Abstract
BACKGROUND: Microvascular invasion (MVI) is a critical determinant of poor prognosis in hepatocellular carcinoma (HCC). Accurate preoperative prediction of MVI is essential for optimizing surgical and therapeutic strategies. This study aims to develop a combined model integrating intratumoral, peritumoral, and clinical features from ultrasound-based radiomics for MVI prediction. METHODS: Ultrasound images of 119 patients with pathologically confirmed HCC were analyzed. A total of 1,414 radiomics features were extracted from intratumoral and peritumoral regions. Feature selection was performed using intraclass correlation coefficient (ICC) analysis, t-tests, and least absolute shrinkage and selection operator (LASSO) regression. Logistic regression, Random Forest, and other machine learning algorithms were applied to construct predictive models. The best-performing intratumoral, peritumoral, and clinical models were combined using logistic regression. SHapley Additive exPlanations (SHAP) analysis, logistic regression coefficients, and partial dependence analysis were employed to evaluate feature contributions and interactions. RESULTS: Both intratumoral and peritumoral models achieved high AUCs (0.781 and 0.792, respectively), with no statistically significant difference between them. The combined model, incorporating tumor size, achieved the highest AUC (0.903, 95% CI: 0.780-1.000) and superior performance across all evaluation metrics. Tumor size exhibited the smallest logistic regression coefficient but the highest SHAP contribution, indicating strong interactions with intratumoral and peritumoral features. Interaction analyses revealed that the combined effects of tumor size and radiomics features significantly enhanced predictive performance. CONCLUSION: This study demonstrates that combining intratumoral, peritumoral, and clinical features enhances the predictive accuracy for MVI in HCC. The findings underscore the value of feature integration and interactions, providing insights for personalized treatment planning and advancing the clinical utility of ultrasound-based radiomics.