Radiomics-Based Machine Learning Model for Differential Diagnosis of Pancreatic Cystic Neoplasms

Introduction : Accurate differential diagnosis of pancreatic cystic lesions is crucial for selecting the optimal management strategy and for timely identification of lesions with malignant potential. A key objective of radiology is the development and implementation of quantitative imaging approaches that can improve diagnostic performance and reduce reliance on subjective image interpretation. Purpose : To evaluate the value of magnetic resonance imaging (MRI) and MRI-based radiomic analysis for the differential diagnosis of pancreatic cystic neoplasms, and to develop machine-learning radiomic models for predicting the malignant potential of pancreatic cystic lesions. Materials and methods : In this retrospective study, MRI examinations of 67 patients with surgically and pathologically confirmed pancreatic cystic neoplasms were analyzed. Lesion segmentation and radiomic feature extraction were performed by a radiologist experienced in abdominal imaging. Seven supervised machine-learning models were trained. Model performance was assessed using area under the receiver operating characteristic curve (ROC-AUC), precision–recall AUC (PR-AUC), accuracy, sensitivity, specificity, and F1-score. The best-performing model was selected based on ROC-AUC. Model interpretability was evaluated using SHapley Additive exPlanations (SHAP). Results : The Random Forest classifier showed the best performance (ROC-AUC = 0.83), followed by LightGBM (ROC-AUC = 0.77). SHAP analysis highlighted the radiomic features with the greatest impact on the model outputs. Conclusion : MRI-based radiomics may support risk stratification of pancreatic cystic lesions. Further studies with larger cohorts are needed to confirm these findings and improve model generalizability.