A Quantitative Lung Computed Tomography Image Feature for Multi-Center Severity Assessment of COVID-19

The COVID-19 pandemic has affected millions and congested healthcare systems globally. Hence an objective severity assessment is crucial in making therapeutic decisions judiciously. Computed Tomography (CT)-scans can provide demarcating features to identify severity of pneumonia —commonly associated with COVID-19—in the affected lungs. Here, a quantitative severity assessing chest CT image feature is demonstrated for COVID-19 patients. An open-source multi-center Italian database¹ was used, among which 60 cases were incorporated in the study (age 27-86, 71% males) from 27 CT imaging centers. Lesions in the form of opacifications, crazy-paving patterns, and consolidations were segmented. The severity determining feature —L_norm was quantified and established to be statistically distinct for the three —mild, moderate, and severe classes (p-value<0.0001). The thresholds of L_norm for a 3-class classification were determined based on the optimum sensitivity/specificity combination from Receiver Operating Characteristic (ROC) analyses. The feature L_norm classified the cases in the three severity categories with 86.88% accuracy. ‘Substantial’ to ‘almost-perfect’ intra-rater and inter-rater agreements were achieved involving expert and non-expert based evaluations (κ-score 0.79-0.97). We trained machine learning based classification models and showed L_norm alone has a superior diagnostic accuracy over standard image intensity and texture features. Classification accuracy was further increased when L_norm was used for 2-class classification i.e. to delineate the severe cases from non-severe ones with a high sensitivity (97.7%), and specificity (97.49%). Therefore, key highlights of this severity assessment feature are accuracy, lower dependency on expert availability, and wide utility across different imaging centers.