COVID-19 lesion discrimination and localization network based on multi-receptive field attention module on CT images

Abstract

Since discovered in Hubei, China in December 2019, Corona Virus Disease 2019 named COVID-19 has lasted more than one year, and the number of new confirmed cases and confirmed deaths is still at a high level. COVID-19 is an infectious disease caused by SARS-CoV-2. Although RT-PCR is considered the gold standard for detection of COVID-19, CT plays an important role in the diagnosis and evaluation of the therapeutic effect of COVID-19. Diagnosis and localization of COVID-19 on CT images using deep learning can provide quantitative auxiliary information for doctors. This article proposes a novel network with multi-receptive field attention module to diagnose COVID-19 on CT images. This attention module includes three parts, a pyramid convolution module (PCM), a multi-receptive field spatial attention block (SAB), and a multi-receptive field channel attention block (CAB). The PCM can improve the diagnostic ability of the network for lesions of different sizes and shapes. The role of SAB and CAB is to focus the features extracted from the network on the lesion area to improve the ability of COVID-19 discrimination and localization. We verify the effectiveness of the proposed method on two datasets. The accuracy rate of 97.12%, specificity of 96.89%, and sensitivity of 97.21% are achieved by the proposed network on DTDB dataset provided by the Beijing Ditan Hospital Capital Medical University. Compared with other state-of-the-art attention modules, the proposed method achieves better result. As for the public COVID-19 SARS-CoV-2 dataset, 95.16% for accuracy, 95.6% for F1-score and 99.01% for AUC are obtained. The proposed network can effectively assist doctors in the diagnosis of COVID-19 CT images.