An Overview of Few-Shot Learning Methods in Analysis of Histopathological Images

Analysis of histopathological images allows doctors to diagnose diseases like cancer, which is the cause of nearly one in six deaths worldwide. Classification of such images is one of the most critical topics in biomedical computing. Deep learning models obtain high prediction quality but require a lot of annotated data for training. The data must be labeled by domain experts, which is time-consuming and expensive. Few-shot methods allow for data classification using only a few training samples; therefore, they are an increasingly popular alternative to collecting a large dataset and supervised learning. This chapter presents a survey on different few-shot learning techniques of histopathological image classification with various types of cancer. The methods discussed are based on contrastive learning, meta-learning, and data augmentation. We collect and overview publicly available datasets with histopathological images. We also show some future research directions in few-shot learning in the histopathology domain.