Machine-learning analysis of contrast-enhanced computed tomography radiomics predicts patients with hepatocellular carcinoma who are unsuitable for initial transarterial chemoembolization monotherapy: A multicenter study

Highlights • Patients who are unsuitable for chemoembolization could progress with extrahepatic spread or vascular invasion after initial chemoembolization monotherapy. • A radiomics signature based on the machine learning algorithm was identified. • The signature combined with clinicoradiologicial predictors could predict TACE-unsuitable patients. • The combined model showed improved predictive performance compared with the model without radiomics signature. • The combined model could stratify patients into three strata with a low, intermediate, or high risk in training and external testing sets.
Introduction Due to the high heterogeneity of hepatocellular carcinoma (HCC), patients with non-advanced disease who are unsuitable for initial transarterial chemoembolization (TACE) monotherapy may have the potential to develop extrahepatic spread or vascular invasion. We aimed to develop and independently validate a radiomics-based model for predicting which patients will develop extrahepatic spread or vascular invasion after initial TACE monotherapy (EVIT). Materials and methods This retrospective study included 256 HCC patients (training set: n = 136; testing set: n = 120) who underwent TACE as initial therapy between April 2007 and June 2018. Clinicoradiological predictors were selected using multivariate logistic regression and a clinicoradiological model was constructed. The radiomic features were extracted from contrast-enhanced computed tomography (CT) images and a radiomics signature was constructed based on a machine learning algorithm. A combined model integrated clinicoradiological predictor and radiomics signature was developed. The predictive performance of the two models was evaluated and compared based on its discrimination, calibration, and clinical usefulness. Results In the training set, 34 (25.0%) patients were confirmed to have EVIT, whereas 26 (21.7%) patients in the testing set had EVIT. When the radiomics signature was added, the combined model showed improved discrimination performance compared to the clinicoradiological model (area under the curves [AUCs] 0.911 vs. 0.772 in the training set; AUCs 0.847 vs. 0.746 in the testing set) and could divide HCC patients into three strata of low, intermediate, or high risk in the two sets. Decision curve analysis demonstrated that the two models were clinically useful, and the combined model provided greater benefits for discriminating patients than the clinicoradiological model. Conclusions This study presents a model that integrates clinicoradiological predictors and CT-based radiomics signature that could provide a preoperative individualized prediction of EVIT in patients with HCC.