ULS23 — Radboudumc Bone & Pancreas (novel subsets)

⚠️ Partial mirror. This repository contains only the two novel, Radboudumc-original subsets of the ULS23 (Universal Lesion Segmentation, 2023) challenge — Radboudumc Bone and Radboudumc Pancreas — as paired CT image + binary segmentation mask VOIs. It is not the full ULS23 aggregator.

What this is / is not

ULS23 is an aggregator that bundles many pre-existing public datasets (DeepLesion, NIH-LN, KiTS21, LiTS, LIDC-IDRI, MSD Tasks 06/07/10, CCC18, LNDb) together with three novel Radboudumc-collected subsets (DeepLesion3D, Radboudumc Bone, Radboudumc Pancreas).

This mirror includes only:

Subset	Lesion VOIs (image+mask)	Series
Radboudumc Bone	697	151
Radboudumc Pancreas	120	119
Total	817	270

All other ULS23 components are deliberately excluded:

• DeepLesion / DeepLesion3D images inherit NIH Clinical Center terms (no explicit public-redistribution grant).
• NIH-LN, CCC18, LIDC-IDRI inherit TCIA terms; LNDb explicitly forbids repackaging.
• KiTS21, LiTS, MSD (06/07/10) carry their own challenge licenses and overlap datasets already present elsewhere in this suite (KiTS23, MSD, LIDC) — an evaluation-leakage risk.

The Radboudumc Bone & Pancreas subsets are net-new Radboudumc clinical data with no patient overlap with any public dataset.

Format

• NIfTI .nii.gz, volume-of-interest (VOI) crops of 256 × 256 × 128 voxels centered on each lesion, at the scan's native spacing (no resampling).
• Images carry a trailing singleton axis → shape (256, 256, 128, 1); squeeze to 3D before use. A small number of masks are stored as 3D (256, 256, 128); both forms squeeze to the same shape and share the image affine (verified: 0 affine mismatches).
• Masks are binary {0, 1}, one target lesion per VOI (sparse — foreground ≈ 0.1–0.2 % of voxels).
• CT intensities are stored in (approximately) Hounsfield units; VOI edge padding can yield values outside the usual HU range — apply a CT window before analysis.

Layout

Radboudumc_Bone/images/<lesion_id>.nii.gz      Radboudumc_Bone/masks/<lesion_id>.nii.gz
Radboudumc_Pancreas/images/<lesion_id>.nii.gz  Radboudumc_Pancreas/masks/<lesion_id>.nii.gz
metadata.csv     # lesion_id, subset, lesion_type, series_id, lesion_index, modality, image, mask

lesion_id (e.g. bone_00002_lesion_01, diag_pancreas_0004_lesion_01) preserves the original ULS23 Radboudumc identifiers; series_id is the per-patient/series group — use it for patient-disjoint splits (multiple lesions can come from one series).

Ground truth

Lesions were identified from clinical radiology reports and segmented in 3D by an experienced radiologist — the single-expert gold standard ULS23 designates for its novel data.

Provenance & counts

• Images: Zenodo ULS23 Part 1 (record 10035161, DOI 10.5281/zenodo.10035161), novel_data/ULS23_Radboudumc_{Bone,Pancreas}/images.zip.
• Masks: official GitHub repo DIAGNijmegen/ULS23, annotations/ULS23/novel_data/ULS23_Radboudumc_{Bone,Pancreas}/labels/ (LFS).
• Zenodo ships 744 bone + 124 pancreas images; the public masks number 697 + 120. The 51 images with no public mask (held-out evaluation lesions) are excluded from this mirror.

Overlap with other datasets

None for these subsets — net-new Radboudumc data. (The excluded ULS23 components overlap KiTS23 / MSD / LIDC and are omitted partly for that reason.)

License

CC BY-NC-SA 4.0 — the Radboudumc-original subsets as released on Zenodo. Non-commercial use only; attribution and ShareAlike required.

Citation

M. J. J. de Grauw et al., "ULS23: Universal Lesion Segmentation in computed tomography," Medical Image Analysis 102:103525 (2025). arXiv:2406.05231. Challenge: https://uls23.grand-challenge.org/ · Data: Zenodo 10.5281/zenodo.10035160 · Code/annotations: https://github.com/DIAGNijmegen/ULS23.