Reference

What everything is, rather than how to do anything. For task-shaped answers, see the how-to guides.

The package

Python API	Reading, writing, amending.
API index	Every public Python name, its summary and where it is documented.
Rust crate	The format engine itself, as a Rust library.
Command line	Every medh5 command.
PyTorch and MONAI	Datasets, samplers, collation, the MetaTensor adapter.

The data

Annotation kinds	Thirteen kinds: six voxel kinds (five encodings and mask), boxes, oriented boxes, keypoints, points, contours, meshes, classification.
Converters	NIfTI, DICOM, DICOM SEG, RTSTRUCT, nnU-Net v2, and 0.x migration.
Curation records	Provenance, quality, agreement, identity, de-identification.
Storage	On-disk layout, codec profiles, chunking, integrity, collections.

Tables

Sample document schema	Every field of /meta.
Clinical schema	The clinical/meta descriptor and the clinical records.
Task manifest schema	Every field of a medh5.task/1 manifest.
Cache manifest schema	Every field of a medh5.cache/1 manifest.
Diagnostic codes	All 100 codes medh5 validate can report.
Cohort check codes	The C1xx–C5xx codes medh5 dataset check reports.
Task and cache codes	The T1xx–T4xx codes medh5 task and medh5 cache report.
Profiles and validation levels	What to check, and how much.
Runnable examples	A complete sample written by following the specification, plus the benchmark scripts.

The specification — with 1.1 for the clinical profile, and the separately versioned task and cache contract — is the normative statement of all of it. Where this section and the specification disagree, one of them is a bug.