FETAL-GAUGE

A Benchmark for Assessing Vision-Language Models in Fetal Ultrasound

---

Overview

Fetal ultrasound is a cornerstone of prenatal care, yet its interpretation is highly operator-dependent and clinically challenging. Although Vision-Language Models (VLMs) have shown impressive results in natural images and other medical imaging domains, their ability to understand fetal ultrasound remains largely unexplored.

FETAL-GAUGE introduces the first large-scale benchmark designed to systematically evaluate vision-language models on clinically relevant fetal ultrasound tasks.

---

Key Features

• First benchmark for evaluating VLMs on fetal ultrasound
• Over 42,000 fetal ultrasound images
• Over 93,000 question–answer pairs
• Covers a wide range of clinically meaningful tasks
• Enables standardized and reproducible evaluation
• Highlights critical performance gaps in current VLMs

---

Benchmark Tasks

FETAL-GAUGE formulates all tasks as visual question-answering (VQA) problems to enable unified evaluation across models.

Task Category	Description
Plane Identification	Identification of standard anatomical planes
Anatomical Recognition	Recognition of fetal organs and structures
Visual Grounding	Localization of anatomical structures
Orientation Understanding	Inference of fetal orientation and position
View Conformity	Assessment of clinical acquisition standards

---

Benchmark Findings

We evaluate a diverse set of general-purpose and medical vision-language models using FETAL-GAUGE.

Key findings include:

• State-of-the-art VLMs struggle with fetal ultrasound understanding
• Best-performing models achieve only ~55% accuracy
• Significant weaknesses in:
  • Fine-grained anatomical reasoning
  • Visual grounding in ultrasound

These results highlight the need for domain-specific multimodal modeling.

---

Data will be published soon!

✒️ Citation

If you find our work useful, please consider giving our repo a star and citing our paper as follows:

@article{alasmawi2025fetal,
  title={FETAL-GAUGE: A Benchmark for Assessing Vision-Language Models in Fetal Ultrasound},
  author={Alasmawi, Hussain and Saeed, Numan and Yaqub, Mohammad},
  journal={arXiv preprint arXiv:2512.22278},
  year={2025}
}