Abstract / Summary
Accurate identification of aggressive tumor phenotypes remains a major challenge in hepatocellular carcinoma (HCC). Here, we developed and validated MAVEN (Multimodal Automated VETC Estimation Network), a fully automated multimodal deep learning system integrating magnetic resonance imaging (MRI), whole-slide histopathology images (WSIs), and clinical variables for predicting vessels encapsulating tumor clusters (VETC) and early recurrence risk. In this multicenter study including 1928 patients from five institutions, MAVEN demonstrated superior performance compared with unimodal and bimodal models, achieving an area under the curve (AUC) of 0.932 in the internal test cohort and 0.879-0.891 across four independent external cohorts. MAVEN-based risk stratification was significantly associated with early recurrence-free survival across all cohorts. Model explainability analyses revealed that both radiologic and histopathologic features contributed to prediction, with consistent spatial localization of high-risk regions. These findings suggest that multimodal integration enables robust and generalizable prediction of tumor aggressiveness and postoperative risk stratification.