Answers are grounded in this paper and cite their sources.
To evaluate the chest CT features of COVID-19 pneumonia and non-COVID viral pneumonias and assess the diagnostic efficacy of radiologists and an AI-based system in distinguishing between the two. This retrospective study included 181 patients (101 COVID-19 and 80 non-COVID). Chest CT images were independently assessed by two radiologists blinded to clinical and PCR data. CT examinations were also analyzed by a deep-learning AI algorithm. CT features were compared between groups using univariate and multivariate logistic regression. Agreement between the two observers and between the AI and COVID-19 RT-PCR test results was calculated using kappa statistics. Age and sex showed no statistically significant differences between groups. Peripheral distribution, nodular ground-glass opacities and a crazy-paving pattern were significantly more prevalent in COVID-19 cases ( p < 0.001). Of the 181 CT examinations, 169 were included in the AI analysis after excluding 12 studies (6.6%) due to motion artifacts affecting > 20% of lung volume. The AI algorithm achieved a sensitivity of 60% and specificity of 83.8% (κ = 0.414). Interobserver agreement between radiologists was almost perfect (κ = 0.834), and both radiologists exhibited substantial agreement with the PCR results (κ = 0.777 and 0.700). Presence of peripheral distribution, nodular GGO and crazy-paving pattern as well as a lack of lobar consolidation, bronchial wall thickening, upper lobe predominance, and tree-in-bud nodules may favor COVID-19 pneumonia with a sensitivity of 91% and specificity of 88%. Several odds ratios in the multivariable analysis had wide confidence intervals, reflecting limited precision due to the modest sample size. Indicative CT characteristics, including peripheral distribution, nodular ground-glass opacities, and crazy paving, may assist in differentiating COVID-19 from non-COVID viral pneumonias. While AI demonstrated potential as a supportive triage tool, the radiologists achieved higher diagnostic accuracy and consistency. These findings support the integration of AI as a complementary decision-support system rather than a standalone diagnostic tool in future pandemic preparedness efforts. All performance estimates derive from internal, single-center validation and should be regarded as exploratory pending external validation.
No takes yet. Share an insight, caveat, or question.
Çoruh et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: