Public Thesis Defense of Maxence Wynen- ICTEAM
sst |
Automating the integration of advanced MRI biomarkers into multiple sclerosis assessment through instance segmentation and machine learning
Friday April 17th, 2026 - 4:30pm - Auditorium BARB94 - Place Ste-Barbe 1, 1348 Louvain-la-Neuve
Multiple sclerosis (MS) is a chronic neurological disease for which magnetic resonance imaging (MRI) plays a central role in diagnosis, monitoring, and research. Lesion-level MRI biomarkers in particular provide essential diagnostic and prognostic information, yet their assessment largely relies on manual or visual interpretation, which is time-consuming, subjective, and poorly scalable. These limitations motivate the development of automated decision-support tools capable of extracting and analyzing lesion-level biomarkers in a reliable and reproducible manner. While numerous automated methods for lesion segmentation have been proposed, most derive individual lesions through connected components post-pro-cessingof semantic segmentation outputs. This approach, however, fails in the presence of confluent lesions, defined as aggregates of pathologically distinct lesions sharing adjacent voxels. As a result, lesion-level analyses become ill-defined and biased. This thesis addresses the challenge of defining, evaluating, and implementing robust automated lesion-level MRI analysis in the presence of confluent lesions, and examines its impact on downstream clinical applications.
The work first establishes an analytical foundation through a critical review of automated methods targeting advanced lesion-level biomarkers and existing lesion instance segmentation strategies. This analysis identifies a key gap in the literature, namely the systematic neglect of confluent lesions, which limits the validity and scalability of automated pipelines. To overcome this limitation, the thesis introduces formal operational definitions of confluent lesions and proposes a confluence-aware evaluation framework for lesion instance segmentation. Within this framework, existing methods are benchmarked, revealing systematic biases that affect lesion detection, delineation, and counting. Building on these methodological contributions, a confluence-aware deep learning framework, ConfLUNet, is developed and shown to improve lesion instance segmentation performance in the presence of confluent lesions.
Finally, the thesis demonstrates the clinical relevance of accurate lesion-level analysis by integrating advanced MRI biomarkers into machine learning models for MS diagnosis, achieving improved diagnostic performance compared to conventional imaging criteria. Overall, this work establishes methodological foundations for reliable automated lesion-level MRI analysis in multiple sclerosis and highlights its importance for enabling robust, interpretable, and clinically meaningful biomarker-driven applications.
Jury members
Prof. Benoît MACQ (UCLouvain), Supervisor
Prof. Meritxell BACH CUADRA (Université de Lausanne) Supervisor
Prof. Laurent FRANCIS (UCLouvain) Chairperson
Prof. Pietro Maggi (UCLouvain) Secretary
Dr. Laurence Dricot (UCLouvain)
Prof. Saïd Mahmoudi (UMons)
Prof. Michel Dojat (Université Grenoble-Alpes)