Anatomically constrained 3D CNN segments pancreatic cancer on CT with robust external validation and outperforms a vision transformer

A 3D CNN model for pancreatic cancer segmentation on CT achieved Dice 0.76 on external validation (n=1859 exams), outperforming a vision transformer (Dice 0.68, p<0.001) and surpassing reader agreement on difficult cases.

• Model performance was stable across acquisition sites, scanner vendors, slice thicknesses, and temporal epochs.
• On a difficulty-enriched subset (n=50), the model achieved Dice 0.71, exceeding reader-pair agreement (Dice 0.57–0.65) and showing high volume concordance (CCC 0.93).
• Prospective validation in clinical trial workflows is pending.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.