Ultrasound COVID-19 Classification Based on the Novel Module-Based Dual-Path Network

Abstract

With a large-scale novel coronavirus pneumonia (COVID-19) outbreak, more and more researchers have acquired convenient and efficient COVID-19 infection status through medical imaging. Here, due to the excellent features of zero-radiation and rapid clinical examination, ultrasound images have been used to assist doctors in COVID-19 diagnosis. To effectively identify the pathological differences between common pneumonia and novel coronavirus pneumonia, an ultrasound COVID-19 classification based on the novel module-based dual-path network (MD-DPNet) is proposed. Specifically, this paper effectively improves the generalization ability by adding the progressive heatmaps intuitively representing the lesion density with the original ultrasound images. Meanwhile, the proposed algorithm creates regular modular sets to reduce the calculating loads and the coupling between each module, which takes advantage of the fact that most pathological features are concentrated in relatively small regions. Furthermore, the proposed MD-DPNet model divides the complete task into multiple sub-tasks to alleviate interclass disequilibrium and the lesion artifact localization is obtained by an unsupervised method. In this paper, experiments show the effectiveness of the proposed innovation, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and Dice reach 0.99, 0.99, 0.99, 0.98, and 0.99, respectively. Experiments on the proposed mixed dataset demonstrate satisfactory results on all the considered tasks, which can provide a new idea for effectively assisting doctors in COVID-19 diagnosis and follow-up treatment.