Towards Better Understanding and Better Generalization of Low-shot Classification in Histology Images with Contrastive Learning

Low-shot learning is an established topic in natural images for years, but few work is attended to histology images, which is of high clinical value since well-labeled datasets and rare abnormal samples are expensive to collect. Here, we pioneer the study of low-shot learning for histology images by setting up three cross-domain tasks that reflect real clinics problems. To enable label-efficient learning and better generalizability, we propose to incorporate contrastive learning (CL) with latent augmentation (LA) to build a few-shot system. CL learns useful representations without manual labels, while LA transfers semantic variations of base dataset in an unsupervised way. These two components fully exploit unlabeled training data and can scale gracefully to other label-hungry problems. In experiments, we find: i) models learned by CL generalize fairly better than supervised learning for histology images, and ii) LA brings consistent gains over baselines. Prior studies of self-supervised representation learning mainly focus on ImageNet-like images, which only present a dominant object in their centers. Recent attention has been paid to images with multi-objects and multi-textures (Chen & Li, 2020). Histology images are a natural choice for such study. We show the superiority of CL over supervised learning in terms of generalization for such data, and provide our empirical understanding for this observation. The findings in this work could contribute to understanding how model generalizes in the context of both self-supervised learning and histology images.