--- license: cc-by-nc-4.0 pretty_name: FluXray tags: [medical, x-ray, segmentation, synthetic] --- # FluXray: label-exact synthetic radiographs 138063 generatively edited digitally reconstructed radiographs with exact, overlapping segmentation masks for 63 structures, rendered from the 1597 annotated whole-body CTs of the MOOSE / ENHANCE-PET 1.6k dataset and restyled with a language-conditioned image editor (FLUX.2 klein), then quality-filtered (see the FleXray paper). Splits: {'train': 124198, 'val': 13865, 'test': 0}. ## Layout `FluXray/thunder_dbs/1.0/` is a ThunderDB (LMDB) that the `flexray` package reads directly; point `GENERATED_DATAPATH` at the root of the `flexray-data` download (the directory containing `FluXray/`): ```bash pip install flexray export GENERATED_DATAPATH=/path/to/flexray-data # training config # data: # Xray: # FluXray: {version: "1.0"} ``` | File | Contents | |---|---| | `data.mdb` | The LMDB database. One entry per sample keyed `_`, holding `img` (float16, 256 x 256, intensities in [0, 1]) and `seg` (uint8, 63 x 256 x 256 binary masks, one channel per structure; channels may overlap). Index entries: `_samples` (ordered keys), `_subjects`, `_splits` (`train`/`val`/`test` key lists), `_metadata` (per-sample pose, split, MOOSE subject, source site, license) and `_attrs` (build provenance, `label_names` in channel order). Read with `thunderpack.ThunderDB.open(path, "r")` or through `flexray`. | | `lock.mdb` | LMDB lock file; recreated automatically and safe to ignore. | | `protocol.yml` | The 63 label names in channel order (`label_names`), the mask thresholds used to binarize the rendered soft masks, and the build recipe. | | `samples.csv` | One row per sample: `sample_id`, MOOSE `subject_id`, `split`, `pose_id`, camera position `x`/`y`/`z` (mm) and rotation `alpha_deg`/`beta_deg`/`gamma_deg`, `visible_label_count`, `source_site`, `license`. | | `filter_kept.csv`, `filter_dropped.csv` | Quality-control scores for the renders that passed and failed the filter. `raw` and `flux` are the soft Dice of a reference segmentation model against the exact labels on the unedited and edited render; `soft_dice_delta_flux_minus_raw` is their difference. Per `pose_id`, `median_delta` and `q_upper_delta` give the median and 0.98 quantile of the delta, `lower_bound = 2 * median - q_upper`, and `keep = delta >= lower_bound`. Pose columns mirror `samples.csv`; `global_index` is build bookkeeping. | | `filter_thresholds.csv` | The per-pose filter statistics behind those decisions: delta mean/median/std/min/max before and after filtering, `n_before`, `n_after`, `n_dropped`, `frac_dropped` and the applied `lower_bound`. | | `../../README.md`, `../../LICENSE` | This card and the license summary (in `FluXray/`). | ## License CC BY-NC 4.0 for the collection as a whole. Each image inherits the license of its MOOSE source site (`license` column in `samples.csv` and per-sample metadata): - CC-BY-4.0: 50282 samples - CC-BY-NC-4.0: 87781 samples License text: https://creativecommons.org/licenses/by-nc/4.0/ (CC BY-NC 4.0) and https://creativecommons.org/licenses/by/4.0/ (CC BY 4.0). Changes from the source: CT volumes were projected to DRRs, restyled with FLUX.2 klein, quality-filtered, and paired with projected masks; no source CT or PET data is included. ## Attribution Please cite the FleXray paper and the ENHANCE.PET 1.6k dataset: Ferrara D. et al. (2026). Sharing a whole-/total-body [18F]FDG-PET/CT dataset with CT-derived segmentations: an ENHANCE.PET initiative. *Scientific Data*. ENHANCE.PET 1.6k - Whole-/Total-Body [18F]FDG-PET/CT with CT-Derived Segmentations was accessed on 2026-02-16 from https://registry.opendata.aws/enhance-pet-1-6k. The CC BY-NC 4.0 images derive from the AutoPET Challenge subset of ENHANCE.PET 1.6k, which additionally requires citing the source collection: Gatidis, S., Kuestner, T. (2022). A whole-body FDG-PET/CT dataset with manually annotated tumor lesions (FDG-PET-CT-Lesions) (Version 2) [Dataset]. The Cancer Imaging Archive. https://doi.org/10.7937/gkr0-xv29