framework of the executed neural network. Thus, the operation performed by the neural network via the neural network engine 184 can represent an "unknown box" based on the features used and the mathematical framework applied by the neural network during execution.

[0090] A spline is a curve that passes through multiple points or vertices, such as a piecewise polynomial parametric curve. A mold parting surface is a 3D mesh that divides the two sides of one or more teeth (e.g., separating the lateral facets of one or more teeth from the lingual facets of one or more teeth). A gingival trimming surface is a 3D mesh that trims the surrounding shell along the gingival margin. The shell refers to a body with nominal thickness. In some examples, the inner surface of the shell matches the surface of the dental arch, and the outer surface of the shell is a nominal offset of the inner surface.

[0091] A face-side band refers to a rigid rib of nominal thickness offset from the outer shell along its surface. A dental window is an opening providing access to the tooth surface, allowing the dental composite material to be placed on the tooth. An incisor is a structure covering the dental window. An incisal ridge provides reinforcement at the incisal edge of the dental appliance 112 and may be derived from the arch. A housing frame strut is a connecting material that couples the parts of the dental appliance 112 (e.g., the lingual portion, the face-side portion, and their sub-components) to the manufacturing housing frame. In this way, the housing frame strut can secure the parts of the dental appliance 112 to the housing frame during manufacturing, protecting the individual parts from damage or wear and / or reducing the risk of misaligned parts.

[0092] In some examples, the neural network executed by neural network engine 184 generates one or more splines based on landmarks. The neural network executed by neural network engine 184 may generate splines based on multiple tooth midpoints and / or the nearest points between adjacent teeth (e.g., contact points or closest proximity points between adjacent teeth). In some instances, the neural network executed by neural network engine 184 generates one spline for each slice. In one instance, neural network engine 184 generates multiple splines for a given slice. For example, neural network engine 184 may generate a first spline for a first subgroup of teeth (e.g., right posterior teeth), a second spline for a second subgroup of teeth (e.g., left posterior teeth), and a third spline for a third subgroup of teeth (e.g., anterior teeth).

[0093] In some scenarios, the neural network engine 184 generates mold parting surfaces based on landmarks. These parting surfaces can be used to separate the surrounding shell used for molding without undercutting. In some examples, the neural network engine 184 generates additional copies of the mold parting surfaces. For instance, a neural network executed by the neural network engine 184 can place one or more copies of the mold parting surfaces with small offsets onto the main parting surface, with the aim of creating interfering conditions during appliance assembly (this could, for example, improve shape adaptation and sealing when applying dental restorative materials to teeth).

[0094] The instrument feature library 164 includes a set of predefined instrument features that may be included in the dental instrument 112. The instrument feature library 164 may include a set of predefined instrument features that define one or more functional characteristics of the dental instrument 112. Examples of predefined instrument features include vents, rear snap clamps, incisor hinges, incisor snaps, incisor registration features, center clamps, custom labels, fabricated housing frames, interdental matrix shanks, and so on. Each vent is configured to allow excess dental composite material to drain from the dental instrument 112.

[0095] The rear latch clamp is configured to couple the facet portion of the dental instrument 112 to the lingual portion of the dental instrument 112. Each incisor hinge is configured to pivotally couple the corresponding incisor to the dental instrument 112. Each incisor latch is configured to hold the corresponding incisor in a closed position. In some examples, the incisor registration feature includes a pair of convex and concave tabs (e.g., along the midline plane) resting on the incisor edge of the dental instrument 112. In one example, the incisor registration feature is used to maintain the vertical alignment of the facet portion and the lingual portion of the dental instrument 112.

[0096] Each central clamp is configured to provide vertical alignment between the lingual and facial portions of the dental appliance 112. Each custom label includes data identifying a part of the dental appliance 112. A manufacturing housing frame is configured to support one or more parts of the dental appliance 112. For example, the manufacturing housing frame can detachably couple the lingual and facial portions of the dental appliance 112 to each other for secure disposal and transport from the manufacturing facility 110 to the clinic 104.

[0097] In some examples, a neural network executed by neural network engine 184 can determine the characteristics of one or more predefined appliance features included in predefined appliance feature library 164. For example, one or more features accessible from predefined appliance feature library 164 can represent part shapes obtained in one or more ways, such as by manual generation (e.g., generated by dentist 106 or via automated generation, such as via the techniques described above in WO2020 / 240351, filed May 20, 2020). Based on the availability and relevance of the patient 106's current dental anatomy, the neural network can be trained (at least in part) using information available from predefined appliance feature library 164.

[0098] In one example, a predefined appliance feature is configured to enable or perform a function attributed to the dental appliance 112. The characteristics of the predefined appliance feature may include one or more of the transformation-related attributes described above (e.g., location, orientation, size) and / or other attributes, such as shape information. A neural network executed by the neural network engine 184 may determine the characteristics of the predefined appliance feature based on one or more rules, such as rules generated and / or refined via machine learning (ML) techniques.

[0099] In some examples, the executed neural network determines the placement information of the posterior snap clamps based on rules. In one example, the neural network engine 184 can generate placement information that positions two posterior snap clamps along an archway during dental restorative procedures, with the two posterior snap clamps positioned at opposite ends of the archway. For example, the first snap clamp may be positioned at one end of the archway during dental restorative procedures, and the second snap clamp may be positioned at the other end of the same archway during dental restorative procedures.

[0100] In some examples, the neural engine 184 may assign the position of a tooth other than the outermost tooth to be restored to one or both of the posterior snap clamps. In some examples, the neural engine 184 positions the concave portion of the posterior snap clamp on the lingual side of the parting surface and the convex portion on the facet side. In some examples, the neural engine 184 determines the placement information of the vent during dental restorative procedures based on rules. For example, the neural engine 184 may assign the vent a position at the midline of the corresponding incisor on the incisal side of the dental appliance 112.

[0101] In some scenarios, the neural network engine 184 determines the placement of the incisor hinges based on rules. In one scenario, the neural network engine 184 assigns a position for each incisor hinge at a corresponding midline of the corresponding incisor. In another scenario, the neural network engine 184 determines that the concave portion of the incisor hinge is anchored to the facet portion of the dental appliance 112 (e.g., toward the incisal edge of the tooth) and positions the convex portion of the incisor hinge to be anchored to the outside of the incisor.

[0102] In one instance, the neural network engine 184 determines the placement of the incisor clip based on rules by positioning it along the midline of the corresponding incisor. In one instance, the neural network engine 184 determines the positioning of the concave portion of the incisor clip, anchoring it to the outside of the incisor and extending downward toward the gum line. In another instance, the positioning determined by the neural network engine 184 anchors the convex portion of the incisor clip to the gum line of the facial band. For example, the incisor clip can secure the incisor in a closed position by latching the convex portion of the incisor clip to the facial band.

[0103] The neural network engine 184 can determine the characteristics of predefined appliance features based on the preferences of the dentist 106. The dentist preference library 168 may include data indicating one or more preferences of the dentist 106. Therefore, the neural network engine 184 can use information related to the dentist 106 from the dentist preference library 168 as training data in the overall training of the neural network to determine the placement or geometric information of the dental appliance 112.

[0104] In various usage scenarios, physician preferences can directly influence the characteristics of one or more instrument features of the dental appliance 112. For example, the physician preference library 168 may include data indicating preferred sizes for various instrument features (such as the size of vents). In some such examples, larger vents may allow the pressure of the dental composite or resin to reach equilibrium more quickly during the filling process, but may result in larger lumps to be trimmed after curing. In these examples, the neural network engine 184 can be trained with scaling information that alters the size of the dental appliance 112 based on preferences attributed to the dentist 106.

[0105] In other examples, physician preferences indirectly influence the properties of appliance features. For instance, physician preference library 168 may include data indicating preferred stiffness or preferred tightness of self-clamping features. Such preference choices can also influence more complex design variations in the degree of activation of the matrix cross-sectional thickness and / or clamping geometry. The neural network engine 184 determines the properties of appliance features by amplifying rules that serve as the basis for training the implemented neural network using the preferences of dentist 106 (or, in some cases, other dentists) available from physician preference library 168. In some examples, neural network engine 184 may amplify rules with physician preference data based on simulations (e.g., Monte Carlo) or finite element analyses performed using physician preference information. In some instances, feature properties may also originate from properties in the material used with the matrix, such as the type of composite material preferred by dentists for use with the appliance.

[0106] Using the output of a neural network executed by neural network engine 184, model assembler 186 generates a digital 3D model of dental appliance 112 for reshaping the dental anatomy of patient 102 (e.g., shaping the current dental anatomy into a proposed dental anatomy). In various examples, model assembler 186 may use custom and / or predefined appliance features that form the output of the neural network executed by neural network engine 184 to generate the digital 3D model. The digital 3D model of dental appliance 112 may include, or be a subset of, one or more point clouds, 3D meshes, or other digital representations of dental appliance 112. In some instances, model assembler 186 stores the digital model of dental appliance 112 in model library 166.

[0107] Model assembler 186 can output a digital model of dental appliance 112 in various ways. In one example, model assembler 186 can (e.g., via network 114 using network interface hardware of communication unit 176) output a digital 3D model of dental appliance 112 to computing device 192 of manufacturing facility 110. By providing the digital 3D model to computing device 192, model assembler 186 enables one or more entities at manufacturing facility 110 to manufacture dental appliance 112. In other examples, computing device 150 can send a digital model of dental appliance 112 to computing system 190 of clinic 104. In these examples, model assembler 186 enables dentist 106 or other entities at clinic 104 to manufacture dental appliance 112 on-site at clinic 104.

[0108] In some examples, computing device 192 may invoke the network interface hardware of communication unit 176 to send a digital 3D model of dental appliance 112 to manufacturing system 194 via network 114. In these examples, manufacturing system 194 manufactures dental appliance 112 based on the digital 3D model of dental appliance 112 formed by model assembler 186. Manufacturing system 194 may use any number of manufacturing techniques, such as 3D printing, chemical vapor deposition (CVD), thermoforming, injection molding, lost-wax casting, milling, machining, or laser cutting, to form dental appliance 112.

[0109] Dentist 106 may receive dental instrument 112 and may use dental instrument 112 to reshape one or more teeth of patient 102. For example, dentist 106 may apply dental composite material to the surface of one or more teeth of patient 102 through one or more incisors of dental instrument 112. Dentist 106 or another clinician at clinic 104 may remove excess dental composite material through one or more vents.

[0110] In some examples, model assembler 186 may store the generated digital 3D model of dental appliance 112 into model library 166. In these examples, model library 166 may provide appliance model heuristics that neural network engine 184 can use as training data when training one or more neural networks. In some examples, model library 166 includes data indicating appliance success criteria associated with each complete instance of dental appliance 112. Neural network engine 184 may augment the neural network training dataset with any appliance success criteria available from model library 166. Appliance success criteria may indicate one or more of manufacturing print yield, physician feedback, patient feedback, customer feedback or ratings, or combinations thereof.

[0111] For example, neural network engine 184 can train the neural network to generate a new or updated placement profile and / or geometry of dental appliance 112 using appliance success criteria determined for previously generated dental appliances via a digital 3D model. As part of the training, the neural network executed by neural network engine 184 can determine whether the appliance success criteria meet one or more threshold criteria, such as threshold manufacturing output, threshold physician-provided rating, threshold patient satisfaction rating, etc.

[0112] In one example, an existing digital 3D model available from model library 166 is a template or reference digital model. In such examples, neural network engine 184 may train the neural network in part based on the template digital model. In various examples, the template digital model may be associated with different characteristics of the current dental anatomy of patient 102, such as a template for a patient with small teeth or whose mouth opening is obstructed beyond a certain width.

[0113] In one example, neural network engine 184 uses a previously generated digital 3D model available from model library 166 to train the neural network. For example, neural network engine 184 utilizes one or more morphological algorithms to adapt the previously generated digital 3D model accessed from model library 166 to the situation represented by dental restorative procedures tailored to the dentition of patient 102 during the training and / or execution of the neural network.

[0114] For example, neural network engine 184 may use morphological algorithms to interpolate the geometry of appliance features and / or generate a new digital model of dental appliance 112 based on the design of an existing digital model. In one instance, the design features of the existing digital model may include dental windows embedded from the periphery, allowing neural network engine 184 to modify the geometry of the existing digital model based on landmarks for different dental anatomy structures.

[0115] Neural engine 184 trains and executes a neural network to perform (and possibly compress) multiple intermediate steps in the process of generating a digital 3D model of dental appliance 112 for placement and geometric purposes. Neural engine 184 uses a 3D mesh of the patient 102's current and / or proposed dental anatomy to generate a feature set describing dental appliance 112. The 3D mesh (or "tooth mesh") and (in available examples) library components form the training input to the neural network.

[0116] As described above, the neural network engine 184 can be trained to automate one or both of the component placement and / or geometry generation of the dental appliance 112. Examples of components include a central clamp registration tab (or "beak"), incisor hinges, incisor clips, incisor vents, rear clip clamps, and various other components. Examples of placement-related factors and / or components that the neural network engine 184 can generate include parting surfaces, gingival trimming, incisors, dental windows, lateral face bands, incisal ridges, lingual shelves, interdental matrix, shell frames, part labels, etc.

[0117] The neural network engine 184 implements a neural network to automate the generation of placement information and / or geometry generation for components (such as incisor hinges and central clamps) that must be placed in specific locations relative to the dental anatomy of patient 102 to perform dental restorations. By utilizing neural network technology to automate the placement and / or geometry information of these components, the procedural turnaround time for the use of dental appliances 112 during dental restorative treatment of patient 102 is significantly reduced.

[0118] Furthermore, by utilizing a neural network trained with a combination of datasets listed in this disclosure, the neural network engine 184 automates placement and / or geometry generation with improved consistency and accuracy, such as through training of a neural network that is updated based on continuous feedback information or other dynamically changing factors.

[0119] The neural network-based automated algorithm of this disclosure, implemented by computing device 150, offers several advantages in the field of restorative dental appliance placement, representing a form of technological improvement. As an example, computing device 150 can invoke neural network engine 184 to generate placement and / or geometric information for dental appliance 112 without explicitly calculating tooth geometric landmarks in each instance.

[0120] Conversely, the neural network engine 184 can execute a neural network trained with the various transformation data and / or baseline ground truth data described above to generate placement and / or geometric information based on these training factors. Furthermore, by continuously improving the output of the neural network based on treatment plans and outcomes from previous patients, feedback from dentist 106 and / or patient 102, and other factors that can be used to fine-tune the neural network using ML techniques, the computing device 150 can improve the accuracy of the data associated with the digital 3D model of the dental appliance 112.

[0121] For example, because the neural network engine 184 can further refine the algorithm by introducing new training data rather than modifying the rule-based logic, the technology disclosed herein can also provide improvements in reusability and computational resource sustainability. While the examples are primarily described in relation to appliances used in dental restorative treatments, it should be understood that in other examples, the neural network engine 184 can be configured to implement the algorithms of this disclosure to similarly generate geometric and / or placement information for other types of dental appliances, such as orthodontic instruments, surgical guides, and bracket bonding templates.

[0122] Although computing device 150 is described herein as performing both training and execution of various neural networks of the present disclosure, it should be understood that in various use cases, the training of a neural network may be performed by a separate device or system from the device performing the training of the neural network. For example, the training system may use a labeled training dataset in the form of... Figure 1 Some or all of the training data described are used to form one or more trained models. Other devices can import the trained models and execute them to produce the various neural network outputs described above.

[0123] Figure 2 This is a flowchart illustrating an exemplary process 200 by which system 100 can generate digital models of dental appliances by executing a neural network trained according to aspects of this disclosure. Process 200 may begin with a training phase 201. As part of training phase 201, neural network engine 184 trains the neural network (202) using transformation data associated with 3D models of various dental anatomy structures. According to various aspects of this disclosure, neural network engine 184 may use backpropagation training techniques to train the neural network.

[0124] In some examples, neural engine 184 may use one 4×4 transformation for each of one or more teeth in the dental anatomy, and one 4×4 transformation to define a baseline truth tuple {position, orientation, size} for dental appliance 112 after placement. In the example where dental appliance 112 represents a central clamp, neural engine 184 may extract transformations for the maxillary central incisor pair (represented as teeth “8 and 9” in the universal numbering system for permanent teeth) or the mandibular central incisor pair (represented as teeth “24 and 25” in the universal numbering system for permanent teeth) from 3D mesh data describing the dentition of the proposed patient 102.

[0125] The baseline truth transformation can represent the "raw" data manually generated by a technician using computer-aided design (CAD) tools to locate and orient the center fixture, or alternatively, it can represent raw data automatically generated in various ways, such as by using the automation techniques described in WO2020 / 240351, submitted on May 20, 2020. By training the neural network using the backpropagation algorithm, the neural network engine 184 generates multiple fully connected layers. That is, a weighted connection exists between a given node in the first layer and each of the corresponding nodes in the next layer.

[0126] The backpropagation algorithm, implemented by Neural Engine 184, adjusts the weights of these node-to-node connections between layers during the training of the neural network, thereby gradually encoding the desired logic into the neural network over multiple training iterations or passes. In some examples of this disclosure, the neural network may include two layers, thereby reducing the computational overhead for both training and final execution.

[0127] While this document describes training a neural network using data from one or more past cases with transformations applied to two teeth, it should be understood that in other examples, neural network engine 184 may also train the neural network using different types and / or amounts of training data. The augmentation of training data may depend on the availability and accessibility of such training data. For example, if accessible from model library 166 or from another source, neural network engine 184 may use transformations for another tooth in the arch of the dental appliance 112 to which it is to be applied, and / or use transformations for one or more teeth in the relative arch to augment the training data for training the neural network.

[0128] In this way, the neural network engine 184 can train the neural network using training data that enables the neural network to determine the positioning information of the dental appliance 112 based on a more comprehensive assessment of the dental anatomy of the patient 102. In these and / or other examples, the neural network engine 184 may augment the training data using relevant preference information obtained from the physician preference library 168, patient feedback information, and / or various other relevant data accessible by the computing device 150.

[0129] After completing the training phase 201 of process 200, the neural network engine 184 may begin execution phase 203. Execution phase 203 may begin (204) when the computing device 150 receives a digital 3D model of the proposed (e.g., post-dental restoration treatment) dental anatomy from the patient 102. In one example, the computing device 150 receives the digital 3D model of the proposed dental anatomy from another computing device, such as the computing system 190 of the clinic 104. The digital 3D model of the proposed dental anatomy for the patient 102 may include a point cloud or 3D mesh of the proposed dental anatomy.

[0130] A point cloud comprises a set of points that represent or define objects in 3D space. A 3D mesh comprises multiple vertices (also called points) and geometric faces (e.g., triangles) defined by these vertices. In one example, dentist 106 generates a physical model of a proposed dental anatomy for patient 102 and uses an image capture system to generate a digital 3D model of the proposed dental anatomy from an image of the captured physical model. In another example, dentist 106 modifies the digital 3D model of patient 102's current anatomy (e.g., by simulating the addition of material to the surface of one or more teeth of the dental anatomy, or by simulating other changes) to generate a digital 3D model of the proposed dental anatomy. In yet another example, computational system 190 may modify the digital model of the current dental anatomy to generate a model of the proposed dental anatomy.

[0131] In some examples, preprocessor 182 preprocesses the 3D model of the proposed dental anatomy to generate a modified model by digitally extending the roots of the initial digital model of the proposed dental anatomy according to the proposed root extensions determined by preprocessor 182, thereby more accurately modeling the complete anatomy of the patient's teeth (206). Step 206 in Figure 2 The dashed boundaries are used to indicate the optional nature of step 206. For example, in some use cases, the preprocessing capabilities provided by step 206 can be categorized as features described herein with respect to neural network engine 184.

[0132] In some examples where preprocessor 182 performs step 206, since the top of the tooth root (e.g., the area furthest from the gum line) can be at different heights, preprocessor 182 can detect the vertices corresponding to the top of the tooth root and then project those vertices along the normal vector, thereby digitally extending the tooth root. In one example, preprocessor 182 (e.g., using the k-means algorithm) groups the vertices into clusters. Preprocessor 182 can compute the average normal vector for each vertex cluster.

[0133] For each vertex cluster, preprocessor 182 can determine the sum of the residual angle difference values ​​between the cluster's average normal vector and the vector associated with each vertex in the cluster. In one example, preprocessor 182 determines which vertex cluster is the top surface of the root based on the sum of the residual angle difference values ​​for each cluster. For example, preprocessor 182 can determine that the cluster with the lowest sum of residual angle difference values ​​defines the top surface of the root.

[0134] The neural engine 184 can obtain one or more tooth transformations (208) based on the proposed dental anatomy represented in the received 3D model. For example, the neural engine 184 can extract corresponding {translation, rotation, scaling} tuples for representing one or more teeth in the 3D model based on the corresponding {position, orientation, size} tuples of teeth in the current dental anatomy and dental restoration result information shown in the 3D model of the proposed (post-restoration) dental anatomy of patient 102.

[0135] The neural network engine 184 can execute a trained neural network to output placement information (210) for the dental appliance 112. In an example where the dental appliance 112 represents a central clamp, the neural network 184 can input two tooth transformations (e.g., transformations describing the position, orientation, and dimensions of two adjacent maxillary central incisors) into a neural network with the aforementioned two layers. The neural network executed by the neural network 184 can output a transformation that positions the central clamp (which may represent a library part) between the two maxillary central incisors and orients it into an overall arch shape perpendicular to the current, intermediate, or proposed dental anatomy of the patient 102.

[0136] The neural network engine 184 can use various sets of basic operations to generate transformations for the dental appliance 112 to be applied to dental restorative procedures for the patient 102 (the output of the trained neural network at execution). As an example, the trained neural network can process a 3D model of a proposed dental anatomy for the patient 102 to automatically detect one or more landmarks within the proposed dental anatomy. In this example, each “landmark” represents a recognizable geometric feature within the 3D model that can be used to determine its position and orientation relative to one or more tooth surfaces. In some examples, the landmarks computed by the trained neural network include one or more slices of the dental anatomy, where each slice may include one or more additional landmarks. For example, the trained neural network can divide a 3D mesh of the proposed dental anatomy into multiple slices and can compute one or more landmarks for each slice, such as the midpoint of each tooth in the slice, the nearest point between two adjacent teeth (e.g., the contact point between two adjacent teeth or the nearest neighbor point between two adjacent teeth), the convex hull of each tooth in the slice, and so on.

[0137] Model assembler 186 generates a 3D model (212) of dental appliance 112. In various examples of this disclosure, model assembler 186 may construct a global 3D mesh for dental appliance 112 based on placement information indicated by the transformation output of a neural network executed by neural network engine 184. For example, model assembler 186 may generate a global 3D mesh for dental appliance 112 based on one or more of the placement properties indicated by the {translation, rotation, scaling} tuples of the transformations of this disclosure.

[0138] In some examples, if the shearing information is available as input data to the executed neural network and / or if the neural network engine 184 otherwise generates shearing information for the dental appliance 112 by executing a trained model of the neural network, the model assembler 186 may also incorporate the shearing or shearing mapping information of the dental appliance 112 into the output transformation.

[0139] In one example, model assembler 186 may extrapolate (e.g., via integration or other similar techniques) one or more properties of the {position, orientation, size} tuple of dental appliance 112 from the transformation output of a neural network. Each {position, orientation, size} tuple generated by model assembler 186 corresponds to a set of appliance properties (e.g., one or both of custom and / or predefined appliance features) for a proposed overall structure of dental appliance 112. In one example, model assembler 186 may determine the position of a custom appliance feature based on the midpoint of a specific tooth.

[0140] For example, model assembler 186 may align or otherwise position the 3D mesh of dental windows and / or incisors (such as exemplary features) based on the midpoint of the teeth. In this way, model assembler 186 may determine the position of predefined appliance features based on transformation information from the output of a neural network executed by neural network engine 184. As an example, model assembler 186 may determine the position of a posterior clamping forceps based on the position of the teeth in the current dental anatomy of patient 102.

[0141] In some instances, the model assembler 186 determines the location of predefined appliance features based on the position of custom appliance features. For example, the model assembler 186 can align a door hinge, door latch, and / or vent with the centerline of the door. Furthermore, the model assembler 186 can adjust the feature orientation, scale, or position based on analysis of the overall model, such as performing finite element analysis to adjust the active clamping force of a latch clamp. The model assembler 186 can also make adjustments (e.g., fine-tuning) based on subsequently anticipated manufacturing tolerances, such as providing appropriate clearance between features.

[0142] Similarly, the model assembler 186 may be tailored to the properties of the material used to generate the physical appliance, such as increasing the thickness when using a more flexible material. In various examples of this aspect of the present disclosure, the model assembler 186 may generate a digital 3D model of the dental appliance 112 to include one or more of a point cloud, a 3D mesh, or other digital representation of the dental appliance 112.

[0143] Computing device 150 outputs a digital 3D model (214) of dental appliance 112. For example, computing device 150 can output the digital 3D model of dental appliance 112 to computing device 192 of manufacturing facility 110 by sending packetized data via network 114 through the network interface hardware of communication unit 176. Manufacturing system 194 manufactures dental appliance 112 (216). For example, computing device 192 can control manufacturing system 194 to manufacture dental appliance 112 such that the dental appliance conforms to placement information generated by a trained neural network (e.g., based on the digital 3D model of dental appliance 112 generated by model assembler 186) executed by neural network engine 184. In various examples, manufacturing system 194 can generate physical dental appliance 112 via 3D printing, CVD, machining, milling, or any other suitable technology.

[0144] In some examples, the computing system 150 receives feedback (218) from the dentist 106 regarding the dental appliance 112. Optional properties of step 218 are described in... Figure 2The text is shown with dashed boundaries. For example, after dentist 106 receives physical dental appliances 112 and uses them for dental restorative treatment of patient 102, dentist 106 can use computing system 190 to send feedback to computing device 150. As an example, computing device 150 can receive data instructing adjustments to the characteristics of future dental appliances (e.g., size, positioning characteristics, orientation characteristics, etc.), designed based on transformation data output by a neural network for a general patient group of patient 102.

[0145] In some examples, the computational system 150 updates the physician preference database 168 (220) based on received physician feedback. Optional properties of step 220 are... Figure 2 The text is shown by dashed boundaries. In some examples, the neural network engine 184 can continuously train the neural network using data available from the physician preference library 168 (which is continuously updated using inbound feedback from physicians).

[0146] Figure 3 This is a flowchart illustrating an exemplary use of a neural network for the placement of library components of a dental appliance 112 according to aspects of this disclosure. An example is described using a dual-hidden-layer neural network to determine the placement of a central clamp registration tab relative to two maxillary central incisors (such as "teeth 8 and 9" as indicated in the universal permanent dentition numbering system) in a specified location and orientation. Figure 3 The starting point of the central clamp will be placed approximately at the midpoint between the two maxillary central incisors and will be oriented such that the vertical (or "Y") axis of the central clamp is perpendicular to the arch of the maxillary incisors.

[0147] The neural network engine 184 can achieve backpropagation-based training of the neural network using a 4×4 transformation for each tooth in the maxillary central incisors and a 4×4 transformation defining the reference ground truth position and orientation of the central fixture after placement. The neural network engine 184 can extract the transformations for the maxillary central incisors from 3D mesh data describing the current dentition of patient 192. The neural network engine 184 can obtain the reference ground truth transformations from various sources, such as those manually produced by technicians using CAD tools for central fixture positioning and orientation, or alternatively automatically generated (e.g., using techniques described in WO 2020 / 240351, submitted May 20, 2020).

[0148] exist Figure 3In the example, neural engine 184 converts each of the 4×4 transformations used for the maxillary central incisors into a corresponding 1×7 quadruplet vector. Neural engine 184 concatenates these two 1×7 quadruplet vectors to produce a single 1×14 feature vector. The 1×14 feature vector corresponds to data from a single patient (hereinafter referred to as a single “case”). The feature vectors of “n” cases can be horizontally concatenated to form an “n”×14 matrix, where “n” represents a non-negative integer value.

[0149] In this way, the neural network engine 184 can encode data from multiple cases into a single matrix, which can be used as training input to train the system. Figure 3 A two-hidden-layer neural network. The neural network engine 184 can use the backpropagation algorithm to train the neural network in some non-limiting examples of this disclosure. The layers of the neural network are fully connected, which means that there are weighted connections between each node i in the first layer and each node j in the next layer.

[0150] The backpropagation training algorithm, executed by the neural network engine 184, adjusts these weights throughout the training process (e.g., through multiple training iterations or laps and their fine-tuning) to progressively encode the desired logic into the neural network over multiple training iterations / laps. According to Figure 3 The specific example shown illustrates that neural network engine 184 uses a fully connected feedforward neural network with two hidden layers. Figure 3 Looking from left to right, the first hidden layer and the second hidden layer have dimensions of 1×32 and 1×64 respectively, and the output dimension is 1×7.

[0151] In other examples consistent with the techniques of this disclosure, the neural network engine 184 may utilize other neural network architectures and techniques, such as recurrent neural networks (RNNs), restricted Boltzmann machines (RBMs), long short-term memory (LSTMs), convolutional neural networks (CNNs), and various other neural network architectures and techniques. In other examples, such as those in which the neural network engine 184 takes a 3D mesh as input, the neural network engine 184 may use a graphical CNN to generate placement information for one or more library components (such as the aforementioned central fixture).

[0152] In other examples, neural engine 184 can use images of the maxillary central incisors (e.g., images of these two teeth captured by rendering) as input to a CNN or a fully connected neural network to generate placement transformations for library components. Neural engine 184 then... Figure 3The weighted connections of the first hidden layer pass 14 input node values. The neural network engine 184 passes each node value in the first hidden layer (with a weighting factor in applicable scenarios) to each corresponding node in the second hidden layer, and so on.

[0153] Once the input has been propagated through the neural network, the neural network 184 transforms the resulting 1×7 vector (interpreted as a quad) into a 4×4 matrix representing the predicted fixture placement matrix. The neural network 184 then computes the layer normalization (“L2 Norm”) of the second layer to obtain the difference between the placement matrix and the baseline truth transformation that serves as the input to the 4×4 transformation. This difference represents the “loss” value, which the neural network engine 184 feeds into the neural network to update the weights via backpropagation.

[0154] After several iterations of this backpropagation process, the neural network is trained to take two tooth transformations (i.e., describing the position and orientation of the two teeth) as input, and output a transformation that positions a library component (e.g., a central clamp) between those teeth and orients it perpendicular to the arch. The training procedure for this particular example flows as follows: Figure 3 As shown in the diagram. Although it is primarily described based on an example using a two-hidden-layer neural network. Figure 3 However, the neural network engine 184 can train and execute more complex computational models in other examples, such as those deep learning models (e.g., generative adversarial networks or "GANs") according to various aspects of this disclosure.

[0155] Figure 4 This is a flowchart illustrating an example of neural network-based component geometry generation according to aspects of this disclosure. The neural network engine 184 is trainable. Figure 4 A neural network-based system is used to generate custom parts, such as dental appliances 112 or their discrete components. In some examples, the neural network engine 184 can use a graphical CNN to generate parts, such as mold parting surfaces. This paper describes the implementation of CNN as a generator network component of a GAN. Figure 4 .

[0156] In this article about Figure 4 In the described example, the dentition of patient 102 (or a patient from a previous case) is described by a set of 3D meshes, where each tooth is represented by its own separate 3D mesh. Each 3D mesh includes a list of vertices and a list of faces describing the relationships between the vertices, provided that the transformation of each tooth is appropriately positioned and oriented to reflect the position of that particular tooth within the arch. In other words, each 3D mesh can specify which vertex is part of which face.

[0157] In these examples, each face is a triangle. The neural network engine 184 inputs the 3D tooth mesh into... Figure 4 The graphical CNN is shown. This graphical CNN then generates the part geometry as output. This output represents the generated part in the form of another 3D mesh, which includes corresponding vertices and faces. The graphical CNN can generate this 3D mesh in one of two ways: 1) by generating a new set of vertices describing the generated part; or 2) by moving a pre-existing set of vertices.

[0158] In the second technique (which involves moving a pre-existing set of vertices), the generator graphical CNN may begin with a template or generalized example of the generated part, and then manipulate the initial set of vertices to adapt the generated part to the dentition of patient 102 (e.g., dental anatomy in current and / or intermediate processing). Subsequently, neural network engine 184 feeds a pair of 3D meshes representing the parts generated by the graphical CNN with the 3D tooth meshes initially input to the graphical CNN into a differentiable discriminator network.

[0159] The neural network engine 184 also feeds a second pairing into the differentiable discriminator network, namely the pairing of the baseline ground truth library component with the 3D tooth mesh initially input into the graphical CNN. The differentiable discriminator network computes the probability of the input pair from the second dataset (i.e., the pairing that includes the baseline ground truth generation component). In other words, the differentiable discriminator network computes the probability that each input dataset corresponds to a baseline ground truth dataset that includes the original tooth mesh and the baseline ground truth mesh of the target geometry.

[0160] A differentiable discriminator network generates gradients, which the neural network engine 184 can use as... Figure 4 The loss function of the generator network (implemented as a CNN in this case) is shown. In the context of machine learning, the loss function quantifies the degree of difference between the machine learning model and the ideal model, and is used to guide the training of the machine learning model. The generator network may also use other loss functions, such as normalization of individual layers (e.g., L1 Norm and / or L2 Norm) and slope distance (which is the sum of positive distances defined for an unsigned distance function). In other examples, the neural network engine 184 may input images of the maxillary central incisors (e.g., images generated from renderings of these two teeth) into a CNN or a fully connected neural network to produce a grid of benchmark truth library components.

[0161] In various examples, the baseline truth generation component can be generated manually using CAD tools or automatically using the techniques described in WO 2020 / 240351, submitted May 20, 2020. In some examples, depending on availability, the neural network engine 184 can use placement information for the dental appliance 112 (such as...) Figure 2 The transformation output of the neural network is used to amplify the training of the generator network.

[0162] Although the discussion primarily focuses on dental appliances (such as dental appliance 112) used for dental restorative procedures, it should be understood that the neural network-based placement and / or geometry generation techniques of this disclosure can also be used in conjunction with other types of dental appliances. Non-limiting examples of other dental appliances that the computing device 150 can customize using the techniques of this disclosure include lingual brackets, integral lingual bracket systems, orthodontic aligners (e.g., clear aligners or transparent aligners), adhesive trays, etc. For example, the neural network engine 184 can train a GAN generator, such as... Figure 4 The GAN's Graph CNN generator is used, but benchmark ground truth components are generated using those found in these other types of dental appliances. In these examples, the trained generator Graph CNN can produce generated features for any of these other types of dental appliances, such as one or more mold parting surfaces. See below for reference. Figure 7 and Figure 8 This section discusses examples of the geometry that neural engine 184 can be used to generate features of tongue-side brackets.

[0163] For example, computing device 150 may invoke neural network 184 to generate placement and / or geometric information for the lingual bracket system, information originally designed by a technician using custom software. As part of the design of the lingual bracket system, neural network engine 184 may generate specifications for the adhesive pad for a specific tooth. Neural network engine 184 may be trained to incorporate various steps of the custom software-based generation process, such as outlining the periphery of a specific tooth, determining the thickness of the formed shell, and removing specific teeth via Boolean operations.

[0164] The neural network engine 184 can train a neural network to select bracket bodies from a library (e.g., a library of appliance features or a model library 166), virtually place the selected bracket bodies on a pad, and combine the pad and the bracket bodies mounted thereon through Boolean addition operations. The neural network can adjust one or more bracket components (e.g., hooks, wings, etc.) to adapt the overall bracket to the specific geometry of a particular tooth and adjacent gingiva. When executed by the neural network engine 184, the neural network can generate a design according to which the adjusted bracket components combine with the bracket bodies to complete the digital design of the entire bracket, and the entire bracket geometry can be derived.

[0165] The neural network engine 184 can encode the entire bracket geometry for export in various ways, such as in the form of a stereolithography (STL) file storing 3D geometric information. To train the neural network for lingual bracket generation, the neural network engine 184 can utilize past cases from a patient group. Using multiple available past cases from multiple patients, the neural engine 184 can train the neural network to automate the design of lingual brackets with relatively little (or no) retraining, thus saving computational resources and achieving improved accuracy in automating the design for the specific dental anatomy characteristics of patient 102, thereby providing improved data precision.

[0166] Examples of custom appliance features that a GAN generator network can generate include information for 3D mesh representations of splines, mold parting surfaces, gingival trimming surfaces, shells, facet bands, lingual shelves, incisors, dental windows, and so on. In one example, the generator network can generate one or more digital meshes representing splines for each slice of a dental anatomy. The GAN generator network can generate splines for a given slice, based on multiple tooth midpoints within the slice and / or the nearest points between adjacent teeth within the slice (e.g., contact points between adjacent teeth or nearest neighbors between adjacent teeth within the slice). In other words, in this example, the generator network accumulates a set of points (e.g., tooth midpoints, contact points between adjacent teeth, nearest neighbors between adjacent teeth, or combinations thereof) for each slice to generate a feature representing a spline for each digital slice.

[0167] In some examples, the generator network generates a mold parting surface as an exemplary feature to be incorporated into the overall 3D model of a dental prosthetic appliance. The neural network engine 184 executes the generator network to generate the mold parting surface based on multiple midpoints and / or the nearest points between adjacent teeth. For example, the generator network can accumulate multiple points for each spline for each slice to generate the mold parting surface. As an example, in an example where the generator network divides the dental anatomy into four slices and generates a single spline for each slice, the points of each of the four splines can be aggregated to generate the mold parting surface.

[0168] In one scenario, neural network engine 184 can feed preference information about dentist 106 from physician preference library 168 into a benchmark truth repository for use as training data augmentation. For example, neural network engine 184 can query physician preference library 168 to determine the preferences of dentist 106. Examples of data stored in physician preference library 168 include preferred sizes, locations, or orientations of predefined appliance features for dentist 106.

[0169] The neural network engine 184 can also train the generator network using data indicating predefined appliance features, such as by accessing and retrieving data from one or more libraries (e.g., stored in a data store, database, data lake, file share, cloud repository, or other electronic repository) representing 3D meshes of predefined features to be incorporated into the entire 3D model of the dental appliance 112. For example, the neural network engine 184 can receive this data by querying the appliance feature library 164. The appliance feature library 164 stores data of 3D meshes defining multiple predefined appliance features such as vents, posterior snap clamps, incisor hinges, incisor snaps, incisor registration features (also referred to as "beaks"), and so on.

[0170] In one example, neural network engine 184 selects one or more predefined appliance features from a plurality of predefined appliance features stored in appliance feature library 164. For example, appliance feature library 186 may include data on a plurality of different predefined appliance features that define a given type of predefined appliance feature. As an example, appliance feature library 164 may include data on different properties (e.g., size, shape, proportion, orientation) of predefined appliance features that define a given type (e.g., data on hinges of different sizes and / or shapes, etc.). In other words, appliance feature library 164 may determine the properties of the predefined appliance features and select features from the predefined appliance library that correspond to the determined properties.

[0171] In some scenarios, the neural network engine 184 performs training data augmentation by selecting predefined appliance features (e.g., a door hinge with a specially set size) from the appliance feature library 164 based on the landmarks of the corresponding teeth, the characteristics of the corresponding teeth (e.g., size, type, location) of the corresponding teeth (e.g., the teeth to be repaired using appliance features when dental appliances are applied to the patient), physician preferences, or both.

[0172] In other examples, the device feature library 164 includes data defining a set of desired predefined appliance features. In some such examples, the neural network engine 184 may retrieve data of 3D meshes representing the predefined features to use each desired predefined feature as additional training data. In such examples, the generator network of the GAN may transform the 3D mesh for inclusion in patient-specific dental appliances. For example, the generator network may rotate or scale (e.g., resize) the 3D mesh for a particular feature based on landmarks of the corresponding tooth, tooth characteristics, and / or physician preferences.

[0173] Figure 5 This is a flowchart illustrating a process 500 that can be implemented by a computing device 150 according to aspects of this disclosure to generate component geometry using GAN. Process 500 generally corresponds to the above description regarding Figure 4 The described technique. Process 500 may begin in training phase 501, where neural engine 184 obtains a 3D mesh of a benchmark ground truth dental appliance component geometry (502). In various examples, neural engine 184 may obtain the 3D mesh of the benchmark ground truth dental appliance component geometry from a source providing a manually generated component geometry or from a source providing a component geometry automatically generated using the technique described in WO 2020 / 240351, submitted May 20, 2020.

[0174] Similarly, as part of the training phase 501, the neural network engine 184 can use a discriminator network to train the generator network using a benchmark ground truth part geometry and a 3D tooth mesh (e.g., Figure 4 (Graphical CNN)(504). For example, the neural network engine 184 can train the generator network by feeding pairs of {generated part geometry, 3D tooth mesh} and {benchmark ground truth part geometry, 3D tooth mesh} into a discriminator network. The neural network engine 184 can execute the discriminator network to compute the probability of each pair, thereby indicating whether the corresponding pair is based on the benchmark ground truth part geometry.

[0175] Although Figure 5 Step 504 is shown as a single step for ease of illustration only, but it should be understood that the neural network engine 184 runs the discriminator network to train the generator network in multiple iterations and continuously fine-tunes the training until the generator network generates sufficiently accurate part geometry to “mimic” the pairing based on the benchmark ground truth geometry relative to the discriminator network.

[0176] Once the neural network engine 184 determines that the generator network has been sufficiently trained, the neural network engine 184 may temporarily suspend or possibly even permanently discard the discriminator network for the execution phase 503 of process 500. In order to begin the execution phase 503, the neural network engine 184 may execute the trained generator network to generate part geometry (506) using a 3D tooth mesh of the patient 102’s current dental anatomy as input.

[0177] In a non-limiting example, neural network engine 184 can execute a trained generator network to generate the mold parting surface of dental appliance 112. Manufacturing system 194 then manufactures dental appliance 112 (508) based on the part geometry generated by the trained generator network. For example, computing device 150 can output the 3D mesh of the part geometry generated by neural network engine 184 to computing device 192 of manufacturing facility 110 by sending packetized data via network 114 through the network interface hardware of communication unit 176.

[0178] Figure 6 This shows the above text about Figure 2 and Figure 3 The rendering of an exemplary central fixture placement performed according to the neural network-based placement technique of this disclosure is discussed. Figure 6 In the two views shown, the central clamp is placed between the two maxillary central incisors (teeth 8 and 9 according to the Universal Numbering System for Permanent Teeth) (e.g., centered or substantially centered) and oriented into an arch shape perpendicular to the proposed dental anatomy of the patient 102.

[0179] Figure 7 This is a rendering showing an example of an adhesive pad (e.g., a lingual bracket) customized to the shape of the corresponding teeth. As mentioned above, regarding... Figure 4 and Figure 5 The described GAN-based technique can be used to generate the geometry of such adhesive pads.

[0180] Figure 8 This is a rendering showing an example of a set of components that make up the tongue-side bracket. (The above is about...) Figure 1-5 The described technology can be used to assemble trays (such as...) Figure 8 (The overall bracket shown), and / or generate bracket placement information for patient 102's teeth.

[0181] Figure 9 This is a flowchart illustrating another example of neural network-based component geometry generation according to aspects of this disclosure. The neural network engine 184 is trainable and executable. Figure 9 The generator network of the GAN shown is used to refine or fine-tune a previously generated dental appliance model to form an updated dental appliance model or an updated model of its components. The neural network engine 184 can refine (or fine-tune or “slightly adjust”) the geometry of a dental appliance model automatically generated using landmark information (e.g., using the techniques described in WO 2020 / 240351, filed May 20, 2020) or the geometry of a dental appliance model manually generated using computer-aided design (CAD) tools to form an updated model (e.g., an updated 3D) of this disclosure.

[0182] The refinement of the previously generated model using a GAN-based generator provides time savings and, in many cases, improves accuracy and data precision regarding the geometric modifications required to make the dental appliance model feasible (the updated model represents feasible dental appliance components for dental prosthetic treatment). While trained sufficiently to fool the discriminator network and / or pass visual inspection, the generator network is configured to progressively modify the component design so that the updated model of the dental appliance geometry is consistent with the design available during the dental prosthetic treatment of patient 102.

[0183] Regarding Figure 4Compared to the described technique (where the component geometry is entirely designed by the generator network of a GAN), Figure 9 The associated techniques will combine computational results from landmark-based automated tools (e.g., the automated tools described in WO 2020 / 240351, submitted May 20, 2020) or manually generated geometry with neural network-based fine-tuning to complete the design (to form an updated part model) by taking advantage of any last-mile minor adjustments that may benefit the dental restoration process for patient 102. Figure 9 The GAN provides fast convergence time, enabling the computing device 150 to take advantage of both landmark-based initial geometry design and neural network-based geometry refinement to generate updated part models in a fast computational manner.

[0184] Figure 9 GANs enable generator networks to be trained even when the number of previously generated models used as training examples is limited. In this way, Figure 9 GANs leverage the engineering elements of the initially designed geometry when training data is limited, while simultaneously offering the benefit of using neural network-based designs for minor adjustments to the original design in the final mile. Figure 4 Compared to GANs, Figure 9 The GAN provides additional input to the generator network (both during the training and execution phases), where the original appliance geometry may need to be slightly adjusted to become the final form (in the form of an updated model or an updated part 3D mesh) for manufacture by the manufacturing system 194 (e.g., by 3D printing).

[0185] Figure 10-12 Aspects relating to this disclosure describe a system configured to display a proposed dental restoration to a patient 102 via an automated design process using generative modeling. According to these aspects of the disclosure, a neural network engine 184 utilizes data collected from a dentist 106 and / or other trained clinicians / technicians to train a neural network configured to generatively model the proposed dental anatomy of the patient 102. For example, the neural network engine 184 may train the neural network to learn the properties of an acceptable dental restoration in a data-driven manner. Examples of the disclosure are described with respect to generating unique two-dimensional (2D) images of the proposed dental anatomy after restoration for a single patient (patient 102 in these examples).

[0186] The following describes the neural network-based generative display technology of this disclosure regarding dental restoration through a non-limiting example of generating 2D images of (post-restoration) dental anatomy. However, it should be understood that the neural network-based generative display technology of this disclosure can also be applied to other fields, such as assisting in the 3D printing of ceramic and / or composite crowns. The goal of the various dental restorative procedures discussed herein is to provide patient 102 with a composite restoration for damaged or unsightly teeth with minimal invasiveness, or to provide dental restorations for other suboptimal conditions associated with the current dental anatomy of patient 102.

[0187] Patient 102 (or any other patient) interested in dental restoration can have their current dental anatomy scanned at clinic 104. The neural network-based generative modeling technique disclosed herein provides a 2D image view of the proposed dental anatomy after restoration, tailored to a given patient (patient 102 in this specific example), which is processed quickly and with high data accuracy. This neural network-based generative modeling technique significantly reduces the delivery time and cost of existing dental restoration planning solutions.

[0188] To improve the data accuracy of generative modeling of post-restorative 2D images of the proposed dental anatomy for patient 102, neural engine 184 can incorporate dental restoration styles (e.g., youth, older, natural, elliptical, etc.) into the training data if information on dental restoration styles is available from past cases. In these and / or other examples, neural engine 184 can incorporate one or more of the following into the training data used to train the neural network: an accepted "golden ratio" criterion for tooth size, an accepted "ideal" indentation shape, patient preference, physician preference, etc. If different styles are available in the training dataset, patient 102 can have the ability to view different restoration options generated by the algorithm in different styles. In other words, neural engine 184 can generate different style options relative to the proposed post-restorative dental anatomy for patient 102 based on different style results from past cases.

[0189] By using this dental restoration-related training data to train a neural network (typically through many training iterations for fine-tuning), the neural network engine 184 improves the accuracy of data regarding the generative modeling of the proposed dental anatomy of patient 102, reduces runtime computational resource consumption (by performing precise training of the neural network), and shortens the overall processing time for generating a dental restoration treatment plan. This reduction is achieved by reducing the number of iterations required for correction or fine-tuning when planning a single round of dental restoration treatment for a given patient.

[0190] Implementing the neural network-based generative modeling techniques of this disclosure via computing device 150, computing devices 150 and 190 also provide various user experience-related improvements. For example, while scanning the current dental anatomy of patient 102, dentist 106 can present patient 102 with a 2D image of the proposed post-restoration dental anatomy by generating a 2D image relatively quickly (and possibly during the same patient encounter). In some examples, dentist 106 can synthesize different post-restoration outcomes (e.g., using different styles or other preference-related factors) to allow patient 102 to view different options to help choose a dental restoration plan.

[0191] In some examples, dentist 106 can provide “pre-approved” goals for generating 3D restoration files, which can be used in the design and / or manufacturing process of dental appliance 112. Providing pre-approved planning information (which neural engine 184 can obtain from dentist preference library 168 or other sources) enables neural engine 184 to train neural networks to generate customized dental restoration models using a reduced number of dentist inputs, thereby compressing the production process for customized products.

[0192] Because patient 102 can visualize the potential post-restorative outcomes of his / her own dental anatomy rather than past cases of other patients, neural network engine 184 also utilizes training data to provide personalization as an improvement to the user experience in these examples. The generative modeling techniques disclosed herein can be applied to areas beyond dental restorations where patients are also interested in unique or customized solutions, such as those related to respirators, bandages, etc.

[0193] According to some examples of this disclosure, neural network engine 184 uses a GAN to generate 2D images of the proposed dental anatomy of patient 102 for use in the post-reconstruction processing stage. As described above, GANs utilize pairings of differentiable functions, typically deep neural networks, with the goal of learning to generate data from unknown data distributions. A first network (called the generator network) generates data samples given some input (e.g., random noise, conditional class labels, etc.). A second network (called the discriminator network) attempts to classify the data generated by the generator from real data points from a real data distribution.

[0194] As part of the training, the generator network continuously deceives (or "fools" or "fools") the discriminator to reclassify the generated data as "real". As the success rate of deceiving the discriminator network using the generated data increases, the training output of the generator network becomes increasingly realistic. In some examples of the generative 2D modeling techniques disclosed herein, the neural network engine 184 uses a conditional GAN ​​(cGAN), where the generator network is conditional on 2D rendered images of 2D scans of the current dental anatomy of patient 102.

[0195] The generator network (in some non-limiting examples, a CNN) takes a rendered 2D image of the current (pre-restoration) dental anatomy of patient 102 as input and generates a 2D image showing how the proposed (post-restoration) dental anatomy of patient 102 will appear, based on the current state of the generator network's adversarial training. The generator network may also accept additional information as input data (depending on availability and / or relevance), such as which teeth will be restored, the restoration style (e.g., youth, older, natural, oval, etc.).

[0196] The discriminator network (which may also be a CNN in some examples) receives a pair of 2D images as input. The first pair of images contains a rendered 2D image of the patient's dental anatomy before restoration and a rendered 2D image of the actual restoration performed by a clinician on the same patient (which is classified as either "real" or "benchmark ground truth" pairings). The second pair of images contains the rendered 2D image before restoration and the restoration generated by the generator network. The generator network and discriminator network are trained simultaneously in an alternating manner, thereby improving each other to achieve the common goal of accurately training the generator network.

[0197] In some implementations, the neural network engine 184 implements the generator network for 2D inpainting of images disclosed herein as an encoder-decoder CNN. In these examples, the generator network reduces the dimensionality of the input image and then restores the original dimensionality (e.g., via a sequence of downsampling and upsampling, or otherwise). The generator network in these examples may also be referred to as a "U-Net". As used herein, "U-Net" refers to a class of encoder-decoder architectures in which feature maps from the encoder are concatenated to corresponding feature maps in the decoder.

[0198] In traditional GANs, the discriminator network receives either a real image (from the input dataset of images) or a synthetic image (generated by a generator). The output of the discriminator network is a probability in the range [0,1], representing the probability that the input image is a real image (from the dataset). In some specific implementations of the 2D inpainting aspect of this disclosure, the discriminator network is a "patchGAN" discriminator network.

[0199] While a typical discriminator network outputs a single value representing the perceptual realism of the input, the patchGAN discriminator network outputs an [n×n] matrix, where each element represents the perceptual realism of the corresponding patch of the input. The perceptual realism represented by each element in the [n×n] output matrix indicates the probability that the corresponding patch of the input image is part of a real image or a benchmark ground truth image. Internally, the discriminator network is implemented as a CNN.

[0200] Figure 10This is a conceptual diagram illustrating the co-training process of a generator network and a discriminator network of a cGAN according to aspects of this disclosure, the cGAN being configured to render a 2D image of a proposed dental anatomy of a patient 102. Figure 10 The document also illustrates how generator networks and discriminator networks are used to process various types of data. Figure 10 In the diagram, "G" represents the generator network of cGAN, and "D" represents the discriminator network of cGAN. The pre-restoration 2D image (which is the input to G and pairs half of the images supplied by G to D) is represented by "x". "G(x)" represents the proposed restoration 2D image generated by G given x as the pre-restoration input. The 2D rendered image of the actual performed dental restoration (or the "true" or "benchmark" image of the dental anatomy after restoration) is represented by "y".

[0201] Figure 10 The specific use case scenario illustrated is associated with unsuccessful iterations during the multi-iteration training process of G. For example... Figure 10 As shown, the combination of D's output G(x) and x is a "false" decision. Conversely, and as expected regarding adversarial cGAN training, D outputs a "true" decision when evaluating the input combination of x and y. In some examples, if D is a fully trained and refined network, then after G is more precisely adversarially trained via cGAN in future iterations, G can generate instances of G(x) that successfully fool D into outputting a "true" decision when fed with x.

[0202] In these examples, after G reaches this training level, neural network 184 can begin to execute G to generate a proposed 2D image of the post-restoration dental anatomy of patient 102 from input x. In some examples, since G and D are trained sequentially, the two networks may not be trained for similar time periods. In these cases, both G and D may undergo training until G passes a qualitative check, such as a visual check performed by dentist 106 or another clinician. In the format used above, “x” represents the 2D pre-restoration image, “y” is the baseline ground truth 2D restored image, and G(x) is the image generated by G given the pre-restoration image input. The total loss term used in some examples is a combination of L1 loss and GAN loss, given by the following formula (1):

[0203]

[0204] The L1 loss is the absolute value of the difference between y and g(x), where the total loss applies to g but not to D. The calculation of the L1 loss is given by the following formula (2), where λ L1 In this particular example, it is 10, but it should be understood that λ L1Other values ​​may be present in other examples consistent with this disclosure.

[0205] L1 loss =λ L1 *abs(yG(x)) ...(2)

[0206] By utilizing a communication connection with computing device 150 via network 114, computing device 190 can provide a chairside application that enables dentist 106 to display one or more 2D renderings of proposed restorative plans to patient 106. This typically occurs during the same patient visit, during which the current dental anatomy is acquired (and sometimes displayed shortly or immediately after the scan is acquired). Instead of displaying past cases of other patients or general models covering hypothetical patients, computing device 190 can use cGANs executed by neural network engine 184 to output customized renderings of proposed dental anatomy for one or more post-restorative scenes, specifically tailored to patient 102's current dental anatomy and treatment plan.

[0207] In other words, the neural network engine 184 enables the generative modeling techniques of this disclosure, allowing the dentist 106 to utilize cloud computing interaction to render 2D images of proposed dental anatomy for one or more dental restoration plans, specifically tailored to the current dental anatomy of the patient 102. From the perspective of the clinic 104, given a scan of the patient 102's current dental anatomy, the computing device 190 processes the scan rapidly (or almost immediately) and utilizes cloud computing capabilities to render 2D images of one or more processed dental anatomy images specific to the dentition and available processing options available to the patient 102.

[0208] In this way, the neural network engine 184 can implement the generative modeling techniques of this disclosure entirely in the image domain, without the need for potentially time-consuming 3D mesh generation. Upon approval (e.g., by patient 102 and / or dentist 106), the computing device 190 can transmit the generated 2D image to the computing device 192 via network 114, enabling the manufacturing system 194 to generate a 3D mesh of the dental appliance 112, or directly manufacture the dental appliance 112.

[0209] In some examples, the neural network engine 184 can generate or regenerate a 2D image of the proposed dental restoration to incorporate patient-specified modifications, such as restoration style selection or other parameters. In one such example, the neural network engine 184 can implement a feedback loop within the cGAN to accommodate modifications provided by the patient or physician regarding restoration styles, tooth shaping, etc.

[0210] In one example, a trained generator network of cGANs enables technicians to create 3D meshes from 2D images output by the trained generator network. In another example, 3D meshes can be automatically generated from 2D images of proposed dental anatomy output by a trained generator network. In one or more of these examples, the 3D mesh can be used as described above regarding... Figure 3 and Figure 4 The system described is used as an input. In some examples, a dentist 106 or other clinician at clinic 104 may use image capture hardware (e.g., a still camera or video camera) to obtain photographs of the current dental anatomy of patient 102. In these examples, computing device 190 may use the captured photographs to generate a rendering of the current dental anatomy of patient 102.

[0211] Therefore, according to various examples of this disclosure, computing devices 190 and 150 can obtain 2D images (whether dental scans or photographs) of a 3D object (in this case, the dentition of patient 102), and use these 2D images to generate another 2D image of the proposed dental anatomy (or a portion thereof) of the proposed dental prosthetic treatment for patient 102. In this way, by using the neural network training mechanism of this disclosure, computing devices 190 and 150 can achieve dental prosthetic modeling in a computationally inexpensive and fast manner, while maintaining the accuracy of the data regarding dental prosthetic modeling.

[0212] In some examples, the neural network engine 184 executes a trained version of the generator network G as a representation of... Figure 3 and Figure 4 The neural network input generation system is shown. For example, neural network engine 184 can use a 2D image of the proposed dental anatomy of patient 102 to amplify the input of the neural network. Figure 3 The transformation matrix input to the neural network and / or the pair Figure 4 The generator network takes the tooth mesh as input. In these cases, considering the larger proportion of the impact on the patient's 102 dental arches used to generate placement and / or geometric information, the neural network engine 184 utilizes the output of a trained version of the generator network G to train in a more holistic manner. Figure 3 Neural networks and / or Figure 4 A generator for graphical CNNs.

[0213] Figure 11A The inputs and outputs of a generator network trained with a cGAN are shown. This generator network is configured to generate a 2D image of a proposed dental anatomy using a 2D rendering of the current dental anatomy of patient 102. Current dental anatomy image 1102 shows a 2D rendering of the current dental anatomy of patient 102. After training with the cGAN (e.g., by successfully deceiving the discriminator network at least a threshold number of times), Figure 10 The generator network (“G”) uses the current dental anatomy image 1102 to generate a proposed dental anatomy image 1104.

[0214] The current dental anatomy image 1102 is a 2D rendering of the anterior row of teeth for restoration in patient 102. The proposed dental anatomy image 1104 is a 2D rendering of a predicted final outcome of a proposed dental restoration treatment plan for patient 102. Thus, Figure 11 illustrates a working example of a use case scenario where the generative modeling techniques of this disclosure are performed iteratively by training the cGAN of the generator network G.

[0215] Figure 11B A comparison is shown between the current dental anatomy image 1102, the proposed dental anatomy image 1104, and the baseline true value repaired image 1106.

[0216] Figure 12 A menu that can be displayed on the computing device 190 is shown as part of a graphical user interface (GUI) that includes the current dental anatomy image 1102 and / or a proposed dental anatomy image 1104. The data menu 1202 presents various options for the dentist 106 or other clinician to manipulate the content generated in the modeling. Figure 12 In the example, data menu 1202 presents options for alternative test cases that can be used to construct a dental restoration plan. Data menu 1202 also includes tooth options that allow dentist 106 to select specific teeth from the current dental anatomy image 1102 to be modeled for reconstruction.

[0217] The visual options menu 1204 allows the dentist 106 to adjust various viewing parameters based on the display of the current dental anatomy image 1102 and / or the proposed dental anatomy image 1104. The dentist 106 can adjust various viewing parameters through the visual options menu 1204 to allow the patient 102 to better see the details of the proposed dental anatomy image 1204.

[0218] In this way, dentist 106 or other clinicians can operate computing device 190 at clinic 104 to achieve cloud interaction via network 114, thereby utilizing the neural network-based generative modeling capabilities provided by computing device 150. By operating data menu 1202, dentist 106 can provide restorative modeling parameters used by neural network engine 184 in generating proposed dental anatomy images 1204. By operating visual options menu 1204, dentist 106 uses computing device 190 as a live display, which customizes the viewing parameters of the proposed dental anatomy images 1204 to suit the viewing needs and preferences of patient 102.

[0219] Figure 13A and Figure 13B This is a conceptual diagram illustrating an exemplary mold parting surface according to various aspects of this disclosure. A neural network engine 184 can generate the mold parting surface 1302 based on landmarks such as the midpoint of each tooth and points between adjacent teeth in each slice (e.g., contact points between adjacent teeth and / or nearest neighbor points between adjacent teeth). In some examples, as part of the neural network-based placement generation technique of this disclosure, the neural network engine can generate a 3D mesh of the mold parting surface 1302. Further details regarding how the mold parting surface 1302 can be used in appliance manufacturing are described in WO 2020 / 240351, filed May 20, 2020.

[0220] Figure 14 This is a conceptual diagram illustrating an exemplary gingival trimming surface according to various aspects of this disclosure. The gingival trimming surface 1402 may include a 3D mesh that represents an enclosing shell between the trimmed gingiva and teeth in a dental anatomy.

[0221] Figure 15 This is a conceptual diagram illustrating an exemplary facet band according to various aspects of this disclosure. Facet band 1502 is a rigid rib with nominal thickness offset from the housing along the face. In some instances, the facet band follows both the arch and the gingival margin. In one instance, the bottom of the facet band does not fall further towards the gingiva than the gingival trimming surface.

[0222] Figure 16 A conceptual diagram illustrating an exemplary tongue-side shelf 1602 according to various aspects of this disclosure is provided. The tongue-side shelf 1602 is a rigid rib with nominal thickness on the tongue side of a mold appliance, which is embedded in the tongue side and follows an arcuate body.

[0223] Figure 17 This is a conceptual diagram illustrating exemplary incisors and windows according to various aspects of this disclosure. Windows 1704A-1704H (collectively referred to as window 1704) include holes providing access to a tooth surface, allowing a dental composite material to be placed on the tooth. An incisor includes a structure covering the window. The shape of the window can be defined as the nominal shape that inserts from the periphery of the tooth when viewed from the face. In some instances, the shape of the incisor corresponds to the shape of the window. The incisor can be inserted to form a gap between the incisor and the window.

[0224] Figure 18This is a conceptual diagram illustrating exemplary rear snap clamps according to various aspects of this disclosure. A neural network 184 can determine one or more characteristics (e.g., placement-related or geometry-related characteristics) of rear snap clamps 1802A and 1802B (collectively, "rear snap clamp 1802"). Rear snap clamp 1802 can be configured to couple a face-side portion of dental appliance 112 to a tongue-side portion of dental appliance 112. Exemplary characteristics include one or more of the size, shape, position, or orientation of rear snap clamp 1802. Position information of rear snap clamp 1802 may be along an arcuate body at opposite ends of the arcuate body (e.g., a first snap clamp is located at one end and a second snap clamp is located at the other end). In some examples, a concave portion of rear snap clamp 1802 may be positioned on the tongue-side of the parting surface, and a convex portion of rear snap clamp 1802 may be positioned on the face-side.

[0225] Figure 19 This is a conceptual diagram illustrating exemplary dental incisor hinges according to various aspects of the present disclosure. According to various aspects of the present disclosure, as part of generating placement and / or geometry information for dental appliance 112, neural network engine 184 may determine one or more characteristics of dental incisor hinges 1902A-1902F (collectively referred to as dental incisor hinges 1902). Dental incisor hinges 1902 may be configured to pivotally couple dental incisors to dental appliance 112. Exemplary characteristics include one or more of the size, shape, position, or orientation of the respective dental incisor hinge 1902. In some non-limiting use cases, the neural network executed by neural network engine 184 may locate the dental incisor hinges 1902 based on the position of another predefined appliance feature. For example, the neural network may locate each dental incisor hinge 1902 at the midline of the corresponding dental incisor. In one usage scenario, the concave portion of the corresponding dental hinge 1902 can be positioned to anchor to the facet portion of the dental appliance 112 (e.g., toward the incisal edge of the corresponding tooth), and the convex portion of the same dental hinge 1902 can be positioned to anchor to the outside of the dental incision.

[0226] Figure 20A and Figure 20BThis is a conceptual diagram illustrating exemplary incisor clips according to various aspects of this disclosure. A neural network engine 184 can determine one or more characteristics of the incisor clips 2002A-2002F (collectively, “incisor clips 2002”), such as placement characteristics and / or geometric characteristics. Exemplary characteristics include one or more of the size, shape, position, or orientation of the incisor clips 2002. In some examples, a neural network executed by the neural network engine 184 can determine the position of the incisor clips 2002 based on the position of another predefined appliance feature. In one example, the neural network can generate a placement profile that positions each incisor clip 2002 at the midline of the corresponding incisor. In one example, the position of a concave portion of a particular incisor clip 2002 can be anchored to the outside of the incisor and extends downward toward the gum line. In another example, a convex portion of a particular incisor clip 2002 can be anchored to the gum line of the facet band.

[0227] Figure 21 This is a conceptual diagram illustrating an exemplary ridge according to various aspects of this disclosure. Ridge 2102 provides reinforcement at the cutting edge.

[0228] Figure 22 This is a conceptual diagram illustrating an exemplary central clamp according to various aspects of this disclosure. The central clamp 2202 aligns the facet and lingual portions of a dental instrument with each other.

[0229] Figure 23 This is a conceptual diagram illustrating exemplary incisors according to various aspects of this disclosure. Incisors 2302A-2302B (collectively referred to as incisors 2302) deliver excess dental composite material out of a dental appliance.

[0230] Figure 24 This is a conceptual diagram illustrating an exemplary dental incision according to various aspects of this disclosure. Figure 24 In the example, the dental appliance includes a dental incisor 2402, a dental incisor hinge 2404, and a dental incisor clip 2406.

[0231] Figure 25 This is a conceptual diagram illustrating an exemplary interdental matrix according to various aspects of the present disclosure. The interdental matrix 2502 includes a handle 2504, a body 2506, and a wrapping portion 2508. The wrapping portion 2508 is configured to fit into the interproximal region between two adjacent teeth.

[0232] Figure 26This is a conceptual diagram illustrating an exemplary manufacturing housing frame and an exemplary dental appliance according to various aspects of this disclosure. The manufacturing housing frame 2602 is configured to support one or more portions of the dental appliance. For example, the manufacturing housing frame 2602 may detachably couple to the lingual portion 2604, the facial portion 2606, and the interdental matrix 2608 of the dental appliance via housing frame struts 2610. Figure 26 In the example, the housing frame strut 2610 attaches or couples the dental appliance parts 2604, 2606 and 2608 to the manufacturing housing frame 2602.

[0233] Figure 27 This is a conceptual diagram illustrating an exemplary dental appliance including custom labels according to various aspects of this disclosure. Custom labels 2702-2708 can be printed on various parts of the dental appliance and include data identifying the respective parts of the dental appliance (e.g., serial number, part number, etc.).

[0234] Various examples have been described. These examples, as well as others, are all within the scope of the following claims.

Claims

1. A computing device, the computing device comprising: An input interface is configured to receive one or more three-dimensional (3D) tooth meshes associated with the current dental anatomy of a dental restoration patient and a 3D part mesh representing the geometry for the generation of dental appliance components; and Neural network engine, the neural network engine being configured as follows: The one or more 3D tooth meshes and the 3D component meshes received by the input interface are provided as input to a neural network trained with training data, which includes the baseline ground value of dental appliance component geometry and 3D tooth mesh of the corresponding dental restoration case. as well as The neural network is executed using the provided inputs to generate an updated model of the dental appliance components relative to the current dental anatomy of the dental prosthesis patient.

2. The computing device of claim 1, wherein the neural network is a generative adversarial network (GAN) comprising a generator network and a discriminator network. Furthermore, in order to execute the neural network to generate an updated model of the dental appliance component relative to the current dental anatomy of the dental prosthetic patient, the neural network engine is configured to provide the generator network of the GAN with the one or more 3D tooth meshes and the 3D component meshes received by the input interface as input.

3. The computing device of claim 2, wherein the neural network engine is configured to train the GAN by: The generator network of the GAN is executed to generate an updated geometry of the dental appliance component using the one or more 3D tooth meshes and the 3D component meshes; The updated geometry of the dental appliance components, the one or more 3D tooth meshes, and the 3D component meshes are provided as input to the discriminator network of the GAN; as well as The discriminator network is executed to output the probability that the updated geometry represents the baseline truth geometry.

4. The computing device according to any one of claims 1 to 3, wherein the neural network is further trained using placement information of the baseline true value dental appliance component geometry, which is part of the training data.

5. The computing device according to any one of claims 1 to 3, wherein the dental instrument component comprises one or more of the following: Mold parting surface, Gingival trimming. Teeth, Tooth window, Side band, Cut the ridge, Tongue-side shelf, Interdental matrix, Shell frame, Parts label; Tooth hinge, Tooth buckle, Dental vent, Clip clamps, or Center clamp.

6. A method, the method comprising: The input interface receives one or more three-dimensional (3D) tooth meshes associated with the current dental anatomy of the dental restoration patient and a 3D part mesh representing the geometry used to generate dental appliance parts. The neural network engine, which is communicatively coupled to the input interface, provides the one or more 3D tooth meshes and the 3D component meshes received by the input interface as input to the neural network trained with training data, which includes the baseline ground truth dental appliance component geometry and 3D tooth mesh of the corresponding dental restoration case. as well as The neural network engine uses the provided input to execute the neural network to generate an updated model of the dental appliance components relative to the current dental anatomy of the dental prosthetic patient.

7. The method of claim 6, wherein the neural network is a generative adversarial network (GAN) comprising a generator network and a discriminator network, and wherein the model executing the neural network to generate an update of the dental appliance component relative to the current dental anatomy of the dental prosthetic patient comprises the generator network of the GAN being provided by the neural network engine with the one or more 3D tooth meshes and the 3D component meshes received from the input interface as input.

8. The method of claim 7, further comprising training the GAN by the neural network engine through the following operations: The generator network of the GAN is executed to generate an updated geometry of the dental appliance component using the one or more 3D tooth meshes and the 3D component meshes; The updated geometry of the dental appliance components, the one or more 3D tooth meshes, and the 3D component meshes are provided as input to the discriminator network of the GAN; as well as The discriminator network is executed to output the probability that the updated geometry represents the baseline truth geometry.

9. The method according to any one of claims 6 to 8, wherein the neural network is further trained using placement information of the baseline true value dental appliance component geometry, which is part of the training data.

10. The method according to any one of claims 6 to 8, wherein the dental appliance component comprises one or more of the following: Mold parting surface, Gingival trimming. Teeth, Tooth window, Side band, Cut the ridge, Tongue-side shelf, Interdental matrix, Shell frame, Parts label; Tooth hinge, Tooth buckle, Dental vent, Clip clamps, or Center clamp.

11. An apparatus, the apparatus comprising: Device for receiving one or more three-dimensional (3D) tooth meshes associated with the current dental anatomy of a dental prosthetic patient and 3D part meshes representing the geometry for the generation of dental appliance components; A device for providing the received one or more 3D tooth meshes and the 3D component meshes as input to a neural network trained with training data, the training data including baseline ground truth dental appliance component geometry and 3D tooth meshes of corresponding dental restoration cases. as well as An apparatus for using the provided input to execute the neural network to generate an updated model of the dental appliance components relative to the current dental anatomy of the dental prosthesis patient.

12. A non-transitory computer-readable storage medium encoded with instructions, which, when executed, cause one or more processors of a computing system to: Receive one or more three-dimensional (3D) tooth meshes associated with the current dental anatomy of the dental restoration patient and a 3D part mesh representing the geometry used to generate dental appliance components; The received one or more 3D tooth meshes and the 3D component meshes are provided as input to a neural network trained with training data, which includes the baseline ground truth dental appliance component geometry and 3D tooth mesh of the corresponding dental restoration case. as well as The neural network is executed using the provided inputs to generate an updated model of the dental appliance components relative to the current dental anatomy of the dental prosthesis patient.

Images