Datasets:

hinge
/

BIC-MAC

license: cc-by-4.0
task_categories:
  - image-to-image
modality:
  - medical-imaging
tags:
  - PET
  - CT
  - MRI
  - attenuation-correction
  - medical-imaging
  - challenge
pretty_name: BIC-MAC Challenge Dataset
size_categories:
  - 100<n<1K

BIC-MAC Challenge Dataset

Dataset for the Big Cross-Modal Attenuation Correction (BIC-MAC) challenge — a medical imaging challenge where participants synthesize pseudo-CT images from non-attenuation-corrected PET (NAC-PET), whole-body MRI, and 2D topograms, enabling CT-less PET reconstruction.

🌐 Challenge website: bic-mac-challenge.github.io
💻 Code & documentation: github.com/bic-mac-challenge/challenge-codebase

Dataset

100 healthy volunteers acquired on a Siemens Biograph Vision Quadra (PET/CT) and MAGNETOM Vida (MRI).

Split	Subjects	Contents
`train/` (labeled)	8	features + CT labels + sinograms + PET labels
`train/` (unlabeled)	67	features + CT labels
`val/`	4	features + sinograms

Each subject directory contains:

features/ — NAC-PET, whole-body DIXON MRI (4 chunks × 2 phases + combined), 2D topogram, demographics
ct-label/ — ground-truth CT in HU, body/organ/face segmentations
recon/ — STIR sinogram files and reconstruction metadata (labeled train + val only)
pet-label/ — ground-truth attenuation-corrected PET, body/organ segmentations (labeled train only)

Download

Full dataset (~650 GB):

from huggingface_hub import snapshot_download

snapshot_download(repo_id="hinge/BIC-MAC", repo_type="dataset", local_dir="./bic-mac-data")

Without sinogram data (~35 GB):

The recon/ folders contain raw sinogram files and account for ~95% of the dataset size. If you only need the imaging data for model training (CT synthesis from NAC-PET/MRI), you can omit them:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="hinge/BIC-MAC",
    repo_type="dataset",
    local_dir="./bic-mac-data",
    ignore_patterns=["*/recon/*"],
)

The sinogram data is required to run the reconstruction pipeline and evaluate PET metrics locally. See the challenge codebase for details.