Habitat radiomics combined with clinical features for predicting tertiary lymphoid structure status in pancreatic ductal adenocarcinoma: development and validation of a preoperative CECT-based model

Background: Tertiary lymphoid structures (TLS) are emerging prognostic biomarkers in pancreatic ductal adenocarcinoma (PDAC), yet no reliable non-invasive tool exists for their preoperative assessment. This study aimed to develop and validate a contrast-enhanced computed tomography (CECT)-based habitat radiomics model for preoperative prediction of intratumoral TLS status in PDAC.Methods: In this retrospective study, 204 surgically resected PDAC patients were enrolled and allocated into training (n=142) and validation (n=62) cohorts. TLS status was determined by fluorescent multiplex immunohistochemistry (mIHC), with positive cases subclassified as primary or secondary follicle-like TLS. K-means clustering partitioned the tumor volume of interest (VOI) into spatially distinct subregions; radiomics features were extracted from each subregion and selected via least absolute shrinkage and selection operator (LASSO) regression. Five machine learning classifiers were evaluated. Independent clinical predictors were identified by multivariate logistic regression (LR) and integrated with the optimal imaging model to construct a combined model.Results: TLS positivity was observed in 32.84% of patients. The LR-based habitat model achieved area under the curve (AUC) of 0.939 and 0.873 in the training and validation cohorts, substantially outperforming the conventional radiomics model (AUCs: 0.815 and 0.701). Tumor differentiation and neural invasion were identified as independent clinical predictors. The combined model yielded the highest AUCs of 0.956 and 0.881, with favorable calibration and superior net clinical benefit on decision curve analysis (DCA).Conclusions: CECT-based habitat radiomics enables robust, non-invasive preoperative prediction of intratumoral TLS in PDAC, offering a practical tool for individualized risk stratification and treatment planning; incorporation of clinical variables yielded numerically higher AUCs but without statistically significant improvement.