Prediction of shapes of anatomic structures using a generative pretrained transformer

Abstract

A computer-implemented method comprises obtaining an input anatomic model that comprises a first 3-dimensional mesh representing a surface of a first portion of an anatomic structure of a patient; generating input tokens based on the input anatomic model; applying a generative pre-trained transformer (GPT) to the input tokens to generate a sequence of output tokens; and generating, based on the output tokens, a predicted anatomic model that includes a second 3D mesh representing a surface of a second portion of the anatomic structure.

Description

[0001]1262-252WO01 / DIG-24-1695-IDF PREDICTION OF SHAPES OF ANATOMIC STRUCTURES USING A GENERATIVE PRETRAINED TRANSFORMER BACKGROUND[ ] This application claims priority to U.S. Provisional Patent Application63 / 664,055, filed June 25, 2024, the entire content of which is incorporated by reference. BACKGROUND[ ] Orthopedic surgeries often involve implanting one or more orthopedic prosthesesinto a patient. For example, in a total shoulder replacement surgery, a surgeon may attach orthopedic prostheses to a scapula and a humerus of a patient. In an ankle replacement surgery, a surgeon may attach orthopedic prostheses to a tibia and a talus of a patient. Placing a prosthesis at a correct location may be critical to achieving a successful outcome for the patient. Placing the prosthesis at an incorrect location may limit the patient’s range of motion, may lead to premature failure of the prosthesis, or other adverse outcomes. Selecting an inappropriately sized, shaped, or anchored prosthesis may likewise lead to adverse outcomes. Accordingly, a surgeon may need to carefully select the prosthesis and plan where to place the prosthesis prior to the surgery. SUMMARY[ ] This disclosure describes example techniques of predicting shapes of anatomicstructures. A user may use a computer-assisted orthopedic system (CAOS) as part of planning an orthopedic surgery. In accordance with the techniques of this disclosure, the CAOS may obtain an input anatomic model that comprises a first 3-dimensional mesh representing a surface of at least a first portion of an anatomic structure of a patient. The CAOS may generate input tokens based on the input anatomic model. Furthermore, the CAOS may apply a generative pre-trained transformer (GPT) to the input tokens to generate a sequence of output tokens. The CAOS may generate, based on the output tokens, a predicted anatomic model that includes a second 3D mesh representing a surface of at least a second portion of the anatomic structure. The CAOS may output the predicted anatomic model for display.[ ] In some examples, the user of the CAOS may want to understand the shape of apatient’s bone prior to onset of a morbidity, such as arthritis or trauma. By understanding the shape of the patient’s bone prior to the onset of the morbidity, the user may be better 1262-252WO01 / DIG-24-1695-IDF able to select and position an orthopedic prosthesis on the bone. In accordance with the techniques of this disclosure, the predicted anatomic model may represent an estimated premorbid state of a patient’s bone. The CAOS may output the premorbid anatomic model for display. In some examples, the CAOS may generate one or more recommendations regarding a prothesis based on the premorbid anatomic model.[ ] In one example, this disclosure describes a computer-implemented methodcomprising: obtaining, by one or more processors, an input anatomic model that comprises a first 3-dimensional (3D) mesh representing a surface of at least a first portion of an anatomic structure of a patient; generating, by the one or more processors, input tokens based on the input anatomic model; applying, by the one or more processors, a generative pre-trained transformer (GPT) to the input tokens to generate output tokens; and generating, by the one or more processors, based on the output tokens, a predicted anatomic model that includes a second 3D mesh representing a surface of at least a second portion of the anatomic structure.[ ] In another example, this disclosure describes a computer-implemented methodcomprising: training, by one or more processors, a machine learning (ML) model to: generate input tokens based on an input anatomic model that comprises a first 3D mesh representing a surface of at least a first portion of an anatomic structure of a patient; apply a generative pre-trained transformer (GPT) of the ML model to the input tokens to generate output tokens; and generate, based on the output tokens, a predicted anatomic model that includes a second 3D mesh representing a surface of at least a second portion of the anatomic structure.[ ] The details of various examples of the disclosure are set forth in the accompanyingdrawings and the description below. Various features, objects, and advantages will be apparent from the description, drawings, and claims. BRIEF DESCRIPTION OF DRAWINGS[ ] FIG. 1 is a block diagram illustrating an example system that may be used toimplement the techniques of this disclosure.[ ] FIG. 2 is a block diagram illustrating example components of a computer assistedorthopedic system, in accordance with one or more techniques of this disclosure.[ ] FIG. 3A is a conceptual diagram illustrating an example input anatomic model, inaccordance with one or more techniques of this disclosure. 1262-252WO01 / DIG-24-1695-IDF[ ] FIG. 3B is a conceptual diagram illustrating an example predicted anatomicmodel, in accordance with one or more techniques of this disclosure.[ ] FIG. 4 is a flowchart illustrating an example operation of a computer-assistedorthopedic system, in accordance with one or more techniques of this disclosure.[ ] FIG. 5 is a block diagram illustrating an example architecture for generatingquantized face embeddings and generating a reconstructed anatomic model, in accordance with one or more techniques of this disclosure.[ ] FIG. 6 is a flowchart illustrating an example operation of a prediction unit togenerate a predicted anatomic model, in accordance with one or more techniques of this disclosure.[ ] FIG. 7 is a flowchart illustrating an example training process for a machinelearning model, in accordance with one or more techniques of this disclosure. DETAILED DESCRIPTION[ ] Orthopedic surgeries often involve implanting one or more orthopedic prosthesesinto a patient. For example, in a total shoulder arthroplasty (TSA), a surgeon may attach orthopedic prostheses to a scapula and a humerus of a patient. In a total ankle replacement (TAR), a surgeon may attach orthopedic prostheses to a tibia and a talus of a patient. In a total knee arthroplasty (TKA) surgery, a surgeon may attach orthopedic prostheses to a femur and tibia of a patient.[ ] One goal of an orthopedic surgery, such as a TSA, TAR, or TKA, is to restore thegeometry of the patient’s joint so that the joint functions in a manner similar to how the joint functioned prior to the onset of a pathology. For example, if a bone of a patient’s bone is eroded due to arthritis, a goal of the orthopedic surgery may be to restore the joint’s ability to move in the same way as before the onset of the arthritis.[ ] Accordingly, when planning an orthopedic surgery, a surgeon may wish to placethe orthopedic prostheses at positions that restore the joint’s ability to move in the same way as before the onset of the pathology. However, it may be difficult for the surgeon to understand the premorbid shape of the patient’s anatomy (e.g., bone or soft tissue), and thereby determine how to restore the joint’s ability to move in the same way as before the onset of the pathology. Accordingly, various techniques have been developed that attempt to reconstruct the premorbid shapes of bones. However, such techniques may suffer from various shortcomings. For example, in some existing techniques, generated premorbid 1262-252WO01 / DIG-24-1695-IDF anatomic models introduce changes to nonpathological portions of the bones. Furthermore, some techniques, such as those involving point clouds, may involve complex and potentially inaccurate post-prediction processes to form the point clouds into 3-dimensional mesh representations of the surface of the premorbid anatomic model. In addition, conventional procedures for generating premorbid anatomic models involve a human, such as a surgeon, looking at medical images of the patient’s current, pathological anatomy, and relying on his or her judgment, manipulate a morbid 3D model of the patient’s anatomy until the 3D model has a plausible premorbid shape. This is a visual and subjective process.[ ] The techniques of this disclosure may address one or more of these shortcomings.As described in this disclosure, a computing system obtains an input anatomic model that comprises a first 3-dimensional mesh representing a surface of at least a first portion of an anatomic structure of a patient. The computing system generates input tokens based on the input anatomic model. The computing system applies a generative pre-trained transformer (GPT) to the input tokens to generate output tokens. The computing system then generates, based on the output tokens, a predicted anatomic model that includes a second 3D mesh representing a surface of at least a second portion of the anatomic structure.[ ] Because the GPT directly generates a 3D mesh, the techniques of this disclosuremay avoid the issues involved in forming point clouds into 3D meshes. Furthermore, the GPT does not rely directly on statistical shape modeling. Additionally, at least in some examples, the GPT does not need to recreate all portions of the anatomic structure, such as faces of a mesh outside a pathological area. This may increase the accuracy of the predicted anatomic models. As described in this disclosure, the predicted anatomic models may be premorbid bone models. The techniques of this disclosure may also be used for other purposes, such as generating completed models of anatomic structures from incomplete models of the anatomic structures. Furthermore, the techniques of this disclosure may avoid the need for the registration process (either rigid or modes-based) that is required in statistical shape model (SSM) processes. Avoiding the registration process may reduce errors attributable to the registration process. Additionally, application of the techniques of this disclosure avoids the need for subjective human determinations of premorbid shapes.[ ] FIG. 1 is a block diagram illustrating an example system 100 that may be used toimplement the techniques of this disclosure. FIG. 1 illustrates computing system 102, 1262-252WO01 / DIG-24-1695-IDF which is an example of one or more computing devices that are configured to perform one or more example techniques described in this disclosure. Computing system 102 may include various types of computing devices, such as server computers, personal computers, smartphones, laptop computers, and other types of computing devices. In some examples, computing system 102 includes multiple computing devices that communicate with each other. In other examples, computing system 102 includes only a single computing device. Computing system 102 includes processing circuitry 104, storage system 106, a display 108, and a communication interface 110. Display 108 is optional, such as in examples where computing system 102 is a server computer.[ ] Examples of processing circuitry 104 include one or more microprocessors, digitalsignal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), discrete logic, software, hardware, firmware or any combinations thereof. In general, processing circuitry 104 may be implemented as fixed- function circuits, programmable circuits, or a combination thereof. Fixed-function circuits refer to circuits that provide particular functionality and are preset on the operations that can be performed. Programmable circuits refer to circuits that can be programmed to perform various tasks and provide flexible functionality in the operations that can be performed. For instance, programmable circuits may execute software or firmware that cause the programmable circuits to operate in the manner defined by instructions of the software or firmware. Fixed-function circuits may execute software instructions (e.g., to receive parameters or output parameters), but the types of operations that the fixed-function circuits perform are generally immutable. In some examples, the one or more of the units may be distinct circuit blocks (fixed-function or programmable), and in some examples, the one or more units may be integrated circuits. In some examples, processing circuitry 104 is dispersed among a plurality of computing devices in computing system 102. In some examples, processing circuitry 104 is contained within a single computing device of computing system 102.[ ] Processing circuitry 104 may include arithmetic logic units (ALUs), elementaryfunction units (EFUs), digital circuits, analog circuits, and / or programmable cores, formed from programmable circuits. In examples where the operations of processing circuitry 104 are performed using software executed by the programmable circuits, storage system 106 may store the object code of the software that processing circuitry 104 receives and executes, or another memory within processing circuitry 104 (not shown) 1262-252WO01 / DIG-24-1695-IDF may store such instructions. Examples of the software include software designed for surgical planning, including image segmentation.[ ] Storage system 106 may be formed by any of a variety of memory devices, suchas dynamic random access memory (DRAM), including synchronous DRAM (SDRAM), magnetoresistive RAM (MRAM), resistive RAM (RRAM), or other types of memory devices. Examples of display 108 include a liquid crystal display (LCD), a plasma display, an organic light emitting diode (OLED) display, or another type of display device. In some examples, storage system 106 may include multiple separate memory devices, such as multiple disk drives, memory modules, etc., that may be dispersed among multiple computing devices or contained within the same computing device.[ ] Communication interface 110 allows computing system 102 to communicate withother devices via network 112. For example, computing system 102 may output medical images and other information for display. Communication interface 110 may include hardware circuitry that enables computing system 102 to communicate (e.g., wirelessly or using wires) to other computing systems and devices, such as a visualization device 114 and an imaging system 116. Network 112 may include various types of communication networks including one or more wide-area networks, such as the Internet, local area networks, and so on. In some examples, network 112 may include wired and / or wireless communication links.[ ] Visualization device 114 may utilize various visualization techniques to displayimage content to a surgeon. In some examples, visualization device 114 is a computing device having computer monitor or display screen. For instance, visualization device 114 may be a computing device that a surgeon uses for planning surgeries. In some examples, visualization device 114 may be a mixed reality (MR) visualization device, virtual reality (VR) visualization device, holographic projector, or other device for presenting extended reality (XR) visualizations. For instance, in some examples, visualization device 114 may be a Microsoft HOLOLENS™ headset, available from Microsoft Corporation, of Redmond, Washington, USA, or a similar device, such as, for example, a similar MR visualization device that includes waveguides. The HOLOLENS ™ device can be used to present 3D virtual objects via holographic lenses, or waveguides, while permitting a user to view actual objects in a real-world scene, i.e., in a real-world environment, through the holographic lenses. In some examples, visualization 114 may have an opaque screen on which both actual objects in the real world and virtual objects are visible. In this 1262-252WO01 / DIG-24-1695-IDF disclosure, the term mixed reality is intended to encompass the term augmented reality. In some examples, there may be multiple visualization devices for multiple users.[ ] Imaging system 116 may comprise one or more devices configured to generatemedical image data. For example, imaging system 116 may include a device for generating CT images. In some examples, imaging system 116 may include a device for generating magnetic resonance imaging (MRI) images. Furthermore, in some examples, imaging system 116 may include one or more computing devices configured to process data from imaging devices in order to generate medical image data. For example, the medical image data may include a 3D image of one or more bones of a patient. In this example, imaging system 116 may include one or more computing devices configured to generate the 3D image based on CT images or MRI images. In other examples, computing system 102 may include one or more computing devices configured to generate the medical image data based on data from devices in imaging system 116.[ ] Storage system 106 of computing system 102 may store instructions that, whenexecuted by processing circuitry 104, cause computing system 102 to perform various activities. For instance, in the example of FIG. 1, storage system 106 may store instructions that, when executed by processing circuitry 104, cause computing system 102 to perform activities associated with a computer-assisted orthopedic system (CAOS) 118. For ease of explanation, rather than discussing computing system 102 performing activities when processing circuitry 104 executes instructions, this disclosure may simply refer to CAOS 118 or components thereof as performing the activities, or may directly describe computing system 102 as performing the activities.[ ] CAOS 118 is a computerized system that assists users with orthopedic surgeries.For example, CAOS 118 may assist users in preoperative planning of an orthopedic surgery. In some examples, CAOS 118 may assist users during an orthopedic surgery. The Stryker BLUEPRINT ™ system is an example of CAOS 118. A user of CAOS 118 may select, design or modify appropriate implant components, determine how best to position and orient the implant components and how to shape the surface of the bone to receive the components, and design, select or modify surgical guide tool(s) or instruments to carry out the surgical plan. The information generated by CAOS 118 may be compiled in a preoperative surgical plan for the patient that is stored in a database at an appropriate location, such as storage system 106, where the preoperative surgical plan can be accessed by a user, including before and during the actual surgery. 1262-252WO01 / DIG-24-1695-IDF[ ] In one or more examples, CAOS 118 may execute completely within a devicewith which a user interfaces. In one or more examples, CAOS 118 may execute completely in a cloud-based network. That is, in such examples, a device with which a surgeon interfaces may upload information to the cloud that CAOS 118 receives. CAOS 118 receives such information, performs example operations, and outputs the results to the device with which the surgeon interfaces. In one or more examples, CAOS 118 may be a combination of cloud-based and executing on the device with which the user interfaces. The example techniques should not be considered limited to a specific configuration.[ ] Additionally, in the example of FIG. 1, storage system 106 stores surgical plans120. Surgical plans 120 may correspond to individual patients. A surgical plan corresponding to a patient may include data associated with a planned or completed orthopedic surgery on the corresponding patient. In the example of FIG.1, a surgical plan corresponding to a patient may include medical image data 126 for the patient, a source anatomic model 128, a predicted anatomic model 132, and other information for the patient. In other examples, source anatomic model 128 and / or predicted anatomic model 132 are not associated with an individual surgical plan.[ ] Medical image data 126 may include CT images of bones of the patient or 3Dimages of bones of the patient based on CT images. In this disclosure, the term “bone” may refer to a whole bone or a bone fragment. In some examples, medical image data 126 may include MRI images of one or more bones of the patient or 3D images based on MRI images of the one or more bones of the patient.[ ] As part of preoperatively planning an orthopedic surgery, CAOS 118 may presentgraphical user interfaces (GUIs) that enable a user to review a patient’s anatomy, select a surgery type, select and position orthopedic prostheses, and perform other tasks for planning the orthopedic surgery. In some examples, CAOS 118 may output manufacturing instructions for patient-specific instruments, patient-specific prostheses, and / or patient-specific models. In some examples, CAOS 118 may output requisition requests for surgical items, including instruments and prosthetic components.[ ] When planning an orthopedic surgery, CAOS 118 may obtain a source anatomicmodel 128. Source anatomic model 128 may represent a surface of one or more bones or other anatomic structures. For instance, the anatomic structure may be one of a scapula, a humerus, a radius, an ulna, a vertebra, a mandible, a femur, a tibia, a fibula, a talus, a cranium, a rib, a pelvis, a bone of a hand or foot, or other bone or soft tissue structure, or 1262-252WO01 / DIG-24-1695-IDF combination of two or more thereof. In some examples, the anatomic structure is a joint and a portion of the anatomic structure included in source anatomic model 128 includes portions of two or more bones of the joint. Generating a predicted anatomic model of two or more bones may be especially helpful in planning joint replacement surgeries because a user may be able to see how the bones interacted prior to onset of a pathology.[ ] In some examples, the source anatomic model represents a morbid bone. A morbidbone is a bone that has one or more pathologies that have changed the shape of the bone from its initial, premorbid shape. Example pathologies include arthritis, trauma, cancer, lesions, and so on. In some examples, source anatomic model 128 represents a first portion of an anatomic structure and not a second portion of the anatomic structure. For instance, image data generated by imaging system 116 might not include all of the bone. Thus, a source anatomic model based on the image data might only represent some of the bone.[ ] Source anatomic model 128 may include a 3D mesh representing a surface of theanatomic structure. A 3D mesh is a collection of vertices, edges, and faces (e.g., polygons) that define a shape of an object. In some examples, each of the faces are defined by the positions of three vertices, resulting in a triangle mesh. In other examples, faces may be defined by other numbers of vertices. Each of the vertices is defined in 3 dimensions. Edges are lines that connect vertices. The 3D mesh may include a large number (e.g., hundreds, thousands, tens of thousands, etc.) of vertices, edges, and faces. Larger numbers of vertices, edges, and faces may increase the accurate of source anatomic model 128. However, the use of such large number of vertices, edges, and faces in a mesh makes performing operations on such a mesh impractical for a human to perform.[ ] CAOS 118 may use ML model 130 to generate a predicted anatomic model 132based on source anatomic model 128. Predicted anatomic model 132 includes a 3D mesh representing a predicted shape of an anatomical structure. Predicted anatomic model 132 may represent a premorbid bone. In some examples where source anatomic model 128 represents a first portion of a bone and not a second portion of the bone, predicted anatomic model 132 may represent the first and second portions of the bone. In some examples where source anatomic model 128 represents at least a portion of a bone, predicted anatomic model 132 represents the portion of the bone and cartilage associated with the bone. In some examples, CAOS 118 may quantify bone loss volume by comparing source anatomic model 128 and predicted anatomic model 132. 1262-252WO01 / DIG-24-1695-IDF[ ] ML model 130 includes a generative pretrained transformer (GPT) 134. MLmodel 130 may have an architecture based on generative pre-trained transformer technology. In some examples, an architecture of GPT 134 is based on an architecture described in Siddiqui et al., “MeshGPT: Generating Triangle Meshes with Decoder-Only Transformers”, arXiv:2311.15475v1 [cs.CV] 27 Nov 2023.[ ] CAOS 118 may process source anatomic model 128 to generate an input anatomicmodel. ML model 130 may generate input tokens based on the input anatomic model. CAOS 118 may provide the input tokens as input to GPT 134. The input tokens represent a least a subset of the faces of the input anatomic model. Thus, this disclosure may refer to the input tokens as face embeddings. In some examples, such as examples where the input anatomic model represents a first portion of a bone and not a second portion of the bone, the input tokens may represent all of the faces of source anatomic model 128. In some examples, such as examples where source anatomic model 128 represents a bone having one or more pathologies, the input tokens may represent fewer than all of the faces of source anatomic model 128. For instance, the input tokens may be limited to those faces representing healthy, nonpathological portions of the bone. GPT 134 may generate output tokens based on the sequence of input tokens. CAOS 118 may generate predicted anatomic model 132 based on the output tokens. Predicted anatomic model 132 may include a second 3D mesh representing a surface of at least a second portion of the anatomic structure.[ ] As mentioned above, CAOS 118 may process source anatomic model 128 togenerate the input anatomic model. In some examples, source anatomic model 128 represents a morbid state of a bone and predicted anatomic model 132 represents a premorbid state of the bone. CAOS 118 may modify source anatomic model 128 to generate the input anatomic model. In this example, source anatomic model 128 includes the non-pathological portion and a pathological portion, and the input anatomic model excludes the pathological portion of the source anatomic model. Excluding the pathological portion of the source anatomic model may have several benefits. For example, excluding the pathological portion of the source anatomic model reduces the number of vertices that a processed as part of applying GPT 134. Reducing the number of vertices reduces the number of memory read and write operations, which may reduce total computation time, reduce electricity usage, and reduce waste heat. Additionally, in some examples, excluding the pathological portion of the source anatomic model may accelerate training. There are wide variances between patients in pathological portions of 1262-252WO01 / DIG-24-1695-IDF bones while the variances between patients are less in non-pathological portions of bones. Thus, it may not be necessary to train GPT 134 on as may training examples when the training examples exclude pathological portions. This may allow GPT 134 to be trained more efficiently, in less time and with less energy consumption.[ ] CAOS 118 may modify source anatomic model 128 in one of a variety of ways.For example, CAOS 118 may determine a position of a geometric primitive shape relative to the source anatomic model. In this example, CAOS 118 may determine retained faces of the source anatomic model and discarded faces of the source anatomic model. The retained faces and discarded faces being on different sides of a boundary defined by the geometric primitive shape. Example types of geometric primitive shapes include spheres, spheres, curved surfaces, polyhedral, and so on. The use of a geometric primitive shape to determine the retained faces of the source anatomic model and the discarded faces of the source anatomic model may provide one or more benefits. For example, it is relatively computationally efficient to determine an intersection of a geometric primitive shape and vertices of a mesh of the source anatomic model. Since the pathological portions of a bone tend to be clustered within a predictable area of the bone, CAOS 118 may easily exclude the pathological portions of the bone by determining the intersection of the geometric primitive shape and the vertices of the mesh of the source anatomic model. Given the ability of GPT 134 to predict vertices of a mesh, more precise ways of excluding the pathological portions may be unnecessary. Thus, the use of the geometric primitive shape may accelerate the generation of predicted anatomic model 132 and may avoid more computationally intensive ways of determining the pathological portions.[ ] In some examples, CAOS 118 may determine positions of one or more anatomicallandmarks on an anatomy of the patient and may determine the position of the geometric primitive shape. For example, CAOS 118 may determine the position of the geometric primitive shape relative to the source anatomic model based on the positions of the one or more anatomical landmarks. In an example where the anatomic structure is a scapula, the anatomical landmarks may include a center of the glenoid fossa, a tip of the coracoid process, points on the rim of the glenoid fossa, and so on. The landmarks may be determined manually or automatically. In some examples, since the input anatomic model is a 3D mesh, the vertices may have attribute that specify the landmarks. This may be computationally efficient since data structures separate from the 3D mesh may be unnecessary for denoting the landmarks. 1262-252WO01 / DIG-24-1695-IDF[ ] Thus, in some examples, the geometric primitive shape is a sphere, the anatomicstructure is a scapula, and as part of determining a position of the geometric primitive shape, CAOS 118 may determine a center of the sphere based on a center of a glenoid fossa of the scapula. After determining the center of the sphere, to determine the position of the geometric primitive shape, CAOS 118 may adjust a position of the sphere medially. In another example, the geometric primitive shape is a plane through source anatomic model 128. For instance, in this example, the anatomic structure may be a scapula and the plane may be a sagittal plane. The use of a sphere or plane may be efficient for determining pathological portions of the scapula because pathological portions of scapulas are typically clustered within a spherical region around the glenoid fossa or are concentrated positioned laterally from a plane, such as a sagittal plane.[ ] In some examples, CAOS 118 may generate the input tokens based on an inputanatomic model that represents the same portions of the anatomic structure as source anatomic model 128. For instance, the input anatomic model may be the same as source anatomic model 128 or simplified version of source anatomic model 128. In this example, CAOS 118 may apply GPT 134 to a limited set of the input tokens. For example, CAOS 118 may apply GPT 134 to a specific percentage of the input tokens (e.g., 80%, 90%, etc.). This solution may provide good results since the meshes’ triangles are reordered given some axis as preprocessing step (e.g., as described in the MeshGPT paper cited above). For instance, CAOS 118 may reorder a mesh representing a scapula to have its first triangle on the lateral side and its last triangle on the medial side. Thus, the mesh’s triangles are ordered along the medio-lateral axis so that if CAOS 118 provides a given percentage (e.g., 90%) of the full mesh tokens as input to GPT 134, it is equivalent as passing the healthy part of the bone by considering that 10% of the tokens represent the pathologic region (e.g., glenoid, tibia mortise, ...) to be completed. Thus, in such examples, CAOS 118 may generate a first set of input tokens based on the input anatomic model and a second set of input tokens based on the input anatomic model. The first set of input tokens may be a specific percentage or quantity of the combined first and second sets of input tokens. CAOS 118 may apply GPT 134 to the first set of input tokens and not the second set of input tokens.[ ] In some examples, CAOS 118 may segment source anatomic model 128 toidentify the pathological portion of source anatomic model 128 and the non-pathological portion of source anatomic model 128. CAOS 118 may modify source anatomic model 128 to exclude the pathological portion from the input anatomic model. For example, 1262-252WO01 / DIG-24-1695-IDF CAOS 118 may transform source anatomic model 128 into a point cloud. CAOS 118 may then apply a point cloud classification model to the point cloud. The point cloud classification model may output attribute values for points in the point cloud. The attribute values for the points in the point cloud indicate whether the point is or is not associated with a pathological portion of the anatomic structure. The point cloud classification model may be implemented using a PointNet architecture, such as that described in Qi et al., “PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation”, arXiv:1612.00593. In this example, CAOS 118 may generate the input anatomic model by modifying source anatomic model 128 to remove faces having vertices corresponding to points classified as being part of the pathological portion of the anatomic structure. In some examples, CAOS 118 may segment source anatomic model 128 using a binary segmentation algorithm. The binary segmentation algorithm may be implemented using a MeshCNN architecture, such as that described in Barda et al., “MeshCNN Fundamentals: Geometric Learning through a Reconstructable Representation”, arXiv:2015.13277v1. In some cases, pathological portions are not easily clustered within a well-defined area such as that encompassed by a geometric primitive shape or use of the geometric primitive shape may exclude too much of the anatomy. In such examples, segmenting source anatomic model 128 in this way may enable GPT 134 to produce more accurate predicted anatomic models, e.g., because GPT 134 would be less confounded with unexcluded pathological areas and may still have sufficient vertices to make accurate predictions.[ ] Modifying source anatomic model 128 to remove pathological areas of theanatomic structure may have various benefits. For example, removing the pathological areas may allow ML model 130 to generate predicted anatomic model 132 with smooth transitions from the current portions of the anatomic structure to premorbid shapes of the anatomic structure. Furthermore, ML model 130 may be significantly simpler because ML model 130 does not need to be trained on training examples that show morbidities, which can vary widely. In other words, training can be faster and simpler. Moreover, training data showing healthy anatomic structures may be easier to obtain than training data showing morbid anatomic structures. Thus, modifying source anatomic model 128 to remove the pathological areas may conserve computing resources and accelerate training of ML model 130.[ ] CAOS 118 may use predicted anatomic model 132 in one or more ways. Forexample, CAOS 118 may output predicted anatomic model 132 for display, e.g., on 1262-252WO01 / DIG-24-1695-IDF display 108 or visualization device 114. In some examples, CAOS 118 may use predicted anatomic model 132 to determine the positions of anatomical landmarks, anatomical axis, or other data characterizing the anatomy of the patient. For example, source anatomic model 128 may represent a distal tibia and not a proximal tibia. In this example, predicted anatomic model 132 may represent both the distal tibia and proximal tibia. Thus, in this example, CAOS 118 may determine anatomical landmarks at the proximal and distal ends of the tibia and calculate a mechanical axis of the tibia based on these anatomical landmarks. CAOS 118 (or a user thereof) may use the mechanical axis of the tibial to select a position of a tibial tray prosthesis during a total ankle replacement surgery. In some examples where visualization device 114 is a MR visualization device, visualization device 114 may render a virtual object based on predicted anatomical model 132 such that the virtual object is superimposed on or in a vicinity of a patient during surgery. This may help the surgeon compare the patient’s current anatomy with the patient’s prepathological anatomy. The surgeon may virtually interact with the virtual object to see the virtual object from multiple angles.[ ] In another example, CAOS 118 may obtain medical image data of a first portionof the anatomic structure, where the medical image data does not include the second portion of the anatomic structure. CAOS 118 may generate the input anatomic model based on the medical image data. For instance, source anatomic model 128 may exclude distal and / or medial portions of a patient’s scapula. In this example, predicted anatomic model 132 may represent the entire scapula, including distal and / or medial portions of the scapula. CAOS 118 may use predicted anatomic model 132 to determine an axis of a scapular spine, or other landmarks, that CAOS 118 can use to determine an inclination or version of a glenoid fossa of the scapula. In another example, the anatomic structure includes a tibia, the first portion of the anatomic structure is a distal portion of the tibia, and a second portion of the anatomic structure included in predicted anatomic model 132 may be a proximal portion of the tibia. A similar process may be performed to predict the patient’s distal humerus from the shape of the patient’s proximal humerus, and thereby obtain information regarding the patient’s elbow.[ ] Generating a predicted anatomic model in this way may allow a user of CAOS118 to see and understand the omitted portion of the anatomic structure. The omitted portion of the anatomic structure may have significance for surgical planning. For instance, the mechanical axis of a tibia may be important in determining a position of a tibial prosthesis but the mechanical axis is defined from landmarks at the proximal and 1262-252WO01 / DIG-24-1695-IDF distal ends of the tibia. However, the proximal end of the tibia is frequently omitted from medical images. Similar considerations apply with respect to the medial and inferior portions of the scapula. These portions of the scapula may have importance for estimating axes and soft tissue attachment portions that contribute to surgical planning. Thus, generating predicted anatomic model 132 in this way may avoid the need for additional medical imaging of the patient, potentially saving time, expense, and radiation exposure.[ ] In another example, source anatomic model 128 may represent a bone having oneor more osteophytes. Because osteophytes are idiosyncratic for individual patients, ML model 130 does not generate predicted anatomic models that include osteophytes. Hence, CAOS 118 may compare source model 128 to predicted anatomic model 132 to identify osteophytes in source model 128. Osteophytes may limit the range of motion of a joint. Hence, in some examples, CAOS 118 may use predicted anatomic model 132 (without osteophytes) to predict a post-surgical range of motion of a joint involving the bone. Since osteophytes are commonly removed during surgery, the predicted post-surgical range of motion may be more accurate than if the predicted post-surgical range of motion were based on an anatomic model that includes the osteophytes. In some examples, estimation of anatomical measurements (e.g., anatomical axes, anatomical landmarks, dimensions of bones, angles of bone surfaces, etc.) may be more accurate when determined based on a model free of osteophytes.[ ] Furthermore, CAOS 118 may generate instructions for fabrication of patient-specific guides. For example, in a total shoulder arthroplasty, a patient-specific guide may be positioned on the patient’s scapula to guide insertion of a pin into the glenoid fossa of the patient’s scapula. In an example involving a total ankle replacement, a patient-specific guide may be positioned on an anterior surface of the patient’s tibia to guide insertion of pins and to guide resection of an area of the distal tibia. Because osteophytes are typically delicate and may interfere with specific steps of the surgery, surgeons frequently remove osteophytes. However, if the patient-specific guides have surfaces shaped to accommodate the osteophytes, removal of the osteophytes prior to using the patient- specific guides may disrupt the relationship between the patient-specific guides and the patient’s bone. Accordingly, CAOS 118 may generate a model of a patient-specific guide based on predicted anatomic model 132, which is free of osteophytes. Thus, resulting patient-specific guide may better match the patient’s bone when the patient-specific guide is being used. Thus, the 3D mesh of the input anatomic model may represent one or more 1262-252WO01 / DIG-24-1695-IDF osteophytes and the 3D mesh of predicted anatomic model 132 does not represent the one or more osteophytes.[ ] As indicated above, predicted anatomic model 132 may include a meshrepresenting cartilage associated with a bone. In other words, the input anatomic model may include a first portion of an anatomic structure (e.g., at least a portion of the bone) and predicted anatomic model 132 may include a second portion of the anatomic structure that includes cartilage associated with the anatomic structure. For example, source anatomic model 128 may represent a scapula and the overlay mesh may represent a labrum. CAOS 118 may display predicted anatomic model 132 to a user, e.g., to help the user have a more complete understanding of the patient’s shoulder joint. Knowing the likely premorbid size and shape of the patient’s labrum may help the user or CAOS 118 select an appropriate prothesis system for the patient, such as anatomic glenoid implant having a concave surface having dimensions similar to the diameter and depth of the patient’s glenoid fossa and labrum. Predicting cartilage may be difficult in other techniques because the cartilage is difficult to detect in typical medical images, such as CT or x-ray images.[ ] In some examples, source anatomic model 128 is generated based on post-operative scans, e.g., CT scans, of the patient after implantation of a prosthesis. Because the prosthesis may contain metal, the CT scan may include various artifacts that distort the shape of a 3D model based on the CT scan. For instance, metal in the prosthesis may create shadows in slices of the CT scan. Such artifacts may disrupt the ability of computerized systems to automatically generate the measurements used for premorbid shape reconstruction using statistical shape modeling techniques. CAOS 118 may exclude such artifacts when generating an anatomic model based on the CT slices. CAOS 118 may then generate a predicted anatomic model based on the source anatomic model. The mesh of predicted anatomic model 132 may include faces corresponding to the areas where the artifacts were present. Thus, in some examples, CAOS 118 may obtain medical image data 126 of the anatomic structure after implantation of an orthopedic prosthesis on the anatomic structure. CAOS 118 may segment the medical image data to identify artifacts in the medical image data attributable to the orthopedic prosthesis. CAOS 118 may then generate the input anatomic model based on portions of the medical image data other than the artifacts. In this way, CAOS 118 may be able to generate a more accurate model of the bone. Use of GPT 134 in ML model 130 may enable CAOS 118 to generate 1262-252WO01 / DIG-24-1695-IDF a predicted anatomic model even if key landmarks areas used in statistical shape modeling cannot be used because of artifacts.[ ] In some examples, CAOS 118 may use predicted anatomic model 132 to improvesegmentation. Segmentation is a process in which an image is partitioned into multiple image segments (i.e., regions). Segmentation of images of anatomic features may be difficult in some circumstances. For example, precise segmentation of the boundaries between closely spaced bones, such as the bones of the foot, wrist, and hand, may be challenging for some segmentation algorithms. Use of ML model 130 with GPT 134 from source models representing portions of such anatomic structures that are easier to segment may result in generation of full models of the anatomic structures. CAOS 118 may then use the full models of the anatomic structures to assist with segmentation of the more difficult to segment portions of the anatomic structures. Thus, in some examples, CAOS 118 obtain medical image data 126 of the anatomic structure and may segment medical image data 126 based on predicted anatomic model 132. Medical image data 126 may include x-ray data, CT data, Magnetic Resonance Imaging (MRI) data, arthroscanner data, whole body CT data, or other types of image data.[ ] In some examples, source anatomic model 128 represents two or more bones of ajoint. For instance, source anatomic model 128 may represent a scapula and a humerus of a patient’s shoulder joint. In another example, source anatomic model 128 may represent a tibia and a talus of a patient’s ankle joint. One or both bones of the joint may have pathological areas. CAOS 118 may generate one or more input anatomic models by modifying the models of the anatomic structures of the joint to remove the pathological areas of one or more of the models. The pathological areas may include areas that are actually pathological and areas adjacent to such areas that are actually pathological. CAOS 118 may provide the input anatomic models as input to machine learning model 130. In some examples, CAOS 118 may provide the modified models as input to machine learning model 130 as a single mesh. Predicted anatomic model 132 generated by machine learning model 130 may include reconstructed premorbid versions of the bones of the joint. Using multiple bones of the joint may help machine learning model 130 to more accurately reconstruct premorbid anatomic structures of the joint than if the premorbid anatomic structures of the joint were predicted independently.[ ] In some examples, the input anatomic model represents a current state of theanatomic structure with an orthopedic prosthesis implanted thereon. For example, the input anatomic model may represent a scapula after implantation of a glenoid prosthesis 1262-252WO01 / DIG-24-1695-IDF onto the scapula. When planning a revision surgery, a user of CAOS 118 may want to understand a shape of the anatomic structure prior to implantation of the orthopedic prosthesis. Accordingly, CAOS 118 may generate the input anatomic model so that the input anatomic model does not represent the orthopedic prosthesis (and, in some examples, areas adjacent to the orthopedic prosthesis), e.g., using a segmentation process or a geometric primitive as described elsewhere in this disclosure. The predicted anatomic model generated by ML model 130 based on the input anatomic model may then represent a state of the anatomic structure prior to implantation of the orthopedic prosthesis on the anatomic structure. Conventional techniques for determining the preimplantation state of the anatomic structure are frequently user-subjective and / or do not produce models that are readily rendered or used for subsequent planning purposes.[ ] In some examples, processing circuitry 104 (e.g., GPU 105) performs a renderingprocess to render predicted anatomic model 132 for display by display 108. For instance, processing circuitry 104 may apply vertex shading to vertices of predicted anatomic model 132. Since predicted anatomic model 132 is already a mesh, it is unnecessary to convert predicted anatomic model 132 into a mesh format, unlike other modeling techniques such as point clouds and statistical shape model-based approaches. Avoiding the conversion to mesh format may save computation resources and avoid creating discrepancies between predicted anatomic model 132 and a model used for rendering. During vertex shading, processing circuitry 104 applies transformations to vertex positions using matrices, such as model, view, and projection matrices, to convert the vertex positions from a local object space to a world space and then to a clip space. Processing circuitry 104 may also perform per-vertex lighting calculations to determine how light interacts with a vertex, including computing the vertex’s color based on light sources and material properties. The rendering process may further include tessellation, geometry shading, primitive assembly, and rasterization. Since predicted anatomic model may include a large number of vertices, the rendering process may perform many such operations in parallel in order to render predicted anatomic model 132 for display. Because predicted anatomic model 132 is natively a mesh-based, it is simpler relative to non-mesh-based models for processing circuitry 104 to re-render predicted anatomic model 132 for display in response to user input to rotate or scale predicted anatomic model 132 or change a position of a light source illuminating predicted anatomic model 132. Thus, CAOS 118 may to be more responsive to the user input, enhancing the experience of the user and saving time. 1262-252WO01 / DIG-24-1695-IDF[ ] In some examples, CAOS 118 determines positioning information for one or moreorthopedic prostheses based on predicted anatomic model 132. For instance, in an example where the patient’s anatomy is a scapula, CAOS 118 may automatically determine, based on predicted anatomic model 132, a center of rotation of the patient’s humerus relative to the patient’s premorbid scapula. In some such examples, CAOS 118 may automatically test multiple positions of the center of rotation to determine a center of rotation that leads to an appropriate range of motion. During testing, CAOS 118 may apply rotations to a mesh-based model of the humerus relative to predicted anatomic model 132 and detect collisions between the models. CAOS 118 may determine a recommended position of an anatomic or reverse glenoid prosthesis such that a center of rotation of a corresponding humeral prosthesis corresponds to the premorbid center of rotation of the patient’s humeral head with respect to the patient’s glenoid fossa. In an example where the patient’s anatomy is a tibia, CAOS 118 may determine, based on predicted anatomic model 132, the premorbid tibial plafond. CAOS 118 may determine a recommended tibial prosthesis and / or talar prosthesis and positions thereof so that, when implanted, the tibial prosthesis and / or talar prosthesis meet at the predicted tibial plafond.[ ] Furthermore, in the example of FIG. 1, system 100 includes a robot 140. Robot140 is configured to assist in performing a surgery. For example, robot 140 may be configured to hold a surgical tool and maneuver the surgical tool so that the surgical tool (or a surgical item attached to the surgical tool) enters the patient’s anatomy at a precise position and orientation specified by a surgical plan. Example surgical tools may include drills, saws, reamers, and so on. Example surgical items may include drill bits, saw blades, pins, wires, and other items used in surgery. CAOS 118 may automatically determine the position and orientation based on predicted anatomic model 132. For example, CAOS 118 may determine a recommended position of a prosthesis as discussed above. In this example, CAOS 118 determines, based on the position of the prosthesis, a position and orientation of a pin used in the process of implanting the prosthesis. In this example, CAOS 118 configures robot 140 to install the pin in the patient at the determined position and orientation. In another example, CAOS 118 determines, based on the position of the prosthesis, a cut plane for resecting a portion of a bone (e.g., humerus, tibia, talus, femur, etc.) as part of a process to implant the prosthesis. In this example, CAOS 118 configures robot 140 to perform the cut along the cut plane. As discussed above, the techniques of this disclosure may increase the accuracy of the predicted, premorbid anatomic models relative to existing techniques. Since accuracy of predicted anatomic model 132 is 1262-252WO01 / DIG-24-1695-IDF important for determining the planned position and orientation of the prosthesis, increased accuracy of predicted anatomic model 132 may lead to better outcomes when using robot 140.[ ] In the example of FIG. 1, system 100 includes a manufacturing system 142. Insome such examples, CAOS 118 configures manufacturing system 142 to manufacture patient-specific hardware based on predicted anatomic model 132. Example patient- specific hardware includes patient-specific instruments, patient-specific prostheses, patient-specific components of prostheses, and so on. Configuring manufacturing system 142 may involve providing a mesh-based model of the patient-specific hardware to an application programming interface (API) of manufacturing system 142, which is further configured to generate and execute control instructions to manufacture the patient- specific hardware. Manufacturing system 142 may use a 3D printing process or other manufacturing process to manufacture the patient-specific hardware. Since accuracy of predicted anatomic model 132 is important for determining the planned position and orientation of the prosthesis, increased accuracy of predicted anatomic model 132 may lead to better adapted patient-specific hardware.[ ] In an example where the patient’s anatomy is a scapula, CAOS 118 determinespositioning information for one or more prostheses based on predicted anatomic model 132. The positioning information includes information on a position and orientation of a base plate of a glenoid prosthesis. In this example, CAOS 118 projects the base plate medially to the input anatomic model to determine a 3-dimensional volume. CAOS 118 configure manufacturing system 142 to manufacture an augment component of the glenoid prosthesis having the 3-dimensional volume. In another example, CAOS 118 determines an insertion axis of a pin based on the planned position of a glenoid prosthesis. CAOS 118 then generates a model of a patient-specific guide for guiding the pin into the patient’s glenoid fossa along the determined insertion axis. To generate the model of the patient-specific guide, CAOS 118 projects points on a guide template from a base plane onto the input anatomic model. The guide template may include a plane and have feet at positions corresponding to user- or computer-selected points on a rim of the glenoid fossa. CAOS 118 then configures manufacturing system 142 to manufacture the patient-specific guide. In another example, the patient-specific guide may guide pins and cutting planes for resection of a portion of a distal tibia. As mentioned above, the patient-specific guide based on predicted anatomic model 132, which may be free of osteophytes. 1262-252WO01 / DIG-24-1695-IDF[ ] FIG. 2 is a block diagram illustrating example components of CAOS 118, inaccordance with one or more techniques of this disclosure. In the example of FIG.2, the components of CAOS 118 include machine learning model 130, a prediction unit 202, a training unit 204, and a feature enhancement unit 206. Machine learning model 200 includes a feature encoder 210, a quantization module 212, a feature decoder 214, a GPT 134, and a codebook 218. In other examples, CAOS 118 may be implemented using more, fewer, or different components. For instance, training unit 204 may be omitted in instances where ML model 130 has already been trained. In some examples, one or more of the components of CAOS 118 are implemented as software modules. Moreover, the components of FIG. 2 are provided as examples and CAOS 118 may be implemented in other ways.[ ] In general, prediction unit 202 applies ML model 130 to an input anatomic modelto generate predicted anatomic model 132. Feature encoder 210 of ML model 130 generates a feature graph and input features based on the input anatomic model. Feature encoder 210 may then apply a graph convolutional encoder to the feature graph and input features to generate face embeddings. Quantization module 212 uses codebook 218 to generate quantized face embeddings based on the face embeddings. Prediction unit 202 may sequence the quantized face embeddings to form a sequence of input tokens for GPT 134. GPT 134 generates a sequence of output tokens (e.g., predicted codebook indices) based on the sequence of input tokens (i.e., the sequence of quantized face embeddings). Prediction unit 202 uses codebook 218 to generate a sequence of quantized face embeddings based on the predicted codebook indices (e.g., the sequence of output tokes). Feature decoder 214 uses the sequence of quantized face embeddings to generate a predicted anatomic model.[ ] Training unit 204 may train ML model 130. Training unit 204 may use a set oftraining examples 230 to train ML model 130. In one or more examples, training unit 204 may be configured to train ML model 130 using a plurality of training epochs. Feature enhancement unit 206 may generate synthetic training examples to enhance training examples 230. In some examples, ML model 130 may be trained by a device or system other than computing system 102. In such examples, computing system 102 may receive ML model 130, including GPT 134.[ ] FIG. 3A is a conceptual diagram illustrating an example input anatomic model300, in accordance with one or more techniques of this disclosure. In the example of FIG. 3A, input anatomic model 300 includes a mesh of faces with a potentially pathological 1262-252WO01 / DIG-24-1695-IDF area removed. Prediction unit 202 may use input anatomic model 300 as an input to ML model 200 for generate a predicted anatomic model.[ ] FIG. 3B is a conceptual diagram illustrating an example predicted anatomicmodel, in accordance with one or more techniques of this disclosure. Prediction unit 202 may use ML model 200 to generate predicted anatomic model 350 based on input anatomic model 300. As shown in the example of FIG.3B, predicted anatomic model 350 includes shaded faces for portions of a scapula not provided in input anatomic model 300. The predicted faces may correspond to a premorbid shape of the scapula.[ ] FIG. 4 is a flowchart illustrating an example operation of CAOS 118, inaccordance with one or more techniques of this disclosure. In the example of FIG. 4, CAOS 118 may obtain an input anatomic model that comprises a first 3D mesh representing a surface of at least a first portion of an anatomic structure of a patient (400). CAOS 118 may generate input tokens based on the input anatomic model (402). CAOS 118 may apply GPT 134 to the input tokens to generate output tokens (404). CAOS 118 may generate, based on the output tokens, a predicted anatomic model that includes a second 3D mesh representing a surface of at least a second portion of the anatomic structure (406).[ ] FIG. 5 is a block diagram illustrating an example architecture for generatingquantized face embeddings and generating a reconstructed anatomic model, in accordance with one or more techniques of this disclosure. In the example of FIG.5, feature encoder 210 includes a feature generator 500 and a graph convolutional encoder 502. Feature decoder 214 includes a decoder 542. The architecture of FIG.5 may be used for training graph convolutional encoder 502 and decoder 542 separately from GPT 134.[ ] Feature encoder 210 obtains source anatomic model 128. Feature generator 500generates a face graph and input features based on an input anatomic model 504. In some examples, input anatomic model 504 may be the same as source anatomic model 128. In some examples, input anatomic model 504 is a subset of faces of the source anatomic model 128.[ ] In the face graph, each face corresponds to a node of a graph. The nodescorresponding to neighboring faces are connected in the graph by undirected edges. Each of the nodes has a feature vector. For each of the nodes, the feature vector of the node may include 9 coordinate values (i.e., 3 coordinate values for each of the 3 dimensions) for each of the vertices of the corresponding face, data defining a face normal vector, data 1262-252WO01 / DIG-24-1695-IDF defining angles between edges of the corresponding face, and data indicating an area of the corresponding face.[ ] Graph convolutional encoder 502 generates face embeddings 508 based on theface graph and input features 506. Graph convolutional encoder 502 performs one or more rounds of message passing. During each round of message passing, graph convolutional encoder 502 collects, for each node of the graph, feature vectors of nodes that neighbor the node. Graph convolutional encoder 502 then aggregates the feature vector of the node with the feature vectors of the nodes that neighbor the node, thereby updating the feature vector of the node. Aggregating the feature vectors may involve determining averages of corresponding features in the feature vectors. After completing the one or more rounds of message passing, graph convolutional encoder 502 may, for each node of the graph, use the updated feature vector of the node as input for a neural network 507 that is trained to generate a face embedding for the face corresponding to the node. In this way, graph convolutional encoder 502 may generate a face embedding for each face of the input anatomic model.[ ] In some examples, neural network 507 includes a series of graph convolutionallayers, such as SAGE-Conv graph convolutional layers. In some such examples, a first layer of neural network 507 takes 16 features as input: 9 features indicating point coordinates of a face, 3 features indicating a normal vector of the face, 3 features indicating angles at the vertices of the face, and feature indicating an area of the face.[ ] Quantization module 212 generates quantized face embeddings 508 based on faceembeddings 506. In the example of FIG. 5, quantization module 212 includes feature splitting unit 520 configured to split the face embeddings 508 of each of the faces into a plurality of sub-vectors 512. Each of sub-vectors 512 may correspond to a vertex of a face. An aggregation unit 522 of quantization module 212 may then aggregate sub-vectors 512 for corresponding vertices into aggregated sub-vectors 523. For instance, each vertex of the model may have an index. Thus, two faces that share a vertex and the information about the faces may indicate the same index for the shared vertex. When aggregating the sub-vectors for corresponding vertices, quantization module 212 may calculate averages of the features in the sub-vectors for vertices having the same index.[ ] For each of the aggregated sub-vectors 523, a codebook lookup unit 524 ofquantization module 212 may look up a codebook index in codebook 218 for the aggregated sub-vector. Codebook 218 includes a plurality of representative vectors. Each of the representative vectors may have the same number of elements as each of the 1262-252WO01 / DIG-24-1695-IDF aggregated sub-vectors. Each representative vector is associated with a unique codebook index. To look up a codebook index in codebook 218 for an aggregated sub-vector, codebook lookup unit 524 may determine which of the representative vectors is closest to the aggregated sub-vector. For instance, codebook lookup unit 524 may treat the aggregated sub-vector and the representative vectors as specifying points in a multi- dimensional space, calculate a Euclidean distance between the aggregated sub-vector and the representative vectors, and determine which of the representative vectors has the lowest Euclidean distance. In this way, codebook lookup unit 524 may output a plurality of codebook indexes 526 for each face. Furthermore, in some examples, codebook lookup unit 524 may apply a tier, residual quantization process that determines multiple codebook indexes for each aggregated sub-vector of a face. In some examples, the representative vectors in codebook 218 are learned. Training unit 204 may use a k-means clustering algorithm to learn the representative vectors. In examples where training unit 204 uses the k-means clustering algorithm to learn the representative vectors, the representative vectors may be cluster-center vectors that indicate the centers of clusters.[ ] A reshaper unit 510 of quantization module 212 uses codebook indexes 526 togenerate quantized face embeddings 528. That is, reshaper unit 510 receives a stack of codebook indexes 526 for each face. For each of the faces, reshaper unit 510 may reduce the stack of codebook indexes for the face to a single feature embedding per face through summation across representative vectors, and concatenation across vertices, thereby generating quantized face embeddings 528 for the faces. For example, reshaper unit 510 may use the following equation to determine quantized face embeddings 528: ,… , In the equation above, indicates a vector containing quantized face embeddings 528 forthe faces, , … indicates the individual quantized face embeddings for faces 1 throughis a d-th codebook index (i.e., a token) of vertex v of face i, indicates the representative vector in codebook 218 corresponding to the codebook index., indicates concatenation of features for vertices v = 0 through v = 2 of face i.[ ] Feature decoder 214 includes a sequence generator 540 and a decoder 542.Sequence generator 540 generates a sequence of quantized face embeddings 544 for the faces. The quantized face embeddings in the sequence of quantized face embeddings 544 may be sequenced in the same way as face embeddings 508. 1262-252WO01 / DIG-24-1695-IDF[ ] Decoder 542 generates a reconstructed anatomic model 546 based on the sequenceof quantized face embeddings 544. For example, decoder 542 outputs a set of 9 coordinate values for each face of reconstructed anatomic model 546. Reconstructed anatomic model 546 is a reconstructed version of input anatomic model 504. Thus, reconstructed anatomic model 546 does not include predicted portions of the anatomic object, e.g., bone. Decoder 542 may be implemented as a 1D ResNet decoder. In some examples, the ResNet decoder includes 34 layers, i.e., decoder 542 may be implemented using a ResNet34 model. In other examples, other decoders may be used, such as ResNet18 model, a ResNet101 model, or another type of neural network model. In general, decoder 542 may be able to reconstruct structures with more variability if decoder 542 has more layers. In examples where ML model 130 is used only for generating a limited number of structures, such as a scapula or tibia, an implementation of decoder 542 with 18 or 34 layers may be sufficient, as opposed to a 101-layer ResNet 101 model. Using fewer layers may conserve computing resources (e.g., processing cycles and memory) if performance is similar. Inference time may also be reduced in models having fewer layers.[ ] FIG. 6 is a flowchart illustrating an example operation of prediction unit 202 togenerate predicted anatomic model 132, in accordance with one or more techniques of this disclosure. In the example of FIG. 6, feature encoder 210 obtains an input anatomic model and generates face embeddings based on the input anatomic model, as described above with respect to FIG.5 (600). Additionally, quantization module 212 applies graph convolutional encoder 506 to generate quantized face embeddings based on the face embeddings, as described above with respect to FIG.5 (602).[ ] Prediction unit 202 may then generate a sequence of features (i.e., a sequence ofinput tokens) based on the quantized face embeddings (606). The sequence of features may include a feature for each of the quantized face embeddings. The feature for a quantized face embedding may be based on the quantized face embedding and a learned discrete positional encoding for the quantized face embedding. The learned discrete positional encoding for the quantized face embedding provides information about the position of the quantized face embedding in the sequence of features and an index of each embedding within the quantized face embedding. In other words, for each embedding within the quantized face embedding, a position encoding value for the embedding is determined and then added to (i.e., summed with) the embedding. In some examples, a function for determining position encoding values is learned as part of a training process for GPT 134. In some examples, functions for determining position encoding values 1262-252WO01 / DIG-24-1695-IDF include sinusoidal functions. Additionally, prediction unit 202 may include a start embedding at the start of the sequence and an end embedding at the end of the sequence. The start embedding indicates the start of the sequence. The end embedding indicates the end of the sequence.[ ] Prediction unit 202 may autoregressively apply GPT 134 to the sequence offeatures to generate output tokens (608). The sequence of features may include the quantized face embeddings. The output tokens may include a sequence of codebook indices for faces of at least a portion of the anatomic structure. For example, prediction unit 202 may apply GPT 134 to the sequence of features to generate a set of predicted codebook indexes for a face. If the set of predicted codebook indices does not include a stop token, prediction unit 202 may use codebook 218 to determine a quantized face embedding for the face. Prediction unit 202 may then update the sequence of quantized face embeddings to include the quantized face embedding. Prediction unit 202 may use the updated sequence of quantized face embeddings as input to GPT 134 to generate a next set of predicted codebook indices. This cycle may continue until the set of predicted codebook indices includes the stop token. GPT 134 may be a GPT-2 medium architecture. That is, GPT transformer may have 24 multi-headed self-attention layers, 16 heads, 768 feature width and a context length of 4608, e.g., as described in Siddiqui et al., “MeshGPT: Generating Triangle Meshes with Decoder-Only Transformers”, arXiv:2311.15475v1 [cs.CV] 27 Nov 2023.[ ] In some examples, prediction unit 202 may apply GPT 134 to first input tokensand second input tokens to generate the output tokens. The first input tokens may be based on the input anatomic model (e.g., the first input tokens may include the quantized face embeddings). The second input tokens may include information describing the patient and not describing the first 3D mesh. The information describing the patient includes one or more of: a gender of the patient, an age of the patient, a glenoid type of a scapula of the patient, a pathology of the anatomic structure, or other information describing the patient. Inclusion of the second input tokens may improve the accuracy of predicted anatomic model 132.[ ] Prediction unit 202 may determine a sequence of quantized face embeddingsbased on the codebook indices (610). Prediction unit 202 may determine the quantized face embeddings in the same as way described above with respect to FIG.5.[ ] Prediction unit 202 may then apply feature decoder 214 to the sequence ofquantized face embeddings to generate predicted anatomic model 132 (612). In some 1262-252WO01 / DIG-24-1695-IDF examples, prediction unit 202 may update the sequence of face embeddings and predicted anatomic model 132 as prediction unit 202 uses GPT 134 to progressively predict additional sets of predicted codebook indices.[ ] FIG. 7 is a flowchart illustrating an example training process for ML model 130,in accordance with one or more techniques of this disclosure. As mentioned above, training unit 204 may train ML model 130. That is, training unit 204 may train ML model 130 to generate input tokens based on the input anatomic model, apply GPT 134 of ML model 130 to the input tokens to generate output tokens; and generate, based on the output tokens, a predicted anatomic model that includes a second 3D mesh representing a surface of at least a second portion of the anatomic structure.[ ] In a first phase of training ML model 130, training unit 204 may train graphconvolutional encoder 502 and feature decoder 214 and learn the representative vectors of codebook 218 (700). In a second phase of training ML model 130, training unit 204 may training unit 204 may train GPT 134 (702).[ ] In the first phase, training unit 204 may perform one or more training epochs. Ineach training epoch, training unit 204 may perform a series of training iterations. In each training iteration, training unit 204 provides an input anatomic model from training examples 230 as input to feature encoder 210. Feature encoder 210, quantization module 212, and feature decoder 214 may generate a reconstructed anatomic model based on the input anatomic model. During the first phase, the input anatomic model may be a complete model of an anatomic structure, such as a healthy bone. After generating the reconstructed anatomic model, training unit 204 may calculate a reconstruction loss value based on a comparison of the input anatomic model and the reconstructed anatomic model. For example, training unit 204 may calculate the reconstruction loss value using a cross-entropy loss on discrete mesh coordinates. Training unit 204 may use a backpropagation method to update parameters (e.g., weights) of feature encoder 210 and feature decoder 214 based on the reconstruction loss value.[ ] In some examples, training unit 204 updates the representative vectors ofcodebook 218 as part of the training process. For example, training unit 204 may add aggregated sub-vectors 523 generated during a training iteration to a training set of aggregated sub-vectors. Training unit 204 may assign a weight to each of the aggregated sub-vectors in the training set such that the aggregated sub-vectors have exponentially less weight the longer the aggregated sub-vectors have remained in the training set. In 1262-252WO01 / DIG-24-1695-IDF some examples, training unit 204 uses the k-means clustering process to update the representative vectors based on the aggregated sub-vectors in the training set.[ ] In some examples, quantization module 212 applies a tiered, residual vectorquantization process. For instance, training unit 204 may initialize a set of k1first-tier representative vectors, assign generated aggregated sub-vectors (i.e., first-tier vectors) to their closest first-tier representative vectors, and then update the first-tier representative vectors based on an average position of the first-tier vectors assigned to the first-tier representative vectors. Training unit 204 may repeat the assignment and updating steps until a termination condition is reached. For each of the first-tier vectors, training unit 204 may calculate second-tier residual vectors indicating differences between the first-tier vectors and their first-tier representative vectors. Training unit 204 then initializes a set of k2 second-tier representative vectors, assigns the second-tier residual vectors to their closes second-tier representative vectors, updates the second-tier representative vectors based on an average position of the second-tier residual vectors assigned to the second- tier representative vectors, and repeats the assignment and updating steps until the termination condition is reached. Training unit 204 may repeat this process for one or more tiers. Thus, when codebook lookup unit 524 receives an aggregated sub-vector (i.e., a first-tier vector), codebook lookup unit 524 determines a codebook index of a first-tier representative vector for the first-tier vector, determines a second-tier residual vector indicating a difference between the first-tier vector and the first-tier representative vector, determines a codebook index of a second-tier representative vector for the second-tier residual vector, determines a third-tier residual vector indicating a difference between the second-tier residual vector and the second-tier representative vector, determines a codebook index of a third-tier representative vector for the third-tier residual vector, and so on. In this way, for each aggregated sub-vector, codebook lookup unit 524 identifies a stack of codebook indexes for the aggregated sub-vector.[ ] In some examples, training unit 204 applies a straight-through estimator (STE)technique to optimize vector quantization. During a training iteration, training unit 204 quantizes face embeddings 508 to generate quantized face embeddings 528. For each of face embeddings 508, training unit 204 calculates a commitment loss value that characterizes the difference between the face embedding and the quantized face embedding. For instance, training unit 204 may calculate the commitment loss value as: 1262-252WO01 / DIG-24-1695-IDF where z is the face embedding, is the quantized face embedding, D is the number of dimensions of the face embedding, and sg is the stop gradient operation. After the reconstructed anatomical model 546 is generated for the training iteration, training unit 204 calculates a general loss value as a sum of the reconstruction loss value (as described above) and the commitment loss values. Training unit 204 uses the backpropagation method to update parameters of feature encoder 210 and feature decoder 214 based on the general loss value. However, the vector quantization function applied by quantization module 212 cannot be directly learned by using backpropagation because the vector quantization function is not differentiable. Hence, during backpropagation, the vector quantization function is regarded as an identity function, allowing modification of parameters of neurons of neural network 507 that generate face embeddings 508. This alters the values in face embeddings 508 that are fed into the vector quantization function applied by quantization module 212. This ultimately has the effect of reducing the commitment loss values produced by the vector quantization function because face embeddings 508 are generated so that aggregated sub-vectors generated from face embeddings 508 are made closer to fixed representative vectors in codebook 218. The STE technique may be used with the tiered residual vector quantization process as described above but the representative vectors at the various tiers are not updated.[ ] In this way, feature encoder 210, quantization module 212, and feature decoder214 may learn to generate mesh triangle models of anatomic structures and feature vectors that provide compact representations of the anatomic structures.[ ] As noted above, training unit 204 may calculate a reconstruction loss value basedon a comparison of an input anatomic model and a reconstructed anatomic model. Training unit 204 may calculate the reconstruction loss value in one of a variety of ways. For example, training unit 204 calculate a mesh regularization loss value (e.g., a chamfer loss, am Earth's Mover Distance loss (also known as Wasserstein loss), an edge regularization loss, a normal consistency loss, or a Laplacian loss. In some examples, training unit 204 may calculate a curvature-weighted chamfer loss, e.g., as described in Bongratz et al., “Vox2Cortex: Fast Explicit Reconstruction of Cortical Surfaces from 3D MRI Scans with Geometric Deep Neural Networks, arxiv:2203.09446v2 [cs.CV] 18 Mar. 2022, where areas with high curvature are present, such as certain areas of a scapula. 1262-252WO01 / DIG-24-1695-IDF[ ] In some examples, training unit 204 calculates different multiple types ofreconstruction loss values and calculates a weighted reconstruction loss value as a weighted sum of the reconstruction loss values. The losses may have a different weight according to the location of the faces or points. For example, training unit 204 may use different weights for the reconstruction loss values for different areas of the anatomic structure. In some examples, training unit 204 calculates a single type of reconstruction loss value for the whole anatomic structure as a sum of reconstruction losses for individual locations on the initial and reconstructed anatomic models. In some such examples, training unit 204 may use different weights for different regions. In this way, more emphasis may be placed on accurate reconstruction of important areas. For example, greater weight can be used for areas of a scapula closer to the glenoid fossa of the scapula than areas further from the glenoid fossa.[ ] As noted above, training unit 204 may train GPT 134 after completing training offeature encoder 210 and feature decoder 214 and learning codebook 218. In other words, the parameters of feature encoder 210, the parameters of feature decoder 214, and learning codebook 218 may be frozen while training GPT 134. Training unit 204 may train GPT 134 using self-attention to autoregressively generate an ordered sequence of triangles which define a final mesh. Given a target token sequence (i.e., a target sequence ofcodebook indices) with , , … , and is thecorresponding predicted sequence element, training unit 204 may train GPT 134 with a loss function of: where N is the number of tokens in the target token sequence, D is the number of sub- vectors per face embedding, |C| is the number of entries on codebook 218, and is the probability of the predicted sequence element being equal to the corresponding token in the target token sequence. Thus, after training feature encoder 210 and learning codebook 218, training unit 204 may use feature encoder 210 and quantization module 212 to generate quantized face embeddings based on training examples 230. As previously mentioned, training examples 230 may include models of healthy or complete anatomical structures. 1262-252WO01 / DIG-24-1695-IDF[ ] Training unit 204 may use training examples 230 to train feature encoder 210,quantization model 212, feature decoder 214, and GPT 134. Training examples 230 may include example anatomic models, such as anatomic models. Training examples 230 may include models of anatomic structures of real patients. For example, in the first training phase, training unit 204 may, for each first training example of a set of first training examples, apply feature encoder 210 to an input anatomic model associated with the training example to generate face embeddings associated with the first training example. Quantization module 212 may apply generate input tokens associated with the first training example based on the face embeddings associated with the first training example. Training unit 204 may apply feature decoder 214 to the input tokens associated with the first training example to generate a reconstructed anatomic model associated with the training example. Training unit 204 may apply a first loss function to generate a loss value associated with the first training example based on the input anatomic model associated with the first training example and the reconstructed anatomic model associated with the first training example. Training unit 204 may perform a first backpropagation process that modifies parameters of the feature encoder and the feature decoder based on the loss value associated with the first training example. In the second training phase following the first training phase, for each second training example of a set of second training examples, training unit 204 may apply feature encoder 210 to generate second face embeddings associated with the second training example. Training unit 204 may apply quantization module 212 to the face embeddings associated with the second training example to generate input tokens based on the second face embeddings. Training unit 204 may apply GPT 134 to generate output tokens associated with the second training example based on the input tokens associated with the second training example. Training unit 204 may apply feature decoder 214 to the input tokens associated with the second training example to generate a predicted anatomic model associated with the training example. Training unit 204 may apply a second loss function to generate a loss value associated with the second training example. Training unit 204 may perform a second backpropagation process that modifies parameters of the GPT based on the loss value associated with the second training example.[ ] In some examples, feature enhancement unit 206 performs data augmentation toimprove the training of machine learning model 130. In some examples, feature enhancement unit 206 performs scaling, rotation, mirroring (e.g., generating a synthetic left scapula model from an actual right scapula model), unit sphere normalization, jitter, 1262-252WO01 / DIG-24-1695-IDF multiresolution remeshing, or other techniques to generates synthetic training examples. Training unit 204 may use the synthetic training examples along with original training examples 230 for training of machine learning model 130.[ ] In some examples, training unit 204 may perform a mesh decimation process onthe anatomic models of training examples 230. For example, training unit 204 may use a Quadric Error Metrics (QEM) mesh simplification algorithm to reduce the number of faces in the anatomic models. Reducing the number of faces may reduce training time. In some examples, training unit 204 may reduce the number of faces in some areas of the anatomic models and not other areas of the anatomic models. For example, training unit 204 may reduce the number of faces in the areas of the anatomic models that are typically healthy while maintaining the number of faces in areas of the anatomic models that are typically pathological. With respect to the scapula, the areas of the anatomic models that are typically healthy include medial portions of the scapula while areas of the scapula that are typically pathological include the glenoid fossa, acromion process, and coracoid process. With respect to the tibia, the areas of the tibia that are typically pathologic include the tibial mortise. This may allow machine learning model 200 to generate fine details at reconstructed premorbid versions of pathological areas while reducing the amount of data passed to processing circuitry 104 (e.g., to GPU 105 and / or CPU). This may reduce a potential bottleneck of processing circuitry capabilities and / or may reduce inference time of the mesh completion algorithm when processing a mesh derived from a high-resolution CT scan.[ ] In some examples, CAOS 118 performs a similar mesh decimation process oninput anatomic model 504 after machine learning model 200 is in the production phase. Performing the mesh decimation process on input anatomic model 504 results in a relatively lower resolution mesh for the healthy areas of the anatomic structure. In such examples, machine learning model 200 generates a relatively high-resolution mesh for the reconstructed premorbid version of the pathological areas of an anatomic structure based on the relatively lower resolution mesh for the healthy areas of the anatomic structure. Because machine learning model 200 is operating on the relatively lower resolution mesh, the process of applying machine learning model 200 may be more computationally efficient. For instance, since fewer vertices and faces are involved, CAOS 118 may generate fewer input tokens. Since GPT 134 is applied to fewer input tokens, application of GPT 134 involves fewer read and write operations from memory. Each read and write operation involves time delay, energy consumption, and heat 1262-252WO01 / DIG-24-1695-IDF production. CAOS 118 may merge the relatively high-resolution mesh for the reconstructed premorbid version of the pathological areas of the anatomic structure with the high-resolution mesh of the healthy areas of original input anatomic model 504 to generate predicted anatomic model 132 having the relatively high resolution. Thus, predicted anatomic model 132 may include faces of the input anatomic model of the first portion and faces of the second portion of the anatomic structure. Having predicted anatomic model 132 be at the relatively high resolution may help a user better understand the shape of the anatomic structure and / or improve the accuracy of automated processes that use predicted anatomic model 132. This process of merging the models may allow CAOS 118 to generate a high resolution predicted anatomical model 132 while reducing demands on computational resources. It is noted that the process of decimating input anatomic model 504 and merging the models typically involves operations on large numbers of faces and vertices. However, these operations are significantly quicker and more computationally efficient than application of ML model 130 to the high-resolution mesh of the healthy areas of original input anatomic model 504.[ ] The following is a non-limiting list of clauses describing techniques of thisdisclosure.[ ] Clause 1. A computer-implemented method comprising: obtaining, by oneor more processors, an input anatomic model that comprises a first 3-dimensional (3D) mesh representing a surface of at least a first portion of an anatomic structure of a patient; generating, by the one or more processors, input tokens based on the input anatomic model; applying, by the one or more processors, a generative pre-trained transformer (GPT) to the input tokens to generate output tokens; and generating, by the one or more processors, based on the output tokens, a predicted anatomic model that includes a second 3D mesh representing a surface of at least a second portion of the anatomic structure.[ ] Clause 2. The computer-implemented method of clause 1, wherein: the firstportion of the anatomic structure is a non-pathological portion of the anatomic structure, and obtaining the input anatomic model comprises modifying, by the one or more processors, a source anatomic model to generate the input anatomic model, wherein the source anatomic model includes the non-pathological portion and a pathological portion, and the input anatomic model excludes the pathological portion of the source anatomic model. 1262-252WO01 / DIG-24-1695-IDF[ ] Clause 3. The computer-implemented method of clause 2, wherein:modifying the source anatomic model comprises: determining, by the one or more processors, a position of a geometric primitive shape relative to the source anatomic model; and determining, by the one or more processors, retained faces of the source anatomic model and discarded faces of the source anatomic model, the retained faces and the discarded faces being on different sides of a boundary defined by the geometric primitive shape.[ ] Clause 4. The computer-implemented method of clause 3, wherein: themethod further comprises determining, by the one or more processors, positions of one or more anatomical landmarks on an anatomy of the patient, and determining the position of the geometric primitive shape comprises determining, by the one or more processors, the position of the geometric primitive shape relative to the source anatomic model based on the positions of the one or more anatomical landmarks.[ ] Clause 5. The computer-implemented method of clause 3, wherein: thegeometric primitive shape is a sphere, the anatomic structure is a scapula, and determining the position of the geometric primitive shape comprises: determining, by the one or more processors, a center of the sphere based on a center of a glenoid fossa of the scapula; and after determining the center of the sphere, adjusting, by the one or more processors, a position of the sphere medially.[ ] Clause 6. The computer-implemented method of clause 3, wherein thegeometric primitive shape is a plane through the source anatomic model.[ ] Clause 7. The computer-implemented method of clause 6, wherein: theanatomic structure includes a scapula, the plane is a sagittal plane.[ ] Clause 8. The computer-implemented method of any of clauses 2-7,wherein modifying the source anatomic model comprises: segmenting, by the one or more processors, the source anatomic model to identify the pathological portion of the source anatomic model and the non-pathological portion of the source anatomic model; and modifying, by the one or more processors, the source anatomic model to exclude the pathological portion from the input anatomic model.[ ] Clause 9. The computer-implemented method of any of clauses 1-6 or 8,wherein the anatomic structure includes one of a scapula, a humerus, a radius, an ulna, a vertebra, a mandible, a femur, a tibia, a fibula, a talus, a cranium, a rib, a pelvis, a bone of a hand or foot. 1262-252WO01 / DIG-24-1695-IDF[ ] Clause 10. The computer-implemented method of any of clauses 1-9,wherein the second portion of the anatomic structure includes cartilage associated with the anatomic structure.[ ] Clause 11. The computer-implemented method of any of clauses 1-10,wherein the anatomic structure is a joint and the first portion of the anatomic structure includes portions of two or more bones of the joint.[ ] Clause 12. The computer-implemented method of any of clauses 1-11,wherein: the method further comprises obtaining, by the one or more processors, medical image data of the first portion of the anatomic structure, wherein the medical image data does not include the second portion of the anatomic structure; and obtaining the input anatomic model comprises generating, by the one or more processors, the input anatomic model based on the medical image data.[ ] Clause 13. The computer-implemented method of clause 12, wherein theanatomic structure includes a scapula and the second portion of the anatomic structure is a medial or distal portion of the scapula.[ ] Clause 14. The computer-implemented method of clause 12, wherein theanatomic structure includes a tibia, the first portion of the anatomic structure is a distal portion of the tibia, and the second portion of the anatomic structure is a proximal portion of the tibia.[ ] Clause 15. The computer-implemented method of any of clauses 1-14,wherein the predicted anatomic model includes faces of the input anatomic model of the first portion and faces of the second portion of the anatomic structure.[ ] Clause 16. The computer-implemented method of clause 15, wherein: themethod further comprises performing, by the one or more processors, a mesh decimation process on the input anatomic model to generate a decimated input anatomic model, generating the input tokens based on the input anatomic model comprises generating the input tokens based on the decimated input anatomic model, and generating the predicted anatomic model comprises merging the faces of the input anatomic model of the first portion and the faces of the second portion of the anatomic structure.[ ] Clause 17. The computer-implemented method of any of clauses 1-16,wherein: the input tokens comprise quantized face embeddings for faces of the input anatomic model, applying the GPT comprises autoregressively applying, by the one or more processors, the GPT to a sequence of features to generate the output tokens, 1262-252WO01 / DIG-24-1695-IDF wherein the sequence of features includes the quantized face embeddings, and the output tokens comprise codebook indices for faces of the second portion of the anatomic structure, generating the predicted anatomic model comprises: determining, by the one or more processors, a sequence of quantized face embeddings based on the codebook indices; and applying, by the one or more processors, a decoder to the sequence of quantized face embeddings to generate the predicted anatomic model.[ ] Clause 18. The computer-implemented method of any of clauses 1-17,wherein the first 3D mesh represents one or more osteophytes of the anatomic structure and the second 3D mesh does not represent the one or more osteophytes.[ ] Clause 19. The computer-implemented method of any of clauses 1-18,wherein obtaining the input anatomic model comprises: obtaining, by the one or more processors, medical image data of the anatomic structure after implantation of an orthopedic prosthesis on the anatomic structure; segmenting, by the one or more processors, the medical image data to identify artifacts in the medical image data attributable to the orthopedic prosthesis; and generating, by the one or more processors, the input anatomic model based on portions of the medical image data other than the artifacts.[ ] Clause 20. The computer-implemented method of any of clauses 1-19,further comprising: obtaining, by the one or more processors, medical image data of the anatomic structure; and segmenting, by the one or more processors, the medical image data based on the predicted anatomic model.[ ] Clause 21. The computer-implemented method of any of clauses 1-20,wherein: the input anatomic model represents a current state of the anatomic structure with an orthopedic prosthesis implanted thereon, and the predicted anatomic model represents a state of the anatomic structure prior to implantation of the orthopedic prosthesis on the anatomic structure.[ ] Clause 22. The computer-implemented method of any of clauses 1-21,wherein: the input tokens are first input tokens, applying the GPT comprises applying, by the one or more processors, the GPT to the first input tokens and second input tokens to generate the output tokens, and the second input tokens include information describing the patient and not describing the first 3D mesh.[ ] Clause 23. The computer-implemented method of clause 22, wherein theinformation describing the patient includes one or more of: a gender of the patient, an 1262-252WO01 / DIG-24-1695-IDF age of the patient, a glenoid type of a scapula of the patient, or a pathology of the anatomic structure.[ ] Clause 24. The computer-implemented method of any of clauses 1-23,wherein: the input tokens are a first set of input tokens, generating the input tokens comprises generating, by the one or more processors, the first set of input tokens and a second set of input tokens based on the input anatomic model, applying the GPT to the input tokens comprises applying, by the one or more processors, the GPT to the first set of input tokens and not the second set of input tokens.[ ] Clause 25. A computer-implemented method comprising: training, by one ormore processors, a machine learning (ML) model to: generate input tokens based on an input anatomic model that comprises a first 3-dimensional (3D) mesh representing a surface of at least a first portion of an anatomic structure of a patient; apply a generative pre-trained transformer (GPT) of the ML model to the input tokens to generate output tokens; and generate, based on the output tokens, a predicted anatomic model that includes a second 3D mesh representing a surface of at least a second portion of the anatomic structure.[ ] Clause 26. The computer-implemented method of clause 25, wherein: theML model includes a feature encoder, a quantization module, a feature decoder, and the GPT, training the ML model comprises: in a first training phase: for each first training example of a set of first training examples: applying, by the one or more processors, the feature encoder to an input anatomic model associated with the first training example to generate face embeddings associated with the first training example; applying, by the one or more processors, the quantization module to generate input tokens associated with the first training example based on the face embeddings associated with the first training example, and applying, by the one or more processors, the feature decoder to the input tokens associated with the first training example to generate a reconstructed anatomic model associated with the training example; applying, by the one or more processors, a first loss function to generate a loss value associated with the first training example based on the input anatomic model associated with the first training example and the reconstructed anatomic model associated with the first training example; performing, by the one or more processors, a first backpropagation process that modifies parameters of the feature encoder and the feature decoder based on the loss value associated with the first training example, and in a second training phase following the first training phase, for each second training example of a set of second 1262-252WO01 / DIG-24-1695-IDF training examples: applying, by the one or more processors, the feature encoder to generate second face embeddings associated with the second training example; applying, by the one or more processors, the quantization module to the face embeddings associated with the second training example to generate input tokens based on the second face embeddings; applying, by the one or more processors, the GPT to generate output tokens associated with the second training example based on the input tokens associated with the second training example; applying, by the one or more processors, the feature decoder to the input tokens associated with the second training example to generate a predicted anatomic model associated with the training example; applying, by the one or more processors, a second loss function to generate a loss value associated with the second training example; and performing, by the one or more processors, a second backpropagation process that modifies parameters of the GPT based on the loss value associated with the second training example.[ ] Clause 27. The computer-implemented method of clause 26, wherein theanatomic structure includes a scapula and the first loss function and the second loss function apply greater weight to areas close to a glenoid fossa of the scapula than other areas of the scapula.[ ] Clause 28. The computer-implemented method of any of clauses 26-27,wherein the first loss function and the second loss function are curvature-weighted chamfer loss functions.[ ] Clause 29. A system comprising: memory; and processing circuitrycommunicatively coupled to the memory, the processing circuitry configured to perform the method of any of clauses 1–28.[ ] Clause 30. A computer-readable storage medium comprising instructionsstored thereon that, when executed by one or more processors, cause the one or more processors to perform the method of any of clauses 1–28.[ ] Clause 31. A system comprising means for performing the method of any ofclauses 1–28.[ ] While the techniques been disclosed with respect to a limited number ofexamples, those skilled in the art, having the benefit of this disclosure, will appreciate numerous modifications and variations there from. For instance, it is contemplated that any reasonable combination of the described examples may be performed. It is intended that the appended claims cover such modifications and variations as fall within the true spirit and scope of the invention. 1262-252WO01 / DIG-24-1695-IDF[ ] It is to be recognized that depending on the example, certain acts or events of anyof the techniques described herein can be performed in a different sequence, may be added, merged, or left out altogether (e.g., not all described acts or events are necessary for the practice of the techniques). Moreover, in certain examples, acts or events may be performed concurrently, e.g., through multi-threaded processing, interrupt processing, or multiple processors, rather than sequentially.[ ] In one or more examples, the functions described may be implemented inhardware, software, firmware, or any combination thereof. If implemented in software, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium and executed by a hardware-based processing unit. Computer-readable media may include computer-readable storage media, which corresponds to a tangible medium such as data storage media, or communication media including any medium that facilitates transfer of a computer program from one place to another, e.g., according to a communication protocol. In this manner, computer-readable media generally may correspond to (1) tangible computer-readable storage media which is non-transitory or (2) a communication medium such as a signal or carrier wave. Data storage media may be any available media that can be accessed by one or more computers or one or more processors to retrieve instructions, code and / or data structures for implementation of the techniques described in this disclosure. A computer program product may include a computer-readable medium.[ ] By way of example, and not limitation, such computer-readable storage media cancomprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage, or other magnetic storage devices, flash memory, or any other medium that can be used to store desired program code in the form of instructions or data structures and that can be accessed by a computer. Also, any connection is properly termed a computer- readable medium. For example, if instructions are transmitted from a website, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of medium. It should be understood, however, that computer-readable storage media and data storage media do not include connections, carrier waves, signals, or other transitory media, but are instead directed to non-transitory, tangible storage media. Disk and disc, as used herein, includes compact disc (CD), laser disc, optical disc, digital versatile disc (DVD), floppy disk and 1262-252WO01 / DIG-24-1695-IDF Blu-ray disc, where disks usually reproduce data magnetically, while discs reproduce data optically with lasers. Combinations of the above should also be included within the scope of computer-readable media.[ ] Operations described in this disclosure may be performed by one or moreprocessors, which may be implemented as fixed-function processing circuits, programmable circuits, or combinations thereof, such as one or more digital signal processors (DSPs), general purpose microprocessors, application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), or other equivalent integrated or discrete logic circuitry. Fixed-function circuits refer to circuits that provide particular functionality and are preset on the operations that can be performed. Programmable circuits refer to circuits that can programmed to perform various tasks and provide flexible functionality in the operations that can be performed. For instance, programmable circuits may execute instructions specified by software or firmware that cause the programmable circuits to operate in the manner defined by instructions of the software or firmware. Fixed-function circuits may execute software instructions (e.g., to receive parameters or output parameters), but the types of operations that the fixed-function circuits perform are generally immutable. Accordingly, the terms “processor” and “processing circuity,” as used herein may refer to any of the foregoing structures or any other structure suitable for implementation of the techniques described herein.[ ] Various examples have been described. These and other examples are within thescope of the following claims.

Claims

1262-252WO01 / DIG-24-1695-IDF CLAIMS:

1. A computer-implemented method comprising: obtaining, by one or more processors, an input anatomic model that comprises a first 3-dimensional (3D) mesh representing a surface of at least a first portion of an anatomic structure of a patient; generating, by the one or more processors, input tokens based on the input anatomic model; applying, by the one or more processors, a generative pre-trained transformer (GPT) to the input tokens to generate output tokens; and generating, by the one or more processors, based on the output tokens, a predicted anatomic model that includes a second 3D mesh representing a surface of at least a second portion of the anatomic structure.

2. The computer-implemented method of claim 1, wherein: the first portion of the anatomic structure is a non-pathological portion of the anatomic structure, and obtaining the input anatomic model comprises modifying, by the one or more processors, a source anatomic model to generate the input anatomic model, wherein the source anatomic model includes the non-pathological portion and a pathological portion, and the input anatomic model excludes the pathological portion of the source anatomic model.

3. The computer-implemented method of claim 2, wherein: modifying the source anatomic model comprises: determining, by the one or more processors, a position of a geometric primitive shape relative to the source anatomic model; and determining, by the one or more processors, retained faces of the source anatomic model and discarded faces of the source anatomic model, the retained faces and the discarded faces being on different sides of a boundary defined by the geometric primitive shape.

4. The computer-implemented method of claim 3, wherein:1262-252WO01 / DIG-24-1695-IDF the method further comprises determining, by the one or more processors, positions of one or more anatomical landmarks on an anatomy of the patient, and determining the position of the geometric primitive shape comprises determining, by the one or more processors, the position of the geometric primitive shape relative to the source anatomic model based on the positions of the one or more anatomical landmarks.

5. The computer-implemented method of claim 3, wherein: the geometric primitive shape is a sphere, the anatomic structure is a scapula, and determining the position of the geometric primitive shape comprises: determining, by the one or more processors, a center of the sphere based on a center of a glenoid fossa of the scapula; and after determining the center of the sphere, adjusting, by the one or more processors, a position of the sphere medially.

6. The computer-implemented method of claim 3, wherein the geometric primitive shape is a plane through the source anatomic model. The computer-implemented method of claim 6, wherein: the anatomic structure includes a scapula, the plane is a sagittal plane.

8. The computer-implemented method of any of claims 2-7, wherein modifying the source anatomic model comprises: segmenting, by the one or more processors, the source anatomic model to identify the pathological portion of the source anatomic model and the non-pathological portion of the source anatomic model; and modifying, by the one or more processors, the source anatomic model to exclude the pathological portion from the input anatomic model.

9. The computer-implemented method of any of claims 1-6 or 8, wherein the anatomic structure includes one of a scapula, a humerus, a radius, an ulna, a vertebra, a1262-252WO01 / DIG-24-1695-IDF mandible, a femur, a tibia, a fibula, a talus, a cranium, a rib, a pelvis, a bone of a hand or foot.

10. The computer-implemented method of any of claims 1-9, wherein the second portion of the anatomic structure includes cartilage associated with the anatomic structure.

11. The computer-implemented method of any of claims 1-10, wherein the anatomic structure is a joint and the first portion of the anatomic structure includes portions of two or more bones of the joint.

12. The computer-implemented method of any of claims 1-11, wherein: the method further comprises obtaining, by the one or more processors, medical image data of the first portion of the anatomic structure, wherein the medical image data does not include the second portion of the anatomic structure; and obtaining the input anatomic model comprises generating, by the one or more processors, the input anatomic model based on the medical image data.

13. The computer-implemented method of claim 12, wherein the anatomic structure includes a scapula and the second portion of the anatomic structure is a medial or distal portion of the scapula.

14. The computer-implemented method of claim 12, wherein the anatomic structure includes a tibia, the first portion of the anatomic structure is a distal portion of the tibia, and the second portion of the anatomic structure is a proximal portion of the tibia.

15. The computer-implemented method of any of claims 1-14, wherein the predicted anatomic model includes faces of the input anatomic model of the first portion and faces of the second portion of the anatomic structure.

16. The computer-implemented method of claim 15, wherein: the method further comprises performing, by the one or more processors, a mesh decimation process on the input anatomic model to generate a decimated input anatomic model,1262-252WO01 / DIG-24-1695-IDF generating the input tokens based on the input anatomic model comprises generating the input tokens based on the decimated input anatomic model, and generating the predicted anatomic model comprises merging the faces of the input anatomic model of the first portion and the faces of the second portion of the anatomic structure.

17. The computer-implemented method of any of claims 1-16, wherein: the input tokens comprise quantized face embeddings for faces of the input anatomic model, applying the GPT comprises autoregressively applying, by the one or more processors, the GPT to a sequence of features to generate the output tokens, wherein the sequence of features includes the quantized face embeddings, and the output tokens comprise codebook indices for faces of the second portion of the anatomic structure, and generating the predicted anatomic model comprises: determining, by the one or more processors, a sequence of quantized face embeddings based on the codebook indices; and applying, by the one or more processors, a decoder to the sequence of quantized face embeddings to generate the predicted anatomic model.

18. The computer-implemented method of any of claims 1-17, wherein the first 3D mesh represents one or more osteophytes of the anatomic structure and the second 3D mesh does not represent the one or more osteophytes.

19. The computer-implemented method of any of claims 1-18, wherein obtaining the input anatomic model comprises: obtaining, by the one or more processors, medical image data of the anatomic structure after implantation of an orthopedic prosthesis on the anatomic structure; segmenting, by the one or more processors, the medical image data to identify artifacts in the medical image data attributable to the orthopedic prosthesis; and generating, by the one or more processors, the input anatomic model based on portions of the medical image data other than the artifacts.

20. The computer-implemented method of any of claims 1-19, further comprising:1262-252WO01 / DIG-24-1695-IDF obtaining, by the one or more processors, medical image data of the anatomic structure; and segmenting, by the one or more processors, the medical image data based on the predicted anatomic model.

21. The computer-implemented method of any of claims 1-20, wherein: the input anatomic model represents a current state of the anatomic structure with an orthopedic prosthesis implanted thereon, and the predicted anatomic model represents a state of the anatomic structure prior to implantation of the orthopedic prosthesis on the anatomic structure.

22. The computer-implemented method of any of claims 1-21, wherein: the input tokens are first input tokens, applying the GPT comprises applying, by the one or more processors, the GPT to the first input tokens and second input tokens to generate the output tokens, and the second input tokens include information describing the patient and not describing the first 3D mesh.

23. The computer-implemented method of claim 22, wherein the information describing the patient includes one or more of: a gender of the patient, an age of the patient, a glenoid type of a scapula of the patient, or a pathology of the anatomic structure.

24. The computer-implemented method of any of claims 1-23, wherein: the input tokens are a first set of input tokens, generating the input tokens comprises generating, by the one or more processors, the first set of input tokens and a second set of input tokens based on the input anatomic model, applying the GPT to the input tokens comprises applying, by the one or more processors, the GPT to the first set of input tokens and not the second set of input tokens.

25. A computer-implemented method comprising: training, by one or more processors, a machine learning (ML) model to:1262-252WO01 / DIG-24-1695-IDF generate input tokens based on an input anatomic model that comprises a first 3-dimensional (3D) mesh representing a surface of at least a first portion of an anatomic structure of a patient; apply a generative pre-trained transformer (GPT) of the ML model to the input tokens to generate output tokens; and generate, based on the output tokens, a predicted anatomic model that includes a second 3D mesh representing a surface of at least a second portion of the anatomic structure.

26. The computer-implemented method of claim 25, wherein: the ML model includes a feature encoder, a quantization module, a feature decoder, and the GPT, training the ML model comprises: in a first training phase: for each first training example of a set of first training examples: applying, by the one or more processors, the feature encoder to an input anatomic model associated with the first training example to generate face embeddings associated with the first training example; applying, by the one or more processors, the quantization module to generate input tokens associated with the first training example based on the face embeddings associated with the first training example, and applying, by the one or more processors, the feature decoder to the input tokens associated with the first training example to generate a reconstructed anatomic model associated with the training example; applying, by the one or more processors, a first loss function to generate a loss value associated with the first training example based on the input anatomic model associated with the first training example and the reconstructed anatomic model associated with the first training example; performing, by the one or more processors, a first backpropagation process that modifies parameters of the feature encoder and the feature decoder based on the loss value associated with the first training example, and1262-252WO01 / DIG-24-1695-IDF in a second training phase following the first training phase, for each second training example of a set of second training examples: applying, by the one or more processors, the feature encoder to generate second face embeddings associated with the second training example; applying, by the one or more processors, the quantization module to the face embeddings associated with the second training example to generate input tokens based on the second face embeddings; applying, by the one or more processors, the GPT to generate output tokens associated with the second training example based on the input tokens associated with the second training example; applying, by the one or more processors, the feature decoder to the input tokens associated with the second training example to generate a predicted anatomic model associated with the training example; applying, by the one or more processors, a second loss function to generate a loss value associated with the second training example; and performing, by the one or more processors, a second backpropagation process that modifies parameters of the GPT based on the loss value associated with the second training example.

27. The computer-implemented method of claim 26, wherein the anatomic structure includes a scapula and the first loss function and the second loss function apply greater weight to areas close to a glenoid fossa of the scapula than other areas of the scapula.

28. The computer-implemented method of any of claims 26-27, wherein the first loss function and the second loss function are curvature-weighted chamfer loss functions.

29. A system comprising: memory; and processing circuitry communicatively coupled to the memory, the processing circuitry configured to perform the method of any of claims 1–28.1262-252WO01 / DIG-24-1695-IDF 30. A computer-readable storage medium comprising instructions stored thereon that, when executed by one or more processors, cause the one or more processors to perform the method of any of claims 1–28.

31. A system comprising means for performing the method of any of claims 1–28.

Images