Abstract
BACKGROUND: The noninvasive diagnosis of pancreatic lesions is a critical clinical challenge. This study aims to create machine learning (ML) radiomic models for differentiating pancreatic lesions and an integrated model for pancreatic ductal adenocarcinoma (PDAC) detection. METHODS: 640 patients with pathologically confirmed malignant (n = 450), borderline (n = 108), or benign (n = 82) lesions were enrolled and divided into training (70%) and validation (30%) cohorts. Radiomic features were extracted from regions of interest on arterial and venous phase CT scans. LASSO logistic regression was used to select 36 features for building ML models, including random forest, logistic regression, support vector machine, and artificial neural networks. An integrated nomogram combining radiomic features and CA19-9 levels was developed to distinguish PDAC from borderline tumors. Model performance was evaluated using receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis. RESULTS: All ML models effectively differentiated the three tumor types. The random forest algorithm showed the best performance, achieving an area under the curve (AUC) of 0.99 and 0.95 in the training and validation sets, respectively. CA19-9 was identified as an independent diagnostic factor for PDAC. The nomogram integrating radiomics and CA19-9 achieved an AUC of 0.89 and accuracy of 0.85 in the training set, with corresponding values of 0.85 and 0.82 in the validation set. CONCLUSIONS: Radiomics-based ML models effectively differentiated benign, borderline, and malignant pancreatic tumors. The nomogram combining radiomic features with CA19-9 demonstrated robust performance, showing considerable potential to streamline the diagnostic process and facilitate timely care planning for patients with suspected pancreatic cancer.