Application of artificial intelligence in the diagnosis of musculoskeletal disorders: a systematic review and meta-analysis

Full Text

Introduction

Musculoskeletal diseases (MSDs) comprise a heterogeneous group of conditions affecting bones, joints, muscles, and connective tissues and are among the leading causes of functional limitations and disability. According to the Global Burden of Disease study, approximately 1.71 billion people worldwide are living with MSDs [1].

The number of patients with osteoarthritis (OA) alone reached 595 million in 2020, with the knee and hip being the most commonly affected joints [2]. Along with inflammatory and degenerative joint diseases, metabolic bone disorders, including osteopenia and osteoporosis, as well as associated fractures, contribute substantially to the overall burden of MSDs [3]. In the Russian Federation, the number of patients with osteoporosis is estimated at 14-16 million, accounting for approximately 10% of the country’s adult population. Among individuals aged 50 years and older, osteoporosis is detected in 34% of women and 27% of men, while the prevalence of osteopenia is 43% and 44%, respectively [4]. According to estimates by the International Osteoporosis Foundation, approximately 37 million low-energy fractures occur annually worldwide among individuals aged 55 years and older [5]. According to a multicenter study performed in the Russian Federation, approximately 112,000 femoral neck fractures and approximately 590,000 osteoporotic fractures overall occurred in 2010 [6].

Medical imaging methods play a key role in the diagnosis of MSDs, assessment of disease severity, risk stratification, and longitudinal follow-up. In osteopenia and osteoporosis, opportunistic screening is of particular importance, whereas in osteoarthritis, inflammatory disorders, and other musculoskeletal conditions, imaging is predominantly performed based on clinical indications [7, 8, 9].

Artificial intelligence-based technologies (AI technologies) are increasingly being used in the diagnosis of MSDs, primarily for the automated detection of pathological changes, quantitative assessment of imaging features, and support of radiological image interpretation. Most published systematic reviews of AI technologies for the diagnosis of MSDs are based predominantly on peer-reviewed publications and largely overlook Russian developments, regulatory documents, and data on implementation in clinical practice. Moreover, in many studies, diagnostic accuracy is assessed using internal datasets, which may lead to the overestimation of results. This high-lights the need for a systematic review integra-ting both peer-reviewed and grey literature, including studies that performed external validation using independent datasets, followed by meta-analysis for individual conditions and assessment of the risk of bias.

The aim — to perform a systematic review of the diagnostic accuracy of artificial intelligence-based technologies for musculoskeletal disorders, including both peer-reviewed publications and grey literature sources, with methodological quality assessment and meta-analysis for conditions with at least 3 comparable independent sources.

Methods

The systematic review and meta-analysis were registered in the International Prospective Register of Systematic Reviews (PROSPERO; CRD420251268735) and conducted in accordance with the PRISMA 2020 guidelines.

Search strategy

The search for relevant sources was conducted using two complementary streams of analysis. Stream 1 included a systematic search for peer-reviewed diagnostic accuracy studies in the PubMed and eLIBRARY bibliographic databases. Stream 2 was aimed at identifying registered or clinically implemented AI technologies and related grey literature sources, including regulatory documents, information from medical device registries, technical reports, materials provided by developers, and data from the Moscow Experiment ¹. In addition, for each AI-based solution identified through grey literature sources, a targeted search for peer-reviewed publications specifically addressing that solution was conducted to identify studies that might not have been retrieved by the initial bibliographic search.

Search for peer-reviewed scientific literature

The primary search for articles was conducted using PubMed and eLIBRARY. The date of the last search was December 18, 2025. The search was restricted to the period from 2023 to 2025, which corresponded to the peak publication activity in this field.

The PubMed search query was as follows:

(“Artificial Intelligence”[Mesh] OR “Machine Learning”[Title/Abstract] OR “Deep Learning”[Title/Abstract])

AND

(“Diagnostic Imaging”[Mesh] OR “computed tomography”[Title/Abstract] OR “CT”[Title/Abstract] OR “magnetic resonance imaging”[Title/Abstract] OR “MRI”[Title/Abstract] OR “radiograph”[Title/Abstract] OR “X-ray”[Title/Abstract] OR “bone densitometry”[Title/Abstract] OR “dual-energy X-ray absorptiometry”[Title/Abstract] OR “DXA”[Title/Abstract] OR “imaging”[Title/Abstract])

AND

(“Musculoskeletal Diseases”[Mesh] OR “Bone Diseases, Metabolic”[Mesh] OR “Tendinopathy”[Mesh] OR osteoarthritis[Title/Abstract] OR osteoporosis[Title/Abstract] OR osteopenia[Title/Abstract] OR fracture[Title/Abstract] OR spondylarthritis[Title/Abstract] OR “spinal stenosis”[Title/Abstract] OR “intervertebral disc degeneration”[Title/Abstract] OR “intervertebral disc herniation”[Title/Abstract] OR osteonecrosis[Title/Abstract] OR sacroiliitis[Title/Abstract] OR bursitis[Title/Abstract] OR “carpal tunnel”[Title/Abstract])

AND

(“Diagnosis”[Mesh] OR “screening”[Title/Abstract] OR “diagnos*”[Title/Abstract] OR “detection”[Title/Abstract] OR “classification”[Title/Abstract])

NOT (“oncology”[Title/Abstract] OR “cancer”[Title/Abstract] OR “carcinoma”[Title/Abstract] OR “neoplasm”[Title/Abstract] OR “tumor”[Title/Abstract] OR “sarcoma”[Title/Abstract] OR “metastases”[Title/Abstract] OR review[pt] OR letter[pt] OR editorial[pt] OR comment[pt]).

The eLIBRARY search was performed in advanced search mode using publication titles, abstracts, and keywords (in Russian). The search strategy comprised three conceptual blocks combined using the Boolean operators AND/OR: (1) AI technologies («искусственный интеллект»); (2) medical imaging modalities («компьютерная томография», «магнитно-резонансная томография», «рентгенография», «УЗИ»); and (3) musculoskeletal disorders and anatomical regions («костно-мышечная система», «опорно-двигательный аппарат», «минеральная плотность костной ткани», «остеопения», «остеопороз», «остеоартроз», «остеоартрит», «спондилоартрит», «дегенерация межпозвоночных дисков», «остеонекроз», «грыжа диска»). No restrictions on publication type were applied at the search stage. Non-original publications, reviews, and materials that did not meet the inclusion criteria were excluded during screening and full-text assessment.

Exclusion criteria:

– non-original studies;

– absence of external validation using an independent dataset;

– studies addressing oncological musculoskeletal diseases, including bone metastases;

– studies in prosthodontics and maxillofacial surgery;

– studies in which AI technologies were used exclusively for anatomical annotation, segmentation, prediction, calibration, quality control, or image post-processing without a diagnostic task;

– studies not focused on the diagnosis or classification of specific MSDs.

Study selection was performed by two re-searchers with 6 and 8 years of experience, respectively. The data extraction form was pilottested on a sample of relevant sources, taking into account the previously applied approach [10]. Data were extracted independently by two researchers. Any disagreements were resolved by involving a third researcher as an adjudicator.

Data extraction from articles included in the systematic review

The following data were extracted from the full texts of the included articles:

– bibliometric data: first author, year of publication, journal, DOI, and country;

– study characteristics: objective, design, condition under investigation, imaging modality, unit of analysis, and availability of external validation;

– sample characteristics: number of patients/images, age, sex, prevalence of the condition, and inclusion and exclusion criteria;

– information on the reference standard and expert physicians;

– AI model characteristics, where available: name, architecture, and code availability;

– diagnostic accuracy measures: specificity, sensitivity, and/or 2×2 contingency table;

– information on potential conflicts of interest.

Search for grey literature sources

The search for sources outside peer-reviewed databases included regulatory documents, medical device registers, and technical reports on AI technologies intended for automated diagnosis or decision support in MSDs. Sources from the European Union, the Russian Federation, the United States, and the Asia-Pacific region were analyzed. Searches were conducted in the Health AI Register, the FDA Artificial Intelligence-Enabled Medical Devices database, the State Register of Medical Devices of Roszdravnadzor, the Unified State Health Information System (EGISZ) portal, and among the services included in the Moscow Experiment. Officially registered or authorized AI services with clinical validation data and reported diagnostic accuracy measures were included. Solutions intended for anatomical annotation and segmentation without a diagnostic task, educational applications, and utilities for image calibration, quality control, and post-processing were excluded. For services included in the Moscow Experiment, monitoring reports covering 6 months of 2025 were analyzed. Studies with technical errors or artifacts that precluded interpretation were excluded. The following information was extracted from grey literature sources: product name, developer and country, imaging modality, clinical validation data, and diagnostic accuracy measures, including specificity, sensitivity, and/or a 2×2 contingency table.

Risk of bias

To assess methodological issues associated with diagnostic accuracy studies, a modified Quality Assessment of Diagnostic Accuracy Studies–Computer-Aided Detection (QUADAS-CAD) tool was used [11, 12]. The assessment was performed across four domains: patient selection, index test, reference standard, and patient flow.

Statistical analysis

Meta-analysis was performed separately for each MSD when at least three independent sources, including grey literature, were available. Only studies with external validation, a patient-level unit of analysis, and available diagnostic accuracy measures, including specificity, sensitivity, and a 2×2 contingency table, were included in the quantitative synthesis. Meta-analysis was performed using MetaDTA v. 2.1.5 and a bivariate random-effects model with joint estimation of specificity and sensitivity [13, 14]. Pooled estimates of specificity and sensitivity with 95% confidence intervals (95% CI) were calculated. When the model was unstable because of the small number of observations, MetaBayesDTA was used. Between-study variability was assessed descriptively using forest plots and hierarchical summary receiver operating characteristic (HSROC) curves. No formal statistical hypothesis testing was performed. The certainty of evidence was assessed using the GRADE approach [15].

Results

Study selection

The main reasons for excluding sources at the full-text assessment stage are presented in the flow chart (Figure 1). To ensure the transparency of the selection process, detailed lists of excluded sources were provided in the supplements. A total of 42 sources (24 peer-reviewed publications and 18 grey literature sources) were included in the systematic review, of which 17 were included in the meta-analysis.

 

Figure 1. Flow diagram of the stepwise source selection process for the systematic review and meta-analysis in accordance with the PRISMA 2020 guidelines

 

Baseline characteristics of articles and grey literature sources included in the systematic review

The review included 24 peer-reviewed publications and 18 grey literature sources. AI technologies were most frequently used for the diagnosis of osteopenia and osteoporosis (8; 33.3%), active sacroiliitis (4; 16.7%), degenerative changes and intervertebral disc herniation (3; 12.5%), knee osteoarthritis, and osteonecrosis of the femoral head. Individual studies addressed vertebral compression fractures, spinal canal stenosis, foot deformities, hip osteoarthritis, adhesive capsulitis, infectious spondylitis, and rheumatoid arthritis.

Almost all studies had a retrospective design; only one study performed a secondary analysis of prospectively collected cohort data [16]. All studies were conducted at two or more centers. External validation was performed using data from independent clinical centers or publicly available datasets that had not been used for model training.

Sample size in terms of the number of patients was reported in almost all articles. In five studies, it exceeded 2500 patients [17, 18, 19, 20, 21]. In terms of the diagnostic task, the studies were distributed approximately equally: 13 studies focused on detecting pathology compared with normal findings/absence of pathology, while 11 focused on classifying the stage, severity, or type of the pathological process. The imaging modality was reported in all studies, with MRI (10; 41.7%), radiography (8; 33.3%), and CT (5; 20.8%) being the most frequently used modalities.

The reference standard was described in all included publications. The diagnostic accuracy of AI technologies was compared with the assessments of expert physicians in 10 (41.7%) studies. The number and experience of physicians were reported in 20 (83.3%) studies.

Most grey literature sources were represented by the services included in the Moscow Experiment, while the remaining sources comprised FDA materials on AI-enabled medical devices.

AI technologies were predominantly used for the diagnosis of knee osteoarthritis, scoliosis, vertebral compression fractures, osteopenia/osteoporosis, intervertebral disc protrusions and herniations, spinal canal stenosis, transverse flatfoot, spondylolisthesis, and hip osteoarthritis. Studies performed within the Moscow Experiment had a prospective design, whereas the FDA materials were retrospective. External validation using data from independent clinical centers or external clinical cohorts was reported in 16 of the 18 sources. Publicly available data-sets were used in 2 sources ².

Information on the reference standard was provided in all sources. During the monitoring of services within the Moscow Experiment, physician assessment and comparison of AI service results with radiologists’ reports served as the reference standard. In international sources, DXA, expert assessment, radiologist consensus, or a predefined reference result was used. The baseline characteristics of the articles and grey literature sources are provided in Supplement 1 (all supplements are available in the electronic version of the article on the journal website).

Given the differences in source types, the results of the risk-of-bias assessment are presented separately for peer-reviewed publications and grey literature sources. This approach accounted for the fact that technical reports and regulatory documents may provide less complete descriptions of methodology than publications in peer-reviewed scientific journals. The QUADAS-CAD quality assessment of the 24 peer-reviewed studies is presented in Figure 2. The QUADAS-CAD quality assessment of the 18 grey literature sources is presented in Figure 3.

 

Figure 2. Risk of bias assessments in peer-reviewed studies using the modified QUADAS-CAD tool

 

Figure 3. Risk of bias assessments in grey literature sources using the modified QUADAS-CAD tool

 

Meta-analyses of individual musculoskeletal diseases

Selection of sources for meta-analysis. Of the 42 sources included in the systematic review, 17 were included in the quantitative synthesis, comprising 8 peer-reviewed publications and 9 grey literature sources. Meta-analysis was per-formed only for conditions for which at least three comparable sources with external validation were available. Ultimately, quantitative synthesis was performed for five conditions: knee osteoarthritis, active sacroiliitis, osteoporosis, vertebral compression fractures, and scoliosis.

Knee osteoarthritis. Three AI services identified from grey literature sources were included in the quantitative synthesis: CVisionRad Knee Arthrosis ³, Oxytech Knee X-ray ⁴, and KOALA ⁵. The distribution of specificity and sensitivity estimates for the individual AI services included in the meta-analysis of knee osteoarthritis is presented in Figure 4. The pooled specificity was 0.728, and the pooled sensitivity was 0.959 (95% post-hoc interval 0.709-0.993).

 

Figure 4. Forest plots of specificity (a) and sensitivity (b) of AI services included in the diagnostic accuracy meta-analysis for the detection of knee osteoarthritis

 

The summary diagnostic accuracy of AI services for knee osteoarthritis and the expected range of between-study variability are presented as an HSROC curve in Figure 5.

 

Figure 5. HSROC curve of the diagnostic accuracy meta-analysis of AI services for the detection of knee osteoarthritis. Black dots represent individual AI services, the transparent diamond indicates the pooled estimate, and the solid line represents the HSROC curve. Two black dots, corresponding to the CVisionRad Knee Arthrosis and Oxytech Knee X-ray services, partially overlap because of their similar sensitivity and specificity values. The dashed line indicates the 95% prediction region, while the grey shaded area represents the 95% posterior region of the pooled estimate calculated within the bivariate model

 

Active sacroiliitis. Three peer-reviewed publications were included in the quantitative syn-thesis: A. Bordner et al. [16], J. Nicolaes et al. [22], and K. Zhang et al. [23]. According to the meta-analysis, the pooled specificity was 0.924 (95% CI 0.642-0.988), and the pooled sensitivity was 0.757 (95% CI 0.521-0.899) (Figures 6, 7).

 

Figure 6. Forest plots of specificity (a) and sensitivity (b) of studies included in the diagnostic accuracy meta-analysis for the detection of active sacroiliitis

 

Figure 7. HSROC curve of the diagnostic accuracy meta-analysis for the detection of active sacroiliitis. The points represent individual studies, the blue square indicates the pooled estimate, and the solid line represents the HSROC curve. The horizontal and vertical lines reflect the uncertainty of specificity and sensitivity estimates in the individual studies

 

Osteoporosis. Five peer-reviewed publications were included in the quantitative synthesis: N. Hong et al. [19], J. Tang et al. [24], C. Huang et al. [25], X. Lin et al. [26], and T. Peng et al. [27]. The distribution of specificity and sensitivity estimates is presented in Figure 8.

 

Figure 8. Forest plots of specificity (a) and sensitivity (b) of studies included in the diagnostic accuracy meta-analysis for the detection of osteoporosis

 

According to the meta-analysis, the pooled specificity was 0.891 (95% CI 0.682–0.969), and the pooled sensitivity was 0.920 (95% CI 0.624-0.987).

The summary diagnostic accuracy of the studies included in the meta-analysis of osteoporosis is presented as an HSROC curve in Figure 9.

 

Figure 9. HSROC curve of the diagnostic accuracy meta-analysis for the detection of osteoporosis. The points represent individual sources, the blue square indicates the pooled estimate, and the solid line represents the HSROC curve. The horizontal and vertical lines reflect the uncertainty of specificity and sensitivity estimates in the individual sources

 

Vertebral compression fractures. Four sources were included in the quantitative synthesis: N. Hong et al. [19], Oxytech Spine Fracture X-ray (2025)⁶, Airway (2025)⁷, and CINA-VCF (2025)⁸. The distribution of specificity and sensitivity estimates is presented in Figure 10. According to the meta-analysis, the pooled specificity was 0.863 (95% CI 0.670-0.952), and the pooled sensitivity was 0.819 (95% CI 0.650-0.917).

 

Figure 10. Forest plots of specificity (a) and sensitivity (b) of sources included in the diagnostic accuracy meta-analysis for the detection of vertebral compression fractures

 

The summary diagnostic accuracy of the included sources for vertebral compression fractures is presented as an HSROC curve in Figure 11.

 

Figure 11. HSROC curve of the diagnostic accuracy meta-analysis for the detection of vertebral compression fractures. The points represent individual studies, the blue square indicates the pooled estimate, and the solid line represents the HSROC curve. The horizontal and vertical lines reflect the uncertainty of specificity and sensitivity estimates in the individual sources

 

Scoliosis. Three AI services identified from grey literature sources were included in the quantitative synthesis: Oxytech Spine X-ray Scoliosis⁹, Esper.Scoliosis¹⁰, and A-rg-scoliosis¹¹. The distribution of specificity and sensitivity estimates for the individual services is presented in Figure 12. According to the Bayesian bivariate meta-analysis, the pooled specificity was 0.420 (95% post-hoc interval 0.152-0.747), and the pooled sensitivity was 0.988 (95% post-hoc interval 0.862-0.997).

 

Figure 12. Forest plots of specificity (a) and sensitivity (b) of AI services included in the diagnostic accuracy meta-analysis for the detection of scoliosis

 

The summary diagnostic accuracy of the AI services and the expected range of between-study variability are presented as an HSROC curve in Figure 13.

 

Figure 13. HSROC curve of the diagnostic accuracy meta-analysis of AI services for the detection of scoliosis. Black dots represent individual studies, the transparent diamond indicates the pooled estimate, and the solid line represents the HSROC curve. The dashed line outlines the 95% prediction region, while the grey shaded area represents the 95% posterior region of the pooled estimate calculated within the bivariate model

 

The results of the leave-one-out sensitivity analysis for each condition are provided in Supplement 2.

Discussion

The present systematic review showed that, when studies and grey literature sources with external validation were included, quantitative synthesis could be performed for only 5 of 12 musculoskeletal diseases (42%). More-over, the certainty of evidence according to GRADE was low for osteoporosis and vertebral compression fractures and very low for active sacroiliitis, knee osteoarthritis, and scoliosis (Supplement 3).

The inclusion of grey literature sources allowed for broader coverage of data on registered and already implemented AI services. However, compared with peer-reviewed publications, these sources often provided less detailed descriptions of methods, study samples, validation procedures, and statistical analyses. Therefore, findings based, among others, on grey literature sources were interpreted with caution and with consideration of the certainty of evidence according to GRADE.

For osteoporosis, our findings are consistent with previously published meta-analyses in which AI technologies demonstrated potentially high diagnostic accuracy; however, the interpretation of the findings was limited by the risk of bias, study heterogeneity, and insufficient external validation [28, 29]. The meta-analysis by L. Gao et al. reported high pooled specificity and sensitivity but also identified limitations related to patient selection [28]. In the study by N. Yamamoto et al. on opportunistic osteoporosis screening, the certainty of evidence was rated as low or very low [29]. Thus, the low certainty of evidence according to GRADE in our study reflects the overall fragmented and methodologically heterogeneous nature of the available evidence base.

For knee osteoarthritis, a previously published meta-analysis by S. Mohammadi et al. also demonstrated promising diagnostic ac-curacy of AI technologies; however, it was based predominantly on studies with internal validation and did not include grey literature sources [30]. The authors reported substantial heterogeneity, wide prediction intervals, and a high risk of bias, which limited the interpretation of the findings. In our review, the meta-analysis of knee osteoarthritis was based exclusively on three grey literature sources, as no peer-reviewed publications met the criteria for quantitative synthesis. Therefore, the certainty of evidence according to GRADE was rated as very low, despite formally high values for individual diagnostic accuracy measures.

For vertebral compression fractures, our findings are consistent with the meta-analysis by J. Jung et al., in which vertebral fractures were considered as part of the overall fracture group [31]. The authors also demonstrated high diagnostic accuracy of AI technologies; however, they included only English-language peer-reviewed publications and did not analyze grey literature sources. Patient selection and characteristics of the reference standard remained the main methodological limitations. Our meta-analysis combined data from peer-reviewed publications and grey literature sources.

Active sacroiliitis was addressed in a systematic review by S.J. Moon et al., in which AI technologies for detecting signs of sacroiliitis on MRI demonstrated potentially high diagnostic accuracy; however, a formal meta-analysis was not performed because of substantial heterogeneity among the studies [32]. External validation was performed in only 3 of the 10 included studies, while a high risk of bias was frequently associated with small sample sizes, the risk of overfitting, and insufficient transparency of validation procedures. Our findings similarly suggest that AI technologies may be considered promising for the detection of active sacroiliitis, although the evidence base remains limited.

For scoliosis, no previous systematic reviews or meta-analyses of the diagnostic accuracy of AI technologies were identified. Published reviews focused primarily on automated Cobb angle measurement, assessment of deformity severity, classification, and prediction of progression [33, 34]. Our meta-analysis partially addresses this gap; however, it was based on a small number of comparable studies, resulting in a very low certainty of evidence according to GRADE.

Potential applications of AI services in musculoskeletal diseases may include opportunistic screening for osteopenia and osteoporosis, detection of vertebral compression fractures, preliminary prioritization of patients based on radiological imaging findings, support for quantitative assessment of pathological changes, use of AI technologies as a “second opinion”, and longitudinal assessment of imaging studies. However, AI services should be regarded as an adjunctive tool to support physicians rather than as a standalone replacement for clinical assessment and expert radiological interpretation.

Study limitations

The search was restricted to the period from 2023 to 2025, which may have resulted in the omission of earlier relevant studies. The strict inclusion criteria reduced the number of eligible sources but allowed the assessment of more reproducible diagnostic accuracy. For grey literature sources, the completeness of reporting on study samples, the index test, and the reference standard was limited. In addition, because of the small number of studies, between-study variability was assessed predominantly descriptively, and some potentially relevant publications may not have been included because their full texts were unavailable.

Conclusion

The results of this systematic review and meta-analysis indicate the potential diagnostic value of artificial intelligence technologies for selected musculoskeletal diseases; however, the low and very low certainty of evidence according to GRADE limits the robustness and generalizability of the overall estimates. Quantitative synthesis could be performed for only some of the conditions because fewer than half of the sources that initially met the search criteria were included in the meta-analysis. For some conditions, meta-analysis became possible only after the inclusion of grey literature sources, the vast majority of which represented Russian AI services. However, this approach was associated with reduced certainty of evidence. Future research should focus on independent external validation, prospective study designs, and the use of standardized criteria for assessing diagnostic accuracy. There is also a need for studies evaluating the cost-effectiveness of artificial intelligence technologies for the diagnosis of musculoskeletal diseases in real-world clinical practice.

Disclaimers

Author contribution

Yafarova A.A. — study concept and design, literature search and review, data acquisition, analysis and interpretation, drafting the manuscript.

Erizhokov R.A. — data analysis and interpretation, editing the manuscript.

Petryaikin A.V. — study concept and de-sign, data analysis and interpretation, editing the manuscript.

Blokhin I.A. — statistical data processing, editing the manuscript.

Reshetnikov R.V. — data analysis and interpretation.

Rodionova L.G. — data acquisition, analysis and interpretation, editing the manuscript.

Varyukhina M.D. — data analysis and interpretation, editing the manuscript.

Omelyanskaya O.V. — study concept and design, editing the manuscript.

Vladzymyrskyy A.V. — study concept and de-sign, editing the manuscript, scientific guidance.

Vasilev Yu.A. — study concept and design, editing the manuscript, scientific guidance.

All authors have read and approved the final version of the manuscript of the article. All authors agree to bear responsibility for all aspects of the study to ensure proper consideration and resolution of all possible issues related to the correctness and reliability of any part of the work.

Funding source. This paper was prepared by a group of authors as a part of the research and development effort titled “Improving the accuracy of diagnostic imaging methods for spinal diseases” (USIS No. 126082024228-9) in accor-dance with the Order No. 1213 dated Novem-ber 27, 2025 “On approval of state assignments funded by means of allocations from the budget of the city of Moscow to the state budgetary (autonomous) institutions subordinate to the Moscow Health Care Department, for 2026 and the planned period of 2027 and 2028” issued by the Moscow Health Care Department.

Disclosure competing interests. The authors declare that they have no competing interests.

Ethics approval. Not applicable.

Consent for publication. Not required.

Use of artificial intelligence. No generative artificial intelligence technologies were used in the preparation of this manuscript.

 

¹ Rydkin S.V., Korepina E.M., Simkin N.V., Pavlov D.S. MOSMED.AI [computer software]. Moscow: Center for Diagnostics and Telemedicine; 2023. State registration certificate for computer software RU 2023614611, registered December 28, 2022.

² IB Lab GmbH. KOALA. 510(k) Premarket Notification. U.S. Food and Drug Administration; 2019 [cited 2026 Mar 16]. Available from: https://www.accessdata.fda.gov/cdrh_docs/pdf19/K192109.pdf; Radiobotics ApS. RBknee. 510(k) Premarket Notification. U.S. Food and Drug Administration; 2021 [cited 2026 Mar 16]. Available from: https://www.accessdata.fda.gov/cdrh_docs/pdf20/K203696.pdf.

³ CVisionRad — Knee Arthrosis. Mosmed.ai: Experiment on the Use of Innovative Technologies in Computer Vision; [cited 2026 Mar 16]. Available from: https://mosmed.ai/service_catalog/cvl-chest-ct-knee/.

⁴ Oxytech Knee Radiography [computer software]. Moscow: Oxytech; 2025.

⁵ IB Lab GmbH. KOALA. 510(k) Premarket Notification. U.S. Food and Drug Administration; 2019 [cited 2026 Mar 16]. Available from: https://www.accessdata.fda.gov/cdrh_docs/pdf19/K192109.pdf.

⁶ Oxytech Spine Fracture X-ray [computer software]. Moscow: Oxygen Technologies Rus LLC; 2025.

⁷ Airway Spine XR Fractures [computer software]. Moscow: Airway AI LLC; 2025.

⁸ Avicenna.AI. CINA-VCF. 510(k) Premarket Notification. U.S. Food and Drug Administration; 2024 [cited 2026 Mar 16]. Available from: https://www.accessdata.fda.gov/cdrh_docs/pdf24/K240612.pdf.

⁹ Oxytech Spine XR Scoliosis. Mosmed.ai: Experiment on the Use of Innovative Technologies in Computer Vision; [cited 2026 Mar 16]. Available from: https://mosmed.ai/service_catalog/oxytech-spine-xr-scoliosis/.

¹⁰ Esper.Scoliosis. Mosmed.ai: Experiment on the Use of Innovative Technologies in Computer Vision; [cited 2026 Mar 16]. Available from: https://mosmed.ai/service_catalog/esperscoliosis/.

¹¹ A-rg-scoliosis [computer software]. Moscow: Airway AI LLC; 2025.

About the authors

Adel A. Yafarova

Center for Diagnostics and Telemedicine

Author for correspondence.
Email: adyafarowa@yandex.ru
ORCID iD: 0000-0003-3002-1067
SPIN-code: 1670-3730

Cand. Sci. (Med.)

Russian Federation, Moscow

Rustam A. Erizhokov

Center for Diagnostics and Telemedicine

Email: ErizhokovRA@zdrav.mos.ru
ORCID iD: 0009-0007-3636-2889
SPIN-code: 2274-6428
Russian Federation, Moscow

Alexey V. Petraikin

Center for Diagnostics and Telemedicine

Email: alexeypetriakin@gmail.com
ORCID iD: 0000-0003-1694-4682
SPIN-code: 6193-1656

Dr. Sci. (Med.)

Russian Federation, Moscow

Ivan A. Blokhin

Center for Diagnostics and Telemedicine

Email: BlokhinIA@zdrav.mos.ru
ORCID iD: 0000-0002-2681-9378
SPIN-code: 3306-1387

Cand. Sci. (Med.)

Russian Federation, Moscow

Roman V. Reshetnikov

Center for Diagnostics and Telemedicine

Email: ReshetnikovRV1@zdrav.mos.ru
ORCID iD: 0000-0002-9661-0254
SPIN-code: 8592-0558

Cand. Sci. (Phys.-Math.)

Russian Federation, Moscow

Larisa G. Rodionova

Center for Diagnostics and Telemedicine

Email: RodionovaLG@zdrav.mos.ru
ORCID iD: 0009-0008-9862-8205
SPIN-code: 9122-6826
Russian Federation, Moscow

Maria D. Varyukhina

Center for Diagnostics and Telemedicine

Email: VaryukhinaMD@zdrav.mos.ru
ORCID iD: 0000-0001-8870-7649
SPIN-code: 7463-4645

Cand. Sci. (Med.)

Russian Federation, Moscow

Olga V. Omelyanskaya

Center for Diagnostics and Telemedicine; MIREA — Russian Technological University

Email: OmelyanskayaOV@zdrav.mos.ru
ORCID iD: 0000-0002-0245-4431
SPIN-code: 8948-6152
Russian Federation, Moscow; Moscow

Anton V. Vladzymyrskyy

Center for Diagnostics and Telemedicine; I.M. Sechenov First Moscow State Medical University (Sechenov University)

Email: VladzimirskiyAV@zdrav.mos.ru
ORCID iD: 0000-0002-2990-7736
SPIN-code: 3602-7120

Dr. Sci. (Med.), Associate Professor

Russian Federation, Moscow; Moscow

Yuriy A. Vasilev

Center for Diagnostics and Telemedicine

Email: npcmr@zdrav.mos.ru
ORCID iD: 0000-0002-5283-5961
SPIN-code: 4458-5608

Dr. Sci. (Med.), Associate Professor

Russian Federation, Moscow

References

Supplementary files