Blimdata

Medical imaging datasets, de-identified before they leave the hospital

A read-only PACS gateway, in-hospital de-identification and a radiologist annotation workspace, feeding a governed data room that licenses African imaging archives to AI teams.

The Blimdata homepage, headlined “Africa's medical imaging archives. Made AI-ready.”, beside a live pipeline panel showing X-ray, CT, MRI and ultrasound studies passing through de-identification.

blimdata.tech(opens in a new tab)

Client: Blimdata
Sector: Medical imaging data
Discipline: Web apps
Service: Web apps
Built with: Next.js, NestJS, Python, Vite, Tailwind CSS

The brief

Diagnostic models are increasingly expected to show they hold up across the populations they will be used on, and the imaging that would demonstrate it is already sitting on hospital PACS across Africa — earning nothing, because no hospital has the engineering capacity to extract DICOM at scale, the compliance tooling to prove de-identification to an international buyer, or a route to the companies that need it. The hard part was never moving files. It is that patient identifiers cannot be allowed to leave the building, which rules out the obvious architecture of uploading studies and cleaning them centrally — so the cleaning has to happen first, on the hospital's own network, in the same process that does the reading.

What we built

A gateway running inside the hospital network, pulling studies from the existing PACS over standard DICOM query/retrieve — read-only and outbound-only, so no inbound firewall rule is opened
De-identification in that same process: all 18 PHI identifiers stripped from DICOM headers, burned-in text detected and redacted from pixel data, and any study failing validation held back rather than transmitted
An annotation workspace where licensed radiologists label pathology, with AI pre-segmentation proposing masks and bounding boxes for a clinician to confirm, correct or reject
A governed data room holding versioned datasets, licensed by subscription or batch, with provenance records and de-identification attestations attached to every release
Role-based access and an immutable audit log across the whole path, AES-256 at rest and TLS 1.3 in transit
A public site carrying a two-sided pitch — hospitals on one side, AI and pharma teams on the other — without either audience having to read the other's argument first