A Deep Learning System to Screen Novel Coronavirus Disease 2019 Pneumonia

Abstract

The real-time reverse transcription-polymerase chain reaction (RT-PCR) detection of viral RNA from sputum or nasopharyngeal swab had a relatively low positive rate in the early stage of coronavirus disease 2019 (COVID-19). Meanwhile, the manifestations of COVID-19 as seen through computed tomography (CT) imaging show individual characteristics that differ from those of other types of viral pneumonia such as influenza-A viral pneumonia (IAVP). This study aimed to establish an early screening model to distinguish COVID-19 from IAVP and healthy cases through pulmonary CT images using deep learning techniques. A total of 618 CT samples were collected: 219 samples from 110 patients with COVID-19 (mean age 50 years; 63 (57.3%) male patients); 224 samples from 224 patients with IAVP (mean age 61 years; 156 (69.6%) male patients); and 175 samples from 175 healthy cases (mean age 39 years; 97 (55.4%) male patients). All CT samples were contributed from three COVID-19-designated hospitals in Zhejiang Province, China. First, the candidate infection regions were segmented out from the pulmonary CT image set using a 3D deep learning model. These separated images were then categorized into the COVID-19, IAVP, and irrelevant to infection (ITI) groups, together with the corresponding confidence scores, using a location-attention classification model. Finally, the infection type and overall confidence score for each CT case were calculated using the Noisy-OR Bayesian function. The experimental result of the benchmark dataset showed that the overall accuracy rate was 86.7% in terms of all the CT cases taken together. The deep learning models established in this study were effective for the early screening of COVID-19 patients and were demonstrated to be a promising supplementary diagnostic method for frontline clinical doctors.

Keywords: COVID-19; Computed tomography; Location-attention classification model.

Fig. 1

Typical transverse-section CT images: (a) COVID-19; (b) IAVP; and (c) no pneumonia manifestations. Both (a) and (b) were taken within 10 d from the onset of the symptoms.

Fig. 2

Process flow chart—take one COVID-19 case as example. HU: Hounsfield unit.

Fig. 3

(a) COVID-19 image with three ground-glass focuses of infections; (b) IAVP image with four focuses of infections; (c) the minimum distance from the mask to the center of the patch (double-headed arrow); (d) diagonal of the minimum circumscribed rectangle of the pulmonary image.

Fig. 4

The network structure of traditional ResNet-18-based classification model (without the relative distance-from-edge mechanism). The location-attention classification model was built on the backbone of ResNet-18 by concatenating the location-attention mechanism in the full-connection layer to improve the overall accuracy rate. Conv2D: convolution 2D.

Fig. 5

Training curve of (a) loss and (b) accuracy rate for the two classification models.

Fig. 6

All CT images (a–c) were from a single CT case. The focuses of infections were pointed out by arrows.

Fig. 7

Examples of two CT cases reports with bounding boxes on the original images to highlight the focus of infections. Images (a)–(c) are from one case of IAVP. Images (d)–(f) are from one case of COVID-19. The segmented region of pulmonary was an image cube and only the center image was marked with a bounding box to facilitate the interpreting of the lesions.