A CT Lesion Foundation Model for Non‐Invasive Assessment of Hepatocellular Carcinoma Aggressiveness and Prediction of Recurrence Risk

Hepatocellular carcinoma (HCC) remains a leading cause of cancer-related mortality worldwide, with high recurrence rates following curative resection. Accurate non-invasive assessment of tumor aggressiveness and recurrence risk is critical for surgical decision-making and postoperative surveillance, yet remains challenging with conventional imaging modalities. Here, we developed a CT lesion foundation model (LFM), trained via self-supervised learning on 32,120 generic lesion and 6,387 HCC-specific CT patches to capture transferable intra-lesional imaging phenotypes. Across one internal and two external validation cohorts, the sequentially pretrained LFM (LFM-Seq) achieved area under the curve values of 0.916, 0.846, and 0.920 for microvascular invasion (MVI) assessment, and 0.833, 0.885 and 0.795 for pathological grade assessment. The model remained robust under scenarios of limited training data and tumor localization perturbations. Furthermore, LFM-Seq-derived risk scores stratified recurrence-free survival (RFS), yielding an externally validated C-index of 0.701, and remained independently associated with RFS after adjusting for clinicopathological factors. Exploratory biological interpretability analyses showed that LFM-Seq-derived risk phenotypes were associated with suppressed lymphocyte-related immune programs. These results suggest that CT lesion foundation models can support non-invasive and robust characterization of HCC aggressiveness, recurrence risk stratification and discovery of interpretable prognostic imaging biomarkers.