CT density–texture radiomics to differentiate mass-like pneumonia from malignancy: an explainable artificial intelligence and machine learning approach

Abstract

Abstract Background Mass-like pulmonary consolidations ≥ 3 cm pose a persistent diagnostic dilemma, as malignancy and pneumonia share overlapping CT features despite fundamentally different management. Although radiomics offers a quantitative approach, most published models lack multivendor external validation, formal radiologist benchmarking, and transparent reproducibility reporting. Objectives To develop and externally validate an explainable CT density–texture radiomics model differentiating mass-like pneumonia from pulmonary malignancy across multivendor CT systems, and to benchmark performance against blinded radiologist interpretation. Methods Consecutive adults with mass-like pulmonary lesions (≥ 3 cm) were prospectively recruited at three centres (January–March 2026). Of 183 assessed, 12 excluded, 171 enrolled. Centre A (n = 71) was the development cohort; Centres B + C (n = 100) the external validation cohort. A fixed 9-mm ROI was placed within the lesion; features were extracted using (PyRadiomics 3.1.0a2). 75 candidate features were provided to an XGBoost classifier. Two blinded radiologists independently reviewed 100 external cases. Results Internal cross-validation AUC was 0.713 ± 0.120 (SD across 100 folds). The locked model achieved external AUC 0.887 (95% CI 0.816–0.948); sensitivity 0.839; specificity 0.841; Brier score 0.135; calibration slope 0.727; O/E ratio 1.076 (0.948–1.228). Decision curve analysis showed positive net benefit from threshold 0.10 to 0.80. In sensitivity analysis, performance was stable under ComBat scanner-batch harmonisation, with no significant difference from the unharmonised model (harmonised AUC 0.801; DeLong p = 0.130). Inter-rater reproducibility was excellent for dominant first-order features (median ICC 0.946). Radiologist AUC significantly higher by DeLong ( p = 0.001–0.002); binary classification errors not significantly different by McNemar ( p = 0.629–1.000). Conclusion In this exploratory multicentre cohort, an explainable CT radiomics model achieved an external AUC of 0.887, with reproducible ROI placement, stable performance across scanner settings in sensitivity analysis, and positive net benefit across a wide range of decision thresholds. Radiologist AUC remained significantly higher. The model is therefore proof-of-concept: a research-stage adjunct intended to support, not replace, expert radiologist interpretation, and one that requires larger independent validation and recalibration before any clinical role can be claimed.