DualCheXNet: dual asymmetric feature learning for thoracic disease classification in chest X-rays

Abstract

Recently, deep convolutional neural networks (DCNNs) such as the famous ResNet and DenseNet have achieved significant improvements in the field of automatic analysis of chest X-rays (CXRs). However, we observe that a wider network can combine characteristics from different DCNNs to improve the ability of object recognition compared with the single networks. In this paper, we focus on the cooperation and complementarity of dual asymmetric DCNNs and present a novel dual asymmetric feature learning network named DualCheXNet for multi-label thoracic disease classification in CXRs. Correspondingly, two asymmetric subnetworks based on the ResNet and DenseNet are combined to adaptively capture more discriminative features of different abnormalities from the raw CXRs. Specifically, the proposed method enables two different feature fusion operations, such as feature-level fusion (FLF) and decision-level fusion (DLF), which exactly form the complementary feature learning embedded in DualCheXNet. Moreover, an iterative training strategy is designed to integrate the loss contribution of the involved classifiers into a unified loss, and optimize the process of complementary features learning in an alternative way. Extensive experiments on the ChestX-ray14 dataset clearly substantiate the effectiveness of the proposed method as compared with the state-of-the-art baselines.