Radiomics analysis using machine learning to predict perineural invasion in pancreatic cancer.

BACKGROUND: Pancreatic cancer is one of the most aggressive and lethal malignancies of the digestive system and is characterized by an extremely low five-year survival rate. The perineural invasion (PNI) status in patients with pancreatic cancer is positively correlated with adverse prognoses, including overall survival and recurrence-free survival. Emerging radiomic methods can reveal subtle variations in tumor structure by analyzing preoperative contrast-enhanced computed tomography (CECT) imaging data. Therefore, we propose the development of a preoperative CECT-based radiomic model to predict the risk of PNI in patients with pancreatic cancer. PATIENTS AND METHODS: This study enrolled patients with pancreatic malignancies who underwent radical resection. Computerized tools were employed to extract radiomic features from tumor regions of interest (ROIs). The optimal radiomic features associated with PNI were selected to construct a radiomic score (RadScore). The model's reliability was comprehensively evaluated by integrating clinical and follow-up information, with SHapley Additive exPlanations (SHAP)-based visualization to interpret the decision-making processes. RESULTS: A total of 167 patients with pancreatic malignancies were included. From the CECT images, 851 radiomic features were extracted, 22 of which were identified as most strongly correlated with PNI. These 22 features were evaluated using seven machine learning methods. We ultimately selected the Gaussian naive Bayes model, which demonstrated robust predictive performance in both the training and validation cohorts, and achieved area under the ROC curve (AUC) values of 0.899 and 0.813, respectively. Among the clinical features, maximum tumor diameter, CA-199 level, blood glucose concentration, and lymph node metastasis were found to be independent risk factors for PNI. The integrated model yielded AUCs of 0.945 (training cohort) and 0.881 (validation cohort). Decision curve analysis confirmed the clinical utility of the ensemble model to predict perineural invasion. CONCLUSION: The combined model integrating clinical and radiomic features exhibited excellent performance in predicting the probability of perineural invasion in patients with pancreatic cancer. This approach has significant potential to optimize therapeutic decision-making and prognostic evaluation in patients with PNI.