An anonymized lumbar computerized tomography (CT) data file was obtained from an anonymized data server of the Korea University Anam Hospital in digital imaging and communication in medicine (DICOM) format. A three-dimensional (3D) polygonised model was generated using 3D Slicer^Ⓡ (version 4.10.2; Harvard University, USA) to obtain an optimized 3D model of the spine. Regarding the vertebrae, 3D models were obtained separately for the five lumbar vertebrae, sacrum, and pelvic bones. In addition, the shapes of the intervertebral discs between the bones were obtained. The vertebrae had approximately 60,000 to 120,000 polygons per bone; however, there were concerns regarding performance degradation when used directly in the program. Therefore, we reduced the number of polygons to 5% for each bone using MeshMixer^Ⓡ (version 3.5; Autodesk Inc., USA).

For DICOM file management in a virtual space, we built a program that could read and write DICOM files in UNITY software^Ⓡ (version 4.2; Unity Technologies, USA) environment. We applied the Grassroots DICOM (GDCM) (version 3.0; http://gdcm.sourceforge.net) library in Visual Studio 2019 (version 16.10.3; Microsoft, USA) and UNITY using CMake (version 3.119.1; Kitware, Inc., USA) and Simplified Wrapper and Interface Generator (version 4.0.2; http://www.swig.org/).

After setting the GDCM, DICOM data were read as binary data and classified into tags such as image type (Tag# 0008, 0008), image orientation (Tag# 0020, 0037), instance number (Tag# 0020, 0013), image position (Tag# 0020, 0032), pixel spacing (Tag# 0028, 0030), windows center (Tag# 0028,1050), windows width (Tag# 0028, 1051) rescale intercept (Tag# 0028, 1052), rescale slope (Tag# 0028, 1053), and pixel data (Tag# 7FE0, 0010) and stored in separate series of arrays.

By adjusting the arrayed pixel data values with the rescale intercept and rescale slope values, the Hounsfield unit (HU) values of each image were obtained. We generated CT images by assembling the pixel colors through adjustment of the HU values obtained based on the windows center and windows width values. The four spatial coordinates of each image were stored and arranged in another array for subsequent virtual X-ray generation. The images were classified as axial, coronal, or sagittal using the image orientation value and then positioned in the virtual space by referencing the image position and pixel spacing values (Fig. 1A). A 3D bone-like image was obtained when only the HU values between 300 and 2,000 were visualized and the remaining values were applied with transparent materials (Fig. 1B).

For a realistic simulation, a 3D C-arm polygon model was created. The C-shaped part, where the X-ray generator and detector were attached for height adjustment, was enabled subsequently for rotation, tilting, and vertical movement. The entire C-arm could execute parallel movement in all four directions. A generator that produced X-rays and a detector that detected them were placed in the virtual space. The detector points were set to be able to be divided by the desired resolution (x * y) between the corner points. The virtual X-rays were generated from the generator, to each divided point in the detector, as shown in Fig. 2A. Using the ratio of the dot product of each X-ray with the reference point of the obtained images, the point of the CT slice images where the collision occurred was calculated. After the corresponding HU values were obtained, the sum of the values was calculated. By dividing this value by the total number of CT slices, X-ray absorption was calculated, and virtual X-ray images were obtained by arranging them sequentially after converting the values into pixel values (Fig. 2B).

The simulation program for the user movement and the C-arm procedure performance was designed to be most compatible with the Oculus Quest 2 (Meta Platforms, USA). However, the simulation was programmed to be compatible with other VR devices using the XR Interaction Toolkit^Ⓡ (Unity, USA).

To ensure a realistic virtual environment, such as the place where actual procedures are performed, the procedure room, surgical table, C-arm, and monitor were created as 3D models and placed in an appropriate virtual space. The procedure room was set in such a way that the position of the patient could be changed, and the learner could perform the procedure on their desired side. The surgical table was divided into several parts to have functions such as height adjustment; the C-arm was also divided into several parts for rotation, tilt, height adjustment, and movement.

The monitor was designed to arrange the real-time X-ray image on the left and the stored image on the right (Fig. 3A). For the trainee’s convenience, two sets of virtual arrows were placed next to the bed such that the C-arm and bed could easily be adjusted (Fig. 3B). The arrows were placed next to the C-arm and the surgical table, where they could be easily accessed by the trainees (Fig. 3C). The table on which the needle and indicators were placed was also situated near the surgical table for the trainee’s convenience (Fig. 3D).

The primary axis of the left controller was designed to easily control the rotation and tilt of the C-arm and programmed to shoot and save the virtual C-arm through the primary/secondary button. By contrast, the primary axis of the right controller was programmed to enable the learner to move freely, and the primary/secondary buttons were programmed to finely move forward and backward when the needle was inserted. The controller’s grip and trigger buttons were set to pick up/grab tools and execute functions.

To enhance the performance and convenience of the simulation, the previously obtained 3D polygonized model and a 3D bone-like image consisting of CT image slices (Figs. 4A and B) were placed in the same spatial position in the virtual simulation space (Figs. 4C and D). Subsequently, the CT image slices were made invisible (Figs. 4E and F). The 3D data of the body contour was also displayed in the correct spatial position (Fig. 4G).

An indicator and needle were created to simulate the procedure. The indicators were created as general bar-type and laser pointer-type (Fig. 5A). The laser pointer-type indicator allowed rays to be emitted; when it touched the skin, a red sphere appeared on the C-arm (Fig. 5B). After being presented with the red sphere, the position could be finely adjusted using virtual arrow buttons (Fig. 5C). When the general bar-shaped indicator, needle, and sphere collided with the X-rays, a value of 2,000 (based on the HU) was added to make it appear as if a metal material appeared on the C-arm (Fig. 5D).

Since it was difficult to support a mesh collider with more than 255 polygons in UNITY, a pelvis and body composed of less than 200 polygons each were separately constructed for efficiency, and collision with a needle- or laser pointer-type indicator was examined (Fig. 6A). When the tip of the needle made contact with the skin, a function to fine-tune the entry point of the needle through the virtual arrow button was implemented to enable precise manipulation (Fig. 6B). Once the needle started to enter the skin, it was not possible to adjust the insertion point. The needle could be advanced while finely adjusting the angle of the portion outside of the skin (Fig. 6C). The position of the needle tip was calculated as the part of the specific CT slice that was touched. When the HU value of that part was 300 or higher, it was judged that the needle had touched the bone and further advancement of the needle was restricted, and the trainee received feedback in the form of a vibration. The trainee was able to perform the procedure by checking the position of the spine and needle through the virtual C-arm. The results screen was configured such that the procedural outcomes could be verified after completing all the procedures (Fig. 6D and Supplementary Video 1). The VRS program could help trainees learn LTFEB by making the skin and bones invisible as well as allowing 3D visualization from multiple directions. This enabled the trainee to gain a clear understanding of the spatial positional relationship between the anatomy and the needle. They were trained to perform needle placement in the posterior to the vertebral body and just antero-lateral to the superior articular process.

The study protocol was reviewed and approved by the clinical trial review board of the Yonsei University Gangnam Severance Hospital (3-2021-0237). This study was conducted in accordance with the ethical principles of the Helsinki Declaration 2013 and followed good clinical practice guidelines. The study was registered at ClinicalTrial.gov (NCT05029219). The results of 20 first- or second-year residents undergoing anesthesiology and pain medicine training with no experience in C-arm-guided spine procedures were included; those already familiar with the procedures were excluded from the study. All participants were randomly allocated to one of the two following groups: the VRS (Group V, n = 10) and control groups (Group C, n = 10). Before group allocation, envelopes containing the group information were numbered sequentially and sealed. The sealed envelopes were opened by an investigator unaware of the trainees’ assessments. Written consent was obtained from all participants.

The entire curriculum flow is summarized in Fig. 7. The following data were gathered for all participants: gender, age, years of residency training, and their previous experience with performing, observing, and assisting C-arm-guided spinal procedures. Subsequently, the participants watched a five-minute video of one of the authors performing LTFEB, single vertebra level, and unilateral procedure, under C-arm guidance. The audio accompanying the video was based on Furman’s [15] book, which described the fundamentals of the C-arm, the direction of the X-ray beam, the relative positions of the needle and beam during the procedure, the anatomical structure of the lumbar spine, and the method for performing the LTFEB.

After receiving a basic introduction, all participants were instructed to perform LTFEB (lumbar 4–5, left) on the phantom as a pre-test. The phantom was prepared according to a previous study [16]. The parameters included in each assessment were the checklist score, global rating score, procedure duration (seconds), number of C-arm taken, and satisfaction score (0–5). The participants in the VRS group (Group V) were individually trained for approximately one hour using a VR headset and an Oculus program. The participants in the control group (Group C) were provided with video material and Furman’s [15] book, which had been provided to all participants in advance to ensure that they could review it and study individually for approximately one hour. The post-test was conducted in the same way as the pre-test after all the training courses were completed. Each performance was assessed by physicians with expertise in the field of pain assessment who were unaware of the group assignment.

The primary outcome was the change in the checklist score (post-test vs. pre-test) to evaluate the proficiency in performing C-arm-guided LTFEB, which has been previously utilized in several studies [16–18]. The checklist score [17], which validated the evaluation of LTFEB consisted of seven task-specific questions, and each question was scored as either ‘yes’ or ‘no’ (Appendix 1). The global rating scale [17,19] consisted of seven questions, and each question was scored out of a total score of five points (Appendix 2). Additionally, the procedural time and the number of C-arm shots were evaluated. The procedure time was defined as the time from when the first X-ray image was taken till the administration of the injectate was completed. For the participant satisfaction evaluation, questions were asked after all the courses were completed; the participants were asked to give an answer ranging from 1 to 5 (1 = unsatisfactory, 5 = satisfactory).

Based on a related previous study [17], the checklist score differed by five points between the control group participating in the didactic session and the low-fidelity group educated with a plastic spine covered by foam. Assuming that the median checklist score is the same as the mean checklist score, we conservatively predicted half of this difference as the average difference of our study. The value obtained by dividing the interquartile range by 1.35 was assumed to be the standard deviation. The number of samples was calculated based on the larger standard deviation of 1.5 among the test and control groups. When the power was 0.9 and the significance level was 0.05, the required target number of participants calculated by the G.power^Ⓡ (Version 3.1.9.7., Germany) was nine per group [20]. Considering a 10% dropout rate, 10 patients per group was deemed to be the required sample size. Continuous variables are presented as the median and interquartile range, depending on normality. Categorical demographic variables are presented as numbers. For continuous variables, comparisons of patient characteristics between the groups were made using the Mann–Whitney U test. Categorical demographic data were analyzed using Pearson’s chi-square test or Fisher’s exact test. The Mann–Whitney U test was performed to test the difference between the absolute values of the pre-test and post-test of the groups for primary and secondary outcomes and to test the difference between the groups and the degree of change of each outcome. To test the differences between before and after education within the group, the Wilcoxon signed-rank test was performed. Statistical significance was set at P < 0.05. All data were analyzed using SPSS Statistics for Windows (Version 24.0, IBM Corp., USA).