Radiomics-enhanced machine learning for differentiating epidural and subdural hematomas on non-contrast CT

Abstract

BACKGROUND: Differentiating epidural hematoma (EDH) from subdural hematoma (SDH) on non-contrast CT (NCCT) is clinically important but can be challenging when imaging features overlap. This study evaluated whether radiomic features combined with machine learning could improve discrimination between EDH and SDH. METHODS: In this retrospective study, 175 adults with surgically confirmed EDH or SDH were analyzed. Conventional morphometric CT features and 107 radiomic features extracted from manually segmented hematomas were assessed. After LASSO-based feature selection, machine learning models were developed and evaluated using area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUCPR), and accuracy. RESULTS: Radiomic features showed slightly better discriminative performance than handcrafted morphometric features alone. Among the evaluated models, XGBoost achieved the best overall performance. The highest performance was obtained using the combined feature set of radiomic and morphometric variables, with an AUC of 0.75, an AUCPR of 0.70, and an accuracy of 0.74. CONCLUSIONS: Integrating radiomic and conventional morphometric CT features improves automated differentiation of EDH and SDH. Radiomics-assisted machine learning may serve as a useful decision-support tool in diagnostically challenging cases, although further external validation is needed before clinical implementation.