CT Image Segmentation for Preoperative Tracker Registration of Robot-Assisted Surgery

Abstract

Preoperative registration of robotic-assisted surgery is a laborious and time-consuming process which reqiures manual operation of a surgeon. To automate this process and improve accuracy, this paper proposed a simplified KiU-Net for passive marker spheres segmentation in CT images with fewer parameters. The architecture of the simplified KiU-Net has two branches: (1) Kite-Net which learns to capture fine details and accurate edges of the passive marker spheres, and (2) U-Net which learns high level features of the passive marker spheres. The dataset contains images of 14 trackers (56 passive marker spheres) scanned by CT. After training the network for 150 epochs, the dice accuracy of validation set reaches 95.2 %. After post-processing, only the complete passive marker spheres are segmented. In this way, the locations of passive marker spheres in preoperative registration can be obtained by CT image segmentation faster with higher accuracy; and the laborious manual operation can be replaced.