Haibin Tu ^1* Jian Wang ² Lin Zhan ² Zhaowang Lin ² Ling Yang ² Min Yu ^3,4 Rong Xie ^3,4 Wanxing Lin ⁵ Yongfei Yang ⁶

• ¹ Department of ultrasound, Mengchao hepatobiliary hospital of Fujian medical university, Fuzhou, China, Fuzhou, Fujian Province, China
• ² Department of radiology Mengchao hepatobiliary hospital of Fujian medical university, Fuzhou, Fujian Province, China
• ³ The United Innovation of Mengchao Hepatobiliary Technology Key Laboratory of Fujian Province, Mengchao Hepatobiliary Hospital of Fujian Medical University, Fuzhou, Fujian Province, China
• ⁴ Mengchao Hepatobiliary Hospital, Fuzhou, Fujian Province, China
• ⁵ Department of Radiology，Datian County General Hospital, Sanming, China
• ⁶ Department of Radiology, Second Hospital of Nanping City,, Nanping, China

Abstract Background: Microvascular invasion (MVI) is a key prognostic factor in solitary hepatocellular carcinoma (HCC), significantly affecting treatment decisions and outcomes. Early prediction of MVI is crucial for enhancing clinical decision-making. Objectives: This study aimed to develop and evaluate four predictive models for MVI: one based on clinical indicators, one on MRI assessments, one using radiomics, and a combined model integrating all data across multiple medical centers. Methods: The study included patients with solitary HCC from three centers (Mengchao Hepatobiliary Hospital, The Second Hospital of Nanping, and Datian County General Hospital). The dataset was divided into an internal training set, validation set, and two external validation sets. Predictive models were built using clinical indicators, MRI, radiomics, and a combination of these. Model performance was assessed through ROC curves, calibration curves, and decision curve analysis (DCA). Lasso regression identified significant features, and SHAP analysis interpreted the model predictions. Results: A total of 319 patients were analyzed: 199 from the internal center, 67 from Nanping, and 53 from Datian. The combined model, which integrated clinical, MRI, and radiomics features, showed superior performance, with an AUC of 0.95(95%CI:0.92-0.98) in the internal training set, 0.92(95%CI:0.83-1.00) in the internal validation set, 0.96(95%CI:0.92-1.00) in Nanping, and 0.94(95%CI:0.88-0.99) in Datian. Calibration curves confirmed the model's accuracy, and NRI/IDI analyses highlighted its advantage over individual models. Key predictive indicators included pseudocapsule, peritumoral enhancement, and wavelet-based MRI features. Conclusion: This multi-center study demonstrates the effectiveness of combining clinical, MRI, and radiomics data in predicting MVI in solitary HCC, with robust results across different medical centers. These models have potential to improve patient management and treatment planning