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

AbstractReal-world large-scale medical image analysis (MIA) datasets have three challenges: 1) they contain noisy-labelled samples that affect training convergence and generalisation, 2) usually an imbalanced distribution of per class, 3) normally comprise a multi-label problem, where can multiple diagnoses. Current approaches are commonly trained to solve subset those problems, but we unaware methods address the problems simultaneously. In this paper, propose new module called Non-Volatile Unbiased Memory (NVUM), which non-volatility stores running average model logits for regularization loss on noisy problem. We further unbias classification prediction in NVUM update learning run extensive experiments evaluate benchmarks proposed by is performed chest X-ray (CXR) sets, formed Chest-Xray14 CheXpert, testing clean CXR OpenI PadChest. Our method outperforms previous state-of-the-art classifiers deal with labels all evaluations. code available at https://github.com/FBLADL/NVUM.KeywordsChest classificationMulti-label classificationImbalanced