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

A radiomics-based machine-learning pipeline built on PCD-CT multispectral reconstructions differentiated benign from malignant hepatic lesions, with the Random Forest model trained on VNC achieving the highest performance at AUC 0.899 (95% CI 0.840-0.951) and accuracy 83.5%. Noninvasive liver lesion classification has largely relied on conventional CT; this work pairs PCD-CT spectral capability (virtual monochromatic and material decomposition images) with radiomics machine learning and reports patient-based classification. A 378-patient cohort from 2 institutions, lesions segmented automatically by nnU-Net and confirmed by an abdominal radiologist, patient-based analysis with lesion-level predictions aggregated per patient, and reported 95% confidence intervals for AUC alongside accuracy.

Cross-domain analysis showed similar mean AUCs across training reconstructions (0.75 to 0.83), but evaluation performance varied by test reconstruction: 110, 140, 180 keV and VNC were most stable (median AUC >= 0.83), while 40 keV, 50 keV, and IDM yielded lower and more variable AUCs (median AUC <= 0.71). Rather than reporting a single best-reconstruction result, the study systematically compared all train-test reconstruction combinations and framed the choice of test reconstruction as an independent factor in performance. Full combinatorial cross-evaluation across nine spectral datasets (40-180 keV, VNC, IDM), using median AUC to characterize stability and separating the roles of training-side and test-side reconstructions.

The authors propose that incorporating high-energy reconstructions and VNC images into radiomic workflows for liver lesion risk stratification may enable noninvasive triage and reduce the need for invasive diagnostics. Translates a technical reconstruction-choice finding into a clinical workflow recommendation aimed at a specific application, noninvasive triage. This is the authors' concluding outlook based on the reported classification performance and stability results; the text reports no prospective validation of a triage pathway or of reduced invasive testing.

The results apply to patients with focal hepatic lesions scanned on PCD-CT at 2 institutions, using a pipeline that depends on nnU-Net automatic segmentation plus radiologist confirmation and on nine specific spectral reconstructions (40-180 keV, VNC, IDM). For radiology departments and researchers building liver lesion radiomics classification pipelines, the work suggests favoring high-energy reconstructions and VNC images as the test domain, and indicates that the choice of training reconstruction has comparatively less influence. The proposed noninvasive triage and reduced need for invasive diagnostics are application directions that require further evaluation within clinical pathways.

Readers should still watch: how to interpret the gap between the mean AUC range across training reconstructions in cross-domain analysis (0.75-0.83) and the AUC 0.899 of the best VNC model; whether the lower and more variable performance of 40 keV, 50 keV, and IDM relates to noise in low-energy images or to how iodine density maps represent features; how patient-level aggregation affects performance given the marked difference in lesion counts between benign and malignant groups (227 versus 2681); and the reproducibility of the high-energy and VNC stability finding at other institutions, with other scan protocols, and in different populations. In addition, this is a fast-parse text without figures or supplementary material, so feature-selection details, model hyperparameters, and the full per-reconstruction performance tables cannot be checked.