Tissue Recognition in Spinal Endoscopic Surgery Using Deep Learning

Abstract

Lumbar intervertebral disc herniation is a common human disease. Nowadays, minimally invasive surgery for the treatment of lumbar disc herniation has been widely carried out. However, endoscopic spine surgery is usually performed by senior and experienced spine surgeons, because serious complications may occur once important tissues damage occurs during surgery. In this research, we developed an algorithm based on YOLOv3 framework to recognize nerve and/or dura mater images under spinal endoscopy. We collected video of surgery from 15 patients with lumbar disc herniation who underwent endoscopic spinal surgery. A total of 4829 images were obtained from these surgery videos, we divided the images into training dataset and test dataset. The training dataset consists of 1385 images of 5 patients, all of which contained images of nerve and/or dura mater. The test dataset consists of 3444 images of 15 patients, 2546 of them contain images of nerve and/or dura mater, and 898 images without nerve and/or dura mater. Three senior endoscopic spine surgeons labeled the nerve and/or dura mater in the training dataset. The results showed that the sensitivity, specificityand accuracy of nerve and dura mater recognition reached 94.27%, 97.55%and 95.12%, respectively. The performance of computeraided diagnosis (CAD) indicates that the system can be effectively identified and recognize nerve and dura mater. The CAD system will be used in endoscopic spinal surgery to assist the endoscopists to identify and recognize nerve and dura mater in the future.