Deep Active Learning for Breast Cancer Segmentation on Immunohistochemistry Images.

摘要: Immunohistochemistry (IHC) plays an essential role in breast cancer diagnosis and treatment. Reliable automatic segmentation of regions on IHC images would be considerable value for further analysis. However, the prevalent fully convolutional networks (FCNs) suffer from difficulties obtaining sufficient annotated training data. Active learning, other hand, aims to reduce cost annotation by selecting informative effective subset labeling. In this paper, we present a novel deep active learning framework images. Three criteria are explicitly designed select samples: dissatisfaction, representativeness diverseness. Dissatisfaction, consisting both pixel-level image-level focuses samples that network does not segment well. Representativeness chooses can mostly represent all unlabeled diverseness makes chosen different those already set. We evaluate proposed method large-scale in-house dataset demonstrate our outperforms state-of-the-art suggestive (SA) [1] representative (RA) [5] two test sets achieves competitive or even superior performance using 40% data full set