Objective:

To evaluate the predictive performance of pre-treatment contrast-enhanced MRI habitat radiomics and deep learning models for identifying vessels encapsulating tumor clusters (VETC) in hepatocellular carcinoma (HCC) and to characterize the associated immune infiltration patterns.

Approach:

• Model Development: Extraction of habitat and deep learning features, followed by LASSO for feature selection to construct intratumoral heterogeneity (ITH) and deep learning (DL) models, which were then integrated into a fusion model to enhance predictive capability.

Key Findings:

• ITH model AUCs: 0.845 (training), 0.806 (validation).
• DL model AUCs: 0.764 (training), 0.745 (validation).
• Fusion model AUCs: 0.901 (training), 0.870 (validation).
• Significant differential gene expression between high-risk and low-risk groups.
• Low-risk group enriched in pathways related to cell cycle, translation, and mitochondrial function.
• High-risk group showed a significant reduction in resting dendritic cells (P < 0.05).

Interpretation:

The CEMRI-based fusion model effectively predicts VETC in HCC and correlates with immune-related transcriptomic and infiltration profiles.

Limitations:

• Retrospective study design may introduce bias.
• Single-center study limits generalizability.
• Further validation in larger cohorts needed.

Conclusion:

A CEMRI-based fusion model integrating ITH and DL features enables accurate prediction of VETC in HCC.