Abstract
PURPOSE: Hepatocellular carcinoma (HCC) is a perilous malignant tumor, and transcatheter arterial chemoembolization (TACE) is a widely adopted treatment technique for advanced HCC. Nevertheless, TACE may not effectively reduce the risk of distant metastases. As emerging and rapidly evolving technologies, radiomics (RAD) and deep learning (DL) have potential for predicting the outcomes post-TACE for HCC patients. This study aimed to develop a predictive model that integrates clinical data, Rad and DL to assess the risk of distant metastasis in HCC patients following TACE, and to evaluate its efficacy for these patients. METHODS: 475 cases were included in our study and were divided into training, testing, and external validation cohorts. Rad, DL, DLR (Deep learning and radiomics) models were developed and compared. The combined model was constructed by combining DLR with clinical data using logistic regression analysis. The calibration and decision curves were generated to assess model performance. RESULTS: Clinical features, including tumor size, node numbers and alpha-fetoprotein (AFP) levels, were associated with the risk of metastasis (p < 0.05). The combined model achieved area under the Receiver Operating Characteristic curve (AUC) values of 0.931, 0.861, and 0.854 in the training, testing, and external validation cohorts. Decision curve analysis (DCA) curves demonstrated the superior clinical utility of these models. CONCLUSION: The combined model can accurately predict distant metastases in HCC patients after TACE. This nomogram model improves personalized clinical decision-making by stratifying TACE-treated HCC patients into distinct risk cohorts, enabling tailored surveillance protocols and adjuvant therapy allocation for high-risk metastasis subgroups.