Automatic Detection of Puncture Needle from CT Image with Deep Learning and Difference of CT Value Along Craniocaudal Direction

We have developed a CT guided needle puncture robot (Zerobot) to assist in interventional radiology surgery. Currently, Zerobot is operated remotely, and the next goal is to perform automatic needle puncture surgery. There is a challenge that automatic detection of puncture needle from CT images for first step of automatic puncture surgery. Because there is the case that the form of puncture needle is curved, it is necessary to detect the needle from CT image instead of estimating the needle position from arm of Zerobot. First, the method detects ROI with ResNet. Next, difference of the CT value is calculated for each pixel in the ROI, and a linear approximation is performed for to detect the needle shape. Images from animal experiments were used to evaluate the learner and image processing. We confirmed that the proposed method can detect needles in a single image and in multiple images.