MRI-Harmonization-BraTS-UPenn

preprocessed multi-site brain MRI dataset used to train and evaluate SA-CycleGAN-2.5D, a self-attention GAN for unpaired cross-scanner MRI harmonization.

paper: SA-CycleGAN-2.5D: Self-Attention CycleGAN with Tri-Planar Context for Multi-Site MRI Harmonization
model: ishrith-gowda/SA-CycleGAN-2.5D

---

dataset overview

property	value
subjects	654 glioma patients
total 2D slices (triplets)	~52,000
modalities	T1, T1ce, T2, FLAIR (4 per subject)
input format	12-channel 2.5D triplets [3 slices × 4 modalities]
image size	128 × 128 px (center-cropped)
domains	2 institutional sites (BraTS, UPenn-GBM)
annotation	unpaired — no cross-site slice correspondence

---

source collections

domain A: BraTS-TCGA-GBM

• source: The Cancer Imaging Archive — TCGA-GBM
• subjects: 88 glioblastoma patients
• scanners: multi-institutional, heterogeneous (1.5T and 3T, multiple vendors)
• citation: Bakas et al., 2017; Menze et al., 2015
• access: public, TCIA data use agreement

domain B: UPenn-GBM

• source: The Cancer Imaging Archive — UPenn-GBM
• subjects: 566 glioblastoma patients
• scanners: 3T Siemens TrioTim (primary), 3T other, 1.5T
• citation: Bakas et al., 2022
• access: public, TCIA data use agreement

---

multi-domain scanner split

for the multi-domain AdaIN journal extension, UPenn-GBM subjects are sub-divided by scanner:

domain id	domain name	subjects	slices (~)
0	brats (multi-institutional)	88	6,538
1	upenn_3t_triotim	434	32,266
2	upenn_3t_other	65	4,863
3	upenn_15t	67	4,990
total		654	~48,657

---

preprocessing pipeline

all volumes preprocessed using the following pipeline (see neuroscope/preprocessing/ in the GitHub repo):

1. skull stripping — FSL BET or pre-stripped (BraTS volumes are pre-stripped)
2. co-registration — T1ce as reference; T1, T2, FLAIR rigidly registered
3. MNI152 atlas registration — affine registration to 1mm isotropic MNI152 space
4. intensity normalization — z-score per modality per subject, clipped at ±3σ, rescaled to [-1, 1]
5. slice extraction — axial slices, excluding top/bottom 10% (non-brain), minimum foreground threshold
6. 2.5D triplet construction — for each valid center slice idx, stack [idx-1, idx, idx+1] across all 4 modalities → [12, H, W] tensor
7. center crop — 128 × 128 px

---

data splits

split	subjects (A)	subjects (B)	slices (~)
train	70	453	41,870
val	9	57	5,230
test	9	56	5,230

splits are subject-level (no slice from a test subject appears in train/val). splits stored in data/metadata/domain_split.json and data/metadata/multi_domain_split.json in the repository.

---

format

subject_dir/
├── t1.nii.gz          # T1-weighted
├── t1ce.nii.gz        # T1 post-contrast
├── t2.nii.gz          # T2-weighted
└── flair.nii.gz       # T2-FLAIR

preprocessed volumes: float32 NIfTI, shape [H, W, D], intensity range [-1, 1].

---

usage

from neuroscope.data.dataset import MRIHarmonizationDataset

dataset = MRIHarmonizationDataset(
    brats_dir="path/to/preprocessed/brats",
    upenn_dir="path/to/preprocessed/upenn",
    split="train",
    image_size=128,
)

sample = dataset[0]
# sample["input"]     — torch.Tensor [12, 128, 128], 2.5D triplet
# sample["target"]    — torch.Tensor [4, 128, 128], center slice
# sample["domain"]    — str, "A" or "B"

see the official repository for full preprocessing scripts and dataloader.

---

access & license

the raw source data is publicly available from TCIA under the TCIA Data Usage Policy. users must agree to the TCIA terms before using this data.

this preprocessed dataset card and the SA-CycleGAN-2.5D model are released under CC BY-NC-ND 4.0 for research use only.

---

citation

if you use this dataset or the associated model, please cite:

@article{gowda2026sacyclegan25d,
  title={SA-CycleGAN-2.5D: Self-Attention CycleGAN with Tri-Planar Context for Multi-Site MRI Harmonization},
  author={Gowda, Ishrith and Liu, Chunwei},
  journal={arXiv preprint arXiv:2603.17219},
  year={2026},
  url={https://arxiv.org/abs/2603.17219},
  doi={10.48550/arXiv.2603.17219}
}

also cite the original TCIA source collections:

@article{bakas2017advancing,
  title={Advancing the cancer genome atlas glioma MRI collections with expert segmentation labels and radiomic features},
  author={Bakas, Spyridon and others},
  journal={Scientific data},
  volume={4},
  number={1},
  pages={1--13},
  year={2017}
}

@article{bakas2022university,
  title={The university of pennsylvania glioblastoma (UPenn-GBM) cohort: advanced MRI, clinical, genomics, \& radiomics},
  author={Bakas, Spyridon and others},
  journal={Scientific data},
  volume={9},
  number={1},
  pages={453},
  year={2022}
}