Identifying cardiomegaly in chest x-rays using dual attention network

The chest X-ray (CXR) is one of the most commonly available radiological examinations for identifying chest diseases. The application of deep learning methods in computer vision is becoming more and more mature, it provides new methods for automatic analysis of medical images and assisting doctors in high-precision intelligent diagnosis. In this paper, we propose a dual attention network to identify cardiomegaly (CXRDANet) on CXR images. CXRDANet is equipped with channel attention module (CAM) and spatial attention module (SAM), which selectively enhance features highly related to lesion area. We select CXR images of cardiomegaly and normal from ChestX-ray14 and NLM-CXR, without overlapping images, as the training set and the test set. Experimental results show that our method attains the accuracy of 0.9050, the sensitivity of 0.9445, the specificity of 0.8610, the F1 score of 0.9059, the AUC of 0.9588, which is a new state-of-the-art performance. In addition, we apply our method to the multi-label CXR image classification, and its performance has reached an excellent level.