SymmAdapt checkpoints (abdomen MRI → CT)

Weights for SymmAdapt: source-free domain adaptation of a SymmFlow (SymmFMSD) segmenter.

Code: github.com/tal-grossman/SymmAdapt_public

These files are the paper MRI→CT run (CHAOS MRI → BTCV CT, four organs: spleen, kidneys, liver).

File	What it is
source_mri_best.pt	Source model trained on labeled MRI (Phase I)
gst_mri_to_ct_best.pt	After generative self-training on unlabeled CT (Phase III)

Load them with the public repo (bash scripts/download_checkpoints.sh, then scripts/run_eval.sh). Do not expect a Hugging Face from_pretrained API — they are PyTorch trainer checkpoints (ema_state_dict).

Not included here:

• LatFM init: ocaetas/SymmFlow (LatFM_coco_beta6.0.pt)
• SD VAE: stabilityai/sd-vae-ft-mse (pulled at runtime)
• Images: DFG / BTCV / CHAOS

License

MIT

Citation

@misc{grossman2026symmadaptsymmetricalflowmatching,
      title={SymmAdapt: Symmetrical Flow Matching for Source-Free Domain Adaptation in Medical Image Segmentation},
      author={Tal Grossman and Noa Cahan and Hayit Greenspan},
      year={2026},
      eprint={2608.22532},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.22532},
}


Also cite SymmFlow if you use these weights:

@article{caetano2025symmetrical,
  title   = {Symmetrical Flow Matching: Unified Image Generation, Segmentation, and Classification with Score-Based Generative Models},
  author  = {Caetano, Francisco and Viviers, Christiaan and De With, Peter HN and van der Sommen, Fons},
  journal = {arXiv preprint arXiv:2506.10634},
  year    = {2025}
}