kdjoumessi/Big-Data-Africa-School-2026

Big Data Africa School 2026 # Big-Data-Africa-School-2026 ## Projet Self-Explainable AI vs. Post-Hoc Explanations: Interpreting AI Decisions in Medical Image Classification. ## Abstract Convolutional Neural Networks (CNNs) have been widely adopted across various sectors, including healthcare, due to their remarkable ability to solve complex medical imaging tasks, often surpassing human performance. However, despite their impressive potential for medical image classification, CNN-based systems are often considered “black boxes,” as their decision-making processes remain opaque to users. This lack of transparency poses a significant barrier to their adoption in clinical settings, where interpretability is essential. Post-hoc explanation techniques, such as Grad-CAM and LIME, attempt to provide visual interpretations of AI decisions. However, these methods face several limitations, including inconsistency, unreliability, and sensitivity to small input perturbations. These flaws can result in differing explanations for similar inputs, ultimately undermining their trustworthiness and clinical applicability. Given the importance of transparency and interpretability in healthcare, there is a clear need for more robust and reliable explainability methods in AI-driven medical image analysis. As an alternative, inherently interpretable models, often referred to as white-box or self-explainable models, have been proposed. These models aim to embed interpretability directly into their architecture. However, they can be more challenging to implement and train compared to widely used black-box models. In this project, participants will train and evaluate white-box CNNs and compare their performance and interpretability to conventional black-box CNNs. Additionally, they will assess the quality of explanations provided by white-box models against popular post-hoc explanation techniques applied to their black-box counterparts, to determine which approach offers more reliable and clinically useful insights. # …