Investigating the Role of Explainable Artificial Intelligence in Biomedical Image Classification, Localization, and Security

This PhD thesis investigates the role of eXplainable Artificial Intelligence (XAI) and Convolutional Neural Networks (CNNs) in classifying and localizing various diseases within biomedical imaging. The primary goal is to address the "black box" nature of deep learning models by employing Class Activation Mapping (CAM) algorithms, such as Grad-CAM and Score-CAM, to provide visual heatmaps that justify AI predictions. This transparency is essential for building clinical trust and transforming AI into a reliable "co-pilot" for medical professionals. The research spans three imaging categories: Skin Surface, Internal Organs (e.g., MRI, RX), and Cellular/Tissue Imaging. Additionally, the thesis examines Adversarial Machine Learning to ensure system resilience against malicious attacks and extends XAI methodologies to audio signal classification, including COVID-19 detection from cough recordings. Ultimately, this work establishes a foundation for trustworthy and secure automated diagnostic systems that prioritize both high accuracy and patient safety.