Overview

Knee Osteoarthritis (OA) is one of the leading causes of chronic pain and disability worldwide. However, current clinical stratification relies primarily on radiographic grading scales that correlate poorly with patient-reported pain and functional outcomes. This structure–symptom discordance suggests that meaningful patient subtypes exist beyond what radiographic grading can capture.Self-supervised contrastive learning offers a principled approach to learning rich representations from medical images without relying on inadequate labels.

This thesis aims to develop a self-supervised learning framework that combines contrastive pre-training on knee MRI with cross-modal knowledge distillation from paired clinical data. The goal is to discover candidate imaging biomarkers for knee OA and evaluate their representation quality, prognostic value, and clinical interpretability.

This project is a collaboration between the Chair of AI in Medicine and the Department of Orthopedics and Sports Orthopedics at the TUM University Hospital.

Objectives

• Pre-train a deep image encoder on knee MRI using a contrastive self-supervised objective to establish a strong baseline representation without annotations.
• Design and implement a cross-modal knowledge distillation framework in which a clinical encoder supervises the alignment of the image encoder with tabular data such as pain scores, functional assessments, and demographics.
• Investigate the effect of negative-free contrastive losses and hard-negative mining strategies on representation quality.
• Evaluate the computational validity of the learned biomarkers, their prognostic value over 10 years, and the clinical interpretability of the learned clusters.

(Tasks can be adapted based on student interests)

Prerequisites

• Advanced knowledge of deep learning with imaging data.
• Strong Python and PyTorch/TensorFlow skills.
• Experience with medical data and self-supervised learning is a plus, but not required.

What We Offer

• High-performance computing resources.
• A collaborative and interdisciplinary research group consisting of computer scientists and clinicians.
• Workspace at the Department of Orthopedics and Sports Orthopedics at the TUM University Hospital.
• Potential for scientific publications.

References

1. Obaido, George, et al. "A Systematic Review of Contrastive Learning in Medical AI: Foundations, Biomedical Modalities, and Future Directions." Bioengineering (2026).
2. Hager, Paul, Martin J. Menten, and Daniel Rueckert. "Best of Both Worlds: Multimodal Contrastive Learning with Tabular and Imaging Data." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.
3. Holland, Robbie, et al. "Deep Learning–Based Clustering of OCT Images for Biomarker Discovery in Age-Related Macular Degeneration (PINNACLE Study Report 4)." Ophthalmology Science 4.6 (2024): 100543.
4. Namiri, Nikan K., et al. "Deep Learning for Large Scale MRI-Based Morphological Phenotyping of Osteoarthritis." Scientific Reports 11.1 (2021): 10915.
5. Rengaraj Rajamohan, Haresh, et al. "Self-Supervised Learning for Knee Osteoarthritis: Diagnostic Limitations and Prognostic Value of Hospital Data." arXiv e-prints (2026): arXiv-2603.

How to Apply

Preferred starting date: July 2026 (flexible). Send an email to amna.gillani@tum.de and christina.valle@tum.de with your CV and transcript