Pancreas segmentation based on an adversarial model under two-tier constraints

Abstract

Pancreas segmentation is vital for the effective diagnosis and treatment of diabetic or pancreatic diseases. However, the irregular shape and strong variability of the pancreas in medical images pose significant challenges to accurate segmentation. In this paper, we propose a novel segmentation algorithm that imposes two-tier constraints on a conventional network through adversarial learning, namely UDCGAN. Specifically, we incorporate a dual adversarial training scheme in a conventional segmentation network, which further facilitates the probability maps from the segmentor to converge on the ground truth distributions owing to the effectiveness of generative adversarial networks (GANs) in capturing data distributions. This novel segmentation algorithm is equivalent to employing adversarial learning on a segmentation network that has been trained in an adversarial manner. Duplex intervention and guidance further refine the loss functions of the segmentor, thus effectively contributing to the preservation of details for segmentation. The segmentation results on the NIH Pancreas-CT dataset show that our proposed model achieves a competitive performance compared with other state-of-the-art methods.