Handling label noise through model confidence and uncertainty: application to chest radiograph classification

Abstract

In this work we analyze the effect of label noise in training and test data when performing classification experiments on chest radiographs (CXRs) with modern deep learning architectures. We use ChestXRay14, the largest publicly available CXR dataset. We simulate situs inversus by horizontal flipping of the CXRs, allowing us to precisely control the amount of label noise. We also perform experiments in classifying emphysema using the ChestXRay14 provided labels that are known to be noisy. Our situs inversus experiments confirm results from the computer vision literature that deep learning architectures are relatively robust but not completely insensitive to label noise in the training data: without or with very low noise, classification results are near perfect; 16% and 32% training label noise only lead to a 1.5% and 4.6% drop in accuracy. We investigate two metrics that could be used to identify test samples that have an incorrect label: model confidence and model uncertainty. We show, in an observer study with an experienced chest radiologist, that both measures are effective in identifying samples in ChestXRay14 that are erroneously labeled for the presence of emphysema.