Automatic Patient-Level Detection of Coronavirus Disease (COVID-19) Using Convolutional Neural Network from Lung CT Scans

ABSTRACT

The outbreak of 2019 novel coronavirus (COVID-19) has caused more than 176 million confirmed cases by June 14, 2021, and this number will continue to grow. Automatic and accurate COVID-19 detection/evaluation from the computed tomography (CT) scans is of great significance for COVID-19 diagnosis and treatment. Due to individual variations of patients and the influx of a large number of patients, the current clinical practices remain subject to shortcomings of potential high-risk and time-consumption issues from radiologists. In this paper, we propose a computer aided detection system to relieve the clinical physicians from tediously reading the CT images of COVID-19 patients. Particularly, a COVID-19 detection network (COVIDNet) is proposed using deep convolutional neural networks (DCNNs) for patient-level COVID-19 detection to distinguish infected and non-infected patients. The underlying method complementarily and comprehensively extract multi-level interplane volumetric correlation features of typical ground glass opacities (GGOs) lesions using 3D multi-Scale Network (MSN). To cover more GGO lesion features and reduce intra-class differences, a Phase Ensemble (PE) is proposed for aggregation of different phases in one CT scan. The proposed method is evaluated on a clinically established COVID-19 database with five-fold cross-validation. Experimental results show that the proposed framework achieves classification performance with the specificity of 1.0000, sensitivity of 0.9700, accuracy of 0.9850, precision of 1.0000, and Area Under the Curve (AUC) of 0.9980. All of these indicate that our method enables an efficient, accurate and reliable patient-level COVID-19 detection for clinical diagnosis. This can significantly improve the work efficiency of clinical physicians on COVID-19 patient diagnosis and evaluation in hospitals and clinics. Impact statement—The proposed method can automatically and accurately distinguish the COVID-19 patients from patient-level CT scan images. On a clinically established large-scale COVID-19 database with five-fold cross-validation, the experimental results show that the proposed framework achieves a classification performance with the specificity of 1.0000, sensitivity of 0.9700, accuracy of 0.9850, precision of 1.0000, and Area Under the Curve (AUC) of 0.9980. It can relieve the clinical physicians from tediously reading the CT images of COVID-19 patients. Thus, it can significantly improve the work efficiency of clinical physicians on COVID-19 patient diagnosis and evaluation in hospitals and clinics, particularly in the pandemic period of COVID-19.