Multicenter Validation of Foundation Model Adaptation for Automated Pancreatic Tumor Delineation on CT Scans

Abstract

Simple Summary Pancreatic cancer is one of the deadliest cancers, and accurate tumor delineation on contrast-enhanced computed tomography (CECT) is essential for staging, treatment planning, and response assessment. Manual tumor segmentation is time-consuming and subject to reader variability. Recently, foundation models have shown promise for medical image segmentation, but their effectiveness for disease-specific tasks such as pancreatic tumor delineation remains uncertain. In this multicenter study, we extended our previously developed Tumor-Adaptive Guidance for SAM (TAGS) framework to a large multicenter cohort and benchmarked it against foundation model-based segmentation approaches. Using 500 CECT examinations from four institutions, we found that fine-tuned TAGS variants outperformed two off-the-shelf foundation models applied without adaptation. We additionally release a multicenter pancreatic tumor CT segmentation dataset to facilitate future research and benchmarking.