Whole slide pathology image patch based deep classification: an investigation of the effects of the latent autoencoder representation and the loss function form

The analysis of whole-slide pathological images is a major area of deep learning applications in medicine. The automation of disease identification, prevention, diagnosis, and treatment selection from whole-slide images (WSIs) has seen many advances in the last decade due to the progress made in the areas of computer vision and machine learning. The focus of this work is on patch level to slide image level analysis of WSIs, popular in the existing literature. In particular, we investigate the nature of the information content present in images on the local level of individual patches using autoencoding. Driven by our findings at this stage, which raise questions about the us of autoencoders, we next address the challenge posed by what we argue is an overly coarse classification of patches as tumorous and non-tumorous, which leads to the loss of important information. We showed that task specific modifications of the loss function, which take into account the content of individual patches in a more nuanced manner, facilitate a dramatic reduction in the false negative classification rate.