Predicting very early recurrence in intrahepatic cholangiocarcinoma after curative hepatectomy using machine learning radiomics based on CECT: A multi-institutional study

Abstract

BackgroundEven after curative resection, the prognosis for patients with intrahepatic cholangiocarcinoma (iCCA) remains disappointing due to the extremely high incidence of postoperative recurrence. MethodsA total of 280 iCCA patients following curative hepatectomy from three independent institutions were recruited to establish the retrospective multicenter cohort study. The very early recurrence (VER) of iCCA was defined as the appearance of recurrence within 6 months. The 3D tumor region of interest (ROI) derived from contrast-enhanced CT (CECT) was used for radiomics analysis. The independent clinical predictors for VER were histological stage, AJCC stage, and CA199 levels. We implemented K-means clustering algorithm to investigate novel radiomics-based subtypes of iCCA. Six types of machine learning (ML) algorithms were performed for VER prediction, including logistic, random forest (RF), neural network, bayes, support vector machine (SVM), and eXtreme Gradient Boosting (XGBoost). Additionally, six clinical ML (CML) models and six radiomics-clinical ML (RCML) models were developed to predict VER. Predictive performance was internally validated by 10-fold cross-validation in the training cohort, and further evaluated in the external validation cohort. ResultsApproximately 30 % of patients with iCCA experienced VER with extremely discouraging outcome (Hazard ratio (HR) = 5.77, 95 % Confidence Interval (CI) = 3.73–8.93, P < 0.001). Two distinct iCCA subtypes based on radiomics features were identified, and subtype 2 harbored a higher proportion of VER (47.62 % Vs 25.53 %) and significant shorter survival time than subtype 1. The average AUC values of the CML and RCML models were 0.744 ± 0.018, and 0.900 ± 0.014 in the training cohort, and 0.769 ± 0.065 and 0.929 ± 0.027 in the external validation cohort, respectively. ConclusionTwo radiomics-based iCCA subtypes were identified, and six RCML models were developed to predict VER of iCCA, which can be used as valid tools to guide individualized management in clinical practice.