Diagnostic accuracy and inter-observer agreement with the CO-RADS lexicon for CT chest reporting in COVID-19

Abstract

Purpose: To measure the diagnostic accuracy and inter-observer agreement with the use of COVID-19 Reporting and Data System (CO-RADS) for detection of COVID-19 on CT chest imaging.

Methods: This retrospective study included 164 consecutive patients with clinical suspicion of COVID-19 in whom a CT chest examination was performed at a single institution between April 2020 and July 2020. Of them, 101 patients was RT-PCR positive for COVID-19. Six readers with varying radiological experience (two each of chest radiologists, general radiologists, and radiologists in training) independently assigned a CO-RADS assessment category for each CT chest study. The Fleiss' K was used to quantify inter-observer agreement. The inter-observer agreement was also assessed based on the duration of onset of symptoms to CT scan. ROC curve analysis was used to determine the diagnostic accuracy of CO-RADS. The area under curve was calculated to determine the reader accuracy for detection of COVID-19 lung involvement with RT-PCR as reference standards. The data sets were plotted in ROC space, and Youden's J statistic was calculated to determine the threshold cut-off CO-RADS category for COVID-19 positivity.

Results: There was overall moderate inter-observer agreement between all readers (Fleiss' K 0.54 [95% CI 0.54, 0.54]), with substantial agreement among chest radiologists (Fleiss' K 0.68 [95% CI 0.67, 0.68]), general radiologists (Fleiss' K 0.61 [95% CI 0.61, 0.61]), and moderate agreement among radiologists-in-training (Fleiss' K 0.56 [95% CI 0.56, 0.56]). There was overall moderate inter-observer agreement in early disease (stages 1 and 2), with cumulative Fleiss' K 0.45 [95% CI 0.45, 0.45]). The overall AUC for CO-RADS lexicon scheme to accurately diagnose COVID-19 yielded 0.92 (95% CI 0.91, 0.94) with strong concordance within and between groups, of chests radiologists with AUC of 0.91 (95% CI 0.88, 0.94), general radiologists with AUC 0.96 (95% CI 0.94, 0.98), and radiologists in training with AUC of 0.90 (95% CI 0.87, 0.94). For detecting COVID-19, ROC curve analysis yielded CO-RADS > 3 as the cut-off threshold with sensitivity 90% (95% CI 0.88, 0.93), and specificity of 87% (95% CI 0.83, 0.91).

Conclusion: Readers across different levels of experience could accurately identify COVID-19 positive patients using the CO-RADS lexicon with moderate inter-observer agreement and high diagnostic accuracy.

Keywords: CO-RADS; COVID-19; SARS-CoV-2.

Fig. 1

Pictorial representation of CT chest studies illustrating typical characteristics of different CO-RADS categories from our study sample. COVID-19, Corona Virus disease 2019; GGO, ground glass opacities

Fig. 2

Flow diagram of patient inclusion and exclusion in the present study. COVID-19, Corona Virus disease 2019; RT-PCR +, reverse transcriptase-polymerase chain reaction positive; RT-PCR−, reverse transcriptase-polymerase chain reaction negative

Fig. 3

Receiver operating characteristic curve a AUC for all readers obtained by plotting of all datasets in ROC space; b AUC comparison for different experience groups. AUC-Area under curve