The platform
From hospital PACSto training-ready dataset.
Brokering raw folders of scans is low margin and high risk. Blimdata is built the other way round: a controlled pipeline that turns an archive into a structured, defensible, premium AI asset — with a gate at every stage.
-
Stage 01
Secure ingestion from the existing PACS.
-
Stage 02
On-premise de-identification and validation.
-
Stage 03
AI-assisted radiologist annotation.
-
Stage 04
Versioned licensing through a governed data room.
End to end
Four stages. Each one a gate.
A study that fails a stage does not proceed to the next one. That is the whole design.
Secure ingestion
A containerised gateway inside the hospital network queries the PACS over standard DICOM C-FIND and retrieves matching studies with C-MOVE, under a read-only application entity title, on a schedule and rate limit the hospital sets.
De-identification
Header identifiers are removed or replaced, dates are shifted consistently per subject, and burned-in text is detected and redacted from pixel data. A validation pass re-scans the output; anything that still trips a rule is quarantined, not transmitted.
Annotation
Radiologists work in a browser workspace with AI pre-segmentation. Labels follow a versioned annotation protocol, a defined proportion is double-read, and disagreements go to senior adjudication.
Licensing
Completed studies are assembled into versioned datasets, published to a governed data room, and released under licences that specify permitted use, training rights, redistribution and retention.
DICOM handling
The format is the hard part.
DICOM is not one format so much as a family of conventions, unevenly implemented across two decades of scanner firmware. Handling it properly is most of the engineering, and it is where naive pipelines leak identifiers.
- High-fidelity pixel extraction across transfer syntaxes, including compressed and multi-frame objects.
- Programmatic metadata modification that preserves diagnostic validity while removing identity.
- Private and vendor-specific tags handled explicitly rather than passed through by default.
- Consistent per-subject date shifting, so temporal relationships survive de-identification.
- Structured cataloguing by modality, body region, study description and acquisition parameters.
Quarantined studies stay inside the hospital network for local review.
Architecture
PHI never leaves the hospital.
The gateway that reads your PACS is the same process that strips the identifiers. Re-identifiable data is never transmitted, never staged in our cloud, and never held by us — because the de-identification step runs before anything touches the wire.
This is not a policy setting that somebody could switch off. It is where the software runs.
Annotation workspace
Ten times faster.Still clinician-decided.
Manual annotation is the cost centre that decides whether a dataset business works. We attack the time per study without removing the clinician from the decision.
- Server-side models pre-segment structures and propose bounding boxes before the reader opens the study.
- The clinician confirms, corrects or rejects every proposal — nothing is auto-accepted into a dataset.
- Keyboard-first review, hanging protocols and study-level batching to remove interface overhead.
- Versioned annotation protocols, so a label always carries the definition it was produced under.
- Inter-reader agreement measured on double-read studies and reported per dataset.
No label enters a licensed dataset without a clinician decision behind it.
Infrastructure
Built for diligence. Not for demos.
The security posture is not a page in a deck; it determines whether a hospital board and a pharma legal team can both say yes.
AES-256 and TLS 1.3
AES-256 for data at rest and TLS 1.3 for everything in transit, including the gateway’s outbound connection from inside the hospital network.
Role-based, least privilege
Annotators, reviewers, hospital administrators and licensees each see only what their role requires. Access is granted per dataset, not per platform.
Immutable audit logs
Every read, export and permission change is written to an append-only log — which is what makes revenue attribution and compliance reporting reconstructible.
Configurable regional storage
Dataset storage region is configurable to meet licensee residency constraints and partner institution requirements.
See it against your own archive.
The fastest way to evaluate this is a written assessment of a single department’s data. No installation, no commitment.