Improved method and system for the automatic processing of bone images for orthopaedic applications, in particular for determining the personalized implant based on groups of bones defects

Abstract

The present invention relates to a computer-implemented method for clustering defects of defective bones and manufacturing a patient-specific orthopaedic implant, including at least the following steps: - acquiring computed tomography CT images of a defective or damaged bone and feeding a database; - given a set of defective bones retrieved from said defective or damaged bone database, reconstructing a synthetic shape of the original shape of the defective bone; - computing a bone erosion depth map of said synthetic shape of the original shape of the defective bone, where said bone erosion depth map, for a given defective bone, comprises a set of erosion values over the region of interest of the reconstructed bone, defining an erosion vector; - transforming said erosion depth map into an erosion matrix which comprises the erosion vectors of the defective bones and the vertices over the surface of the region of interest of said synthetic shape of the original shape of the defective bone; - feeding said erosion matrix to an unsupervised machine learning algorithm that groups data into a tree of nested clusters, to identify clusters of relevant bone defects and to generate synthetic three-dimensional bone models representative of each defect cluster; - associating a defect label to each cluster; - training a defect classifier using said erosion matrix and said defect label to classify clusters of defects of the bones; - feeding a CAD environment with the outputs of said defect classifier; - creating the orthopaedic implant based on the outputs fed to the CAD environment.

Description

[0001] Title: Improved method and system for the automatic processing of bone images for orthopaedic applications, in particular for determining the personalized implant based on groups of bones defects

[0002] DESCRIPTION

[0003] CROSS-REFERENCE TO RELATED APPLICATIONS

[0004] This application claims priority to and benefit of Italian patent application no. 102024000008995 filed on 19.04.2024, Italian patent application no. 102024000009004 filed on 19.04.2024, Italian patent application no. 102024000009007 filed on 19.04.2024, Italian patent application no. 102024000009010 filed on 19.04.2024, and Italian patent application no. 102024000009013 filed on 19.04.2024, which are all incorporated by reference herein.

[0005] Field of Invention

[0006] This invention relates to an improved method for the automatic processing of computed tomography (CT) images and three-dimensional bone models enabling the following orthopaedic-related applications:

[0007] - bone morphometric analyses aimed at improving the design and manufacturing of orthopaedic implants enhancing bone-implant fitting and biomechanical performances;

[0008] - implant positioning to test the bone fitting of several implant sizes and shapes leading to a more effective implant design;

[0009] - bone-implant interaction analyses to improve the primary stability of the implants;

[0010] - reconstructions of severely damaged bones to enhance and streamline the design and manufacturing of personalized implants.

[0011] The present disclosure further relates to an integrated system for the automatic processing of bone images for orthopaedic applications and, more specifically, for designing and manufacturing orthopaedic prostheses.

[0012] In particular, the present disclosure concerns a method for determining the personalized implant based on groups of bones defects. Background Art

[0013] In this specific technical field, it is known to take advantage of the virtual reconstruction of bone shapes to help surgeons during their interventions.

[0014] For example, a Statistical Shape Model (SSM) can be used to virtually reconstruct a glenoid bone defect and to predict the inclination, version, and center position of the native glenoid.

[0015] In some instances, the SSM may be created on the basis of a plurality of healthy scapulae. Then, artificial bone defects may be created in all scapulae and reconstructed using the SSM-based reconstruction method. For each bone defect, the reconstructed surface may be compared with the original surface. Furthermore, the inclination, version, and glenoid center point of the reconstructed surface may be compared with the original parameters of each scapula.

[0016] This solution has been used for small glenoid bone defects. However, this SSM-based reconstruction method, while being able to accurately reconstruct the native glenoid surface, is limited to the preoperative planning of shoulder arthroplasty.

[0017] As another example, a 3D glenoid vault model may be as a template to predict normal glenoid version. Computed tomography scans of both shoulders may be obtained in subjects with unilateral glenohumeral osteoarthritis.

[0018] Custom-developed graphic software may be used to create a 3-D reconstruction of each scapula. Measurement differences between the glenoid and vault model may be analyzed by repeated-measures analysis of variance. That technique may be used in correcting the pathologic glenoid version due to arthritis.

[0019] Even this technique, however, is limited to the study of the pathologic evolution of the scapula but does not allow obtaining data for implementing orthopaedic-related applications such as the designing and manufacturing of shoulder prosthesis. As another example, an image of a patient’s shoulder can be segmented to develop a 3D shoulder model. Virtual surgery can be performed on the 3D shoulder model to generate a modified shoulder model. The virtual surgery can include resecting and reaming a virtual humerus of the 3D shoulder model. Moreover, a virtual representation of the humeral implant can be implanted on the virtual humerus and a virtual representation of the glenoid implant on the virtual glenoid to virtually update the modified shoulder model.

[0020] Based on the above, it is clear that a possible assessment of glenoid bone shape would be important to select the optimal glenoid component design during shoulder arthroplasty planning and implantation.

[0021] However, all the aforementioned existing methodologies are based on the availability of a 3D healthy glenoid model but in the real-life situations, the surgeons have to face the problem that the shoulder joint can undergo degenerative changes caused by various issues, such as rheumatoid arthritis, osteoarthritis, Cuff tear arthropathy, vascular necrosis, or bone fracture.

[0022] During a shoulder arthroplasty surgery, the components of the prosthesis are matched with the anatomy of the patient in an effort to maintain or restore a natural range of motion of a healthy shoulder joint. Patient specific instrumentation can assist a surgeon in planning and implementing a shoulder arthroplasty to restore natural movement. However, even with the multitude of advances in prosthetic components and patient specific instrumentation, restoring a full range of motion can remain difficult. In some cases, the range of motion of a patient following a successful procedure is limited more than is desirable to some patients.

[0023] There do exist some methods to locate and measure surgically relevant anatomic features and propagate these measurements to different populations using a programmable data processing system, and methods for implant design primarily centered on aligning existing or prototype implants with three-dimensional bone data. However, such methods rely on establishing point correspondences among populations of bones, computing principal components, and evaluating implant fit through virtual placement and range-of-motion simulations. The focus lies on iteratively refining implant geometry based on matching the implant to the averaged or principal modes of variation of a given population, with emphasis on virtual resection, surgical landmark detection, and comparisons of different implant designs across multiple bones. These methods address mainly healthy bones, and do not address a method for reconstructing severely eroded bones. These methods provide a robust suite for analyzing healthy bones or known implant prototypes, computing point correspondences, and running virtual fitting or kinematic tests on a range of designs. With these methods, implants are assessed across a population to iteratively modify generic designs or prototypes, producing a broad morphological matching. In this case, implants are assessed across a population to iteratively modify generic designs or prototypes.

[0024] There also exist methods in which a proposed implant fits across the breadth of an anatomical population, building an SSM by leveraging PCA from healthy bones, then sampling that model in order to locate the region of shape space that yields an acceptable fit. A specialized level- set approach in PCA space constrains the analysis to shapes that meet a specified error threshold, helping identify how design modifications might expand the percentage of bones well-served by the implant. These methods implement the implant-fitting task primarily around geometric correspondence in shape space, without incorporating bone-density information or finite element analyses. Furthermore, the statistical shape model is derived from a cohort of healthy tibiae, allowing an evaluation of implant shape fit across an assumed healthy range.

[0025] There also exist methods that focus on analyzing the variation in proximal femoral canal shape in patients with primary hip osteoarthritis (OA). Using radiographic images and Active Shape Modeling (ASM), major shape modes may be identified, and these major shape modes may be clustered into distinct morphological categories (10 main clusters). The goal is to highlight how anatomical differences in the proximal femur may influence hip arthroplasty component design. These methods focus on a single anatomical region (i.e., the proximal femur) in primary hip OA, using 2D radiographs supplemented by some 3D measurements (e.g., CT-based femoral anteversion). The entire pipeline aims to capture and classify femoral canal shape for hip arthroplasty considerations. Furthermore, these methods rely on ASM to cluster femurs based on morphological differences (e.g., canal flare, neck- shaft angle). After identifying shape modes, they perform hierarchical clustering to segment patients into distinct “types” of femurs. These methods primarily focus on how these shape clusters correlate with simple demographic or geometrical parameters (offset, neck- shaft angle, cortical flare indices). These methods are also limited to the femoral canal only.

[0026] There also exist methods which focuses on the development of a deconvolution algorithm for the automatic mapping of the cortical bone thickness and density.

[0027] There also exist methods which focus on an automated osteoporosis diagnosis from X-ray images by focusing primarily on texture-based descriptors (both appearance-based and feature-based methods). Two key approaches — one using Principal Component Analysis (PCA), the other employing Linear Discriminant Analysis (LDA) — provide extracted features, which are subsequently classified by different machine learning models (kNN, Naive Bayes, and SVM).

[0028] There also exist methods which describe a two-stage pipeline specifically designed to perform medical image segmentation in clinical settings. The training phase combines a level-set segmentation algorithm (driven by a “pathologically modeled” energy functional) with window-based feature extraction and principal component analysis (PCA). A support vector machine (SVM) is then trained on those extracted features. During the clinical segmentation stage, the SVM directly classifies new images into the modeled regions.

[0029] Summary of the Invention

[0030] A scope of the present disclosure is that of enabling a virtual reconstruction of real eroded glenoid vaults for the final purpose of manufacturing of orthopaedic prosthesis.

[0031] Another purpose of the present disclosure is that of providing a new method and system for the automatic processing of computed tomography (CT) images and three-dimensional bone models further enabling orthopaedic- related applications in an automated manner, fully automated or in part for some specific applications.

[0032] The present disclosure is focused on a method describing a glenoid bone loss using 3-dimensional measurements but without the need for a healthy contralateral reference scapula.

[0033] It is understood that the present disclosure can also be used, with appropriate modifications, for different anatomical compartments.

[0034] A first embodiment of the present disclosure relates to a method for the automatic processing of computed tomography (CT) images and three- dimensional bone models starting from: acquisition of several patient’s data from anonymized non- pathological CT scan images; further storage of patient data performed in parallel with the storage of three-dimensional bone models and calibration data; the method further including at least the following phases: performing morphometric analyses of bones by automatically detecting anatomical landmarks and regions of interest by means of machine learning techniques; feeding a CAD environment with the outputs of said bone models, detected landmarks and predicted regions of interest for designing and manufacturing an orthopaedic implant on the basis of said at least one shape thus improving bone-implant fitting and biomechanical performances.

[0035] Advantageously, the phase of storing the three-dimensional bone models includes the storing of outer cortical models of various bones, such as at least one bone of the following group: calcaneus, clavicle, femur, fibula, hip, humerus, navicular, patella, radius, scapula, talus, tibia and ulna.

[0036] Moreover, the outer cortical models are obtained through an automatic three-dimensional outer cortical segmentation performed by an algorithm capable of viewing and interacting with 3D anatomical models comparing 3D models with medical image data through a contour overlay of the model on the images.

[0037] Further, the phase of storing the three-dimensional bone models includes the storing of inner cortical models of various bones, such as at least one bone of the following group: femur, humerus, scapula, tibia and ulna.

[0038] More specifically, the inner cortical models include the boundary of the relative bone computed by means of a deconvolution algorithm.

[0039] It should be noted that the mentioned calibration data are obtained by a calibration phantom phase converting Hounsfield units from the CT images to bone mineral density; wherein the Hounsfield units or CT numbers are used as quantitative scale for describing radiodensity when a bone mineral density phantom was scanned alongside the patient.

[0040] It is remarkable that the storage phase is automatically updated once a new CT image is available.

[0041] What’s more, the availability of the non-rigidly registered bones allows computing a statistical shape model (SSM), using principal component analysis, to capture their variation in shape .

[0042] In this content, a plurality of synthetic shapes is drawn from the statistical shape model (SSM) and are used as training dataset for a supervised machine learning algorithm for point cloud segmentation, i.e., for automatically detecting regions of interest of an unseen shape.

[0043] Moreover, a non-rigid registration of 3D bone models includes creating at least one shape in correspondence of each bone model starting from a reference shape which is rigidly registered onto the other shape enabling the automatic detection of anatomical landmarks.

[0044] The above processes enable at least the following improvement over the known solution: bone morphometric analysis; implant positioning;

[0045] - bone-implant interaction analysis;

[0046] - more efficient bone reconstruction;

[0047] - personalized orthopaedic implant.

[0048] Advantageously, the system comprises two foundational subprocesses:

[0049] • The storage of patient data, three-dimensional bone models and calibration data (calibration of CT images, i.e. conversion of Hounsfield units to bone mineral density).

[0050] • Automatic anatomical landmarks and regions of interest detection.

[0051] The outputs of the above subprocesses, for instance outer and inner cortical segmentations, landmarks, machine learning and statistical shape models, etc., enable the execution of morphometric analyses of bones, automatic positioning of implants, analyses of the bone-implant interaction in terms of micromotions and potential risk of bone resorption, reconstruction of bones and personalization of implants.

[0052] As an example, the native shape of the glenoid is reconstructed by fitting a statistical shape model (SSM) of the scapula. The total vault loss percentage, local vault loss percentages, defect depth, defect area percentage, and subluxation distance and region are computed based on a comparison of the reconstructed and eroded glenoids.

[0053] Further features and advantages of the method and system according to the present disclosure will appear from the following description of at least an implementing embodiment given by way of non-limiting illustrative example with reference to the enclosed drawings wherein:

[0054] Brief Description of the Drawings

[0055] - Fig. 1 is a schematic illustrative view of both a method and a system according to the present disclosure for the automatic processing of bones images for orthopaedic applications; Fig. 2 is a schematic view of a flow chart showing as flow of method steps a subprocess of the method of Figure 1 bringing to an update of the system of the present disclosure;

[0056] - Fig. 3 is a schematic view of a flow chart showing as flow of steps a further subprocess of the method of Figure 1 bringing to an update of the system of the present disclosure;

[0057] - Fig. 4 is another schematic view of a flow chart showing as flow of steps a further subprocess of the method of Figure 1 bringing to an update of the system of the present disclosure;

[0058] - Fig. 5 is a further schematic view of a flow chart showing as flow of steps a further subprocess of the method of Figure 1 bringing to an update of the system of the present disclosure;

[0059] - Fig. 6 is a further schematic view of a flow chart showing as flow of steps a further subprocess of the method of Figure 1 bringing to an update of the system of the present disclosure;

[0060] - Fig. 7 is another schematic view of a flow chart showing as flow of steps a further subprocess of the method of Figure 1 bringing to an update of the system of the present disclosure;

[0061] - Fig. 8 is another schematic view of a flow chart showing as flow of steps a further subprocess of the method of Figure 1 bringing to an update of the system of the present disclosure;

[0062] - Fig. 9 is another schematic view of a flow chart showing as flow of steps a further subprocess of the method of Figure 1 bringing to an update of the system of the present disclosure;

[0063] - Fig. 10 is another schematic view of a flow chart showing as flow of steps a further subprocess of the method of Figure 1 bringing to an update of the system of the present disclosure.

[0064] - Fig. 11 is a graphical representation of the step of clustering the defects of a bone;

[0065] - Fig. 12 is a graphical representation of the step of the ray casting algorithm; - Fig. 13 is a graphical representation of the step of the reconstruction of a defective bone, where the white regions show the defective bone, the grey regions show the reconstructed bone and the crosses, axes and coordinate systems are the reference features.

[0066] - Fig. 14 is a schematic view of the modules of the system capable of carrying out the method.

[0067] Detailed Description

[0068] A comprehensive method and system for the automatic processing of bone images for orthopaedic applications will be disclosed hereinafter with reference to the above reported list of figures.

[0069] The method according to the present disclosure is based on a group of phases that can be performed sequentially or in parallel.

[0070] The main phases will be disclosed hereinafter with reference to enclosed drawings.

[0071] We will start considering the acquisition of several patient’s data from anonymized non-pathological CT scan images, measured in Hounsfield units or CT numbers that are used in this field as quantitative scale for describing radiodensity when a bone mineral density phantom was scanned alongside the patient.

[0072] For a better understanding of the following disclosure, it should be remarked that a computed tomography scan, also named CT scan or computed axial tomography scan, is a medical imaging technique used to obtain detailed internal images of a body.

[0073] This technique has been originally developed in the 1970s and since that date CT scanning has proven to be a versatile imaging technique.

[0074] Generally speaking, CT scanners use a rotating X-ray tube and a row of detectors placed in a gantry to measure X-ray attenuations by different tissues inside the body. The multiple X-ray measurements taken from different angles are then processed on a computer using tomographic reconstruction algorithms to produce tomographic (i.e.: cross-sectional) images that appear like virtual "slices" of a body or body component. According to the specific kind of image acquisition and procedures various types of scanners are available on the market.

[0075] The result of a CT scan is a volume of voxels (i.e.: values on a regular grid in a three-dimensional space), which may be presented to a human observer by various methods, for instance: slices of varying thickness; projection including maximum or average intensity projection; or volume rendering (VR) that may become projections when viewed on a 2 -dimensional display.

[0076] Moreover, epitomes of volume rendering models feature a mix of for example coloring and shading in order to create realistic and observable representations.

[0077] Keeping in mind the availability of the above-mentioned medical images, the method of the present disclosure comprises at least a couple of fundamental subprocesses:

[0078] • the storage of patient data performed in parallel with the storage of three- dimensional bone models and calibration data, for instance: calibration of CT images (i.e., conversion of Hounsfield units to bone mineral density);

[0079] • the automatic detection of anatomical landmarks and regions of interest performed through machine learning techniques that are integrated into the system of the present disclosure.

[0080] We will see hereinafter the steps or passages bringing to the results of the above mentioned subprocesses.

[0081] At the end, the outputs of the above subprocesses enable morphometric analyses of bones, automatic positioning of implants, analyses of the boneimplant interaction in terms of micromotions and potential risk of bone resorption, reconstruction of bones and personalization of implants.

[0082] Let’s start from a process flow chart 20 shown in the schematic view of Figure 2 and representing the first subprocess.

[0083] A process flow 200 is started with an extraction step of anonymized non- pathological CT images. Those CT images are available in a single database or in a group of independent databases that may be identified to pick-up or select patient data.

[0084] For instance, a database of this kind is represented by images collected with time and made available by medical centres, according to specific agreements.

[0085] The step 200 is implemented by three independent and parallel process phases shown by three legs or flow of steps 201, 202, 203 reported in Figure 2 a corresponding list of actions.

[0086] A main flow 201 evolves from a first automatic outer cortical segmentation 210 performed on the available CT images to a second automatic inner cortical segmentation 220 of the same CT images.

[0087] First of all, the available CT images are stored.

[0088] Secondly, three-dimensional outer cortical models of various bones are stored as well: calcaneus, clavicle, femur, fibula, hip, humerus, navicular, patella, radius, scapula, talus, tibia and ulna. The segmentation of the outer cortical model is performed by a specific algorithm.

[0089] Moreover, three-dimensional inner cortical models of femur, humerus, scapula, tibia and ulna are stored. The inner cortical boundary is computed using a deconvolution algorithm.

[0090] The first automatic three-dimensional outer cortical segmentation is preferably performed by a tool for engineers and physicians (i.e. an algorithm) capable of viewing and interacting with 3D anatomical models.

[0091] A tool of this kind is for instance a 3D medical image segmentation software. Several software are available on the market, by way of example only, not by way of limitation: Materialise Mimics Viewer, Thermo ScientificTM Amira, 3D Slicer, MedSeg, and others.

[0092] This tool enables for instance a receiver to compare 3D models with a medical image data through a contour overlay of the model on the images. The second automatic three-dimensional inner cortical segmentation is preferably performed by another software tool using a deconvolution algorithm based on the cortical bone mapping (CBM) technique. CBM allows to map both the cortical bone thickness and density.

[0093] A manual check phase 230 and refinement might be necessary for checking the quality of the automatic segmentations and refine the output, provided by a 3D medical image segmentation software, if needed; however, a skilled person in this art understands that such a phase is absolutely optional.

[0094] At the end of this main segmentation process 201 several 3D bone models are obtained in the step 240.

[0095] Those bone models are ready for mass storage in a specific database indicated by the numeral 250.

[0096] The storage in the database 250 is implemented through a data collecting node 245 further receiving the respective output of the process flows 202 and 203.

[0097] As to process flow 202, it evolves through an extraction phase 218 of calibration data always taken from the available anonymized non- pathological CT Images of step 200. Calibration data are retrieved from a calibration phantom to convert Hounsfield units to bone mineral density.

[0098] More specifically, the phase 218 extracts Hounsfield units (HU) from the CT images. The scale of the Hounsfield units is a linear transformation and a quantitative scale for describing radiodensity. It is frequently used in CT scans, where its value is also termed CT number.

[0099] The extraction phase 218 allows obtaining a bone mineral density conversion equation if the CT images are calibrated, that is to say: if a bone mineral density phantom was scanned alongside the patient.

[0100] The output results of this phase 218 are calibration data 228 available for a subsequent storage in the database 250 through the data collecting node 245. Similarly, the process flow 203 evolves through an extraction phase 215 of patient data always taken from the available anonymized non-pathological CT Images of step 200.

[0101] The data extracted may be at least: age, gender, ethnicity, height and weight, even if the list of these data is not exhaustive. Additional data are stored as well: name and country of the facilities where CT images were taken and slice thickness of the CT images.

[0102] The output results of this phase 215 are patient data 225 available for a subsequent storage in the database 250 through the data collecting node 245.

[0103] At the end of the first subprocess 20 we have obtained a specific database 250 of non-pathological CT images, patient data and bone model data.

[0104] Differently from known solutions, the database 250 is populated in an automatic manner and the data storage is automatically updated once new CT images are available.

[0105] Moreover, the storage of calibration data and the computation and storage of the inner cortical boundary is a new and essential feature of the system and model of the present disclosure to enrich the patient data available for the subsequent process phases.

[0106] This database 250 will be used and updated by other phases or subprocesses of the method according to the present disclosure, as we will see hereinafter.

[0107] More particularly, as shown in Figure 3, the database 250 including non- pathological CT images, patient data and bone model data is used in a process flow or subprocess 30 for the automatic detection of anatomical landmarks and regions of interest.

[0108] As to the landmark detection, two independent data sources are necessary and picked up from the database 250.

[0109] First, three-dimensional bone models (or shapes) of a specific anatomy (e.g. femur) are used after a download step in block 240. In parallel, a source of reference bone (also denoted as shape) is obtained in step 340.

[0110] A rigid registration of bones is performed in step 310.

[0111] This step is implemented by rigidly registering the bones either by aligning their principal axes of inertia to a common coordinate system or, as an alternative, by using the Iterative Closest Point algorithm.

[0112] Different techniques for the creation of 3D models allow their greater application in medical training and biomedical sciences. Recently developed digital technologies have revolutionized anatomy education through 3D visual reconstruction of the human body.

[0113] The implementation and application of 3D models can be divided into three main categories: visualization, simulation, and physical replication to obtain a functioning model.

[0114] In relation to the simulation and physical replication aspects, the density and accuracy of the polygonal mesh are more important factors than the resolution, or presence of the texture map. Data demonstrate that the microcomputed tomography technique is the most preferable way to obtain high density polygonal mesh, as well as this technique produces the most accurate results in terms of topology. Another crucial aspect is the presence of inner structures that are impossible to capture using any other technique.

[0115] It is important to note that the first requirement for the creation of an anatomically accurate 3D model is a volumetric dataset. Most models are created from computed tomography (CT) data, but it is possible to create models from datasets with modalities such as magnetic resonance imaging (MRI) and ultrasound.

[0116] Access to virtual images can offer a unique possibility for acquiring information about normal anatomy and / or anatomical variations. Differentiation of tissue density (from low to high) allows the segmentation technique to be used for different anatomical structures. Subsequently, a non-rigid registration of 3D bone models is performed in step 320.

[0117] This further step 320 is implemented creating shapes in correspondence of each bone model starting from the reference shape 340 which is non-rigidly registered onto the other shapes using a Gaussian process-based algorithm.

[0118] The result of these two operative steps 310 and 320 is a set of shapes having the same number of nodes which are mapped to the nodes of the reference shape enabling the automatic detection of anatomical landmarks (e.g. tip of the lesser trochanter) in step 350.

[0119] The selection of these landmarks onto the reference shape is manually performed once by an experienced orthopaedic surgeon. This selection is indicated by the block 360.

[0120] The result of the non-rigid registration of bones of the step 320 is also used in the step 40, to perform an automatic anatomical region of interest detection, which runs in parallel with the step 350 and independently from each other.

[0121] As to the region of interest detection, the availability of the non-rigidly registered shapes obtained in step 320 allows computing a statistical shape model (SSM) using principal component analysis.

[0122] Statistical Shape Models (SSMs) are geometric models that describe a collection of semantically similar objects in a very compact way. SSMs capture the variation in shape of many three-dimensional objects.

[0123] The creation of SSMs requires a correspondence mapping, which can be achieved e.g. by parameterization with a respective sampling. If a corresponding parameterization over all shapes can be established, variation between individual shape characteristics can be mathematically investigated.

[0124] In the specific application here disclosed the regions of interest (for instance a femoral neck) are manually labelled once onto the reference shape. Then, several synthetic shapes are drawn from the SSM and serve as training dataset for a supervised machine learning algorithm for point cloud segmentation. The trained model allows to automatically detect regions of interest in an unseen shape.

[0125] The result of the whole process flow 40 is an automatic anatomical regions of interest detection using a combined statistical shape modelling- supervised machine learning methodology.

[0126] Let’s start with Figure 4 wherein it is shown the schematic flow chart 40 of the method for calculating the anatomical regions.

[0127] Starting from the non-rigidly registered bones of the step 320, a statistical shape model (SSM) is computed by leveraging the principal component analysis, as shown in step 410. The statistical shape model output 411 is then stored in the model database 450. The statistical shape model can be retrieved to accomplish, for instance, bone reconstruction as per process flow 80.

[0128] Anatomical regions of interest are manually labelled once on the reference bone by an experienced orthopaedic surgeon, in step 440 and fed to the block 420 where regions of interest are automatically labelled over the synthetic bones drawn from the SSM.

[0129] The reconstructed bone is generated from a bone-specific SSM, e.g., the scapula SSM; the hip SSM and similar. The synthetic bone shapes drawn from the SSM and their corresponding automatically labelled anatomical regions of interest serve as input for the supervised machine learning model training 460.

[0130] A supervised machine learning is a subcategory of machine learning and artificial intelligence. As it is known, a machine learning is a branch of artificial intelligence (Al) and computer science which focuses on the use of data and algorithms to imitate the way that humans learn.

[0131] In other words, the machine learning is the use and development of computer systems that are able to learn and adapt without following explicit instructions, for instance by using algorithms and statistical models to analyze and draw inferences from patterns in data.

[0132] In this context, the supervised machine learning implemented in the present disclosure may be defined as the use of labeled datasets to train algorithms that take care of classifying data or predict outcomes accurately.

[0133] As input data is fed into the model, it adjusts its weights until the model has been fitted appropriately, which occurs as part of a cross-validation process. Supervised learning helps organizations solve a variety of real- world problems at scale, such as classifying spam in a separate folder from your inbox.

[0134] Well, in the disclosed example the phase 460 may be considered a paradigm in a machine learning procedure wherein input objects (for example, a vector of predictor variables) and a desired output value (also known as human-labeled supervisory signal train a model.

[0135] The training data is processed, building a function that maps new data on expected output values. An optimal scenario will allow for the algorithm to correctly determine output values for unseen instances. This requires the learning algorithm to generalize from the training data to unseen situations in a "reasonable" way (also known as inductive bias). This statistical quality of an algorithm is measured through the so-called generalization error.

[0136] For completeness’s sake, and for a better understanding of the present disclosure, the phase 460 may be considered including a series of substeps, for instance: a first decision step for determining the type of training examples wherein the user should decide what kind of data is to be used as a training set.

[0137] The training set needs to be representative of the real-world use of the function. Thus, a set of input objects is gathered and corresponding outputs are also gathered, either from human experts or from measurements.

[0138] Moreover, the input feature representation of the learned function must be determined. The accuracy of the learned function depends strongly on how the input object is represented. Typically, the input object is transformed into a feature vector, which contains a number of features that are descriptive of the object. The number of features should not be too large; but should contain enough information to accurately predict the output.

[0139] Then, the learning algorithm is executed on the gathered training set. Some supervised learning algorithms require the user to determine certain control parameters. These parameters may be adjusted by optimizing performance on a subset of the training set, or via a cross-validation.

[0140] Finally, the accuracy of the learned function is evaluated. After parameter adjustment and learning, the performance of the resulting function should be measured on a test set that is separate from the training set.

[0141] Even if the above list of steps may appear complex, a wide range of supervised learning algorithms are nowadays available.

[0142] The end of this sub-process 40 is represented by the phase 470 of Figure 4 wherein the supervised machine learning model is fed to the database 450 of SSM and ML models.

[0143] With reference to the drawing Figure 5, the process flow or subprocess or module 50 allows an automatic morphometric analysis of bones.

[0144] The flow starts with three inputs: the three-dimensional bone model, either retrieved from database 250 or reconstructed from a defective bone as per process flow 80, automatically detected landmarks computed using the non-rigid registration technique as per phase 350 and the anatomical regions of interest 530 predicted by the supervised machine learning model 470. The combined inputs are addressed through a node 540 to a computational phase wherein a computation of anatomical planes, axes, cross-sections, angles are performed as well as other measurements of interest.

[0145] The results of this computational phase 540 may represent the input of CAD environment 560 wherein a specific CAD software is capable to provide an aided implant design.

[0146] An implant may be created based on the results of the computational phase 540 that are provided to the CAD environment 560. The implant can be created manually or by 3D printing techniques, such as Electron Beam Melting - EBM or Selective Laser Malting - SLM.

[0147] But the results of the computational phase 540 are also used as input to generate spreadsheets for data analyses in block 570.

[0148] The output block 570 can be also leveraged as a corpus of data to ingest into a pre-trained Large Language Model (LLM) for a faster and easier retrieval of morphometric data via text prompts eliminating the timeconsuming process of manual searches in the spreadsheets depicted in the block 570. The user can interact with the LLM through a user-friendly graphical user interface by entering queries in natural language and receive accurate, contextually relevant answers, thus streamlining data access and simplifying the interaction with complex bone morphometric datasets.

[0149] Let’s now consider the other flows of the method of the present disclosure with reference to the Figure 6 and following.

[0150] Automatic implant positioning

[0151] With reference to the drawing Figure 6, the process flow or subprocess or module 60 allows to calculate an automatic implant positioning, in particular an anatomy specific optimization-based positioning of implants.

[0152] Given a three-dimensional bone model 240, either retrieved from database 250 or reconstructed from a defective bone as per process flow 80, and corresponding anatomical features 600, such as landmarks, regions of interests, planes and axes, etc., the optimal size, position and orientation of a three-dimensional implant model are computed using an optimization algorithm 610, while satisfying one or more constraints such as the boneimplant interpenetration. Typical objective functions are but not limited to, the implant coverage over the bone resection or the volumetric filling ratio, depending on the specific implant. The optimization algorithms 610 are provided by open source libraries and are both gradient-based or not depending, for instance, on the nature of the objective functions.

[0153] The automatic positioning of a single implant requires just few seconds, thus enabling the processing of a large set of implants in a short period of time. The efficiency of the positioning is achieved by reducing the dimensionality of the problem from three to two-dimensions. The implantbone interpenetration is analysed by computing cross-sections of both the outer or inner cortical bone and the implant; cross-sections are then converted to polygons allowing the optimization algorithms to run faster. This procedure can be extended to any bone-implant pair. The two- dimensional approach for implant-bone interpenetration can be used to assess the optimal shape or size of implant features, such as pegs or stems.

[0154] The output of the optimization algorithms 610 are the optimal values of the position and orientation of the implant 611 and the optimal size of the implant 612 which allow to position and orient the implant with respect to the bone. The optimal size 612 is used as input for the implant database 630, where the implants are stored. The output of the implant database 630 and the optimal size of the implant 612 are used to realize the bone-implant assembly 620, which is the assembly available in the CAD.

[0155] The bone-implant assembly 620 is checked in a step 640 by the user, such as an engineer or a surgeon and manually improved if necessary.

[0156] The bone-implant assembly 620 is stored in the local memory of the computer of the user.

[0157] The module 60 can be used for different kind of implants, such as shoulders, knee and / or hip.

[0158] The module 60 can then be integrated into a CAD (Computer Aided Design) software, to make virtual surgeries less time consuming and scalable.

[0159] Based on the automatic implant positioning, the implant may be positioned within a patient, such as during a surgical procedure. The soft tissues are incised, any previous implants are removed, a fixation is performed and the implant is connected to any other functioning implants, already fixed in the body of the patient.

[0160] Bone-implant interaction analysis

[0161] With reference to the Figure 7, the process flow or subprocess or module 70 allows to assign density and elastic modulus values to the bone model. Given a calibrated CT image set and the meshed model of the bone-implant assembly 620, the quantitative computed tomography - QCT methodology applied to the finite element method allows to assign density and elastic modulus values to the bone model.

[0162] In particular, the quantitative computed tomography - QCT is a medical technique that measures bone mineral density - BMD using a standard X- ray Computed Tomography - CT scanner with a calibration standard to convert the Hounsfield Units - HU of the CT image to bone mineral density values. QCT scans are primarily used to evaluate all bones mineral density.

[0163] In particular, according to the prior art, the quantitative computed tomography (QCT) is a medical technique for measuring bone mineral density (BMD) in the axial spine and peripheral skeleton.

[0164] In contrast to other densitometric techniques, QCT allows to separate between cortical and trabecular bone. In particular, with respect to the dual-energy X-ray absorptiometry (DXA), which provides an areal density (aBMD) measured in g / cm2, the BMD measured using QCT is a true density measured in g / cm3. Local BMD values, estimated by the QCT methodology, can be mapped in subject-specific finite element models (FEA) to accurately predict, for instance, strain distribution in the bone and fracture risk under physiological loading.

[0165] The development of patient-specific QCT / FEA models requires two relationships to assess bone mechanical properties. The first equation converts raw CT attenuation data to BMD. In the context of synchronous calibration, a device called phantom is scanned simultaneously with the patient and the CT scanner response to the region of interests (ROIs) of the phantom is measured. Calibration phantoms consist of inserts (from two to six) having varying concentrations of calcium hydroxyapatite (HA) or dipotassium phosphate (K2HPO4) and provide equivalent BMD expressed in units of mgHA / cm3or mgK2HPO4 / cm3. Other calibration approaches, namely the asynchronous and the phantom-less calibration methodologies are available, however the synchronous one is the most used and validated methodology. The second relationship allows to convert BMD values to bone mechanical properties, such as elastic modulus, and is experimentally derived.

[0166] Starting from the database 250, including non-pathological CT images, patient data and bone model, CT images are taken from, in a step 720.

[0167] In general, solid phantoms placed in a pad under the patient during CT image acquisition are used for calibration. These phantoms contain materials that represent a number of different equivalent bone mineral densities. For example, usually either calcium hydroxyapatite (CaHAP) or potassium phosphate (K2HPO4) are used as the reference standard. Hence, it is necessary to check if the data are calibrated.

[0168] Therefore, in a parallel step 730 a check on the availability of the calibration of the data is performed. If the check has a negative outcome, a step 740 is performed, therefore an equivalent set of calibration data is computed using existing calibration data stored in the patient and bone model database. On the other side, if the check has a positive outcome, next step 750 is performed.

[0169] On the other side, the bone-implant assembly 620 are also fed to the step 750.

[0170] In the step 750, the quantitative computed tomography - QCT is performed, to assign bone material properties.

[0171] Further, a Finite Element Analysis - FEA is performed in step 760, using a third-party software, obtaining FEA results 770, which is at least a text file.

[0172] For sake of clarity, the FEA is the simulation of any given physical phenomenon using the numerical technique called the Finite Element Method - FEM. Generally, Partial Differential Equations are used to understand and quantify any physical phenomena.

[0173] The results from the finite analysis, performed with a third-party software, are processed to compute the primary stability of the implant, in terms of micromotions, performed in step 780 and the potential for bone remodelling and risk of bone resorption, performed in step 790. In particular, the text file coming as output from step 770, is fed to a graphical interface in step 780 which computes the micromotions occurring in the bone-implant interface; micromotions are computed as the boneimplant relative displacements at the contact interface, according to known art techniques.

[0174] Moreover, the text file coming as output from step 770, is fed to a graphical interface in step 790 which computes the potential for bone remodelling and risk of bone resorption are assessed by computing the strain energy density changes between the bone-implant assembly and the intact bone model, according to known art techniques.

[0175] Therefore, is evident that an advantage of the process 70 is the computation of the micromotions and risk of bone resorption, in particular the combination of the QCT performed in step 750 and the FEA performed in the step 760 allows this kind of computation.

[0176] Bone reconstruction

[0177] With reference to the Figure 8, the process flow or subprocess or module 80 implements a statistical-based methodology for the reconstruction of bones. Taking a scapula as an example, the reference scapula is rigidly registered to the defective one and corresponding points are computed and fed to a statistical shape model to compute a synthetic shape which represents the premorbid scapula. The methodology can be extended to any other bone.

[0178] The module 80 can then be integrated into a CAD (Computer Aided Design) software, for easy-to-use reconstruction of defective bones.

[0179] Starting from the model database 450 including SSM and / or ML models, constructed through a Statistical Shape Model computation, a Statistical Shape Model - SSM computation 410 is performed.

[0180] Starting from the defective bones database 810, a defective bone 811 is chosen. The defective bones database 810 comprises CT images of unhealthy bones, comprising deformations or traumas, of patients which need personalized implants. The database 810 has been built over years. Starting from the database 250, including non-pathological CT images, patient data and bone model, a reference bone 251 is chosen.

[0181] The collecting node 820 receives the defective bone 811 and the reference bone 251, in order to produce a rigid registration 830.

[0182] In a step 840, hundreds of corresponding points on the reference and defective bones are identified and allow to compute a deformation field. These correspondences define how each selected point of the statistical model should be deformed to its corresponding point on the defective bone (SSM-based reconstruction 860).

[0183] The collecting node 850 receives the reference-defective bone point mapping 840, the Statistical Shape Model 410 in order to perform the step of SSM- based defective bones reconstruction 860, ensuring anatomically realistic representations and preserving statistical variability, having a reconstructed bone 870, which is a synthetic shape which represents the premorbid bone. Considering a scapula, the system also returns to the user the values of the inclination and version angles of the glenoid and the width and height of the glenoid. Moreover, the following geometrical entities are generated as well and the user can interact with them through the CAD software, for instance: the Friedman line, the infero-superior and anteroposterior axes, the landmarks defining the glenoid rim, the glenoid center, the landmarks corresponding to the "inferior angle" and "trigonum spinae" and finally the scapular and glenoid reference systems.

[0184] Clustering of defects and implant personalization

[0185] With reference to the Figures 9 and 10, the process flow or subprocess or module 90 is a bone defect clustering methodology, performed in order to obtain an implant personalization, for example a glenoid implant personalization, based on classification of bone defects, in order to gather the bone defects in clusters. It allows an implant personalization based on the defect clustering.

[0186] The principal process flow 90 is shown in Figure 9; the flow F shown in Figure 10 is a standing alone method which is used to obtain the defect classifier 952. Starting with Figure 9, the defective bones database 810 is used to extract a single defective bone 811 which is reconstructed in step 860, in order to obtain a reconstructed bone 870; as an example, a large sample of defective scapulae is reconstructed in step 870.

[0187] Before going on with the description of the process 90, it is necessary describing the process F to obtain a defect classifier 952, making reference to the Figure 10.

[0188] In step 910, bone erosion depth map is computed, for example glenoid erosion depth maps are computed in step 910, using a known ray-casting algorithm, for example a 3D STL model. In particular, for each vertex of the region of interest of the reconstructed bone, a ray is traced parallel to a clinically relevant direction (e.g., the normal of the glenoid plane for the scapula; a radius of the best-fit sphere modelling the acetabulum and passing through the centre of this sphere for the hip) which intersects the surface of the defective bone, according to the ray-casting approach. The distance along the ray between the vertex of the reconstructed bone and the intersection point with the surface of the defective scapula provides an estimate of the bone erosion.

[0189] For a given defective bone, the set of erosion values over the region of interest defines an erosion vector.

[0190] All the erosion maps are transformed into a matrix in step 920 of Figure 10 and fed to an unsupervised machine learning algorithm to find clusters of defects in step 930. In particular, by stacking together the erosion vectors of TV defective bones, an Nx M erosion matrix is obtained in step 920, where M represents the number of vertices over the surface of the region of interest of the reconstructed bone. Thus, the z-th row of the erosion matrix represents the z-th defective bone and its j-th column represents the j-th vertex over the surface of the region of interest of the reconstructed bone. As already disclosed, the reconstructed bone is generated from a bonespecific SSM, consequently, all reconstructed bones of the same type, derived from the same SSM, have the same number of vertices, and vertices across different reconstructed bones, stemming from the same SSM, are in correspondence. Then, hierarchical clustering, an unsupervised machine learning (ML) algorithm that groups data into a tree of nested clusters, is leveraged to discover bone defect clusters, in step 930. First, the erosion matrix is pre-processed by scaling its values using a min-max algorithm. Then its dimensionality is reduced by using the Principal Component Analysis technique by imposing the preservation of a pre-defined amount of variance (e.g., 95%). The pre-processed erosion matrix is fed to the aforementioned hierarchical clustering algorithm, and a similarity metric is chosen (e.g., correlation, cosine similarity). Results of the hierarchical clustering are displayed in a dendrogram diagram showing a nested hierarchy of clusters at different levels of similarity.

[0191] The analysis, of the dendrogram, combined with expert knowledge 940, allows to identify a number of clinically relevant glenoid defects and to generate synthetic three-dimensional bone models representative of each defect cluster.

[0192] In particular, the determination of the optimal number of clinically relevant clusters, which translates to the determination of the optimal threshold where to cut the dendrogram, is based on a hybrid approach combining expert knowledge 940 with data-driven cluster estimation methods. Specifically, the hierarchical clustering is performed iteratively on progressively expanded datasets, starting with 500 bones and increasing in increments (e.g., 1000, 1500, etc.). The process is continued until a plateau is observed, where further dataset expansion no longer leads to the formation of new meaningful clusters, i.e. clinically relevant bone defects, indicating convergence. Once clusters are defined, to each defective bone in the database is assigned a label corresponding to its defect (e.g., posterior glenoid erosion for the scapula), thus the N x M erosion matrix is paired with a Nx 1 cluster label vector, in step 950.

[0193] Therefore, the output of the clustering with expert feedback methodology are N defect labels (for instance, “anterior” defect label) associated to each defective scapula.

[0194] Then, the erosion matrix 920 and the cluster label vectors 950 are used to train a defect classifier which is a supervised machine learning algorithm 951 which is able to assign labels to never seen scapulas. To reduce overfitting during training, the training dataset is augmented with synthetic erosion maps, using Gaussian copula-based augmentation, enabling the generation of synthetic defects while preserving the statistical distribution of the real bone dataset. Hence, the trained bone defect classifier 952 allows to automatically compute the defect labels of an unseen defective scapula.

[0195] In particular, the erosion matrix 920 and the cluster label vectors 950 are split into train, validation and test datasets. The train and validation datasets are fed to a supervised machine learning algorithm 951, specifically a Random Forest classifier, which, in contrast to a neural network, provides better interpretability allowing to better understand which features contribute most to classification, and allows faster training and inference time. The train dataset is also augmented by creating synthetic defects, i.e. new rows of the erosion matrix are generated, by leveraging the Gaussian copulas which allows to learn the overall distribution of the erosion matrix, train a model and generate new synthetic rows. The augmented dataset is used to train the bone defect classifier 952. Then, the bone defect classifier 952 is stored in the SSM / ML model database 450 to be later retrieved for inference tasks. As shown in Fig. 9, given a new unseen defective bone, its reconstruction is computed in the reconstruction step 860 and the bone erosion is assessed in step 910, i.e. its erosion vector is computed, then pre-processed (scaling and dimensionality reduction), and finally fed to the bone defect classifier 952 to predict the bone defect label in step 955.

[0196] Turning back to the description of the process 90, with reference to the Figure 9, from the N labels, N parameterized defect-specific bone implants are designed and stored in a database 960.

[0197] Once the defect label is available, a parametric orthopedic implant is picked in step 970 from a database of parametric defect-specific implants 960 whose parameters are automatically tuned tuned to fit the size and shape of the defective glenoid, in general are tuned to the defect-specific implant in step 980 to obtain a patient / defect-specific implant 990. These parameters include for instance the length of a central peg for a glenoid implant, or the stem direction of a custom iliac stem for an acetabular implant. Given for example an unseen damaged scapula, its defect label is predicted in step 955 and the corresponding glenoid implant is selected in step 970, from the database of data- specific implants 960.

[0198] Finally, the implant parameters are tuned to fit the size and shape of the defective glenoid, in general are tuned to the defect- specific implant, in step 980, in order to obtain the patient- specific implant 990.

[0199] The presented method is able to automatically analyse bone defects using 3 -dimensional measurements without the need for a healthy contralateral bone.

[0200] System

[0201] Figure 14 shows a system 1400 used to perform any of the methods described herein. In one or more embodiments, the system 1400 may include one or more user devices 1402 (which may be associated with one or more users 1401), one or more imaging devices 1405, one or more computing systems 1406, and / or one or more databases 1412. However, these components of the system 1400 are merely exemplary and are not intended to be limiting in any way. For simplicity, reference may be made hereinafter to, user device 1402, imaging device 1405, computing device 1406, database 1412, etc., however, this is not intended to be limiting and may still refer to any number of such elements. For example, reference to a computing device 1406 may also refer to any other number of computing device 1406. Likewise, reference to computing devices 1406 may also refer to a single computing device 1406.

[0202] The user device 1402 may be any type of device, such as an e-reader, personal assistant device, speaker, gaming console, smartphone, desktop computer, laptop computer, tablet, smart television (for example, a television with Internet connectivity, the capability to install applications, etc.), and / or any other type of device. A display device 1403 (for example, a computer monitor, a television, etc.) may also be included in the system 1403 that may present information to the user 1401.

[0203] The imaging device 1405 may be any type of device used to capture images of a patient (or parts of a patient, such as bones), as described herein. The computing device 1406 may be a “backend” system (however, the computing device 1406 may also be a local device), such as a server (or other type of device with processing capabilities) that performs many or all of the processing tasks described herein. For example, the computing device 1406 may include one or more processor(s) 1407, memory 1408, and / or any other components used to perform the processing tasks described herein. Any of the other elements of the system may also include such components as well (for example, the user device 1402 may also include one or more processors and memory, etc.).

[0204] While reference is made to processing primarily being performed on the computing device 1406, it should be noted that some of the processing may be shared with the user device 1402 as well. That is, some of the processing may be performed locally by the user device 1402 (or all of the processing may be performed locally by the user device 1402, in some instances). The processing may also be performed by (or shared with) any other device, system, etc.

[0205] The processor 1407 (or processors) may include any suitable processing unit capable of accepting digital data as input, processing the input data based on stored computer-executable instructions, and generating output data. The computer-executable instructions may be stored, for example, in the data storage and may include, among other things, operating system software and application software. The computer-executable instructions may be retrieved from the data storage and loaded into the memory 1408 as needed for execution. The processor may be configured to execute the computer-executable instructions to cause various operations to be performed. Each processor may include any type of processing unit including, but not limited to, a central processing unit, a microprocessor, a microcontroller, a Reduced Instruction Set Computer (RISC) microprocessor, a Complex Instruction Set Computer (CISC) microprocessor, an Application Specific Integrated Circuit (ASIC), a System- on-a-Chip (SoC), a field-programmable gate array (FPGA), and so forth.

[0206] The computing device 1406 (or any other computing device) may also include data storage. The data storage may store program instructions that are loadable and executable by the processor 1407, as well as data manipulated and generated by one or more of the processor 1407 during execution of the program instructions. The program instructions may be loaded into the memory 1408 as needed for execution. The memory 1408 may be volatile memory (memory that is not configured to retain stored information when not supplied with power) such as random access memory (RAM) and / or non-volatile memory (memory that is configured to retain stored information even when not supplied with power) such as read-only memory (ROM), flash memory, and so forth. In various implementations, the memory 1408 may include multiple different types of memory, such as various forms of static random access memory (SRAM), various forms of dynamic random access memory (DRAM), unalterable ROM, and / or writeable variants of ROM such as electrically erasable programmable readonly memory (EEPROM), flash memory, and so forth.

[0207] Various program modules, application may be stored in data storage that may comprise computer-executable instructions that when executed by one or more of the processor 1407 cause various operations to be performed. The memory 1408 may have loaded from the data storage one or more operating systems (O / S) that may provide an interface between other application software (for example dedicated applications, a browser application, a web-based application, a distributed client-server application, etc.) executing on the server and the hardware resources of the server. More specifically, the O / S may include a set of computer-executable instructions for managing the hardware resources of the server and for providing common services to other application programs (for example managing memory allocation among various application programs). The O / S may include any operating system now known or which may be developed in the future including, but not limited to, any mobile operating system, desktop or laptop operating system, mainframe operating system, or any other proprietary or open-source operating system.

[0208] In one or more embodiments, any of the elements of the system 1400 (e.g., the user device 1402, imaging device 1405, computing device 1406, database 1403, etc.) may be configured to communicate via a communications network 1450. The communications network 1450 may include, but not limited to, any one of a combination of different types of suitable wired and / or wireless communications networks such as, for example, broadcasting networks, cable networks, public networks (e.g., the Internet), private networks, wireless networks, cellular networks, or any other suitable private and / or public networks. Further, the communications network 1450 may have any suitable communication range associated therewith and may include, for example, global networks (e.g., the Internet), metropolitan area networks (MANs), wide area networks (WANs), local area networks (LANs), or personal area networks (PANs). In addition, communications network 1450 may include any type of medium over which network traffic may be carried including, but not limited to, coaxial cable, twisted-pair wire, optical fiber, a hybrid fiber coaxial (HFC) medium, microwave terrestrial transceivers, radio frequency communication mediums, white space communication mediums, ultra- high frequency communication mediums, satellite communication mediums, or any combination thereof.

[0209] Following from the above description and invention summaries, it should be apparent to those of ordinary skill in the art that, while the methods and apparatuses herein described constitute exemplary embodiments of the present invention, the invention contained herein is not limited to this precise embodiment and that changes may be made to such embodiments without departing from the scope of the invention as defined by the claims. Additionally, it is to be understood that the invention is defined by the claims and it is not intended that any limitations or elements describing the exemplary embodiments set forth herein are to be incorporated into the interpretation of any claim element unless such limitation or element is explicitly stated. Likewise, it is to be understood that it is not necessary to meet any or all of the identified advantages or objects of the invention disclosed herein in order to fall within the scope of any claims, since the invention is defined by the claims and since inherent and / or unforeseen advantages of the present invention may exist even though they may not have been explicitly discussed herein.

Claims

CLAIMS1. A computer-implemented method (90) for clustering defects of defective bones and manufacturing a patient-specific orthopaedic implant, including at least the following steps:- acquiring computed tomography CT images of a defective or damaged bone (810, 811) and feeding a database;- given a set of defective bones retrieved from said defective or damaged bone database (810, 811), reconstructing a synthetic shape of the original shape of the defective bone (860);- computing a bone erosion depth map (910) of said synthetic shape of the original shape of the defective bone (860), where said bone erosion depth map (910), for a given defective bone, comprises a set of erosion values over the region of interest of the reconstructed bone, defining an erosion vector;- transforming said erosion depth map (910) into an erosion matrix (920) which comprises the erosion vectors of the defective bones and the vertices over the surface of the region of interest of said synthetic shape of the original shape of the defective bone (860);- feeding said erosion matrix (920) to an unsupervised machine learning algorithm (930) that groups data into a tree of nested clusters, to identify clusters of relevant bone defects and to generate synthetic three- dimensional bone models representative of each defect cluster;- associating a defect label (950) to each cluster;- training a defect classifier (951, 952) using said erosion matrix (920) and said defect label (950) to classify clusters of defects of the bones;- feeding a CAD environment (50) with the outputs of said defect classifier (951, 952);- creating the orthopaedic implant based on the outputs fed to the CAD environment.

2. Method according to claim 1, further comprising the steps of:- predicting a defect label (955) of an unseen defective bone by said defect classifier (951, 952);- given said defect label (955), selecting a corresponding bone implant (970), from a database of defect-specific implants (960);- tuning (980) the implant parameters to fit the size and shape of the defectspecific implant (960) in order to obtain the patient-specific implant (990).

3. Method according to anyone of claims 1 or 2, wherein the computing bone erosion depth maps (910) are also computed using a ray-casting algorithm.

4. Method according to anyone of claims 1-3, wherein said unsupervised machine learning algorithm (930) is an hierarchical clustering algorithm whose results are displayed in a dendrogram diagram showing a nested hierarchy of clusters at different levels of similarity, capable to identify a number of clinically relevant bone defects and to generate synthetic three- dimensional bone models representative of each defect cluster.

5. Method according to anyone of claims 2-4, wherein said step of assigning a label to said cluster of defects (930) is aided with the validation of an expert in step (940).

6. Method according to anyone of claims 1-5, wherein said defect classifier (951, 952) is a supervised machine learning model.

7. Method according to anyone of claims 1-6, wherein said defective bone is extracted from a database (810) comprising CT images of unhealthy bones, comprising deformations or traumas, of patients which need personalized implants.

8. Data processing apparatus comprising means for carrying out the steps of the method of anyone of the preceding claims 1-7.

9. A computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the method of anyone of the preceding claims 1-7.

10. System (1400) capable of carrying out the method (90) according to any one of the preceding claims 1-7 comprising at least one user device (1402), at least imaging device (1405), at least one computing system (1406) and at least one databases (1412), configured to communicate with each other via at least one communications network (1450).

Images