Autonomous Spinal Robotic System for Transforaminal Lumbar Epidural Injections: A Proof of Concept of Study

Abstract

Study Design

Phantom study

Objective

The aim of our study is to demonstrate in a proof-of-concept model whether the use of a marker less autonomous robotic controlled injection delivery system will increase accuracy in the lumbar spine.

Methods

Ideal transforaminal epidural injection trajectories (bilateral L2/3, L3/4, L4/5, L5/S1 and S1) were planned out on a virtual pre-operative planning software by 1 experienced provider. Twenty transforaminal epidural injections were administered in a lumbar spine phantom model, 10 using a freehand procedure, and 10 using a marker less autonomous spinal robotic system. Procedural accuracy, defined as the difference between pre-operative planning and actual post-operative needle tip distance (mm) and angular orientation (degrees), were assessed between the freehand and robotic procedures.

Results

Procedural accuracy for robotically placed transforaminal epidural injections was significantly higher with the difference in pre- and post-operative needle tip distance being 20.1 (±5.0) mm in the freehand procedure and 11.4 (±3.9) mm in the robotically placed procedure (P < .001). Needle tip precision for the freehand technique was 15.6 mm (26.3 – 10.7) compared to 10.1 mm (16.3 – 6.1) for the robotic technique. Differences in needle angular orientation deviation were 5.6 (±3.3) degrees in the robotically placed procedure and 12.0 (±4.8) degrees in the freehand procedure (P = .003).

Conclusion

The robotic system allowed for comparable placement of transforaminal epidural injections as a freehand technique by an experienced provider, with additional benefits of improved accuracy and precision.

Introduction

Since the 1950s, epidural steroid injections have been utilized for conservative treatment of lower back pain caused by radiculopathy secondary to disc herniation or stenosis.^1-3 This is currently the most frequently performed procedure in pain medicine in the United States. ⁴ The 3 main approaches for administering these injections in the lumbar spine include transforaminal, interlaminar, and caudal approaches. ⁴ The main advantage of the transforaminal approach is the presumed ability to deliver medications as close as possible to the lumbar nerve roots. ⁴ These injections are frequently delivered under fluoroscopic, computed tomography (CT), or ultrasound guidance in order to increase the accuracy of needle placement. To our knowledge, there are no differences in outcomes between these modalities, although fluoroscopy is the most commonly utilized.^4,5

There is wide variability in the literature, ranging from 0-100%, regarding the efficacy of lumbar epidural injections for pain control.^4,6,7 However, the most highly cited prospective randomized control trial demonstrated an efficacy of 84% (defined as pain reduction greater than 50% 1 year after treatment). ⁷ Factors associated with variable success may include spinal instability, chronicity and grade of nerve root compression, and procedure technique and needle tip accuracy.^4,7,8 Furthermore, although rare, multiple epidural injection cases have been linked to spinal cord or neural injuries. ⁴ These complications are frequently related to inadvertent vascular injection (up to 23% of cases) of corticosteroids.^9,10

We designed a robotic system capable of performing spinal injections. The aim of our study is to demonstrate a proof-of-concept model for the use of an autonomous robotic controlled injection delivery system as it pertains to safety and accuracy of lumbar transforaminal epidural injections. The purpose of this study is to compare the accuracy of freehand transforaminal epidural injections by an expert provider to our spinal robotics system on a phantom model. We hypothesized that the robotic system would have a higher degree of accuracy compared to the conventional freehand method by the expert provider.

Materials and Methods

Study Design Overview

In this phantom study, we performed 20 transforaminal epidural injections; 10 using a freehand transforaminal procedure under fluoroscopic guidance by 1 expert provider (A. C.) and 10 using a robotic targeting system. To determine sample size, an a priori power analysis was performed using the statistical power analysis program G*power 3.1, including a t-test, an alpha set at .05, an effect size Cohen’s d  =  1.4, and a power of .8. ¹¹ This resulted in a total sample size of 20, or 10 injections per group. A custom software pre-operative planning module was developed for this study where the provider was able to plan their ideal transforaminal needle trajectories in a 3D space. These pre-operative trajectories were then compared to the actual physical trajectories performed by the provider and the robotic system. The primary metrics of the study were the distance and angulation between the pre-operative planned and actual post-operative needle tips and trajectories.

Phantom

The phantom of the lumbosacral spine was made by following the method proposed by Bellingham and Peng, using a radiopaque adult-size spine model consisting of only bony elements from T12-sacrum (Sawbones, Washington, USA). ¹² Sugar-free Metamucil (P&G, Cincinnati, OH, USA) as described by Park et al. was also added to ensure that the gelatin layer was opaque. ¹³ Thus, the bone, needles, and targets were only visible with fluoroscopy and CT images, and not with the naked eye. A CT image of this phantom model was then acquired (Figure 1).

Figure 1.

CT reconstruction of the radiopaque sawbones lumbar spine model: (A) Posterior (B) Oblique, and (C) anterior views of the model are demonstrated.

Pre-Operative Planning Software

A custom pre-operative planning module was developed for this study using 3D Slicer. ¹⁴ In this software module, needle trajectories could be planned on a CT image (Figure 2). The CT volume was displayed with standard anatomical slices as well as with a 3D rendering. A plan was created by placing a target point at the intended needle tip position during injection, and an entry point where the needle should enter the body. This plan was created by our expert interventional pain management provider (A. C.) for all trajectories. This was made for bilateral foraminal trajectories targeting L2/3, L3/4, L4/5, L5/S1 and S1 foramina, for a total of 10 “ideal” trajectories. As described by Mandell et al. with a transforaminal posterolateral approach, the final needle tip position on the axial view in this pre-operative planning module was located within the posterior margin of the neural foramen. ⁴ A model of the needle, and a line representing its trajectory, was displayed on slice and 3D views, and an option to view the volume resliced down the axis of the injection was provided to visualize and confirm a collision-free trajectory. These representations were updated dynamically with any change in the planned points.

Figure 2.

Preoperative Software modeling: (A) Segmented CT image of the phantom lumbar spine model demonstrating pre-operative planning for transforaminal epidural injections for bilateral L2/3, L3/4, L4/5, L5/S1 and S1 trajectories (B) Demonstrating the custom software application for selecting entry and target points on the phantom model CT image.

Provider Technique

Our expert interventional pain management provider (A. C.) performs transforaminal epidural spinal injections on over 500 patients per year and prefers performing injections under fluoroscopic guidance. Similarly described by Mandell et al., a lateral oblique was obtained first to confirm needle entry site, slightly more inferior than the traditional safe triangle approach. ⁴ A lateral radiograph was then taken to confirm the needle tip position with the needle tip remaining in the posterior half of the neural foramen, followed by an anteroposterior view (Figure 3A). This was performed for bilateral L2/3, L3/4, L4/5, L5/S1 and S1 trajectories, for a total of 10 injections. A post-operative CT image of this phantom model with the needles was then obtained (Figure 3B).

Figure 3.

Freehand needle placement: (A) Anteroposterior radiograph of the phantom model after the freehand technique demonstrating needle tip and trajectory relationship to the radiopaque sawbones lumbar spine (B) CT image of the phantom model with the needles in place.

Robotic Technique

The needles were subsequently withdrawn from the phantom model and this process was repeated with the robotic targeting system. For the robotic technique, a UR10 robotic arm (Universal Robots, Odense, Denmark) was used with an attached custom-built injection device. The preoperative phantom spine CT images were acquired and were digitally segmented. Anatomical landmarks, such as the spinous and transverse process, were manually annotated in CT images. The injection device was pre-calibrated to the robot arm end effector. The spine phantom and the robotic injection device are kept static during registration. Intraoperative imaging of the spine and the injection device with radiographs were then obtained in multiple viewpoints, and corresponding anatomical landmark targets of the spine were then annotated. The corresponding anatomical landmarks are used to estimate an initial pose of the phantom model in the C-arm source frame by solving a PnP problem. ¹⁵ As previously described by Gao et al., a marker less 2D/3D pipeline for registration was obtained. ¹⁶ A 2D/3D image-based registration algorithm was then used to produce spine and injector pose estimations with respect to the extrinsic imaging device, the C-arm (Figure 4A). 3D phantom model and the robot injection device model are jointly registered by optimizing a similarity score between the simulated digitally reconstructed radiograph∼(DRR) images and the real X-ray images. The similarity metric was chosen to be patch-based normalized gradient cross correlation (Grad-NCC). ¹⁷ The optimization strategy was selected as “Covariance Matrix Adaptation: Evolutionary Search” (CMA-ES) due to its robustness to local minima. ¹⁸ The relative pose transformation between the phantom and the robot is obtained from the registration outcomes. This transformation is integrated to the robotic kinematics chain using the robot pre-calibration result. The robotic arm was then utilized to precisely orient the injection device. Using the location information from the registration, the robot was automatically moved to align with the planned trajectory as defined by the skin entry and target needle tip position in the planning software. The robot then inserted the needle along this trajectory, to the target point (Figure 4B). ¹⁹ After all 10 needles were placed, an additional CT of the phantom model with the 10 spinal needles was obtained. The placement of the first needle did not influence placement of subsequent needles, as both the provider and robotic arm did not bump into or alter the trajectories of previously placed needles.

Figure 4.

Robotic Injection System: (A) Schematic of the relationship for registration between the spine phantom model, the C-arm, and the robotic arm with the injection device. (B) Actual image of the robotic arm with the attached injection device.

Statistical Analysis

The post-operative CT images from the freehand fluoroscopic guidance technique and the robotic technique were then incorporated into the 3D Slicer software and compared with the pre-operative trajectory plan. Procedural accuracy, defined as the absolute difference between pre-operative planning and actual post-operative needle tip distance (mm) in 3D space and angular orientation (degrees), were assessed between the freehand and robotic procedures utilizing independent Student’s t test, with statistical significance set at P < .05. For needle tip distance measurements, precision was reported by the range or difference between the lowest and highest absolute distances in 3D space within the freehand and robotic technique groups as compared to their ideal target points. Analyses were performed using SPSS, version 23.0, software (IBM Corp. Chicago, IL, USA).

Results

Table 1 demonstrates the needle tip distance (mm) of the post-operative robotic and freehand technique compared to the pre-operative plan. Procedural accuracy for robotically placed transforaminal epidural injections was significantly higher with the difference in pre- and post-operative needle tip distance being 20.1 (±5.0) mm in the freehand procedure and 11.4 (±3.9) mm in the robotically placed procedure (P < .001, Table 1). Needle tip precision for the freehand technique was 15.6 mm (26.3 – 10.7) compared to 10.1 mm (16.3 – 6.1) for the robotic technique (Figure 5). Differences in needle angular orientation deviation were 5.6 (±3.3) degrees in the robotically placed procedure and 12.0 (±4.8) degrees in the freehand procedure (P = .003) (Table 2, Figure 6).

Table 1.

Comparison of Needle Tip Error Distance Between Freehand Fluoroscopic Technique vs Robotic Technique. All Numeric Values Represent Needle Tip Distance (mm) Error Between the Actual Needle and the Planned Trajectory.

Transforaminal Injection Level	Freehand Fluoroscopic Technique		Robotic Technique		P-Value
	Left	Right	Left	Right
L2/L3	16.29	18.68	16.01	9.58
L3/L4	18.63	26.24	15.08	16.28
L4/L5	22.87	23.70	9.49	8.91
L5/S1	10.66	24.45	15.56	8.23
S1	15.09	23.91	8.43	6.14
Average error (SD)	20.05 (4.99)		11.37 (3.88)		<.001

Figure 5.

Scatter plot demonstrating differences in precision of the needle tip in mm between the postoperative freehand fluoroscopic (red) and robotic technique (blue): (A) x- and y-axis (Axial view), (B) x- and z-axis (AP View), and (C) y- and z-axis (Sagittal view). (D) Denotes the orientation of the XYZ plane in relation to the phantom model. L and R denote Left and Right followed by the level of the transforaminal injection. Each circular ring is spaced out by 6.25 mm with a total diameter of 25 mm of the outer circle.

Table 2.

Comparison of Trajectory Angulation Error Distance Between Freehand Fluoroscopic Technique vs Robotic Technique. All numeric Values Represent Trajectory Angulation (Degrees) Error Between the Actual Needle and the Planned Trajectory.

Transforaminal Injection Level	Freehand Fluoroscopic Technique		Robotic Technique		P-Value
	Left	Right	Left	Right
L2/L3	19.95	16.36	9.00	6.85
L3/L4	17.80	11.74	10.20	10.00
L4/L5	11.11	4.61	3.16	4.47
L5/S1	6.67	12.26	4.11	2.02
S1	9.65	9.69	1.26	4.56
Average error (SD)	11.98 (4.84)		5.56 (3.26)		.003

Abbreviations: SD = Standard deviation. Bold denotes statistical significance.

Figure 6.

Differences from planned trajectory: Differences in trajectories between the pre-operative planned software trajectories (yellow) and the actual post-operative freehand (red) and robotic techniques (purple) are demonstrated.

Discussion

Robotic-assisted surgical treatment continues to gain popularity for a variety of fields including general surgery, urology, orthopaedics, and spine surgery. Here we demonstrate a proof-of-concept model for the use of an autonomous robotic controlled injection delivery system for enhancing the safety and accuracy of lumbar transforaminal epidural injections.

There is limited literature on the use of robotics for guiding spinal injections. In 2016, Beyer et al. performed a phantom model experiment for comparing robot-assisted to freehand facet joint puncture using the iSYS 1.3 (iSYS Medizintechnik GmbH, Kitzbuehel, Austria) robotic targeting system. ²⁰ They demonstrated that robot-assisted puncture of the facet joints allowed more accurate positioning of the needle with a lower number of needle readjustments. ²⁰ Additionally, Li et al. demonstrated the use of a body-mounted robotic system for Magnetic Resonance Imaging (MRI) guided lumbar spine injections within a closed bore magnet. ²¹ They demonstrated, through a cadaveric study, that a robot-assisted approach is able to provide more accurate and reproducible cannula placements than the freehand procedure, as well as a reduction in the number of insertion attempts. ²¹ Unlike our robotics system, their robotics system relied on radiopaque markers for registration, provided a semi-autonomous system by providing needle guidance to the correct location while the provider manually inserted the needle, and did not compare post-operative trajectories to pre-operatively planned trajectories. Our robotic system demonstrates the autonomous entry of the needle to the desired depth and target point, rather than serving as a guide or tube holder for manual insertion.

There is currently no commercially available robotic platform system that can administer spine injections. However, robotics has started to gain popularity in the field of spine surgery by aiding in the placement of pedicle screws. The first commercial application was in 2004 with the SpineAssist (Mazor Robotics Ltd., Caesarea, Israel). ²² Since then, other iterations of spinal robotic systems have been developed such as the Renaissance® (Mazor Robotics Ltd., Caesarea, Israel) in 2011 and Mazor X® (Mazor Robotics Ltd., Caesarea, Israel) in 2016. Additional commercial competitors include: ROSA® SPINE (Zimmer Biomet Robotics, Montpellier, France) in 2016 and the Excelsius GPS® (Globus Medical, Inc., Audubon, Pennsylvania) in 2017. ²³ Skilled control of robot-assisted spine surgery has been shown to improve the accuracy of pedicle screw placement and decrease radiation exposure to surgical staff. ²³ However, current robotic technology has many disadvantages including high cost, steep learning curves, semi-autonomous nature, limited surgical indications, and technological glitches.^23,24

Currently, robot-assisted spine surgery is mainly restricted to instrumentation procedures with pedicle screw insertion. All of these systems are semi-active robotic systems. Meaning, they will guide and assist the surgeon in placing spinal implants, as opposed to a fully automatic system that performs the surgical operation autonomously. ²² Hence, once aligned, the surgeon will then utilize a combination of guidewires, drills, and dilators to place pedicle screws manually to a desired depth. ²³ Our robotic system expands the robotic framework by making the entire process autonomous in nature so that the provider does not have to manually advance and inject the needle to a desired target. Our robotic system not only guides the injection to the correct location but also controls for depth. Exact positioning of the needle with minimal 3D deviation from the pre-operatively planned trajectory might increase the therapeutic efficacy of epidural injections. For example, a provider may attempt to target the traditional safe or Kambin’s triangle to administer an epidural injection. However, in reality, their needle tips may not reach this desired anatomical location.

In order to have an acceptable clinical outcome, the needle tip must be placed anywhere within a triangle shaped boundary as determined in the safe triangle, posterolateral, or Kambin’s triangle approach. ⁴ If the needle tip is within this triangular boundary, the injection should theoretically provide relief. ⁴ For reference, Kambin’s triangle height and width from L1-L5 ranges from 12-18 mm and 10-12 mm respectively, or an area of 60-108 mm² in the lumbar spine. ²⁵ Our study has shown improved accuracy by the robotic platform which translates to appropriate needle placement directly resulting in improved patient outcomes. Further clinical studies must be conducted to confirm this benefit.

In addition to being autonomous, the proposed robotic system further advances spine robotics because it is also marker less or fiducial-less. In general, current spine robotics systems require a preoperative or intraoperative CT scan of the spine. In the operating room, a bone pin fiducial marker is placed on the patient. An intraoperative imaging device, such as the O-arm ® (Medtronic Sofamor Danek, Inc., Memphis, TN, USA), or other form of imaging, such as radiographs are utilized to capture both the surgical area of interest and the fiducial marker. This is used to perform a registration of the intraoperative imaging to the pre-operative CT scan and produce an intraoperative pose estimation. The surgeon can then plan a 3D trajectory on the reconstructed images and the robot will be able to align with this preplanned trajectory. However, if the fiducial markers are accidentally displaced during surgery, the robot would register this as movement by the patient and this would result in improper screw placement.

Major limitations of this study are that we utilized a phantom model, the small number of needle injections performed, and that 1 expert provider was utilized for all freehand injections and trajectory planning. The distance that the experienced provider missed the pre-operative target by, as compared to the robotic system, could have been caused by the discrepancy in tactile feedback between the sawbones model compared to an actual patient that the provider is accustomed to performing the procedure on. However, fluoroscopy was ultimately used to determine the placement of the needle tip by the provider. Although all post-operative trajectories (freehand and robotics groups) were compared to the provider’s ideal pre-operative trajectory planning on the software module, it is still highly dependent on the experience of the provider and might therefore vary considerably. Ideally, the robotic system would have been able to target each planned trajectory point flawlessly with no errors. Errors within the hand-eye calibration, the 2D-3D registration, and the hollow needle-steering and gelatin interface may have accounted for some errors. We hope to investigate this in future research. Additionally, the phantom model must be static while taking intraoperative radiographs for registration in our model. Future work will include refining registration and accounting for patient movement.

This study indicates that robotic assistance may be beneficial in enhancing the accuracy of transforaminal epidural injections. Although there are still many challenges, we demonstrate the potential of a marker less autonomous spinal robotic system.

ORCID iDs

Henry Phalen https://orcid.org/0000-0002-6854-6582

Amit Jain https://orcid.org/0000-0002-9983-3365