Abstract / Summary
Background: Lung Ultrasound (LUS) is increasingly used for pneumonia severity assessment due to its accessibility, cost-effectiveness, and diagnostic accuracy. However, most existing deep learning approaches formulate severity estimation as a nominal classification problem, neglecting the inherent ordinal progression of pulmonary involvement. This limitation may lead to clinically significant misclassification errors, particularly between distant severity levels. To address this gap, we investigate whether explicitly modeling the ordinal structure of LUS severity scores can improve diagnostic reliability. Methods: Using the Italian COVID-19 Lung Ultrasound Database (ICLUS-DB), a multicentric dataset comprising four ordered severity levels, we propose an ordinal learning framework based on a ResNet18 backbone coupled with a Cumulative Link Model (CLM) and optimized using the Quadratic Weighted Kappa (QWK) loss. We further introduce a dedicated patient-wise cross-validation protocol and systematically compare the proposed approach against nominal and alternative ordinal learning strategies. Results & Conclusions: The findings from our investigation highlight the advantages of ordinal classification for LUS severity assessment. Among the evaluated approaches, the proposed CLM framework achieved the best ordinal performance, reaching a QWK of 0.739, a Spearman correlation coefficient of 0.750, and a 1-Off Accuracy of 0.967. These results demonstrate an improved ability to preserve severity ordering and reduce clinically relevant misclassification errors compared with its nominal counterpart. In addition to frame-level analysis, we extended the evaluation to the video level to better reflect clinical practice, further supporting the robustness and applicability of the proposed framework.