A machine-learning model based on dynamic contrast-enhanced MRI for preoperative differentiation between hepatocellular carcinoma and combined hepatocellular–cholangiocarcinoma

Abstract

To establish a machine-learning model based on dynamic contrast-enhanced (DCE) magnetic resonance imaging (MRI) to differentiate combined hepatocellular-cholangiocarcinoma (cHCC-CC) from hepatocellular carcinoma (HCC) before surgery. Clinical and MRI data of 194 patients with histopathologically diagnosed cHCC-CC (n=52) or HCC (n=142) were analysed retrospectively. ITK-SNAP software was used to delineate three-dimensional (3D) lesions and extract high-throughput features. Feature selection was carried out based on Pearson's correlation coefficient and least absolute shrinkage and selection operator (LASSO) regression analysis. A radiomics model (radiomics features), a clinical model (i.e., clinical-image features), and a fusion model (i.e., radiomics features+clinical-image features) were established using six machine-learning classifiers. The performance of each model in distinguishing between cHCC-CC and HCC was evaluated with the receiver operating characteristic (ROC) curve, the area under the ROC curve (AUC), sensitivity, and specificity. Significant differences in liver cirrhosis, tumour number, shape, edge, peritumoural enhancement in the arterial phase, and lipid were identified between cHCC-CC and HCC patients (p<0.05). The AUC of the fusion model based on logistic regression was 0.878 (95% CI: 0.766-0.949) in the arterial phase in the test set, and the sensitivity/specificity was 0.844/0.714; however, the AUC of the clinical and radiomics models was 0.759 (95% CI: 0.663-0.861) and 0.838 (95% CI: 0.719-0.921) in the test set, respectively. The fusion model based on DCE-MRI in the arterial phase can significantly improve the diagnostic rate of cHCC-CC and HCC as compared with conventional approaches.