DiFUSAL: Diffusion-Based Fetal Ultrasound Synthesis with Active Learning

Maryam Arjemandi, Salma Hassan, Hu Wang, Saudabi Valappi, Mohammad Yaqub

Abstract: The availability of high-quality fetal ultrasound data is vastly scarce, constrained by privacy concerns and the rarity of publicly accessible datasets. In addition, ultrasound is noisy, operator-dependent, and sensitive to patients’ variability. These limitations hinder the development of robust diagnostic models in maternal-fetal healthcare. We introduce DiFUSAL, a generative framework that integrates diffusion models with self-guided active learning for the automated synthesis of fetal ultrasound images. DiFUSAL estimates key biometric features such as gestational age and uses them to generate structured clinical-report-style prompts. These prompts are used to condition the diffusion model, enabling targeted and realistic image generation aligned with real-world diagnostic scenarios. This approach removes the need for human feedback or manual fine-tuning, making the process scalable and adaptable. Our method achieves an average LPIPS of 0.3660 ± 0.0442, outperforming standard baselines in perceptual similarity. We further demonstrate the utility of the generated data in two downstream tasks: fetal plane classification and gestational age prediction. DiFUSAL enhances model generalization and performance across both tasks. By generating clinically meaningful synthetic data, DiFUSAL contributes to automated interpretation, ultrasound training, point-of-care support, and foundation model development. Code and data will be made available at GitHub.

External IDs:dblp:conf/miccai/ArjemandiHWVY25