NVUM: Non-Volatile Unbiased Memory for Robust Medical Image Classification

Real-world large-scale medical image analysis (MIA) datasets havethree challenges: 1) they contain noisy-labelled samples that affect training con-vergence and generalisation, 2) they usually have an imbalanced distribution ofsamples per class, and 3) they normally comprise a multi-label problem, wheresamples can have multiple diagnoses. Current approaches are commonly trainedto solve a subset of those problems, but we are unaware of methods that ad-dress the three problems simultaneously. In this paper, we propose a new trainingmodule called Non-Volatile Unbiased Memory (NVUM), which non-volatilitystores running average of model logits for a new regularization loss on noisymulti-label problem. We further unbias the classification prediction in NVUMupdate for imbalanced learning problem. We run extensive experiments to eval-uate NVUM on new benchmarks proposed by this paper, where training is per-formed on noisy multi-label imbalanced chest X-ray (CXR) training sets, formedby Chest-Xray14 and CheXpert, and the testing is performed on the clean multi-label CXR datasets OpenI and PadChest. Our method outperforms previous state-of-the-art CXR classifiers and previous methods that can deal with noisy labelson all evaluations.