Prediction of 3-year recurrence rate of hepatocellular carcinoma after resection based on contrast-enhanced CT: a single-centre study

Abstract

Objective: We present a new artificial intelligence-powered method to predict 3-year hepatocellular carcinoma (HCC) recurrence by analysing the radiomic profile of contrast-enhanced CT (CECT) images that was validated in patient cohorts. Methods: This retrospective cohort study of 224 HCC patients with follow-up for at least 3 years was performed at a single centre from 2012 to 2019. Two groups of radiomic signatures were extracted from the arterial and portal venous phases of pre-operative CECT. Then, the radiological model (RM), deep learning-based radiomics model (DLRM), and clinical & deep learning-based radiomics model (CDLRM) were established and validated in the area under curve (AUC), calibration curve, and clinical decision curve. Results: Comparison of the clinical baseline variables between the non-recurrence (n = 109) and recurrence group (n = 115), three clinical independent factors (Barcelona Clinic Liver Cancer staging, microvascular invasion, and α-fetoprotein) were incorporated into DLRM for the CDLRM construction. Among the 30 radiomic features most crucial to the 3 year recurrence rate, the selection from deep learning-based radiomics (DLR) features depends on CECT. through the Gini index. In most cases, CDLRM has shown superior accuracy and distinguished performance than DLRM and RM, with the 0.98 AUC in the training cohorts and 0.83 in the testing. Conclusion: This study proposed that DLR-based CDLRM construction would be allowed for the predictive utility of 3-year recurrence outcomes of HCCs, providing high-risk patients with an effective and non-invasive method to possess extra clinical intervention. Advances in knowledge: This study has highlighted the predictive value of DLR in the 3-year recurrence rate of HCC.