Comparing Visual Scoring of Lung Injury with a Quantifying AI-Based Scoring in Patients with COVID-19

Biebau, Charlotte; Dubbeldam, Adriana; Cockmartin, Lesley; Coudyze, Walter; Coolen, Johan; Verschakelen, Johny; De Wever, Walter

OBJECTIVES: Fast diagnosis of Coronavirus Disease 2019 (COVID-19), and the detection of high-risk patients are crucial but challenging in the pandemic outbreak. The aim of this study was to evaluate if deep learning-based software correlates well with the generally accepted visual-based scoring for quantification of the lung injury to help radiologist in triage and monitoring of COVID-19 patients. MATERIALS AND METHODS: In this retrospective study, the lobar analysis of lung opacities (% opacities) by means of a prototype deep learning artificial intelligence (AI)-based software was compared to visual scoring. The visual scoring system used five categories (0: 0%, 1: 0–5%, 2: 5–25%, 3: 25–50%, 4: 50–75% and 5: >75% involvement). The total visual lung injury was obtained by the sum of the estimated grade of involvement of each lobe and divided by five. RESULTS: The dataset consisted of 182 consecutive confirmed COVID-19 positive patients with a median age of 65 ± 16 years, including 110 (60%) men and 72 (40%) women. There was a correlation coefficient of 0.89 (p < 0.001) between the visual and the AI-based estimates of the severity of lung injury. CONCLUSION: The study indicates a very good correlation between the visual scoring and AI-based estimates of lung injury in COVID-19.