Pulmonary X-Ray Image Classification Using Deep Learning with Coordinate Attention and Meta-ACON Activation Function

Pneumonia is a serious health risk that usually examined by lung X-rays. Traditional deep learning models neglect image location information, fail to establish long-range dependencies and the limited ability to handle pneumonia features. In this paper, we propose MCANet model based on ResNet50 architecture by combining coordinate attention module and introducing a new activation function (Meta-ACON). The coordinate attention module in this model is used to capture positional information and establish long-range dependencies on feature maps. Meta-ACON activation function can adaptively select whether each neuron is activated, which contributes to dynamically handling pneumonia features. This paper performs comparison experiments on the public dataset ChestXray2020. Accuracy, Precision, Recall, and F1-score are used as evaluation metrics. Experiment results show that the MCANet improves each metrics compared with the original ResNet50 by 4.14%, 3.63%,4.25% and 3.99% respectively.