Multimodal CT for Predicting Microvascular Invasion in Solitary cHCC-CCA: Dual-Center External Validation

Preoperative prediction of microvascular invasion (MVI) in combined hepatocellular-cholangiocarcinoma (cHCC-CCA) remains difficult, and externally validated CT-based tools are scarce. To develop and externally validate a multimodal model for MVI prediction in solitary cHCC-CCA using portal venous-phase quantitative CT features combined with multiphasic CT semantic features, and to compare intratumoral, 10-mm peritumoral, and combined intratumoral + peritumoral segmentation strategies.

This retrospective dual-center study included 184 patients with pathologically confirmed solitary cHCC-CCA who underwent contrast-enhanced CT within 1 month before resection (Center 1: n = 139; Center 2: n = 45). Center 1 was randomly split into training (n = 97) and internal test (n = 42) cohorts, and Center 2 served as an independent external validation cohort. Clinical variables and CT semantic features were assessed on multiphasic CT, whereas radiomics and deep learning features were extracted exclusively from portal venous-phase images using three segmentation strategies (intratumoral, 10-mm peritumoral, and combined intratumoral + peritumoral). Feature selection, hyperparameter tuning, and calibration were performed in the training cohort. Operating thresholds were selected using training-cohort out-of-fold (OOF) predictions to target a sensitivity around 0.80 and were then fixed for external validation. Model performance was evaluated in the external cohort using area under the receiver operating characteristic curve (AUC; 95% CI), calibration, and decision curve analysis, with SHapley Additive exPlanations (SHAP) used for interpretability.

Interobserver agreement for key semantic features was almost perfect (κ = 0.81-0.84), and overall semantic agreement was high (mean κ = 0.84). Radiomics and deep learning features showed good reproducibility (ICCs > 0.80). In external validation, discrimination was moderate and consistent across segmentation strategies (AUC, 0.761-0.800). The 10-mm peritumoral strategy achieved the numerically highest AUC (0.800; 95% CI: 0.658-0.916) and high sensitivity at a prespecified sensitivity-oriented operating threshold; differences versus other strategies were modest and not statistically significant (all P > 0.05). SHAP analyses consistently highlighted rim arterial-phase hyperenhancement and widened perilesional enhancement as major contributors to MVI-positive predictions.

A multimodal approach combining portal venous-phase quantitative CT features with multiphasic semantic CT features enabled externally validated preoperative MVI risk estimation in solitary cHCC-CCA. Peritumoral modeling showed a consistent but modest numerical advantage without statistically proven superiority. Findings are limited by retrospective design, small external cohort size, and restricted population diversity; prospective multicenter validation is warranted to confirm its role as a decision-support tool for preoperative MVI risk stratification.