Abstract
Hepatocellular carcinoma (HCC) shows marked spatial heterogeneity, limiting biopsy-based Edmondson-Steiner (ES) grading. We developed a multicenter radiogenomic framework to noninvasively predict ES grade and explore underlying molecular mechanisms. Arterial-phase DCE-MRI from 295 patients and The Cancer Imaging Archive (TCIA) cases were analyzed using three tumor regions (body, edge, and out). An integrated volume-of-interest (VOI) random forest (RF) model was trained with selected features and externally validated. Radiogenomic analysis correlated radscore with TCIA transcriptomic profiles using weighted gene co-expression network analysis (WGCNA). The model achieved high discrimination (area under the curve [AUC] 0.959 internally; 0.860 externally). Radscore-associated modules revealed ribosomal dysregulation and immune exhaustion. A derived prognostic signature stratified and The Cancer Genome Atlas (TCGA) patients into distinct risk groups and independently predicted survival (hazard ratio [HR] 3.95, p < 0.0001; C index 0.643). This integrated radiogenomic approach enables noninvasive ES grading and provides insight into biologically relevant tumor heterogeneity.