SA-CycleGAN-2.5D

SA-CycleGAN-2.5D: Self-Attention CycleGAN with Tri-Planar Context for Multi-Site MRI Harmonization

Authors: Ishrith Gowda (UC Berkeley EECS), Chunwei Liu (Purdue University)

Paper: SA-CycleGAN-2.5D: Self-Attention CycleGAN with Tri-Planar Context for Multi-Site MRI Harmonization

---

overview

multi-site neuroimaging analysis is fundamentally confounded by scanner-induced covariate shifts, where the marginal distribution of voxel intensities P(x) varies non-linearly across acquisition protocols while the conditional anatomy P(y|x) remains constant. this is particularly detrimental to radiomic reproducibility, where acquisition variance often exceeds biological pathology variance.

SA-CycleGAN-2.5D is a domain adaptation framework that reduces Maximum Mean Discrepancy (MMD) by 99.1% (1.729 → 0.015) across two institutional MRI domains (BraTS and UPenn-GBM) without paired training data, collapsing a trained scanner domain classifier to near-chance accuracy (59.7%).

---

key results

---

architecture

SA-CycleGAN-2.5D integrates three architectural innovations:

1. 2.5D tri-planar manifold injection

• input: 12-channel volume (3 adjacent axial slices × 4 MRI modalities: T1, T1ce, T2, FLAIR)
• preserves through-plane gradients ∇z at O(HW) complexity, bridging 2D efficiency and 3D spatial consistency
• eliminates the slice-discontinuity artifacts common in 2D-only harmonization approaches

2. U-ResNet generator with dense voxel-to-voxel self-attention

• surpasses the O(√L) receptive field limit of standard CNNs
• models global scanner field biases — critical for field-strength-induced intensity shifts
• CBAM (Convolutional Block Attention Module) applied at resolution layers [3, 4, 5]
• 9 residual blocks, ngf=64

3. spectrally-normalized discriminator

• constrains Lipschitz constant K_D ≤ 1 for stable adversarial optimization
• prevents mode collapse without gradient penalty overhead

journal extension contributions (5 novel modules)

---

intended use

intended uses:

• multi-site MRI harmonization for clinical/research neuroimaging pipelines
• pre-processing step before radiomic feature extraction, segmentation, or classification
• privacy-preserving harmonization via the federated variant (no raw data sharing)
• bandwidth-constrained deployment via the compression variant

out-of-scope uses:

• modalities other than brain MRI (not validated)
• clinical diagnosis (research use only)
• sites with fewer than ~50 subjects per scanner type (insufficient domain coverage)

---

training details

---

usage

import torch
from neuroscope.models.generator import UResNetGenerator
from neuroscope.models.cyclegan import SACycleGAN25D

# load model
model = SACycleGAN25D.from_pretrained("ishrith-gowda/SA-CycleGAN-2.5D")
model.eval()

# input: [B, 12, H, W] — 3 adjacent axial slices × 4 modalities (T1, T1ce, T2, FLAIR)
# normalized to [-1, 1] per modality
input_25d = torch.randn(1, 12, 128, 128)  # replace with real volume

with torch.no_grad():
    harmonized = model.generator_A2B(input_25d)  # [B, 4, H, W] — center slice, all modalities

see the official repository for full preprocessing pipeline, dataset loading, and inference scripts.

---

data

trained on:

• BraTS-TCGA-GBM — The Cancer Imaging Archive, multi-institutional glioblastoma cohort
• UPenn-GBM — University of Pennsylvania GBM cohort (multiple scanner configurations)

see ishrith-gowda/MRI-Harmonization-BraTS-UPenn for the preprocessed dataset card.

---

ethical considerations

• data privacy: all training data is de-identified per TCIA data use agreements
• federated variant: enables cross-site harmonization with zero raw data transfer, directly addressing patient privacy in multi-institutional settings
• clinical use: this is a research tool. outputs should not be used for clinical diagnosis without further validation
• bias: evaluated exclusively on adult glioma patients. performance on pediatric, non-glioma, or non-brain MRI is unknown

---

citation

@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}
}

---

acknowledgments

data provided by The Cancer Imaging Archive (TCIA) under the BraTS-TCGA-GBM and UPenn-GBM collections. compute provided by Chameleon Cloud (NSF-funded research infrastructure). research conducted under the supervision of Dr. Chunwei Liu.