Comparison of Machine Learning Models Using Diffusion-Weighted Images for Pathological Grade of Intrahepatic Mass-Forming Cholangiocarcinoma.

Is the radiomic approach, utilizing diffusion-weighted imaging (DWI), capable of predicting the various pathological grades of intrahepatic mass-forming cholangiocarcinoma (IMCC)? Furthermore, which model demonstrates superior performance among the diverse algorithms currently available? The objective of our study is to develop DWI radiomic models based on different machine learning algorithms and identify the optimal prediction model. We undertook a retrospective analysis of the DWI data of 77 patients with IMCC confirmed by pathological testing. Fifty-seven patients initially included in the study were randomly assigned to either the training set or the validation set in a ratio of 7:3. We established four different classifier models, namely random forest (RF), support vector machines (SVM), logistic regression (LR), and gradient boosting decision tree (GBDT), by manually contouring the region of interest and extracting prominent radiomic features. An external validation of the model was performed with the DWI data of 20 patients with IMCC who were subsequently included in the study. The area under the receiver operating curve (AUC), accuracy (ACC), precision (PRE), sensitivity (REC), and F1 score were used to evaluate the diagnostic performance of the model. Following the process of feature selection, a total of nine features were retained, with skewness being the most crucial radiomic feature demonstrating the highest diagnostic performance, followed by Gray Level Co-occurrence Matrix lmc1 (glcm-lmc1) and kurtosis, whose diagnostic performances were slightly inferior to skewness. Skewness and kurtosis showed a negative correlation with the pathological grading of IMCC, while glcm-lmc1 exhibited a positive correlation with the IMCC pathological grade. Compared with the other three models, the SVM radiomic model had the best diagnostic performance with an AUC of 0.957, an accuracy of 88.2%, a sensitivity of 85.7%, a precision of 85.7%, and an F1 score of 85.7% in the training set, as well as an AUC of 0.829, an accuracy of 76.5%, a sensitivity of 71.4%, a precision of 71.4%, and an F1 score of 71.4% in the external validation set. The DWI-based radiomic model proved to be efficacious in predicting the pathological grade of IMCC. The model with the SVM classifier algorithm had the best prediction efficiency and robustness. Consequently, this SVM-based model can be further explored as an option for a non-invasive preoperative prediction method in clinical practice.