AI That Reads Medical Scans Works — But Hospitals Can't Plug It In

Here's a finding that might surprise you: when hospitals try to use AI to read medical scans, the AI is rarely the problem. A team of researchers from Canada and the United States deployed imaging AI at six hospitals using PACS-AI, a free platform anyone can host on their own servers. It connects to PACS — the picture archiving system where hospitals store X-rays, CT scans, and MRIs (think of it as the hospital's giant photo library). The team's conclusion was blunt: the binding constraint is not how accurate the models are, but the infrastructure to route studies, display results, capture feedback, and audit what runs.

In plain terms, the AI can spot things in images, but someone has to make sure the right scan reaches it, the answer gets back to the right doctor quickly, and there's a record of what the AI said. That's unglamorous work, and it's where projects stall. At one hospital testing AI on angiography images (X-rays of blood vessels), the system completed 515 of 607 jobs — about 85%. The failures weren't the AI being wrong. They were cases where the necessary diagnostic views simply weren't captured in the first place.

Doctors seemed to like it. Of 638 clinician ratings collected, 78.1% were positive. That matters because trust is the currency of medical AI. A tool that annoys radiologists gets quietly ignored, no matter how clever it is underneath.

The paper also makes a point worth remembering: the team published honest readiness levels for every model — essentially a public label saying how mature or reliable each one is. They argue that being upfront about what a model can and can't do is itself a form of governance. In a field crowded with hype, that's refreshing. For patients, the takeaway is simple: AI in your hospital may arrive slower than the headlines suggest, not because the technology fails, but because the wiring takes time.