Shap-interpretable predictive modeling of microvascular invasion and early recurrence in hepatocellular carcinoma using MRI habitat imaging combined with clinical features.

OBJECTIVE: To develop and validate an integrated model combining Gd-EOB-DTPA-enhanced MRI habitat imaging with clinical features for preoperative prediction of microvascular invasion (MVI) and early recurrence in hepatocellular carcinoma (HCC). METHODS: This retrospective study enrolled 230 pathologically confirmed HCC patients, classified as MVI-positive or -negative. Radiomics features were extracted from the total tumor volume and a 3-mm peritumoral region. Tumor regions were segmented into three spatial habitats via K-means clustering, and habitat-specific features were obtained. Key features were selected using least absolute shrinkage and selection operator (LASSO) regression. Seven machine learning algorithms were trained; the intratumoral heterogeneity (ITH) score model showed optimal performance. Four models were developed: Clinical, Peritumoral Radiomics, ITH Score, and Combined (ITH + Peritumoral Radiomics + Clinical). Performance was assessed with ROC analysis, calibration, decision curve analysis (DCA), and SHapley Additive exPlanations (SHAP). A prognostic model was developed using the DeepSurv network to assess early recurrence-free survival (RFS) following HCC resection, with Kaplan-Meier curves plotted for evaluation. RESULTS: Extreme Gradient Boosting (XGBoost) achieved the best performance for ITH and traditional radiomics models. The ITH score model outperformed the Peritumoral Radiomics model. The Combined model achieved the highest performance on the training set (AUC: 0.925; sensitivity: 0.926; specificity: 0.785; accuracy: 0.863; F1-score: 0.882). Calibration and DCA confirmed reliability and clinical benefit. SHAP analysis clarified feature contributions.The HCC prognostic model-defined MVI-High risk patients, who exhibited significantly different risk scores, also had a significantly poorer early RFS per Kaplan-Meier analysis (P < 0.001). CONCLUSION: The integrated MRI habitat-clinical model outperformed standalone approaches, showing promise for individualized surgical planning and recurrence risk stratification in HCC.