A combined radiomics and clinical model for preoperative differentiation of intrahepatic cholangiocarcinoma and intrahepatic bile duct stones with cholangitis: a machine learning approach

Abstract

Background: This study aimed to develop and validate a predictive model integrating radiomics features and clinical variables to differentiate intrahepatic bile duct stones with cholangitis (IBDS-IL) from intrahepatic cholangiocarcinoma (ICC) preoperatively, as accurate distinction is crucial for determining appropriate treatment strategies.

Methods: A total of 169 patients (97 IBDS-IL and 72 ICC) who underwent surgical resection were retrospectively analyzed. Radiomics features were extracted from ultrasound images, and clinical variables with significant differences between groups were identified. Feature selection was performed using LASSO regression and recursive feature elimination (RFE). The radiomics model, clinical model, and combined model were constructed and evaluated using the area under the curve (AUC), calibration curves, decision curve analysis (DCA), and SHAP analysis.

Results: The radiomics model achieved an AUC of 0.962, and the clinical model achieved an AUC of 0.861. The combined model, integrating the Radiomics Score with clinical variables, demonstrated the highest predictive performance with an AUC of 0.988, significantly outperforming the clinical model (p < 0.05). Calibration curves showed excellent agreement between predicted and observed outcomes, and the Hosmer-Lemeshow test confirmed a good model fit (p = 0.998). DCA revealed that the combined model provided the greatest clinical benefit across a wide range of threshold probabilities. SHAP analysis identified the Radiomics Score as the most significant contributor, complemented by abdominal pain and liver atrophy.

Conclusion: The combined model integrating radiomics features and clinical data offers a powerful and reliable tool for preoperative differentiation of IBDS-IL and ICC. Its superior performance and clinical interpretability highlight its potential for improving diagnostic accuracy and guiding clinical decision-making. Further validation in larger, multicenter datasets is warranted to confirm its generalizability.

Keywords: intrahepatic bile duct stones; intrahepatic cholangiocarcinoma; intrahepatic lithiasis; nomogram; radiomics.

Figure 1

Flowchart of included and excluded patients.

Figure 2

ROI delineation on ultrasound images. (A) Original grayscale ultrasound image of a patient with intrahepatic cholangiocarcinoma (ICC) combined with bile duct stones. (B) The region of interest (ROI) was manually delineated (red area) along the tumor margin. This case demonstrates the challenge in differentiating ICC with bile duct stones from intrahepatic bile duct stones with cholangitis (IBDS-IL) based on imaging alone.

Figure 3

Receiver operating characteristic curve analysis of the clinical model.

Figure 4

Recursive Feature Elimination (RFE) selected feature importance. The plot displays the top 10 features selected using RFE with Random Forest as the evaluation model. Feature importance values are represented along the x-axis, with individual features listed on the y-axis.

Figure 5

Receiver operating characteristic curve analysis of the modeling methods. The Random Forest model showed the best diagnostic performance, with AUC values of 1.0 (1.0–1.0) in the training group (A) and 0.962 (0.904–1) in the validation group (B).

Figure 6

(A) The nomogram for the combined model integrates clinical factors (Abdominal Pain and Liver Atrophy) and the Radiomics Score to predict the probability of intrahepatic bile duct stones with cholangitis (IBDS-IL) and intrahepatic cholangiocarcinoma (ICC). (B) Calibration curve of the combined model. The dotted line represents the apparent performance, the solid line indicates the bias-corrected results, and the dashed line represents the ideal performance.

Figure 7

(A) Receiver Operating Characteristic (ROC) curves for the three models. The combined model achieved the highest AUC (0.988), followed by the radiomics model (0.962) and the clinical model (0.861), demonstrating superior predictive performance of the combined model. (B) Decision Curve Analysis (DCA) for the three models. The combined model (yellow line) provided the greatest net benefit across a wide range of threshold probabilities, indicating its superior clinical utility compared to the radiomics model (red line) and clinical model (blue line). (C) Radar chart comparing key performance metrics (precision, specificity, sensitivity, AUC, F1 score, accuracy, and recall) for the three models. (D) Confusion matrices for the clinical model, radiomics model, and combined model. The combined model showed the best classification performance, with fewer misclassifications, particularly in identifying ICC cases (0 misclassified).

Figure 8

SHapley Additive exPlanations (SHAP) summary plot illustrating the contribution of individual features to the combined model’s output. The x-axis represents the SHAP values, reflecting the impact of each feature on the model’s predictions. Positive SHAP values indicate a higher likelihood of predicting ICC, while negative SHAP values correspond to IBDS-IL.