Advancing Fetal Ultrasound Image Quality Assessment in Low-Resource Settings

This repository contains code for assessing fetal ultrasound image quality using the ACOUSLIC-AI 2024 blind-sweep dataset.

Requirements

• Python 3.9 or higher

Install required Python packages:

pip install -r requirements.txt

1. Download Data

Option A: Download Raw Dataset

We use the ACOUSLIC-AI dataset. You can download it by running:

bash download_data.sh

After running the script, the data directory should have the following structure:

data/
└── acouslic-ai/
    ├── circumferences/
    │   └── fetal_abdominal_circumferences_per_sweep.csv
    ├── images/
    │   └── stacked_fetal_ultrasound/
    │       └── *.mha
    └── masks/
        └── stacked_fetal_abdomen/
            └── *.mha

Option B: Download Preprocessed Dataset (Recommended)

Downloading the raw dataset may take hours based on our experience. To save time, we provide a link to the preprocessed data:

Preprocessed Acouslic AI dataset

Once downloaded, place the zip file in the project’s root directory and run:

unzip acouslic-ai-train-set_preprocessed.zip

After running the script, the data directory should have the following structure:

data/
└── acouslic-ai/
    └── workshop/
        ├── train/
        │   └── *.npz
        ├── val/
        │   └── *.npz
        ├── test/
        │   └── *.npz
        └── meta_info.csv

2. Download Model Weights

Download FetalCLIP model weights:

• FetalCLIP weights

Place the weight under the project root directory.

3. Preprocessing

Skip this step if you use the preprocessed dataset.

To preprocess the raw data, including train, validation, and test splits as well as data augmentation, run:

python preprocess.py

After preprocessing, the data folder structure will be:

data/
└── acouslic-ai/
    ├── circumferences/
    │   └── fetal_abdominal_circumferences_per_sweep.csv
    ├── images/
    │   └── stacked_fetal_ultrasound/
    │       └── *.mha
    ├── masks/
    │   └── stacked_fetal_abdomen/
    │       └── *.mha
    └── workshop/
        ├── train/
        │   └── *.npz
        ├── val/
        │   └── *.npz
        ├── test/
        │   └── *.npz
        └── meta_info.csv

Each .npz file contains:

{
    "image": numpy.ndarray,
    "mask": numpy.ndarray
}

4. Reproduce Experiments

Classification

python main.py --config config/classification.yml

Modify the model_name field in the YAML config file to experiment with different models.

Segmentation

python main.py --config config/segmentation.yml

Results

Architecture	Models	Accuracy	F1 Score	Precision	Recall	# Trainable Parameters
CNN	DenseNet	0.9516	0.7024	0.7805	0.6420	7.0 M
	EfficientNet	0.9537	0.7253	0.7725	0.6855	4.0 M
	VGG	0.9510	0.7084	0.7580	0.6671	134 M
Transformer	Swin	0.9565	0.7429	0.7864	0.7113	1.7 M
	DEIT	0.9554	0.7466	0.7619	0.7363	2.4 M
	ViT400M	0.9560	0.7506	0.7657	0.7417	2.4 M
	FetalCLIPCLS	0.9575	0.7570	0.7782	0.7397	2.4 M

Model performance on fetal ultrasound image quality assessment (IQA). Metrics reported as mean over five runs. Best scores are bolded.

Related Articles

@misc{he2025advancingfetalultrasoundimage,
      title={Advancing Fetal Ultrasound Image Quality Assessment in Low-Resource Settings}, 
      author={Dongli He and Hu Wang and Mohammad Yaqub},
      year={2025},
      eprint={2507.22802},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.22802}, 
}

@misc{maani2025fetalclipvisuallanguagefoundationmodel,
      title={FetalCLIP: A Visual-Language Foundation Model for Fetal Ultrasound Image Analysis}, 
      author={Fadillah Maani and Numan Saeed and Tausifa Saleem and Zaid Farooq and Hussain Alasmawi and Werner Diehl and Ameera Mohammad and Gareth Waring and Saudabi Valappi and Leanne Bricker and Mohammad Yaqub},
      year={2025},
      eprint={2502.14807},
      archivePrefix={arXiv},
      primaryClass={eess.IV},
      url={https://arxiv.org/abs/2502.14807}, 
}