Machine learning to preoperatively determine histological differentiation of mass-forming intrahepatic cholangiocarcinoma

Abstract

Objective We aim to establish a machine learning model by extracting the radiomic features from CT images and integrating clinical features to preoperatively predict the histological differentiation of mass-forming ICC. Material and Method: Clinical data and CT images were retrospectively collected from 109 ICC patients (male to female = 63:46) in our hospital from January 2013 to October 2023. The machine learning classification algorithms used in this study were random forest (RF), XGBoost (Extreme Gradient Boosting), k-nearest neighbors (KNN), and logistics regression (LR). The area under the curve of the receiver operating characteristic (AUROC) of the model on each class, as well as the macro and micro averages were calculated to comprehensively evaluate the model performance.SHapley Additive exPlanations (SHAP) was used to explain the output of the optimal model. Results Concomitant cirrhosis was more likely to occur in poorly differentiated ICC (p < 0.01), while elevated ALT and AST were more common in moderately differentiated ICC (p = 0.02). The RF model constructed based on radiomic features had moderate performance, with a macro-averaged AUC of 0.72, and an AUC of 0.69 for poorly differentiated ICC, 0.7 for moderately poorly differentiated ICC, and 0.71 for moderately differentiated ICC. Both the clinical features model and the fusion model of clinical + radiomic features performed relatively poorly, with a macro-averaged AUC of 0.51 and 0.57, respectively. Conclusion We directly classified the three histological differentiations of 109 ICC cases and found that the radiomics model performed moderately well. This suggested that a radiomics feature model alone might perform better in classification than a fusion model and that the addition of clinical features.