Towards Trustworthy AI in Healthcare: Transparent and Interpretable Models for Medical Imaging

This PhD thesis focuses on creating trustworthy Artificial Intelligence (AI) systems for healthcare, specifically in medical imaging for breast cancer classification. The research addresses the need for transparent and interpretable models by exploring Explainable AI (XAI) and its relationship with causality, utilizing Deep Learning (DL) techniques and emphasizing user-centric design through clinician feedback. The thesis investigates ante-hoc explainability with ProtoPNet on mammography images, introduces an end-to-end explainable AI framework for Digital Breast Tomosynthesis (DBT) images, and proposes a novel explainable-by-design Neural Network (NN) architecture using hybrid prototypes for improved interpretability and accuracy on DBT data, ultimately aiming to enhance the reliability of AI in healthcare and contribute to better clinical outcome.