An Interpretable MRI-Based Machine Learning Model for Preoperative Identification of the Vascular Dissemination Phenotype in Hepatocellular Carcinoma: A Dual-Center Study
Source: PubMed Central Open Access, NCBI / U.S. National Library of Medicine
Objective To develop and externally validate an interpretable MRI-based machine-learning model for preoperative identification of the vascular dissemination phenotype in hepatocellular carcinoma (HCC), defined by vessels encapsulating tumor clusters (VETC) and/or microvascular invasion (MVI). Materials and Methods This dual-center retrospective study included 642 patients with surgically confirmed HCC who underwent preoperative contrast-enhanced MRI. Patients from Institution I (n = 435) and Institution II (n = 207) formed the training and independent external validation cohorts, respectively. Clinical and conventional MRI predictors were selected using univariable logistic regression, collinearity assessment, and recursive feature elimination. Nine machine-learning models were developed and externally validated. Performance was assessed using area under the curve (AUC), sensitivity, specificity, calibration, decision-curve analysis, and subgroup analyses. SHAP was used for model interpretation, and transcriptomic analysis was performed in 30 patients. Results Among the nine machine-learning models, XGBoost achieved the highest observed AUCs in the training cohort (0.85; 95% CI: 0.81, 0.88) and external validation cohort (0.82; 95% CI: 0.75, 0.88), with higher AUCs than logistic regression in both cohorts (< 0.001 and= 0.047, respectively). In external validation, sensitivity and specificity were 71.1% and 85.5%, respectively. Calibration and decision-curve analyses supported
Abstract
Objective To develop and externally validate an interpretable MRI-based machine-learning model for preoperative identification of the vascular dissemination phenotype in hepatocellular carcinoma (HCC), defined by vessels encapsulating tumor clusters (VETC) and/or microvascular invasion (MVI). Materials and Methods This dual-center retrospective study included 642 patients with surgically confirmed HCC who underwent preoperative contrast-enhanced MRI. Patients from Institution I (n = 435) and Institution II (n = 207) formed the training and independent external validation cohorts, respectively. Clinical and conventional MRI predictors were selected using univariable logistic regression, collinearity assessment, and recursive feature elimination. Nine machine-learning models were developed and externally validated. Performance was assessed using area under the curve (AUC), sensitivity, specificity, calibration, decision-curve analysis, and subgroup analyses. SHAP was used for model interpretation, and transcriptomic analysis was performed in 30 patients. Results Among the nine machine-learning models, XGBoost achieved the highest observed AUCs in the training cohort (0.85; 95% CI: 0.81, 0.88) and external validation cohort (0.82; 95% CI: 0.75, 0.88), with higher AUCs than logistic regression in both cohorts (< 0.001 and= 0.047, respectively). In external validation, sensitivity and specificity were 71.1% and 85.5%, respectively. Calibration and decision-curve analyses supported model performance. Subgroup discrimination remained acceptable for tumors ≤ 5.0 cm and BCLC stage 0 or A disease, although sensitivity was lower for tumors ≤ 5.0 cm. SHAP identified intratumoral artery, nonsimple nodular growth type, and necrosis or severe ischemia as leading contributors. Transcriptomic analysis suggested exploratory enrichment of cell-cycle and metabolic pathways. Conclusion The interpretable MRI-based XGBoost model showed favorable performance for identifying the vascular dissemination phenotype in HCC, with SHAP-based interpretability and exploratory transcriptomic context.