SNOW: Semi-Supervised, Noisy And/Or Weak Data For Deep Learning In Digital Pathology

Digital pathology produces a lot of images. For machine learning applications, these images need to be annotated, which can be complex and time consuming. Therefore, outside of a few benchmark datasets, real-world applications often rely on data with scarce or unreliable annotations. In this paper, we quantitatively analyze how different types of perturbations influence the results of a typical deep learning algorithm by artificially weakening the annotations of a benchmark biomedical dataset. We use classical machine learning paradigms (semi-supervised, noisy and weak learning) adapted to deep learning to try to counteract those effects, and analyze the effectiveness of these methods in addressing different types of weakness.