High Speed, High Density Intraoperative 3D Optical Topographical Imaging with Efficient Registration to MRI and CT for Craniospinal Surgical Navigation

Intraoperative image-guided surgical navigation for craniospinal procedures has significantly improved accuracy by providing an avenue for the surgeon to visualize underlying internal structures corresponding to the exposed surface anatomy. Despite the obvious benefits of surgical navigation, surgeon adoption remains relatively low due to long setup and registration times, steep learning curves, and workflow disruptions. We introduce an experimental navigation system utilizing optical topographical imaging (OTI) to acquire the 3D surface anatomy of the surgical cavity, enabling visualization of internal structures relative to exposed surface anatomy from registered preoperative images. Our OTI approach includes near instantaneous and accurate optical measurement of >250,000 surface points, computed at >52,000 points-per-second for considerably faster patient registration than commercially available benchmark systems without compromising spatial accuracy. Our experience of 171 human craniospinal surgical procedures, demonstrated significant workflow improvement (41 s vs. 258 s and 794 s, p < 0.05) relative to benchmark navigation systems without compromising surgical accuracy. Our advancements provide the cornerstone for widespread adoption of image guidance technologies for faster and safer surgeries without intraoperative CT or MRI scans. This work represents a major workflow improvement for navigated craniospinal procedures with possible extension to other image-guided applications.

per screw than traditional fluoroscopy, necessitating additional anesthetic and operating room time [21][22][23] . Matched-point registration protocols are also unable to account for variances in spinal anatomy due to changes in patient positioning from CT gantry to operating table, critical in trauma and deformity cases. While this may be overcome with intraoperative 3D fluoroscopy or CT, it is at the cost of significant capital expense, increased patient radiation burden particularly with multilevel fusions, and substantial setup time [24][25][26] . Our experimental navigation system confers significant benefit over the currently available navigation techniques, implementing a simple point picking protocol and ultrafast optical radiation-free imaging registration to fuse the intraoperative surface anatomy with the preoperative CT. Rapid repeat registration allows for sequential segmental registration, minimizing intersegmental deviation from preoperative imaging to intraoperative positioning that would otherwise have required an intraoperative imaging device. While we validate our system here for craniospinal navigation, the technology is immediately applicable to non-neurosurgical navigation applications, with rapid repeat registration lending itself well to future soft-tissue applications.

Results
Ex-vivo feasibility of our experimental optical navigation technology was studied in 6 adult human cadavers, resulting in the integrated design of navigation with surgical lighting. In-vivo proof-of-principle validation of OTI was performed on 10 anaesthetized ventilated adult swine models, where interference between optical illumination for surgical lighting and OTI, both in the visible spectrum, were studied and minimized. Optical imaging of subperiosteal dissection planes between soft and bony tissues, cluttered by bleeding and carbonization effects from electro-cautery using standard surgical techniques, was performed to demonstrate pre-clinical applicability and establish required specifications for 3D imaging speed (<0.5 seconds to acquire the entire operative field using standard surgical suction to clear pooling blood) and for maximal tolerated animal anatomy movement speed (<2 mm/second) using IR tracking. To study navigation accuracy, 71 thoracic and lumbar pedicle screws were inserted and quantified by comparing intraoperative trajectory data to true screw placement based on postoperative CT imaging. Median (95%) translational and angular error of the experimental navigation system in Registered reconstructed surface data to pre-acquired imaging data with tool tracking capabilities. Verification of the system's accuracy is conducted by sliding a passively tracked probe along boney landmarks of the anatomy and confirming the system is reporting the tool's spatial location correctly. (C) Grey-scale stereoscopic cameras acquire surface images: light patterns are projected onto the surface, images are captured and 3D reconstructions and thresholded point-clouds are created representing the bony surface of the spine. (D) Registration of the acquired 3D-point cloud to pre-acquired imaging data (i.e. CT, MRI, OTI) using an iterative closest point (ICP) algorithm based on a three-point picking protocol.
SCIENTIFIC RePoRTS | (2018) 8:14894 | DOI:10.1038/s41598-018-32424-z the adult swine model was 1.67 mm (5.12 mm) and 4.37° (12.95°) in the axial plane and 1.63 mm (7.81 mm) and 6.50° (17.76°) in the sagittal plane. Following engineering optimization based on the cadaver and swine data, human clinical trials were commenced for cranial and spinal surgical procedures. At the time of writing, 171 human craniospinal surgical procedures have been performed using our experimental navigation system. Relative displacement drift between stereo-cameras over time was found to degrade navigation accuracy during the validation phase, aggravated by larger thermal expansion coefficients of the 3D printed plastic material used in the experimental system. An active calibration protocol was developed to account for camera drift which showed statistically improved cranial screw coordinate measurements.
In regards to cranial accuracy, Euclidean (3D) translational error was quantified based on the location of cranial screws fastening cranial fixation plates as measured on postoperative CT, relative to the intraoperative screw head location as reported by the experimental navigation system (Fig. 4). Median (95%) 3D translational errors for all cranial screws (N = 216 screws, 19 patients) was 2.49 mm (5.53 mm) with 1.18 mm (3.16 mm), 1.11 mm (3.95 mm), 1.15 mm (4.32 mm) in the X, Y, and Z directions respectively. Significant improvements in all axes were seen following software integration of the active calibration algorithm to account for camera drift over time (p < 0.001), reducing the median (95%) 3D translational error from 4.01 mm (5.  In regards to spinal surgical accuracy, previous work for spinal image-guided navigation techniques has focused on anatomic screw location (screw grading) rather than true in-vivo quantification of error (translational or angular deviations of the true screw trajectory vs. the intraoperative trajectory). In this study, translational and angular deviations were quantified for both the benchmark navigation systems and the experimental navigation system in the axial and sagittal planes (Fig. 5, Table 1). Absolute error measurements were quantified based on postoperative CT scans of implanted pedicle screws, ranging from 4.5-6.5 mm in diameter. Metallic CT artefacts amplify variation in absolute error measurements, particularly from large-diameter pedicle screws appropriate for our thoracolumbar series. The true engineering accuracy of the experimental navigation system is therefore underestimated. Clinical grading of screws, using the established Heary grading system 27 , was performed independently by two neuroradiologists (CH, PH), three neurosurgeons (DWC, NP, & LC) and two orthopedic spine surgeons (RG, GT) ( Table 2). A mean clinical major breach (Heary III-IV) rate of 5.7% for all navigated screws was reported, with strong intraclass correlation (ICC: 0.725; p < 0.001) and fair inter-rater agreement (Fleiss' Kappa (95% CI): 0.248 (0.243-0.254); p < 0.0001). Furthermore, there were no significant differences in major screw breaches between the experimental and benchmark systems (6.8% vs. 5.3%, p = 0.99).
Univariate analysis, accounting for age, gender, surgical navigation method (i.e. benchmark navigation, experimental navigation) and screw location (cervical, thoracic, lumbar, sacrum) identified age as a predictor for increased axial distance error, and sacral screw location as a predictor of increased axial angle error vs. the thoracic and lumbar spine. Generalized linear regression confirmed sacral screw location as a predictor of increased axial angle error (p = 0.009; Table 1) and advanced age as an indicator of increased axial distance error (p = 0.005). Surgical workflow improvement, as measured by setup and registration time to enable navigation, was compared between the experimental and benchmark systems. For cranial procedures, the median total intra-operative setup and registration time interquartile range (IQR) for the experimental navigation system was 38 s (26-76) vs. 96 s (65-120) for the benchmark system (p = 0.001). For spinal procedures, experimental navigation system median (IQR) time was 41 s (25-68) vs. 258 s (143-355) for matched-point based (p < 0.001) and 794 s (609-1136) intraoperative CT based benchmark systems (p = 0.001) (Fig. 2D). Time for pre-operative data loading/processing was not measured, however is on the order of <5 minutes, with perhaps a few minutes of time savings for the experimental vs. matched-point based systems due to eliminating the need to select 6-10 points pre-operatively. This suggests that intraoperative navigation confers significant accuracy benefit compared to freehand and fluoroscopy techniques, and that the remarkable gains in surgical workflow facilitated by the experimental navigation system do not come at the expense of surgical accuracy.

Discussion
This work represents a major shift in the current surgical paradigm through the introduction of ultra-fast optical topographical imaging and registration. We have demonstrated the implementation of an optical topographical imaging modality in craniospinal surgery, with thorough clinical and engineering data analysis to ensure surgical accuracy. By using optical imaging based surface point acquisition and GPU based parallel computing processing, we perform registration of intraoperative anatomy to preoperative MRI or CT imaging at speed orders of magnitude faster than current point-matching navigation systems. The form factor of our experimental surgical navigation system has been designed to integrate into the existing operating room environment, with the benefit  of performing imaging and registration tasks considerably faster than existing technologies. While the navigation optics are integrated into an overhead surgical lighting unit, this is readily moved out of the surgical field following registration to maintain instrument tracking, hence functioning well even with a microscope in the field. We believe these significant innovations eliminate the workflow restrictions that have traditionally led some surgeons to forgo navigation in favor of freehand approaches. While the utility of the present study is apparent in the context of craniospinal procedures, the same optical topographical imaging technology is suitable for a variety of applications. In the immediate-term, this technology may be readily translated to open lateral skull base approaches. The accuracy of surface mapping in this context via manual tracer digitization has been described 28 . Optical topographical imaging would allow significantly faster automated surface registration for open lateral skull base procedures, while maintaining functionality using an operating microscope. In the short-term, rapid optical topographical imaging allows for frequent repeat registrations, minimizing the significant target registration errors seen with existing neuronavigation technologies as a result of progressive brain shift during the procedure 29 . Frameless stereotactic navigation is also employed routinely in otolaryngology, with growing applications in orthopedic, abdominal, and craniomaxillofacial procedures 15,[30][31][32] . The utility of optical topographical imaging techniques is evident in these non-neurosurgical applications, particularly in those with significant soft-tissue manipulation or deformation, where rapid repeat registration is required to maintain accurate correlation to preoperative imaging. High frequency re-acquisition of intraoperative optical images also lends itself well to augmented-reality, with co-registered images overlaid onto operating microscopes or other displays employed commonly in multiple surgical disciplines. More speculatively in the longer-term, applications of optical topographical imaging to endoscopic procedures may be considered. Miniaturization of a structured-light projection and imaging device into an endoscope form factor has been described previously 33 . However, significant work is required to enable accurate registration in the context of blood and mucous membranes present in a typical endoscopic endonasal skull base approach, and to maintain registration accuracy throughout the procedure.
The salient findings of our study are, first, that intraoperative navigation based on OTI is accomplished significantly faster than existing technologies. These differences are particularly pronounced when compared to newer-generation devices employing intraoperative CT scanning, relative to techniques requiring point-matching registration to preoperative imaging (Fig. 2D). Such benefit was enabled by the efficient GPU algorithm, as demonstrated by the computation time for optical image acquisition and registration at 5.07 ± 1.83 seconds measured over 476 craniospinal registrations, each consisting of over 250,000 surface points, with average throughput of over 52,000 points per second, representing an improvement on the current clinical paradigm, whereby spatial accuracy is maintained while vastly improving registration time and workflow. Second, for spinal procedures, absolute translational and angular accuracy of intraoperative navigation is comparable to benchmark technologies (Fig. 6). The accuracy measured in this study is the total surgical application accuracy, stacking the navigation system's accuracy with the surgeon's ability to utilize 3D navigation data in placing pedicle screws, where surgeon's experience and anatomical knowledge also contribute 27 .
For cranial applications, the speed advantage of OTI relative to benchmark systems remains albeit less pronounced than in the spine. Multiple techniques for frameless stereotactic patient-to-image cranial registration have been developed, including surface-based or point-matching using anatomic landmarks, skin adhesive markers or bone-implanted fiducials. The accuracy of the OTI system in our cranial cohort is on par with that achieved using bone-implanted skull fiducials by current frameless navigation systems, and superior to the accuracy obtained using more common registration techniques such as surface or point matching on the scalp 34 . Remarkably, the accuracy of OTI is also on par to that reported for frame-based stereotactic localization, the current gold-standard for cranial navigation, which ranges from 2.5-3.5 mm in 3D error in both phantom and in-vivo studies [35][36][37] . With the introduction of software improvements of the active calibration protocol adjusting for camera drift over time, a significant improvement in median (95%) 3D accuracy from 4.01 mm (5.99 mm) to 1.89 mm (3.40 mm) was seen, without using bone-implanted fiducial screw based registration 34 which would disrupt existing surgical workflows.
The introduction of the active calibration protocol is crucial as it facilitates rapid intraoperative re-registration without exposing the patient to additional radiation. While the exact cause of camera drift is unknown, several possible factors have been identified including thermal expansion of the aluminum and acrylonitrile butadiene styrene (ABS) camera housing due to heat produced by the LED surgical lights, optical drift occurring within the cameras, and torsion stemming from the structural design of the experimental system and various screw connection sites. These inaccuracies, while still present in spine, are more prominent in cranial applications where a significant time lag between registration and screw placement exists, during which surgical steps such as mechanical drilling, cutting, musculocutaneous flap traction, and patient movement can all introduce navigation error due to relative displacement between the patient's cranium and the navigation reference clamp. The active calibration protocol obviates the need for additional intraoperative imaging while maintaining the required surgical accuracy. While the current analysis comprises the experience of a number of surgeons, the majority of screws navigated with the experimental system were either directly placed or supervised by one surgeon (V.X.D.Y.), representing a single-surgeon influence. Larger studies, involving multiple surgeons, are therefore underway to fully evaluate the evolution of a novice user to a skilled operator using the experimental navigation system, with multicenter studies representing the subsequent logical progression.

Methods
System Design. Our experimental navigation system consists of a small camera gantry integrated with surgical lighting. The camera gantry is composed of two cameras, a digital micromirror device pico-projector and an infrared optical tracking system ( Fig. 2A,B). The projector illuminates a patterned light of known structure and periodicity (e.g. grids, repeating bars) that is recorded by the cameras and used to reconstruct the 3D surface of the patient (Fig. 3A) 38 . The patterns enable image correspondence between the stereo cameras to be established. With calibrated cameras, the stereo images allow for 3D mapping of the surface correlating to the various height disparities. The experimental navigation system registers the acquired 3D-point cloud to pre-acquired imaging data (i.e. CT, MRI) using a surface registration algorithm that is based on the iterative closest point (ICP) algorithm (Fig. 3C,D). The experimental navigation system also utilizes an infrared (IR) tracking system containing SCIENTIFIC RePoRTS | (2018) 8:14894 | DOI:10.1038/s41598-018-32424-z two IR cameras surrounded by IR LEDs to illuminate the tracking volume. The IR system is currently used to track surgical tools using passive-reflective markers (Fig. 3D).

Pre-and Intra-Operative Workflow.
Pre-operatively, the pre-acquired imaging data (CT/MRI) must be loaded onto the experimental navigation system, as with current benchmark systems registered to pre-acquired imaging. A DICOM (Digital Imaging and Communications in Medicine) stack is loaded via disc/USB/Ethernet, and thresholded appropriately by the user to maximize visible bony landmarks while minimizing noise. For spinal procedures, the spinal level to be registered is pre-identified. Intra-operatively, the navigation system is introduced into the surgical field as a standard overhead surgical light (Video 1). Following clamping of the dynamic reference frame to the target spinal level, registration proceeds by selecting three approximate points on the pre-identified level for registration, followed by structured-light acquisition of the 3D surface of the exposed field (Video 1). Registration proceeds similarly for cranial procedures. In its current iteration, OTI registration is not continuously renewed during the surgical procedure. Following registration, if a surgical microscope is required for the procedure, the navigation system head can be moved out from its overhead position and aimed towards the reference frame, as with current benchmark optical navigation systems, to maintain registration and allow tool tracking with a microscope in the field. Clinical Trial Design. To maximize safety, the clinical trials begin with a lead-in phase where the experimental system was used to obtain optical topographical imaging and perform registration only (recruiting 09/2013-08/2014), followed by a validation phase where the benchmark navigation system provided surgical guidance while the experimental system measured established guidance trajectories (recruiting 03/2015-10/2015), a cross-over phase where the roles of the benchmark and experimental systems were switched (recruiting 1/2016-10/2016), and finally using the experimental navigation system only for surgical guidance with post-operative imaging analysis to verify accuracy (recruiting 11/2016 -present). Following engineering optimization based on our swine data, human clinical trials were commenced for craniospinal procedures. Human trials were benchmarked against two commercial point matching-based navigation systems -NAV3/3i (Stryker; Portage, MI, USA), and StealthStation S7 (Medtronic Sofamor Danek; Memphis, TN, USA), for cranial procedures using the point merge and/or surface tracer protocols. For spinal procedures, benchmark systems consisted of NAV3/3i using point merge and surface tracer protocols, or StealthStation S7 in conjunction with O-arm for automated registration to intraoperative imaging. 3D Navigation was used to guide thoracolumbar spinal pedicle screw insertion, and map the skull projections of intracranial lesions to optimize craniotomy boundaries. At the time of writing, 171 craniospinal surgical procedures have been performed using our experimental optical navigation system comprising 476 registrations. Inclusion criteria for all human clinical trials were as follows: >18 years of age, with no history of prior surgery in the area to be operated upon (cranial or spinal, as appropriate).
Engineering and Clinical Analysis. Experimental Navigation System -Animal Cohort. Translational and angular deviation in the axial and sagittal planes were analyzed for 71 thoracic and lumbar pedicle screws. Screws were placed using the experimental navigation system. Experimental Navigation System -Cranial Cohort. Accuracy was quantified by placing a tracked probe on the screws of fixed cranial plates and comparing the reported position of the tooltip to the coordinates of the cranial screws on a post-operative CT registered to the pre-operative image (Fig. 4). Data were obtained for 216 cranial screws. Subsequent to the integration of an active calibration algorithm utilizing a known geometry to apply appropriate transformations, cranial screws were dichotomized into active calibration (N = 102) vs. non-active calibration (N = 114).
Image Post-Processing. Post-operative CT images were co-registered to pre-operative CT using an iterative closest point (ICP) algorithm. The location of the cranial screw was determined as the point where the center of the screw comes into contact with the skull. The screw location recorded by the OTI system was compared to the location of the centre of the screw on post-operative CT. The post-operative actual screw location was located using an OSIRIX 64 bit workstation (Version 10.9.5, PIXMEO SARL, Switzerland).
Engineering Analysis. Application accuracy was compared between screws that were inserted with and without active calibration using a Mann-Whitney univariate test. p < 0.05 was considered statistically significant.
Benchmark Navigation System -Clinical Spine Cohort. Translational and angular deviation in the axial and sagittal planes were computed for 209 cervical, thoracic and lumbar pedicle screws. Screws were placed using commercially available navigation systems. All screws were graded clinically using the Heary Grading System. Experimental Navigation System -Clinical Spine Cohort. After the lead-in phase (which included freehand or fluoroscopy guided screws, N = 21 screws), translational and angular deviation in the axial and sagittal planes were measured for 162 screws, which were placed using commercially available benchmark navigation or OTI navigation. Image Post-Processing. Pre-and post-operative CT images were resliced to 0.3 mm thickness and dynamically resliced using multiplanar reconstruction corresponding to the axial and sagittal co-ordinates of the intra-operative and post-operative screw trajectories. For benchmark spine screws, the corresponding axial and sagittal co-ordinates were localized manually using distinct anatomical landmarks. The entry point of the screw was determined as the point where the center of the screw comes into contact with the vertebral body. The distance from the axis of symmetry perpendicular to the point of entry, as well as the angle between the screw trajectory and the perpendicular distance of the entry point, were recorded on the pre-and post-operative axial and sagittal multiplanar reconstructions. All measurements were performed using an OSIRIX 64-bit workstation (Version 10.9.5, PIXMEO SARL, Switzerland) (Fig. 5).
Clinical Grading. All screws were graded clinically using the method introduced by Heary et al. 39 . where Grade I denotes the screw is entirely contained within pedicle; Grade II the screw violates lateral pedicle but screw tip is contained within the vertebral body; Grade III indicates the screw tip penetrates anterior or lateral vertebral body; Grade IV indicates a medial or inferior breach of the pedicle; Grade V involves a violation of the pedicle or vertebral body endangering the spinal cord, nerve root, or great vessels. Clinical grading was performed independently by two neuroradiologists (CH, PH), three neurosurgeons (DWC, NP, & LC) and two orthopedic spine surgeons (RG, GT) ( Table 2). Heary grades were assessed for inter-rater reliability using the intra-class correlation coefficient. Clinical grades were dichotomized into no breach/minor breach (Heary Grade ≤2) vs. major breach (Heary Grade >2) and reliability of agreement was measured using the Fleiss' Kappa test.
Engineering Analysis. Engineering analysis was performed as previously outlined by Jakubovic et al. 40 .
Translational and angular deviations in the axial and sagittal planes in the benchmark spine cohort were compared with the corresponding deviations in the experimental navigation system spine cohort. Statistical analysis was not performed on the lead-in phase as the experimental navigation system was not used for guidance. Translational and angular deviations were compared using the Kruskal-Wallis test, visualized on Bland-Altman plots and tested using a generalized linear model. Age, gender, screw location, and guidance method (navigation, fluoroscopy, or freehand), were considered as covariates. p < 0.05 was considered statistically significant. All statistical analyses were performed using IBM SPSS 17.0 statistical software.

Data Availability
The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.