Reducing the Annotation Cost of Whole Slide Histology Images using Active Learning

Histopathology serves as the gold standard for tumor diagnosis. Whole slide scanners have made computer vision-based methods available for pathologists to locate regions of high diagnostic significance. An essential step of whole slide image (WSI) diagnosis is the segmentation of the tumor region by generating a tumor probability heatmap. Most WSI diagnosis methods use patch-based classifiers or segmentation models, they both require a large set of training patches from annotated WSIs. Annotating WSIs is time-consuming and laborious. Active learning is a method that can suggest the most informative unlabeled data for annotation, but traditional active learning methods are not directly applicable for WSIs. Meanwhile, unannotated WSIs also contain rich information that can be further exploited by self-supervised learning. By utilizing unannotated data alongside active learning, we proposed a self-supervised active learning framework for tumor region segmentation of WSIs. The proposed method is evaluated on the public available CAMELYON dataset and achieved satisfying performance using 3% of the annotated data.