From Privacy to Progress: xAILab at MICCAI 2025 in Daejeon, South Korea

10/07/2025

This September, our xAILab Bamberg had the pleasure of participating in the prestigious MICCAI 2025 (Medical Image Computing and Computer-Assisted Intervention) conference, hosted for the very first time in Daejeon, South Korea. Set in the heart of Korea’s high-tech research hub, this year’s MICCAI brought together a record-breaking 3,000 participants and over 1,000 accepted papers, marking a new milestone in the conference’s nearly three-decade history.

Under the dual themes “From Concept to Clinical” and “Pan-Asian Connections,” MICCAI 2025 celebrated innovation, collaboration, and inclusivity across the global medical imaging community and our lab was proud to contribute to this inspiring gathering.

Spotlight on Our Contribution

Our PhD candidate Francesco Di Salvo represented xAILab Bamberg with a poster presentation of the paper “Embedding-Based Federated Data Sharing via Differentially Private Conditional VAEs”, co-authored with Hanh Huyen My Nguyen and Prof. Dr. Christian Ledig.

In this work, the team tackles one of the central challenges in modern medical AI: how to enable collaborative model training across institutions while preserving patient privacy. Traditional Federated Learning (FL) approaches, while privacy-conscious, often face high communication costs and limited adaptability. To overcome these barriers, Francesco and colleagues proposed a Differentially Private Conditional Variational Autoencoder (DP-CVAE) framework that allows institutions to share privacy-preserving data embeddings instead of raw images or models.

By leveraging foundation model embeddings as compact, informative representations of medical images, the method drastically reduces redundancy and computation. The DP-CVAE collaboratively learns a global, privacy-aware data distribution, enabling downstream tasks to be trained efficiently and securely. Compared to standard FL classifiers, the approach demonstrated improved scalability and privacy guarantees while generating higher-fidelity embeddings with five times fewer parameters than existing DP generative models.

The poster drew strong interest from attendees, sparking discussions around privacy-preserving data sharing, generative modeling, and the evolving role of foundation models in medical imaging.

You can explore the full paper here.

Recognizing Excellence in Peer Review

We are also thrilled to celebrate our PhD student Sebastian Doerrich, who was recognized among the top 15 reviewers of MICCAI 2025 with an Outstanding Reviewer Award.
This distinction highlights Sebastian’s commitment to maintaining the high scientific standards that define the MICCAI community and reflects our lab’s broader dedication to fostering open, rigorous, and fair research practices.

Science, Collaboration, and Korean Hospitality

Beyond the poster sessions, Daejeon offered a perfect setting for scientific exchange and cultural discovery. Our team enjoyed reconnecting with colleagues and friends from around the world, discussing research during the day and sharing ideas (and local cuisine!) over dinner in the evenings. The conference venue, surrounded by Daejeon’s futuristic skyline and vibrant university district, captured the spirit of MICCAI’s 2025 themes: technological innovation with a human touch.

MICCAI 2025 also underscored the growing importance of clinical translation and regional collaboration, featuring four visionary keynotes that bridged neuroscience, medical robotics, responsible AI, and clinical radiology. These talks, together with over 100 workshops, challenges, and tutorials, made this year’s conference a true celebration of diversity in both ideas and people.

Looking Ahead

Our participation at MICCAI 2025 reaffirmed xAILab Bamberg’s mission to advance trustworthy, privacy-aware AI for medical imaging. From Francesco’s innovative contribution to Sebastian’s recognition as an outstanding reviewer, our lab continues to push the boundaries of what’s possible when technical excellence meets ethical responsibility.

We return from Daejeon inspired by the science, the culture, and the community and look forward to continuing our collaborative journey toward more inclusive and secure medical AI.