Techniques for assessment of musculoskeletal health
The integration of AI models with CT and MRI data for bone and muscle health assessment addresses the limitations of DXA, offering comprehensive and efficient musculoskeletal evaluations.
Patent Information
- Application Number
- PCT/EP2025/062475
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-05-07
- Publication Date
- 2026-11-12
AI Technical Summary
Current diagnostic methods, such as DXA, fail to provide a comprehensive assessment of bone and muscle health, particularly in terms of bone microarchitecture and muscle quality, leading to incomplete fracture risk evaluation and operational inefficiencies in clinical settings.
A computed-implemented method using artificial intelligence models, trained on imaging data from CT and MRI, integrates Hounsfield units, bone microstructural indices, and muscle quality indices to determine bone fragility and muscle quality, enabling a comprehensive musculoskeletal assessment.
This approach enhances diagnostic efficiency by providing real-time, integrated assessments of bone and muscle health, reducing the need for separate imaging sessions and improving pre-surgical planning and clinical decision-making.
Smart Images

Figure EP2025062475_12112026_PF_FP_ABST
Abstract
Description
DESCRIPTIONTitle: Techniques for assessment of musculoskeletal healthTechnical field
[0001] The present invention relates to techniques for evaluating musculoskeletal health, including bone and muscle quality, using imaging data, such as for instance computed tomography, x-ray, and magnetic resonance imaging.
[0002] The invention applies to various clinical settings, including fracture liaison services, pre-surgical planning, trauma care, routine diagnostics, and opportunistic screening using routine imaging studies.Background
[0003] Osteoporosis, a systemic skeletal bone fragility disorder characterized by reduced bone mineral density (BMD) and impaired bone microarchitecture, remains a leading cause of fractures among older adults and at-risk populations such as postmenopausal women. Current diagnostic approaches, particularly dual-energy x-ray absorptiometry (DXA), primarily focus on measuring BMD. However, these techniques are inherently limited in their ability to provide a comprehensive assessment of bone health.
[0004] Beyond osteoporosis, the evaluation of bone quality, a broader concept encompassing microarchitecture, material properties, and mechanical competence, has emerged as a critical focus in fields such as orthopaedics, spinal neurosurgery, sports medicine, and geriatric care. Bone quality is particularly relevant in contexts requiring a robust assessment of skeletal integrity, such as pre-surgical planning for orthopaedic implants, joint replacement, or spinal surgeries. Optimizing bone health pre-surgery can significantly enhance surgical outcomes by reducing the risk of implant failure and improving postoperative recovery.
[0005] Similarly, the evaluation of muscle quality has gained attention as a key factor in musculoskeletal health. Sarcopenia, characterized by reduced muscle mass and function, often coexists with osteoporosis, forming the condition known as osteosarcopenia. Muscle health directly impacts bone health through mechanical loading, which influences bone density and strength. Reduced muscle quality, particularly from ectopic fat infiltration, contributes to frailty, increased fall risk, and poor functional outcomes in aging populations. Computed tomography (CT) enables precise quantification of muscle cross-sectional area (CSA) and the degree of ectopic fat infiltration, providing a detailed assessment of muscle quality. Conventional diagnostic methods, such as DXA, primarily focus on bone mass and fail to capture the critical muscle-bone dynamics essential for assessing overall fracture risk and functional capacity.
[0006] Current limitations in bone and muscle health assessment also extend to understanding localized adaptations in athletes, individuals undergoing high-impact physical activity, or patients with conditions such as cancer-related bone metastases. Traditional imaging modalities such as DXA or standard radiographs do not provide adequate insights into the structural and functional aspects of bone and muscle quality critical for such applications.Description of the invention
[0007] A purpose of the present invention is to provide a new diagnostic technique for assessing and / or monitoring bone and muscle health in all individuals, regardless of their age, sex, or health status. In particular, the invention aims to improve early detection of bone and muscle disorders, as well as their management and care. The invention also aims to reduce the risk of fractures, surgical complications and other negative consequences associated with bone and muscle disorders.
[0008] In a non-limiting way, the invention also aims to address one or more of the following challenges with DXA and existing techniques:- Incomplete assessment of bone fragility: While BMD is a critical determinant of bone strength, it does not capture trabecular microarchitecture, a key factor in fragility. Consequently, many individuals at risk of fractures are not identified using DXA alone. - TBS, derived from 2-dimensional DXA images, offers additional insights into bone microarchitecture and complements DXA-BMD for bone health evaluation. However, both BMD and TBS remain reliant on DXA imaging and separate processing, adding complexity and delays to the diagnostic workflow.- Limitations in special populations: In spine surgery patients or individuals with degenerative changes such as osteophytes, vertebral fractures, or sclerotic endplates, DXA-based BMD often overestimates bone quality. This misrepresentation can lead to inappropriate surgical planning or treatment delays.- Patients with advanced spinal pathologies often require immediate evaluation, where DXA is not readily accessible. Furthermore, sarcopenia in these populations exacerbates surgical risks and recovery challenges, necessitating integrated assessments of muscle and bone health.- Incomplete assessment of muscle quality: Traditional imaging techniques, including DXA, do not measure muscle quality or ectopic fat infiltration. This gap leads to a limited understanding of the functional implications of sarcopenia and its contribution to fracture risk.- Operational and logistical barriers: DXA imaging requires dedicated sessions separate from other diagnostic workflows. In pre-surgical and urgent care settings, this introduces delays, complicates workflows, and may limit timely decision-making. Similarly, muscle health assessments are rarely integrated into routine diagnostic workflows, leaving a significant gap in comprehensive musculoskeletal care.
[0009] A first aspect of the invention relates to the training of an artificial intelligence model as described below.
[0010] According to said first aspect, the invention covers a computed-implemented method for training an artificial intelligence model, comprising:- obtaining imaging data representing at least one region of the body of one or more individuals,- determining one or more musculoskeletal parameters from said imaging data, - forming ground truth data from one or more of said musculoskeletal parameters, - generating one or more training datasets that comprise said ground truth data and one or more of said imaging data,- training the artificial intelligence model using one or more of said training datasets.
[0011] Said one or more musculoskeletal parameters are preferably chosen from a list including Hounsfield units, one or more bone microstructural index parameters, and one or more muscle quality index parameters.
[0012] Of course, the above list is not limitative and may also include other parameters.
[0013] In a preferred embodiment, the step for determining one or more musculoskeletal parameters comprises determining a bone fragility index parameter using said Hounsfield units and at least one of said bone microstructural index parameters.
[0014] According to this preferred embodiment, the ground truth data may comprise said bone fragility index parameter.
[0015] In other words, said bone fragility index could typically be a composite parameter designed to reflect the combined contribution of bone quantity and bone quality. For example, the bone fragility index may integrate a bone quantity index typically assessed via Hounsfield units from computed tomography scan, and a bone quality index that can be reflected by one or more bone microstructural index parameters, such as trabecular bone score or microarchitectural measures derived from image processing of high-resolution modalities.
[0016] In one embodiment, the bone fragility index is computed by categorizing both bone quantity (e.g., Hounsfield units) and bone quality (e.g., one or more bone microstructural index parameters) into strata - such as tertiles or clinical interpretation thresholds - and mapping theircombination to predefined bone fragility index levels, for example "Normal", "Medium", "High", or "Very High". Such a method enables a person skilled in the art to determine bone fragility index from known, measurable quantities.
[0017] In an embodiment, said imaging data comprise imaging data acquired using computed tomography scan and / or imaging data acquired using magnetic resonance imaging.
[0018] Such imaging data may be retrieved from a data storage system, such as a "picture archiving and communication system" (PACS).
[0019] In one embodiment, concerning classification of bone quantity, Hounsfield units (HU) values are derived from routine computed tomography imaging. Bone quantity categories may be determined as follows:- Manufacturer-derived thresholds, e.g.:■ Normal: HU > 150;■ Osteopenia: HU between 100 and 150;■ Osteoporosis: HU < 100;- Data-driven thresholds (tertiles): These can be derived from a representative cohort used to train or validate the model. For example:■Tl: HU < 110;■T2: 110 < HU < 140;■T3: HU > 140.
[0020] Of course, the above thresholds are modifiable based on the cohort's demographics or anatomical region.
[0021] In one embodiment, concerning classification of bone quality:- Using trabecular bone score (TBS):■TBS, derived from DXA images, reflects bone microarchitecture. Clinically validated thresholds can typically be:• Normal: TBS > 1.32;• Partially Degraded: 1.23 < TBS < 1.32;• Degraded: TBS < 1.23;- Using other microarchitectural indices (e.g., from QCT or HR-pQCT): Bone microstructural index parameters (BMSI) may also include quantitative structural parameters such as:■Trabecular Thickness (Tb.Th) - typically in mm (e.g., normal values "'0.12-0.18 mm);■ Trabecular Number (Tb.N) - in 1 / mm (e.g., ~1.2-2.0 / mm);■ Trabecular Separation (Tb.Sp) - in mm (e.g., ~0.3-0.5 mm);■ Bone Volume Fraction (BV / TV) - dimensionless (e.g., ~0.12-0.24);■ Structure Model Index (SMI) - dimensionless (lower is better, e.g., 0-2);■ Connectivity Density (Conn.D) - in mm-3;Each of the above parameters can be stratified into tertiles or risk categories based on normative data from population cohorts;- Deriving a Combined BMSI: If multiple microarchitecture parameters are used, an adaptive BMSI score may be obtained from available standardized parameters using training reference data, or a composite BMSI from linear combination such as:BMSIadaptive=with zi representing all available microarchitecture parameters;BMSIcombined= o * TBS + oc2* Tb. Th + oc3* Tb. N + — F ocn* Conn. D, where the weights oq can be derived from empirically or via regression techniques from training data.
[0022] In one embodiment, the bone fragility index is obtained by cross-classifying HU and BMSI into a matrix.
[0023] Said matrix may be adapted to the number and types of BMSI used. If only TBS is used, then a table as exemplified below can apply directly. If other microstructural parameters are included, BMSI categories can be defined accordingly (e.g., tertiles of the composite score).
[0024] In an embodiment, the step for determining one or more musculoskeletal parameters comprises calculating one or more experimental variograms from said imaging data.
[0025] As known per se in the art, for example from the document W02023051909, experimental variograms provide texture metrics allowing, in a non-limitative way, to quantify directly trabecular structure, and to capture critical details such as trabecular spacing, orientation, and connectivity.
[0026] The teaching of document W02023051909 can be used to deduce experimental variogram calculations and derive a trabecular bone score. Its adaption to 3D may uses the same global formula with adaptation of the directional vector to 3D:with< :sin(0) cos(<p)U9,<Pk sin(0) sin(<p) directional vector on a sphere;cos(0)9 e [0, n], polar angle (Z-axis);<p e [0, 2n], azimuthal angle (in the X-Y-plane);h(Pi) intensity or value at voxel Pi;r: radius (distance).
[0027] The step for determining one or more musculoskeletal parameters may comprise determining at least one of said musculoskeletal parameters, in particular a trabecular bone score, using one or more of the calculated experimental variograms.
[0028] Said trabecular bone score typically form one of said bone microstructural index parameters.
[0029] Said trabecular bone score may be used to form the ground truth data.
[0030] Although not limited to this type of architecture, the artificial intelligence model described above with reference to said first aspect of the invention may typically comprise a convolutional neural network.
[0031] The invention also covers a data processing apparatus comprising means for carrying out a method according to said first aspect.
[0032] The invention also covers a computer program comprising executable instructions which, when the instructions are executed by a computer, cause the computer to carry out a method according to said first aspect.
[0033] The invention also covers a non-transitory computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out a method according to said first aspect.
[0034] A second aspect of the invention relates to the use of an artificial intelligence model previously trained according to said first aspect of the invention.
[0035] According to said second aspect, the invention more specifically covers a computed-implemented method for determining one or more musculoskeletal parameters from imaging data representing at least one region of the body of at least one individual, the method comprising a step for providing the imaging data as input to an artificial intelligence model trained with a method according to the above described first aspect of the invention, an output of the artificial intelligence model forming said one or more musculoskeletal parameters.
[0036] Said musculoskeletal parameters are preferably chosen from a list including Hounsfield units, one or more bone microstructural index parameters, one or more bone fragility index parameters, and one or more muscle quality index parameters.
[0037] In an embodiment, the imaging data used in the method according to said second aspect comprise imaging data acquired using x-ray imaging.
[0038] Such imaging data may be retrieved from a data storage system (e.g., PACS).
[0039] Although not limited to this type of architecture, the artificial intelligence model described above with reference to said second aspect may typically comprise a convolutional neural network.
[0040] The invention also covers a data processing apparatus comprising means for carrying out a method according to said second aspect.
[0041] The invention also covers a computer program comprising executable instructions which, when the instructions are executed by a computer, cause the computer to carry out a method according to said second aspect.
[0042] The invention also covers a non-transitory computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out a method according to said second aspect.
[0043] One or more of the data processing apparatus, computer program, and computer-readable medium described with reference to said second aspect may be the same as, part of, or other than, those described above with reference to said first aspect.
[0044] A third aspect of the invention relates to the training of an artificial intelligence model as described below.
[0045] According to said third aspect, the invention covers a computed-implemented method for training an artificial intelligence model, comprising:- obtaining one or more musculoskeletal parameters and one or more musculoskeletal outcomes related to one or more individuals,- forming ground truth data from one or more of said musculoskeletal outcomes, - generating one or more training datasets that comprise said ground truth data and one or more of said musculoskeletal parameters,- training the artificial intelligence model using one or more of said training datasets.
[0046] Within the scope of said third aspect, said one or more musculoskeletal parameters are preferably chosen from a list including Hounsfield units, one or more bone microstructural index parameters, and one or more muscle quality index parameters.
[0047] Said one or more musculoskeletal outcomes may typically be representative of a musculoskeletal risk associated with one or more of said individuals.
[0048] Although not limited to this type of architecture, the artificial intelligence model described above with reference to said third aspect may typically comprise a machine learning model.
[0049] The invention also covers a data processing apparatus comprising means for carrying out a method according to said third aspect.
[0050] The invention also covers a computer program comprising executable instructions which, when the instructions are executed by a computer, cause the computer to carry out a method according to said third aspect.
[0051] The invention also covers a non-transitory computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out a method according to said third aspect.
[0052] One or more of the data processing apparatus, computer program, and computer-readable medium described with reference to said third aspect may be the same as, part of, or other than, those described above with reference to said first aspect and / or said second aspect.
[0053] A fourth aspect of the invention relates to the use of an artificial intelligence model previously trained according to said third aspect of the invention.
[0054] According to said fourth aspect, the invention covers a computed-implemented method for determining one or more musculoskeletal outcomes related to at least one individual, the method comprising:- obtaining input data comprising one or more musculoskeletal parameters related to the at least one individual,- providing said input data to an artificial intelligence model trained with a method according to said third aspect, an output of the artificial intelligence model forming one or more of said musculoskeletal outcomes.
[0055] Within the scope of said fourth aspect, said musculoskeletal outcomes may typically be representative of a musculoskeletal risk associated with the individual.
[0056] In an embodiment, said input data further comprise clinical risk factors data related to the at least one individual.
[0057] In a first implementation of the method according to said fourth aspect, said musculoskeletal parameters used in the method according to said fourth aspect comprise musculoskeletal parameters previously determined using a method according to said second aspect which implements a first artificial intelligence model, the artificial intelligence model used in the method according to said fourth aspect being a second artificial intelligence model.
[0058] In a second implementation of the method according to said fourth aspect, said input data used in the method according to said fourth aspect comprise one or more musculoskeletal parameters from imaging data representing at least one region of the body of the at least one individual, in particular one or more musculoskeletal parameters directly determined from said imaging data, i.e. through direct measurement without artificial intelligence prediction.
[0059] In an embodiment, the imaging data constituting the input data used in the method according to said second implementation comprise imaging data acquired using computed tomography scan and / or magnetic resonance imaging.
[0060] Said imaging data constituting the input data used in the method according to said second implementation may be retrieved from a data storage system (e.g., PACS).
[0061] In an embodiment, the musculoskeletal parameters forming the input data used in the method according to said second implementation are chosen from a list including Hounsfield units, one or more bone microstructural index parameters, one or more bone fragility index parameters, and one or more muscle quality index parameters.
[0062] In an embodiment, the musculoskeletal parameters forming the input data used in the method according to said second implementation comprise a bone fragility index parameter determined using said Hounsfield units and at least one of said bone microstructural index parameters.
[0063] According to said second implementation, the invention therefore covers, for example, a computed-implemented method for determining one or more musculoskeletal outcomes related to at least one individual, one or more of said musculoskeletal outcomes being preferably representative of a musculoskeletal risk associated with the at least one individual, the method comprising:- obtaining input data comprising one or more musculoskeletal parameters related to the at least one individual, one or more of said musculoskeletal parameters being directly obtained from imaging data representing at least one region of the body of the at least one individual, these imaging data may have been acquired using computed tomography scan and / or magnetic resonance imaging, said input data optionally comprising clinical risk factors data related to the at least one individual, - providing said input data to an artificial intelligence model, in particular a model trained with a method according to said third aspect, an output of the artificial intelligence model forming one or more of said musculoskeletal outcomes.
[0064] Said first implementation and second implementation can be combined with each other, for example to enable selection of one or the other of these implementations according to the available input data and musculoskeletal parameters. In such a combination, said input data used in the method according to said first implementation may be first input data, said input data used in the method according to said second implementation may be second input data, said musculoskeletal parameters forming said first input data may be first musculoskeletalparameters, and said musculoskeletal parameters forming said second input data may be second musculoskeletal parameters.
[0065] Although not limited to this type of architecture, the (second) artificial intelligence model described above with reference to said fourth aspect may typically comprise a machine learning model.
[0066] The invention also covers a data processing apparatus comprising means for carrying out a method according to said fourth aspect.
[0067] The invention also covers a computer program comprising executable instructions which, when the instructions are executed by a computer, cause the computer to carry out a method according to said fourth aspect.
[0068] The invention also covers a non-transitory computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out a method according to said fourth aspect.
[0069] One or more of the data processing apparatus, computer program, and computer-readable medium described with reference to the fourth aspect of the invention may be the same as, part of, or other than, those described above with reference to said first aspect and / or said second aspect and / or said third aspect.
[0070] A fifth aspect of the invention relates to a method for determining one or more musculoskeletal parameters from imaging data representing at least one region of the body of at least one individual, said one or more musculoskeletal parameters being chosen from a list including Hounsfield units, one or more bone microstructural index parameters, and a bone fragility index parameter, said bone fragility index parameter being determined using said Hounsfield units and at least one of said bone microstructural index parameters.
[0071] In an embodiment, the imaging data used in the method according to said fifth aspect comprise imaging data acquired using computed tomography scan and / or imaging data acquired using magnetic resonance imaging.
[0072] The invention also covers: a data processing apparatus comprising means for carrying out a method according to said fifth aspect; a computer program comprising executable instructions which, when the instructions are executed by a computer, cause the computer to carry out a method according to said fifth aspect; and a non-transitory computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out a method according to said fifth aspect.
[0073] One or more of the data processing apparatus, computer program, and computer-readable medium described with reference to said fifth aspect may be the same as, part of, orother than, those described above with reference to said first aspect and / or said second aspect and / or said third aspect and / or said fourth aspect.
[0074] Among other advantages, the invention introduces a novel, multi-faceted approach for the comprehensive assessment of musculoskeletal health, makes it possible to enhance the diagnostic value of imaging modalities, and facilitates opportunistic screening and pre-surgical planning, offering real-time assessments of musculoskeletal health. It may eliminate the need for DXA scans, making bone fragility and muscle quality assessments more accessible and efficient. In addition, the invention enables a comprehensive musculoskeletal assessment that integrates bone, muscle and clinical risk factors information, providing a complete bone health assessment and musculoskeletal risk profile of an individual or patient, addressing a critical gap in clinical diagnostics. The invention applies to various clinical settings, including fracture liaison services, pre-surgical planning, trauma care, routine diagnostics, and opportunistic screening using routine imaging.
[0075] As described above, the invention may combine the use of Hounsfield units (HU) derived from computed tomography (CT) scans with directly measured bone texture metrics, Al-predicted surrogates of bone microarchitecture (e.g., trabecular bone score (TBS)), and / or clinical risk factors. The invention may integrate direct and / or predicted muscle quality metrics, such as ectopic fat infiltration and muscle density, derived from CT, magnetic resonance imaging (MRI), or predicted from x-ray using Al model(s). This approach leverages advanced analytical techniques and Al to provide a unified and clinically efficient solution for musculoskeletal health evaluation and subsequent risk assessment across a variety of imaging modalities and anatomical sites.
[0076] The invention provides the following, non-limiting, key features and advancements:- Combination of HU and directly measured bone texture metrics. The technique of the invention may utilize HU values as a measure of bone density, readily derived from CT scans. An advancement is in the current technique through which HU values are then combined with directly measured trabecular bone texture metrics obtained from the CT image through variogram analysis. The experimental variogram applied to CT allows for the direct quantification of trabecular structure, capturing critical details such as trabecular spacing, orientation, and connectivity. Additionally, muscle quality may be assessed through direct quantification of ectopic fat infiltration and muscle density using CT, providing complementary insights into musculoskeletal health. Such a direct measurement approach bypasses the need for predictive models, providing standalone assessments of both bone and muscle quality.- Integration of Al-predicted metrics: To further enhance the evaluation of bone health, the technique of the invention may incorporate Al-predicted microarchitectural metrics such as TBS from DXA and micro-CT, combined with CT-derived HU. Al model(s) may be trained on ground truth data, including DXA-derived TBS, micro-CT imaging, and / or histological datasets, enabling the prediction of bone quality from routine CT, x-ray or MRI imaging. Additionally, Al model(s) may predict muscle quality metrics such as ectopic fat infiltration and muscle density from x-ray images, with ground truth data derived from CT or MRI. Such a predictive capability adds a layer of depth to bone and muscle health assessments, capturing information that HU values alone cannot provide.- Extension to x-ray imaging: The invention may also extend bone and muscle assessment capabilities to x-ray imaging. By leveraging ground truth data from CT or DXA, the invention enables the prediction of HU, TBS-bone microarchitecture equivalent and muscle quality metrics from x-ray images.- Extension to MRI: The invention may also extend bone fragility assessment capabilities to MRI. By leveraging ground truth data from CT or DXA, the invention enables the prediction of HU and TBS-bone microarchitecture equivalent metrics, combining these with directly measured proton-density fat-fraction (PDFF) from MRI images.- Versatility across multiple anatomical sites: The invention is applicable to a wide range of anatomical sites, including but not limited to spine (e.g., vertebrae), pelvis (e.g., iliac crest, sacrum), hip (e.g., femoral neck, proximal femur), long bones (e.g., femur, tibia, humerus, radius, ulna), and muscle groups associated with these regions (e.g., paraspinal, gluteal, quadriceps). The invention can thus be utilized across various clinical disciplines, including osteoporosis, orthopedics, neurosurgery, obesity interventions, endocrinology, and general trauma care.- Opportunistic screening: The invention enables opportunistic bone health assessments during routine imaging, such as abdominal or chest CTs performed for other diagnostic purposes. By extracting HU values and applying Al-predicted metrics or variogram analysis to the same imaging data, the invention may eliminate the need for additional imaging sessions. This facilitates rapid assessments, particularly in pre- surgical or emergency settings where timely decision-making is crucial.- Pre-surgical and routine clinical applications: In pre-surgical settings, such as spine or joint replacement surgery, the invention may provide critical insights into bonefragility, helping surgeons tailor their approach. For example: Initiating osteoporosis, anabolic treatments prior to surgery; Modifying surgical fixation techniques, implant choices, or surgical approaches based on bone quality; Identifying sarcopenia-related risks that could impact post-surgical recovery. In routine clinical practice, the invention may support assessment across diverse populations, including those with limited access to DXA.- Broader clinical utility and impact: The ability to combine HU, bone quality, and muscle quality provides clinicians with a more holistic view of bone health, addressing limitations in current diagnostic approaches. By eliminating reliance on separate DXA scans and stand-alone muscle health assessments, the invention reduces diagnostic delays and operational inefficiencies, ensuring that at-risk patients receive timely intervention. Such an integrated approach is particularly valuable in aging populations, where both osteoporosis and sarcopenia significantly contribute to morbidity and mortality.
[0077] The present invention, and all its aspects, embodiments and advantages associated therewith will become more readily apparent in view of the detailed description of particular embodiments provided below, including the accompanying drawings.Brief description of the drawings
[0078] The following, non-limiting, embodiments of the invention are described with reference to the accompanying drawings in which:- Figure 1 is a schematic representation of an embodiment of the invention in which an artificial intelligence model is used to predict musculoskeletal parameters from imaging data, which may include x-ray imaging data;- Figure 2 is a schematic representation of an embodiment of the invention in which an artificial intelligence model, for example the model of figure 1, is trained using training datasets generated with imaging data which may include CT and / or MRI imaging data; - Figure 3 is a schematic representation of an embodiment of the invention in which an artificial intelligence model is used to predict musculoskeletal outcomes from input data which may comprise musculoskeletal parameters predicted with an artificial intelligence model according to figure 1 and / or musculoskeletal parameters directly determined from CT and / or MRI imaging data;- Figure 4 is a schematic representation of an embodiment of the invention in which an artificial intelligence model, for example the model of figure 3, is trained using training datasets generated with input data which may comprise musculoskeletal parameterspredicted with an artificial intelligence model according to figure 1 and / or musculoskeletal parameters directly determined from CT and / or MRI imaging data; Figure 5 is a schematic representation of a computing device according to the invention.Detailed description of embodiments
[0079] The embodiments described in the present section implement various aspects of the invention. These embodiments are not limitative and the invention does cover techniques as defined in the claims which can implement other combinations of the different aspects of the invention, including any combinations of the aspects described above and below.
[0080] 1) Determination of MSK parameters with a deep learning model
[0081] In the embodiment illustrated in figure 1, parameters 1, generally referred to in this document as "musculoskeletal parameters", are determined from input imaging data 2 using an artificial intelligence model 3.
[0082] 1.1) Imaging data
[0083] In this non-limiting example, imaging data 2 are acquired using x-ray imaging providing data in digitized image form, representing a region of an individual's body containing a bone structure and its surrounding soft tissues.
[0084] Imaging data 2 may be retrieved from a data storage system, such as a "picture archiving and communication system", which typically stores images from various medical imaging instruments and optionally meta-data including for example clinical data such as age and / or sex and / or acquisition parameters.
[0085] Imaging data 2 may typically be subjected to pre-processing including one or more steps chosen from a list including segmenting, normalizing, pixel-spacing, and filtering the image data. For example, a segmentation step may be implemented to extract regions of interest from bone or soft tissues (e.g. trabecular bone, cortical bone, muscle, etc.), while normalization and filtering steps may be used to prepare the images for use as input to the artificial intelligence model 3.
[0086] 1.2) Musculoskeletal parameters
[0087] In this non-limiting example, musculoskeletal (MSK) parameters 1 may be chosen from a list including Hounsfield units (HU), one or more bone microstructural index parameters (BMSI), one or more bone fragility index parameters (BFI), a trabecular bone score (TBS), and one or more muscle quality index parameters (MQI).
[0088] As known per se in the prior art, HU may be derived from the absorption attenuation of the bone shown on an image, typically quantifying the attenuation of x-rays by tissue relativeto water (0 HU) and air (-1000 HU). HU provide a practical and immediate measure of bone density, which typically facilitates assessments during routine imaging or pre-surgical planning.
[0089] HU can be used to assess bone density, fat infiltration, or muscle quality, as described for example in the following documents:- Pickhardt PJ, Lee LJ, del Rio AM, Lauder T, Bruce RJ, Summers RM, Pooler BD, Binkley N. Simultaneous screening for osteoporosis at CT coIonography: bone mineral density assessment using MDCT attenuation techniques compared with the DXA reference standard. J Bone Miner Res. 2011 Sep;26(9):2194-203. doi: 10.1002 / jbmr.428. PMID: 21590738; PMCID: PMC3304444.- Aubrey J, Esfandiari N, Baracos VE, Buteau FA, Frenette J, Putman CT, Mazurak VC. Measurement of skeletal muscle radiation attenuation and basis of its biological variation. Acta Physiol (Oxf). 2014 Mar;210(3):489-97. doi: 10.1111 / apha.12224. PMID: 24393306; PMCID: PMC4309522.- Mourtzakis M, Prado CM, Lieffers JR, Reiman T, McCargar LJ, Baracos VE. A practical and precise approach to quantification of body composition in cancer patients using computed tomography images acquired during routine care. Appl Physiol Nutr Metab. 2008 Oct;33(5):997-1006. doi: 10.1139 / H08-075. PMID: 18923576.
[0090] The computation of HU may be based on a CT scan, as well as the definition of region of interest (ROI) on axial slices of bone (e.g. vertebrae) or muscle (e.g. psoas or paraspinal muscles). HU computation can be achieved using conventional techniques embedded in CT systems or softwares, including 3D slicer, ImageJ, OsiriX, etc.
[0091] The mean HU value within the ROI can be calculated to further derive bone or muscle quality indices, using tertiles or existing thresholds for clinical interpretation.
[0092] Typially, for muscle HU interpretation:- a HU range > 40 indicates Lean muscle, with low fat infiltration;- 30-40 HU corresponds to normal or average attenuation;- < 30 HU indicates high fat infiltration and lowered muscle quality;- < 0 HU indicates severe fat degeneration.
[0093] Typically, for bone HU interpretation:- HU > 150 indicates dense or normal bone quantity;- HU in [100, 150] indicates Osteopenia or moderate bone quantity;- HU < 100 indicates Osteoporosis or low bone quantity.
[0094] MQI may typically depend on the muscle surrounding the bone shown on the image.
[0095] MQI can be obtained using CT-derived HU measures on muscle tissues using the above HU interpretation ranges, and also from MRI images, for example as described in the following documents:- Marina Codari MSc, PhD, Moreno Zanardo MSc, Maria Eugenia di Sabato MD, Elisabetta Nocerino MD, Carmelo Messina MD, Luca Maria Sconfienza MD, PhD, Francesco Sardanelli MD. MRI-Derived Biomarkers Related to Sarcopenia: A Systematic Review. 2019 Sep; 51(4):1117-1127. doi: https: / / doi.org / 10.1002 / jmri.26931.- Hooijmans MT, Schlaffke L, Bolsterlee B, Schlaeger S, Marty B, Mazzoli V. Compositional and Functional MRI of Skeletal Muscle: A Review. J Magn Reson Imaging. 2024 Sep;60(3):860-877. doi: 10.1002 / jmri.29091. Epub 2023 Nov 6. PMID: 37929681; PMCID: PMC11070452.
[0096] From CT, MQI can be obtained from muscle attenuation measures (HU), muscle area (cm2) to compute skeletal muscle index (SMI = muscle area / height), muscle radiodensity heterogeneity (standard deviation of HU within ROI) or fat-to-muscle ratio using threshold-based segmentation. Each of these parameters can be stratified into tertiles or risk categories based on normative data from population cohorts.
[0097] From MRI, MQI can be obtained from several measurements including:- Fat quantification (intramuscular fat) with Dixon technique;- Proton Density Fat Fraction (PDFF);- Texture analysis or radiomics;- Muscle volume (can be standardized by Age, for example).
[0098] One or more of said BMSI parameters may combine TBS and / or other microarchitecture parameters.
[0099] TBS may typically depend on a texture of the trabecular part of the bone shown on an image. In a non-limitative way, TBS may additionally depend on HU and / or on one or more of said MQI parameters.
[0100] According to an advantageous aspect of the invention, musculoskeletal parameters 1 may comprise a BFI parameter determined from HU and / or BMSI (e.g. TBS) and / or MQI.
[0101] More generally, musculoskeletal parameters 1 preferably reflects structural and quality attributes of a bone shown on imaging data 2.
[0102] 1.3) Deep learning model
[0103] In this non-limiting example, the artificial intelligence model 3 comprises a convolutional neural network. In a non-limiting way, the model 3 may alternatively comprise another type of deep learning model or a hybrid deep-learning architecture.
[0104] The artificial intelligence model 3 can be designed for either a classification task or a regression task, depending on the desired output.
[0105] For example, in a classification setup, the model may assign the individual's BFI to discrete risk categories (e.g., low, moderate, high) using a ground-truth class derived from HU, BMSI, and MQI values. Such a classification setup may typically employ fully connected layers with "SoftMax" or "Sigmoid" activation in the final layer and be optimized using a cross-entropy loss function to improve categorical predictions. Architectures such as "ResNet" (Residual Neural Networks), "EfficientNet", or "Vision Transformers" (ViT) can be leveraged for feature extraction, followed by a multi-layer perceptron (MLP) classifier to map extracted features to discrete BFI categories.
[0106] In a regression setup, the model may predict continuous values of musculoskeletal parameters 1 (e.g. HU, BMSI, MQI, ...), providing a more detailed and individualized assessment. Such a regression setup may typically use a fully connected output layer with linear activation, trained using Mean Squared Error (MSE) or Mean Absolute Error (MAE) loss functions to minimize prediction errors. Popular architectures for regression tasks may be used for the model 3, for example "DenseNet", "U-Net" (if spatial information is considered), "ResNet", "VGG", or transformer-based models like "Swin Transformer", which enable the network to learn fine-grained structural variations.
[0107] 1.4) Deep learning model: Training
[0108] In this embodiment, with reference to figure 2, training the artificial intelligence model 3 comprises the following steps:- obtaining training imaging data 2A,- determining one or more musculoskeletal parameters 1A from said training imaging data 2A,- forming ground truth data 10 from one or more of said musculoskeletal parameters 1A,- generating one or more training datasets 11 that comprise said ground truth data 10 and one or more of said training imaging data 2A,- training the model 3 using one or more of said training datasets 11.
[0109] In this non-limiting example, the training imaging data 2A are acquired using computed tomography scan and imaging data acquired using magnetic resonance imaging,providing computed tomography (CT) and magnetic resonance imaging (MRI) data in digitized image form, representing a region of the body of individuals, this region corresponding to that of the imaging data 2, thus containing in this example a bone structure and its surrounding soft tissues.
[0110] Imaging data 2A may be retrieved from the data storage system mentioned above or from another data storage system.
[0111] Imaging data 2A may typically be subjected to pre-processing including one or more steps chosen from a list including segmenting, normalizing, and filtering the CT and / or MRI data. These processes may extract gold-standard musculoskeletal measures from bone and muscle tissues, for example HU, microarchitecture parameters (e.g., TBS), muscle cross-sectional area, muscle attenuation, intramuscular fat, bone marrow fat, peri / epi-cardial fat from CT, and / or proton-density fat-fraction (PDFF), abdominal fat, fat volume, fat density, and / or intramuscular fat from MRI. Pre-processing may also implement an automatic fracture detection module, identifying fracture type (e.g., biconcave, intra / inter-trochanteric, crush), grade (e.g., grade 1, grade 2, or grade 3), and site (e.g., LI vertebra, L5 vertebra).
[0112] CT imaging enables the direct measurement of muscle quality, particularly through the quantification of ectopic fat infiltration and muscle density. These metrics provide critical insights into sarcopenia, complementing bone density evaluations. By simultaneously assessing bone and muscle quality, CT allows for a comprehensive analysis of osteo-sarcopenia, a condition of growing clinical significance in aging populations and in those with rapid weight loss. The CT repository may comprise a collection of CT images and / or meta-data related to the associated subject and acquisition parameters.
[0113] MRI imaging adds another dimension to musculoskeletal health assessment by providing PDFF, a gold-standard measure of ectopic fat infiltration in muscles. MRI-based PDFF complements CT-derived metrics by offering radiation-free, highly sensitive assessments of muscle quality. MRI also holds potential for indirect evaluation of bone health by predicting HU and bone microarchitecture (e.g., TBS) through Al models trained on CT ground truth. The MRI repository may comprise MRI images and / or meta-data related to the associated subject and acquisition parameters. MRI images may be used to extract gold-standard muscle parameters, such as proton-density fat-fraction (PDFF) and / or other muscle quality surrogates.
[0114] In this non-limiting example, musculoskeletal parameters 1A may be chosen from a list including HU, one or more BMSI, one or more bone fragility index parameters (BFI), a trabecular bone score (TBS), and one or more muscle quality index parameters (MQI).
[0115] HU may be extracted from the measured attenuation coefficients of the bone tissue and / or soft tissue region of interest displayed in CT images. HU determination may be performed manually or by third-party software.
[0116] In this example, one or more of said MQI parameters (e.g. proton density fat fraction) may be determined from CT images (e.g. muscle cross sectional area, intramuscular fat, bone marrow fat, peri / epi-cardial fat, etc.) and / or MRI images (proton density fat fraction, abdominal fat, fat volume, fat density, intramuscular fat, etc.). MQI may be directly extracted from metadata and / or calculated using pre-defined or existing thresholds / tertiles for metrics such as muscle cross-sectional area, muscle attenuation, intramuscular fat, bone marrow fat, peri / epi-cardial fat for CT images, or proton-density fat-fraction, abdominal fat, fat volume, fat density, or intramuscular fat for MRI. The muscle quality index may be compared to a reference population, producing a unitless score for interpretation.
[0117] TBS may be determined using one or more of the experimental variograms calculated from CT imaging data 2A. Calculations of these variograms may be performed using one slice of the CT scan (e.g., the sagittal slice) or using two or more slices. Gray-level variations may be evaluated on a 2-D or 3-D-based image, providing insights and ground-truth information about bone quality, microarchitecture, and texture. Examples of variogram calculations are described in the document WQ2023051909A1, whose contents are incorporated here by reference.
[0118] Advantageously, musculoskeletal parameters 1A may comprise a BFI parameter determined from HU and at least one of said BMSI parameters. Using HU values and microarchitecture metrics (e.g., CT-TBS), this process may be configured to categorize the data into predefined classes or tertiles based on established thresholds for bone quantity and quality. The integration of HU and TBS classifications results in a BFI, which may be stratified into different risk levels - such as normal, moderate, high, or very high - according to predefined thresholds or tertiles.
[0119] The table below shows an evaluation table designed to determine BFI of an individual based on previously determined HU and a bone microarchitecture parameter such as TBS.
[0120] This example table is structured as a two-dimensional matrix:- Hounsfield Units (HU) are arranged in columns, with each column representing a categorical evaluation, for example: "Normal", "Osteopenia", and "Osteoporosis". -TBS is arranged in rows, with each row also representing a similar categorical evaluation, such as "Normal", "Partially degraded", and "Degraded".
[0121] Each intersection of a row and a column corresponds to a combination of the two parameters. This combination provides an evaluation of the patient's bone fragility. For instance, a "Normal" bone fragility may be determined for a combination of high HU ("Normal") and high TBS ("Normal"), whereas a "Very High BFI" could result from a combination of low HU ("Osteoporosis") and low TBS ("degraded"). Other combinations are possible, for example:- HU "Osteopenia" +TBS "Normal" => Medium BFI- HU "Osteoporosis" + TBS "Normal" => High BFI- HU "Normal" + TBS "Partially degraded" => Medium BFI- HU "Osteopenia" + TBS "Partially degraded" => Medium BFI- HU "Osteoporosis" + TBS "Partially degraded" => High BFI- HU "Normal" +TBS "Degraded" => Medium BFI- HU "Osteopenia" +TBS "Degraded" => High BFI
[0122] Depending on the position in the table, specific information may be provided, such as a clinical recommendation, an intervention threshold, or a general diagnosis of the patient's bone quality. This table may provide a practical tool for integrating quantitative imaging data (HU) with qualitative bone microarchitecture data (TBS), thus enabling a comprehensive and objective assessment of bone quality.
[0123] The ground truth data 10 may be established by integrating gold-standard measures of bone fragility (e.g., HU + TBS) and / or microarchitecture parameters and / or muscle quality index (MQI), and advantageously the BFI, for example in the form of a ground-truth vector.
[0124] With reference to figure 2, the artificial intelligence model 3 may be trained using supervised learning techniques with ground-truth data derived from such CT and / or MRI data.
[0125] For example, backpropagation may be applied to minimize the difference between a predicted MSK parameter vector and a reference vector, adjusting weights through gradient descent and optimization techniques such as "Adam" or "RMSprop". The model 3 may be further refined through data augmentation (e.g., rotation, contrast adjustments) to enhance generalization and be validated on independent datasets to ensure robustness. This Al-driven approach enables the extraction of clinically relevant MSK parameters from standard x-rayimages, providing a widely accessible tool for assessing bone and muscle health without requiring CT or MRI scans.
[0126] 1.4.1) Deep learning model training: Continuous representation
[0127] According to a first illustrative example of ground truth representation, ground truth data 10 are structured as a multi-dimensional vector where each parameter is stored as a continuous value, and the model 3 is submitted to a regression-based training, to predict from an x-ray image input exact numerical values of one or more of the following parameters:- Bone Mineral Density (BMD) may surrogate HU attenuation coefficient (continuous, in Hounsfield Units);- BMSI: A continuous index derived from trabecular and cortical microarchitecture parameters, incorporating:■ TBS (continuous, unitless scale, from experimental variogram calculation); ■ Trabecular Bone Microarchitecture:• Trabecular separation (Tb.Sp) (continuous, mm);• Trabecular thickness (Tb.Th) (continuous, mm);• Trabecular number (Tb.N) (continuous, mm-1);• Trabecular bone volume fraction (BV / TV) (continuous, %);• Trabecular connectivity density (Conn.D) (continuous, mm-3);• Degree of anisotropy (DA) (continuous, unitless);• Structure model index (SMI) (continuous, unitless);■ Cortical Bone Microarchitecture:• Cortical thickness (Ct.Th) (continuous, mm);• Cortical porosity (Ct.Po) (continuous, %);- Muscle Quality Parameters:■ Cross-sectional area (CSA) of muscle (continuous, cm2);■ Muscle density (attenuation in HU);■ Intramuscular fat fraction (IMFF, %).
[0128] Examples of values corresponding to a healthy individual at low risk are: HU = 180; TBS = 1.45; Tb.Sp = 0.18 mm ; Tb.Th = 0.14 mm; Tb.N = 1.90 mm-1; BMSI = 82; CSA = 45 cm2; Muscle Attenuation (MA) = 38 HU; PDFF = 4.5%.
[0129] Examples of values corresponding to an osteopenic individual at moderate risk are: HU = 120; TBS = 1.20; Tb.Sp = 0.22 mm; Tb.Th = 0.12 mm; Tb.N = 1.65 mm-1; BMSI = 65; CSA = 38 cm2; MA = 32 HU; PDFF = 9.5%.
[0130] Examples of values corresponding to an osteoporotic individual at high risk are: HU = 75; TBS = 1.00; Tb.Sp = 0.30 mm; Tb.Th = 0.10 mm; Tb.N = 1.40 mm’1; BMSI = 50; CSA = 30 cm2; MA = 25 HU; PDFF = 18.0%.
[0131] 1.4.2) Deep learning model training: Categorical representation
[0132] According to a second illustrative example of ground truth representation, ground truth data 10 are structured as a categorical vector, grouping bone and muscle quality into discrete risk categories using tertiles derived from a training population, or existing thresholds for results interpretation. This classification-based approach enables categorization of individuals into clinically relevant risk profiles, supporting stratification-based decision-making. Instead of using continuous values, discrete classes based on threshold-based binning can be assigned, for example for one or more of the following parameters:- BMSI: Class 1 (Normal), Class 2 (Partially Degraded), Class 3 (Degraded);- HU: Class 1 (Normal), Class 2 (Osteopenia), Class 3 (Osteoporosis);- BFI: A 9-category ordinal index based on HU and BMSI: BFI = HU Class + BMSI Class -> {1-9}, where 1 = Best Bone Quality, 9 = Worst Bone Quality;- MQI: A categorical index derived from muscle CSA, density, and IMFF: Low, Medium, High Muscle Quality (tertile-based classification).
[0133] Examples of classes corresponding to a healthy individual at low risk are: HU Class = Normal (> 150 HU); TBS Class = Normal (> 1.32); Microarchitecture Class = Normal; BMSI = High (> 75); BFI = Class 1 (Healthy); MQI = Tertile 1 (High Muscle Quality).
[0134] Examples of classes corresponding to an individual at moderate fragility risk are: HU Class = Osteopenia (90-149 HU); TBS Class = Partially Degraded (1.10-1.34); Microarchitecture Class = Mild Degradation; BMSI = Moderate (55-74); BFI = Class 5 (Moderate Fragility); MQI = Tertile 2 (Intermediate Muscle Quality).
[0135] Examples of classes corresponding to an individual at high fragility risk are: HU Class = Osteoporosis (< 90 HU); TBS Class = Degraded (< 1.10); Microarchitecture Class = Severe Degradation; BMSI = Low (< 55); BFI = Class 9 (Severe Fragility); MQI = Tertile 3 (Low Muscle Quality).
[0136] 1.5) Deep learning model: Prediction
[0137] Referring to figure 1, the artificial intelligence model 3 trained as described above is provided with the imaging data 2, to predict one or more of said musculoskeletal parameters 1.
[0138] In this example, the input data 2 comprise x-ray imaging, which is one of the most widely available imaging modalities in clinical practice, but currently underutilized for bone fragility assessments due to the lack of direct measures for bone density or microarchitecture.
[0139] Of course, musculoskeletal parameters 1 can be used "as-is", in their original continuous representations, or transformed, for example transformed into classes.
[0140] 2) Determination of MSK outcomes with a machine learning model
[0141] In the embodiment illustrated in figure 3, output 20, generally referred to in this document as "musculoskeletal outcomes", are determined from input data 21 using an artificial intelligence model 22.
[0142] 2.1) Input data
[0143] In this non-limiting example, input data 21 generally comprise one or more musculoskeletal parameters (HU, BMSI, BFI, TBS, MQI, etc.) related to an individual, and may include:- First type of input data 21: One or more musculoskeletal parameters previously determined using an artificial intelligence model 3 as described above (e.g., one or more musculoskeletal parameters 1 determined from x-ray imaging data 2 with the model 3); and / or- Second type of input data 21: One or more musculoskeletal parameters directly determined from imaging data (e.g. CT and / or MRI images). The word "directly" refers to obtaining measurements without relying on Al predictions from an image, but rather through direct measurement. For example, HU may be extracted from the measured attenuation coefficients of the bone tissue and / or soft tissue region of interest displayed in a CT image.
[0144] Concerning said second type of input data 21, CT and / or MRI imaging data may be acquired and subjected to pre-processing, for example as described in section 1.4 above.
[0145] Input data 21 may also comprise clinical risk factors data (CRF) related to the individual, providing critical context to imaging-based assessments.
[0146] For example, predicted and / or detected prevalent fractures, or surgical implants, may be used as detected clinical risk factors.
[0147] Clinical risk factors may generally be extracted from an electronic medical records (EMR), and may include one or more, preferably at least two, of the following factors: age, sex, body mass index (BMI), previous history of fracture, parental history of hip fracture, current smoking status, glucocorticoid use in the last three months, medication use or comorbidities (e.g., diabetes or rheumatoid arthritis), other types of secondary osteoporosis, high alcohol intake (e.g. three or more doses per day), prevalence of fractures, metabolic outcomes derived from EMR or identified during pre-processing.
[0148] Integrating clinical risk factors with imaging metrics enables better stratification of individuals according to their risk of adverse musculoskeletal events, such as fragility fractures or osteo-sarcopenia-related functional declines.
[0149] 2.2) Musculoskeletal outcomes
[0150] In this non-limiting example, musculoskeletal outcomes 20 may generally be representative of a musculoskeletal risk associated with the individual.
[0151] Musculoskeletal outcomes 20 may comprise, or may be used to determine, a likelihood score related, but not restricted, to at least one of osteoporosis, sarcopenia, osteosarcopenia, fragility fractures, surgical implant loosening, surgical failure, peri-prosthetic fracture.
[0152] A comprehensive report may be generated to detail the musculoskeletal outcomes 20.
[0153] 2.3) Machine leaning model
[0154] In this non-limiting example, the artificial intelligence model 22 comprises a machine learning model, which can be used as a risk assessor.
[0155] Different machine learning (ML) architectures can be applied to perform multifactorial predictions. For binary classification (event / no event), suitable models include a "Multilayer Perceptron" (MLP) with fully connected layers ("ReLU" activation, dropout, and batch normalization), a "Random Forest" (RF) model for robust feature importance analysis, or an "Extreme Gradient Boosting" ("XGBoost") model optimized for structured tabular data. Given the time-to-event nature of the prediction, ML-based survival analysis models can be used as alternative approaches, including "Cox Proportional Hazards" models with ML enhancements, "Random Survival Forests" (RSF) for non-linear feature interactions, or "DeepSurv", a deeplearning adaptation of the Cox model capable of capturing complex dependencies over time.
[0156] The training process preferably follows a supervised learning framework, using backpropagation and gradient-based optimization (for neural networks) or decision-tree-based gradient boosting (for RF and "XGBoost" models). The model may be trained to minimize loss functions adapted to the prediction task, such as binary cross-entropy for classification or negative log-likelihood for survival models. The training pipeline may include data preprocessing (handling missing values, normalizing continuous features, encoding categorical variables), hyperparameter tuning (grid search or Bayesian optimization), and validation / testing (k-fold cross-validation, independent dataset evaluation). The performance may be assessed using "AUC-ROC" for classification models, "Concordance Index" (C-index) for survival models, and Brier score for event probability calibration.
[0157] By integrating bone, muscle and CRF parameters into a single predictive model, this ML-driven approach enables a personalized and multi-factorial assessment of musculoskeletal health, improving risk stratification and early detection of high-risk individuals. This allows for proactive clinical decision-making, such as early interventions, treatment adjustments, or enhanced patient monitoring, without requiring invasive testing beyond standard imaging and EMR data.
[0158] 2.4) Machine leaning model: Training
[0159] In this embodiment, with reference to figure 4, training the artificial intelligence model 22 comprises the following steps:- obtaining input data 21A related to individuals,- obtaining one or more musculoskeletal outcomes 20A from one or more of said input data 21A,- forming ground truth data 30 from one or more of said musculoskeletal outcomes 20A,- generating one or more training datasets 31 that comprise said ground truth data 30 and one or more of said input data 21A,- training the model 22 using one or more of said training datasets 31.
[0160] In this non-limiting embodiment, input data 21A may form a vector comprising CRF parameters obtained from EMR and musculoskeletal parameters as described in section 2.1 above. Therefore, musculoskeletal parameters may include HU density, TBS, bone microarchitecture parameters, bone marrow fat, muscle cross-sectional area, muscle attenuation, intramuscular fat, PDFF, and other muscle quality surrogates, and these musculoskeletal parameters can either be directly measured from CT / MRI imaging data using conventional methods (second type of input data 21), when available, or predicted from another Al model (CNN classifier or regressor, see sections 1.3-1.5 above) output, providing features based on x-ray (first type of input data 21).
[0161] The ground truth data 30 for training the ML model may be established from longitudinal patient follow-up data, which document whether the individual experienced at least one of predicted adverse musculoskeletal events. Ground truth labels can be represented as binary vector or one-hot encoded matrix, depending on whether the prediction task is a singlelabel (binary classification) or multi-label (multiple event types).
[0162] For survival analysis models: The ground truth 30 can be represented as a time-to-event vector, where the first column contains the time (e.g., days or months) until the event occurs or until censoring, and the second column contains a censoring indicator (1 if the eventoccurred, 0 if the patient was lost to follow-up or event-free at the last observation). This real-world survival status is derived from clinical registries, patient histories, surgical revision records, and post-fracture assessments.
[0163] By leveraging these diverse data sources, the ML model 22 may learn to identify complex interactions and risk patterns that contribute to musculoskeletal deterioration, moving beyond single-variable assessments to a multi-dimensional prediction framework. This structured approach ensures the ML model 22 can be trained using the appropriate loss functions, whether for binary classification (cross-entropy), multi-label learning (binary crossentropy per event), or survival analysis (negative log-likelihood or Cox partial likelihood).
[0164] Such a machine learning model processes a comprehensive, multi-factorial input vector integrating musculoskeletal parameters and clinical risk factors to predict whether an individual will experience at least one adverse musculoskeletal event, such as osteoporosis, fragility fracture, secondary fracture, implant loosening, implant failure, sarcopenia, or peri-prosthetic fracture. The outcome may be structured as a binary survival status (yes / no), indicating whether the event occurs within a defined follow-up period. This multi-factorial approach ensures that the prediction is not based on a single parameter but rather on a holistic assessment of bone, muscle, and patient-specific clinical risk factors.
[0165] 2.5) Machine leaning model: Prediction
[0166] Referring to figure 3, the artificial intelligence model 22 trained as described above is provided with input data 21, to predict one or more of said musculoskeletal outcomes 20.
[0167] 2.6) Machine learning model: Binary outcome
[0168] According to a first type of implementation, the machine learning model 22 is used for binary classification (yes / no musculoskeletal event). "Random Forest", "XGBoost" and "MultiLayer Perceptron" (MLP) architectures can be considered as suitable models. This approach predicts whether an individual will or will not experience an adverse musculoskeletal event (fragility fracture, implant failure, etc.).
[0169] As a quantified example, the model 22 may predict as output 20 a survival status prediction equal to 1 (Fragility Fracture Event) from an input vector with the following combination of musculoskeletal parameters and CRF:- Musculoskeletal parameters (bone & muscle quality): HU = 90; BMSI = 55 (Moderate); TBS = 1.200; Trabecular Separation = 0.40 mm; Trabecular Thickness = 0.12 mm; Trabecular Number = 1.90 mm-1; MQI = 2nd tertile (Moderate);- CRF: Age = 74 years; Sex = Female; BMI = 21.5; Prior Fracture = Yes (Hip fracture, 4 years ago); Parental History of Hip Fracture = No; Glucocorticoid Use = No;1Rheumatoid Arthritis = No; Smoking = Yes (Former); Alcohol Intake = Low (<1 drink / day).
[0170] 2.7) Machine learning model: Continuous outcome
[0171] According to a second type of implementation, the machine learning model 22 is used to provide a continuous risk score prediction, indicating the of experiencing an adverse musculoskeletal event within a given timeframe. Gradient Boosting, Logistic Regression or MLP architectures can be considered in this embodiment.
[0172] As a quantified example, the model 22 may predict as output 20 a continuous risk score equal to 0.23 (23% probability of musculoskeletal event in the next 5 years) from an input vector with the following combination of musculoskeletal parameters and CRF:- Musculoskeletal parameters: HU = 110; BMSI = 65 (Moderate-High); TBS = 1.310; Trabecular Separation = 0.35 mm; Trabecular Thickness = 0.14 mm; Trabecular Number = 2.10 mm-1; MQI = 3rd tertile (Best);- CRF: Age = 68 years; Sex = Male; BMI = 25.0; Prior Fracture = No; Parental History of Hip Fracture = No; Glucocorticoid Use = No; Rheumatoid Arthritis = No; Smoking = No; Alcohol Intake = Moderate (<2 drinks / day).
[0173] 2.8) Machine learning model: Time-to-event prediction
[0174] According to a third type of implementation, the machine learning model 22 is used for time-to-event prediction (i.e. a time until an adverse musculoskeletal event occurs). Survivalbased architectures can be considered, including but not restricted to Cox Proportional Hasards, DeepSurv, or Random Survival Forests.
[0175] As a quantified example, the model 22 may predict as output 20 a time-to-event of 1.8 years and a survival probability at 5 years equal to 12%, from an input vector with the following combination of musculoskeletal parameters and CRF:- Musculoskeletal parameters: HU = 75; BMSI = 45 (Low); TBS = 1.100; Trabecular Separation = 0.45 mm; Trabecular Thickness = 0.10 mm; Trabecular Number = 1.70 mm-1; MQI = 1st tertile (Worst);- CRF: Age = 79 years; Sex = Female; BMI = 19.5; Prior Fracture = Yes (Vertebral fracture, 2 years ago); Parental History of Hip Fracture = Yes; Glucocorticoid Use = Yes (Chronic use); Rheumatoid Arthritis = Yes; Smoking = Yes (Current); Alcohol Intake = High (>3 drinks / day).
[0176] 3) System
[0177] Sections 1 and 2 above describe computed-implemented methods, namely:- methods for training artificial intelligence models, e.g. deep learning models, using imaging data and musculoskeletal parameters,- methods for determining musculoskeletal parameters either with an artificial intelligence model thus trained or with musculoskeletal parameters directly determined from imaging data,- methods for training artificial intelligence models, e.g. machine learning models, using musculoskeletal parameters as input data and musculoskeletal outcomes, - methods for determining musculoskeletal outcomes using an artificial intelligence model thus trained.
[0178] These methods can be used independently or in combination, and may be used to perform one or more of the following non-limiting operations:- Measuring HU directly from CT imaging;- Predicting HU from x-ray images with CT ground truth,- Measuring directly bone microarchitecture or texture metrics from CT (e.g., variograms and TBS derived from CT);- Predicting bone microarchitecture or texture indices with DXA and / or CT and / or micro-CT ground truth (e.g., TBS);- Measuring directly ectopic fat from CT;- Measuring directly PDFF from MRI;- Predicting muscle quality from x-ray with CT and / or MRI ground truth;- Obtaining CRF from EMR.
[0179] Of course, the invention also covers systems for implementing one or more of these methods, and in particular data processing apparatus - also called computing device or computer -, computer program, and non-transitory computer-readable medium.
[0180] Methods according to the invention can be implemented by any computing device, for example a computing device 40 as illustrated in figure 5. The device 40 may typically comprise a memory 41 and a processor 42 that are interconnected. Memory 41 may be configured to store a program 43 comprising executable instructions, the processor 42 configured to execute these instructions to carry out one or more methods according to the invention.
[0181] More specifically and with reference to the above description, the system may comprise:- One or more input device (e.g., a PACS imaging device) configured to retrieve several data modalities, for example x-ray, CT, MRI, and / or EMR. These modalities may be automatically identified and prepared for further processing.- A data processing apparatus configured to perform one or more of the following: ■ X-ray and / or CT and / or MRI and / or EMR pre-processing,■ Deep learning training and / or analysis,■ Machine learning training and / or analysis,■ Musculoskeletal report generation.
[0182] The system may generally be configured to perform one or more of the steps described above either locally or in a cloud-based environment.
[0183] In different embodiments of the invention, machine learning and deep learning models can predict trabecular microarchitectural metrics, such as TBS, directly from CT, x-ray or MRI images. These Al models leverage ground truth data to bridge the gap between traditional imaging techniques and advanced musculoskeletal health evaluations. The integration of Al techniques across major components of musculoskeletal health assessment, including bone health, muscle health, and clinical risk factors, enables the evaluation of feature contributions within each component to adverse MSK outcomes. By quantifying these contributions, Al uncovers the underlying drivers of fragility fractures, surgical risks or sarcopenia, for example. Explainable Al transforms complex outputs into interpretable insights, empowering clinicians to define precise risk profiles and make evidence-based decisions tailored to the unique needs of each patient.
[0184] The invention advances the state of the art by at least one of the following:- Combining Al-predicted or directly measured HU with microarchitectural surrogates such as TBS provides a clinically relevant assessment of bone fragility. This approach bridges the gap between traditional imaging metrics and actionable insights for fragility and sarcopenia risks.- Introducing direct texture analysis using variograms derived from CT data bypasses the reliance ground truth datasets. The invention enables immediate evaluation of trabecular structure, offering a novel approach to texture-based bone health assessments.- Enhancing x-ray imaging by incorporating Al-predicted microarchitectural indices (e.g., TBS) and muscle quality metrics, trained from CT ground truth, which extends the utility of widely available modalities. This innovation allows for robust MSK assessments even in settings where advanced imaging like CT is unavailable. - Integrating MRI-based PDFF measurements provides a radiation-free method to evaluate muscle quality and ectopic fat infiltration.- Incorporating clinical risk factors (e.g., age, fracture history, comorbidities, etc.) into the assessment framework ensures that patient-specific biological and systemic variables are considered. This integration enhances the precision and relevance of risk evaluations for MSK outcomes.- Providing an integrated evaluation of osteo-sarcopenia by synthesizing bone health, muscle quality, and clinical risk factors fills a critical gap in MSK assessments. This holistic perspective enables clinicians to understand the interplay between these domains, delivering personalized and actionable recommendations in both routine and specialized care.
[0185] The invention addresses limitations in current practice by integrating multiple capabilities into a single, seamless workflow. This allows for bone and muscle quality assessments during routine imaging and in pre-surgical or urgent care scenarios, eliminating the need for additional imaging.
[0186] Non-limiting advantages of the invention, which address inefficiencies in current methodologies, may include the following:- Combining bone density from HU with bone microarchitecture measurements, and muscle quality assessments, into a single, cohesive analysis, offering holistic insights into musculoskeletal health.- Eliminating the need for separate DXA scans, enabling real-time assessments during routine or pre-surgical imaging.- Extending bone fragility and muscle quality evaluation and screening to widely available x-ray modalities, broadening accessibility and utility.- Incorporating both direct texture analysis (variograms) and Al-predicted metrics, ensuring robust and versatile evaluations.- Serving diverse patient populations and clinical scenarios, from trauma care to osteoporosis, obesity and chronic disease management.
[0187] To date, there is no established technique that combines HU bone density and muscle quality, Al-predicted bone microarchitecture metrics, direct texture analysis and / or clinical risk factors into a single diagnostic framework. Current practices either rely on standalone DXA scans, fragmentary analyses of HU for BMD, or isolated assessments of muscle mass without evaluating quality. These limitations reduce their applicability in diverse clinical contexts and add time to the clinical workflow.
Claims
CLAIMS1. A computed-implemented method for training an artificial intelligence model (3), comprising:- obtaining imaging data (2A) representing at least one region of the body of one or more individuals,- determining one or more musculoskeletal parameters (1A) from said imaging data (2A), said one or more musculoskeletal parameters being chosen from a list including Hounsfield units, one or more bone microstructural index parameters, and one or more muscle quality index parameters,- forming ground truth data (10) from one or more of said musculoskeletal parameters (1A),- generating one or more training datasets (11) that comprise said ground truth data (10) and one or more of said imaging data (2A),- training the artificial intelligence model (3) using one or more of said training datasets (11).
2. A method according to claim 1, wherein the step for determining one or more musculoskeletal parameters (1A) comprises determining a bone fragility index parameter using said Hounsfield units and at least one of said bone microstructural index parameters, the ground truth data (10) comprising said bone fragility index parameter.
3. A method according to claim 1 or 2, wherein the imaging data (2A) comprise imaging data acquired using computed tomography scan and / or imaging data acquired using magnetic resonance imaging.
4. A method according to any one of claims 1 to 3, wherein the step for determining one or more musculoskeletal parameters (1A) comprises calculating one or more experimental variograms from said imaging data (2A), and determining a trabecular bone score using one or more of the calculated experimental variograms, said trabecular bone score forming one of said bone microstructural index parameters and being used to form the ground truth data (10).
5. A computed-implemented method for determining one or more musculoskeletal parameters (1) from imaging data (2) representing at least one region of the body of at least one individual, said musculoskeletal parameters (1) being chosen from a list including Hounsfield units, one or more bone microstructural index parameters, one or more bone fragility index parameters, and one or more muscle quality index parameters, the method comprising a step for providing the imaging data (2) as input to an artificial intelligence model (3) trained with amethod according to any one of claims 1 to 5, an output (1) of the artificial intelligence model (3) forming said one or more musculoskeletal parameters.
6. A method according to claim 5, wherein the imaging data (2) comprise imaging data acquired using x-ray imaging.
7. Acomputed-implemented method fortraining an artificial intelligence model (22), comprising:- obtaining one or more musculoskeletal parameters (21A) and one or more musculoskeletal outcomes (20A) related to one or more individuals, said one or more musculoskeletal parameters being chosen from a list including Hounsfield units, one or more bone microstructural index parameters, and one or more muscle quality index parameters, said one or more musculoskeletal outcomes being representative of a musculoskeletal risk associated with one or more of said individuals, - forming ground truth data (30) from one or more of said musculoskeletal outcomes (20A),- generating one or more training datasets (31) that comprise said ground truth data (30) and one or more of said musculoskeletal parameters (21A),- training the artificial intelligence model (22) using one or more of said training datasets (31).
8. A computed-implemented method for determining one or more musculoskeletal outcomes (20) related to at least one individual, said musculoskeletal outcomes being representative of a musculoskeletal risk associated with the individual, the method comprising:- obtaining input data (21) comprising one or more musculoskeletal parameters related to the at least one individual, said musculoskeletal parameters being previously determined using a method according to claim 5 or 6 which implements a first artificial intelligence model (3),- providing said input data (21) to a second artificial intelligence model (22) trained with a method according to claim 7, an output (20) of the second artificial intelligence model (22) forming one or more of said musculoskeletal outcomes.
9. A method according to claim 8, wherein said input data (21) further comprise clinical risk factors data related to the at least one individual.
10. A method according to claim 8 or 9, wherein said musculoskeletal parameters are first musculoskeletal parameters and said input data are first input data, the method further comprising:- obtaining second input data comprising one or more second musculoskeletal parameters from imaging data representing at least one region of the body of the at least one individual, said second musculoskeletal parameters being chosen from a list including Hounsfield units, one or more bone microstructural index parameters, one or more bone fragility index parameters, and one or more muscle quality index parameters,- providing said second input data to said second artificial intelligence model (22), an output (20) of the second artificial intelligence model (22) forming one or more of said musculoskeletal outcomes.
11. A method according to claim 10, wherein said second musculoskeletal parameters forming the second input data comprise a bone fragility index parameter determined using said Hounsfield units and at least one of said bone microstructural index parameters.
12. A method according to claim 10 or 11, wherein the imaging data comprise imaging data acquired using computed tomography scan and / or magnetic resonance imaging.
13. A data processing apparatus (40) comprising means for carrying out a method according to any one of claims 1 to 4, and / or a method according to claim 5 or 6, and / or a method according to claim 7, and / or a method according to any one of claims 8 to 12.
14. A computer program (43) comprising executable instructions which, when the instructions are executed by a computer, cause the computer to carry out a method according to any one of claims 1 to 4, and / or a method according to claim 5 or 6, and / or a method according to claim 7, and / or a method according to any one of claims 8 to 12.
15. A non-transitory computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out a method according to any one of claims 1 to 4, and / or a method according to claim 5 or 6, and / or a method according to claim 7, and / or a method according to any one of claims 8 to 12.
