Abstract 1633: Pretreatment prediction of recurrence-free survival in hepatocellular carcinoma after surgical resection by machine-learning based imaging analysis

Abstract

Abstract In hepatocellular carcinoma (HCC), patients at early stages are recommended to accept surgical resection as primary treatment by internationally endorsed guidelines. Nevertheless, post-surgical recurrence is one of the main threats that lead to death, with 5-year recurrence rate of over 70%. How to select patients at high risk of recurrence is essential to guide systematic follow-up and provide basis for alternative treatment strategy making, thus to prolong overall survival. In this study, we aimed to predict recurrence free survival (RFS) for HCC patients undergoing surgical resection using MR-based radiomics analysis. A cohort of 176 patients diagnosed with HCC was enrolled with complete recurrence follow-ups. A total of 4404 radiomic features including shape and size, intensity, textural, and wavelet features were extracted on the segmented tumor lesion. Mutual correlation among the features were initially assessed to reduce redundant features by Pearson’s correlation analysis. Lasso-cox regression modeling was further applied to select recurrence-related key features and generate the radiomics signature with 10-fold cross validation. In addition, we explored the predictive ability of conventional clinical factors as comparable predictors with radiomics signature by univariate analysis and constructed clinical model by cox proportional hazard regression modeling. The result turned out that 4 T2WI-based radiomic features were identified as most correlated predictors for the recurrence prediction modeling, which were correspondingly gabor10_glszm_SZLGE, gabor12_glcm45_cluster_prominence, W5L5_fos_kurtosis, and W5E5_fos_skewness. The radiomics signature could successfully stratify patients into high-risk and low-risk groups with p-value < 0.001 and of 0.019 in the training and validation cohorts, respectively, by log-rank test. The radiomics-based model yielded C-indexes of 0.761 and 0.719 in the training and validation cohorts, respectively for RFS prediction. Venous invasion and AFP were selected as effective clinical predictors. The clinical model presented with C-index of 0.551 and 0.583 in the training and validation cohorts, respectively for RFS prediction. When adding clinical predictors into radiomics model, it did not show significant improvement for the prediction with Delong-test p-value > 0.05 in both training and validation cohorts. Our study revealed T2WI-based archetypal radiomic features related with RFS and highlighted the radiomics model as an effective noninvasive tool for RFS pretreatment prediction in HCC after surgical resection, which would beyond doubt provide reliable basis for individualized treatment decision making in HCC management. Citation Format: Jingwei Wei, Yuqi Han, Xia Wu, Yushen Jin, Jie Tian. Pretreatment prediction of recurrence-free survival in hepatocellular carcinoma after surgical resection by machine-learning based imaging analysis [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 1633.