Multi-label long-tailed disease recognition on Chest X-ray images

Abstract

Machine learning and deep learning recently have many big achievements in computer vision and there is a trend to apply deep learning in diagnostic medical images, for example, Chest-Xray classification. Since most large-scale image classification benchmarks contain single-label images with a mostly balanced distribution of labels, many standard deep learning methods fail to accommodate the class imbalance and cooccurrence problems posed by the long-tailed multi-label nature of tasks like disease diagnosis such as Chest-Xray classification. Compared to conventional single-label classification problem, multi-label recognition is often more challenging due to issues called the dominant of negative samples (when we treat multi-label classification as series of binary classification) and the long tail distribution of positive samples. In this thesis, we modified the orignial binary cross entropy loss to get a new loss function called class-aware balanced loss which can solve two previous problems in ChestXray14 dataset. We train Swin Transformer model on Chest-Xray14[1] dataset with our new loss and archive the best AUC score compared to other SOTA algorithms