scripts/gen-annotations/README.md

Regenerating MedVision data

These scripts rebuild the preprocessed images/masks and the benchmark-plan annotations for any dataset in the catalogue, from the original public sources. They are the record of how the published data was actually produced.

They do not upload anything. Publishing to HuggingFace is a separate, deliberate step.

setup-env.sh        install medvision_ds
dataset_specs.py    per-dataset recipe: download entrypoint, steps, reorientation policy
build_dataset.py    the driver

Quick start

./setup-env.sh /path/to/data-dir

# see what would run, without running it
python build_dataset.py --data_dir /path/to/data-dir --dataset KiTS23 --dry_run

# build it
python build_dataset.py --data_dir /path/to/data-dir --dataset KiTS23

--data_dir is your MedVision_DATA_DIR; datasets are built under <data_dir>/Datasets/. Use --all for the whole catalogue, --steps to run a subset (download,segmentation,detection,biometry), and --max_workers for the downloads that support parallelism.

• download: data retrieval and preprocessing
• segmentation: generate MaskSize annotations
• detection: generate bounding box annotations
• biometry: generate tumor/lesion size or angle/distance annotations

Two modes, and only two

Reproduce what is published — the default, also spelled --latest. Each (dataset, task) pair (i.e., config) resolves to the newest version declared for it in _ANNOTATION_INDEX (in MedVision.py), and that value is passed down as --annotation_version. Config of one dataset can sit at different versions and are handled in a single run: KiTS23 emits 1.0.0 for segmentation and detection and 1.1.1 for biometry.

Publish a new annotation version — --new-annotation-version, for maintainers. Bump src/medvision_ds/__version__.py first; the driver then passes no version and the planner stamps __version__. The run is refused unless that version is strictly above every affected pair's newest published version, so you cannot overwrite published data by forgetting to bump.

There is no third mode. In particular you cannot ask for an older version, because the current codebase cannot produce one: the generation code changes between annotation versions — that is why the version is bumped. v1.1.0 changed the tumour/lesion cluster-size threshold (200px → 20px), v1.1.1 corrected the transposed in-plane spacing in the ellipse fit, v1.2.1 pinned float promotion against NEP 50. Naming an older version would emit a file with that name but with today's values. To rebuild an older version, check out the git commit with "[release]" in the commit title and install that src/.

Credentials

Read from the environment only. Never put a token in a file that a script reads.

Variable	Needed for
SYNAPSE_TOKEN	BCV15, BraTS24, FeTA24, MAMA-MIA	required; the download fails without it
HF_TOKEN	SKM-TEA, ToothFairy2	only if their HuggingFace mirror is private for you
MedVision_SKMTEA_HF_ID, MedVision_ToothFairy2_HF_ID, BiometricVQA_KiPA22_HF_ID	those three	optional; override a working public default

The driver checks the required ones before running anything and names the missing variable.

These tokens are for reproduction, not for loading. Ordinary load_dataset use needs no credentials across most of the catalogue: 21 of the 30 datasets fetch an already-preprocessed copy, so nothing is retrieved from the gated source. BCV15, BraTS24 and MAMA-MIA are gated upstream but ship such a copy, and therefore load without a token.

Three datasets are the exception, because no redistributable copy exists and the loader must build them from source at load time too:

Dataset	Also needed when loading
FeTA24	SYNAPSE_TOKEN
SKM-TEA, ToothFairy2	access to their HuggingFace mirror — HF_TOKEN only if it is private for you

Reorientation

Ground truth is read in voxel space, so annotations must be computed against images that are already in RAS+. Two places can do that, and dataset_specs.py records which one each dataset uses:

• 8 datasets reorient inside their own download_raw.py (AFIDs, DEEP-PSMA, LIDC-IDRI, LNQ2023, MAMA-MIA, PDDCA, PI-CAI, VerSe) and are given no --reorient2RAS. It has to happen there because AFIDs, PDDCA and VerSe derive landmark voxel indices at download time, and the landmark-biometry planner cannot reorient at all.
• The rest are reoriented by the planner, so --reorient2RAS is passed to every step that accepts it. Repeating it is free: the reorienter early-returns on an already-RAS file and writes nothing.

Ceph-Biometrics-400 receives the flag nowhere, which is correct rather than special-cased — its only task is landmark biometry, whose planner has no such parameter, and its 2D X-rays sit on an identity affine already.

dataset_specs.py carries the full rationale as comments.

Tumour/lesion biometry: v1.1.0 and v1.1.1 are not reproduced here

For KiTS23, BraTS24, MSD, autoPET-III, HNTSMRG24 and KiPA22 the newest biometry annotation is 1.1.1, and it was not produced by preprocess_biometry.py. Its lineage is:

version	produced by
1.0.0	preprocess_biometry.py — what this driver runs
1.1.0	scripts/regenerate_tl_annotation_v1.1.0.py
1.1.1	scripts/regenerate_tl_annotation_v1.1.1.py + scripts/align_tl_split_to_v1.0.0.py

Those two regeneration scripts remain in scripts/ and are described in doc/release-v1.1.1.md. They are deliberately not wired into this driver: align_tl_split_to_v1.0.0.py rewrites already-published plans in place, which is the practice the annotation-identity policy exists to prevent, and it should stay a deliberate one-off rather than a routine step.

Limitations worth knowing

• Split parameters are fixed catalogue-wide, not per-dataset choices. Random seed 1024 and split ratio 0.7 are the default in all 72 preprocess_*.py, and the same two values are declared as RAMDOM_SEED and SPLIT_TRAIN_RATIO in MedVision.py. The 11 tumour/lesion biometry scripts likewise all default to --shrunken_bbox_scale 0.9 and --enlarged_bbox_scale 1.1. The driver overrides none of them, so a rebuild inherits exactly the constants the published data was built with — this is not an assumption to be recovered. Verified end to end: regenerating PI-CAI segmentation and detection into a scratch directory reproduces the published plans byte for byte, split membership and order included.
• --force_uint16_mask is not passed. Whether it was used per dataset is not recoverable; it is a mask-dtype hygiene step that does not change annotation values.
• Downloads are large. Several datasets are tens of gigabytes. --dry_run first.
• Existing plan files are never overwritten without --force.