Classification of combined hepatocellular and cholangiocarcinoma and hepatocellular carcinoma using contrast-enhanced CT based radiomics and machine-learning methods

Abstract

Abstract Purpose To explore the diagnostic performance of contrast-enhanced CT radiomics combined with a large panel of machine-learning methods in the classification of combined hepatocellular and cholangiocarcinoma (CHC) and hepatocellular carcinoma (HCC). Methods 48 CT radiomic features manually extracted using Local Image features Extraction (LIFEx) software from 264 patients diagnosed with HCC (n) and CHC (n) and treated in West China Hospital from January 2012 to December 2017 were retrospectively analyzed. A total of 45 diagnostic models were built based on 5 selection methods (DC, RF, Lasso, Xgboost and GBDT) and 9 classification algorithms (LDA, SVM, RF, Adaboost, KNN, GaussianNB, LR, GBDT, and DT). The area under the curve (AUC), accuracy, sensitivity and specificity of these models were evaluated, based on which the optimal model was determined. Results The ROC analysis revealed that all contrast-enhanced CT radiomic-based machine-learning models showed promising ability in the classification of HCC and CHC with 21 out of 45 models showing the classification AUC over 0.95. The best discriminative performance was observed in the combination of “GBDT + GBDT”, with the AUCs of 1.000 and 0.978 and in the training and validation groups. The accuracy, sensitivity and specificity of “GBDT + GBDT” in the validation group were 0.918, 0.867, 0.944 respectively. Conclusion Contrast-enhanced CT radiomic-based machine learning models show potential to be applied in differentiating HCC and CHC, and among all the models built GBDT + GBDT was identified to be the optimal model in our analysis.