Dynamic feature learning for COVID-19 segmentation and classification

Abstract

Since December 2019, coronavirus SARS-CoV-2 (COVID-19) has rapidly developed into a global epidemic, with millions of patients affected worldwide. As part of the diagnostic pathway, computed tomography (CT) scans are used to help patient management. However, parenchymal imaging findings in COVID-19 are non-specific and can be seen in other diseases. In this work, we propose to first segment lesions from CT images, and further, classify COVID-19 patients from healthy persons and common pneumonia patients. In detail, a novel Dynamic Fusion Segmentation Network (DFSN) that automatically segments infection-related pixels is first proposed. Within this network, low-level features are aggregated to high-level ones to effectively capture context characteristics of infection regions, and high-level features are dynamically fused to model multi-scale semantic information of lesions. Based on DFSN, Dynamic Transfer-learning Classification Network (DTCN) is proposed to distinguish COVID-19 patients. Within DTCN, a pre-trained DFSN is transferred and used as the backbone to extract pixel-level information. Then the pixel-level information is dynamically selected and used to make a diagnosis. In this way, the pre-trained DFSN is utilized through transfer learning, and clinical significance of segmentation results is comprehensively considered. Thus DTCN becomes more sensitive to typical signs of COVID-19. Extensive experiments are conducted to demonstrate effectiveness of the proposed DFSN and DTCN frameworks. The corresponding results indicate that these two models achieve state-of-the-art performance in terms of segmentation and classification.