Using deep learning in pathology image analysis: A novel active learning strategy based on latent representation

Most countries worldwide continue to encounter a pathologist shortage, significantly impeding the timely diagnosis and effective treatment of cancer patients. Deep learning techniques have performed remarkably well in pathology image analysis; however, they require expert pathologists to annotate substantial pathology image data. This study aims to minimize the need for data annotation to analyze pathology images. Active learning (AL) is an iterative approach to search for a few high-quality samples to train a model. We propose our active learning framework, which first learns latent representations of all pathology images by an auto-encoder to train a binary classification model, and then selects samples through a novel ALHS (Active Learning Hybrid Sampling) strategy. This strategy can effectively alleviate the sample redundancy problem and allows for more informative and diverse examples to be selected. We validate the effectiveness of our method by undertaking classification tasks on two cancer pathology image datasets. We achieve the target performance of 90% accuracy using 25% labeled samples in Kather's dataset and reach 88% accuracy using 65% labeled data in BreakHis dataset, which means our method can save 75% and 35% of the annotation budget in the two datasets, respectively.