Abstract
OBJECTIVE: This study aims to develop models for predicting microvascular invasion (MVI) in hepatocellular carcinoma (HCC) patients prior to surgery using two-dimensional (2D) and three-dimensional (3D) radiomics features from contrast-enhanced computed tomography (CT). The study compares the predictive performance of various models and explores the potential of radiomics to capture tumor spatial heterogeneity. MATERIALS AND METHODS: A total of 150 hepatocellular carcinoma (HCC) patients who underwent contrast-enhanced CT examination and curative resection were included in this study. 2D features from the largest cross-sectional slice, as well as 3D radiomic features, were extracted from the non-contrast (NC), arterial phase (AP), portal venous phase (PVP), and balanced phase (BP) images. Feature selection was performed using the least absolute shrinkage and selection operator (LASSO) algorithm, and predictive models were constructed using logistic regression and XGBoost machine learning algorithms. The predictive performance of the models was evaluated using the area under the receiver operating characteristic curve (AUC). RESULTS: The 2D BP model (AUC = 0.801) and 3D PVP model (AUC = 0.876) showed superior performance among single-sequence models. The 2D multi-sequence model (AUC = 0.851) outperformed the 3D combined model (AUC = 0.811). Radiomics-based models outperformed clinical feature-based models, and combining radiomics scores with clinical features improved prediction accuracy. However, 3D models did not significantly outperform 2D models. CONCLUSION: Both 2D and 3D radiomics models are effective for predicting MVI in HCC patients preoperatively. While the 3D model captures spatial heterogeneity, the 2D model excels at capturing local texture features. This study provides new insights into radiomics in HCC, contributing to its clinical application and standardization.