Sample self-selection using dual teacher networks for pathological image classification with noisy labels