Depth Estimation for Instrument Segmentation from a Single Laparoscopic Video toward Laparoscopic Surgery Support
Takuya Suzuki, Keisuke Doman, Yoshito Mekada · 2019
It is necessary to extract surgical instruments such as forceps from laparoscopic images in order to improve the safety of laparoscopic surgery using a surgery support system. For image segmentation for surgical instruments, a deep learning technique such as a fully-convolutional neural network (FCN) is effective. It is known that the segmentation accuracy can be improved by using a stereo camera, because the depth information as well as color information on surgical instruments should be useful. This paper proposes a FCN-based depth estimation method from a single laparoscopic image captured by a monocular camera. And also proposes a U-Net-based image segmentation method using on the estimated depth information as well as color information. In experiments with the dataset of the MICCAI challenge, our method improved both the average IOU and Dice coefficient by about 2%, comparing with a comparative method using only color information. We confirmed the effectiveness of our method.