Efficient detection & classification of digital histopathology imagery-applications in medical diagnostics for oncology

Abstract

Lung cancer remains one of the most prevalent and lethal forms of cancer worldwide, necessitating accurate and timely diagnosis to guide effective therapeutic decisions. Among existing diagnostic modalities, histopathological whole-slide images (WSIs) stained with Hematoxylin and Eosin (H&E) remain the gold standard for subtype classification, offering rich morphological context. However, the application of computational techniques to analyze WSIs at scale continues to face significant challenges. These include the absence of detailed, pixel-wise annotations, intra-class heterogeneity and inter-class ambiguity, staining variations, and the computational burden posed by the extremely high resolution of WSIs, often resulting in inefficient, resource-intensive pipelines. Furthermore, the distribution of disease-relevant patterns within a slide is highly imbalanced, with diagnostically critical regions occupying only a small portion of the tissue. This thesis addresses these pressing limitations by introducing two novel, dataefficient, and weakly supervised learning frameworks—E-GloConNet and AttenEViT-HDMIL—that collectively advance the state of lung cancer subtype classification from histopathology. These frameworks are designed to operate effectively without reliance on pixel-level annotations, thereby aligning with real-world clinical data availability.