Skip to content

CVPR 2025 BiomedSegFM, 3-D validation set: 2 144 CT / MRI / PET / 3-D ultrasound / microscopy volumes as npz

pl.load("hf_biomedsegfm_3d_val") loads this dataset. The validation volumes of the CVPR 2025 Foundation Models for Interactive 3D Biomedical Image Segmentation challenge (Ma et al.), pooled from 30+ public datasets (AMOS, AbdomenAtlas, TotalSegmentator, BraTS, ...). Each case is an .npz with 'imgs' (the volume), 'gts' (the label map) and 'spacing'; the spacing key cannot be carried by the layout, so the grids load with unit spacing (read s.path('image') with numpy for the real spacing). The field modality is CT, MRI, PET, US3D or Microscopy.

import pardesline as pl

ds = pl.load("hf_biomedsegfm_3d_val")
print(ds.describe())  # roles, frames, units, labels, licence, citation
s = ds[0]
obj = s.image  # fetched now, cached under ~/.cache/pardesline/datasets/hf_biomedsegfm_3d_val

Provenance and licence
Source Hugging Face Hub repo junma/CVPR-BiomedSegFM, paths 3D_val_npz
Scheme hf
Size of the whole source 19.5 GB
Licence CC-BY-NC-SA-4.0
Commercial use not allowed (non-commercial licence)
Redistribution allowed
Licence note CC BY-NC-SA 4.0 on the Hub card (checked 2026-10-14); each pooled source dataset keeps its own terms, TO CONFIRM per case.
Terms https://huggingface.co/datasets/junma/CVPR-BiomedSegFM
Homepage https://www.codabench.org/competitions/5263/
Cite Ma J., et al. CVPR 2025 Challenge: Foundation Models for Interactive 3D Biomedical Image Segmentation. 2025.
Samples grouped by id

Roles (what each sample contains)

Role Type Frame Units Labels / layout Pattern or file
image volume from the file header from the file header layout: (Z,Y,X) 3D_val_npz/{modality}_{id}.npz
The volume ('imgs' array; spacing in the 'spacing' key of the same file)
mask labels from the file header from the file header layout: (Z,Y,X) 3D_val_npz/{modality}_{id}.npz
Label map ('gts' array); class ids per source dataset

Standardized layout

Files are fetched on first access into the dataset cache and stay in their original format; ds.export(folder) writes every sample as standard files (PLY for points and meshes, MetaImage for volumes and label maps) with a .pardesline.json sidecar carrying frame, units, labels, licence and provenance, plus a manifest for the folder. Label maps become exact surfaces with to_mesh(label=k), DICOM series become one volume, raw arrays are decoded through the layout above.

ds = pl.load("hf_biomedsegfm_3d_val").take(2)
ds.export("./standardized")        # PLY / MetaImage + sidecars + manifest.json

Algorithms and models that run on it

Generated from src/pardesline/data/manifests/hf_biomedsegfm_3d_val.json on 2026-10-14.