Prompt2Echo (P2Echo)

Text-Prompted Multi-View Echocardiography Segmentation with One Model

Official pretrained weights for P2Echo, a unified framework for text-prompted segmentation of multiple cardiac structures across standard echocardiographic views.

P2Echo uses structured, view-aware text prompts to control segmentation through a single model. Its Dynamic Image-Text Alignment (DyITA) module adaptively conditions decoder features on the requested view and anatomical structure.

Resources

• 📄 Paper: https://arxiv.org/abs/2611.13425
• 💻 Code: https://github.com/moeinheidari7829/P2Echo
• 🌐 Project Page: https://moeinheidari7829.github.io/p2echo/
• 🤗 Model Repository: https://huggingface.co/moein99/P2Echo

Pretrained Weights

Download P2Echo Checkpoint (Google Drive)

Download the checkpoint and place it at:

checkpoints/p2echo.pth

The model supports four echocardiographic views (2CH, 3CH, 4CH, PSAX) and five cardiac structures (LV, MYO, LA, RV, RA).

Installation

git clone https://github.com/moeinheidari7829/P2Echo.git
cd P2Echo
pip install -r requirements.txt

Evaluation

P2Echo is evaluated on an internal multi-view dataset and three held-out external datasets. Detailed quantitative comparisons, ablation studies, and implementation details are available in the paper.

Citation

@article{heidari2026p2echo,
  title={Prompt2Echo (P2Echo): Text-Prompted Multi-View Echocardiography Segmentation with One Model},
  author={Heidari, Moein and Kimbowa, Alvin and Tabatabaei, Mahdi and Bozorgpour, Afshin and Ghorbani Kolahi, Sina and Jiang, River and Chen, Wenjin and Foran, David J. and Grewal, Jasmine and Merhof, Dorit and Hacihaliloglu, Ilker},
  journal={arXiv preprint arXiv:2611.13425},
  year={2026}
}

Intended Use

This model is intended for research in medical image analysis and echocardiographic segmentation. It has not been validated for clinical decision-making.