OADA: An Online Data Augmentation Method for Raw Histopathology Images

Deep learning-based automatic medical diagnosis is intensively studied in recent years. Abundant clinical raw records can be utilized, but we demonstrate that mixed and unknown magnification scales and staining conditions of raw histopathology images greatly hinder many successful deep models in this task. To address this problem, this paper proposes an Online Adaptive Data Augmentation method (OADA). In each training epoch, OADA adaptively selects base images and determines the personalized augmentation size of each image based on the current training status. The chosen images are augmented to update the training set. Extensive experiments show that OADA-empowered deep models obtain significant improvement compared to their bare versions, and OADA outperforms a suite of data augmentation baselines and state-of-the-art competitors.