A depthwise separable dense convolutional network with convolution block attention module for COVID-19 diagnosis on CT scans

Coronavirus disease 2019 (COVID-19) has caused more than 3 million deaths and infected more than 170 million individuals all over the world. Rapid identification of patients with COVID-19 is the key to control transmission and prevent depletion of hospitals. Several networks have been proposed to assist radiologists in diagnosing COVID-19 based on CT scans. However, CTs used in these studies are unavailable for other researchers to do deeper extensions due to privacy concerns. Furthermore, these networks are too heavy-weighted to satisfy the general trend applying on a computationally limited platform. In this paper, we aim to solve these two problems. Firstly, we establish an available dataset COVID-CTx, which contains 828 CT scans positive for COVID-19 across 324 patient cases from three open access data repositories. To our knowledge, it has the largest number of publicly available COVID-19 positive cases compared to other public datasets. Secondly, we propose a light-weighted hybrid neural network: Depthwise Separable Dense Convolutional Network with Convolution Block Attention Module (AM-SdenseNet). AM-SdenseNet synergistically integrates Convolutional Block Attention Module with depthwise separable convolutions to learn powerful feature representations while reducing the parameters to overcome the overfitting problem. Through experiments, we demonstrate the superior performance of our proposed AM-SdenseNet compared with several state-of-the-art baselines. The excellent performance of AM-SdenseNet can improve the speed and accuracy of COVID-19 diagnosis, which is extremely useful to control the spreading of infection.