Fetal Ultrasound Plane Classification: A Cross-Domain Generalization Study for Low-Resource Clinical Settings

A research study applying deep learning to fetal ultrasound image classification, with a specific focus on measuring how well models generalize across different ultrasound machines and operators. This project is a hands-on precursor to my independent small-scale research interest in cross-domain generalization in deep learning-based fetal ultrasound image quality assessment.

Motivation

Most deep learning models for fetal ultrasound are trained and evaluated on data from a single hospital or machine, and often perform worse when applied to images from a different device, operator, or clinical setting. This project tests that idea directly: train classifiers on one subset of a public dataset, then explicitly measure how much accuracy drops when evaluating on a held-out subset that differs by acquisition machine, and compare whether pretraining changes how large that drop is.

Dataset

This project uses the FETAL_PLANES_DB dataset (Burgos-Artizzu et al., 2020), a public dataset of 12,400 fetal ultrasound images from 1,792 patients, collected at two hospitals in Barcelona using multiple ultrasound machines and operators.

• Source: https://zenodo.org/record/3904280
• The dataset includes a CSV of metadata (FETAL_PLANES_DB_data.csv) with columns including Image_name, Patient_num, Plane (the anatomical plane label), US_Machine, Operator, and a provided Train split.

Setup: download the dataset from Zenodo, unzip the images into data/Images/, and place the CSV at data/FETAL_PLANES_DB_data.csv. See data/README.md for exact expected structure.

I have not redistributed the dataset itself in this repository — only code that operates on it — in line with the original dataset's terms of use.

Method

1. Task: multi-class classification of the ultrasound Plane (e.g. abdomen, brain, femur, thorax, other).
2. Models: a ResNet18 backbone (pretrained on ImageNet), fine-tuned on fetal ultrasound images, compared against a simple from-scratch CNN baseline — to isolate how much pretraining contributes versus a basic architecture with no transfer learning.
3. Generalization evaluation (the core idea of this project):
   • In-distribution (ID) test set: held-out images from the same US_Machine values seen during training.
   • Out-of-distribution (OOD) test set: held-out images from a US_Machine (or Operator) not seen during training.
   • I report accuracy on both and compute the generalization gap (ID accuracy − OOD accuracy) as the key metric of interest, rather than just overall accuracy.

This ID/OOD split design is a small-scale version of the evaluation approach used in published work on fetal ultrasound generalization to low-resource settings (see Related Work), and reflects the kind of question I'd like to continue exploring in future research.

Related Work

This project builds on a growing body of work on fetal ultrasound deep learning. Burgos-Artizzu et al. (2020) introduced the FETAL_PLANES_DB dataset used here, evaluating CNN classifiers for standard-plane detection. A 2022 systematic review (arXiv:2201.12260) surveyed 145 papers on fetal ultrasound deep learning and identified generalization across imaging sites and devices as a persistent open problem. Sendra-Balcells et al. (2023) directly evaluated fetal ultrasound model generalization to low-resource settings across five African countries, using held-out-device evaluation similar in spirit to the ID/OOD design used here, though at larger scale and with real cross-country data rather than a single public dataset. This project is a small, self-contained test of the same underlying question — how much do pretraining and architecture choice affect generalization under domain shift — rather than a replication of either study.

Project structure

fetal-us-quality-gen/
├── README.md
├── requirements.txt
├── .gitignore
├── data/
│   └── README.md          # dataset download/setup instructions
├── src/
│   ├── dataset.py          # PyTorch Dataset + ID/OOD split logic
│   ├── model.py             # ResNet18 and simple CNN baseline
│   ├── train.py              # training loop
│   ├── evaluate.py           # ID vs OOD evaluation + generalization gap
│   └── utils.py               # seeding, transforms, plotting helpers
└── results/                # training logs, saved plots
    ├── best_model.pt        # ResNet18 checkpoint
    ├── generalization_gap.png
    └── simple_cnn/           # simple CNN baseline checkpoint + plot

How to run

pip install -r requirements.txt

# Train ResNet18 (pretrained)
python src/train.py --data_dir data/ --epochs 15 --model resnet18 --output_dir results/

# Train simple CNN baseline
python src/train.py --data_dir data/ --epochs 15 --model simple_cnn --output_dir results/simple_cnn/

# Evaluate either model's ID vs OOD generalization gap
python src/evaluate.py --data_dir data/ --checkpoint results/best_model.pt --output_dir results/
python src/evaluate.py --data_dir data/ --checkpoint results/simple_cnn/best_model.pt --output_dir results/simple_cnn/

Results

Both models were trained for 15 epochs, holding out one ultrasound machine entirely as the out-of-distribution (OOD) test set.

Model	ID Accuracy	OOD Accuracy	Generalization Gap
Simple CNN (from scratch)	87.2%	78.4%	8.9 points
ResNet18 (pretrained)	96.5%	94.7%	1.7 points

The pretrained ResNet18 not only achieved higher accuracy overall but showed a substantially smaller generalization gap than the from-scratch CNN (1.7 points vs. 8.9 points) — suggesting that ImageNet pretraining improves not just raw accuracy but specifically the model's robustness to domain shift across imaging devices. This is a more informative comparison than accuracy alone: a model could achieve high overall accuracy while still degrading sharply under distribution shift, which the CNN baseline illustrates directly.

Overall accuracy also masks per-class differences: for both models, the "Other" class showed noticeably weaker precision than the overall accuracy suggests, indicating that specific anatomical categories may be more vulnerable to domain shift than others even when aggregate accuracy looks strong.

This result is a useful but limited signal: FETAL_PLANES_DB was collected across only two hospitals, so the diversity of machines/settings is narrower than a true cross-country domain shift (e.g. testing on real low-resource clinical data, as in Sendra-Balcells et al., 2023). A small gap for the pretrained model here does not rule out much larger generalization failures under more extreme distribution shifts, which remains an open question worth exploring further with more diverse, real-world clinical data.

Limitations

• Single public dataset from one geographic region (Spain); the broader motivation behind this project — generalization to genuinely low-resource settings such as Bangladesh — would need real out-of-region clinical data, which this project does not include.
• Small-scale, single-run experiments; no hyperparameter search or statistical significance testing.
• Scoped as a focused study of one specific question (pretraining's effect on generalization), rather than a comprehensive evaluation across multiple architectures, datasets, or domain shift types.

Why this project

Access to expert ultrasound review is not evenly distributed — in some clinical settings, only one specialist may be available to interpret scans, which can mean longer waits and more room for inconsistent image quality to affect results. That real-world constraint is part of what drew me to this specific question, though this project itself is a small, exploratory step: it does not solve that problem, but it reflects the kind of technical question — can a model trained in one setting generalize reliably to another — that would need to be addressed before AI tools could responsibly assist in such settings.

References

Burgos-Artizzu, X.P., Coronado-Gutiérrez, D., Valenzuela-Alcaraz, B. et al. Evaluation of deep convolutional neural networks for automatic classification of common maternal fetal ultrasound planes. Sci Rep 10, 10200 (2020).

Sendra-Balcells, C. et al. Generalisability of fetal ultrasound deep learning models to low-resource imaging settings in five African countries. Sci Rep 13, 2728 (2023).

A Review on Deep-Learning Algorithms for Fetal Ultrasound-Image Analysis. arXiv:2201.12260 (2022).