Selected Research
Theme 1: Fetal ultrasound FMs and clinical applications
I previously led an effort to build the to-date largest foundation model for fetal ultrasound trained entirely on Danish healthcare data (UltraDINO, MICCAI25). My ongoing work is focussed on the evaluation of UltraDINO in clinically motivated settings, under aspects such as robustness to domain shift and subgroup fairness (MICCAI26). We are actively working on clinical validation of UltraDINO-based applications through prospective validation and RCTs.
General Methods Make Great Domain-Specific Foundation Models: A Case-Study on Fetal Ultrasound
MICCAI 2025 · Jakob Ambsdorf*, Asbjørn Munk*, Sebastian Llambias, Anders N. Christensen, Kamil Mikolaj, Randall Balestriero, Martin G. Tolsgaard, Aasa Feragen, Mads Nielsen
Introduces UltraDINO and demonstrates the value of domain-specific pretraining by reaching state-of-the-art results across three public fetal ultrasound datasets.
Are Foundational Models Less Biased Than Specialized Models? An Ultrasound Study
MICCAI 2026 (to appear) · Joris Fournel, Jakob Ambsdorf, Benjamin Jønch Jürgensen, Christopher Boland, Aya Elgebaly, Kamil Mikolaj, Paraskevas Pegios, Emilie Pi Fogtmann Sejer, Martin Tolsgaard, Anders Nymark Christensen, Mads Nielsen, Aasa Feragen
Compares foundation models to task-specific supervised models for spontaneous preterm birth prediction and fetal weight estimation, highlighting that foundation models often inherit finetuning data biases, and demonstrating how task-aligned pretraining can improve both overall performance and subgroup fairness.
Theme 2: Brain MRI foundation models, datasets, and benchmarks
I have contributed to brain MRI foundation model research through methods (AMAES, ADSMI@MICCAI24), open datasets (FOMO300K, Scientific Data), and benchmarks (FOMO challenge, MICCAI25/26). I am a co-lead of the FOMO Foundation Model Challenge for Brain MRI at MICCAI 2026, which provides a benchmark for foundation models on real clinical data in challenging domain-shift settings.
AMAES: Augmented Masked Autoencoder Pretraining on Public Brain MRI Data for 3D-Native Segmentation
ADSMI at MICCAI 2024 · Jakob Ambsdorf*, Asbjørn Munk*, Sebastian Llambias, Mads Nielsen
Introduces an efficient pretraining framework for 3D U-Net-style models and the BRAINS-45K collection of 44,756 public brain MRI volumes.
Towards Brain MRI Foundation Models for the Clinic: Findings from the FOMO25 Challenge
arXiv preprint, 2026 · Asbjørn Munk, Stefano Cerri, Vardan Nersesjan, Christian Hedeager Krag, Jakob Ambsdorf, Pablo Rocamora García, Julia Machnio, Peirong Liu, …, Michael Eriksen Benros, Juan Eugenio Iglesias, Mads Nielsen
Summarizes the results of the FOMO26 challenge across 19 foundation models submitted by 16 teams, and studies how pretraining objectives, model scale, and training data affect generalization to clinical brain MRI tasks.
A Large-Scale Heterogeneous 3D Magnetic Resonance Brain Imaging Dataset for Self-Supervised Learning
Scientific Data, 2026 · Stefano Cerri*, Asbjørn Munk*, Sebastian Nørgaard Llambias, Jakob Ambsdorf, Julia Machnio, Vardan Nersesjan, Christian Hedeager Krag, Peirong Liu, Pablo Rocamora García, Mostafa Mehdipour Ghazi, Mikael Boesen, Michael Eriksen Benros, Juan Eugenio Iglesias, Mads Nielsen
Introduces FOMO300K, a large and heterogeneous public brain MRI collection for self-supervised learning.
FOMO26: Foundation Model Challenge for Brain MRI
MICCAI 2026 · Jakob Ambsdorf*, Stefano Cerri*, Asbjørn Munk*, Sebastian Nørgaard Llambias, Julia Machnio, Pablo Rocamora García, Zahra Sobhaninia, Alice Schiavone, …, Michael Eriksen Benros, Juan Eugenio Iglesias, Peirong Liu, Melanie Benjamin Ganz, Risheng Xu, Mads Nielsen
A community benchmark evaluating brain MRI foundation models across seven clinically motivated downstream tasks, using few-shot and out-of-domain evaluation settings.