Explainability-based graph augmentation for out-of-distribution robustness in digital pathology

Whole slide images (WSIs), which are high-resolution digital representations of tissue samples, present significant challenges for processing because of their gigapixel scale. Recent studies show that graph neural networks (GNNs), which leverage neighborhood information, can enhance cancer classification accuracy in WSIs. However, GNN performance is affected by out-of-distribution (OOD) data, which occurs when the training and testing data are from different sources. Detecting OOD samples in graph data is especially challenging due to its complexity, which makes GNNs vulnerable to performance degradation. To address this issue, we propose a novel data augmentation framework to improve GNN robustness against OOD samples. Our approach augments node features by sampling important subgraphs, simulating potential OOD scenarios during training. Experiments on three public WSI datasets demonstrate significant improvements in graph classification tasks on OOD samples. In this work, one dataset serves as the in-distribution benchmark, while the others represent OOD scenarios. These results highlight the potential of data augmentation to enhance GNN robustness against OOD samples, improving cancer classification in WSIs.