Decoding Tumor Heterogeneity with Imaging Biomarkers Predicts Microvascular Invasion and Postoperative Risk Stratification in Hepatocellular Carcinoma

Acad Radiol. 2026 Aug 12:S1076-6332(26)00576-3. doi: 10.1016/j.acra.2026.07.068. Online ahead of print.

ABSTRACT

RATIONALE AND OBJECTIVES: Microvascular invasion (MVI) is a major driver of recurrence and poor outcomes in hepatocellular carcinoma (HCC), yet biopsy-based pathologic assessment is invasive and susceptible to sampling bias and delayed availability. We aimed to develop and validate a noninvasive preoperative model integrating multiphase magnetic resonance imaging-derived habitat/intratumoral heterogeneity (ITH) features and clinical variables for MVI prediction and postoperative recurrence-risk stratification.

METHODS: We retrospectively analyzed 978 patients with HCC from four tertiary centers. Seven models were constructed: four unimodal, two bimodal, and one trimodal (Decision). Model performance was assessed using area under the curve (AUC), sensitivity, specificity, calibration, and decision curve analysis in the internal and external validation cohorts. Prognostic value for recurrence-free survival (RFS) was evaluated using the concordance index (C-index), time-dependent receiver operating characteristic analysis, and Kaplan-Meier analysis, and compared with clinicopathologic risk stratification. Subgroup analyses stratified by hepatitis B virus (HBV) status and histologic differentiation were used to assess robustness.

RESULTS: The decision model achieved AUCs of 0.912, 0.875, and 0.882 in the training, internal validation, and external validation cohorts, respectively, and significantly outperformed the clinical model, all unimodal models, and the Rad-Clinical model (p < 0.001). Model ablation analyses demonstrated incremental gains from incorporating clinical variables with imaging features and from jointly modeling subregional composition using habitat phenotyping and ITH metrics. In the external validation cohort, the model outperformed the clinicopathologic model for recurrence-risk stratification, with a higher C-index (0.776 vs. 0.702) and higher AUCs at 2 years (0.810 vs. 0.729) and 5 years (0.876 vs. 0.779). Performance was consistent across HBV- and histologic grade-based subgroups. SHapley Additive exPlanations analysis supported model interpretability.

CONCLUSION: The preoperative decision model demonstrated robust performance for MVI prediction and postoperative recurrence-risk stratification, supporting its potential role as an adjunct tool for preoperative risk assessment.

PMID:42586894 | DOI:10.1016/j.acra.2026.07.068