PGE-UNet

PGE-UNet

Official pretrained checkpoints for:

Integrating physics guidance into deep learning for cine cardiovascular segmentation

PGE-UNet is a physics-guided deep learning framework for cine cardiovascular magnetic resonance image segmentation. The framework combines a Noise-Aware Encoder, a Physics-Regularized Decoder, and a simulation-based pseudo-(B_1^+) prior.

Paper

• Journal: Computerized Medical Imaging and Graphics
• Volume: 134
• Article: 102798
• Year: 2026
• DOI: https://doi.org/10.1016/j.compmedimag.2026.102798
• Source code: https://github.com/QuocKhanhLuong/PGE-UNet

Available Checkpoints

The checkpoints contain PyTorch model state_dict objects. They must be used with the model implementation and preprocessing pipeline provided in the official GitHub repository.

Despite its filename, best_model_acdc_no_anatomical.pth is the official ACDC checkpoint used for the reported evaluation. The original filename is retained for reproducibility and traceability with the experimental setup.

Download

Install the Hugging Face Hub CLI:

pip install --upgrade huggingface_hub

Download all checkpoints:

hf download QuocKhanhLuong/PGE-UNet \
  best_model_acdc_no_anatomical.pth \
  best_model_mnm.pth \
  best_model_scd.pth \
  SHA256SUMS \
  --local-dir weights

Evaluation

Clone the official repository:

git clone https://github.com/QuocKhanhLuong/PGE-UNet.git
cd PGE-UNet

Evaluate ACDC:

python scripts/evaluate_acdc.py \
  --weights weights/best_model_acdc_no_anatomical.pth

Evaluate M&Ms:

python scripts/evaluate_mnm.py \
  --weights weights/best_model_mnm.pth

Evaluate SCD:

python scripts/evaluate_scd.py \
  --weights weights/best_model_scd.pth

Important Notes

• The models expect 2.5D inputs consisting of five adjacent cine CMR slices.
• The class ordering differs between ACDC and M&Ms.
• Input images must follow the preprocessing procedure provided in the official repository.
• These models are intended for research use only.
• The checkpoints have not been validated for clinical diagnosis or clinical decision-making.

Citation

@article{trannguyen2026pgeunet,
  title   = {Integrating physics guidance into deep learning for cine cardiovascular segmentation},
  author  = {Tran-Nguyen, My Anh and Luong, Quoc Khanh and Kha, Minh Bao and Dang, Tuan Linh},
  journal = {Computerized Medical Imaging and Graphics},
  volume  = {134},
  pages   = {102798},
  year    = {2026},
  doi     = {10.1016/j.compmedimag.2026.102798},
  url     = {https://doi.org/10.1016/j.compmedimag.2026.102798}
}

License

The source code and pretrained checkpoints are provided for research purposes only. For commercial use, please contact the authors.