Multiphasic CT Delta Radiomics for Four-Class Prediction of Combined Vetc/Mvi Phenotypes In Hepatocellular Carcinoma
Abstract
Purpose
Preoperative phenotyping of vessels that encapsulate tumor clusters (VETC) and microvascular invasion (MVI) is clinically important in hepatocellular carcinoma (HCC) diagnosis and treatment. While most studies rely on MRI, CT-based prediction remains limited despite multiphasic contrast-enhanced CT being widely available. This study aims to develop and evaluate multiphase CT radiomics models for four-class prediction of combined VETC/MVI phenotypes using single-phase (arterial [AP], delayed [DP], portal venous [PVP]) or inter-phase delta features.
Methods
This retrospective study included 366 patients with HCC: VETC-/MVI- (148), VETC+/MVI+ (82), VETC-/MVI+ (61), and VETC+/MVI- (75). Triphasic CT images and tumor masks were resampled and registered. PyRadiomics features were extracted with fixed discretization (binWidth=25). Models were trained using single-phase features (AP, DP, PVP) and delta features (AP-DP, AP-PVP, DP-PVP). Feature selection followed a five-step pipeline with six method combinations, comprising SIS, within-category correlation filtering or VIF, relevance confirmation, and embedded selection using RF/ElasticNet/Lasso. Class imbalance was addressed using class-weighted loss. Ten types of classifiers (e.g., SVM, RF, LR, XGBoost) were evaluated using class-wise and macro AUC. The dataset was split into 90% for stratified five-fold cross-validation and 10% for independent testing.
Results
Delta radiomics outperformed single-phase models for the four-class VETC/MVI prediction tasks. The optional delta radiomics model (DP-PVP; VIF-RF; XGBoost) achieved a final macro AUC of 0.847, with the strongest class-wise discrimination for MVI+/VETC- (AUC=0.957) and VETC+/MVI+ (AUC=0.905). Among single-phase models, AP showed the highest macro AUC of 0.764 (Correlation-LASSO; LDA), whereas DP and PVP both achieved macro AUC of 0.745 (VIF-RF; Random Forest).
Conclusion
Inter-phase delta radiomics improved four-class VETC/MVI classification task performance compared to single-phase features, suggesting that cross-phase image-based biomarkers may capture additional vascular invasion-related information. Class-wise performance showed some heterogeneity, suggesting differentiated separability across phenotypes. Multi-institutional validation and robustness analyses across acquisition settings are warranted to enhance generalizability.