ishrith-gowda/SA-CycleGAN-2.5D-demo

SA-CycleGAN-2.5D: Multi-Site MRI Harmonization Demo

Interactive demo for SA-CycleGAN-2.5D, a self-attention CycleGAN with tri-planar 2.5D context for multi-site brain MRI harmonization.

What This Does

Multi-site MRI data suffers from scanner-induced domain shifts that confound downstream analysis. This model translates brain MRI volumes between institutional scanning protocols (BraTS ↔ UPenn-GBM) while preserving anatomical structure.

Architecture highlights:

• —2.5D input: 3 adjacent axial slices × 4 MRI modalities (T1, T1ce, T2, FLAIR) = 12-channel input
• —Self-attention bottleneck: captures long-range spatial dependencies for anatomical consistency
• —CBAM attention: channel and spatial attention in encoder/decoder paths
• —Multi-scale discriminator: spectral-normalized PatchGAN at 2 scales

Usage

1. 1.Upload a grayscale brain MRI slice (PNG/JPG) or use the provided examples
2. 2.Select harmonization direction (BraTS → UPenn-GBM or UPenn-GBM → BraTS)
3. 3.Click Submit to run inference

The uploaded image is replicated across 4 modality channels and 3 slice positions to form the 12-channel input expected by the model. For research use, see the full pipeline in the GitHub repository.

Links

• —Paper: arXiv:2603.17219
• —Code: github.com/ishrith-gowda/SA-CycleGAN-2.5D
• —Model: huggingface.co/ishrith-gowda/SA-CycleGAN-2.5D
• —Dataset: huggingface.co/datasets/ishrith-gowda/MRI-Harmonization-BraTS-UPenn

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},
  doi={10.48550/arXiv.2603.17219}
}