USFetal: Tools for Fetal Brain Ultrasound Compounding

---

Overview

USFetal is an open-source toolbox designed for fetal brain ultrasound compounding. It integrates a variety of techniques to combine multiple 3D ultrasound volumes of the fetal brain into a single, high-quality volumetric representation.

The toolbox is designed to support research, evaluation, and benchmarking of different compounding strategies under a unified and configurable framework.

Supported Techniques

Traditional Methods

• Multi-scale (e.g., Difference of Gaussians – DoG)
• Transformation-based (e.g., Principal Component Analysis – PCA, Independent Component Analysis – ICA)
• Variational (e.g., optimization-based methods and Total Variation (TV) regularization)

Deep Learning Methods

• Self-Supervised Learning (SSL) – Scan-specific training approach that leverages complementary information across views.
• Plug-and-Play (PnP) – A training-free framework that integrates learned priors into the compounding process.

---

Installation

Follow the steps below to set up the environment and run the toolbox.
Tested on Ubuntu 22.04, Python 3.9, and CUDA 12.4.

1) Clone the repository

git clone https://github.com/mkhateri/USFetal-Compounding-Tools.git
cd USFetal-Compounding-Tools
cd USFetal

2) Create and activate the Conda environment

conda env create -f environment.yml
conda activate USFetal

3) Explicit PYTHONPATH step

export PYTHONPATH=$(pwd):$PYTHONPATH

---

Data Structure

Each subject folder inside the dataset (e.g., data/) contains multiple 3D ultrasound volumes and their corresponding (optional) binary brain masks. Each volume–mask pair (e.g., view1) corresponds to a distinct acquisition view of the same subject.

data/
├── subject_01/{volumes/, masks/}
│   ├── volumes: view1.nii.gz, view2.nii.gz, …, viewN.nii.gz
│   └── masks:   view1.nii.gz, view2.nii.gz, …, viewN.nii.gz
├── subject_02/{volumes/, masks/}
│   ├── volumes: view1.nii.gz, view2.nii.gz, …, viewM.nii.gz
│   └── masks:   view1.nii.gz, view2.nii.gz, …, viewM.nii.gz
└── ...

---

Usage

Each subdirectory provides scripts and configuration files for running a specific framework:

USFetal/
  ├── configs/              # Configuration files
  ├── evaluation/           # Image quality assessment
  ├── traditional/          # Classical compounding methods
  ├── ssl_scan_specific/    # Scan-specific self-supervised learning framework
  └── pnp/                  # Plug-and-play compounding

Users can either:

-Edit configuration files in configs/, or

-Override parameters directly from the command line.

Traditional Methods

Run classical compounding methods using a configuration file:

python traditional/main.py \
            --methods pca:standard ica:patchwise variational:dog

Override data paths and key parameters from the command line:

python traditional/main.py \
  --methods pca:patchwise \
  --data_parent ./data \
  --output_dir ./output_traditional \
  --patch_size 96 96 96 \
  --overlap 48 48 48

Self-Supervised Learning (SSL)

The SSL framework supports scan-specific self-supervised training and inference.

python ssl_scan_specific/scripts/train.py 

During training:

• The full configuration is saved to the experiment directory

• The best, last, and interval-based model checkpoints are stored automatically

Run inference

python ssl_scan_specific/scripts/inference.py 

During inference:

• The trained model checkpoint is loaded automatically

• Model architecture and parameters are recovered from the training experiment

Plug-and-Play (PnP) Framework

Before running the PnP framework, download the pretrained denoiser models from the model zoo:

python pnp/main_download_pretrained_model.py

This script automatically downloads and prepares the required pretrained denoisers. You can then use the default configuration from configs/config_pnp.py:

python pnp/main.py

Or specify paths and parameters directly from the command line:

python pnp/pnp.py \
  --data_parent ./data \
  --output_dir ./output_pnp \
  --iter_num 10

Image Quality Assessment

The evaluation module computes quantitative metrics (e.g., SSIM, PSNR, CC, MI, Entropy) for both the input views and the compounded volumes.

It accepts the path to:

• The parent data folder containing input views and masks

• The output directory containing compounded results

python evaluation/iq_assess.py \
  --parent_data <parent_data> \
  --output_dir <output_dir>

The output_dir format is as follows:

output/
├── subject_01/
│   ├── subject_01_pnp.nii.gz
│   ├── subject_01_pca.nii.gz
│   ├── subject_01_ssl.nii.gz
│   └── ...
├── subject_02/
│   ├── subject_02_pnp.nii.gz
│   ├── subject_02_pca.nii.gz
│   ├── subject_02_ssl.nii.gz
│   └── ...
└── ...

Results are automatically saved under <output_dir> with .csv, .tex, and .txt formats.

---

Citation

🙏 If you find this toolbox useful in your work, please consider starring the repo ⭐ and citing our paper:

@article{khateri2026usfetal,
  title={USFetal: Tools for Fetal Brain Ultrasound Compounding},
  author={Khateri, Mohammad and Ghahremani, Morteza and Valencia, Sergio and Jaimes, Camilo and Sierra, Alejandra and Tohka, Jussi and Grant, P Ellen and Karimi, Davood},
  journal={arXiv preprint arXiv:2601.06726},
  year={2026}
}

---

Acknowledgments

The PnP compounding implementation adapted from Deep Plug-and-Play Image Restoration by Zhang et al. We gratefully acknowledge the authors and the DPIR repository.