SU-UNet: A Novel Self-Updating Network for Hepatic Vessel Segmentation in CT Images

Abstract

Hepatectomy is currently one of the most commonly used treatment methods for malignant liver tumors. It is of great significance to clinical surgery to perform accurate hepatic vessel segmentation in preoperative CT images. However, due to the complex structure of hepatic vessels and low contrast in the CT images, it is difficult for experienced doctors to perform accurate manual labeling. Based on this, the labels of the existing public datasets are noisy. In this paper, we propose a double UNet structure based on the soft-constraint method to more accurately segment the vessels from the noisy annotation dataset. First, two different Unet output different segmentation predictions. Then a Self-updating module (SUM) is designed to optimize the noisy vessel label based on segmentation predictions so that the optimized label can better guide the network training. This method can guide the network to get better segmentation predictions. Extensive experiments using a noisy public dataset demonstrate the superiority of our method.