Using deep learning in diagnostic decision-making for explainable image-based cancer screening

Oral cancer remains a major global health concern with high mortality rates due to late-stage diagnosis. Early detection is critical, yet traditional screening is often limited by resource constraints and the need for specialist expertise. This thesis investigates automated oral cancer screening using deep learning (DL) to develop accessible tools that utilize standard photographic images rather than specialized equipment. This research approaches screening through both performance-driven and human-centric perspectives. First, we develop DL-based methods for classification, lesion detection, and semantic segmentation, proving that pathological patterns can be identified in standard photographs. Second, we address the "black box" nature of AI by introducing a human-centered interpretability framework. We propose a novel IDL-CBR approach that integrates Informed Deep Learning (IDL) with Case-Based Reasoning (CBR) to align model outputs with clinical logic. The contributions of this work are fourfold: the design of an explainable screening system that prioritizes clinical trust; the development of robustness strategies to mitigate image noise and acquisition variability; a comprehensive benchmarking of state-of-the-art DL architectures for oral pathology; and the release of a curated, public dataset to foster collaborative research.