Highlights
• A novel AI-powered framework reveals an inverted U-shaped non-linear relationship between tumor-stroma ratio (TSR) and mortality in hepatocellular carcinoma, with critical risk thresholds identified at 0.188 and 0.268
• The Token-Guided Multimodal Fusion architecture integrates whole-slide imaging, TSR quantification, and clinical variables as high-dimensional tokens, achieving area under the curve exceeding 0.80 for prognosis prediction
• Biological validation through transcriptomics demonstrates that the high-risk TSR phenotype is characterized by active tumor proliferation, stromal activation, and tumor microenvironment crosstalk
• This study represents a paradigm shift from manual TSR estimation to AI-driven semantic reasoning in computational pathology
