Automatic Diagnosis of COVID-19 Medical Images based on Graph Attention Network

Abstract

In view of the COVID-19 pandemic and its highly infectious characteristic, traditional artificial diagnosis based on medical imaging, though capable of detecting pulmonary lesion in human body, is found of lower efficiency. Therefore, it is particularly urgent that we design a set of accurate and automatic pneumonia diagnosis methods with aid of artificial intelligence technology, so that pneumonia in patients can be diagnosed and treated early. This study first introduces DenseNet to the Convolutional Neural Network (CNN) structure to improve sharing of characteristic information of lung image in convolutional layers and thus obtain more accurate image features. Secondly, characteristics of pneumonia disease are discriminated rapidly using the Graphic Attention Network (GAT). The authors adopt the X-ray dataset in Radiological Society of North America (RSNA) Pneumonia Detection Challenge released by Kaggle to train and verify the network. According to experimental results, the accuracy of COVID-19 diagnosis and F-Score both reach 98%. The method provides CT doctors with an end-to-end deep learning technology for pneumonia diagnosis.