Comparison of Different Augmentation Techniques for Improved Generalization Performance for Gleason Grading

The fact that deep learning based algorithms used for digital pathology tend to overfit to the site of the training data is well-known. Since an algorithm that does not generalize is not very useful, we have in this work studied how different data augmentation techniques can reduce this problem but also how data from different sites can be normalized to each other. For both of these approaches we have used cycle generative adversarial networks (GAN); either to generate more examples to train on or to transform images from one site to another. Furthermore, we have investigated to what extent standard augmentation techniques improve the generalization performance. We performed experiments on four datasets with slides from prostate biopsies, stained with H&E, detailed annotated with Gleason grades. We obtained results similar to previous studies, with accuracies of 77% for Gleason grading for images from the same site as the training data and 59% for images from other sites. However, we also found out that the use of traditional augmentation techniques gave better performance compared to when using cycle GANs, either to augment the training data or to normalize the test data.