Before an AI can detect a pancreatic tumor, it first needs to precisely locate and outline the pancreas within a scan — a task called segmentation. The pancreas is one of the most challenging organs to segment because of its irregular shape, small size relative to surrounding structures, and variable appearance between patients.
While CT-based pancreas segmentation has advanced considerably, MRI-based segmentation is far less developed, largely due to a lack of large, publicly available MRI datasets with expert annotations. Yet MRI has important advantages: it avoids radiation and provides superior soft tissue contrast, making it especially valuable for monitoring high-risk patients over time.
Researchers assembled a dataset of 767 MRI scans from 499 participants across five medical centers collected between 2004 and 2022. This included both T1-weighted and T2-weighted sequences, covering patients referred for pancreatic cystic lesions and suspected cancers. For CT benchmarking, 1,350 scans from publicly available sources were also included.
Expert radiologists manually annotated the pancreas in all MRI scans using standardized protocols. On average, annotating one scan took about 25 minutes. The multi-center origin of the data is critically important — models trained on data from a single center often fail when applied in new hospitals with different scanners and imaging protocols.
The study introduced PanSegNet, a deep learning segmentation algorithm built upon the established nnUNet framework but with specific modifications to handle the unique challenges of pancreatic segmentation in both CT and MRI. The architecture incorporates transformer-based attention mechanisms that help the model focus on relevant regions while accounting for the pancreas's complex and variable shape.
PanSegNet was designed to work across both CT and MRI modalities in a single unified framework, which is unusual — most existing tools work on only one imaging type. This cross-modality capability is important for clinical workflows where patients may receive different types of scans at different points in their care.
PanSegNet achieved state-of-the-art performance on CT segmentation, competitive with the best existing methods, while also delivering substantially improved MRI segmentation compared to all prior approaches. The model maintained strong performance across the different centers in the dataset, demonstrating robustness to variations in scanner hardware and imaging protocols.
The multi-center validation is particularly important: an AI that works well in one hospital but fails in another is of limited clinical value. The cross-site performance of PanSegNet suggests it could be deployed across diverse healthcare settings without requiring extensive retraining for each new institution.
Precise pancreas segmentation is not just an academic exercise — it has direct clinical consequences. Accurate segmentation enables reliable measurement of pancreas volume over time, which is an early indicator of disease progression. It also improves the accuracy of AI tumor detection models that depend on knowing exactly where the pancreas is before analyzing it for abnormalities.
For surgical planning, detailed segmentation maps help surgeons understand the three-dimensional anatomy of the pancreas and its relationship to surrounding blood vessels and organs. For patients undergoing surveillance for high-risk lesions, automated segmentation could reduce the radiologist time required per scan, making monitoring programs more scalable.
MRI has been underused in pancreatic cancer surveillance partly because automated analysis tools were lacking. With PanSegNet providing reliable MRI segmentation, the door opens for AI-driven MRI screening programs for high-risk individuals — who could benefit from radiation-free monitoring over many years.
The publicly available dataset created in this study is itself a major contribution to the field. By sharing the annotated MRI data, the authors enable other researchers worldwide to develop and improve segmentation tools without having to collect and annotate their own data from scratch, accelerating progress across the entire field.