Prognostic value of the extent of affected lung parenchyma in COVID-19 pneumonia patients: Visual estimation versus automatic quantification by artificial intelligence

Abstract

Objective

To compare the prognostic impact of the extent of lung disease detected on computed tomography (CT) when assessed visually by an expert radiologist compared to automatically by an artificial intelligence (AI) system in patients with COVID-19 pneumonia.

Material and methods

A retrospective study of patients with clinical suspicion of COVID-19 pneumonia which assessed the extent of lung involvement visually and by AI. Patients were divided into poor (death/ICU) and good (discharge) prognosis groups. Univariate and multivariate analyses (logistic regression) were performed on the variables that demonstrated significant differences between both groups.

Results

Patients with a poor prognosis more frequently had greater lung involvement visually (stages 3–4; 37.5% vs 14.3%; p = 0.001) and by AI (stages 3–4; 35% vs 6.2%; p < 0.001). The radiologist-AI agreement correlation coefficient was excellent (0.905; p < 0.001). High blood pressure (OR 4.26; p < 0.001), alterations in levels of creatinine (OR 5.63; p < 0.001), lactate dehydrogenase (OR 11.69; p < 0.001) and D-dimer (OR 5.68; p < 0.001), and the extent of affected lung parenchyma assessed visually (stage 1vs4 OR 10.36; p = 0.001) and by AI (stage 1vs4 OR 25; p = 0.001) were the variables with the greatest prognostic impact in the univariate analysis. The multivariate analysis models considering the extent assessed visually and by AI did not demonstrate any significant differences (AUC 0.876 vs 0.870; p = 0.278).

Conclusion

The extent of affected lung parenchyma on CT images demonstrates prognostic value both on their own and in conjunction with clinical factors and blood levels in patients with COVID-19 pneumonia. No significant differences were observed between the radiologist's visual estimate and the AI-based automatic detection system used in this study.