API index¶

Every public name of the Python package, one table per module: what it is, the first sentence of its docstring, and the section of these pages that documents it. A name exported by more than one module is listed under each. The full signatures are in the docstrings (help(name)), and medh5._core is the compiled extension the modules re-export from, not an API of its own.

`medh5`¶

Name	Kind	Summary	Documented in
`__format_version__`	constant (`str`)
`__version__`	constant (`str`)		medh5 › Versioning
`Activity`	class	One activity of the provenance graph: what an agent did, with its inputs and outputs (§11.1).	Curation records › Editing a record
`Agent`	class	An agent of the provenance graph: a person, software or organization that acts (§11.1).
`Agreement`	class	One agreement measurement of a quality record, such as a Dice against another rater (§11.2).	Curation records › Agreement
`amend`	function	Copy-on-write amend: build a new file from the old and replace it.	Python API › Amending
`Annotation`	class	What every annotation answers, whatever its kind.	Python API › Timepoints
`audit_splits`	function	Read every file's claims and cross-check them (spec §12.3).	Curation records › Split claims
`CODES`	constant (`dict`)	Every diagnostic code, keyed by code string.	Python API › Validation
`Cohort`	class	The `cohort` member of the sample document: dataset, site, scanner and split group (§12.2).	Curation records › Editing a record
`Collection`	class	A read-only mapping of `sample_key -> Sample` over one shard.	Python API › Opening
`compare_annotations`	function	Compare two annotations, choosing the comparison their kinds support.	Curation records › Agreement
`create`	function	Create a new sample; use as a context manager, or call `commit()`.	Annotation kinds › Closure
`Deidentification`	class	The `deidentification` member: how identifiers were removed from the sample (§11.4).	Curation records › Editing a record
`FeatureCache`	class	An open cache.	Python API › Tasks and caches
`FORMAT_VERSION`	constant (`str`)		medh5 › Versioning
`Grid`	class	The lattice of a stored array: its shape, axes and index-to-world geometry (§3.2).	Annotation kinds › Boxes
`grid_patches`	function	Deterministic sliding-window cover of a volume — the inference path.	PyTorch and MONAI › GridPatchDataset — deterministic tiling, for inference
`Identity`	class	The `identity` member of the sample document: who and what the sample is about (§12.1).	Python API › Identity and structure
`Image`	class	Lazy access to one image and its geometry.	Python API › Timepoints
`Instance`	class	One physical object: a box, a class, an id and optionally a cropped mask.
`Issue`	class	A known defect recorded in a quality record: its code, severity, classes and note (§11.2).	Curation records › Quality
`LabelClass`	class	One semantic class (§5.2).	Annotation kinds › Closure
`LabelSet`	class	A controlled vocabulary (§5.1).	Annotation kinds › Label sets
`MEDH5Error`	exception	Base of the package's own exceptions: a file, version, schema, validation or integrity error.	Python API › Exceptions
`MEDH5FileError`	exception	A file cannot be opened, is not HDF5, or is structurally unreadable.	Python API › Amending
`MEDH5FileNotFoundError`	exception	The file does not exist.	Python API › Exceptions
`MEDH5IntegrityError`	exception	A digest or `content_id` does not match the data it covers.
`MEDH5IOError`	exception	A read or write the engine could not complete: HDF5 failing a damaged chunk, say.	Python API › Exceptions
`MEDH5IsADirectoryError`	exception	The path names a directory, not a file.	Python API › Exceptions
`MEDH5PermissionError`	exception	The file exists but may not be read (or written).	Python API › Exceptions
`MEDH5SchemaError`	exception	The sample document is absent, is not JSON, or violates its JSON Schema.
`MEDH5ValidationError`	exception	Input rejected by a writer, or a validation failure raised as an exception.	Python API › Opening
`MEDH5VersionError`	exception	The file's `medh5_version` is one this operation cannot work with.	Python API › Amending
`Observation`	class	One object seen once: a row of one annotation.	Python API › Tracking
`open`	function	Open a `.medh5` sample file, read-only.	Python API › Opening
`open_collection`	function	Open a `.medh5c` shard, read-only.	Python API › Opening
`open_sample`	function	Open a `.medh5` sample file, read-only.	Python API › Opening
`pack`	function	Copy sample files into one collection shard (spec §2.2).	Python API › Collections
`Patch`	class	One draw: where to read, and how the location was chosen.	Python API › Sampling patches
`PatchSampler`	class	Choose patch windows in a volume (spec §14.3).	Python API › Sampling patches
`PROFILES`	constant (`tuple`)		Profiles and validation levels › Profiles — what to hold the file to
`Provenance`	class	Who did what (§11.1): agents and the activities they performed.	Curation records › Provenance
`QualityRecord`	class	What is known about an annotation's trustworthiness: status, confidence, agreement, issues (§11.2).
`Sample`	class	A read-only view of one sample root (a file, or a collection member).	Python API › Sample
`SampleDocument`	class	The whole document, typed.	Python API › Identity and structure
`SampleWriter`	class	Builder for one sample.	Python API › Writing
`SplitAudit`	class	What a cohort's split claims say, and where they disagree.
`SplitClaim`	class	One entry of `splits`: a claim that the sample belongs to a partition of a split set (§12.3).	Build and split a cohort › What the digest covers
`TaskManifest`	class	A `medh5.task/1` manifest: build it, check it, save it, preflight it.	Python API › Tasks and caches
`Timeline`	class	The sample's timepoints, in acquisition order; indexable by position or id.	Python API › Timepoints
`Timepoint`	class	One timepoint of the sample: an observation occasion, in DICOM terms usually one study (§3.7).	Curation records › Editing a record
`TimepointPair`	class	One ordered pair of visits, and the change label spanning it, if any.
`TimepointPairSampler`	class	Enumerate the visit pairs a longitudinal model trains on (§3.7, §9).	PyTorch and MONAI › PairedPatchDataset — the same place at two visits
`Track`	class	Every observation of one physical object, ordered by timepoint.	Curation records › Tracking
`Tracking`	class	`{instance_id: Track}` with the coverage needed to tell resolved from unexamined, and the reason for every absence.	Curation records › Tracking
`unpack`	function	Extract sample roots back into standalone files (spec §2.2).	Python API › Collections
`validate_cache`	function	Validate a cache: its checksums, every source pin, and — given a task — that it was built for this task, fitted on this task's training partition, and (`check_rows`) that each row's entry was built at its cutoff, from versions it admits, of its own subject's sources.	Python API › Tasks and caches
`VoxelAnnotation`	class	The uniform read contract of the five voxel encodings (spec §7).

`medh5.annotations`¶

Name	Kind	Summary	Documented in
`Annotation`	class	What every annotation answers, whatever its kind.	Python API › Timepoints
`ANNOTATION_KINDS`	constant (`tuple`)
`AnnotationHeader`	class	The fixed attribute header every annotation carries (spec §6.2).
`AnnotationPayload`	class	Datasets and kind-specific attributes for one encoded annotation.
`Assertion`	class	One label assertion: a class, its value, and what it is about.	Classification and change labels › Hierarchical labels
`BoxesAnnotation`	class	Axis-aligned boxes, `(N, S, 2)` `[lo, hi]` at voxel edges (§8.2).
`ClassificationAnnotation`	class	Reader for `kind = "classification"`.
`ContoursAnnotation`	class	Planar polygons with an offset table (§8.6) — RTSTRUCT-shaped.
`encode_boxes`	function	Pack axis-aligned boxes, `(N, S, 2)` in `[lo, hi]` at voxel edges, as a `boxes` payload (§8.2).
`encode_classification`	function	Pack label assertions as a `classification` payload over one `scope` (§9).
`encode_contours`	function	Concatenate planar polygons into a `contours` payload with an offset table (§8.6).
`encode_keypoints`	function	Pack `(N, K, S)` keypoints of `N` objects as a `keypoints` payload (§8.4).
`encode_mesh`	function	Pack a triangle surface mesh, `(V, 3)` vertices and `(F, 3)` faces, as a `mesh` payload (§8.7).
`encode_obb`	function	Pack oriented boxes as centres, full edge lengths and rotation matrices, an `obb` payload (§8.3).
`encode_points`	function	Pack an `(N, S)` point set — landmarks or seeds — as a `points` payload (§8.5).
`encode_voxels`	function	Encode class masks, choosing the encoding by measurement when asked to.
`GEOMETRIC_KINDS`	constant (`tuple`)
`GeometricAnnotation`	class	Base for the §8 kinds: N objects whose coordinates live in one space.
`Instance`	class	One physical object: a box, a class, an id and optionally a cropped mask.
`KeypointsAnnotation`	class	`(N, K, S)` keypoints with per-slot classes and visibility (§8.4).
`MeshAnnotation`	class	A triangle surface mesh, optionally several submeshes (§8.7).
`ObbAnnotation`	class	Oriented boxes: centre, full edge lengths, rotation (§8.3).
`open_annotation`	function	The reader class for an annotation handle's `kind`.
`PointsAnnotation`	class	A point set: landmarks, seeds, or half a correspondence (§8.5).
`Polygon`	class	One planar polygon handed to `encode_contours`.	Annotation kinds › Contours
`READERS`	constant (`dict`)	`kind` -> reader class, for every kind.
`RESERVED_KINDS`	constant (`tuple`)	Kinds a 1.0 writer must not write (spec §16).
`SCOPES`	constant (`tuple`)
`select_encoding`	function	Choose an encoding by measurement: `(kind, stats)`.
`SPACES`	constant (`tuple`)
`TASKS`	constant (`tuple`)
`transcode`	function	Convert an open voxel annotation to another encoding (§7.6).	Annotation kinds › Transcoding
`VOXEL_KINDS`	constant (`tuple`)
`VoxelAnnotation`	class	The uniform read contract of the five voxel encodings (spec §7).

`medh5.annotations.base`¶

Name	Kind	Summary	Documented in
`Annotation`	class	What every annotation answers, whatever its kind.	Python API › Timepoints
`ANNOTATION_KINDS`	constant (`tuple`)
`AnnotationHeader`	class	The fixed attribute header every annotation carries (spec §6.2).
`AnnotationPayload`	class	Datasets and kind-specific attributes for one encoded annotation.
`DEFAULT_TASK_FOR_KIND`	constant (`dict`)
`GEOMETRIC_KINDS`	constant (`tuple`)
`Instance`	class	One physical object: a box, a class, an id and optionally a cropped mask.
`instance_id_dtype`	function	`uint32` unless an id needs the wider form (spec §7.4, §8.2).
`RESERVED_KINDS`	constant (`tuple`)	Kinds a 1.0 writer must not write (spec §16).
`SPEC_ANNOTATION_ATTRS`	constant (`tuple`)
`TASKS`	constant (`tuple`)
`VOXEL_KINDS`	constant (`tuple`)
`VoxelAnnotation`	class	The uniform read contract of the five voxel encodings (spec §7).

`medh5.annotations.classification`¶

Name	Kind	Summary	Documented in
`Assertion`	class	One label assertion: a class, its value, and what it is about.	Classification and change labels › Hierarchical labels
`assertion_rows`	function	`(classes, values, scope_ids, schemes, scheme_values)` from a mapping or assertion rows (§9).
`check_scope`	function	Return `scope` if it is a classification scope, else raise E412 (§9).
`ClassificationAnnotation`	class	Reader for `kind = "classification"`.
`encode_classification`	function	Pack label assertions as a `classification` payload over one `scope` (§9).
`Labels`	type	What a classification is built from: `class -> value`, or assertion rows.
`SCOPES`	constant (`tuple`)

`medh5.annotations.geometric`¶

Name	Kind	Summary	Documented in
`BoxesAnnotation`	class	Axis-aligned boxes, `(N, S, 2)` `[lo, hi]` at voxel edges (§8.2).
`check_slice_index`	function	Return what is wrong with a `slice_index` column of one plane per box, or `None` (§8.2).
`check_space`	function	Return `space` if it is `index` or `world`, else raise E412 (§8.1).
`CONTOUR_ROLES`	constant (`tuple`)
`ContoursAnnotation`	class	Planar polygons with an offset table (§8.6) — RTSTRUCT-shaped.
`encode_boxes`	function	Pack axis-aligned boxes, `(N, S, 2)` in `[lo, hi]` at voxel edges, as a `boxes` payload (§8.2).
`encode_contours`	function	Concatenate planar polygons into a `contours` payload with an offset table (§8.6).
`encode_keypoints`	function	Pack `(N, K, S)` keypoints of `N` objects as a `keypoints` payload (§8.4).
`encode_mesh`	function	Pack a triangle surface mesh, `(V, 3)` vertices and `(F, 3)` faces, as a `mesh` payload (§8.7).
`encode_obb`	function	Pack oriented boxes as centres, full edge lengths and rotation matrices, an `obb` payload (§8.3).
`encode_points`	function	Pack an `(N, S)` point set — landmarks or seeds — as a `points` payload (§8.5).
`GEOMETRIC_READERS`	constant (`dict`)
`GeometricAnnotation`	class	Base for the §8 kinds: N objects whose coordinates live in one space.
`KeypointsAnnotation`	class	`(N, K, S)` keypoints with per-slot classes and visibility (§8.4).
`MeshAnnotation`	class	A triangle surface mesh, optionally several submeshes (§8.7).
`ObbAnnotation`	class	Oriented boxes: centre, full edge lengths, rotation (§8.3).
`PointsAnnotation`	class	A point set: landmarks, seeds, or half a correspondence (§8.5).
`Polygon`	class	One planar polygon handed to `encode_contours`.	Annotation kinds › Contours
`ROTATION_TOL`	constant (`float`)
`SPACES`	constant (`tuple`)
`VISIBILITY`	constant (`dict`)

`medh5.annotations.voxel`¶

Name	Kind	Summary	Documented in
`analyse`	function	Measure a set of class masks: counts, the overlap graph, a colouring and localization (§7.6).
`annotation_to_masks`	function	Every class of a voxel annotation as a full mask.
`AnnotationPayload`	class	Datasets and kind-specific attributes for one encoded annotation.
`BitmaskAnnotation`	class	Classes as bits of `uint64` planes; position from `bit_class_ids` (§7.3).
`BITS_PER_PLANE`	constant (`int`)
`check_roundtrip`	function	Whether transcoding a payload to `to_kind` keeps every class mask and its ignore region (§7.6).	Annotation kinds › Transcoding
`checked_class_id`	function	A class id in the writable range `1..65534`, or a coded refusal: E303 for a reserved id (0 or 65535), E306 for any other outside the range.
`contains_at`	function	`values >= threshold`, decided in the stored precision (§7.5).
`cost_model`	function	Return the raw, uncompressed bytes each voxel encoding would take for the analysed masks.
`CostModel`	class	Raw (pre-compression) bytes per encoding.
`DEFAULT_THRESHOLD`	constant (`float`)
`encode_bitmask`	function	Pack class masks into `uint64` bitplanes, one bit per class, as a `bitmask` payload (§7.3).
`encode_instances`	function	Pack objects into boxes, ids and bit-packed crops (spec §7.4).
`encode_labelmap`	function	Pack mutually exclusive class masks into one integer volume (§7.1).
`encode_layers`	function	Colour the overlap graph and pack each colour into one labelmap (§7.2); `ignore` and `ignore_id` are checked and painted as `encode_labelmap` does.
`encode_mask`	function	Wrap one boolean volume as a `mask` payload.	Annotation kinds › Transcoding
`encode_masks`	function	Encode class masks into the voxel encoding `kind` and return the payload.
`encode_probmap`	function	Pack per-class probability planes (spec §7.5).
`encode_voxels`	function	Encode class masks, choosing the encoding by measurement when asked to.
`greedy_colour`	function	Colour an overlap graph greedily, highest degree first: a dict of class id to layer (§7.6).
`IN_BAND_IGNORE_KINDS`	constant (`tuple`)
`Instance`	class	One physical object: a box, a class, an id and optionally a cropped mask.
`InstanceInput`	class	One object handed to `encode_instances`.	Longitudinal studies › Join the objects
`instances_from_masks`	function	One object per class mask, minting ids from start_id; an id past the largest `uint64` is refused (E404).
`InstancesAnnotation`	class	Objects: each a class, a sample-scoped id, a box and maybe a mask (§7.4).
`label_dtype_size`	function	Return the bytes per voxel of a labelmap for these class ids: 1 (`uint8`) or 2 (`uint16`).
`LabelmapAnnotation`	class	Mutually exclusive classes in one volume; `65535` marks ignore (§7.1).
`layers_from_colouring`	function	Group a colouring of class id to layer into one tuple of class ids per layer, ascending.
`LayersAnnotation`	class	Classes coloured onto layers so no two that overlap share one (§7.2).
`LOCALIZED_BBOX_FRACTION`	constant (`float`)
`MaskAnnotation`	class	A region, not a class: what a field of view or an ignore mask is (§7.7).
`Masks`	type	class id -> boolean occupancy over the grid's spatial shape.
`masks_equal`	function	Whether two mask dicts hold the same classes with identical masks: the losslessness check.
`normalize_masks`	function	Coerce a mask mapping to `bool` arrays of one agreed shape.
`OverlapStats`	class	Measured properties of a set of class masks, and what they imply (§7.6).	Segmentation › Write the masks
`payload_to_masks`	function	Decode any voxel payload into a dict of class id to boolean mask.
`ProbmapAnnotation`	class	Soft labels; a voxel contains a class at or above `threshold` (§7.5).
`READERS`	constant (`dict`)	`kind` -> reader class, for the voxel kinds.
`select_encoding`	function	Choose an encoding by measurement: `(kind, stats)`.
`SLAB_BYTES`	constant (`int`)	Read budget for a scan that only needs a yes/no or a count.
`SPARSE_FILL`	constant (`float`)
`storage_dtype`	function	The narrowest of `float16`/`float32` that keeps every voxel's containment decision (§7.5); E411 when neither does — a value within `float32` rounding of the threshold, on its other side.
`STORAGE_DTYPES`	constant (`tuple`)
`TRANSCODABLE`	constant (`tuple`)		Python API › Annotations
`transcode`	function	Convert an open voxel annotation to another encoding (§7.6).	Annotation kinds › Transcoding
`transcode_payload`	function	Re-encode a payload; the same payload comes back when it already has `to_kind`.	Annotation kinds › Transcoding

`medh5.bench`¶

Name	Kind	Summary	Documented in
`benchmark_file`	function	Every targeted metric on one file but the many-class draw, an image patch read, and the paired read when the file is longitudinal.
`many_class_measurement`	function	The foreground draw on a many-class sample: O(1) whatever the classes.
`MANY_CLASSES`	constant (`int`)
`Measurement`	class	One timed metric, and whether it met its target.
`report`	function	The text table the `medh5 bench` command prints.	Python API › Validation
`synthetic_many_class_sample`	function	A sample with classes small classes, for the many-class draw.
`synthetic_pair`	function	A two-timepoint sample with a registration, for the paired metrics.
`synthetic_sample`	function	Write a sample shaped like the one the published numbers were measured on: a CT, `classes` overlapping organs, a sampling index.	Profiles and validation levels › Levels — how much to check
`TARGETS`	constant (`dict`)	Metric -> (upper bound in ms, description): the performance guide's targets.
`throughput`	function	Sustained patches/s through the real dataloader.
`timed`	function	Median milliseconds per call.

`medh5.cache`¶

Name	Kind	Summary	Documented in
`build_document_cache`	function	An event-level cache of every document event of the task's sources.	Python API › Tasks and caches
`CacheEntry`	class	One cached feature and everything it depends on.
`CacheReport`	class	What validating a cache found.	Python API › Tasks and caches
`CacheWriter`	class	Write a cache; it appears at `path` only on `commit`.	Python API › Tasks and caches
`event_entry_id`	function	The entry id of an event-level feature of `(content_id, event_id)`.
`FeatureCache`	class	An open cache.	Python API › Tasks and caches
`fitted_on`	function	The `fitted_on` record for learned preprocessing over `partition` (default: the task's training partition — the split's first).	Python API › Tasks and caches
`fitted_on_mismatches`	function	Why preprocessing recorded as fitted was not fitted on task's training partition (T405): another task, split (`set_id`), membership or partition.	Python API › Tasks and caches
`HashingTextEncoder`	class	Signed feature hashing of lower-cased words into `dim` floats.	Python API › Tasks and caches
`LEVELS`	constant (`tuple`)
`open_cache`	function	Open a cache file, read-only.
`SCHEMA`	constant (`str`)
`schema_text`	function	The `medh5.cache/1` manifest JSON Schema.	Cache manifest schema
`SUFFIX`	constant (`str`)
`validate_cache`	function	Validate a cache: its checksums, every source pin, and — given a task — that it was built for this task, fitted on this task's training partition, and (`check_rows`) that each row's entry was built at its cutoff, from versions it admits, of its own subject's sources.	Python API › Tasks and caches

`medh5.cli`¶

Name	Kind	Summary
`command_tree`	function	The grammar as data: `{"options", "positionals", "commands"}`, recursively — what documentation is checked against.
`EXIT_ERROR`	constant (`int`)
`EXIT_OK`	constant (`int`)
`EXIT_USAGE`	constant (`int`)
`main`	function	Run `medh5` on argv (`sys.argv[1:]` by default); the exit code.

`medh5.clinical`¶

Name	Kind	Summary	Documented in
`ASSESSMENT_SYSTEM`	constant (`str`)
`augment`	function	Add clinical records to a sample, in place or into `out` (1.1 §10).	Python API › Clinical history (format 1.1)
`baseline_day_clock`	function	The clock `imaging_events_from_timepoints` measures on.	Clinical history beside the images › Add a history to a 1.0 file
`Bounds`	type	Inclusive `(lo, hi)` bounds on one instant, in clock microseconds.
`Clinical`	class	One sample's clinical profile, read: `Sample.clinical`.	Python API › Clinical history (format 1.1)
`ClinicalRecords`	class	A logical-record bundle: what `augment` adds and `records()` reads.	Python API › Clinical history (format 1.1)
`Clock`	class	The subject clock every clinical time is measured on (1.1 §3).	Python API › Clinical history (format 1.1)
`clock_day`	function	The clock day holding `t_us`: its midnight to the last microsecond before the next, counted from the clock's zero (the engine's `Bounds::day_containing`).	Python API › Clinical history (format 1.1)
`CLOCK_REFERENCES`	constant (`tuple`)
`COMPARATORS`	constant (`tuple`)		PyTorch and MONAI › Clinical tasks (format 1.1)
`DAY`	constant (`int`)		Train on clinical tasks › Describe the task
`day_bounds`	function	Every instant of a calendar `day` on a `utc` clock, inclusive.	Python API › Clinical history (format 1.1)
`days`	function	`value` days, in clock microseconds, rounded half to even, as the engine rounds a visit's `days_from_baseline`.	Clinical history beside the images › Write a history
`Document`	class	Canonical, de-identified source text (1.1 §6).	Python API › Document
`DocumentInfo`	class	A document's metadata; its text is read when asked for.	Python API › Clinical history (format 1.1)
`ENDPOINT_TYPES`	constant (`tuple`)
`Event`	class	One immutable version of information about the subject (1.1 §5).	Python API › Clinical history (format 1.1)
`EVENT_KINDS`	constant (`tuple`)
`HOUR`	constant (`int`)		Python API › Clinical history (format 1.1)
`hours`	function	`value` hours, in clock microseconds, rounded half to even — as the engine rounds `medh5 clinical select --cutoff-hours` and every time given in hours, so a cutoff is one instant from every frontend.
`imaging_events_from_timepoints`	function	`imaging` events from `days_from_baseline`, known to the day.	Python API › Clinical history (format 1.1)
`LESION_PRESENCE`	constant (`str`)
`LESION_VALUES`	constant (`tuple`)
`Link`	class	A typed relationship between sample-relative objects (1.1 §7).	Python API › Clinical history (format 1.1)
`MICROSECOND`	constant (`int`)
`MINUTE`	constant (`int`)		From clinical history to a training batch › Write two subjects
`PROFILE`	constant (`str`)
`records_from`	function	A bundle from lists.
`RELATIONS`	constant (`tuple`)
`SCHEMA`	constant (`str`)
`schema_text`	function	The clinical descriptor and logical-record JSON Schema.	Clinical schema
`SECOND`	constant (`int`)
`select`	function	Select at `cutoff_us` from records not read from a file.	Python API › Clinical history (format 1.1)
`SelectedEvent`	class	One event version a selection admits, in input order.
`Selection`	class	What a cutoff admits (1.1 §9).	Python API › Clinical history (format 1.1)
`SELECTION_POLICIES`	constant (`tuple`)
`SelectionPolicy`	class	How inputs are chosen at a cutoff (task-and-cache contract §3.4).	Python API › Clinical history (format 1.1)
`STATUSES`	constant (`tuple`)
`strip`	function	Write the imaging projection of `path` to `out`: a different sample, and a reported loss.	Python API › Clinical history (format 1.1)
`TEMPORAL_TYPES`	constant (`tuple`)
`utc_us`	function	An aware `datetime` on a `utc` clock: microseconds since 1970-01-01T00:00:00Z.	Python API › Clinical history (format 1.1)

`medh5.collection`¶

Name	Kind	Summary	Documented in
`Collection`	class	A read-only mapping of `sample_key -> Sample` over one shard.	Python API › Opening
`default_key`	function	The sample key a file gets when none is supplied: its name up to its first dot (medh5 3.0 takes the stem, the name less its last suffix).
`extract`	function	Extract one member of a collection into a standalone sample file.	Pack small samples into collections › Pack and unpack
`is_collection`	function	Whether a root declares itself a collection.
`open_any`	function	Open a file whatever its kind, resolving key inside a collection.	Python API › Opening
`open_collection`	function	Open a `.medh5c` shard, read-only.	Python API › Opening
`pack`	function	Copy sample files into one collection shard (spec §2.2).	Python API › Collections
`SAMPLES_GROUP`	constant (`str`)	Where sample roots live inside a collection (spec §2.2).
`SUFFIX`	constant (`str`)	Conventional extension for a collection file (spec §2.1).
`unpack`	function	Extract sample roots back into standalone files (spec §2.2).	Python API › Collections

`medh5.conformance`¶

Name	Kind	Summary	Documented in
`build_corpus`	function	Write every case (or those names) and an `expected.json` beside them; returns the manifest's path.
`Case`	class	One corpus entry.	Migrate from 0.x › 3. Read the report
`case_by_name`	function	The conformance case called name; a `KeyError` names an unknown one.
`CaseResult`	class	What running one case produced.
`CASES`	constant (`tuple`)
`check_checksums`	function	Files whose checksum does not match `SHA256SUMS` (empty when intact).	Run the conformance suite › Score it
`CHECKSUMS`	constant (`str`)
`load_manifest`	function	The suite's `expected.json`.	Run the conformance suite › Publish a suite
`publish`	function	Write the whole suite into outdir — the cases, `expected.json`, the code table, the JSON Schemas (the sample document, the clinical profile, the task and cache contract), the task-and-cache fixtures under `companion/` when the clinical cases they read are among the cases, a checksum file and instructions, everything a third-party implementation needs; returns outdir.	Run the conformance suite › Publish a suite
`run_corpus`	function	Build the corpus into outdir and check this validator against it.
`SCHEMA`	constant (`str`)
`score`	function	Score a foreign validator's results against the published expectations.	Annotation kinds › Classification
`score_companion`	function	Score another implementation of the task-and-cache contract against the fixtures under `companion/` of a published suite.	Run the conformance suite › Score a task-and-cache implementation
`SEED`	constant (`int`)
`summarize`	function	The `--json` summary of a run or a score.	Run the conformance suite › Score it

`medh5.curation`¶

Name	Kind	Summary	Documented in
`Activity`	class	One activity of the provenance graph: what an agent did, with its inputs and outputs (§11.1).	Curation records › Editing a record
`ACTIVITY_FIELDS`	constant (`frozenset`)
`ACTIVITY_TYPES`	constant (`tuple`)
`Agent`	class	An agent of the provenance graph: a person, software or organization that acts (§11.1).
`AGENT_FIELDS`	constant (`frozenset`)
`AGENT_TYPES`	constant (`tuple`)
`Agreement`	class	One agreement measurement of a quality record, such as a Dice against another rater (§11.2).	Curation records › Agreement
`audit_splits`	function	Read every file's claims and cross-check them (spec §12.3).	Curation records › Split claims
`build_tracks`	function	Join `instance_id` across a sample's timepoints (`Sample.tracks`).
`carries_instance_ids`	function	Whether an annotation carries object identity (`instance_ids`) to join on.
`check_timestamp`	function	Return `value` if it is an RFC 3339 timestamp, else raise E604 (§11.1).
`Cohort`	class	The `cohort` member of the sample document: dataset, site, scanner and split group (§12.2).	Curation records › Editing a record
`compare`	function	Compare two annotations, choosing the comparison their kinds support.	Curation records › Agreement
`compare_instances`	function	Object-level agreement between two instance-carrying annotations.	Curation records › Agreement
`compare_voxel`	function	Per-class Dice or IoU between two voxel annotations on the same grid.	Curation records › Two vocabularies
`DATE_PATTERN`	constant (`Pattern`)	An ISO-8601 calendar date prefix, as `date` must start (the engine's).
`Deidentification`	class	The `deidentification` member: how identifiers were removed from the sample (§11.4).	Curation records › Editing a record
`dice_agreement`	function	Return a mean-Dice `Agreement` from a mapping of class id to that class's Dice (§11.2).
`ID_SOURCE`	constant (`str`)
`Identity`	class	The `identity` member of the sample document: who and what the sample is about (§12.1).	Python API › Identity and structure
`InstanceAgreement`	class	Object-level agreement: what matched, what did not, and how.
`Issue`	class	A known defect recorded in a quality record: its code, severity, classes and note (§11.2).	Curation records › Quality
`ISSUE_SEVERITY`	constant (`tuple`)
`LATERALITY_VALUES`	constant (`tuple`)
`Observation`	class	One object seen once: a row of one annotation.	Python API › Tracking
`PARTITIONS`	constant (`tuple`)
`PRESENT`	constant (`str`)
`Provenance`	class	Who did what (§11.1): agents and the activities they performed.	Curation records › Provenance
`PSEUDONYM_SOURCE`	constant (`str`)
`QUALITY_FIELDS`	constant (`frozenset`)
`quality_from_json`	function	Parse the `quality` member of a sample document into a dict of key to `QualityRecord` (§11.2).
`QUALITY_STATUS`	constant (`tuple`)
`quality_to_json`	function	Serialize a mapping of key to `QualityRecord` (or its JSON form) as the `quality` member (§11.2).
`QualityRecord`	class	What is known about an annotation's trustworthiness: status, confidence, agreement, issues (§11.2).
`REASONS`	constant (`tuple`)	Why an unseen object is or is not a finding (§7.4), in precedence order.	Curation records › Tracking
`RESOLVED`	constant (`str`)
`RFC3339`	constant (`Pattern`)	The timestamp form §11.1 requires; `check_timestamp` is the rule.
`SEX_VALUES`	constant (`tuple`)
`SplitAudit`	class	What a cohort's split claims say, and where they disagree.
`SplitClaim`	class	One entry of `splits`: a claim that the sample belongs to a partition of a split set (§12.3).	Build and split a cohort › What the digest covers
`splits_from_json`	function	Parse the `splits` member of a sample document into a tuple of `SplitClaim`s (§12.3).
`STATES`	constant (`tuple`)		Curation records › Tracking
`Timeline`	class	The sample's timepoints, in acquisition order; indexable by position or id.	Python API › Timepoints
`Timepoint`	class	One timepoint of the sample: an observation occasion, in DICOM terms usually one study (§3.7).	Curation records › Editing a record
`TIMEPOINT_FIELDS`	constant (`frozenset`)
`Track`	class	Every observation of one physical object, ordered by timepoint.	Curation records › Tracking
`Tracking`	class	`{instance_id: Track}` with the coverage needed to tell resolved from unexamined, and the reason for every absence.	Curation records › Tracking
`UNEXAMINED`	constant (`str`)
`VoxelAgreement`	class	Per-class agreement between two voxel annotations.

`medh5.curation.agreement`¶

Name	Kind	Summary	Documented in
`box_iou`	function	IoU of two `(S, 2)` boxes in the same space.
`ClassMap`	type	Which class of `b` each class of `a` is compared with: by id (one vocabulary), by key (two inline label sets), or an explicit one-to-one map.	Curation records › Two vocabularies
`compare`	function	Compare two annotations, choosing the comparison their kinds support.	Curation records › Agreement
`compare_instances`	function	Object-level agreement between two instance-carrying annotations.	Curation records › Agreement
`compare_voxel`	function	Per-class Dice or IoU between two voxel annotations on the same grid.	Curation records › Two vocabularies
`DEFAULT_IOU`	constant (`float`)
`dice`	function	Sørensen--Dice, or `None` when both masks are empty.	Curation records › Agreement
`InstanceAgreement`	class	Object-level agreement: what matched, what did not, and how.
`iou`	function	Intersection over union, or `None` when both masks are empty.	Curation records › Agreement
`MatchBy`	type	How objects are paired: by `instance_id` within one sample where both sides share ids (`auto`), always by IoU, or by id (one sample only).	Curation records › Agreement
`OBJECT_KINDS`	constant (`tuple`)	Kinds whose objects carry an axis-aligned box that IoU can be taken over.
`VoxelAgreement`	class	Per-class agreement between two voxel annotations.

`medh5.curation.scrub`¶

Name	Kind	Summary	Documented in
`apply`	function	Act on the actionable findings and write the §11.4 attestation.	Converters › DICOM
`DATE_KEYS`	constant (`frozenset`)
`Finding`	class	One place an identifier may live.	Curation records › De-identification
`IDENTIFYING_KEYS`	constant (`frozenset`)	Attributes to remove, from the DICOM PS3.15 E.1 basic profile.
`IDENTITY_RULES`	constant (`tuple`)	Findings about what the sample is called.
`MAX_DEPTH`	constant (`int`)	How deep the walk goes before it reports rather than descends.
`PROFILES`	constant (`tuple`)		Curation records › De-identification
`PSEUDONYM_PREFIX`	constant (`str`)	What `pseudonymise` produces, and therefore what the rules skip.
`pseudonymise`	function	A stable pseudonym for a UID: same input, same output, everywhere.	De-identify and publish › 2. Apply, with a salt
`QUASI_IDENTIFYING_KEYS`	constant (`frozenset`)	Attributes that identify in combination — and that some pipelines need.
`scan`	function	Find identifiers in one file.	Curation records › De-identification
`scan_document`	function	Every rule, over every string in one sample document.
`scrub_paths`	function	`scan` each of paths, or, with apply_changes, `apply` the other options to each.	De-identify and publish › 2. Apply, with a salt
`ScrubReport`	class	What was found, what was changed, and what was never looked at.
`STRICT_RULES`	constant (`tuple`)	Rules that only `--profile strict` acts on.
`UID_KEYS`	constant (`frozenset`)	Keys that hold a UID.
`UNFIXABLE_LOCATIONS`	constant (`tuple`)	Findings `--apply` never acts on by itself, however hard the profile looks.

`medh5.curation.splits`¶

Name	Kind	Summary	Documented in
`anatomy_units`	function	Grouping key -> every grouping key sharing anatomy with it, sorted.
`audit_splits`	function	Read every file's claims and cross-check them (spec §12.3).	Curation records › Split claims
`Conflict`	class	One `set_id` claimed against more than one manifest (W906).
`Leak`	class	One unit of anatomy appearing in more than one partition of one split.
`Membership`	class	One file's claim to one partition of one split.
`SplitAudit`	class	What a cohort's split claims say, and where they disagree.

`medh5.dataset`¶

Name	Kind	Summary	Documented in
`Assignment`	class	Where one group went.
`check`	function	Every cohort-level check, over a manifest.	Cohort check codes
`CHECK_CODE_SEVERITY`	constant (`dict`)	Each cohort code's severity.	Cohort check codes
`CHECK_CODES`	constant (`dict`)	Cohort codes.	Cohort check codes
`ClassStats`	class	How often a class occurs, over the samples that examined it.
`CohortReport`	class	What `medh5.dataset.check` found in a cohort: the manifest it read, the number of samples, every finding and each class's coverage.
`compute_stats`	function	Stream a whole cohort, optionally across workers threads.	Pack small samples into collections › Members in a cohort
`DatasetStats`	class	Everything one streaming pass over a cohort learned.
`entries_for`	function	The manifest entries in one file: one per sample, so a collection fans out into one entry per member (each with its `key`).
`Entry`	class	One sample's metadata, as far as a cohort needs it.	Partial labels and coverage › What this buys you at training time
`find`	function	Every sample and collection under root, in a stable order.
`Finding`	class	One cohort check's finding: its `C`-code, severity, message and the samples or paths it concerns.	Cohort check codes
`make_splits`	function	Assign every group in manifest to a partition, or to a fold.	Train on clinical tasks › Describe the task
`Manifest`	class	A cohort as metadata, with the digest that makes a split checkable.	The cohort manifest › Paths
`Moments`	class	Count, mean and M2 for one image key — a Welford accumulator over the finite values folded in.
`scan`	function	Build a manifest from a directory tree, and report what would not open.	The cohort manifest › Paths
`Split`	class	A complete assignment of a cohort's groups.	Curation records › Split claims
`write_claims`	function	Stamp each sample with its partition and the manifest digest (§12.3).	Build and split a cohort › Split it

`medh5.dataset.check`¶

Name	Kind	Summary	Documented in
`check`	function	Every cohort-level check, over a manifest.	Cohort check codes
`CHECK_CODE_SEVERITY`	constant (`dict`)	Each cohort code's severity.	Cohort check codes
`CHECK_CODES`	constant (`dict`)	Cohort codes.	Cohort check codes
`CohortReport`	class	What `medh5.dataset.check` found in a cohort: the manifest it read, the number of samples, every finding and each class's coverage.
`Finding`	class	One cohort check's finding: its `C`-code, severity, message and the samples or paths it concerns.	Cohort check codes
`SEVERITIES`	constant (`tuple`)

`medh5.dataset.manifest`¶

Name	Kind	Summary	Documented in
`counts`	function	How many entries carry each value of by — the stratification tally.	Python API › Tasks and caches
`entries_for`	function	The manifest entries in one file: one per sample, so a collection fans out into one entry per member (each with its `key`).
`Entry`	class	One sample's metadata, as far as a cohort needs it.	Partial labels and coverage › What this buys you at training time
`find`	function	Every sample and collection under root, in a stable order.
`GROUPABLE`	constant (`tuple`)	Fields worth grouping or stratifying on — all single-valued and scannable.
`load`	function	Read a manifest written by `Manifest.save` or `medh5 dataset index`; see `Manifest.load`.	The cohort manifest › Paths
`Manifest`	class	A cohort as metadata, with the digest that makes a split checkable.	The cohort manifest › Paths
`PathStyle`	type	How `Manifest.save` writes paths.	The cohort manifest › Paths
`scan`	function	Build a manifest from a directory tree, and report what would not open.	The cohort manifest › Paths
`SCHEMA`	constant (`str`)	What every manifest `Manifest.save` writes names under `schema`.	The cohort manifest › Paths
`SUFFIXES`	constant (`tuple`)

`medh5.dataset.split`¶

Name	Kind	Summary	Documented in
`Algorithm`	type	How `make_splits` deals groups to partitions: `"medh5.split/2"` moves few when the cohort grows, and is medh5 3.0's default.	Build and split a cohort › The algorithm, and continuing a split
`Assignment`	class	Where one group went.
`DEFAULT_RATIOS`	constant (`dict`)
`load`	function	Read a split written by `medh5 dataset split -o`.	The cohort manifest › Paths
`make_splits`	function	Assign every group in manifest to a partition, or to a fold.	Train on clinical tasks › Describe the task
`Split`	class	A complete assignment of a cohort's groups.	Curation records › Split claims
`write_claims`	function	Stamp each sample with its partition and the manifest digest (§12.3).	Build and split a cohort › Split it

`medh5.dataset.stats`¶

Name	Kind	Summary	Documented in
`ClassStats`	class	How often a class occurs, over the samples that examined it.
`compute_stats`	function	Stream a whole cohort, optionally across workers threads.	Pack small samples into collections › Members in a cohort
`DatasetStats`	class	Everything one streaming pass over a cohort learned.
`Moments`	class	Count, mean and M2 for one image key — a Welford accumulator over the finite values folded in.
`stats_for`	function	One file's contribution, read plane by plane rather than whole.

`medh5.document`¶

Name	Kind	Summary	Documented in
`META_DATASET`	constant (`str`)
`new_document`	function	Start a `SampleDocument` with its two required members, `identity` and `timepoints` (§2.4).
`read_document`	function	The sample document under a sample root, parsed: a later minor's as its supported projection (1.1 §2.2).
`read_document_text`	function	The raw `/meta` text under a sample root (`sample.root`).
`SampleDocument`	class	The whole document, typed.	Python API › Identity and structure
`schema`	function	The sample-document JSON Schema, parsed: a fresh dict on every call, so a caller's edit stays the caller's.	Cache manifest schema
`schema_text`	function	The sample-document JSON Schema (draft 2020-12), as published: the engine's embedded copy of `medh5-sample-1.0.schema.json`.	Sample document schema
`validate_against_schema`	function	Validate a document, returning human-readable messages (empty when valid).

`medh5.errors`¶

Name	Kind	Summary	Documented in
`Code`	class	One diagnostic code.
`code`	function	Look up a diagnostic code, raising `KeyError` for unknown codes.	Diagnostic codes › Reading a code
`CODES`	constant (`dict`)	Every diagnostic code, keyed by code string.	Python API › Validation
`codes_for`	function	Every code in one domain, in table order.
`CODES_REVISION`	constant (`int`)	The revision of the code table this engine implements (§15.2).	Profiles and validation levels › Levels — how much to check
`Domain`	type	The area of the specification a diagnostic code belongs to (§15.2).
`MEDH5Error`	exception	Base of the package's own exceptions: a file, version, schema, validation or integrity error.	Python API › Exceptions
`MEDH5FileError`	exception	A file cannot be opened, is not HDF5, or is structurally unreadable.	Python API › Amending
`MEDH5FileNotFoundError`	exception	The file does not exist.	Python API › Exceptions
`MEDH5IntegrityError`	exception	A digest or `content_id` does not match the data it covers.
`MEDH5IOError`	exception	A read or write the engine could not complete: HDF5 failing a damaged chunk, say.	Python API › Exceptions
`MEDH5IsADirectoryError`	exception	The path names a directory, not a file.	Python API › Exceptions
`MEDH5PermissionError`	exception	The file exists but may not be read (or written).	Python API › Exceptions
`MEDH5SchemaError`	exception	The sample document is absent, is not JSON, or violates its JSON Schema.
`MEDH5ValidationError`	exception	Input rejected by a writer, or a validation failure raised as an exception.	Python API › Opening
`MEDH5VersionError`	exception	The file's `medh5_version` is one this operation cannot work with.	Python API › Amending
`Severity`	type	A diagnostic code's severity: an error fails validation, a warning does not.

`medh5.geometry`¶

Name	Kind	Summary	Documented in
`affine_summary`	function	Return a dict describing an affine: `spacing`, `origin`, `direction`, `determinant`, `orthonormal`.
`apply_affine_to_box`	function	Return the `(S, 2)` axis-aligned bounds of a box's corners mapped through an affine.
`AXIS_KINDS`	constant (`tuple`)
`box_corners`	function	Return the `2S` corners of an `(S, 2)` axis-aligned box as a `(2S, S)` array.
`box_to_slices`	function	Convert one `(S, 2)` box at voxel edges to a tuple of `slice(start, stop)`, one per axis (§8.1).	Detection and boxes › The half-voxel, once
`build_affine`	function	Return the `(S+1, S+1)` index-to-world affine built from spacing, origin and direction (§3.3).
`check_orthonormal`	function	Return `direction` as a float array, raising E102 unless it is orthonormal to within `tol`.
`check_pyramid`	function	Return a message for each pyramid level whose geometry breaks the derivation rule (§4.3).
`COORD_SYSTEMS`	constant (`tuple`)		Python API › Geometry
`decompose_affine`	function	Split an index-to-world affine back into `(spacing, origin, direction)` arrays (§3.3).
`derive_level_grid`	function	Derive the `Grid` of one pyramid level from level 0 and its per-axis downsample factors (§4.3).
`DOWNSAMPLE_METHODS`	constant (`tuple`)
`GEOMETRY_RTOL`	constant (`float`)
`Grid`	class	The lattice of a stored array: its shape, axes and index-to-world geometry (§3.2).	Annotation kinds › Boxes
`index_to_world`	function	Map continuous index coordinates to world coordinates through an `(S+1, S+1)` affine (§3.3).
`is_orthonormal`	function	Whether `direction` is a square matrix with `direction.T @ direction == I` to within `tol`.
`is_proper_rotation`	function	Whether `matrix` is orthonormal to within `tol` and has determinant +1.
`KNOWN_UNITS`	constant (`tuple`)
`LABEL_SAFE_METHODS`	constant (`tuple`)
`MAX_SPATIAL`	constant (`int`)
`MIN_SPATIAL`	constant (`int`)
`MM_PER_UNIT`	constant (`dict`)		Python API › Annotations
`ORTHONORMAL_TOL`	constant (`float`)		Converters › NIfTI
`Pyramid`	class	The multiscale declaration of an image group: its levels, factors, method and grids (§4.3).
`pyramid_factors`	function	Recover the `(L, S)` downsample factors of pyramid levels from their spacings relative to `base`.
`read_grid`	function	One grid group read back (`sample.root["grids/ct"]`).
`read_grids`	function	Every grid under a sample root, by id.
`slices_to_box`	function	Convert `slice(start, stop)` objects to one `(S, 2)` `float32` box at voxel edges (§8.1).
`SPEC_GRID_ATTRS`	constant (`tuple`)
`TIME_UNITS`	constant (`tuple`)
`voxel_volume`	function	Return the volume of one voxel: the product of `spacing`, in the grid's units to the power `S`.
`world_to_index`	function	Map world coordinates back to continuous index coordinates through an `(S+1, S+1)` affine (§3.3).

`medh5.image`¶

Name	Kind	Summary	Documented in
`check_value_type`	function	A known `value_type`, or E203.
`Image`	class	Lazy access to one image and its geometry.	Python API › Timepoints
`is_probability`	function	Whether every value lies in `[0, 1]` (and there is at least one).
`lossless_as_int16`	function	Whether a float array would survive `int16` storage unchanged (W907).
`SPEC_IMAGE_ATTRS`	constant (`tuple`)
`VALUE_TYPES`	constant (`tuple`)

`medh5.integrity`¶

Name	Kind	Summary	Documented in
`array_digest`	function	Return the `"<algo>:<hex>"` digest of an in-memory array stored at `path` (§13.1).
`ATTESTED_GROUPS`	constant (`tuple`)
`attrs_digest`	function	Return the digest of the canonical JSON of an object's named attributes (§13.2).
`canonical_attrs`	function	Canonical JSON over an object's named attributes (§13.2).
`collect_digests`	function	The `digest` of every dataset under a sample root that carries one.
`compute_content_id`	function	The Merkle root over stored digests, `meta` and canonical attributes.	Python API › Integrity
`dataset_digest`	function	The digest of a stored dataset over its decompressed content (§13.1).
`DEFAULT_ALGO`	constant (`str`)
`diagnose`	function	What is stale, mismatched or missing in one file.
`Diagnosis`	class	What is wrong with one file, before anything is done about it.
`DIGEST_ALGOS`	constant (`tuple`)
`digest_bytes`	function	Return the `"<algo>:<hex>"` digest of a byte string.
`fix`	function	Repair what was asked: rebuild stale indices, rewrite digests (with a reason, recorded as a provenance activity).	Storage › Atomic writes
`fix_paths`	function	`fix` over many files, every one of them: a file that cannot be repaired stops no other.
`group_digest`	function	The digest of an object group: its datasets and canonical attributes.
`parse_digest`	function	Split a `"<algo>:<hex>"` digest into `(algo, hex)`, raising E703 when it is malformed (§13.1).
`raw_chunks`	function	Return every stored chunk of a dataset as `bytes`, exactly as it sits on disk.
`relative_path`	function	An object's path relative to a sample root (the file root by default).
`Repair`	class	What was actually done to one file.
`stale_index_entries`	function	Return the names of the sampling index entries whose `source_digest` no longer matches (§13.3).
`STREAM_BYTES`	constant (`int`)	Datasets are hashed in slabs of at most this many bytes.
`subtrees_identical`	function	Return the paths at which two group trees differ byte for byte; empty means a pure copy.
`verify_object`	function	Whether the dataset at `path` under `root` matches its stored `digest` attribute (§13.1).
`verify_root`	function	Verify a sample root (`Sample.root`, or a collection member's).
`VerifyResult`	class	Outcome of a verification pass.	Python API › Integrity

`medh5.io`¶

Name	Kind	Summary	Documented in
`augment_from_fhir`	function	`read_fhir` source and add the records to the sample at path (`medh5.clinical.augment`), in place or into out: the augment report and the conversion's.	Converters › MEDS and FHIR R4
`augment_from_meds`	function	`read_meds` source and add the records to the sample at path (`medh5.clinical.augment`), in place or into out: the augment report and the conversion's.	Converters › MEDS and FHIR R4
`build_label_set`	function	Mint one label set covering a whole cohort's mask names and box labels.	Converters › 0.x files
`ConversionReport`	class	Every note from one conversion, plus what it produced.
`convert_world`	function	Re-express a 3-D index→world affine between RAS and LPS.
`from_dicom`	function	Convert a DICOM tree into one sample per subject (or per study).	Converters › DICOM
`from_dicom_seg`	function	Add a DICOM SEG's segments to an existing sample (§7).	Converters › DICOM SEG
`from_itk_transform`	function	Add the ITK transform at path to sample, from the frame of grid fixed to that of grid moving — ITK's fixed and moving images.	Converters › ITK transforms
`from_nifti`	function	Write one sample from a set of co-registered NIfTI volumes.	Converters › Conversion reports
`from_nnunetv2`	function	Convert an nnU-Net v2 dataset into one `.medh5` per case.	Converters › nnU-Net v2
`from_rtstruct`	function	Import an RTSTRUCT's ROIs as contours, and optionally rasterise them.	Converters › RTSTRUCT
`group_by_subject`	function	Group occasions into subjects (`mode="subject"`) or leave them apart.
`import_seg_nifti`	function	Add NIfTI masks to an existing sample, checking they sit on its grid.	Converters › NIfTI
`ItkTransform`	class	One ITK transform as a homogeneous matrix in LPS millimetres, mapping fixed points to moving ones (`x_moving = T(x_fixed)`).	Converters › ITK transforms
`merge_reports`	function	One report over a whole cohort.
`migrate`	function	Migrate one 0.x file into one 1.0 sample (Appendix B).	Converters › 0.x files
`migrate_paths`	function	Migrate a cohort, minting one label set for all of it.	Converters › 0.x files
`Note`	class	One thing a conversion did that the source did not fully determine.
`read_fhir`	function	The Observations and DiagnosticReports of subject (a reference, such as `Patient/123`) in source, as clinical records on clock, and what the conversion decided.	Converters › MEDS and FHIR R4
`read_itk_transform`	function	Read an ITK transform file: text (`.tfm`, `.txt`) or ANTs' MATLAB v4 `.mat`, holding one of the linear types or a composite of them.	Converters › ITK transforms
`read_meds`	function	The rows of MEDS subject subject_id in source, as clinical records on clock, and what the conversion decided.	Converters › MEDS and FHIR R4
`read_nifti`	function	One NIfTI file as `(array, geometry)` in MEDH5 conventions.	Converters › NIfTI
`scan_dicom`	function	Walk a directory tree and collect its image series.	Converters › DICOM
`Series`	class	One DICOM series, resolved into a volume and its geometry.
`SubjectGroup`	class	Occasions belonging to one subject, in timepoint order.
`to_dicom_seg`	function	Export one voxel annotation as a DICOM SEG against its source slices.	Converters › DICOM SEG
`to_itk_transform`	function	Write transform transform_id of sample as ITK reads it: its linear part to out (`.tfm`, `.txt` or `.mat`), and a displacement field applied first, as ANTs pairs them, to warp_out as a DISPVECT NIfTI (intent 1006) whose vectors are in the vectors convention, required.	Converters › ITK transforms
`to_nifti`	function	Export one image, or one class of one annotation, as a NIfTI file.	Converters › NIfTI
`to_nnunetv2`	function	Export samples as an nnU-Net v2 dataset.	Converters › nnU-Net v2
`to_rtstruct`	function	Export a `contours` annotation as an RT Structure Set.	Converters › RTSTRUCT
`write_itk_transform`	function	Write transform as ITK reads it, picking the format by suffix: `.tfm` or `.txt` text, or `.mat` as ANTs writes it.	Converters › ITK transforms

`medh5.io.dicom`¶

Name	Kind	Summary	Documented in
`from_dicom`	function	Convert a DICOM tree into one sample per subject (or per study).	Converters › DICOM
`read_series`	function	One series as `(volume, geometry)`, ordered and measured (§3.3).	Converters › DICOM
`require_pydicom`	function	Import pydicom, or raise an `ImportError` naming the extra to install.
`scan_dicom`	function	Walk a directory tree and collect its image series.	Converters › DICOM
`select_series`	function	One series per modality — the thinnest, or the first seen.	Converters › DICOM
`Series`	class	One DICOM series, resolved into a volume and its geometry.
`SPACING_TOLERANCE`	constant (`float`)	How far a slice may sit from where the regular stack puts it, as a fraction of the slice spacing, and never less than 0.01 mm.
`STORED_TAGS`	constant (`tuple`)	Acquisition parameters copied to `/meta → acquisition` (§4.5).

`medh5.io.dicom_seg`¶

Name	Kind	Summary	Documented in
`BINARY`	constant (`str`)		Converters › DICOM SEG
`FRACTIONAL`	constant (`str`)		Converters › DICOM SEG
`FRACTIONAL_TYPES`	constant (`tuple`)	What `to_dicom_seg`'s fractional_type takes (DICOM's `SegmentationFractionalType`).	Converters › DICOM SEG
`frames_agree`	function	Whether a SEG's frame of reference is a grid's; `None` when either states none, so there is nothing to compare.	Converters › DICOM SEG
`from_dicom_seg`	function	Add a DICOM SEG's segments to an existing sample (§7).	Converters › DICOM SEG
`MAX_FRACTIONAL_VALUE`	constant (`int`)	The stored value of a probability of 1 in a FRACTIONAL SEG this package writes: values are quantised to 1/255, so within 1/510 of what was given.	Converters › DICOM SEG
`place_frames`	function	Put each frame on the slice of grid that its position names (§3.3).
`read_dicom_seg`	function	A SEG file as `{segment number: volume}` plus its geometry and segments.	Converters › DICOM SEG
`read_dicom_seg_frames`	function	Every frame of a SEG, each with the segment and the position it carries.	Converters › DICOM SEG
`require_highdicom`	function	Import highdicom, or raise an `ImportError` naming the extra to install.
`SEGMENTATION_TYPES`	constant (`tuple`)	What `to_dicom_seg`'s segmentation_type takes.	Converters › DICOM SEG
`SLICE_TOLERANCE`	constant (`float`)	How far off a slice centre a frame may sit and still be that slice.
`to_dicom_seg`	function	Export one voxel annotation as a DICOM SEG against its source slices.	Converters › DICOM SEG

`medh5.io.fhir`¶

Name	Kind	Summary	Documented in
`augment_from_fhir`	function	`read_fhir` source and add the records to the sample at path (`medh5.clinical.augment`), in place or into out: the augment report and the conversion's.	Converters › MEDS and FHIR R4
`COMPARATORS`	constant (`dict`)	`Quantity.comparator` -> `value_comparator` (`ad` has none).	Converters › MEDS and FHIR R4
`read_fhir`	function	The Observations and DiagnosticReports of subject (a reference, such as `Patient/123`) in source, as clinical records on clock, and what the conversion decided.	Converters › MEDS and FHIR R4
`STATUSES`	constant (`dict`)	FHIR R4 Observation and DiagnosticReport status -> event status (`registered` is left out).	Converters › MEDS and FHIR R4
`UCUM`	constant (`str`)		Converters › MEDS and FHIR R4

`medh5.io.grouping`¶

Name	Kind	Summary
`build_occasions`	function	Adapt any sequence of source objects into occasions.
`contradictions`	function	Demographics on which occasions disagree: `{field: [values]}`.
`FALLBACK_PREFIX`	constant (`str`)
`group_by_subject`	function	Group occasions into subjects (`mode="subject"`) or leave them apart.
`note_instance_ids`	function	Record that objects were not joined across merged studies (§7.4).
`Occasion`	class	One study/visit of one subject, before it becomes a timepoint.
`output_name`	function	A unique filename stem for one group; it is added to used.
`SubjectGroup`	class	Occasions belonging to one subject, in timepoint order.

`medh5.io.itk`¶

Name	Kind	Summary	Documented in
`from_itk_transform`	function	Add the ITK transform at path to sample, from the frame of grid fixed to that of grid moving — ITK's fixed and moving images.	Converters › ITK transforms
`ItkTransform`	class	One ITK transform as a homogeneous matrix in LPS millimetres, mapping fixed points to moving ones (`x_moving = T(x_fixed)`).	Converters › ITK transforms
`read_itk_transform`	function	Read an ITK transform file: text (`.tfm`, `.txt`) or ANTs' MATLAB v4 `.mat`, holding one of the linear types or a composite of them.	Converters › ITK transforms
`to_itk_transform`	function	Write transform transform_id of sample as ITK reads it: its linear part to out (`.tfm`, `.txt` or `.mat`), and a displacement field applied first, as ANTs pairs them, to warp_out as a DISPVECT NIfTI (intent 1006) whose vectors are in the vectors convention, required.	Converters › ITK transforms
`write_itk_transform`	function	Write transform as ITK reads it, picking the format by suffix: `.tfm` or `.txt` text, or `.mat` as ANTs writes it.	Converters › ITK transforms

`medh5.io.legacy`¶

Name	Kind	Summary	Documented in
`BOX_SHIFT`	constant (`float`)	0.x `[min, max)` integer boxes sit at voxel edges once shifted (§8.1).
`build_label_set`	function	Mint one label set covering a whole cohort's mask names and box labels.	Converters › 0.x files
`is_legacy`	function	True when path is a readable 0.x file.
`legacy_meta`	function	A 0.x file's metadata, without its arrays.
`LegacyMeta`	class	0.x metadata, as far as a migration needs it.
`LegacySample`	class	A whole 0.x file in memory.
`LegacySpatial`	class	0.x geometry: one grid, shared by every image in the file.
`load_sidecar`	function	Read a reviewed label-set sidecar back.
`migrate`	function	Migrate one 0.x file into one 1.0 sample (Appendix B).	Converters › 0.x files
`migrate_paths`	function	Migrate a cohort, minting one label set for all of it.	Converters › 0.x files
`read_legacy`	function	A whole 0.x file.
`SCHEMA_VERSION`	constant (`str`)	The only 0.x schema version that ever shipped.
`SUFFIX`	constant (`str`)
`write_sidecar`	function	Write the minted label set for review before a cohort-wide migration.

`medh5.io.meds`¶

Name	Kind	Summary	Documented in
`augment_from_meds`	function	`read_meds` source and add the records to the sample at path (`medh5.clinical.augment`), in place or into out: the augment report and the conversion's.	Converters › MEDS and FHIR R4
`BIRTH`	constant (`str`)	The MEDS code of a subject's birth.	Converters › MEDS and FHIR R4
`BOTH_VALUES`	constant (`tuple`)		Converters › MEDS and FHIR R4
`COLUMNS`	constant (`tuple`)	The columns `columns=` may rename: MEDS's own five, and the two MEDS lacks, which are read only when declared.	Converters › MEDS and FHIR R4
`read_meds`	function	The rows of MEDS subject subject_id in source, as clinical records on clock, and what the conversion decided.	Converters › MEDS and FHIR R4

`medh5.io.nifti`¶

Name	Kind	Summary	Documented in
`convert_world`	function	Re-express a 3-D index→world affine between RAS and LPS.
`COORD_SYSTEMS`	constant (`tuple`)		Python API › Geometry
`from_nifti`	function	Write one sample from a set of co-registered NIfTI volumes.	Converters › Conversion reports
`import_seg_nifti`	function	Add NIfTI masks to an existing sample, checking they sit on its grid.	Converters › NIfTI
`RAS_TO_LPS`	constant (`ndarray`)	World-axis sign flip.
`read_nifti`	function	One NIfTI file as `(array, geometry)` in MEDH5 conventions.	Converters › NIfTI
`require_nibabel`	function	Import nibabel, or raise an `ImportError` naming the extra to install.
`to_nifti`	function	Export one image, or one class of one annotation, as a NIfTI file.	Converters › NIfTI

`medh5.io.nnunetv2`¶

Name	Kind	Summary	Documented in
`BACKGROUND`	constant (`int`)
`cases`	function	Case identifiers present in `imagesTr` — and, with include_test, in `imagesTs`.	Run the conformance suite › Publish a suite
`from_nnunetv2`	function	Convert an nnU-Net v2 dataset into one `.medh5` per case.	Converters › nnU-Net v2
`read_dataset_json`	function	Load and check an nnU-Net v2 `dataset.json`.
`REQUIRED_KEYS`	constant (`tuple`)
`to_nnunetv2`	function	Export samples as an nnU-Net v2 dataset.	Converters › nnU-Net v2

`medh5.io.report`¶

Name	Kind	Summary
`ConversionReport`	class	Every note from one conversion, plus what it produced.
`merge_reports`	function	One report over a whole cohort.
`Note`	class	One thing a conversion did that the source did not fully determine.
`SEVERITIES`	constant (`tuple`)	`decision` was determined by the data; `guess` was not.

`medh5.io.rtstruct`¶

Name	Kind	Summary	Documented in
`from_rtstruct`	function	Import an RTSTRUCT's ROIs as contours, and optionally rasterise them.	Converters › RTSTRUCT
`MIN_POLYGON_VERTICES`	constant (`int`)
`read_rtstruct`	function	An RTSTRUCT as `{roi number: [contour, ...]}` in patient coordinates.
`to_rtstruct`	function	Export a `contours` annotation as an RT Structure Set.	Converters › RTSTRUCT

`medh5.labels`¶

Name	Kind	Summary	Documented in
`available`	function	Every vocabulary name `load` accepts, bundled and registered.	Annotation kinds › Label sets
`BACKGROUND_ID`	constant (`int`)
`canonical_json`	function	Return the canonical JSON bytes of a document: sorted keys, no whitespace, UTF-8 (§5.1).
`check`	function	Whether sample files share one vocabulary, as `medh5 labels check --json`.	Choose and map a label set › Find where a cohort disagrees
`check_class_id`	function	Return `class_id` if it is an assignable class id, 1 to 65534 (§5.3).
`CLOSURES`	constant (`tuple`)
`describe`	function	Name -> `{id, version, classes, sha256}`.	Python API › Sub-packages
`FORMS`	constant (`tuple`)
`from_keys`	function	Mint an inline `LabelSet` from bare class keys, with ids counting up from `start`.
`IGNORE_ID`	constant (`int`)
`INLINE_REQUIRED_BELOW`	constant (`int`)
`LabelClass`	class	One semantic class (§5.2).	Annotation kinds › Closure
`LabelSet`	class	A controlled vocabulary (§5.1).	Annotation kinds › Label sets
`load`	function	A bundled or registered vocabulary by name; E305 for an unknown one.	Annotation kinds › Label sets
`load_file`	function	A vocabulary from a JSON file on disk.	Converters › NIfTI
`MAX_CLASS_ID`	constant (`int`)
`OntologyCode`	class	A binding of a class to a code of an external vocabulary, such as SNOMED-CT or RadLex (§5.2).
`register`	function	Make a vocabulary loadable by name for the rest of the process.	Curation records › Provenance
`Relation`	class	A non-`is_a` edge between two classes of a label set, such as `part_of` (§5.4).
`Skeleton`	class	A keypoint topology declared by a label set: keypoint class ids and edges between them (§5.5).
`unregister`	function	Drop a registered vocabulary; bundled ones cannot be removed.

`medh5.monai`¶

Name	Kind	Summary	Documented in
`affine_for`	function	The index→world affine of one image, as MONAI receives it: optionally re-expressed in space, and in millimetres (see `_monai_affine`).	PyTorch and MONAI › MONAI
`available`	function	Whether MONAI can be imported.	PyTorch and MONAI › MONAI
`convert_affine`	function	Re-express an index→world affine in another world convention.	PyTorch and MONAI › A reader for `LoadImage`
`from_metatensor`	function	A `MetaTensor` back to `(array, geometry)` a writer can consume.	PyTorch and MONAI › MONAI
`LoadMedH5d`	class	`medh5.monai.to_dict` as a MONAI dictionary transform.	PyTorch and MONAI › A reader for `LoadImage`
`MedH5Reader`	class	A MONAI `ImageReader` for `.medh5` samples and collection members.	PyTorch and MONAI › A reader for `LoadImage`
`meta_dict`	function	The metadata MONAI carries alongside a tensor.	PyTorch and MONAI › MONAI
`require_monai`	function	Import MONAI, or explain how to install it.
`SPACES`	constant (`tuple`)		Python API › Geometry
`to_dict`	function	A MONAI dictionary-transform item: `{image_id: MetaTensor, ...}`.	PyTorch and MONAI › MONAI
`to_metatensor`	function	One image (or ROI of it) as a MONAI `MetaTensor`.	PyTorch and MONAI › MONAI

`medh5.nodes`¶

Name	Kind	Summary	Documented in
`Attrs`	class	An object's attributes, as a mutable mapping.
`Dataset`	class	A stored dataset: shape, dtype, chunks, filters, attributes and `ds[...]`.	Python API › Stored objects
`Group`	class	A stored group: members by name (or path), and attributes.

`medh5.sample`¶

Name	Kind	Summary	Documented in
`amend`	function	Copy-on-write amend: build a new file from the old and replace it.	Python API › Amending
`annotation_id`	function	An annotation id from a reference that may be written as a path.
`AnnotationCollection`	class	The sample's annotations, with task- and timepoint-aware views.
`create`	function	Create a new sample; use as a context manager, or call `commit()`.	Annotation kinds › Closure
`FORMAT_VERSION`	constant (`str`)		medh5 › Versioning
`FRAME_ATTRS`	constant (`tuple`)	Where frame-of-reference uids are named: `(group, attribute names)`.
`ImageCollection`	class	The sample's images, with timepoint- and modality-aware views.
`MANAGED_ROOT_ATTRS`	constant (`tuple`)	Root attributes `commit` writes itself; an amend does not copy them.
`open_sample`	function	Open a `.medh5` sample file, read-only.	Python API › Opening
`PROFILES`	constant (`tuple`)		Profiles and validation levels › Profiles — what to hold the file to
`ROOT_DIGEST_ATTRS`	constant (`tuple`)	Root attributes covered by `content_id`.
`Sample`	class	A read-only view of one sample root (a file, or a collection member).	Python API › Sample
`SampleWriter`	class	Builder for one sample.	Python API › Writing
`STANDARD_GROUPS`	constant (`tuple`)
`TimepointView`	class	Everything in the sample that belongs to one timepoint.

`medh5.sampling`¶

Name	Kind	Summary	Documented in
`BLEND_MODES`	constant (`tuple`)	How overlapping patch predictions are weighted when stitched back: `constant` averages them, `gaussian` trusts each patch most at its centre, where its model saw the most context.	PyTorch and MONAI › GridPatchDataset — deterministic tiling, for inference
`blend_weights`	function	The weight of each voxel of a patch when patches are stitched back.	PyTorch and MONAI › GridPatchDataset — deterministic tiling, for inference
`coerce_patch_size`	function	Broadcast a scalar patch size across ndim spatial axes.
`grid_patches`	function	Deterministic sliding-window cover of a volume — the inference path.	PyTorch and MONAI › GridPatchDataset — deterministic tiling, for inference
`iter_patches`	function	The patches of `PatchSampler.draws`, drawn one per `next()`.	Python API › Sampling patches
`PAIR_MODES`	constant (`tuple`)
`PairReport`	class	What a paired dataset kept and what it skipped.
`Patch`	class	One draw: where to read, and how the location was chosen.	Python API › Sampling patches
`PatchAggregator`	class	Stitch patch predictions back into one volume.	PyTorch and MONAI › GridPatchDataset — deterministic tiling, for inference
`PatchSampler`	class	Choose patch windows in a volume (spec §14.3).	Python API › Sampling patches
`Seed`	type	What seeds a draw: an integer, a sequence of integers — `(seed, epoch, index)` — a `numpy.random.Generator`, which supplies one seed and advances, or `None` for fresh entropy.	Python API › Sampling patches
`STRATEGIES`	constant (`tuple`)
`TimepointPair`	class	One ordered pair of visits, and the change label spanning it, if any.
`TimepointPairSampler`	class	Enumerate the visit pairs a longitudinal model trains on (§3.7, §9).	PyTorch and MONAI › PairedPatchDataset — the same place at two visits
`window_around`	function	Slices covering patch voxels around center, plus the padding needed.

`medh5.storage`¶

Name	Kind	Summary	Documented in
`BLOSC2_FILTER_ID`	constant (`int`)
`BLOSC_FILTER_ID`	constant (`int`)
`build_index`	function	Compute the sampling index of one voxel annotation (not written).	Python API › Derived data
`BUILTIN_FILTER_IDS`	constant (`frozenset`)
`BULK_MIN_BYTES`	constant (`int`)	Datasets at least this large are bulk: W902 wants them compressed.
`CACHE_SAFETY`	constant (`float`)
`chunk_report`	function	Return a dict describing a chunking: `chunks`, `chunk_bytes`, `chunk_mib` and `n_chunks`.
`Codec`	class	One codec setting.	Storage › Codec profiles
`CodecProfile`	class	A named pairing of codecs for image data and for label/field data.
`COMPRESS_MIN_BYTES`	constant (`int`)	Datasets smaller than this are stored contiguous and unfiltered.
`dataset_layout`	function	The layout the writer gives one dataset: `{}` when it is stored contiguous (small or empty), else `{"chunks": ..., "codec": ...}`, the codec being profile's for role — a `CodecProfile` of your own as much as a named one.	Storage › Codec profiles
`DEFAULT_L3_BYTES`	constant (`int`)
`DEFAULT_MAX_COORDS`	constant (`int`)
`DEFAULT_OCCUPANCY_FACTOR`	constant (`int`)
`DEFAULT_PATCH`	constant (`int`)
`DEFAULT_PROFILE`	constant (`str`)
`describe_filters`	function	A dataset's actual filter pipeline, e.g.
`detect_l3_bytes`	function	Return one CPU's share of the L3 cache in bytes, best effort, with a fallback of ~1.375 MiB.
`field_chunks`	function	Return the chunks of a displacement field on `grid`: one vector component per chunk, or `None`.
`fit_chunks`	function	Return the trailing `len(shape)` entries of `proposed` clipped to `shape`, or `None`.
`grid_chunks`	function	Return the chunk shape the writer gives an array laid out on `grid` (§14.1).
`IndexPayload`	class	The datasets of one annotation's index entry (§14.3).
`is_bulk`	function	Whether a dataset is large enough for the W902 warning.
`MAX_CHUNK_BYTES`	constant (`float`)
`MIN_CHUNK_BYTES`	constant (`float`)
`occupancy`	function	The occupancy map of a mask: one bit per `factor`-cube of voxels (`factor` at least 1).	Storage › The sampling index
`optimize_chunks`	function	Return a full chunk shape for an array whose axes are described by `axis_kinds` (§14.1).
`OVERSHOOT_LIMIT`	constant (`float`)
`profile_family`	function	`portable` when every dataset of the file needs only HDF5's own filters, else `balanced` — what an amend of it defaults to.
`PROFILES`	constant (`dict`)		Storage › Codec profiles
`read_indices`	function	Every stored index entry under a sample root (`Sample.root`, or the `Sample`), by annotation id.
`recompress`	function	Rewrite path (or write out) with every bulk dataset re-encoded under profile; `rechunk` also re-derives chunk shapes from the grids.	Storage › Atomic writes
`recompress_paths`	function	Recompress many files, one result each: every one of them — a file that cannot be recompressed stops no other.
`RecompressResult`	class	What one file's re-encoding did.	Storage › Integrity
`resolve_profile`	function	A profile by name (`None` is `balanced`), or the profile given.	Storage › Codec profiles
`Role`	type	What a dataset holds, which picks its codec profile (§14.2).
`SamplingIndex`	class	A reader for one `index/<ann_id>` entry: the sampling index of the `training` profile (§14.3).	Python API › Sampling index
`spatial_chunk_for`	function	Return chunk extents for the spatial axes alone, sized to a training patch and the L3 cache.

`medh5.task`¶

Name	Kind	Summary	Documented in
`CENSORING`	constant (`tuple`)	What a target does with a row it cannot label (contract §5).
`CODES`	constant (`dict`)	The companion contract's finding codes (T1xx manifest, T2xx identity and splits, T3xx sources, T4xx caches) and what each means.	Python API › Tasks and caches
`concept_token`	function	An event's concept token: `kind\|code_system\|code`, or `kind\|` when it is uncoded.	Python API › Tasks and caches
`Confine`	type	Where a runner lets a task's sources be opened from (contract §2): `True` confines them to the base relative URIs resolve against, a path to that directory, `False` opens wherever the URIs point.	Python API › Tasks and caches
`EventTable`	class	One subject's merged event versions, as columns.
`Finding`	class	One thing wrong with a task, its sources, or a cache.	Python API › Tasks and caches
`LinkTable`	class	One subject's links, every fragment's, as columns; indexing builds a `medh5.clinical.Link`, and `fragment(i)` says whose it is.
`Note`	class	Something a manifest should state and does not.	Python API › Tasks and caches
`Packed`	class	A packed UTF-8 column: one buffer, `int64` offsets, and a validity mask (`None` when every cell is valid) — the clinical tables' own encoding (1.1 §4).
`Preflight`	class	A task's preflight: manifest and source findings, every subject's merged history, and every row's view.	Python API › Tasks and caches
`preflight`	function	Preflight a task (a manifest, or the path of one); `confine` says where sources may be opened (`Confine`).	Python API › Tasks and caches
`Reconciled`	class	An event version several fragments hold, and the digest they share.
`Repinned`	class	One source pinned to a new version.	Python API › Tasks and caches
`RepinReport`	class	What `TaskManifest.repin` did, or would do.	Python API › Tasks and caches
`ROI`	constant (`tuple`)	What a slot may request its region be centred on (contract §3.5).	Segmentation › Read them back
`ROIS`	constant (`tuple`)	How a slot's region was centred: `ROI`, and `center_fallback` when a slot asked for an eligible instance and the row has none.
`Row`	class	One example: a subject at a cutoff.
`ROW_STATUSES`	constant (`tuple`)
`RowView`	class	Everything a task admits for one row, and nothing else.	Python API › Tasks and caches
`SCHEMA`	constant (`str`)	The manifest's own version.
`schema_text`	function	The `medh5.task/1` JSON Schema.	Task manifest schema
`SELECTION_STATUSES`	constant (`tuple`)
`Slot`	class	A stable modality slot, filled per row by the newest eligible image of its modality — never by a filename (contract §3.5).	Python API › Tasks and caches
`SlotFill`	class	How one slot was filled for one row — or that it was not.
`SourceRef`	class	One sample — a file, or a member of a `.medh5c` — at a pinned version.	Python API › Tasks and caches
`Subject`	class	One subject: its fragments, its clock and its partition.
`SubjectHistory`	class	One subject's history, merged across its fragments: what every row of the subject indexes.
`Target`	class	What a row is labelled with (contract §5): an event concept whose final `value_text` falls in `positive` or `negative` inside `(cutoff, cutoff + horizon_us]`, read from the full history.	Python API › Tasks and caches
`TARGET_STATUSES`	constant (`tuple`)
`TargetLabel`	class	A row's target: `positive` (1.0), `negative` (0.0), or unobserved (`censored`, `prevalent`, `none`) with the reason.
`TaskManifest`	class	A `medh5.task/1` manifest: build it, check it, save it, preflight it.	Python API › Tasks and caches

`medh5.torch`¶

Name	Kind	Summary	Documented in
`ALIGNMENTS`	constant (`tuple`)
`AVAILABLE`	constant (`bool`)
`CACHE`	constant (`HandleCache`)		PyTorch and MONAI › Keeping a file's items together
`ClinicalTaskDataset`	class	One item per admitted row of a task.	PyTorch and MONAI › Clinical tasks (format 1.1)
`collate`	function	`DataLoader(collate_fn=...)` for the MEDH5 item shape.	PyTorch and MONAI › Keeping a file's items together
`collate_clinical`	function	`DataLoader(collate_fn=...)` for `ClinicalTaskDataset` items.	PyTorch and MONAI › Clinical tasks (format 1.1)
`ConceptVocabulary`	class	Concept indices, per-concept value statistics and the categories each concept's text values took, fitted on one partition of one task — and saying which.	PyTorch and MONAI › Clinical tasks (format 1.1)
`FileGroupedSampler`	class	Shuffle files, and yield every item of a file before the next file.	PyTorch and MONAI › Keeping a file's items together
`grid_patches`	function	Deterministic sliding-window cover of a volume — the inference path.	PyTorch and MONAI › GridPatchDataset — deterministic tiling, for inference
`GridPatchDataset`	class	Deterministic sliding-window cover of every file — the inference path.	PyTorch and MONAI › GridPatchDataset — deterministic tiling, for inference
`GridStitcher`	class	Reassemble a `medh5.torch.GridPatchDataset`'s predictions.	PyTorch and MONAI › GridPatchDataset — deterministic tiling, for inference
`HandleCache`	class	PID-scoped, thread-safe LRU cache of open samples.	PyTorch and MONAI › Keeping a file's items together
`LABEL_FORMATS`	constant (`tuple`)
`open_cached`	function	Open a sample — or a collection member — through the per-worker cache.
`PairedPatchDataset`	class	Corresponding patches from two visits of one subject (§3.7, §10).	PyTorch and MONAI › PairedPatchDataset — the same place at two visits
`PairReport`	class	What a paired dataset kept and what it skipped.
`Patch`	class	One draw: where to read, and how the location was chosen.	Python API › Sampling patches
`PatchDataset`	class	`samples_per_volume` patches from each file, drawn by a sampler.	PyTorch and MONAI › PatchDataset — random patches
`PatchSampler`	class	Choose patch windows in a volume (spec §14.3).	Python API › Sampling patches
`require_torch`	function	Without PyTorch, raise an `ImportError` that says how to install it — or, when it is installed and fails to import, why.
`set_cache_size`	function	Resize this process's cache: how many files each worker keeps open.	PyTorch and MONAI › Keeping a file's items together
`stack_images`	function	Channel-stack several images into one `(B, C, *patch)` tensor.
`TimepointPair`	class	One ordered pair of visits, and the change label spanning it, if any.
`TimepointPairSampler`	class	Enumerate the visit pairs a longitudinal model trains on (§3.7, §9).	PyTorch and MONAI › PairedPatchDataset — the same place at two visits
`VolumeDataset`	class	One item per file: the whole volume, eagerly read.	PyTorch and MONAI › VolumeDataset — whole volumes
`worker_init_fn`	function	`DataLoader(worker_init_fn=...)`: drop handles inherited across `fork` (in a process that opened them, close them; see `HandleCache.clear`).	PyTorch and MONAI › The DataLoader

`medh5.torch.clinical`¶

Name	Kind	Summary	Documented in
`ClinicalTaskDataset`	class	One item per admitted row of a task.	PyTorch and MONAI › Clinical tasks (format 1.1)
`collate_clinical`	function	`DataLoader(collate_fn=...)` for `ClinicalTaskDataset` items.	PyTorch and MONAI › Clinical tasks (format 1.1)
`concept_of`	function	An event's concept token: `kind\|code_system\|code` (`kind\|` uncoded), a `\` or `\|` in the system or code escaped (`medh5.task.concept_token`).
`ConceptVocabulary`	class	Concept indices, per-concept value statistics and the categories each concept's text values took, fitted on one partition of one task — and saying which.	PyTorch and MONAI › Clinical tasks (format 1.1)
`PAD`	constant (`int`)	Concept index of padding.
`UNKNOWN`	constant (`int`)	Concept index of a concept the vocabulary was not fitted on.

`medh5.torch.collate`¶

Name	Kind	Summary	Documented in
`collate`	function	`DataLoader(collate_fn=...)` for the MEDH5 item shape.	PyTorch and MONAI › Keeping a file's items together
`stack_images`	function	Channel-stack several images into one `(B, C, *patch)` tensor.

`medh5.torch.datasets`¶

Name	Kind	Summary	Documented in
`ALIGNMENTS`	constant (`tuple`)
`GridPatchDataset`	class	Deterministic sliding-window cover of every file — the inference path.	PyTorch and MONAI › GridPatchDataset — deterministic tiling, for inference
`LABEL_FORMATS`	constant (`tuple`)
`PairedPatchDataset`	class	Corresponding patches from two visits of one subject (§3.7, §10).	PyTorch and MONAI › PairedPatchDataset — the same place at two visits
`PatchDataset`	class	`samples_per_volume` patches from each file, drawn by a sampler.	PyTorch and MONAI › PatchDataset — random patches
`VolumeDataset`	class	One item per file: the whole volume, eagerly read.	PyTorch and MONAI › VolumeDataset — whole volumes

`medh5.torch.handles`¶

Name	Kind	Summary	Documented in
`CACHE`	constant (`HandleCache`)		PyTorch and MONAI › Keeping a file's items together
`DEFAULT_MAXSIZE`	constant (`int`)
`HandleCache`	class	PID-scoped, thread-safe LRU cache of open samples.	PyTorch and MONAI › Keeping a file's items together
`open_cached`	function	Open a sample — or a collection member — through the per-worker cache.
`set_cache_size`	function	Resize this process's cache: how many files each worker keeps open.	PyTorch and MONAI › Keeping a file's items together
`worker_init_fn`	function	`DataLoader(worker_init_fn=...)`: drop handles inherited across `fork` (in a process that opened them, close them; see `HandleCache.clear`).	PyTorch and MONAI › The DataLoader

`medh5.torch.samplers`¶

Name	Kind	Summary	Documented in
`FileGroupedSampler`	class	Shuffle files, and yield every item of a file before the next file.	PyTorch and MONAI › Keeping a file's items together

`medh5.torch.stitch`¶

Name	Kind	Summary	Documented in
`GridStitcher`	class	Reassemble a `medh5.torch.GridPatchDataset`'s predictions.	PyTorch and MONAI › GridPatchDataset — deterministic tiling, for inference

`medh5.transforms`¶

Name	Kind	Summary	Documented in
`AffineTransform`	class	A homogeneous `(S+1, S+1)` matrix in world coordinates.	Converters › ITK transforms
`basis`	function	The B-spline basis weights at `t`: `(order + 1, *t.shape)`.
`BSplineTransform`	class	A free-form deformation evaluated with a uniform B-spline basis.
`ChainTransform`	class	Several transforms applied in order — a resolved path.	Registration between visits › Write one
`check_transform_id`	function	Return `transform_id` if it is a valid object id, else raise E003.
`CompositeTransform`	class	Components applied in order; each one's `to_frame` is the next's `from_frame`.	Converters › ITK transforms
`cubic_sample`	function	Interpolate `(C, *spatial)` data with cubic splines at `(N, S)` continuous indices: `(N, C)`.
`DEFAULT_ORDER`	constant (`int`)
`DisplacementTransform`	class	`T(x) = x + u(x)` with `u` sampled on `field_grid`.
`encode_affine`	function	Pack a homogeneous world-to-world `(S+1, S+1)` matrix as an `affine` transform payload (§10.3).
`encode_bspline`	function	Pack `(S, *cp_shape)` B-spline control-point coefficients as a `bspline` payload (§10.5).
`encode_composite`	function	Declare a `composite` transform: two or more component transform ids, applied in order (§10.5).
`encode_displacement`	function	Pack a `(S, spatial)` displacement field (§10.4), cast to dtype*.
`encode_identity`	function	Return the payload of an identity transform, which stores nothing but its endpoints.
`EXTRAPOLATIONS`	constant (`tuple`)
`FLOAT16_SAFE_VOXELS`	constant (`float`)
`folding_fraction`	function	Return the fraction of Jacobian determinants that are `<= 0`: where the transform folds.	Python API › Transforms
`frame_graph`	function	Frame -> frames one hop away, as `resolve_between` would walk them: a reverse edge exists only where the inverse can be evaluated, and a transform and the stored inverse its `inverse_id` names are one edge each way (§10.1).
`frames_of_timepoint`	function	The frames of every grid of one timepoint, in grid order.	Registration between visits › When there is no path
`IdentityTransform`	class	Two frames that coincide, said explicitly.	Converters › ITK transforms
`inside_extent`	function	Return a boolean array saying which points lie within a lattice's voxel extent, edges included.
`INTERPOLATIONS`	constant (`tuple`)
`InverseTransform`	class	The inverse of another transform, evaluated (not stored).	Registration between visits › Write one
`jacobian_determinant`	function	Return `det(I + du/dx)` per voxel of a displacement field on `grid`; values `<= 0` mark folding.	Python API › Transforms
`LAST_ROW_TOL`	constant (`float`)
`linear_part`	function	Return a grid's `direction @ diag(spacing)`: the `(S, S)` map from index to world displacement.
`linear_sample`	function	Interpolate `(C, *spatial)` data multilinearly at `(N, S)` continuous indices: an `(N, C)` array.
`refuse_outside`	function	Raise `MEDH5ValidationError` if any of the `inside` flags is false: `extrapolation='error'`.
`resolve_between`	function	The transform relating two frame uids, or `None` when none does.
`sample_field`	function	Interpolate a field at `(N, S)` continuous indices with the declared `interpolation` (§10.4).
`SPEC_TRANSFORM_ATTRS`	constant (`tuple`)
`SUPPORTED_ORDERS`	constant (`tuple`)
`target_registration_error`	function	TRE `‖T(p_F) − p_M‖` over world landmarks with matching row order (§10.6): `{mean, median, max, n}`.	Python API › Transforms
`to_world_vectors`	function	Convert `(N, S)` displacement vectors in `vector_space` on `grid` to world units (§10.4).
`Transform`	class	What every transform answers, whatever its kind.	Python API › Transforms
`TRANSFORM_KINDS`	constant (`tuple`)
`TransformHeader`	class	The attribute header every transform carries (spec §10.1).
`VECTOR_SPACES`	constant (`tuple`)
`wrap_transform`	function	The facade class for an engine transform handle.

`medh5.validate`¶

Name	Kind	Summary	Documented in
`Diagnostic`	class	One finding.	Python API › Validation
`Level`	type	A validation level (§15.1); each includes the checks of the levels before it.
`LEVEL_ORDER`	constant (`dict`)
`LEVELS`	constant (`tuple`)
`merge`	function	Combine several reports, prefixing each diagnostic's location with its file.
`Report`	class	Everything a validation pass found.
`rules_for`	function	The rules a pass at level runs, in order.
`validate_file`	function	Validate one `.medh5` sample or `.medh5c` collection file.	Python API › Validation
`validate_paths`	function	Validate many files, one report each, as `validate_file` does.	Python API › Validation
`validate_root`	function	Validate an open `medh5.sample.Sample` — a file, or a member of a collection — held to the profiles it declares and to profiles besides.

API index¶

medh5¶

medh5.annotations¶

medh5.annotations.base¶

medh5.annotations.classification¶

medh5.annotations.geometric¶

medh5.annotations.voxel¶

medh5.bench¶

medh5.cache¶

medh5.cli¶

medh5.clinical¶

medh5.collection¶

medh5.conformance¶

medh5.curation¶

medh5.curation.agreement¶

medh5.curation.scrub¶

medh5.curation.splits¶

medh5.dataset¶

medh5.dataset.check¶

medh5.dataset.manifest¶

medh5.dataset.split¶

medh5.dataset.stats¶

medh5.document¶

medh5.errors¶

medh5.geometry¶

medh5.image¶

medh5.integrity¶

medh5.io¶

medh5.io.dicom¶

medh5.io.dicom_seg¶

medh5.io.fhir¶

medh5.io.grouping¶

medh5.io.itk¶

medh5.io.legacy¶

medh5.io.meds¶

medh5.io.nifti¶

medh5.io.nnunetv2¶

medh5.io.report¶

medh5.io.rtstruct¶

medh5.labels¶

medh5.monai¶

medh5.nodes¶

medh5.sample¶

medh5.sampling¶

medh5.storage¶

medh5.task¶

medh5.torch¶

medh5.torch.clinical¶

medh5.torch.collate¶

medh5.torch.datasets¶

medh5.torch.handles¶

medh5.torch.samplers¶

medh5.torch.stitch¶

medh5.transforms¶

medh5.validate¶