Improving multi-label chest X-ray disease diagnosis by exploiting disease and health labels dependencies

The widely used ChestX-ray14 dataset addresses an important medical image classification problem and has the following caveats: 1) many lung pathologies are visually similar, 2) a variant of multiple diseases including lung cancer, tuberculosis, and pneumonia are present in a single scan at the same time, i.e. multiple labels. Existing literature uses state-of-the-art deep learning models being transfer learned where output neurons of the networks are trained for individual diseases to cater for multiple disease labels in each image. However, most of them don't consider the label relationship explicitly between present and absent classes. In this work we have proposed a pair of novel error functions that can be employed for any deep learning model, Multi-label Softmax Loss (MSML) and Correlation Loss (CorLoss), to specifically address the properties of multiple labels and visually similar data. Moreover, we provide a fine-grained perspective into this problem and use bilinear pooling as an encoding scheme to increase discrimination of the model. The experiments are conducted on the ChestX-ray14 dataset. We first report improvements using our proposed loss with various backbone networks. After that, we extend our experiments to prove the rich disparity being learned by the model with our proposed losses, which can be fused with other models to improve the overall performances.