Multimodal AI in Hepatocellular Carcinoma: Bridging Diagnosis, Therapy and Prognosis Toward Precision Oncology

This narrative review summarizes recent advances in multimodal AI for hepatocellular carcinoma (HCC), a highly heterogeneous malignancy. Traditional singlemodal data (eg imaging, serology, pathology) are insufficient for accurate diagnosis and personalized treatment. Multimodal technology, which integrates four categories of data (imaging, pathology, molecular, and clinical) through artificial intelligence (AI), has emerged as a promising solution. This review aims to summarize the sources, characteristics, and standardization of multimodal data in HCC, and to comprehensively update its clinical applications in diagnosis, treatment decision-making, and prognosis prediction.

We searched the literature in PubMed, Web of Science, Cochrane Library, and China National Knowledge Infrastructure (CNKI) for articles on multimodal AI applications in HCC published up to 2025. We also critically appraised the included studies based on study design (retrospective vs prospective), validation strategies (internal vs external), sample size, multicenter status, and risk of bias (eg data leakage, overfitting, class imbalance).

Multimodal models have shown promising performance across key clinical tasks, including the diagnosis of early-stage and atypical HCC, preoperative prediction of microvascular invasion (MVI), and assessment of treatment response. In prognostic evaluation, multimodal integration has also demonstrated potential in predicting early recurrence and overall survival. However, challenges remain, including multicenter data heterogeneity, limited model interpretability, and a lack of prospective validation. Notably, no prospective studies or randomized controlled trials have yet been conducted to validate the clinical benefit of multimodal AI in HCC, highlighting a critical gap between current research and clinical implementation.

Multimodal AI shows substantial promise in improving HCC management, including earlier detection, more accurate risk stratification, and individualized treatment planning. However, most current evidence remains retrospective, and prospective validation is needed before routine clinical adoption. Going forward, we should focus on deeper integration across different data types, building standardized datasets, advancing digital twin technology, and facilitating translation of these tools into clinical care to ultimately improve patient outcomes.