Beyond Accuracy: Pneumonia Severity Grading in Chest X-Rays Using RSNA 2018 Bounding-Box Extent Metadata and ViT

Chest X-ray pneumonia assessment is commonly formulated as a binary classification problem that does not explicitly represent radiographic burden. In this work, we reinterpret the RSNA 2018 Pneumonia Detection Challenge dataset by using expert bounding-box annotations to derive three burden-oriented groups: low/mild (Severity 1), moderate (Severity 2), and severe (Severity 3). We evaluate a classical binary baseline, a unified four-class classifier, and three severity-specific binary specialist branches. The selected binary Vision Transformer achieved 95.07±0.05% maximum validation accuracy, while the unified four-class model reached 79.16±0.63%. To complement exact-match accuracy, we report three task-specific ordinal measures: Severity Consistency Score (SCS), probability-aware Severity Consistency Score (pSCS), and Adjacent-Class Accuracy (ACA). These reached 91.80%, 89.18%, and 96.10%, respectively, indicating that many non-exact predictions remained close to the reference burden category. The specialist branches achieved maximum validation accuracies of 94.59±0.33%, 97.91±0.05%, and 99.03±0.08% for Severity 1, Severity 2, and Severity 3, respectively. At the joint-system level, maximum-probability fusion of the three specialist outputs achieved a Severity Binary Accuracy (SBA) of 98.17±0.17% on the common evaluation cohort. Overall, the results show that RSNA bounding-box extent can support a reproducible, annotation-derived radiographic burden analysis, while ordinal-aware measures provide complementary information to strict four-class accuracy.