Tailoring automated data augmentation to H&E-stained histopathology

Convolutional neural networks (CNN) are sensitive to domain shifts, which can result in poor generalization. In medical imaging, data acquisition conditions differ among institutions, which leads to variations in image properties and thus domain shift. Stain variation in histopathological slides is a prominent example. Data augmentation is one way to make CNNs robust to varying forms of domain shift, but requires extensive hyperparameter tuning. Due to the large search space, this is cumbersome and often leads to sub-optimal generalization performance. In this work, we focus on automated and computationally efficient data augmentation policy selection for histopathological slides. Building upon the RandAugment framework, we introduce several domain-specific modifications relevant to histopathological images, increasing generalizability. We test these modifications on H&E-stained histopathology slides from Camelyon17 dataset. Our proposed framework outperforms the state-of-the-art manually engineered data augmentation strategy, achieving an area under the ROC curve of 0.964 compared to 0.958, respectively.