From Noise to Clarity: Generating Realistic Fetal Ultrasound Images with Diffusion Models
Arjemandi, Maryam
Arjemandi, Maryam
Author
Supervisor
Department
Computer Vision
Embargo End Date
2026-05-30
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Ultrasound imaging is a critical tool in maternal-fetal healthcare, aiding in fetal development assessment and anomaly detection. However, the effectiveness of deep learning mod els in fetal ultrasound analysis is often hindered by the scarcity of highquality datasets. The challenges of patient variability, operator dependency, and ethical constraints further exacerbate this issue, limiting the development of robust AI-based diagnostic tools. Existing generative models attempt to bridge this gap, but they frequently require manual intervention and finetuning to ensure anatomical accuracy and clinical relevance. While some recent approaches aim to automate the generative process, such as self-supervised learning and active learning techniques, they often struggle with generating anatomically diverse images or maintaining clinical fidelity across gestational stages. These limitations highlight the need for a more adaptive and autonomous framework capable of continu ously improving its outputs without human oversight. To address these challenges, this work presents a novel generative framework integrating diffusion-based synthesis and self supervised learning to enhance fetal ultrasound image generation. A key component of this work is DiFUSAL: Diffusion Based Fetal Ultrasound Synthesis with Active Learning, a self-improving diffusion model that iteratively refines its output through an active learning paradigm. Unlike prior models, DiFUSAL autonomously evaluates and enhances its synthetic images, ensuring both anatomical consistency and gestational age variations, ultimately enabling the creation of a diverse and scalable synthetic dataset. We evaluate the clinical utility of our approach by assessing the impact of DiFUSAL generated images on two key downstream tasks: fetal plane classification and gestational age estimation. Extensive experiments demonstrate that training with synthetic data en hances classification accuracy and prediction robustness, reducing performance disparities caused by limited realworld datasets. DiFUSAL achieves a superior perceptual similar ity score (LPIPS: 0.3660 ± 0.0442), outperforming conventional generative models and reinforcing the reliability of synthetic fetal ultrasound imaging. Beyond synthetic image generation, this thesis explores complementary techniques such as domain adaptation and uncertainty-aware modeling to further improve model generalization. The findings con tribute to advancing deep learining fetal imaging by providing scalable, privacy-preserving, and clinically applicable solutions. By enabling highquality data synthesis, this research lays the foundation for broader adoption of AI in maternal-fetal healthcare and beyond.
Citation
Maryam Arjemandi, “From Noise to Clarity: Generating Realistic Fetal Ultrasound Images with Diffusion Models,” Master of Science thesis, Computer Vision, MBZUAI, 2025.
Source
Conference
Keywords
Fetal Ultrasound, Gestational Age, Denoising Diffusion Probabilistic Model, Stable Diffusion, Learned Perceptual Image Patch Similarity, Variational Autoencoder