New AI model FetalMind boosts fetal ultrasound diagnosis with expert-guided clinical workflow

Fetal ultrasound interpretation presents unique challenges for medical AI systems—multiple viewing angles, complex disease presentations, and significant image variability across patients demand specialized approaches. Most existing medical vision-language models were designed around structured adult imaging and struggle to generalize to this critical domain.

Researchers have introduced FetalMind, a new medical AI system purpose-built for fetal ultrasound that combines automated report generation with diagnostic support. The work introduces Salient Epistemic Disentanglement (SED), a technique that embeds expert clinical knowledge directly into the model's architecture to reduce variability and improve reliability.

How the system works#

Rather than learning view-disease patterns passively, SED injects an expert-curated knowledge graph that explicitly maps relationships between ultrasound views and specific conditions. Reinforcement learning then steers the model toward clinically sound reasoning pathways, ensuring its outputs align with obstetric practice rather than merely matching statistical patterns in training data.

This architecture reduces learning bottlenecks and helps the model handle the heterogeneity present across different viewing angles and disease presentations—a major source of error in previous approaches.

Training at scale with real-world data#

To support development, the team curated FetalSigma-1M, the first large-scale corpus of fetal ultrasound reports spanning over 20,000 reports from twelve medical centers. This addresses a chronic scarcity of high-quality domain-specific training data that has limited prior research.

Results and clinical impact#

Evaluation across all gestational stages showed FetalMind outperformed both open-source and proprietary baseline models by an average of 14 percentage points. On critical conditions requiring immediate clinical attention, accuracy improved by 61.2%—a meaningful gain for high-stakes diagnostic scenarios. The system remained efficient, stable, and scalable throughout testing.