Artificial intelligence in musculoskeletal ultrasound: muscle injuries, peripheral nerves, and hyaluronic acid

How artificial intelligence is being studied to grade muscle injuries, automatically measure peripheral nerves, and support ultrasound-guided hyaluronic acid injections, and what the evidence already shows versus where it remains a promise.

Artificial intelligence (AI) is already part of daily practice in some areas of ultrasound, for example in triaging thyroid and breast nodules, but in musculoskeletal (MSK) ultrasound the picture is quite different: still early-stage, with few studies, almost no external validation, and no clinically established algorithm. A systematic review published in 2024 in Radiologia Medica identified only 16 studies on AI in MSK ultrasound published between 2020 and 2023, most using internal validation only, without confirmation in other populations or centers. This article reviews what current research shows in three areas of direct clinical relevance: grading muscle injuries, evaluating peripheral nerves, and ultrasound-guided procedures with hyaluronic acid, making clear where the technology already helps and where it remains a promise.

Muscle injuries: from the BAMIC classification to computer-assisted quantification

The British Athletics Muscle Injury Classification (BAMIC), described by Pollock and colleagues in 2014, has become the most widely used language for grading muscle injuries in athletes. The system combines the anatomical location of the injury, myofascial, proximal, distal or central myotendinous junction, or intratendinous, with the grade of tissue damage, from 0 to 4, providing a severity estimate that correlates with expected return-to-sport time. Although originally described using magnetic resonance imaging, BAMIC is also widely applied to ultrasound for the initial evaluation of acute muscle injuries, since ultrasound is fast, accessible, and can be performed pitch-side. A recognized limitation is that ultrasound grading depends heavily on the examiner's experience and image quality, and has historically shown more inter-observer variability than MRI-based grading.

It is precisely this variability that drives interest in quantitative and AI-based tools. A 2024 study published in Insights into Imaging showed that shear wave speed, measured by shear wave elastography, remains reduced in the hamstring after injury even in athletes already cleared to return to sport, suggesting a quantitative biomarker of tissue recovery that goes beyond visual inspection of the image. Software such as MyoVision-US, described by Wen and colleagues in Scientific Reports in 2025, demonstrates that neural networks can automatically extract muscle thickness, cross-sectional area, and echogenicity with excellent agreement to manual measurement, with intraclass correlation coefficients between 0.85 and 0.99, cutting image-analysis time from hours to seconds. These tools do not yet automatically grade acute muscle injuries or replace clinical interpretation, since they were validated for general muscle architecture rather than acute injury, but they point toward a future of AI-assisted quantification of injury severity, complementing BAMIC grading performed by the examiner.

Peripheral nerves: from the manual caliper to automatic segmentation

Sonographic evaluation of peripheral nerves is now an essential part of diagnosing entrapment neuropathies, such as carpal tunnel syndrome, at the median nerve at the wrist, and ulnar neuropathy at the elbow. Measuring the nerve's cross-sectional area is the most-used parameter: a nerve enlarged at the compression point suggests edema and axonal distress, often before relevant changes appear on nerve conduction studies. This measurement, however, is done manually, requires specific training, and takes examiner time, which is driving the development of automatic segmentation algorithms.

A recent example is the work of Moser and colleagues, published in Scientific Reports in 2024, who trained a U-Net-shaped neural network on 2,355 manually segmented images from 51 wrists, among carpal tunnel patients and healthy controls, to automatically locate and measure the median nerve. The algorithm reached a Dice coefficient of 0.76 versus manual segmentation, with a median difference of about 11 percent in the calculated area. When this automatic measurement was used to classify wrists as normal or pathological, diagnostic accuracy was 75 percent, versus 81 percent when the measurement was made manually by an expert. In other words, AI already comes close to human performance but still trails slightly behind it, and continues to struggle with anatomically atypical cases, such as bifid nerves or low-contrast images.

Placing this finding in a broader context, the systematic review cited earlier found that most AI studies in MSK ultrasound, including those on peripheral nerves, tendons, muscle, and cartilage, used only internal validation, with no multicenter external validation study reported at that time. In practice, this means automatic nerve-segmentation algorithms are, for now, promising research tools for standardizing the screening of entrapment neuropathies, not validated substitutes for an examination performed by an experienced sonographer.

Ultrasound-guided hyaluronic acid: solid evidence for joint preservation

Unlike the two previous topics, intra-articular hyaluronic acid injection, or viscosupplementation, is not an AI-based novelty. It is an established practice, used for decades as a strategy for pain control and joint preservation in patients with osteoarthritis, particularly before considering arthroplasty. What has changed in recent years is increasingly robust evidence that guiding the needle by real-time ultrasound, rather than relying only on palpable anatomical landmarks, the so-called blind technique, improves both procedural accuracy and clinical outcomes.

A systematic review with meta-analysis published in 2024 in the Australasian Journal of Ultrasound in Medicine pooled four randomized controlled trials, with 338 participants in total, all in knee osteoarthritis, and found a consistent advantage for ultrasound guidance: 96 percent intra-articular placement accuracy versus 78 percent with the blind technique, a large effect on reducing procedural pain, a large effect on functional improvement, and higher immediate patient satisfaction, with a score of 4.9 versus 4.1 in one study, with p equal to 0.01.

It is important to note where this evidence is concentrated: the four trials included in the meta-analysis dealt exclusively with the knee, and no comparative study of equivalent quality was identified for the hip or hand joints. This does not mean image guidance is dispensable in those other joints. On the contrary, for technically deeper or harder-to-access joints such as the hip, shoulder, or ankle, anatomical logic and current clinical practice favor ultrasound use. But in terms of randomized-trial evidence specifically on viscosupplementation, the knee is by far the best-studied joint today.

Where technology-assisted guidance is heading

Real-time guidance with conventional ultrasound is therefore already a mature, proven technology. What remains in the research phase is the use of AI to assist that guidance. Recent narrative reviews, such as one published in 2026 in Frontiers in Artificial Intelligence on shoulder injection, describe deep segmentation networks, with architectures such as nnU-Net and U-Net, able to identify periarticular structures, such as the humeral head and rotator cuff, with Dice coefficients above 0.90 in feasibility studies, along with real-time needle-tracking prototypes and navigation systems already registered with regulatory agencies, mainly in China. In a related field, a pilot randomized controlled trial tested AI-based needle tracking in ultrasound-guided regional anesthesia training, showing technical feasibility in that related context.

The important caveat is that the authors of these reviews themselves acknowledge a lack of direct clinical studies measuring, for example, the first-pass success rate of AI-guided joint injections compared with conventional ultrasound technique. The most optimistic figures cited in this literature are generally projections or results from small-scale feasibility studies, not confirmed outcomes from robust clinical trials. In today's practice, the tool with real, ready-to-use evidence remains real-time ultrasound itself in the hands of a trained examiner. AI-assisted navigation is a research direction worth following, not a requirement of current practice.

What this means in clinical practice

Bringing the three topics together, a common pattern emerges. In grading muscle injuries, BAMIC remains the clinical reference standard, and quantitative tools such as elastography help refine prognosis but do not replace clinical examination and expert interpretation. In peripheral nerve evaluation, automatic segmentation algorithms already come close to human accuracy and may, in the future, help standardize screening for entrapment neuropathies, but they still need external validation before entering routine practice. For joint pain and joint preservation, on the other hand, ultrasound-guided hyaluronic acid injection, with no need for AI at all, is already evidence-based practice with strong support, particularly for the knee, and should be the standard technique whenever available.

For those who practice or teach musculoskeletal ultrasound, the practical lesson is twofold: master today what already has robust evidence, structured muscle-injury grading, careful peripheral nerve measurement, and precise ultrasound guidance for injections, while closely following AI research in these three areas, because it is quite likely that, in the coming years, part of these tasks will gain real computational support that has been validated and tested outside the research lab that created it.