Researchers from the CHUV Department of Radiology, the Faculty of Biology and Medicine at UNIL, and members of the CIBM Center for Biomedical Imaging have won three artificial intelligence competitions in medical imaging at the MICCAI 2025 international conference in Daejeon, South Korea.
This recognition highlights the teams’ expertise in developing reliable, accessible, and clinically useful AI tools. Their work advances research on AI learning methods, an essential step toward trustworthy, accessible, and clinically relevant imaging technologies.
Today, in medical imaging, AI models still require large amounts of high-quality data, for instance, to accurately predict the location of a brain tumor or to assist doctors in quickly choosing the best treatment for a stroke. Traditional learning methods rely on examples manually annotated by expert radiologists, a demanding, time-consuming, and costly process.
A promising new approach, self-supervised learning (SSL), now paves the way for models that can exploit vast, heterogeneous medical image datasets without initially requiring expert annotations. Conceptually, this is the same principle behind large language models such as ChatGPT.
Building on this approach, the CHUV–CIBM–UNIL research teams demonstrated the effectiveness of self-supervised learning (SSL) and synthetic data generation in competitions held during the MICCAI 2025 international conference in South Korea.
AI Innovations for the medical imaging of tomorrow
At the Foundation Model Challenge for Brain MRI (FOMO2025), focused on optimizing self-supervised learning for multi-organ MRI analysis—the FOMO2JOMO team won 1st place in the Method track and 2nd place in the Open track.
The winning solution combines advanced representation learning with causal conditioning mechanisms, integrating information from brain structure volumes and patients’ clinical status. This approach enhances model generalization across scanners and imaging protocols.
In the Self-Supervised Learning for 3D Medical Imaging Challenge (SSL3D), the team won 1st place for developing a method that effectively extracts 3D anatomical representations from unlabeled MRI data. This allows AI models to learn from large, heterogeneous datasets while maintaining accuracy on new images.
The Low-field Pediatric Brain Magnetic Resonance Image Segmentation and Quality Assurance Challenge 2025 (LISA 2025) posed an additional challenge—segmenting brain structures in very low-field pediatric MRI, where signal quality and resolution are limited.
The goal of LISA is to develop and evaluate AI algorithms that can accurately analyze such images, a crucial step toward making MRI more affordable and accessible worldwide.
In this challenge, the University of Lausanne team won 1st place for hippocampus segmentation and 3rd place for basal ganglia segmentation, using an innovative approach based on synthetic and simulated MRI data to enhance performance under complex imaging conditions.
“Competing with participants from top universities around the world and industry, these are remarkable results in the fast-moving field of AI for medicine, which demonstrates the cutting-edge skills of our teams, and the strategic importance of public science funding in Switzerland”, said Dr Jonas Richiardi, Head of the CIBM Data Science CHUV-HUG Imaging for Precision Medicine Section & Director of the Translational Machine Learning Laboratory, Department of Medical Radiology, CHUV
These innovative approaches mark a major milestone toward the future of medical imaging. They pave the way for faster diagnostics, personalized treatments, and broader access to high-quality imaging, even in regions with limited resources.
Winning teams
FOMO2025 and SSL3D Challenges
Dr. Pedro M. Gordaliza¹², Dr. Jaume Banus Cobo³, Nataliia Molchanova²⁵, Maxence Wynen⁶, Benoit Gérin⁶, Prof. Meritxell Bach Cuadra¹², Dr. Jonas Richiardi³⁴
LISA 2025 Challenge
Vladyslav Zalevskyi¹², Dr. Thomas Sanchez¹², Busra Bulut¹², Prof. Meritxell Bach Cuadra¹²
Affiliations
¹ CIBM Center for Biomedical Imaging – Signal Processing CHUV-UNIL, Trustworthy Medical Image Analysis Section
² Medical Image Analysis Laboratory, Faculty of Biology and Medicine, UNIL
³ Translational Machine Learning Laboratory, Department of Medical Radiology – CHUV
⁴ CIBM Center for Biomedical Imaging – Data Science CHUV-HUG, Imaging for Precision Medicine Section
⁵ University of Applied Sciences and Arts Western Switzerland (HES-SO)
⁶ Université Catholique de Louvain
The winning teams express their gratitude to the challenge organizers and funding bodies, including the Swiss National Science Foundation (SNSF) and the Responsible AI program of the Hasler Stiftung.