What can AI actually tell your surgeon that a scan can't?

Your MRI shows cartilage thinning — but how fast is it changing, and is the way you move making it worse?

That gap between what a scan reveals and what is actually happening to a joint under load sits at the heart of modern orthopaedic assessment. A standard MRI captures anatomy at a single moment: it can show worn cartilage, a narrowed joint space, or a fraying meniscus, but it cannot show how smoothly the knee bends as you rise from a chair, whether one hip is compensating for the other, or whether the loading pattern across your joint is accelerating the damage. Two patients can present with near-identical imaging findings and follow completely different clinical trajectories — because the missing variable is usually function, not structure.

Conventional examination helps, but it is observer-dependent and difficult to reproduce precisely across appointments or between clinicians. Measurable, objective data on how a joint moves — and how that movement is changing over time — has largely been absent from the routine consultation.

That is the gap domain-specific AI exists to fill. The tools that address it bear little resemblance to general-purpose language models: they analyse spatial and biomechanical data that text-based systems cannot perceive. The MSK Computer Vision Lab, working out of the ENRICH-MSK research programme, has built and clinically validated two such tools — and both are already in active use.

Why a language model can't read your knee

Think of a language model like ChatGPT as an extraordinarily well-read librarian. Ask it to summarise the literature on knee osteoarthritis, draft a patient letter, or explain what a Kellgren–Lawrence grade means — and it will do so fluently. What it cannot do is look up from the desk and watch you walk across the room.

That limitation is structural, not a matter of how 'intelligent' the system is. Language models are built to process text: they learn by predicting the next word across billions of written documents, and their output is always language. There is no mechanism in that architecture for perceiving a three-dimensional MRI volume, distinguishing articular cartilage from subchondral bone across hundreds of image slices, or computing the angle of knee flexion from a video frame. Those tasks require computer vision — a fundamentally different class of AI that reads pixels, tracks anatomical landmarks in space, and quantifies motion over time.

Orthopaedic assessment sits squarely in the visual and spatial domain. Cartilage lives in a 3D structure that changes in ways invisible to the naked eye across consecutive scans. Biomechanical risk accumulates in movement patterns — subtle shifts in loading, compensatory hip drift, reduced smoothness of knee flexion — that no text corpus can encode. Answering those clinical questions demands AI that can see, measure, and compare spatial data directly.

This is not a criticism of general-purpose language models; they have genuine roles in healthcare administration and clinical summarisation. The point is simply that the clinical questions that matter most in orthopaedic care — how is this joint loaded, and is the cartilage changing? — are spatial problems. They need spatial tools.

The research behind the MSK Computer Vision Lab

The academic infrastructure behind this work dates to 2020, when Prof Paul Lee established the ENRICH-MSK partnership — a formal collaboration between the University of Lincoln and MSK Doctors spanning engineering, computer science, sports science, and life sciences. More than £1 million in research and innovation funding followed, supporting a dedicated MSK research centre that includes Lincolnshire's only musculoskeletal open MRI scanner.

The lab's principal peer-reviewed output is a pair of chapters in Musculoskeletal Regeneration Medicine (Springer, 2025), a volume edited by Prof Lee — giving the research both authorial and editorial credibility in a single publication. One chapter addresses MRI cartilage segmentation; the other, markerless motion capture for knee osteoarthritis. Research from the lab was also published in the Journal of Arthritis in 2022, formalising the clinical case for the motion capture approach. The programme is ongoing: the 2026/27 sports science placement at MSK Doctors explicitly includes building AI-supported movement analysis workflows alongside the research team in clinic.

Two complementary strands run through this body of work. The first reads imaging data — training algorithms to segment cartilage from MRI volumes with quantified, reproducible precision. The second reads movement data — extracting biomechanical biomarkers from video, without sensors or laboratory equipment. What distinguishes this programme from many academic groups working on knee AI is that it pursues both strands in parallel and has taken each from prototype to active clinical use on hardware — a single consumer-grade GPU — that is realistic outside specialist academic centres.

Measuring cartilage on a scan — more precisely than the human eye alone

Imagine a radiologist marking the outline of cartilage on every image slice of your MRI — hundreds of them, each demanding the same concentration as the last. The MSK Computer Vision Lab's nnAtrousU-Net does something closely analogous: it traces the exact boundary of articular cartilage from bone across each slice of a 3D knee scan, automatically, in seconds.

Accuracy matters here, and the lab has measured it. Benchmarked against the OAI-ZIB dataset — an established academic standard for knee MRI cartilage segmentation — nnAtrousU-Net achieved a Femoral Cartilage Dice Similarity Coefficient (DSC) of 90.46±2.86, outperforming all published comparators in the comparison table at the time of publication. A DSC above 90% means the AI's cartilage outline closely matches what expert human annotators would draw. The companion architecture, nnAtrousTransFormer-2, reached 90.08, making both models competitive. These results come from a controlled academic benchmark rather than live clinical deployment — a distinction worth keeping in view when interpreting what the technology currently offers.

For a patient, the primary value is reproducibility across serial scans. Human readers of the same MRI can measure cartilage thickness differently on different days or at different centres. An algorithm gives the same answer to the same image every time. For someone being monitored over months or years — whether cartilage is stable, thinning gradually, or responding to treatment — that consistency changes what the data can tell a surgeon in ways that a single, subjective read cannot.

The clinical expression of this research is onMRI™, MSK Doctors' AI-driven MRI analysis platform (patent-pending), which applies quantitative segmentation to support more reproducible cartilage assessment over time.

What watching you sit down reveals about your joints

A 15-second clip of someone rising from a chair looks unremarkable. What the AI extracts from it is not.

Using a standard RGB camera — no reflective markers, no wearable sensors — the system tracks joint positions frame by frame during sit-to-stand and squat movements. It is not recording what a person looks like; it is computing how each joint moves through the task: the arc of knee flexion, how smoothly that arc unfolds, how the upper body shifts as the leg engages. These micro-patterns, invisible to the consulting-room eye, carry clinical information.

A 2022 paper in the Journal of Arthritis (Armstrong, Wen, Zhang, Ye, and Lee) formalised what the lab's markerless approach can reliably detect. Two biomechanical signals emerged as statistically significant markers for knee pain: the smoothness of the knee's maximum flexion curve during the movement, and the cumulative acceleration of elbow flexion alongside it. Both are too subtle for a clinician watching in real time to notice unaided — yet both correlated with knee pain status at p<0.05.

MAI Motion® — the deployed clinical platform, registered with the UKCA and MHRA — applies this methodology in practice. From movement data captured during the assessment, it derives a 'Motion Age' biological age score benchmarked against population norms, giving patients and clinicians an objective, longitudinal picture of how movement is changing over months. The initial assessment is clinician-supervised; subsequent re-scans can be completed via a smartphone app from home, so the measurement stays consistent whether a patient is seen in Sleaford or reassessing themselves six weeks later.

Looking further ahead, the same RGB pose-estimation method could in principle support community-level gait screening — identifying patterns associated with early joint deterioration before someone has thought to seek a consultation.

What this means if you're a patient booking now

The science described above reaches a patient as something much simpler: you arrive, you move, and the clinician has numbers where before there were impressions.

At an MSK Doctors movement assessment, MAI Motion® captures joint kinematics as you perform functional tasks — sit-to-stand, squat, or gait — translating the same biomechanical signals the lab has validated into an objective report that feeds directly into treatment planning. Where imaging is also indicated, the group's Sleaford site (NG34) houses Lincolnshire's only dedicated musculoskeletal open MRI scanner. For patients who need a more detailed biomechanical picture, the complementary 3D infrared motion capture service applies millimetre-precision joint tracking to surface movement compensations that standard examination typically misses.

Neither pathway requires a GP referral. MSK Doctors accepts patients directly at its Sleaford (NG34) and Grantham (NG31) clinics, without NHS-style waiting lists. London-based patients can access equivalent assessment through the group's London arm, the London Cartilage Clinic. To arrange a movement or joint assessment, visit mskdoctors.com.