Ph.D. Dissertation Defense: Mohammed Alali
Monday, April 13
9 AM
Zoom only: https://unl.zoom.us/j/98961632196
"Histopathology Image Classification Using Machine Learning"
Histopathology image classification is vital for cancer diagnosis, but the gigapixel scale of Whole-Slide Images (WSIs) and severe domain shifts present major computational hurdles. This dissertation proposes a comprehensive, end-to-end machine learning framework designed to bridge the gap between theoretical accuracy and real-world clinical deployment.
First, to manage massive data dimensionality, this research develops a robust feature extraction pipeline. By combining heuristic tile filtering with a ConvNeXt architecture and attention-based pooling, the system successfully compresses gigapixel WSIs into dense feature vectors. This approach eliminates physical artifacts and resolves class imbalances, matching or outperforming state-of-the-art extractors.
Building on these representations, the dissertation introduces a dynamic feature-space graph utilizing EdgeConv and a supervised gating mechanism. Evaluated on complex multi-class subtyping tasks—specifically differentiating Adenocarcinoma (LUAD), Squamous Cell Carcinoma (LUSC), and normal tissue—the framework demonstrated exceptional reliability. The gating mechanism provided extreme robustness against domain shifts and noise, significantly outperforming baselines like TransMIL and CLAM.
Finally, to enable deployment in resource-constrained medical facilities, the framework applies model quantization to reduce computational requirements. By optimizing models for low-power edge devices, this dissertation delivers an accurate, scalable, and highly accessible AI-assisted diagnostic tool.
Commitee:
Dr. Jitender Deogun, Advisor
Dr. Lisong Xu
Dr. Juan Cui
Dr. Etsuko Moriyama