PhD Position (E13, 100%, 3+ years)

PhD Position: Representation Learning & Next-Generation Medical Foundation Models (E13, 100%, 3+ Years)

30.08.2026, Academic staff

For our AI-Assisted Healthcare Lab at TUM School of Medicine and Health, we are seeking an outstanding PhD student (f/m/d) to advance the theoretical and methodological foundations of medical deep learning. The project focuses on novel self-supervised objectives, information geometry, mitigating representation bias for rare pathological findings, and building next-generation multimodal foundation architectures. Full-time, TV-L E13, 100%, fixed-term for 3+ years.

For our AI-Assisted Healthcare Lab at TUM School of Medicine and Health, we are seeking an outstanding PhD student (f/m/d) to advance the theoretical and methodological foundations of medical deep learning and develop next-generation foundation model architectures for precision medicine.

Self-supervised objectives like masked autoencoders and joint-embedding predictive architectures bias representations toward frequent patterns. Rare but informative signals get suppressed because the training loss does not reward encoding them. In medicine, rare findings are often the most important, as a small change can be the difference between normal and pathological. We want to understand this bias and correct it.

This is a fully funded position (TV-L E13, 100%) for a PhD student with a strong background in mathematics, computer science, or machine learning. The work has a strong focus on developing new objectives or new architectures for medical deep learning and on new ways to measure what representations contain or miss. The position is embedded in a well-funded, multi-year research project providing a unique multimodal and longitudinal setting for this research.

Research Directions

• Self-supervised and predictive representation learning
• Information content and geometry of learned representations
• Multimodal representation learning and information decomposition
• Learning with missing, partial, and irregular observations
• Temporal and longitudinal representation learning
• The role of objectives, architectures, and context in shaping representations
• New learning objectives and architectures for foundation models

The work will combine methodological development with large-scale experiments, aiming for contributions at leading machine learning and computer vision venues such as NeurIPS, ICML, ICLR, CVPR, and MICCAI, as well as high-impact medical journals.

Research Environment

The PhD will be embedded at TUM University Hospital, offering direct access to large-scale multimodal clinical datasets, high-end GPU and storage infrastructure, international research collaborations, and dedicated funding for international conference participation.

Supervision

The PhD candidate will be supervised by:

• Dr. rer. nat. Cosmin Bercea
  Senior Researcher in Generative AI and Medical Imaging, TUM University Hospital
  Google Scholar Profile →
• Prof. Dr. med. Keno Bressem
  Radiologist & Medical AI Lead, TUM University Hospital
  Google Scholar Profile →
• Prof. Dr. med. Lisa Adams
  Professor of Radiology & Deputy Director of Radiology, TUM University Hospital
  Google Scholar Profile →

Your Profile

We are looking for candidates with:

• A strong BSc and MSc degree in computer science, machine learning, mathematics, physics, engineering, or a related field
• Strong foundations in machine learning and deep learning
• Experience with PyTorch or similar frameworks
• Interest in representation learning, self-supervised learning, multimodal learning, foundation models, or ML theory
• Strong programming and analytical skills
• Excellent written and spoken English

Application

Please submit your application including a CV, cover letter, and complete academic transcripts (Bachelor’s and Master’s) via email directly to:

Kontakt: keno.bressem@tum.de, lisa.adams@tum.de, cosmin.bercea@tum.de

More Information

https://radiologie.mri.tum.de/en/ai-assisted-healthcare