Work

Histopathology Image Classification Using Machine Learning

Year: 2026
Type: article
Abstract: Histopathology image classification is a critical component of cancer diagnosis. However, the gigapixel scale of Whole-Slide Images (WSIs) and the high variability in tissue staining and scanner quality across medical centers present significant computational challenges. This dissertation proposes a comprehensive machine learning framework to address these challenges, bridging the gap between theoretical models and practical clinical deployment. First, to manage the massive dimensionality and noise inherent in WSIs, this research develops a robust feature extraction methodology. The pipeline implements a stringent tile filtering technique to eliminate physical artifacts and resolve severe class imbalances. It integrates ConvNeXt alongside an attention-based pooling mechanism to isolate the most informative tissue regions. Evaluated against state-of-the-art extractors like CLAM, KimiaNet, CTransPath, and Lunit-DINO, this approach successfully compresses gigapixel images into dense, representative vectors while maintaining superior diagnostic accuracy. Building upon these representations, this dissertation introduces a dynamic feature-space graph utilizing EdgeConv and a supervised gating mechanism for WSI classification. The model was rigorously evaluated on binary and complex multi-class subtyping tasks, specifically differentiating Adenocarcinoma (LUAD) from Squamous Cell Carcinoma (LUSC) and normal tissue. Compared against baselines such as CLAM, HipoMap, and TransMIL, the proposed framework demonstrated exceptional clinical reliability. The gating mechanism provided extreme robustness against domain shifts and scanner noise, significantly outperforming baselines that suffered severe degradation under similar conditions. Finally, to address hardware limitations in real-world medical facilities, this work investigates model deployment in resource-constrained environments. Using the PCam and MHIST datasets, model quantization was applied to reduce the memory footprint and computational requirements of the classification frameworks. This enables advanced diagnostic inference on low-power edge devices, making the technology accessible to remote clinics lacking expensive server infrastructure. Ultimately, this dissertation provides a robust, end-to-end computational pathology framework. By solving for data dimensionality, domain shift, and hardware limitations, this research paves the way for accurate, scalable, and practical AI-assisted diagnostics in clinical settings.
Language: English
Open Access status: green
OpenAlex ID: W7171751057
FWCI:
Cites: 116
Cited by:
Related to: 10