Development and validation of a Random Forest prediction model integrating clinical and ultrasound radiomics features for microvascular invasion in hepatocellular carcinoma patients: a retrospective study.

Background: Microvascular invasion (MVI) is a pivotal driver of postoperative recurrence in hepatocellular carcinoma (HCC), yet its preoperative prediction remains a significant clinical challenge. Accurate preoperative MVI assessment is crucial for guiding surgical decisions, as patients with MVI-positive HCC may benefit from more aggressive resection margins or adjuvant therapies. Existing prediction models based on computed tomography (CT) or magnetic resonance imaging (MRI) radiomics are limited by high cost, limited availability in resource-constrained settings, radiation exposure (for CT), and contraindications (for MRI). In contrast, ultrasound, as the first-line imaging modality for HCC screening, offers unique advantages including real-time imaging capability, cost-effectiveness, portability, no radiation exposure, and widespread availability. However, predictive models based on ultrasound radiomics for MVI remain scarce. This study aimed to develop and validate a Random Forest model integrating clinical characteristics and ultrasound radiomics features for preoperative prediction of MVI in HCC patients. Methods: A total of 224 HCC patients who underwent hepatectomy at the First Affiliated Hospital of Xinjiang Medical University between January 2023 and June 2025 were retrospectively enrolled. Patients were randomly split into training (n=157) and test (n=67) cohorts at a 7:3 ratio. MVI was diagnosed according to the Liver Cancer Pathology Group of China (LCPGC) consensus criteria. After standardizing ultrasound image preprocessing and tumor segmentation, 1,046 radiomic features were extracted. Following feature reduction via 10-fold cross-validated Least Absolute Shrinkage and Selection Operator (LASSO) regression, three machine learning models [Random Forest, backpropagation neural network (BPNN), and Nomogram] were developed and compared. Results: The study cohort comprised 224 patients (mean age 57.9±6.2 years; 70.5% male), with MVI confirmed in 74 patients (33.0%). Patients were randomly split into training (n=157, MVI 31.8%) and test (n=67, MVI 35.8%) cohorts at a 7:3 ratio. Multivariate analysis identified maximum tumor diameter, alpha-fetoprotein (AFP), and vascular endothelial growth factor (VEGF) as independent clinical predictors of MVI (all P<0.05). LASSO regression selected four robust radiomics features. In the test cohort, the Random Forest model achieved an area under the receiver operating characteristic curve (AUC) of 0.908, with sensitivity of 79.2%, specificity of 81.4%, and accuracy of 80.6%, significantly outperforming both the BPNN model (AUC =0.675) and Nomogram model (AUC =0.755). Decision curve analysis confirmed the higher clinical utility of the Random Forest model. Conclusions: The Random Forest model integrating clinical characteristics and ultrasound radiomics features demonstrated favorable predictive performance for the preoperative assessment of MVI in HCC patients. However, given the retrospective, single-center design and relatively small sample size, these findings should be interpreted with caution. This model holds potential as a cost-effective, noninvasive tool to support individualized surgical planning. Further external validation in multi-center cohorts is warranted before clinical implementation.