AUTHOR=Baldi Dario , Punzo Bruna , Colacino Andrea , Aiello Marco , Franzese Monica , La Grutta Ludovico , Saba Luca , Bossone Eduardo , Passaro Emanuela , Mantini Cesare , Cavaliere Carlo , Berti Sergio , Celi Simona , Meloni Antonella , Cademartiri Filippo , Maffei Erica 

TITLE=Severity assessment of COVID-19 disease: radiological visual score versus automated quantitative CT parameters using a pneumonia analysis algorithm

JOURNAL=Frontiers in Medicine

VOLUME=Volume 12 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/medicine/articles/10.3389/fmed.2025.1606771

DOI=10.3389/fmed.2025.1606771

ISSN=2296-858X

ABSTRACT=ObjectiveEvaluate the impact of the use of an AI-assisted quantitative tool for assessing stratification of patients with acute lung involvement from coronavirus (COVID-19) compared to a semi-quantitative visual score made by the radiologist.MethodsWe retrospectively enrolled 611 patients with respiratory distress and suspected pneumonia admitted between 27 February and 27 April 2020. Demographic, imaging, and clinical data were collected. Lung involvement was visually assessed using a 5-class severity scale and compared with automated AI-based CT analysis (CT Pneumonia Analysis 2.5.2, SyngoVIA Siemens), which quantified volume and density of alterations. Patients were assigned to severity classes for concordance analysis. Subgroup analysis across biweekly periods assessed changes in visual rater performance. Correlation with SpO₂ and diagnostic performance (accuracy, sensitivity, specificity, AUC) of both methods in predicting RT-PCR results were evaluated.ResultsHigh concordance was found between visual and quantitative assessments (k = 0.73, p < 0.001), with most discordances in low-severity (classes 0–1, k = 0.71), while agreement was excellent for higher severity (classes 2–4, k = 0.91). Misclassifications were mainly for mild cases; concordance was strong in severe, life-threatening presentations. Temporal analysis showed a progressive improvement in agreement over time (k = 0.62, 0.61, 0.54, 0.73). A mild but significant negative correlation emerged between quantitative assessment and SpO₂ values (r = −0.13, p = 0.02). Diagnostic performance between the two methods was similar: visual (AUC = 0.55, Accuracy = 44%, Sensitivity = 27%, Specificity = 78%) and quantitative (AUC = 0.56, Accuracy = 45%, Sensitivity = 27%, Specificity = 79%). Neither method showed strong predictive power for RT-PCR COVID-19 positivity. Nonetheless, assessing lung involvement remains essential for managing respiratory distress, regardless of confirmed infection, particularly for identifying patients with >25% parenchymal involvement who may require hospitalization.ConclusionVisual and AI-based CT assessments showed high concordance and similar accuracy, especially in patients with >25% lung involvement. This study demonstrates the utility of AI-based algorithms to improve the diagnostic efficiency and the reliability, highlighting their value in routine COVID-19 pneumonia evaluation and management.