Pancreatic cancer kills over 51,000 Americans each year, with more than 85% of cases diagnosed at an unresectable stage. Emerging research shows that subtle changes in the shape and texture of a normal-looking pancreas on CT scans can actually signal early cancer up to a year before clinical symptoms appear. However, extracting these signals requires extremely precise measurements of the entire pancreas volume.
Manual pancreas segmentation (the process of tracing the organ's outline on each CT image slice) is time-consuming, highly variable between radiologists, and practically impossible to do at scale for population screening. An accurate, fully automated AI-based segmentation tool would unlock the ability to analyze millions of scans for early cancer signals.
Researchers at Mayo Clinic curated 3,031 CT scans from 3,031 unique patients, far larger than any previous single-institution dataset for this purpose. Each scan was carefully selected to show a normal pancreas, ensuring real-world diversity including minor incidental findings like small cysts or calcifications. Five fellowship-trained radiologists each segmented their assigned scans using a standardized protocol to create consistent ground truth labels.
The team trained a 3D nnU-Net model, a self-configuring deep learning architecture that automatically optimizes its own preprocessing, training, and architecture for each new task. Using five-fold cross-validation, five separate models were trained, and their predictions were averaged together using soft majority voting to create the final ensemble model. Extensive data augmentation including rotation, flipping, noise addition, and brightness adjustments made the model robust to real-world imaging variability.
On the internal test set of 452 scans, the model achieved a mean Dice Similarity Coefficient of 0.94 (out of 1.0), meaning it agreed with radiologist segmentations in 94% of the pancreas volume on average. The Concordance Correlation Coefficient of 0.95 confirmed the model's volumetric measurements closely matched expert measurements.
On the independent external dataset AbdomenCT-1K, drawn from 12 medical centers worldwide, the model achieved an even higher Dice score of 0.96 and a Concordance Correlation Coefficient of 0.98. This improved performance on diverse multi-institutional data demonstrates the model's robustness and strong generalizability across different scanner types, hospitals, and patient populations.
Accurate automated segmentation makes it feasible to apply radiomic analysis to enormous repositories of existing CT scans to identify patients at high risk for early pancreatic cancer. Research shows that new-onset diabetes patients with specific clinical scores have nearly 20 times the general population's risk of developing the disease, and screening these individuals with AI-enhanced CT analysis could detect cancer while it is still curable.
Beyond cancer, precise volumetric pancreas measurements have implications for studying diabetes, metabolic syndrome, and chronic pancreatitis, all conditions linked to pancreatic volume changes. By making the AbdomenCT-1K segmentations publicly available, the team also provides a benchmark resource that will help other researchers develop and compare future AI tools.
This fully automated segmentation model establishes new performance benchmarks and demonstrates the feasibility of deploying AI for scalable, reproducible pancreas imaging analysis across diverse clinical environments. The model's performance on multi-institutional external data is particularly encouraging for real-world adoption.
The researchers envision this tool being integrated into existing radiology workflows to automatically flag patients with abnormal pancreatic features for closer follow-up, potentially catching cancers at stage I or II when surgical cure is still possible. Future studies will directly test whether radiomic features extracted using this segmentation tool can predict pancreatic cancer risk in prospective cohorts.