Turning a million open-access papers into a million usable medical image-text pairs sounds, described in one sentence, like a data processing task.

Every team building AI for healthcare runs into the same wall eventually: the data that would make their model genuinely useful is the data they're least able to access.

Real clinical records actual patient scans, actual case notes, actual outcomes sit behind two barriers that exist for good reason. Privacy law protects patients from being identifiable in a dataset that might circulate for years. Licensing protects the hospitals and institutions that generated the data in the first place, and often comes with a price tag most research teams can't justify.

So researchers do what's rational: they go where the data is open. And for medical AI, that almost always means PubMed Central (PMC) the free, full-text archive of biomedical literature. It's enormous, it's public, and every figure, chart, and clinical image embedded in millions of papers is technically accessible.

The problem is that "accessible" and "usable" aren't the same thing.

When we looked closely at one of the most widely referenced PMC-derived datasets 24 million image-text pairs we found that 80.3% of the images inside it were not medical at all. Bar charts. Line graphs. Screenshots of software interfaces. Photos that happened to appear in a biomedical paper but had nothing to do with a clinical finding. All of it labeled, structurally, the same way as an actual diagnostic image.

This matters because of how these datasets get used. A model trained on a dataset like this doesn't know that 4 out of 5 of its "medical" examples aren't medical. It just learns from whatever it's given, and if most of what it's given is noise, the model absorbs that noise as if it were signal. The result is a system that looks like it was trained on a huge medical corpus, but was functionally trained on a much smaller, uncontrolled slice of it without anyone knowing exactly how small, or how uncontrolled.

This is the quiet failure mode in medical AI data sourcing. Nobody sets out to build a low-relevance dataset. It happens by default, because filtering PMC properly actually verifying which images are clinically meaningful is slow, expensive, and requires people with medical training, not just an automated script.

We think that step shouldn't be optional. It's the difference between a dataset you can defend in a methods section and one that quietly undermines whatever gets built on top of it.

The fix isn't complicated in concept, even if it's hard in practice: manual review, by people who can actually tell a diagnostic image from a chart, at a scale that matches the size of the corpus. We'll walk through exactly what that looked like and the relevance numbers it produced in the next piece in this series.

For now, the takeaway for any team sourcing medical training data is simple: check composition before you trust scale. A 24-million-pair dataset and a 1-million-pair dataset aren't comparable on size alone if one of them is mostly not medical.