Noise Robust Learning with Hard Example Aware for Pathological Image classification

Label noise among the datasets can be memorized by deep neural networks, which degenerates the robustness of the learned model. However, recent studies proposed to deal with label noise are difficult in distinguishing informative hard examples from harmful noise ones, especially for medical image analysis. In this paper, we propose a noise robust learning algorithm based on hard example aware and Co-weighting to circumvent label noise. Firstly, we identify and exclude noisy samples according to the statistical features of predictive history, which allows retaining informative hard examples. Secondly, we adopt Co-weighting strategy for robust training, which improves Co-teaching algorithm through weighting samples between peer networks. Compared with extensive state-of-the-art methods, our strategy achieves the best performance on two colon tumor histology image datasets under different ratios of noise.