Research

Our Research

In recent years, foundation models have transformed the analysis of imaging data. Yet in biomedical settings, their adoption remains limited due to challenges such as data scarcity, sparse annotations, and inherent uncertainty, even among experts. Our group aims to develop novel and trustworthy AI methods that go beyond existing approaches, translating advances in machine learning into reliable tools for both biomedical research and clinical practice.

A key part of our research vision is self-supervised representation learning, exploring innovative strategies to adapt foundation models to biomedical images for segmentation and classification with minimal supervision. We also aim to develop uncertainty-aware AI models that explicitly account for annotation ambiguity, leading to more calibrated and safer predictions. More broadly, our group is committed to advancing machine learning approaches that tackle fundamental challenges in biomedical image analysis. Through these efforts, we strive to create AI that is not only accurate but also robust and trustworthy.

Our research projects

This section introduces current projects of the AI in Biomedical Imaging group:

Duration:

2025-2028
Funding source:

Industrial Fellowship from the Luxembourg National Research Fund (FNR)
Researchers:
Partners:

Becker Investment Research & Development (BIRD) S. A.
Description:

Having healthy teeth is not the same as having aligned teeth. While the public often equates dental health with an attractive smile, orthodontic specialist focuses on functional aspects (medical), i.e., how the dental arches fit together. Various methods exist for orthodontic assessment, including ICON, OGS, KIG, and the PAR Index. Among these, the PAR Index is widely trusted for providing a standardized measure of treatment quality by comparing pre- and post-treatment scores.

Advances in digital dentistry and AI have shown promise for improving orthodontic practice, yet automated PAR assessment from intra-oral images remains challenging and largely manual. This project hypothesizes that AI can enhance orthodontic planning by automatically categorizing patients according to the PAR Index using standardized intra-oral photographs. Such AI could improve efficiency, consistency, and precision, while also serving as a training tool for junior practitioners and a platform for self-improvement or knowledge sharing among orthodontists. We aim to leverage a uniquely large and diverse dataset of intra-oral images collected by BIRD S.A. across multiple clinical sites. This dataset surpasses any previously published collection used for orthodontic AI, offering a rich foundation to develop robust and clinically relevant models.
Project details (PDF):

Duration:
Funding source:

LCSB
Researchers:
Partners:

This project is a collaboration with two internal teams within LCSB: the Neuroinflammation group, which defined the research need, and the AI Modelling and Prediction group, whose PhD student carries out the implementation.
Description:

A major challenge in biomedical imaging is understanding how individual cells, molecular deposits, or other biological structures change or behave over time. This is especially difficult when objects change shape or when imaging data is affected by shifts in the camera or by deformable tissues.

This project explores the development of robust computational methods to track these dynamic objects in both 2D and 3D imaging data. By combining techniques that correct for movement and advanced algorithms that can handle morphological changes, we aim to maintain consistent identification of objects over time.

This research has broad potential applications. In neuroscience, tracking beta-amyloid plaques in live tissue could improve understanding of Alzheimer’s disease progression. In cell biology, following individual cells within 3D tissue cultures may reveal new insights into cellular behavior and proliferation. In medical imaging, monitoring tumor progression over time could provide valuable information on cancer development and treatment response.

Through this project, we aim to create tools that make long-term tracking of dynamic biological systems more accurate, efficient, and accessible enabling translational applications in biomedicine.
Project details (PDF):