Interpretable CT-based radiomics nomogram integrating clinical and semantic features for preoperative grading of...

Preoperative grading of pancreatic neuroendocrine neoplasms (PNENs) is crucial for personalized management. This study aimed to develop and validate an interpretable machine-learning framework based on CT imaging to predict PNENs pathological grades. We retrospectively analyzed 125 patients with pathologically confirmed PNENs, divided into training and test cohorts. A comprehensive radiomics pipeline was implemented, involving the extraction of high-dimensional features from CT images. We constructed three predictive models: a clinical model, a radiomics model, and a fusion model. To ensure methodological robustness, seven machine-learning algorithms were employed and cross-validated to identify the optimal classifier. Furthermore, the SHAP (SHapley Additive exPlanations) method was utilized to decode the "black box" of the radiomics signature. The fusion model-integrating CT radiomic signatures with semantic hallmarks-consistently outperformed the standalone clinical and radiomics models, achieving a superior AUC of 0.891 in the validation cohort. SHAP analysis revealed that microscopic texture complexity (e.g., Zone Entropy) and macro-geometric irregularity (e.g., Sphericity) were the dominant predictors, aligning with the biological aggressiveness of high-grade lesions. By integrating multi-dimensional CT features with a rigorous machine-learning consensus, our study provides a highly accurate and interpretable "virtual biopsy" tool.