Preoperative 3D morphological phenotyping of hepatocellular carcinoma via a foundation model: a non-invasive biomarker for recurrence risk stratification.

Background: Imaging features play a central role in hepatocellular carcinoma (HCC) staging, but current systems remain limited in prognostic precision and treatment guidance. Pathological gross classification, which categorises HCC into four subtypes based on tumour margin morphology, represents a morphological surrogate of aggressive tumour behaviour encompassing histological and molecular features and has demonstrated value in predicting prognosis and guiding adjuvant locoregional therapy selection. Importantly, gross classification is assessable through imaging, but its evaluation utterly depends on postoperative pathology and remains subjective, time-consuming and poorly reproducible. We aimed to develop a three-dimensional (3D) Foundation Model-driven system to automatically perform non-invasive morphological phenotyping and to evaluate its prognostic utility. Methods: A cohort of 551 patients who underwent hepatectomy after gadoxetic acid-enhanced MRI from two medical centres was retrospectively analysed. Pathological gross classification served as the reference standard. The 3D MorphStratify Model was built to automatically phenotype 3D gross classification with a dedicated 3D Foundation Model encoder and a channel-aware fusion module. A combined model integrating 3D MorphStratify and Conventional Imaging Predictors (Morph-CIP) was constructed and validated for recurrence risk stratification. The biological rationale linking imaging morphological phenotyping to underlying biology was preliminarily explored. Results: Eligible patients were divided into a development dataset (n=291), a validation dataset (n=192) and an external test set (n=68). The 3D MorphStratify Model achieved a micro-averaged area under the receiver operating characteristic curve (AUC) of 0.932, significantly outperforming senior radiologists (AUC 0.825) with about 40-fold improved efficiency. The imaging-based 3D morphological phenotyping (3D MorphStratify) independently predicted poor recurrence-free survival (RFS, HR 2.16, 95% CI 1.20 to 3.87, p=0.010). The Morph-CIP Model enables risk stratification through training set (AUC 0.836, 0.768, 0.765 for 1-year, 2-year, 3-year RFS), internal test set (AUC 0.816, 0.690, 0.711) and external test set (AUC 0.737, 0.709, 0.702). The high-risk group is characterised by the enrichment of glycolysis and hypoxia signalling pathways. Conclusion: 3D MorphStratify is the first automated 3D morphological phenotyping system for HCC, transforming subjective pathological gross classification into an objective, non-invasive imaging biomarker. The integration of 3D MorphStratify with conventional imaging predictors enables recurrence risk stratification, potentially providing complementary information for prognostic evaluation.