AnyCXR: Human Anatomy Segmentation of Chest X-ray at Any Acquisition Position using Multi-stage Domain Randomized Synthetic Data with Imperfect Annotations and Conditional Joint Annotation Regularization Learning
Paper • 2512.17263 • Published
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This release contains CT-derived synthetic research data. Please respect the source license, non-commercial use and privacy conditions.
Log in or Sign Up to review the conditions and access this dataset content.
AnyCXR DRR training materials
Our publicly released AnyCXR training data: self-generated synthetic radiographs (DRRs) and overlapping anatomical supervision, with losslessly compressed arrays and original-resolution synthetic images. Code and project · Model weights · Paper
Contents
| Component | Contents |
|---|---|
ct_msdr/ |
31,565 CT-derived cases; 1,136,340 DRRs (9 appearance variants × AP/PA/LA/45° OB); paired 54-channel anatomy supervision, availability and QC arrays |
d3_original_synthetic.zip |
20,711 original-resolution synthetic images and available per-structure masks |
d3_frontal/ |
The existing frontal synthetic training cache: 5,961 geometry records, 17,028 image/flavor slots, 54-channel soft targets and availability |
d3_multiview/ |
10,385 paired lateral/oblique synthetic images with soft targets, availability and frozen source splits |
training_membership.json |
The current multiview training/development source-case membership; this is a selected subset of the full cache contents |
class_names.json, arrays.json, SHA256SUMS |
Ordered anatomy channels, shapes/views, exact byte sizes and compressed/uncompressed SHA-256 checksums |
The array payload is 82.42 GB compressed, restoring 203.57 GB of native arrays (decimal GB; metadata is additional). An additional 2.85 GB contains the original-resolution synthetic-only ZIP; the complete release is approximately 85.90 GB including metadata. These counts describe stored projections/cache records, not distinct patients; do not add cache slots to the primary corpus as independent new cases.
Download and unpack
After agreeing to the access conditions and logging in to Hugging Face:
python -m pip install huggingface_hub zstandard
hf download agaresd/anycxr-drr --repo-type dataset --local-dir ./anycxr-drr
cd anycxr-drr
sha256sum -c SHA256SUMS
python unpack.py .
unpack.py verifies every restored array against its original SHA-256. Keep enough space for both compressed and decompressed files. No raw CT volume or radiology report is included.
Array and geometry conventions
CT DRRs are uint8 arrays at 384 × 384; image shards have dimensions [case, 9 variants, 4 views, height, width]. Anatomical targets have 12 geometry slots (standard, portable 0, portable 1 for each view) and 54 independent foreground channels at 96 × 96. Masks are packed along width, using big-endian bit order; numpy.unpackbits(..., axis=-1, count=96, bitorder="big") restores them. Missing or rejected organs have availability zero; they are not background negatives. Quality-control arrays are retained without altering source masks.
ct_msdr/release_index.jsonl joins each source case to its image shard/local index and target shard/index. Shard roots in ct_msdr/shards.json are relative to ct_msdr/; resolve them against your download directory. Metadata with machine-local file paths has been removed or set to null. The stable release mapping, array order and source-case identifiers remain available.
The d3_frontal/ and d3_multiview/ targets are uint8 soft masks, divided by 255 when loaded. Their native source orientation is retained: the canonical-view training adapter flips their PA images and masks together. Do not apply that source-specific flip to the CT/MSDR arrays. Lateral and oblique names are nominal acquisition buckets; source groups can mix physical left/right directions, so they do not establish calibrated continuous azimuths.
The frontal cache stores more image/flavor slots than geometry records. Follow each CSV record's base_image_index, flavor_start_index, flavor_count and mask_index; a flavor slot is not an additional patient. training_membership.json records the selected multiview subset; the source cache's full train partition is not an instruction to enlarge a frozen experiment. Development sets have been used for model selection and must not be described as untouched clinical tests.
License and citation
Released under CC BY-NC-SA 4.0. See LICENSE for terms and required source attribution. Please respect the privacy conditions and do not attempt re-identification.
The original-resolution synthetic archive and the training caches overlap; they must not be counted as independent additional patients. Compression is lossless and preserves the released masks.
Please cite AnyCXR, arXiv:2512.17263. Required attribution for the underlying CT source is provided in LICENSE.
- Downloads last month
- 22