MedCorpora standards

A defensible medical dataset standard

A defensible medical AI dataset connects six layers: documented permission, privacy protection, technically correct imaging structure, traceable clinical ground truth, measurable quality and source provenance. Removing names from DICOM is necessary, but it is not the complete product.

01Permission
02Privacy
03Structure
04Clinical truth
01

The buyer must be able to inspect the evidence

Every material claim about case count, sequence, label, demographic field and commercial right should have a source. The dataset card records definitions, exclusions, missingness and known limitations. Supplier agreements and privacy records remain available for diligence under appropriate confidentiality controls.

02

Quality is task-specific

A technically readable image may still be unsuitable for a target model. Acceptance criteria therefore cover the intended task: sequence completeness for MRI, reconstruction and field of view for CT, annotation standards for segmentation, and cohort separation for validation.

03

Derived data remains traceable

NIfTI volumes, previews, masks and machine-readable metadata are linked back to pseudonymous study and series identifiers. The process should be reproducible without preserving direct patient identifiers.

04

Questions, answered directly.

Is de-identification enough for commercial use?

No. The supplier must also have authority to process and license the data for the agreed commercial purpose.

Does every dataset need segmentation masks?

No. The required ground truth depends on the model task. Segmentation masks add value only when clinically appropriate and reliably produced.

What is provenance?

Provenance is the documented history of where data came from, how it was transformed and how each label or derived file was produced.

Institutional engagement

Build the programme.

hello@medcorpora.com