Multilabel Classification and Localization of Rare Pulmonary Diseases using Deep Learning

Abstract

Chest radiography is the most common radiological examination used for the diagnosis of thoracic diseases. Currently, automated classification of radiological images is abundantly used in clinical diagnosis. However, each pathology has its own response characteristic receptive field regions, which is a key problem during the classification of chest diseases. In addition to extreme class imbalance, cases labelled as uncertain in the dataset further complicate this task. To solve this problem, we propose a semi-supervised learning approach known as U-SelfTrained. In this scheme, uncertain labels are left unlabeled in the dataset; first, the model is trained on labelled instances and then on unlabeled instances relabeling them with labels having a higher probability. Comprehensive experimentation was carried out on the CheXpert dataset, which consists of 223,816 frontal and lateral view CXR images of 64,740 patients with 14 diseases. The testing accuracy is 0.877 on the CheXpert dataset, which is currently the highest score achieved to date. This validates the effectiveness of this method for chest radiography classification. The practical significance of this study is the implementation of AI algorithms to assist radiologists in improving their diagnostic accuracy.