Clinical imaging datasets for analysis of pancreatic cancer increasingly aggregate scans collected under heterogeneous workflows and annotation strategies. Deep learning models for medical image segmentation are typically evaluated using overlap metrics such as Dice scores, which assumes training data is drawn from heterogeneous distributions. While state-of-the-art segmentation frameworks such as nnU-Net achieve strong benchmark performance, little is known about how data provenance influences the anatomical representations learned by these models. Understanding these effects is critical for interpretability, robustness, and safe deployment in clinical settings. This project aims to investigate whether pancreas CT segmentation models trained on different data sources learn systematically different anatomical priors, even when standard accuracy metrics are similar. To evaluate these effects, we train multiple source-specific nnU-Net models on curated subsets of the PANORAMA pancreas dataset that reflect distinct data collection strategies. We will compare outputs via Dice scores and anatomical descriptors such as predicted volume, connected components, centroid location, spatial extent, and voxel-wise inter-model disagreement maps. Ongoing analysis aims to quantify these differences and demonstrate disagreement mapping as a computationally efficient proxy for anatomical uncertainty.
