README.md

metadata

license: other
license_name: sliver07-research-only
license_link: https://sliver07.grand-challenge.org/Rules/
task_categories:
  - image-segmentation
tags:
  - medical
  - ct
  - liver
  - segmentation
  - sliver07
  - 3d
size_categories:
  - n<1K
configs:
  - config_name: preview
    data_files:
      - split: train
        path: preview/train-*
      - split: test
        path: preview/test-*
dataset_info:
  config_name: preview
  features:
    - name: patient_id
      dtype: string
    - name: official_split
      dtype: string
    - name: num_slices
      dtype: int32
    - name: image
      dtype: image
    - name: mask
      dtype: image
    - name: overlay
      dtype: image
  splits:
    - name: train
      num_bytes: 5705343
      num_examples: 20
    - name: test
      num_bytes: 2790343
      num_examples: 10
  download_size: 8499887
  dataset_size: 8495686

SLIVER07 — MICCAI 2007 Liver Segmentation Challenge (re-mirror)

Re-host of the SLIVER07 training + test releases from the Zenodo open-access mirror, restructured into the same dataset/case_XXXXX/ + train.jsonl layout we use for KiTS23 / KiPA22 / AbdomenCT1K so a single Base3DDataset subclass can load it.

Composition

Split	Cases	With mask
train	20	yes
test	10	no (GT held server-side at sliver07.grand-challenge.org)

case_00000..case_00019 are the 20 training volumes (liver-orig001..020 in the upstream naming) with paired ground-truth liver masks. case_00020..case_00029 are the 10 test volumes (liver-orig001..010) — the masks are withheld by the challenge organizers for online scoring. Use the train split for benchmarking.

File layout

dataset/case_00000/
  imaging.mhd          # MetaImage header (ElementDataFile = imaging.raw)
  imaging.raw          # binary CT volume
  segmentation.mhd     # MetaImage header (ElementDataFile = segmentation.raw)
  segmentation.raw     # binary 0/1 liver mask
...
dataset/case_00029/    # test cases have only imaging.{mhd,raw}
train.jsonl
test.jsonl
README.md

train.jsonl / test.jsonl list one entry per case with image, mask, label, modality, dataset, official_split, patient_id keys. Image/mask paths are prefixed with data/nii/SLIVER07/ so they slot directly into the EasyMedSeg Base3DDataset.HF_JSONL_PREFIX convention. mask is null for test entries.

Mask labels

CT integer labels:

Value	Class
0	background
1	liver

Single binary class — the official SLIVER07 GT is a single curated reference mask per volume (verified by a radiologist).

CT voxel intensity

Raw HU values are preserved (MET_SHORT element type for images). Per-volume spacing varies (0.5–5 mm slice spacing, 0.54–0.86 mm in-plane); read from each .mhd header rather than assuming a fixed spacing.

License

This mirror inherits the SLIVER07 challenge terms, which permit research use only and forbid commercial use or redistribution to non-registered parties. See https://sliver07.grand-challenge.org/Rules/ for the canonical license. The full upstream license.txt is reproduced at the repo root.

Cite the canonical paper:

@article{heimann2009comparison,
  title   = {Comparison and evaluation of methods for liver segmentation from CT datasets},
  author  = {Heimann, Tobias and van Ginneken, Bram and Styner, Martin A. and others},
  journal = {IEEE Transactions on Medical Imaging},
  volume  = {28},
  number  = {8},
  pages   = {1251--1265},
  year    = {2009},
  doi     = {10.1109/TMI.2009.2013851}
}