PhD Thesis Defense: Adrien MEYER
Date and Time: Friday, 2 October 2026, 1:00 PM
Location: Salle Hygie, Institut Hospitalo-Universitaire (IHU) Strasbourg, 1 place de l'Hôpital (Nouvel Hôpital Civil)
The defense will be held in English.
Jury :
Director: Prof. Nicolas Padoy, ICube, Université de Strasbourg
Co-Director: the late Prof. Didier Mutter, HUS, Université de Strasbourg
Reviewers:
- Dr. Sandrine Voros, Université Grenoble Alpes
- Prof. Purang Abolmaesumi, The University of British Columbia
- Prof. Sophia Bano, University College London
Abstract:
Computer-aided ultrasound analysis could make examinations more consistent, accessible, and less operator-dependent. However, its development still relies on expert annotations that are costly, application-specific, and difficult to produce at scale. Sparse prompts offer a way to reduce dense labeling by letting clinicians provide targeted inputs, such as points, from which prompt-conditioned models can infer boxes or masks. Yet conventional annotation tools and prompts impose workflows that are model-driven and poorly aligned with clinical practice, contributing to expert fatigue and limiting sustained engagement. In this thesis, we co-design annotation strategies and prompt-conditioned models around sparse expert interactions that are both aligned with clinical practice and structured for model interpretation. We first study this idea in video annotation, where frame-by-frame labeling makes the process slow and discontinuous. We show that on-the-fly point-based interaction can make annotation more fluid while providing an effective signal for object detection. We then broaden this perspective to ultrasound representation learning, showing that heterogeneous open-access annotations can be leveraged despite the fragmentation of available datasets. Finally, by grounding prompt design in sonographic practice, we show that interactive segmentation can become more natural for clinicians and more informative for models. More broadly, this thesis shows that scalable ultrasound image analysis depends on annotation workflows that reflect how clinicians work while making their intent explicit to models.
