WAW-TACE — Warsaw TACE HCC Multiphase CT Dataset

A multiphase abdominal CT dataset of 233 treatment-naive hepatocellular carcinoma (HCC) patients receiving trans-arterial chemo-embolization (TACE) monotherapy at the Medical University of Warsaw, with 377 hand-crafted liver tumor masks plus auto-generated whole-abdomen organ masks, radiomics features, and clinical outcomes.

Dataset Summary

Field	Details
Modality	Multiphase abdominal CT (NIfTI)
Body Part	Liver (HCC tumors) + 104 abdominal organ VOIs
Subjects	233 treatment-naive HCC patients (TACE monotherapy)
CT volumes	854 (precontrast=200, arterial=230, portal_venous=231, delayed=193)
Tumor masks	378 hand-crafted, expert-validated (gold)
Organ masks	854 TotalSegmentator-generated, no manual review (silver)
License	CC-BY-4.0
Source	https://zenodo.org/records/12741586

Phases

CT and mask filenames encode the phase as an integer suffix:

ID	Phase
0	Precontrast / native
1	Late arterial
2	Portal venous
3	Delayed

Not every patient has every phase. See metadata/ct_hcc_metadata_v2.csv for per-patient phase availability.

Data Structure

WAW-TACE/
├── README.md
├── images/
│   └── {patient_id}_{phase}.nii.gz             # 854 CT volumes
├── tumor_masks/
│   └── {patient_id}_{phase}_{tumor_idx}.nrrd   # 378 hand-crafted tumor masks
├── organ_masks/
│   └── {patient_id}_{phase}.nii.gz             # 854 TotalSegmentator masks (104 VOIs)
└── metadata/
    ├── clinical_data_wawtace_v2_15_07_2024.xlsx
    ├── ct_hcc_metadata_v2.csv
    ├── radiomics_data_wawtace_09_05_2024.xlsx
    └── supplementary_table_s1_definitions_v2.xlsx

{tumor_idx} indexes multiple tumors annotated on the same CT (e.g. patient 102, phase 1 has six tumor masks: 102_1_0.nrrd … 102_1_5.nrrd). Each tumor mask is a binary 3D NRRD aligned to the corresponding images/{patient}_{phase}.nii.gz.

Mask Provenance — Recommended Ground Truth

Tumor masks (gold). TotalSegmentator produced initial tumor guesses; the primary radiologist (K.B., 6 yrs experience, 2 yrs in segmentation) manually corrected each in 3D Slicer; two senior radiologists (K.K. — 14 yrs, K.L. — 11 yrs in TACE/abdominal radiology) independently validated and modified the corrections; final 3 mm Gaussian smoothing was applied. Use these for tumor segmentation benchmarks.

Organ masks (silver). 104 abdominal VOIs (liver, spleen, kidneys, etc.) produced per CT phase by TotalSegmentator (nnU-Net pretrained). The original paper extracted radiomics features from these uncorrected masks. Treat as weak / auxiliary labels, not gold.

Splits

The released dataset has no official train/val/test split. All 233 patients form a single pool — define your own split downstream.

Citation

@article{bartnik2024wawtace,
  title   = {WAW-TACE: A Hepatocellular Carcinoma Multiphase CT Dataset
             with Segmentations, Radiomics Features, and Clinical Data},
  author  = {Bartnik, Krzysztof and Bartczak, Tomasz and Krzyzi{\'n}ski, Mateusz
             and Korzeniowski, Krzysztof and Lamparski, Krzysztof and W{\k{e}}grzyn, Piotr
             and Lam, Eric and Bartkowiak, Magdalena and Wr{\'o}blewski, Tadeusz
             and Mech, Krzysztof and Januszewicz, Magdalena and Biecek, Przemys{\l}aw},
  journal = {Radiology: Artificial Intelligence},
  year    = {2024},
  doi     = {10.1148/ryai.240296},
  pmid    = {39441110}
}

Sources

• Zenodo: https://zenodo.org/records/12741586
• Paper: https://pubs.rsna.org/doi/full/10.1148/ryai.240296
• PMC: https://pmc.ncbi.nlm.nih.gov/articles/PMC11605144/