Multi-label Chest X-Ray Image Classification via Label Co-occurrence Learning

Abstract

Existing multi-label medical image datasets generally exist a large number of co-occurring label pairs. This phenomenon is interpreted as label co-occurrence. Obviously, label co-occurrence is available and its effect should be taken into account for automated medical diagnosis. In this paper, we propose a novel label co-occurrence learning (LCL) method for multi-label chest X-ray (CXR) image classification. By taking advantage of the dependencies between pathologies, the proposed LCL module in our model is designed to generate a set of inter-dependent weighting adapters for exploring the potential pathologies. Specifically, these classifiers are initialized with the weight coefficients extracted from the co-occurrence matrix of training data. In the training phase, the LCL module can accurately tweak the multi-label outputs with their corresponding object weights, and further predict additional abnormal findings. Moreover, the LCL module can be directly integrated into any Convolutional Neural Networks (CNNs) with end-to-end training. Extensive experiments on the ChestX-ray14 dataset substantiate the effectiveness of the proposed method as compared with the state-of-the-art baselines.