Noninvasive prediction of perineural invasion in intrahepatic cholangiocarcinoma by clinicoradiological features and computed tomography radiomics based on interpretable machine learning: a multicenter cohort study

Background: Perineural invasion (PNI) of intrahepatic cholangiocarcinoma (ICC) is a strong independent risk factor for tumor recurrence and long-term patient survival. However, there is a lack of non-invasive tools for accurately predicting the PNI status. We develop and validate a combined model incorporating radiomics signature and clinicoradiological features based on machine learning for predicting PNI in ICC, and used the Shapley Additive explanation (SHAP) to visualize the prediction process for clinical application. Methods: This retrospective and prospective study included 243 patients with pathologically diagnosed ICC (training, n=136; external validation, n=81; prospective, n=26, respectively) who underwent preoperative contrast-enhanced CT between January 2012 and May 2023 at three institutions (three tertiary referral centers in Guangdong Province, China). The ElasticNet was applied to select radiomics features and construct signature derived from CT images, and univariate and multivariate analyses by logistic regression were used to identify the significant clinical and radiological variables with PNI. A robust combined model incorporating radiomics signature and clinicoradiological features based on machine learning was developed and the SHAP was used to visualize the prediction process. A Kaplan–Meier survival analysis was performed to compare prognostic differences between PNI positive and negative groups and was conducted to explore the prognostic information of the combined model. Results: Among 243 patients (mean age, 61.2 y ± 11.0 (SD); 152 men and 91 women), 108 (44.4%) were diagnosed as PNI-positive. The radiomics signature was constructed by seven radiomics features, with areas under the curves (AUCs) of 0.792, 0.748, and 0.729 in the training, external validation, and prospective cohorts, respectively. Three significant clinicoradiological features were selected and combined with radiomics signature to construct a combined model using machine learning. The eXtreme Gradient Boosting (XGBoost) exhibited improved accuracy and robustness (AUCs of 0.884, 0.831, and 0.831, respectively). Survival analysis showed the construction combined model could be used to stratify relapse-free survival (hazard ratio, 1.933; 95% confidence interval (CI): 1.093–3.418; P =0.021). Conclusions: We developed and validated a robust combined model incorporating radiomics signature and clinicoradiological features based on machine learning to accurately identify the PNI statuses of ICC, and visualize the prediction process through SHAP for clinical application.