UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Domain adaptation and multi-scale relational graph neural network in classification of prostate cancer histopathology images Bazargani, Roozbeh


Most current deep learning models for hematoxylin and eosin (H&E) histopathology image analysis lack the power of generalization to datasets collected from other institutes due to the domain shift in the data. While graph convolutional neural networks have shown significant potential in natural and histopathology images, their use in histopathology images has only been studied using a single magnification or multi-magnification with late fusion. In this thesis, we study the domain shift problem with multiple instance learning (MIL) on prostate cancer datasets collected from different centers. First, we develop a novel center-based H&E color augmentation for cross-center model generalization. While previous work used methods such as random augmentation, color normalization, or learning domain-independent features to improve the robustness of the model to changes in H&E stains, our method first augments the H&E color space of the source dataset to color space of both datasets and then adds random color augmentation. Our method covers the larger range of the color distribution of both institutions resulting in a better generalization. Next, to leverage the multi-magnification information and early fusion with graph convolutional networks, we handle different embedding spaces at each magnification by introducing the Multi-Scale Relational Graph Convolutional Network (MSRGCN) as a novel MIL method. We model histopathology image patches and their relation with neighboring patches and patches at other magnifications as a graph. To pass the information between different magnification embedding spaces, we define separate message-passing neural networks based on the node and edge type. Our proposed color adaptation method improves the model performance on both the source and target datasets, and has the best performance on the unlabeled target dataset compared to State-Of-The-Art (SOTA), showing promise as an approach to learning generalizable features for histopathology image analysis. We also compare our MS-RGCN with multiple SOTA methods with evaluations on several source and held-out datasets. Our method outperforms the SOTA on all of the datasets and image types consisting of tissue microarrays, whole-mount slide regions, and whole-slide images. Through an ablation study, we test and show the value of the pertinent design features of the MS-RGCN.

Item Citations and Data


Attribution-NonCommercial-NoDerivatives 4.0 International