Noninvasive prediction of future macrovascular invasion occurrence in hepatocellular carcinoma based on quantitative imaging analysis: A multi-center study.
e14623 Background: To investigate whether preoperative imaging-based analysis could help to predict future macrovascular invasion (MaVI) occurrence in hepatocellular carcinoma (HCC). Methods: A cohort of 224 patients with HCC was enrolled from five independent medical centers (training cohort: n = 154; independent validation cohort: n = 70). Predictive clinical factors were primarily selected by uni- and multi-variable analysis. CT-based imaging analysis was performed based on extraction of 1217 radiomic features. Recursive feature elimination and random forest (RF) were chosen as the optimal radiomics modelling algorithms. A clinical-radiomics integrated model was constructed by RF modelling. Cox-regression analyses further selected risk independent factors. Risk stratification was explored by Kaplan-Meier analysis with log-rank test, regarding to MaVI occurrence time (MOT), progression free survival (PFS) and overall survival (OS). Results: The clinical-radiomics integrated model could successfully predict MaVI occurrence with areas under curve of 0.920 (training cohort, 95% confidence index [CI]: 0.875-0.965) and 0.853 (validation cohort, 95% CI: 0.737-0.970). The radiomics signature added significant improvement to the integrated model in both training and validation cohorts with p-value of 0.009 and 0.008, respectively. Radiomic features: N25_ori_gldzm_IN (hazard ratio [HR]: 0.44; p = 0.001) and N25_Coif1_ngldm_DE (HR: 0.60; p = 0.016) were selected as independent risk factors associated with MaVI occurrence time. The cox-regression model could stratified patients into high-risk and low-risk groups in MOT (p < 0.001), PFS (p = 0.003), and OS (p = 0.007). Conclusions: The noninvasive quantitative imaging analysis could enable preoperative prediction of future MaVI occurrence in HCC with prognosis implication.