From black-box prediction to transparent insight: the status quo and paradigm shift of explainable artificial intelligence in hepatocellular carcinoma research

Abstract

Hepatocellular carcinoma (HCC) is a malignancy with high global incidence and mortality, whose significant heterogeneity and poor prognosis pose severe clinical challenges. While artificial intelligence (AI) shows potential in HCC imaging, pathology, and prognosis, its “black-box” nature limits clinical adoption. Explainable AI (XAI) aims to reveal the decision-making logic of AI models. This narrative review synthesizes recent advances of XAI across four key domains of HCC research. In imaging diagnosis, techniques such as Grad-CAM and SHAP have enabled semantic alignment between AI outputs and clinical standards like LI-RADS, enhancing interpretability. In biomarker discovery, XAI has progressed from identifying single markers to revealing functional gene modules and molecular subtypes through multi-omics integration. In treatment efficacy prediction, XAI-based models have quantified feature contributions to therapeutic responses, supporting individualized treatment stratification. In prognosis assessment, XAI has enabled dynamic risk stratification by integrating clinical, imaging, and pathological features. However, three cross-cutting limitations persist across these domains: explanations remain predominantly correlational rather than causal, a semantic gap exists between pixel-level heatmaps and high-level clinical reasoning, and most models are static, unable to adapt to evolving clinical data. Current research is moving toward causal inference frameworks, concept-driven interpretability, and interactive, dynamic systems. In summary, XAI is transitioning from a retrospective explanation tool toward a prospective clinical decision partner, yet bridging the gap between explanation and actionable decision support remains the central challenge.