Abstract: Objective To investigate the value of radiomics features derived from contrast⁃enhanced CT and MRI arterial phase images in predicting the microvascular invasion (MVI) status in hepatocellular carcinoma (HCC). Methods HCC patients who underwent surgery at Liuzhou People's Hospital and The First Affiliated Hospital of Guilin Medical University from January 2022 to January 2024, and had completed contrast⁃enhanced CT and MRI examinations within 1 month before surgery were retrospectively enrolled and divided into a training cohort and a test cohort. Patients were classified as MVI (-) and MVI (+) groups according to postoperative pathological results. Clinical predictors of MVI were screened by multivariable Logistic regression to construct a clinical model. Arterial⁃phase CT and MRI images were imported into the Yizhun Medical⁃Darwin research platform, and two radiologists manually delineated three⁃dimensional regions of interest (ROIs) along the margins. Radiomics features were automatically extracted and preprocessed with min⁃max normalization, features with good intraclass consistency (>0.80) were retained. The LASSO regression classifier in the platform was used for feature screening and dimensionality reduction. The optimal module identified five valuable radiomics features to construct CT, MRI, and CT+MRI models. The radiomics score (Radscore) of the CT+MRI model was calculated using the Logistic regression component in the platform. The clinical parameters related to MVI were combined with Radscore to construct a combined prediction model. Model performance was evaluated using the receiver operating characteristic curve (ROC), the goodness of fit was evaluated using the calibration curve, and the net benefit of predicting MVI was compared among the models using decision curve analysis (DCA). Results A total of 118 HCC patients was included, with 82 in the training cohort and 36 in the test cohort. Multivariate Logistic regression analysis showed that tumor size, AST, and Radscore were independent predictors of MVI (all P<0.05). In both the training cohort and the test cohort, the combined model demonstrated highest performance in predicting MVI (AUC: 0.892, 0.803), outperforming the other four models. The Delong curve showed that only the comprehensive model had better predictive performance than the clinical model in both the training and the test cohorts (P=0.001, 0.038). The calibration curve indicated the comprehensive model achieved a good fit, while DCA revealed that the comprehensive model had a higher clinical net benefit in predicting MVI  than other models (0.20⁃0.60 in the training cohort and 0.20-0.58 in the test cohort). Conclusions The construction of Radscore based on radiomics features from enhanced CT and MRI arterial phase images, when it combined with tumor size and AST, the comprehensive model yielded a high predictive value for the MVI status in HCC, supporting its utility in preoperative clinical decision⁃making.

Key words: Hepatocellular carcinoma, Microvascular invasion, Radiomics, Contrast?enhanced CT, Magnetic resonance imaging