From community-acquired pneumonia to COVID-19: a deep learning-based method for quantitative analysis of COVID-19 on thick-section CT scans

Abstract

Objective: To develop a fully automated AI system to quantitatively assess the disease severity and disease progression of COVID-19 using thick-section chest CT images.

Methods: In this retrospective study, an AI system was developed to automatically segment and quantify the COVID-19-infected lung regions on thick-section chest CT images. Five hundred thirty-one CT scans from 204 COVID-19 patients were collected from one appointed COVID-19 hospital. The automatically segmented lung abnormalities were compared with manual segmentation of two experienced radiologists using the Dice coefficient on a randomly selected subset (30 CT scans). Two imaging biomarkers were automatically computed, i.e., the portion of infection (POI) and the average infection HU (iHU), to assess disease severity and disease progression. The assessments were compared with patient status of diagnosis reports and key phrases extracted from radiology reports using the area under the receiver operating characteristic curve (AUC) and Cohen's kappa, respectively.

Results: The dice coefficient between the segmentation of the AI system and two experienced radiologists for the COVID-19-infected lung abnormalities was 0.74 ± 0.28 and 0.76 ± 0.29, respectively, which were close to the inter-observer agreement (0.79 ± 0.25). The computed two imaging biomarkers can distinguish between the severe and non-severe stages with an AUC of 0.97 (p value < 0.001). Very good agreement (κ = 0.8220) between the AI system and the radiologists was achieved on evaluating the changes in infection volumes.

Conclusions: A deep learning-based AI system built on the thick-section CT imaging can accurately quantify the COVID-19-associated lung abnormalities and assess the disease severity and its progressions.

Key points: • A deep learning-based AI system was able to accurately segment the infected lung regions by COVID-19 using the thick-section CT scans (Dice coefficient ≥ 0.74). • The computed imaging biomarkers were able to distinguish between the non-severe and severe COVID-19 stages (area under the receiver operating characteristic curve 0.97). • The infection volume changes computed by the AI system were able to assess the COVID-19 progression (Cohen's kappa 0.8220).

Keywords: Artificial intelligence; COVID-19; Deep learning; Disease progression.

Fig. 1

Lesion segmentation for three consecutive axial CTs from a severe patient. First row: original image; second row: lesion segmentation image

Fig. 2

Lesion segmentation for three consecutive axial CTs from a non-severe patient. First row: original image; second row: lesion segmentation image

Fig. 3

Box-plot of POI (a) and iHU (b) for the severe and non-severe patients. POI, portion of infection; iHU, average infection Hounsfield unit

Fig. 4

Receiver operating characteristic (ROC) curves of the model. AUC, area under the receiver operating characteristic curve; POI, portion of infection; iHU, average infection Hounsfield unit

Fig. 5

The lesion segmentation of six adjacent CT scans that taken from Jan. 27 to Feb. 12 for a severe patient. The red dot corresponds to the time given for the “severe” diagnosis and the green point corresponds to the time given for the “non-severe” diagnosis

Fig. 6

The false-positive segmentation from an exam with motion artifacts