Development and validation of machine learning models integrating spectral computed tomography-derived three-dimensional quantitative parameters for predicting histopathological grade in pancreatic ductal adenocarcinoma.

Background: Pancreatic ductal adenocarcinoma (PDAC) carries a dismal prognosis, with histopathological grade closely linked to treatment selection, therapeutic response, and clinical outcomes. However, tumor grading currently relies on postoperative pathological assessment of resected specimens, limiting its utility in guiding preoperative clinical decision-making. The objective of this study was to assess the potential of integrating spectral computed tomography (CT)-derived three-dimensional (3D) quantitative parameters with clinical data using multiple machine learning (ML) algorithms to preoperatively predict histopathological grade in patients with PDAC. Methods: This retrospective study enrolled 285 patients with pathologically confirmed PDAC who underwent preoperative spectral CT, of whom 129 were ultimately included after excluding 156 patients due to prior anticancer therapy, suboptimal image quality, incomplete data, or concurrent primary malignancies. All patients were randomly assigned to either a training cohort (70%, n=90) or a testing cohort (30%, n=39). Based on postoperative histopathological findings, well- and moderately differentiated tumors were classified as low-grade (LG), while poorly differentiated tumors were defined as high-grade (HG). We collected twelve clinical features along with six 3D quantitative parameters extracted from the portal venous phase. Variable selection was performed using univariate analysis and the least absolute shrinkage and selection operator (LASSO) regression. Predictive models were developed using three ML algorithms: the light gradient boosting machine (LightGBM), logistic regression (LR), and multilayer perceptron (MLP) classifier. We evaluated the predictive models using the receiver operating characteristic (ROC) curve and decision curve analysis (DCA). The accuracy, area under the ROC curve (AUC), sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) were all calculated. Results: This study comprised 129 patients with PDAC (mean age: 61.87±10.31 years; 73 males, 56 females). Based on histopathological evaluation, patients were stratified into LG (n=92) and HG (n=37) groups. Nine variables were selected into the final model, namely 3D-conventional, 3D-iodine concentration (IC), 3D-effective atomic number (Zeff), 3D-40 kiloelectron volt (keV), 3D-100 keV, 3D-slope of the spectral attenuation curves, body mass index (BMI), carbohydrate antigen 125 (CA125), neutrophil to lymphocyte ratio. LightGBM, LR, and MLP models demonstrated satisfactory diagnostic efficacy, with the AUCs of 0.872, 0.866, 0.870 and 0.856, 0.859, 0.796 in the training and testing cohorts, respectively. The DCA confirmed that all models provided a positive net benefit in the testing cohort. Conclusions: We developed three ML models integrating 3D spectral CT quantitative parameters and clinical features for preoperative prediction of PDAC tumor grade. All models demonstrated robust accuracy, validating spectral CT-derived 3D quantitative parameters as reliable imaging biomarkers for assessing histopathological grade. These findings suggest that spectral CT biomarkers combined with ML algorithms offer a noninvasive tool for preoperative risk stratification, with important implications for personalized treatment planning in patients with PDAC.