EDM2 diffusion model synthesizes fetal ultrasound, boosting classification to 93.36%

Prenatal ultrasound is vital for fetal health assessment, but AI development is bottlenecked by scarce, privacy-restricted, and difficult-to-annotate datasets. To address this, the authors introduce a foundational synthetic imaging framework built on EDM2, a state-of-the-art diffusion architecture. The model is trained on multiple public datasets and generates 512×512 pixel images spanning six anatomical fetal planes. The approach delivers improved image quality over previous methods, reflected in lower FID scores, and when fine-tuned for downstream classification, an ensemble achieves 93.36% accuracy—outperforming models trained only on real images.

Despite these quantitative gains, the clinical reality check is sobering. A fetal ultrasound specialist with over 10 years of experience scored 100 generated images at a mean realism of 2.67/5, with real images rated significantly higher. Common artefacts include smoothing, speckle irregularities, and anatomical inconsistencies, which limit direct clinical deployment. Still, the authors have released code, models, and data to support further research, positioning this as a practical foundation for addressing data scarcity in medical imaging. The results highlight both the promise of generative models for privacy-preserving dataset augmentation and the gap still needed to close for realistic synthetic medical imagery.