Two-center PCD-CT spectral radiomics in 378 patients: VNC-trained model separates benign from malignant liver lesions at AUC 0.899, with high-keV and VNC test reconstructions most stable
In a two-center study of 378 patients with focal hepatic lesions (190 benign, 227 lesions; 188 malignant, 2681 lesions), radiomic features were extracted from nine PCD-CT spectral datasets (40, 50, 70, 90, 110, 140, 180 keV, virtual non-contrast [VNC], and iodine density maps [IDM]) after nnU-Net segmentation confirmed by an abdominal radiologist, multiple machine-learning models were trained with lesion-level predictions aggregated per patient, and the Random Forest model trained on VNC achieved the highest performance (AUC 0.899; 95% CI 0.840-0.951; accuracy 83.5%), while cross-domain analysis showed similar mean AUCs across training reconstructions (0.75 to 0.83) and test reconstruction choice mattered: 110, 140, 180 keV and VNC were most stable (median AUC >= 0.