Dataset qualification & delivery
Most medical imaging data fails the spec. We find the studies that pass.
A written acceptance specification, a verdict for every study with the criteria it failed, cure lists for the supplier, documented de-identification and a hashed manifest per batch. For data we supply, or for data you already hold.
The problem
A terabyte of DICOM is not a dataset
Hospital exports and vendor deliveries arrive as folders. Inside them are lossy re-exports, mixed studies under one report, missing portal-venous series, 12-bit images where 16 were promised and the occasional patient name burned into the pixels. None of that is visible from a file listing, and all of it surfaces after the invoice.
Qualification reads every study before you pay for it, and tells the supplier exactly what to fix.
of studies in a typical hospital export fail at least one criterion of a strict portal-venous CT specification
studies in a raw teleradiology feed has been lossy-compressed at some point in its history
of case folders in one source mixed several studies under a single report
Observed across source batches processed through our pipeline in 2026.
What we deliver
Six things a qualification engagement produces
Engage us for the full chain from specification to delivered batch, or for a single step against data you already hold.
Specification capture
We translate your inclusion and exclusion criteria into a versioned, machine-checkable requirement set: modality, anatomy, phase, slice, kVp, bit depth, compression, report pairing, demographics and coverage. Every change is dated and the eligible count is recomputed.
Indexing and profiling
Every study is read at header and pixel level and profiled on more than thirty attributes. Duplicates are found by content hash and clinical near-match. Reports are paired by accession and identifier; unmatched reports are held, not guessed.
Per-study acceptance verdicts
Each study is marked eligible, review or reject, with the exact criterion codes it failed and whether the failure is recoverable by re-export. Studies whose contrast phase cannot be read from metadata go to a radiologist rather than being inferred.
Cure lists for data suppliers
Suppliers receive a ranked list of what to fix and how many studies each fix would unlock: re-export losslessly, split mixed folders, supply the missing report. It turns a rejection pile into a work order.
De-identification and packaging
Headers scrubbed to a documented profile, UIDs remapped through a sealed crosswalk, burned-in annotation detected and quarantined, re-verification after de-identification, and a hashed manifest for every batch.
Delivery audit and disputes
You load your own pass/fail results and we reconcile them against the delivery ledger. Disputes are opened against a delivered study and settled on the recorded evidence. Withdrawn studies are erased and certified.
Worked example
What a specification looks like
A portal-venous CT abdomen specification we currently qualify against, written the way the pipeline reads it. Every line becomes a criterion code that appears on each study's verdict.
- Modality CT; anatomy includes abdomen; pelvis preferred, recorded but not required
- Portal-venous series present, stated in metadata or confirmed by a radiologist
- No lossy compression at any point in the study history
- Portal-venous slice thickness above 0 and at most 2.5 mm
- kVp populated and above 100
- Bits allocated at least 16
- Exactly one matched clinical report, in the agreed structure
- One study per folder; no duplicate patient or study identifiers in the batch
What the client receives
- Versioned requirement set with change history
- Verdict ledger: every study, its decision and its failure codes
- Batch manifests with SHA-256 per file and per batch
- Dataset datasheet: composition, demographics, scanner mix, known gaps
- De-identification methodology and residual-risk statement
- Supplier cure report with studies unlocked per fix
- Reconciliation report against your own audit
- Erasure records and destruction certificates
Since mid-2026 this process has delivered more than two thousand qualified CT studies to a US diagnostic AI developer in batches of around fifty, each verified against its manifest, with client audits reconciled against the ledger.
How it runs
From specification to verified batch
Specification
We turn your inclusion criteria into a machine-checkable requirement set: modality, anatomy, contrast phase, slice thickness, kVp, bit depth, compression, report pairing and demographic constraints.
Index & profile
Every study is read at header and pixel level and profiled on more than thirty attributes, including phase pattern, scanner model, coverage, compression history and report match.
Verdict per study
Each study receives an eligible, review or reject verdict with the exact criterion codes it failed. Studies whose phase cannot be read from metadata go to a radiologist, not a guess.
De-identify & deliver
Eligible studies are de-identified, re-verified, packaged with a hashed manifest and shipped to your bucket in batches. You load your own pass/fail list and we reconcile against it.
Consent & permitted use
What this data may be used for, stated up front
Every study in the catalogue was obtained under a signed institutional research agreement with the source imaging network, with a data processing agreement on file. The permitted-use scope travels with the data and is written into every delivery manifest.
Permitted purpose
Development, validation and testing of diagnostic AI. Each delivery is matched to a declared purpose before a single study is packaged.
Permitted jurisdictions
United States and India. Deliveries to other jurisdictions require a fresh agreement with the source, which we will tell you plainly if it is not available.
Not permitted
General-purpose or foundation-model pretraining, and onward resale or sublicensing to third parties. We do not sell around the scope of the source agreement.
Retention
Time-limited retention is part of the agreement. Deliveries carry a retention term and we issue destruction certificates when a term ends or a study is withdrawn.
De-identification
DICOM headers are scrubbed to a documented profile, study, series and instance UIDs are remapped through a sealed crosswalk, and studies with burned-in annotation are detected and quarantined rather than shipped.
Provenance
Each batch ships with a hashed manifest, the acceptance verdict for every study and the source-agreement reference, so your counsel can trace the chain of rights.
Need the de-identification methodology and source-agreement summary for a vendor review? Request the compliance pack.
FAQ
Frequently asked questions
Can you qualify data we already hold?
Yes. The same indexing, profiling and verdict process runs against your own exports or a cloud bucket you grant us read access to. Many teams start by learning what fraction of a vendor delivery actually meets their specification.
What does a verdict actually contain?
A decision of eligible, review or reject; the criterion codes failed; whether the failure is recoverable by re-export or resupply; the series that was judged as the target phase, with its slice thickness, kVp, bit depth and compression; anatomy coverage; and the report-match status.
How do you treat studies that were once lossy but are lossless now?
As lossy, unless your specification says otherwise. Decompressing a lossy image does not restore the lost information, so we record compression history and let you decide whether history matters. Most diagnostic specifications say it does.
Who decides the contrast phase of an ambiguous series?
A radiologist. Where the DICOM metadata states the phase we use it. Where it can only be inferred from timing or appearance, the study is held for human review. Inference is never silently promoted to a label.
What happens when we reject a delivered study?
You open a dispute against the delivered study reference. We compare your finding with the recorded verdict and the de-identified pixel data, settle it on the evidence, and if the study is withdrawn we erase it from your delivery, record the erasure and issue a certificate.
How is this priced?
Per accepted study for data we supply, and per study profiled for data you hold, with the specification work scoped as a fixed fee. You pay for studies that pass, not for the terabytes we read to find them.
Bring a specification, or bring a bucket.
If you have criteria, we will tell you how many studies in our catalogue pass. If you have data, we will tell you how much of it meets your own bar, and what the supplier needs to fix.
Or email [email protected] · Mohali, India