The Recurrent U-Net for Needle Segmentation in Ultrasound Image-Guided Surgery

In minimally invasive surgery, poor needle visualization under ultrasound has been one of the challenges of the surgery. To improve the resulting puncture error, it is effective to use deep learning from the image level to assist the surgeon in locating the needle tip position and presenting the needle trajectory angle. In this paper, a network structure based on the U-shaped network and convolutional gated recurrent unit is proposed, which could capture the information about the movement of the needle and improve the accuracy of the needle segmentation. The input of the network are three temporally consecutive ultrasound images, and the output is the segmentation of the last frame. The network is trained and tested on our bovine liver dataset containing 192 videos captured from 20 different livers. Our proposed network achieved 85% precision, 86% recall, 85% F1-score, and 76% IoU, and could localize the needle with a mean tip error of 0.66mm and a mean trajectory angle error of 0.72°. This work demonstrates the effectiveness of convolutional neural networks in improving needle visibility for minimally invasive surgical procedures. The incorporation of needle motion information into the network framework significantly boosts the segmentation precision.