Training a neural network based on unreliable human annotation of medical images

Building classification models from clinical data often requires human experts for example labeling. However, it is difficult to obtain a perfect set of labels due to the complexity of the medical data and the large variability between experts. In this study we present a neural-network training strategy that is more robust to unreliable labeling by explicitly modeling the label noise as part of the network architecture. Our method is demonstrated on breast microcalcifications classification into benign and malignant categories, given multi-view mammograms. We show that the proposed training procedure outperforms standard training methods that ignore the existence of label noise.