README.md

metadata

license: cc-by-4.0
task_categories:
  - image-segmentation
tags:
  - medical
  - ct
  - liver
  - hepatocellular-carcinoma
  - tace
  - tumor-segmentation
  - dicom
  - tcia
pretty_name: HCC-TACE-Seg
size_categories:
  - n<1K
dataset_info:
  features:
    - name: patient_id
      dtype: string
    - name: study_uid
      dtype: string
    - name: ct_series_uid
      dtype: string
    - name: seg_series_uid
      dtype: string
    - name: num_ct_slices
      dtype: int32
    - name: slice_index
      dtype: int32
    - name: liver_voxels
      dtype: int64
    - name: mass_voxels
      dtype: int64
    - name: portal_vein_voxels
      dtype: int64
    - name: aorta_voxels
      dtype: int64
    - name: image
      dtype: image
    - name: mask
      dtype: image
    - name: overlay
      dtype: image
  splits:
    - name: preview
      num_bytes: 26847001
      num_examples: 104
  download_size: 26843221
  dataset_size: 26847001
configs:
  - config_name: default
    data_files:
      - split: preview
        path: data/preview-*

HCC-TACE-Seg

Multimodality annotated hepatocellular carcinoma data set including pre- and post-TACE multiphasic contrast-enhanced CT with voxel-level segmentations of liver, tumor mass, portal vein, and abdominal aorta.

Dataset Details

Field	Value
Modality	CT (multiphasic contrast-enhanced)
Body part	Liver (hepatocellular carcinoma + intrahepatic vessels + abdominal aorta)
Task	3D multi-class segmentation (4 foreground classes)
Patients	105 (68 M / 37 F)
Studies	~210 (pre-TACE + post-TACE per patient)
CT series	572
SEG series	105 (one STAPLE-fused consensus mask per patient on the pre-TACE study)
Images	51,968 DICOM slices
Format	DICOM (images) + DICOM SEG (segmentations)
License	CC BY 4.0

Each patient was scanned 1–12 weeks (avg 3 weeks) before their first TACE procedure (pre-TACE) and at follow-up (14 weeks post-TACE). Only the pre-TACE study is segmented. Three independent radiology residents semi-automatically segmented each case in AMIRA; senior body radiologist K.M.E. (20+ yrs experience) reviewed all annotations. The released mask is the STAPLE-fused consensus of the three rater segmentations.

Segmentation Labels

The DICOM SEG file for each patient contains four binary segments:

1. Liver — liver parenchyma
2. Mass — entire tumor (necrotic + viable HCC combined)
3. Portal vein — intrahepatic vessels
4. Abdominal aorta

Structure

images/<PatientID>/<StudyInstanceUID>/<SeriesInstanceUID>/*.dcm
segmentations/<PatientID>/<StudyInstanceUID>/<SeriesInstanceUID>/*.dcm
series_to_patient.json                              # series-level metadata index

PatientID ranges from HCC_001 to HCC_105. Segmentation DICOMs are DICOM SEG objects and reference their source CT series via the ReferencedSeriesSequence (top-level CT SeriesInstanceUID) and PerFrameFunctionalGroupsSequence → DerivationImageSequence → SourceImageSequence (per-frame source CT SOPInstanceUID), enabling loss-less alignment to the matched CT slice grid.

Source

• TCIA collection: https://www.cancerimagingarchive.net/collection/hcc-tace-seg/
• DOI: 10.7937/TCIA.5FNA-0924
• Released: 2022-08-17 (now fully public, no registration required)

Citation

@article{moawad2023hcctaceseg,
  author  = {Moawad, Ahmed W. and Fuentes, David and ElBanan, Mohamed G. and
             Shalaby, Ahmed S. and Guccione, Jeffrey and Kamel, Serageldin
             and Jensen, Corey T. and Elsayes, Khaled M.},
  title   = {Multimodality annotated hepatocellular carcinoma data set
             including pre- and post-TACE with imaging segmentation},
  journal = {Scientific Data},
  volume  = {10},
  number  = {33},
  year    = {2023},
  doi     = {10.1038/s41597-023-01928-3}
}

@misc{hcctaceseg2022tcia,
  author    = {Moawad, A. W. and Fuentes, D. and ElBanan, M. G. and
               Shalaby, A. S. and Guccione, J. and Kamel, S. and
               Jensen, C. T. and Elsayes, K. M.},
  title     = {HCC-TACE-Seg [Dataset]},
  year      = {2022},
  publisher = {The Cancer Imaging Archive},
  doi       = {10.7937/TCIA.5FNA-0924}
}