Attention Mechanism in Radiologist-Level Thorax Diseases Detection

Abstract Chest X-ray images play an important role in diagnosing thorax diseases. They can show the complete state of the human chest, including the heart, lungs, ribs, etc. However, the diagnosis of diseases is mainly based on the location and morphology of diseased tissues, which requires our model to pay more attention to the lesion, so we choose to introduce the attention mechanism. In recent years, the development of deep neural networks has gradually shifted from increasing the depth of the network to increasing the complexity of the network, which is to improve the capacity of the network by introducing attention mechanism with almost no parameter increase. Multiple attention mechanisms have been proposed for visual tasks to augment convolutions. They have been proved to be powerful in the public datasets but which kind of attention is more effective on X-ray images has not been studied. This paper explores effectiveness of spatial attention and channel attention mechanisms on Chest X-ray 14 dataset and combines them together to improve the disease recognition capability of model. Our AUROC beat the benchmark by 7.7%.