ArchitectureBuilding the
system up.
Software for radiology departments. Eight layers, added one at a time — each only once you've seen what's missing without it.
A radiology department is already about seven systems.
Hospital IT ordinary software
Imaging speaks its own protocol
Different vendors, bought at different times — and nothing here talks to anything else directly.
The standard is the store, not a layer on top of it.
Most vendors design their own database and offer an export. We don't translate into FHIR at the boundary — the record is already in that form.
One standard interface, not one program per system.
The routing is HL7's own publish/subscribe, built into the data store. Anything that speaks the standard connects as it is — anything that doesn't gets one adapter, written to that same interface.
A custom router — a program per connection
Following a standard buys you nothing here. Every link is written from scratch, then maintained on its own.
One standard interface — HL7 pub/sub, in the data store
Standard systems need no integration work at all. The exception costs one adapter — and the next like it reuses it.
The record can now tell the other systems.
The rules sit beside the data, and the repository delivers. The systems on the right are the hospital's own — unchanged.
The scanners speak a protocol from the nineties.
A seven-figure machine installed fifteen years ago isn't being replaced to suit our software. So we implement its services ourselves.
A network-facing parser for a binary protocol — the kind of code where memory safety does real work.
One identifier runs through all of it.
The accession number — the one already printed on the requisition. It threads the order, the appointment, the images, the report and the billing.
Inventing our own would mean keeping a mapping table. But at two in the morning, the person reconciling two systems is holding a piece of paper.
Native desktop, not a browser tab.
A floating dictation window is an OS window. Scrolling hundreds of CT slices needs frame timing a tab won't promise. And one installed binary is easier to approve than a browser we'd then be patching.
The radiologist talks. The system drafts. They sign.
Everything before this makes us comparable to what they already have. This is why they'd move.
The model runs on hardware inside the hospital — a significant share of sites won't let patient data leave the building at all, which rules a hosted model out entirely.
A separate system per jurisdiction.
The usual approach is one deployment for everyone, with a column marking who each row belongs to. A legal boundary shouldn't be a filter inside a query.
Within a region, which hospital a user belongs to comes from their signed login.
Enforced in the server, not the screen.
Before an MRI there's a safety screening — metal implants, pacemakers, anything the magnet would move.
In a lot of systems that's a checkbox and a disabled button. Here the server refuses the operation, so no screen, script or API call gets around it.