Explainable machine learning-based multiphase contrast-enhanced CT radiomics for noninvasively predicting GPC3 expression in hepatocellular carcinoma: a bicentric study.
Authors
Affiliations (2)
Affiliations (2)
- Department of Radiology, The Second Affiliated Hospital of Wenzhou Medical University Wenzhou, Zhejiang, China.
- Department of Radiology, The First People's Hospital of Yancheng City Yancheng, Jiangsu, China.
Abstract
Accurate pre-operative prediction of Glypican-3 (GPC3) profiles in hepatocellular carcinoma (HCC) is essential for selecting candidates for targeted therapy and individualized treatment. This study aimed to develop and externally validate an explainable machine learning model for noninvasively prediction of GPC3 expression in HCC. A total of 301 HCC patients were retrospectively enrolled from two centers (Center 1: 241; Center 2: 60) between January 2016 and September 2024. Radiomics features were extracted from arterial (AP), portal venous (PVP), and delayed phase (DP) images. Robust features were selected via a sequential pipeline including interobserver agreement screening (intraclass correlation coefficient [ICC] > 0.8), Spearman correlation analysis (Ρ > 0.85), and recursive feature elimination (RFE). Single-phase (AP, PVP, DP), combined-phase (APD) radiomics, clinical, and clinical-radiomics models were constructed. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), and SHAP analysis was applied for model interpretability. The combined clinical-radiomics model outperformed the clinical model, achieving an AUC of 0.886 in the training cohort, 0.880 during internal validation cohort, and 0.839 during external validation cohort (P < 0.05). The APD radiomics model showed performance non-inferior to that of the combined model (P > 0.05) and outperformed all single-phase models. AFP > 7 ng/mL (OR=3.257) and intratumoral necrosis (OR=2.947), as confirmed by multivariate logistic regression, were independent predictors for GPC3 positivity. SHAP analysis further identified radiomics score as the most influential predictor (mean |SHAP|=1.84). The multiphase CECT radiomics-based combined model demonstrates good cross-center generalizability and enables accurate noninvasive prediction of GPC3 expression status in HCC. This model may provide a reliable imaging-based tool for screening candidate patients for GPC3-targeted therapies.