Artifact Identification in Digital Histopathology Images Using Few-Shot Learning

The advent of deep learning methods has led to breakthroughs in many digital histopathology image analysis tasks. However, automatic analysis is often impacted by the presence of various artifacts introduced during different tissue and slide processing stages. Therefore, it is desirable to have a generic artifacts identification algorithm to automatically exclude the artifacts regions in the downstream analysis. In this paper, considering the wide diversity of artifacts that present in histopathology images, and the difficulty to obtain a large amount of training data, we frame the artifacts identification task as a tile-based image classification problem and explore the feasibility of using a few-shot learning technique, specifically, prototypical network, for the task. We demonstrate that the use of prototypical network can effectively identify image tiles that contain various artifacts using a very small set of training images. The trained model is also able to generalize well to unseen artifacts. We validate the approach by applying it on both immunohistochemistry and H&E stained tissues images, showing that it is a more favorable approach compared to standard transfer learning for this application.