Li Liang,1,2,* Jinshu Pang,1,* Bulin Zhang,2 Qiao Que,1 Ruizhi Gao,1 Yuquan Wu,1 Jinbo Peng,1 Wei Zhang,2 Xiumei Bai,1 Rong Wen,1 Yun He,1 Hong Yang1 

1Department of Medical Ultrasound, The First Affiliated Hospital of Guangxi Medical University, Nanning, Guangxi Zhuang Autonomous Region, People’s Republic of China; 2Department of Medical Ultrasound, Liuzhou People’s Hospital, Liuzhou, Guangxi Zhuang Autonomous Region, People’s Republic of China

*These authors contributed equally to this work

Correspondence: Hong Yang, Department of Medical Ultrasound, The First Affiliated Hospital of Guangxi Medical University, No. 6 Shuangyong Road, Nanning, Guangxi Zhuang Autonomous Region, People’s Republic of China, Tel +8607715356706, Email yanghong@gxmu.edu.cn Yun He, Department of Medical Ultrasound, The First Affiliated Hospital of Guangxi Medical University, No. 6 Shuangyong Road, Nanning, Guangxi Zhuang Autonomous Region, People’s Republic of China, Tel +8607715356706, Email heyun@gxmu.edu.cn

Purpose: To develop and validate a machine learning (ML) model for predicting early postoperative recurrence in hepatocellular carcinoma (HCC) patients by integrating contrast-enhanced ultrasound (CEUS) features with Liver Imaging Reporting and Data System (LI-RADS) classification.
Materials and Methods: A retrospective analysis was conducted on data from 279 patients who underwent surgical resection for HCC. CEUS-derived features, including the LI-RADS classification, were integrated with clinical and pathological variables to construct predictive models. Patients were randomly assigned to training (n = 196) and validation (n = 83) cohorts in a 7:3 ratio. Feature selection was performed using univariate Cox regression (p ≤ 0.05), and four ML algorithms—Random Survival Forest (RSF), Gradient Boosting Machine (GBM), CoxBoost, and XGBoost—were applied to develop recurrence prediction models. Model performance was evaluated using the concordance index (C-index), area under the curve (AUC), calibration curves, decision curve analysis (DCA), and Kaplan–Meier (KM) survival analysis.
Results: Five significant features identified by univariate Cox regression were included in model development: microvascular invasion (MVI), tumor size, LI-RADS classification, tumor necrosis, and arterial enhancement patterns. Among the four ML algorithms, GBM achieved the best overall performance, with the following results. The C-index for 1-year and 2-year recurrence prediction was 0.802 and 0.735 in the training cohort, and 0.804 and 0.710 in the validation cohort, respectively. The corresponding AUCs were 0.820 and 0.764 in the training cohort, and 0.817 and 0.716 in the validation cohort. Feature importance analysis identified LI-RADS classification, MVI, and tumor size as the top three prognostic indicators, while KM survival analysis confirmed the model’s ability to stratify patients into distinct risk groups (training cohort: p < 0.001; validation cohort: p = 0.003).
Conclusion: The GBM-based ML model integrating CEUS imaging features and LI-RADS classification demonstrates potential for predicting early postoperative recurrence of HCC, which may assist in guiding follow-up strategies.

Keywords: hepatocellular carcinoma, CEUS, LI-RADS, early recurrence, machine learning, prognostic modeling