From Promise to Utility of Medical Image Foundation Models: Methods, Evaluation, and Fetal Ultrasound Applications
PhD defence by Jakob Ambsdorf
Assessment Committee
Professor Erik B Dam, DIKU, University of Copenhagen (Chairperson)
Professor Julia Schnabel, TUM
Professor and Dean of the Medical School Wiro Niessen, University of Groningen
Supervisors
Professor Mads Nielsen
Professor Aasa Feragen
Professor Martin Grønnebæk Tolsgaard
Department
Department of Computer Science
Place
Building: Statens Naturhistoriske Museum, Room: Auditoriet (AUD), Sølvgade 83, 1307 København K
Email address to gain access to the thesis: jakob.ambsdorf@gmail.com.
You will either receive a copy of the thesis or be informed where you can read a physical copy.
Recipients of copies of the thesis are not allowed to share or distribute it due to copyright compliance.
Short description of the thesis
Medical image analysis faces two persistent bottlenecks: expert-labeled data are costly and scarce, while models often fail to generalize across scanners, populations, and acquisition settings. Foundation models promise to address both by learning reusable representations from large-scale unlabeled data. But how much of this promise is realized in clinically relevant settings?
This thesis introduces two domain-specific foundation models: AMAES for 3D brain MRI and UltraDINO for 2D fetal ultrasound. It first examines how pretraining methods from general computer vision transfer to specialized medical imaging domains.
It then demonstrates foundation model utility beyond benchmarks by evaluating UltraDINO for fetal aneuploidy detection, growth estimation, and spontaneous preterm birth prediction. Together, these studies examine when the promise holds in clinically motivated tasks, what adaptations are needed, and where limitations remain.