Radiomics Analysis Using Machine Learning for Predicting Perineural Invasion in Pancreatic Cancer

Background Pancreatic cancer is one of the most aggressive and lethal malignancies in the digestive system, characterized by an extremely low five-year survival rate. The perineural invasion (PNI) status in pancreatic cancer is positively correlated with adverse prognoses, including overall survival and recurrence-free survival. Emerging radiomics can reveal subtle tumor structural variations by analyzing preoperative contrast-enhanced computed tomography (CECT) imaging data. Building on this, we propose to develop a preoperative CECT-based radiomics model to predict the risk of PNI in 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 (ROI). Optimal radiomic features associated with PNI were selected to construct a radiomics score (RadScore). The model’s reliability was comprehensively evaluated by integrating clinical and follow-up information, with SHAP(SHapley Additive exPlanations)-based visualization for interpretation of decision-making processes. Results A total of 167 pancreatic malignancy patients were included. From CECT images, 851 radiomic features were extracted, with 22 identified as most strongly correlated with PNI. These 22 features were evaluated using seven machine learning methods, ultimately selecting the Gaussian Naive Bayes model, which demonstrated robust predictive performance in both training and validation cohorts, achieving area under the ROC curve (AUC) values of 0.899 and 0.813, respectively. Among clinical features, maximum tumor diameter, CA-199 levels, blood glucose concentration, and lymph node metastasis were 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 in predicting perineural invasion. Conclusion The combined model integrating clinical and radiomic features exhibits excellent performance in predicting the probability of perineural invasion in pancreatic cancer patients. This approach holds significant potential for optimizing therapeutic decision-making and prognostic evaluation in patients with PNI.