STAffEcho

Official implementation of Anatomy-Guided Spatiotemporal Affinity Learning for Unsupervised Domain Adaptation in Echocardiography Segmentation.

The paper is published in IEEE Journal of Biomedical and Health Informatics.

• IEEE Xplore: https://ieeexplore.ieee.org/document/11531132
• DOI: https://doi.org/10.1109/JBHI.2026.3694986

---

Project Structure

.
|-- train.py                         # Training entry point
|-- context_cluster.py               # Anatomical context mapping strategy determination
|-- infer_geometrics.py              # Inference for geometric metrics
|-- infer_clinical_metrics.py        # Inference for clinical metrics
|-- geom_stats_wilcoxon.py           # Statistical significance analysis for geometric metrics
|-- clinical_stats_bootstrap_ci.py   # Statistical significance analysis for clinical metrics
|-- geom_aggregate_metrics.py        # Aggregation of geometric metrics from multiple runs
|-- clinical_aggregate_metrics.py    # Aggregation of clinical metrics from multiple runs
|-- config/
|   |-- STAffEcho/                   # STAffEcho experiment settings
|   |   |-- cm2cd.json               # CAMUS -> CardiacUDA
|   |   |-- cd2cm.json               # CardiacUDA -> CAMUS
|   |   `-- cd2echo.json             # CardiacUDA -> EchoNet-Dynamic
|   |-- UpperBound/                  # Supervised upper-bound settings
|   `-- woDA/                        # Source-only settings without domain adaptation
|-- datasets/                        # Dataset loaders
|-- models/                          # Network and affinity modules
|-- utils/                           # Losses, schedulers, logging, and helpers
`-- requirements.txt

---

Requirements

Create a Python environment and install the dependencies:

conda create -n STAffEcho python=3.8
conda activate STAffEcho
pip install -r requirements.txt

The listed PyTorch package targets CUDA 11.8. If your CUDA version differs, install the matching PyTorch build from the official PyTorch instructions, then install the remaining dependencies.

---

Dataset

The code supports the following datasets:

• CAMUS
• EchoNet-Dynamic
• CardiacUDA

Before running experiments, update dataset roots and output directories in the JSON files under config/.

Example:

"source": {
  "dataset": "Camus",
  "config": {
    "root": "/path/to/CAMUS"
  }
}

Recommended practice is to replace all local absolute paths with your own dataset, checkpoint, log, and result directories.

---

Training

Training is implemented in train.py. Experiment settings are defined by JSON files under config/.

Before running training, update the basedir variable in train.py to the configuration folder you want to use, for example:

basedir = "config/STAffEcho"

Then run:

python train.py

By default, train.py iterates over all configuration files in basedir and trains with the seeds defined in the script.

---

Evaluation and Analysis

Update checkpoint, dataset, and output paths before running evaluation or analysis scripts.

python infer_geometrics.py
python infer_clinical_metrics.py
python geom_aggregate_metrics.py
python clinical_aggregate_metrics.py
python geom_stats_wilcoxon.py
python clinical_stats_bootstrap_ci.py

---

Acknowledgements

This codebase was developed with reference to GraphEcho and Pixel-Embedded-Affinity.

We thank the authors for making their code and ideas publicly available.

---

Citation

If you find this repository useful, please cite:

@ARTICLE{11531132,
  author={Fang, Xinyan and Ma, Xinghua and Liu, Yang and Zhang, Jiahao and Wang, Xiaodi and Li, Xiangyu and Kow, Alfred Wei Chieh and Luo, Gongning and Wang, Guohua},
  journal={IEEE Journal of Biomedical and Health Informatics},
  title={Anatomy-Guided Spatiotemporal Affinity Learning for Unsupervised Domain Adaptation in Echocardiography Segmentation},
  year={2026},
  volume={},
  number={},
  pages={1-13},
  keywords={Echocardiography;Modeling;Videos;Context;Labeling;Computers;Conferences;Pixel;Training;Modules (abstract algebra);Echocardiography;Unsupervised domain adaptation;Affinity;Segmentation},
  doi={10.1109/JBHI.2026.3694986}
}

Please replace volume, number, and pages with the final IEEE Xplore metadata if IEEE has assigned them.