Bone density is not bone quality and the x-ray you already took might know more than the report says

I like to think of a good recipe as more than a list of ingredients. Flour, egg, and a bit of patience will get you noodles, but the ratio and the technique are what make them good noodles. Bone health has worked the same way for decades. The standard recipe for fracture risk has leaned almost entirely on one ingredient: How much bone a patient has, measured as bone mineral density (BMD) on a DXA scan. But how much bone you have is only part of the story. How that bone is organised — its microarchitecture, its quality — matters just as much, and it’s something BMD alone cannot capture.

That’s the idea behind Trabecular Bone Score (TBS), the technology Medimaps has spent over a decade building the evidence for. It’s also the idea behind our newest project, TBS Reveal, the subject of a study my colleagues and I recently published in eClinicalMedicine, part of The Lancet Discovery Science.[1]

TBS Reveal received CE MDR certification as a Class IIa medical device in June 2026, making it available for clinical use across Europe. I want to use this piece to walk through what we actually found, and — since a paper can only say so much — what I think it means once you step outside the methods section, particularly for departments in the UK.

Why this matters to radiology, not just endocrinology

Two things make osteoporosis a hard problem for a health system to catch early. First, screening pathways rely heavily on DXA, which isn’t universally accessible or routinely ordered. Second, BMD alone doesn’t fully capture fracture risk, as many fragility fractures occur in patients with osteopenia, or even normal BMD.

TBS Reveal was built to sit at the intersection of those two gaps: It derives a composite Bone Fragility Index from routine x-rays of the spine, abdomen, chest, or pelvis that already include two or more lumbar vertebrae — no new scan, no new visit, no new radiation dose.

That last point is the one I think Radiology will find most interesting. This isn’t a new screening pathway competing for scanner time. It’s a second read on an image that already exists.

What we actually tested, and why the dataset mattered

We worked with five clinical sites (four university hospitals and one private imaging centre, spanning Europe and the United States) and were deliberate about the order we brought data in. We started with two European sites (Italy and Austria) with broadly similar demographics, then deliberately went looking for the opposite: US sites chosen specifically to stress-test the model against more ethnically diverse populations than our largely European internal dataset.

When we broke results down by ethnicity, performance held up: Stable across groups that were meaningfully under-represented in training. That’s one of the results I’m proudest of, because a model is only as generalisable as the data it was tested against, not just trained on.

Across the full multinational dataset (more than 18,800 paired radiographs and DXA scans from over 11,100 adults) TBS Reveal showed accuracy of up to 90% in external validation, specificity of up to 96% across external cohorts, and an AUC of 0.83–0.86 across all cohorts.

What the numbers don’t say on their own

Here’s a limitation we were upfront about in the paper: Every patient in our dataset had already been referred for a DXA scan, meaning they were more likely than the general population to have a bone health concern already flagged. It wasn’t a pure screening population. The prevalence of very-high fragility risk we detected — around 10% across cohorts — landed close to published real-world reference values for that referred population, which gave us real confidence. But it also tells us exactly what to test next: How TBS Reveal performs in a true opportunistic setting, where someone has an x-ray for a completely unrelated reason and has never had a DXA at all.

It’s also worth being precise about what TBS Reveal isn’t. It’s not another tool that flags ‘low bone density’ in a population where most people getting a DXA already have some degree of osteopenia. That tells a clinician little they didn’t already suspect. What we built combines bone density and TBS-based microarchitecture into a single Bone Fragility Index aimed specifically at the smaller group of patients at very high fragility risk: The ones who most need to move to the front of the queue.

Where I think this actually fits into a UK pathway

I’m the AI person, not a radiologist, so I’ll say this with appropriate humility rather than as a clinical recommendation. But two settings seem like a natural fit.

1. Fracture liaison services see a steady stream of first-fracture patients and have to make quick calls about who gets a DXA referral versus immediate treatment. A tool flagging very-high fragility risk directly from imaging already being done could help prioritise that queue.
2. Orthopaedic and emergency settings: If a patient presents with a fracture and there’s underlying bone fragility no one has accounted for, hardware placed into that bone may not hold, and the patient can return months later with a more serious complication. Flagging that risk before surgery, from an x-ray already taken, is a meaningfully different conversation.

Longer term, I keep coming back to the fact that osteoporosis is a silent disease. By the time someone gets a first DXA, often a decade or more of risk has gone unnoticed. x-rays are common in ways DXA scans simply aren’t. If tools like this can responsibly extend earlier detection to younger adults, before the first fracture rather than after, that’s the version of this work I find most worth building toward.

TBS Reveal is intended exclusively for healthcare professionals and is not available for purchase by the general public. TBS Reveal is a Class IIa medical device, CE marked under EU MDR 2017/745 (CE 2460). FDA clearance is pending; TBS Reveal is not available for clinical use in the United States. Manufacturer: Medimaps Group SA, Chemin du Champ-des-Filles 36A, 1228 Plan-les-Ouates, Geneva, Switzerland. Always follow the instructions for use.

Reference

1. Gatineau, G., De Gruttola, M., Hind, K., et al. (2026). Validation of a deep learning model for bone fragility detection from conventional radiographs: an international cohort study. eClinicalMedicine, 95, 103974. doi:10.1016/j.eclinm.2026.103974

The content on this page is provided by the individuals concerned and does not represent the views or opinions of RAD Magazine.