Coalitional Active Learning in Deep Learning for Clinical Decisions, Addressing Challenges in Data Annotation, Malicious Annotator Detection, Coalition Stability, and Model Performance

Details

Supervisors

Macq, Benoît ; Legay, Axel

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en cybersécurité, à finalité spécialisée: conception et analyse de systèmes

Abstract

Deep learning models play a crucial role in clinical decision-making, heavily relying on hospital data and expert annotations. However, existing models often remain static and fail to adapt to the evolving nature of medical practice. In this work, we focus on a Coa- litional Active Learning (CAL) framework a concept that involves leveraging annotations from coalitions of healthcare professionals rather than relying solely on individual institu- tions. Although CAL as a framework has been proposed in the literature, my contribution lies in enhancing its application to clinical decision-making by specifically addressing the challenges associated with annotation quality. In this context, I propose a methodology within the CAL framework that identifies and mitigates the influence of spurious and malicious annotators, who can degrade model performance and destabilize the coalition. By incorporating reinforcement learning techniques, I simulate and analyze the behavior of these annotators, enabling their detection and reducing their potential adverse effects. This emphasis on annotation quality not only ensures the integrity and reliability of clinical decision-making processes but also fosters more effective collaboration among healthcare professionals.