Comprehensive Multitask Ensemble Segmentation and Clinical Interpretation of Pancreatic and Peripancreatic Anatomy With Radiomics and Deep Learning Features

ABSTRACT To develop and validate a multitask deep learning framework for the simultaneous segmentation and clinical classification of pancreatic and peripancreatic anatomical structures in contrast‐enhanced CT imaging, enabling robust, automated diagnostic assessment and TNM staging. In this retrospective multicenter study, 3019 contrast‐enhanced abdominal CT scans from patients with confirmed or suspected pancreatic disease were analyzed. Six anatomical structures were manually annotated: tumor, parenchyma, pancreatic duct, common bile duct, peripancreatic veins, and arteries. An ensemble model combining nnU‐Net, TransUNet, and Swin‐UNet was trained for segmentation. Post‐segmentation, 215 radiomic and 2560 deep features were extracted and filtered via ICC, correlation, and harmonization procedures. Feature selection was performed using LASSO, MI, and ANOVA. Clinical classification was conducted using XGBoost, MLP, and TabTransformer models. Performance was evaluated through five‐fold cross‐validation and tested on independent internal and external datasets. The ensemble model achieved high segmentation accuracy (mean DSC: 0.89–0.94 across structures) and superior boundary precision (HD95: < 3 mm). For classification tasks, the best‐performing models attained AUCs of 95.5% for tumor malignancy, 94.7% for parenchymal condition, 94.8% for ductal status, and 94.1% for vessel invasion. Feature reproducibility was confirmed with ICC ≥ 0.75 for 198 radiomic and 2112 deep features. External validation confirmed high accuracy and generalizability, with minimal performance degradation across clinical sites. Our multitask AI framework offers comprehensive and clinically actionable insights from CT imaging, combining precise anatomical segmentation with diagnostic classification. The system supports automated TNM staging and demonstrates strong potential for clinical integration. Future studies should explore multimodal imaging, longitudinal data, and genomics integration to further expand its diagnostic and prognostic capabilities.