Training Region Selector for Gram Stained Slides with Limited Data: A Data Distillation Approach

In this paper, we tackle the region selection task for the purpose of whole slide image (WSI) analysis with a focus on Gram stained slides. Typically, the number of tiles required to capture using high magnification objective is extremely large. Thus, in practice, we need to perform region of interest selection in low magnification objectives to choose only the best candidate regions for the later high magnification analysis. With the fast development of computer vision and deep learning, it is possible to train an accurate convolutional neural network (CNN) based classifier to do this region selection task which would normally be done by experienced scientists and pathologists. However, data collection and labelling during the training stage are very time-consuming, labour intensive and expensive tasks. Researchers starting to ask a much more challenging question: how can we use unlabeled data to boost the system performance? In this paper, to answer this question and to reduce labelling effort, we propose a data distillation training framework to train a CNN classifier on limited labeled data with the help of unlabeled data. The unlabeled data will be carefully selected by a teacher model via a data distillation process and put into the training set for a student model. Extensive experiments show that the proposed framework achieves a notable gain in accuracy.