CALM-CXR: a calibration-aware lung-masked workflow for hierarchical chest X-ray classification

Chest X-ray classification studies often report endpoint performance without isolating how anatomical guidance, staged decisions, probability calibration, and error propagation influence the final prediction. CALM-CXR is a reproducible protocol that operationally couples lung-mask-conditioned regional expert modeling with separately trained normal-abnormal and bacterial-viral stages, stage-specific temperature scaling, calibration-partition threshold selection, and coherent three-class probability composition. Its methodological advance is the controlled, auditable coupling of these established components: each stage retains its own probability and decision parameters, final class probabilities remain mathematically coherent, and errors can be attributed to screening or subtype classification. The protocol was evaluated on 4,840 images using filename-derived group-disjoint train, validation, calibration, and test partitions. Across seeds 42, 52, and 62, the full protocol achieved accuracy 0.7920 (SD 0.0248), macro-F1 0.7999 (SD 0.0165), macro one-vs-rest AUROC 0.9063 (SD 0.0088), and macro PR-AUC 0.8616 (SD 0.0068). Removing lung-mask conditioning produced the largest mean decrease in accuracy (0.0234) and macro-F1 (0.0206) among the component ablations. A matched flat ConvNeXt-Tiny classifier achieved accuracy 0.7955 (SD 0.0135) and macro-F1 0.7982 (SD 0.0128), and paired comparisons did not establish consistent overall superiority of the hierarchical organization. CALM-CXR should therefore be interpreted as an auditable evaluation protocol, not as a new individual primitive or a clinically validated diagnostic system, independent external validation remains necessary. ➢ Defines an auditable two-stage evaluation protocol that separates normal-abnormal screening from bacterial-viral subtyping and retains stage-specific probabilities and errors. ➢ Tests lung-mask conditioning, learned regional routing, anatomy and routing-entropy penalties, temperature scaling, calibration-only thresholds, and hierarchical probability composition through controlled ablations. ➢ Includes matched flat three-class baselines, paired statistical comparisons, calibration metrics, bootstrap uncertainty, and filename-derived group-overlap auditing.