With the appearance of advanced surgical equipment, such as high-resolution endoscopy and endoscopic high-speed diamond burr, percutaneous spinal endoscopy has been gradually applied to the surgical treatment for disc herniation, spinal stenosis, and metastatic tumors of the vertebral body[4]. Under endoscopic spine endoscopic, spinal tissue structures such as the annulus fibrosus, nucleus pulposus, lamina, and ligamentum flavum can be visualized intraoperatively.During spinal weight-bearing, the intervertebral disc is subjected to longitudinal compression forces in which the nucleus pulposus squeezes the surrounding annulus fibrosus due to distributed stress. Long-term weight-bearing causes degenerative changes in the patient's spine that cause the annulus fibrosus to tear and protrude, compressing the spinal nerve roots and dural sac, which results in low back pain symptoms[3, 14] .Compared to conventional open decompression surgery with or without fusion, the PELD surgical incision is only 7–9 mm, resulting in a minimally invasive procedure for LDH. This procedure also reduces surgical blood loss and avoids damaging the posterior muscle and ligament complex of the lumbar spine[15].However, the learning curve of PELD is steep for beginners. The procedure consists of two parts: (1)a working channel was established by transforaminal or interlaminar approach under fluoroscopic guidance, and (2) endoscopic lumbar nucleus pulposus removal with epidural decompression, which requires the operator to be familiar with spinal anatomy and skilled in endoscopic instrumentation.Studies have divided patients into two groups based on surgeons' experience with PELD surgery–early and late. The early group was defined as the initial patients treated before the cut-off number, and the late group as the remaining patients after the first cut-off point. According to the cut-off number, the late group had a better operation time and failure rate than the early group, and the operator's PELD learning curve had reached a plateau[6].Yong Ahn [7]et al reported that the cut-off number used to distinguish early from late groups was 20 cases, but this still does not represent the plateau when operator operation levels reached the learning curve.
In recent years, artificial intelligence and machine learning have been widely used and developed in the field of medical images, such as radiology[8, 16].Image recognition and classification are two of its most frequent tasks. It is able to learn the feature distribution and patterns from the provided data, reduces the require to prior rules, and matches required process through nonlinear deep neural network[17].CT and MRI medical imaging technology are applied in diagnosing of lumbar disc herniation and lumbar spinal stenosis because they are safe and fast and have several imaging angles.Haixing [18]et al proposed a new spinal MRI image segmentation method, which was accurately segmented the vertebral body, lamina, and dural sac in MRI images,and developed a multiscale attention network MANet to diagnose lumbar spinal stenosis.Feng Gao [19]et al used the deep learning-based BEMRI algorithm to remove the Rician noise in the MRI images and explore the relationship between lumbar facet joints and lumbar disc herniation to analyze the postoperative clinical symptoms of patients with LDH.During spinal surgery, a deep learning computer-assisted navigation system can simultaneously build an interactive 3D image in real time to synchronize patient anatomy and surgical instruments.Guoxin Fan[10] et al performed U-net-based deep learning on preoperative lumbar CT plain images of patients diagnosed with L5/S1 single-segment disc herniation to rapidly and automatically segment and reconstruct the lumbosacral structures and calculate the Kambin triangle area to assess the surgical difficulty of patients undergoing percutaneous endoscopic lumbar discectomy though the foraminal approach .Jens Fichtner [11]et al in a retrospective study of 2232 patients undergoing thoracolumbar internal fixation surgery found that the secondary revision rate (0.40%) in the 3D fluoroscopic navigation group was significantly lower (P < 0.01) compared to the freehand placement group (1.14%), which reduced the radiation dose exposure to the operator and the patient and shortened the operative time during internal fixation surgery.
In the previous review of the literature, the combination of artificial intelligence and endoscopic surgery manifested itself mainly in the placement of computer-navigated endoscopic working channels[20, 21].In contrast, research reports on spinal endoscopic surgical images are blank.Peng Cui [22]et al developed a CAD system based on YOLOv3 that is a deep learning algorithm architecture to identify nerve roots and dural sac in spinal endoscopic surgical images,which the sensitivity, specificity and accuracy can reach 90.90%, 93.68%, and 92.29% respectively. However, the algorithm they developed identified few structures, and the selected pictures failed to fully cover the entire endoscopic surgical procedure, such as ligamentum flavum exposure, radiofrequency electrode hemostasis, bite removal of lamina.
Mask R-CNN first appeared in 2017 as an efficient model for object detection and instance segmentation extended by Faster R-CNN[13, 17, 23]. It can quickly and accurately detect multiple targets in the same image and segment each target separately to achieve pixel-level high-precision labeling of the object contour[18, 24].He Kaiming [13]et al proposed a residual neural network (ResNet), which enables the model to extract features of the input image effectively, and helps the model to comprehensively extract the image features to improve the accuracy of the network for image recognition.In this study, we used computer deep learning technology to identify and categorize tissue types and surgical instruments from a sizable number of endoscopic spinal pictures. To accomplish this, we use residual network as the backbone of Mask R-CNN ——ResNet101 and ResNet50 respectively.Among them, ResNet101 was worse in recognition than ResNet50 in anatomical tissue, especially in ligamentum flavum, and nucleus pulposus, but was generally superior to ResNet50 in surgical instrument recognition.The two models were generally similar in terms of precision, recall, and specificity. However, the mean average precision of ResNet101 is better than that of ResNet50, which has better stability and accuracy.
Our team developed the artificial intelligence algorithm based on Mask R-CNN.The process of using requires attention: (1) the operator adjusts the brightness of the endoscope to avoid exposure and interfere with the field of vision.(2)stop the bleeding in time and keep the operation field clear.The innovative aspect of this study was the development of a convolutional neural network to recognize and categorize various elements within the field of view during spine endoscopic spinal surgery. By combining artificial intelligence and spinal endoscopic surgery navigation, we can track the entire procedure, including ligament exposure to the flavum, decompression of the lumbar spinal canal, nerve exploration, and removal of the nucleus pulposus,so as to achieve the purpose of intraoperative real-time identification of each element.
There are still some limitations to this study. Firstly, all included patients received percutaneous endoscopic lumbar discectomy though the interlaminar approach with a single operation.Patients receiving percutaneous endoscopic though the foraminal approach or other approaches should be included in the next stage to improve the application scope of artificial intelligence algorithm in the safe navigation of endoscopic spinal surgery.Furthermore,at present, this algorithm has not been used in clinical practice, but it is in the identification of spinal endoscopic pictures and videos, which can fully assist real-time intraoperative spinal endoscopic surgery in the future