Abstract
Pancreatic ductal adenocarcinoma (PDAC) remains one of the most aggressive and treatment-resistant cancers, highlighting the urgent need for personalized therapeutic strategies. This study aims to develop a machine learning model capable of predicting tumor response in PDAC patients using setup images acquired as part of the routine radiotherapy workflow. Aim I investigates the predictive capabilities of delta-radiomic features in forecasting tumor response during treatment, with particular focus on features calculated between treatment fractions. Among the features analyzed, Grey Level Non-Uniformity and Run-Length Non-Uniformity consistently emerged as key predictors. Aim II investigates the integration of KRAS mutation status with radiomic features to predict treatment response. This analysis revealed that KRAS-mutated status was associated with poorer response outcomes. Delta-radiomic analysis showed strong predictive performance, but combining KRAS status with these features further improved predictive accuracy suggesting that genetic markers like KRAS may enhance delta-radiomic models. Aim III demonstrated that delta-radiomic analyses examined the variability across different Magnetic Resonance Imaging machines and hospital sites, inquiring as to their robustness and generalizability. Despite differences in treatment settings, results from both internal and external cohorts were promising, with predictive performance remaining strong across institutions. In Aim IV, it was demonstrated that preprocessed feature maps significantly improved the clarity and detail of radiomic structures. Feature mapping based on delta-radiomic formulations, especially with smaller kernel sizes and ring-based volumes, consistently outperformed traditional texture analysis models. Overall, the results underscore the potential of delta-radiomics in supporting adaptive radiotherapy strategies and advancing personalized cancer treatment for PDAC patients.