U-Net for Pancreas Segmentation in Abdominal CT Scans

Abstract

In this paper, we propose a U-Net based approach for pancreas segmentation. Under the same setting where bounding boxes are provided, this method outperforms previously reported results with a mean Dice Coefficient of 86.70% for the NIH dataset with 4-fold cross validation. Results show that a network trained from scratch with medical images can achieve a better performance with much less training time compared to fine-tuning the models that are pretrained on natural images.