– Written by Nor-Eddine Regnard, France, and Ali Guermazi, USA

INTRODUCTION

Sports imaging is a specialized subset of musculoskeletal imaging that focuses on acute and chronic pathologies commonly seen in athletes¹. Sports imaging primarily involves traumatic bone and joint injuries, including fractures, fissures, dislocations, and joint effusions, as well as muscular, myotendinous and myoaponeurotic, tendinous, cartilaginous, meniscal and labral lesions.

These abnormalities can of course be detected using standard imaging modalities: X-ray radiographs for fractures, dislocations, and joint effusions; ultrasound for joint effusions, bone fissures, and tendinous, muscular, myotendinous, and myoaponeurotic lesions; CT scans for fractures and bone fissures; MRI for fractures, bone fissures, muscular, myotendinous, and myoaponeurotic lesions, as well as tendon, meniscus, cartilage, and labral injuries; scintigraphy for bone fissures and fractures; and PET-CT for synovitis and bone marrow lesions.

Here, we will describe some of the work done on artificial intelligence and musculoskeletal imaging and its ramifications for sports imaging.

CURRENT SOFTWARE FOR AI AND MUSCULOSKELETAL RADIOLOGY

Numerous AI and musculoskeletal imaging software programs are in use worldwide (Table 1). They offer several key benefits:

• Time efficiency: Accelerating the detection and quantification of abnormalities and even assisting in report generation.
• Enhanced diagnostic accuracy: Improving sensitivity and specificity across different readers.
• Triage and prioritization: Streamlining workflow by prioritizing urgent cases.
• Decision support: Providing radiologists with a reliable second opinion.
• Error reduction: Minimizing diagnostic mistakes.
• Broader healthcare benefits: Saving time in emergency departments, reducing costs, and improving communication between clinicians and radiologists.

Artificial Intelligence applications in the field of musculoskeletal radiology are primarily focused on fracture detection in standard X-rays and CT scans. Several commercially available software packages are designed for the automatic detection of fractures in the limbs, pelvis, spine, and ribs in both adults and children using standard radiography.

BoneView by Gleamer is the best-validated AI tool in this field, (Figure 1) with the most scientific evidence. In several publications, it has been shown to reduce the number of missed fractures by over 30%, as well as saving significant time in the interpretation of standard bone X-rays^2,3. These findings have been confirmed in a pediatric population as well⁴.

In a natural prevalence study, BoneView software enabled the detection of a significant number of fractures, joint dislocations, joint effusions and focal bone lesions compared to standard radiologists’ reports⁵. The negative predictive value for fractures in traumatic radiography was 99.5%. Another article reported that BoneView enhanced MSK radiologists’ ability to detect wrist fractures, including occult fractures, when compared to gold standard CT images⁶. Aidoc’s C-Spine software detects cervical spine fractures in CT scans. In a recent study⁷, the software detected fewer fractures than the attending radiologists (sensitivity of 71.5% versus 88.2%) but detected many fractures.

 missed by the radiologists, including those in need of stabilizing therapy. Avicenna.AI also offers cervical vertebral fracture detection software available in Europe and the United States (Figure 2). Some software programs are available on the market for rib fracture detection in CT scans although they lack peer reviewed evidence (InferRead CT Bone by Infervision, RibFx by Aidoc).

New software programs such as Gleamer’s BoneCT (Figure 3) and Guerbet’s Liflow software detect focal bone lesions that are likely to be metastatic on thoraco-abdominopelvic scans. However, these programs are not currently CE or FDA approved. Several software packages are available for automatic calculation of bone age using the Greulich and Pyle method: BoneXpert by Visiana, BoneAge by Gleamer, PANDA by ImageBiopsy Lab, BoneAge by VUNO Med. A recent study showed that these software programs perform well, with fairly similar accuracy⁸.

Some software can automatically calculate orthopedic measurements for the spine (Cobb angle calculation, sagittal balance), the lower limbs, hips, feet, shoulders and knees. Gleamer’s BoneMetrics (Figure 4) and ImageBiopsyLab’s LAMA, HIPPO, FROG and SQUIRREL are the software packages with the strongest scientific support, demonstrating high measurement accuracy with significant time savings⁹. Several software programs can detect signs of knee osteoarthritis on standard radiography and quantify them using the Kellgren-Lawrence method. KOALA from ImageBiopsy Lab and RBKnee from Radiobiotics are two available software packages with good scientific support¹⁰.

Two software packages for MRI of the knee are largely restricted to the European market. Incepto’s KEROS software (Figure 5), which has demonstrated good performance in detecting ruptures of the anterior cruciate ligament¹¹ and in detecting and characterizing meniscus lesions; and Mediaire’s mdknee software, which mainly grades cartilage damage on MRI, but has not yet been validated in peer-reviewed journals.

Smart Soft’s CoLumbo software is the only MRI software for the lumbar spine to be marked in Europe and the USA. It detects and measures herniated discs, bulging discs, canal narrowing, nerve compression and spondylolisthesis, with an acceptable but improvable level of performance¹². Another program from Caerus Medical looks promising, with detection of degenerative disc disease (Pfirrman and Modic grades), disc herniation, disc bulging, vertebral fracture, posterior interapophyseal osteoarthritis, canal narrowing (Schizas grade) and foraminal narrowing (Figure 6). This software is not yet branded and there is no peer-reviewed scientific evidence about it.

This is a rapidly changing field, but to the best of our knowledge, there are no other commercially available software programs for other musculoskeletal imaging applications.

RESEARCH WORK IN AI AND MUSCULOSKELETAL RADIOLOGY

In this section, we will look at areas of musculoskeletal imaging methods that have been published, but which have no current commercial applications. Unlike most commercial software, these models are often trained on smaller, single-center datasets and focus on binary or highly specific tasks. As a result, they are not readily applicable to routine clinical practice and may face challenges with generalizability. They may, however, provide interesting approaches in the future.

Several studies have shown promising results for detecting hip arthritis and knee osteoarthritis on standard X-rays. Most of the data came from highly standardized clinical studies, such as OAI and MOST, which follow specific acquisition protocols and rigorous techniques that do not necessarily correspond to the average quality of X-rays performed everyday around the world¹³. External validation studies should be encouraged. In a study using the OARSI Atlas system for the hip, the deep learning model had accuracies of 83% for detecting femoral osteophytes, 65% for acetabular osteophytes, 81% for joint space narrowing, 89% for subchondral sclerosis, and 91% for subchondral cysts¹³. A deep learning model for assessing carpal instability on standard radiographs—measuring the scapholunate joint, scapholunate and capitolunate angles, and interruptions of carpal arches—has shown promising results, with accuracy surpassing that of most clinicians in detecting arc interruptions¹³. 

In knee MRI, approximately ten studies have investigated AI models for detecting meniscal tears in a binary manner (presence or absence), while only three have addressed multiclass classification (specific tear types). These studies suggest that AI models more easily detect complex meniscal tears than simple ones. The AUC for distinguishing between normal menisci and horizontal or radial meniscal tears was approximately 0.813.

Several studies have explored the ability of deep learning models to detect rotator cuff tendon injuries on MRI. Most relied on the coronal T2 sequence with fat saturation, but some used other sequences. The models demonstrated high accuracy in identifying complete supraspinatus tendon ruptures, with slightly lower performance for partial supraspinatus tears and infraspinatus or subscapularis tendon injuries. Most studies, however, used radiologist interpretations as the gold standard rather than surgical arthroscopic findings¹³. One study assessed a deep learning model’s ability to detect labral lesions on MRI, reporting good sensitivity (90%) but moderate specificity (73%).

In lumbar spine MRI, most deep learning models show a significant drop in performance during external validation studies. While they generally maintain good sensitivity, specificity is often lower, particularly for detecting disc bulges and foraminal stenosis. Performance appears to be more reliable when identifying disc herniations, canal stenosis, lateral recesses, and facet arthropathy¹³. Some studies have had promising results using deep learning models in hip MRI to detect labral lesions, cartilage defects, bone edema, and subchondral cysts. One of the studies evaluated a deep learning model’s ability to detect anterior talofibular and calcaneofibular ligament sprains, with encouraging results. Another study investigated the detection of triangular fibrocartilage complex lesions in wrist MRI, and also reported good performance.

RESEARCH WORK IN SPORTS IMAGING AND AI

These injuries result from repeated overuse of muscle groups or joints, as well as acute trauma. The anatomical location and types of injuries tend to be sport specific. The primary goals of the medical team managing professional athletes include early injury detection, assessment of severity, treatment guidance, estimation of recovery time, and prevention of recurrence. Imaging plays a crucial role in this process, helping radiologists select the most appropriate modality to establish a diagnosis, identify abnormalities, correlate findings with clinical symptoms, differentiate between old and recent injuries, assess severity, and monitor healing. The primary imaging modalities used in sports trauma are ultrasound, CT, and MRI.

Despite efforts to standardize reports, significant variation remains in how abnormalities are described in sports imaging. Additionally, imaging in athletes with pain is often normal despite thorough evaluation, sometimes leading to the monitoring of borderline abnormalities. Advanced quantitative imaging techniques, such as ultrasound elastography and diffusion tensor imaging in MRI, can offer additional insights in certain sports-related injuries. They are not widely available in all imaging centers, however, and come with their own limitations.

Artificial intelligence has the potential to enhance this field by standardizing the detection and characterization of lesions. By integrating imaging data with clinical information—such as symptoms and history of relapse—AI could provide more precise, validated, and reproducible predictions regarding the necessary immobilization period for optimal healing and relapse prevention¹.

Currently, there is very little literature on AI applications in this specific field. Several factors contribute to this gap:

Developing AI algorithms requires large, diverse datasets, but SRI cases are relatively rare and dispersed across different centers with varying imaging protocols.

Sports imaging is highly specialized, and training AI models would require expert radiologists for accurate annotation.

A lack of precise, integrated clinical data makes it difficult to develop reliable predictive algorithms.

Lastly, there is relatively limited collaboration between sports medicine, imaging specialists, and engineering research laboratories

CONCLUSION

Artificial intelligence in musculoskeletal imaging is a rapidly evolving field, with well-established applications primarily in conventional radiography, and which is widely used worldwide. AI is also being applied to MRI, particularly for the knee and lumbar spine, and is developing quickly.

AI in sports imaging is still in its early stages, however, with significant gaps in fundamental research to fill before practical applications can be realized. To bridge this gap, large-scale, multisite, and multidisciplinary data collection efforts are essential. Collaboration with engineering research laboratories will also be crucial to amass the critical volume of data needed to train robust AI models and to optimize data processing for clinical use.

Nor-Eddine Regnard, MD 

Reseau d’Imagerie Sud Francilien

Evry, France

Ali Guermazi, MD, PhD, MSC

Department of Radiology

Boston University School of Medicine

Boston, USA

Contact: guermazi@bu.edu

References

1. Martín-Noguerol T, Rossi I, Tol JL, Bencardino J, Guermazi A, Luna A. Past, present, and future in sports imaging: how to drive in a three-lane freeway. Eur Radiol 25 oct 2022. https://doi.org/10.1007/s00330-022-09193-5
2. Guermazi A, Tannoury C, Kompel AJ, Murakami AM, Ducarouge A, Gillibert A, Li X, Tournier A, Lahoud Y, Jarraya M, Lacave E, Rahimi H, Pourchot A, Parisien RL, Merritt AC, Comeau D, Regnard NE, Hayashi D. Improving radiographic fracture recognition performance and efficiency using artificial intelligence. Radiology 2022;302(3):627-36. https://doi.org/10.1148/radiol.210937
3. Duron L, Ducarouge A, Gillibert A, Lainé J, Allouche C, Cherel N, Zhang Z, Nitche N, Lacave E, Pourchot A, Felter A, Lassalle L, Regnard NE, Feydy A. Assessment of an AI aid in detection of adult appendicular skeletal fractures by emergency physicians and radiologists: a multicenter cross-sectional diagnostic study. Radiology 2021;300(1):120-9. https://doi.org/10.1148/radiol.2021203886
4. Hayashi D, Kompel AJ, Ventre J, Ducarouge A, Nguyen T, Regnard NE, Guermazi A. Automated detection of acute appendicular skeletal fractures in pediatric patients using deep learning. Skelet Radiol 2022. https://doi.org/10.1007/s00256-022-04070-0
5. Regnard NE, Lanseur B, Ventre J, Ducarouge A, Clovis L, Lassalle L, Lacave E, Grandjean A, Lambert A, Dallaudière B, Feydy A. Assessment of performances of a deep learning algorithm for the detection of limbs and pelvic fractures, dislocations, focal bone lesions, and elbow effusions on trauma X-rays. Eur J Radiol 2022:110447. https://doi.org/10.1016/j.ejrad.2022.110447
6. Jacques T, Cardot N, Ventre J, Demondion X, Cotten A. Commercially available AI algorithm improves radiologists’ sensitivity for wrist and hand fracture detection on X-ray, compared to a CT-based ground truth. Eur Radiol 2023. https://doi.org/10.1007/s00330-023-10380-1
7. Ruitenbeek HC, Oei EH, Schmahl BL, Bos EM, Verdonschot RJ, Visser JJ. Towards clinical implementation of an AI-algorithm for detection of cervical spine fractures on computed tomography. Eur J Radiol 2024:111375. https://doi.org/10.1016/j.ejrad.2024.111375
8. Gräfe D, Beeskow AB, Pfäffle R, Rosolowski M, Chung TS, DiFranco MD. Automated bone age assessment in a German pediatric cohort: agreement between an artificial intelligence software and the manual Greulich and Pyle method. Eur Radiol 2023. https://doi.org/10.1007/s00330-023-10543-0
9. Hayashi D, Regnard NE, Ventre J, Marty V, Clovis L, Lim L, Nitche N, Zhang Z, Tournier A, Ducarouge A, Kompel AJ, Tannoury C, Guermazi A. Deep learning algorithm enables automated Cobb angle measurements with high accuracy. Skelet Radiol 2024. https://doi.org/10.1007/s00256-024-04853-7
10. Brejnebøl MW, Hansen P, Nybing JU, Bachmann R, Ratjen U, Hansen IV, Lenskjold A, Axelsen M, Lundemann M, Boesen M. External validation of an artificial intelligence tool for radiographic knee osteoarthritis severity classification. Eur J Radiol 2022;150:110249. https://doi.org/10.1016/j.ejrad.2022.110249
11. Tran A, Lassalle L, Zille P, Guillin R, Pluot E, Adam C, Charachon M, Brat H, Wallaert M, d’Assignies G, Rizk B. Deep learning to detect anterior cruciate ligament tear on knee MRI: multi-continental external validation. Eur Radiol 2022. https://doi.org/10.1007/s00330-022-08923-z
12. Lehnen NC, Haase R, Faber J, Rüber T, Vatter H, Radbruch A, Schmeel FC. Detection of degenerative changes on MR images of the lumbar spine with a convolutional neural network: a feasibility study. Diagnostics 2021;11:902. https://doi.org/10.3390/diagnostics11050902
13. Ruitenbeek HC, Oei EH, Visser JJ, Kijowski R. Artificial intelligence in musculoskeletal imaging: realistic clinical applications in the next decade. Skelet Radiol 2024;53:1849-1868. https://doi.org/10.1007/s00256-024-04684-6
14. Guermazi A, Omoumi P, Tordjman M, Fritz J, Kijowski R, Regnard NE, Carrino J, Kahn CE, Knoll F, Rueckert D, Roemer FW, Hayashi D. How AI may transform musculoskeletal imaging. Radiology 2024;310:e230764. https://doi.org/10.1148/radiol.230764

Header Image by John Martinez (Cropped)