Radiomics-Based Clinical Imputation and Anatomical Region Selection for Enhanced Classification of Pancreatic Ductal Adenocarcinoma

Accurate and early detection of pancreatic ductal adenocarcinoma (PDAC) is essential for improving patient outcomes but remains difficult due to subtle radiological signs and incomplete clinical information in real-world datasets. In this study, we propose a radiomics-based framework to impute missing clinical variables directly from contrast-enhanced CT scans using the PANORAMA dataset. We evaluate the influence of both original and imputed clinical variables on PDAC classification, using gradient boosting models to assess their discriminative contribution and statistical relevance. Our findings demonstrate that clinical imputation significantly enhances both data completeness and classification performance (AUC =0.988). Moreover, we explore multiple label configurations and show that including tumor regions allows the model to capture informative features of adjacent venous structures, potential early markers of PDAC, and to learn lesion-related patterns. Interestingly, some non-tumoral regions also exhibit strong discriminative value, underscoring the importance of considering anatomical context beyond the tumor itself. These results advocate for the integration of radiomics-based imputation and region-focused modeling to improve PDAC detection in heterogeneous imaging datasets.