Out-of-Distribution Generalization of Gigapixel Image Representation

Abstract

This thesis addresses the significant challenge of improving the generalization capabilities of artificial deep neural networks in the classification of whole slide images (WSIs) in histopathology across different and unseen hospitals. It is a critical issue in AI applications for vision-based healthcare tasks, given that current standard methodologies struggle with out-of-distribution (OOD) data from varying hospital sources. In histopathology, distribution shifts can arise due to image acquisition variances across different scanner vendors, differences in laboratory routines and staining procedures, and diversity in patient demographics. This work investigates two critical forms of generalization within histopathology: magnification generalization and OOD generalization towards different hospitals. One chapter of this thesis is dedicated to the exploration of magnification generalization, acknowledging the variability in histopathological images due to distinct magnification levels and seeking to enhance the model's robustness by learning invariant features across these levels. However, the major part of this work focuses on OOD generalization, specifically unseen hospital data. The objective is to leverage knowledge encapsulated in pre-existing models to help new models adapt to diverse data scenarios and ensure their efficient operation in different hospital environments. Additionally, the concept of Hospital-Agnostic (HA) learning regimes is introduced, focusing on invariant characteristics across hospitals and aiming to establish a learning model that sustains stable performance in varied hospital settings. The culmination of this research introduces a comprehensive method, termed ALFA (Exploiting All Levels of Feature Abstraction), that not only considers invariant features across hospitals but also extracts a broader set of features from input images, thus maximizing the model's generalization potential. The findings of this research are expected to have significant implications for the deployment of medical image classification systems using deep models in clinical settings. The proposed methods allow for more accurate and reliable diagnostic support across various hospital environments, thereby improving diagnostic accuracy and reliability, and paving the way for enhanced generalization in histopathology diagnostics using deep learning techniques. Future research directions may build on expanding these investigations to further improve generalization in histopathology.