Pre-treatment CEMRI habitat radiomics and deep learning for noninvasive prediction of the VETC pattern in hepatocellular carcinoma: an exploratory radiogenomic analysis.

Background: Vessels encapsulating tumor clusters (VETC) constitute an aggressive morphological subtype of hepatocellular carcinoma (HCC), characterized by marked angiogenesis and immune infiltration, and can potentially be identified noninvasively using radiomics. Objective: This study aimed to evaluate the predictive performance of pre-treatment contrast-enhanced MRI (CEMRI) habitat radiomics and deep learning models for identifying VETC in HCC, and to characterize the associated immune infiltration patterns. Methods: We retrospectively analyzed CEMRI scans from 336 patients with HCC (336 lesions). Habitat features and deep learning (DL) features were extracted from the CEMRI scans, and the LASSO was applied for feature selection to construct an intratumoral heterogeneity (ITH) model and a deep learning (DL) model. Robust ITH and DL features were subsequently integrated to develop a fusion model with enhanced predictive capability, and model performance was assessed using the area under the receiver operating characteristic curve (AUC). Using radiomics features derived from the fusion model, 47 HCC patients with baseline liver MRI scans in The Cancer Imaging Archive (TCIA) and matched RNA-seq profiles in The Cancer Genome Atlas (TCGA) were stratified into a high-risk group (n = 23) and a low-risk group (n = 24). Transcriptomic profiles were then analyzed using gene set variation analysis (GSVA) to compare immune-related pathway activity between groups. Differential gene expression analysis (DGEA), gene set enrichment analysis (GSEA), and cell type identification by estimating relative subsets of RNA transcripts (CIBERSORT)-based immune infiltration analysis were further performed to elucidate potential differences in the tumor immune microenvironment and their biological implications. Results: The ITH model achieved AUCs of 0.845 and 0.806 in the training and validation cohorts, respectively, whereas the DL model yielded AUCs of 0.764 and 0.745. The fusion model demonstrated superior performance, with AUCs of 0.901 and 0.870 in the respective cohorts. Transcriptomic analyses identified significant differential gene expression between the high- and low-risk groups. GSEA indicated that the low-risk group was enriched in pathways related to the cell cycle, translation, and mitochondrial function. Immune profiling revealed a significant reduction in resting dendritic cells in the high-risk group ( Conclusions: A CEMRI-based fusion model integrating ITH and DL features enables accurate prediction of VETC in HCC and shows preliminary associations with immune-related transcriptomic and infiltration profiles linked to this vascular pattern.