In this paper, we propose a deep reinforcement learning-based approach, called Sequential Attention-based Sampling for Histopathological Analysis (SASHA), for efficient analysis of large-scale histopathological images (WSI). SASHA learns information-rich features using a lightweight hierarchical attention-based multi-instance learning (MIL) model, and selectively samples high-resolution patches, corresponding to 10-20% of the entire image, to perform reliable diagnosis. It achieves comparable performance to conventional high-resolution full-image analysis methods with much lower computational cost and memory usage, and outperforms conventional sparse sampling methods. This presents an efficient solution to the problem of automated diagnosis of large-scale medical images where diagnostic information is limited to only a portion of the image.
