Why cartilage is so difficult to measure on a standard MRI

Your MRI report arrives and says 'mild cartilage thinning'. Your symptoms, on the other hand, have been anything but mild. That gap — between what a scan report says and what your knee actually feels like — is not unusual, and it points to a genuine limitation in how cartilage has traditionally been assessed.

The cartilage lining your knee joint is only 1–6 mm thick, making it one of the hardest tissues for an MRI to resolve precisely. Unlike bone, which shows up in sharp contrast, cartilage blends into the surrounding soft tissue. Worse, it has no direct blood supply, so it cannot repair itself the way most other tissues can. By the time structural thinning becomes clearly visible, meaningful degeneration has often been under way for some time.

The standard method for grading what an MRI shows — a system called MOAKS (MRI Osteoarthritis Knee Score) — asks a trained radiologist to read the scan by eye and assign semi-quantitative scores. That process is skilled, but it is inherently subjective. Two experienced radiologists reviewing the same images can reach different conclusions, and the same radiologist re-reading the same scan months later may score it differently. This variability is not a failing of the individuals involved; it is a structural limitation of manual assessment applied to tissue this thin and this subtle.

Osteoarthritis, the disease most directly tracked through cartilage MRI, affects an estimated 240 million people worldwide and ranks as the fourth leading cause of disability globally. The inability to track early cartilage loss objectively — before joints deteriorate significantly — remains a recognised clinical gap.

What the new AI models actually do inside an MRI scan

Two distinct models sit at the heart of the May 2026 Springer chapter by Wen and Ye: the nnAtrousU-Net and the nnAtrousTransFormer. Both are trained to read knee MRI images and automatically trace — or 'segment' — the boundaries of cartilage and bone, slice by slice through the scan volume.

The nnAtrousU-Net uses a technique called atrous, or dilated, convolution. Rather than examining each pixel only in the context of its immediate neighbours, the model samples a wider neighbourhood across the image while keeping its resolution intact. The analogy is stepping back from a painting: you take in the full composition and the sweep of the brushwork without losing the ability to see individual strokes. In a knee MRI, this matters because cartilage is thin and irregular — the model needs both fine-grained detail and broader anatomical context to draw an accurate boundary.

The nnAtrousTransFormer adds a second capability borrowed from language-processing AI: a Transformer attention mechanism. This allows the model to link a fragment of cartilage in one corner of the scan to a related fragment elsewhere in the same volume — the way a reader connects a pronoun back to the noun it refers to, several sentences earlier. For thin, sometimes discontinuous cartilage running across multiple MRI slices, that global awareness improves boundary precision.

Both models were trained on the publicly available OAI-ZIB knee MRI dataset, with deliberate augmentation — rotation, noise injection, contrast variation, and resolution shifts — to prepare them for real-world inconsistencies between different scanners and acquisition settings. Training ran on consumer-grade NVIDIA RTX 3090 graphics cards, not specialist supercomputing infrastructure, which is relevant when considering eventual deployment in a clinical setting.

What the accuracy numbers actually mean for a patient

Numbers such as '90.46 ± 2.86' carry little weight without a translation. The Dice Similarity Coefficient — DSC — is essentially an overlap score: when the AI draws its cartilage boundary and the expert radiologist draws theirs independently, a DSC of 90% means the two outlines agree on nine-tenths of the cartilage area. Think of two clinicians sketching the same border on a lightbox, then holding the sheets up together; at 90%, almost all of one drawing lies directly on top of the other.

The Average Surface Distance figure sharpens that picture further. On femoral cartilage, the nnAtrousU-Net's predicted boundary sits on average just 0.17 mm from the expert's line — and no more than 0.21 mm for the tibial surface. Given that the full cartilage layer is only a few millimetres deep, a disagreement of less than a quarter of a millimetre is, in practical terms, negligible. It places the model's precision within the same order of magnitude as natural biological variation between scans.

Bone segmentation scores ~98.76 DSC on the same benchmark. Bone edges are sharp, high-contrast, and unambiguous on MRI — so that figure is expected, and it confirms the pipeline is working correctly. The gap between ~99% for bone and ~90% for cartilage reflects exactly the difficulty described earlier: cartilage is thin, low-contrast, and structurally irregular. Closing that gap to 90% — and holding it — represents meaningful progress.

For someone being monitored for early osteoarthritis, the reproducibility point may matter most. The AI draws the same boundary each time the same scan is analysed, removing the reader variability that affects manual scoring. A 0.5–1 mm reduction in cartilage thickness over 12 months can be a significant early signal; an objective, consistent measurement method is what makes that signal detectable.

Thickness and shape over time — the real clinical prize

Knowing that an AI model can match an expert's outline on a single scan only matters clinically if the same tool produces the same outline six months later — on a different scanner, processed by a different workstation, reviewed by whoever happens to be on call.

What the models presented in the 2026 Wen and Ye chapter actually output is not a simple 'normal' or 'abnormal' label. Across both axial and sagittal planes, they generate full quantitative maps of cartilage thickness and three-dimensional shape. That distinction is significant. A clinician monitoring early osteoarthritis does not need to know that cartilage is thin; they need to know whether it is thinner than it was, by how much, and precisely where the change is occurring. Millimetre-scale reductions tracked consistently over 12 or 24 months — with the same boundary-tracing logic applied each time, without the inter-reader drift described earlier — are what turn a series of MRI images into a true monitoring record. Thickness and shape are, accordingly, the parameters that clinicians following OA progression or outcomes after regenerative treatment most want to quantify longitudinally.

Structural thinning, though, is only part of the cartilage picture. Techniques such as T2 mapping and T1ρ imaging detect biochemical changes — shifts in water content and collagen organisation — that may precede visible structural loss. These compositional methods depend on an accurate spatial framework: AI segmentation provides it, defining precisely where the cartilage sits so that the biochemical signal can be mapped within those boundaries. One technique builds on the other, though routine clinical integration of this combined approach has yet to be established.

The gap between a research dataset and your next MRI appointment

The OAI-ZIB dataset is one of the most rigorous cartilage benchmarks available, but it is a curated research collection — not a cross-section of every scanner model, field strength, and patient population found in everyday clinical practice. Achieving a DSC of 90.46% on that benchmark is the proof that the architecture works at the required precision. It is not, by itself, evidence of clinical readiness.

The steps between that proof and routine use are specific. Prospective validation across diverse patient cohorts — different ages, body compositions, disease stages, and scanner hardware — comes first. Integration with radiology reporting workflows follows. In Great Britain, AI software classified as a medical device requires UKCA certification under MHRA oversight; no such submission is described in the 2026 Wen and Ye chapter, nor should one be expected at this stage of research. What the chapter does establish is something more concrete: that the segmentation precision required for meaningful cartilage monitoring is now achievable on consumer-grade hardware, removing one barrier that once made clinical deployment look distant. Regulatory and prospective validation work remains the realistic frontier — and knowing precisely what that pathway demands is what distinguishes publishable science from speculative promise.

What this research means for patients seen at MSK Doctors

Research like the Wen and Ye chapter shapes how AI-assisted MRI analysis is developed beyond the academic setting. MSK Doctors' Computer Vision Lab in Sleaford sits within this same tradition — applying computer vision and deep learning to musculoskeletal MRI with the goal of producing objective, reproducible cartilage measurements rather than the kind of summary impressions that vary between readers.

That work feeds into onMRI™, the group's AI-driven MRI analysis platform (patent-pending), which addresses the same structural segmentation and quantitative measurement problem the Springer chapter examines: making serial MRI reads reproducible across sessions and clinicians, so that imaging becomes a true monitoring record rather than a point-in-time snapshot. The same caveats apply — the platform is in active development, and prospective validation across diverse patient populations remains the appropriate standard for any AI imaging tool.

For patients who cannot tolerate a closed-bore scanner, the Open MRI at Sleaford extends access without sacrificing quantitative analysis; onMRI™ is designed to work with low-field as well as conventional imaging. For anyone monitoring early osteoarthritis or tracking outcomes after regenerative treatment, that represents an additional objective data layer alongside clinical assessment.

MSK Doctors accepts patients without a GP referral and without a waiting list. If you want to understand what your MRI can tell you about your cartilage health, a consultation can be booked directly at mskdoctors.com.

1. [1] Articular Cartilage Damage — Wikipedia. https://en.wikipedia.org/?curid=19057920 https://en.wikipedia.org/?curid=19057920
2. [2] Cartilage — Wikipedia. https://en.wikipedia.org/?curid=166945 https://en.wikipedia.org/?curid=166945