Abstract
Purpose.
Intraoperative 2D virtual long-film (VLF) imaging is investigated for 3D guidance and confirmation of the surgical product in spinal deformity correction. Multi-slot-scan geometry (rather than a single-slot “topogram”) is exploited to produce parallax views of the scene for accurate 3D colocalization from a single radiograph.
Methods.
The multi-slot approach uses additional angled collimator apertures to form fan-beams with disparate views (parallax) of anatomy and instrumentation and to extend field-of-view beyond the linear motion limits. Combined with a knowledge of surgical implants (pedicle screws and/or spinal rods modeled as “known components”), 3D-2D image registration is used to solve for pose estimates via optimization of image gradient correlation. Experiments were conducted in cadaver studies emulating the system geometry of the O-arm (Medtronic, Minneapolis MN).
Results.
Experiments demonstrated feasibility of multi-slot VLF and quantified the geometric accuracy of 3D-2D registration using VLF acquisitions. Registration of pedicle screws from a single VLF yielded mean target registration error of (2.0±0.7) mm, comparable to the accuracy of surgical trackers and registration using multiple radiographs (e.g., AP and LAT).
Conclusions.
3D-2D registration in a single VLF image offers a promising new solution for image guidance in spinal deformity correction. The ability to accurately resolve pose from a single view absolves workflow challenges of multiple-view registration and suggests application beyond spine surgery, such as reduction of long-bone fractures.
Keywords: Intraoperative imaging, image-guided surgery, linear slot-scanning radiography, 3D-2D image registration
1. INTRODUCTION AND PURPOSE
Many mobile intraoperative imaging systems are developing the capability for virtual long-film (VLF) radiography suitable for imaging of the extended views of the spine or long bones.1–3 Such systems can accomplish VLF imaging by way of single-slot scanning (analogous to a topogram), multi-slot scanning, or small angle rotations of the source. For example, the O-arm system in Figure 1a (Medtronic, Littleton MA) could potentially acquire VLFs via longitudinal motion of the gantry and a multi-slot collimator – not available in current clinical systems, but a topic of research investigated in this work. X-ray imaging systems in current practice, including the O-arm, offer a limited longitudinal (z-direction) field of view (FOV, ~20–30 cm) covering only 4–6 vertebrae in a single radiographic view or 3D scan. The ability to visualize longer lengths of the spinal column beyond the usual FOV could provide a basis for measurement of spinal alignment,4 planning corrective approach, and assessment of deformity correction in the operating room.
Figure 1.
Virtual long-film imaging. (a) Illustration of the O-arm at −zmax and +zmax (transparent overlay) showing the longitudinal extent of the gantry travel. (b) For illustrative purposes in preliminary studies – a coronal slice of stitched cone-beam CT reconstructions, formed from 4 overlapping scans. (c) Simulated AP and lateral VLF images obtained from forward projections of the reconstruction in (b).
A rough impression of long-length imaging can be obtained from manual acquisition of individual 2D radiographs,5 but the uncontrolled motion of the imager and/or patient table can introduce unintended parallax, and the parallax discrepancy in large-area views can challenge synthesis of a geometrically accurate long-length image. Methods exist and are under continuing development to register and stitch standard 2D radiographs for parallax-free panoramic views.6 Methods that employ the slot-scanning technique via one or more collimators, on the other hand, offer additional reduction of patient dose compared to stitched radiographs.1,7
The work summarized below investigates the utility of 2D VLF imaging for 3D quantitative guidance via 3D-2D image-based registration of patient anatomy and instrumentation (known-components) therein. While 3D-2D registration requires multiple parallax views, a multi-slot VLF geometry is exploited to achieve 3D pose estimation from a single image acquisition. The method is applied in the context of spine surgery to resolve 3D poses of pedicle screws for rapid verification of placement.
2. METHODS
2.1. Virtual long-film imaging
Slot-scanning and projection geometry.
A general schematic of the slot-scan geometry is illustrated in Figure 2a. For the O-arm, we used a system model that includes a PaxScan 4030CB flat-panel detector (Varex Imaging, Palo Alto CA) with 1536×2048 pixels at 0.194 mm pitch, yielding 39.7×29.8 cm2 detector area. A multi-slot collimator was simulated with beam apertures that are narrow in the (z) scan direction with optional tilting of the slots. The diagram in Figure 2a shows an example hypothetical configuration featuring 3 slots with the two edge slots tilted by ±ϕ. A fixed source-to-detector distance (SDD) of 116.8 cm enforces ϕ≤7°, beyond which the fan-beams would fall outside the detector. The slot-scan is executed by linear motion of the gantry in z, such that the same detector rows are exposed to the collimated beams as the source is translated. When combined to produce a single VLF, the tilted beams extend the FOV beyond the travel limits of the gantry (zmax = ±18 cm), thus increasing the effective length of the detector from 29.8 cm to 2(mzmax + SDD tan ϕ), which for ϕ=7° and an object at isocenter (i.e., magnification at m=1.8) amounts to ~93 cm.
Figure 2.
Virtual long-film imaging geometry. (a) Schematic of the VLF projective geometry, showing the fan-beams cast by a 3-slot collimator. (b) Individual radiographs with parallax obtained from a single VLF scan for ϕ = 7°, with regions marked for extent of union (∪z) and intersection (∩z) of the radiographs.
Parallax views for 3D colocalization.
An exciting prospect of a multi-slot collimator, in addition to producing extended radiographs, is that the parallax views can be exploited to drive 3D localization of anatomy and instruments via 3D-2D registration. To differentiate the two tasks, the union of individual views is defined as ∪z, and their intersection as ∩z, where extra views (at ±ϕ) of the same scene is made available as shown in Figure 2b. The extent of both ∪z and ∩z are subject to collimator angle (ϕ) and magnification (m).
Virtual long-film simulation.
A prototype O-arm (model 1000) was used as the basis for experimentation. The system did not incorporate a multiple-slot collimator; rather, in the current work, VLFs were simulated by first stitching multiple cone-beam CT volumes and forward-projecting according to the geometry of the O-arm with the 3-slot collimator of Figure 2a. Using cadaver torsos as subjects, 4 overlapping FOVs (Figure 1c) were captured by z translation of the gantry and 3D filtered back-projection (FBP) reconstructions were performed as described in previous work.8 The reconstructions were stitched into a single extended volume (μ) using the encoded motor positions. A custom forward projector was defined, where given a homogeneous projective transform of the desired view (e.g.,P for an AP view) obtained from a geometric calibration of the imager, the set of projective transforms for individual fan-beams are defined as {Pϕ} = {P T(z) R(ϕ) : |z| ≤ zmax}, which effectively rotates (R) and translates (T) the image volume as opposed to the gantry. Finally, Poisson noise was added to the projected images to simulate low-dose radiographic imaging conditions.
2.2. Known-component registration
A 3D-2D registration algorithm (KC-Reg) established in prior work9 was adapted to work with VLFs. The method works by optimizing the parameters of a known component (e.g., preoperative patient CT or pedicle screws) such that the gradient correlation (GC) between an acquired VLF and simulated projections of the known components is maximized. Earlier work using standard radiographic / fluoroscopic image acquisition relied on multiple views (e.g., AP and LAT) obtained via rotation of the gantry. In the proposed approach, the objective function for registration of a single screw (k) using a single (i.e., slot collimated) VLF can be expressed as:
| (1) |
where VLFϕ is one of the constituent radiographs (Figure 2b) acquired from a single slot tilted at ϕ, and is the simulated VLF that is obtained by projecting the input 3D component using {Pϕ} projective transforms. The objective function is solved using covariance matrix adaptation evolution strategy (CMA-ES), which optimizes the pose parameters (λ) of the screws.
2.3. Cadaver experiments with induced deformation
Virtual long-film imaging and registration of surgical instrumentation were evaluated in experiments conducted on a cadaveric human torso (77 year old male, medium body habitus). Ten pedicle screws were placed (2 unilateral and 8 bilateral) across the thoraco-lumbar spine. The registrations used vendor-specific CAD models (Solera, Medtronic, Littleton MA), which included a rigid polyaxial screw and its articulating tulip head, and collision avoidance was enforced for bilateral cases (Figure 3a).9 Four 3D scans were acquired to obtain extended reconstructions with the specimen in prone, supine, and left lateral decubitus pose and with extra padding positioned to induce further deformation. Target clinical task was confirmation of pedicle screw placement following their delivery. The experiments assessed the fidelity of radiograph and VLF simulation with respect to target registration error (TRE) computed at the screw tips.
Figure 3.
Geometric accuracy of pedicle screw registration. TRE at the screw tip (mm) and angular approach (deg) when using: (a) 2 (AP + LAT) acquired (R); (b) 2 simulated (); (c) 2 VLF radiographs; and (d) a single VLF radiograph taking advantage of the collimated parallax views.
3. RESULTS AND BREAKTHROUGH WORK
Experiments presented in Figure 3b conservatively used two views (AP + LAT) as in prior work and compared the geometric accuracy of 3D-2D registration using: standard radiographs (R); simulated radiographs (, forward-projection of the stitched cone-beam CT image); two VLFs (AP + LAT); and a single AP VLF using 3-slot collimation. All cases demonstrated <2 mm mean TRE, and the differences in error distributions were not statistically significant, thus validating the fidelity of radiograph and VLF simulation.
Registration based on a single (AP) VLF in Figure 3d took advantage of the collimated parallax views and yielded results that are comparable to that of 2-view registrations (Figure 3a–c). The collimator angle was set to ϕ = 7°, limited by the current geometry (source-detector distance) and detector size, as larger angles would not intersect the detector. Given that the separation of two extreme views is 2ϕ, the results appear consistent with prior work that showed 10–15° view separation to provide performance equivalent to 90° separation (AP + LAT).10 The resulting registration are also shown below in Figure 4, as trajectory overlays on post-operatively acquired 3D scans.
Figure 4.
Confirmation of pedicle screw placement. Overlay of screw principal axes on screw-axis-aligned (a) axial and (b) sagittal slices of the stitched reconstruction (μ).
4. CONCLUSIONS
The current work demonstrated accurate 3D confirmation and guidance of tasks related to screw placement in spinal deformity correction using VLF imaging. The 3D-2D registration method was shown to be accurate and robust using a single VLF acquisition by taking advantage of a multi-slot projective geometry, thus overcoming the additional time, dose, and workflow disruption required to obtain multiple- (e.g., AP and LAT) radiographs. A limitation of the current work was in projective simulation of VLFs from a stitched 3D volume, whereas eventual implementation would use a multi-slot collimator technique. The stitched volume method supported initial investigation of 3D-2D registration feasibility from a single VLF. We anticipate that future work may improve further on the geometric accuracy demonstrated here, because the volume stitching method incurred a loss in spatial resolution associated with 2×4 detector pixel binning, whereas multi-slot VLF imaging could be acquired with 1×1 (full-resolution) detector readout.
The concept of known components can also extend beyond surgical implants and instruments, such that prior information on anatomy may be used to extract similar 3D pose estimation of anatomical structures. Motivated by extended views presented by VLFs, automatic methods for measuring global spinal alignment (e.g., sagittal or coronal curvature) will be investigated to anatomically assess and guide deformity correction.11 While the presented work was applied to spine surgery, future work will consider alternative applications, such as correction of long-bone fractures, that may benefit from this approach.
ACKNOWLEDGEMENTS
This research was supported by NIH grant R01-EB-017226, NIH T32 AR-067708, and research collaboration with Medtronic. The authors thank Mr. Ronn Wade (Anatomy Board, University of Maryland), and Drs. Rajiv Iyer and Camilo Molina (Department of Neurosurgery, Johns Hopkins University) for assistance with cadaver specimen.
REFERENCES
- [1].Geijer H, Verdonck B, Beckman K-W, Andersson T and Persliden J, “Digital radiography of scoliosis with a scanning method: radiation dose optimization.,” Eur. Radiol 13(3), 543–551 (2003). [DOI] [PubMed] [Google Scholar]
- [2].Deschênes S, Charron G, Beaudoin G, Labelle H, Dubois J, Miron M-C and Parent S, “Diagnostic imaging of spinal deformities: reducing patients radiation dose with a new slot-scanning X-ray imager.,” Spine (Phila. Pa. 1976) 35(9), 989–994 (2010). [DOI] [PubMed] [Google Scholar]
- [3].Bittersohl B, Freitas J, Zaps D, Schmitz MR, Bomar JD, Muhamad AR and Hosalkar HS, “EOS imaging of the human pelvis: reliability, validity, and controlled comparison with radiography.,” J. Bone Joint Surg. Am 95(9), e58 (2013). [DOI] [PubMed] [Google Scholar]
- [4].Glaser DA, Doan J and Newton PO, “Comparison of 3-dimensional spinal reconstruction accuracy: biplanar radiographs with EOS versus computed tomography.,” Spine (Phila. Pa. 1976) 37(16), 1391–1397 (2012). [DOI] [PubMed] [Google Scholar]
- [5].Wright JG, Treble N and Feinstein AR, “Measurement of lower limb alignment using long radiographs.,” J. Bone Joint Surg. Br 73(5), 721–723 (1991). [DOI] [PubMed] [Google Scholar]
- [6].Wang L, Traub J, Weidert S, Heining SM, Euler E and Navab N, “Parallax-free intra-operative X-ray image stitching.,” Med. Image Anal 14(5), 674–686 (2010). [DOI] [PubMed] [Google Scholar]
- [7].Barnes GT and Brezovich IA, “The design and performance of a scanning multiple slit assembly.,” Med. Phys 6(3), 197–204 (1979). [DOI] [PubMed] [Google Scholar]
- [8].Uneri A, Zhang X, Stayman JW, Helm P, Osgood GM, Theodore N and Siewerdsen JH, “Advanced image registration and reconstruction using the O-Arm system: dose reduction, image quality, and guidance using known-component models,” SPIE Med. Imaging, Webster RJ and Fei B, Eds., 43, SPIE; (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
- [9].Uneri A, De Silva T, Goerres J, Jacobson MW, Ketcha MD, Reaungamornrat S, Kleinszig G, Vogt S, Khanna AJ, Osgood GM, Wolinsky J-P and Siewerdsen JH, “Intraoperative evaluation of device placement in spine surgery using known-component 3D-2D image registration,” Phys. Med. Biol 62(8), 3330–3351 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
- [10].Uneri A, Otake Y, Wang AS, Kleinszig G, Vogt S, Khanna AJ and Siewerdsen JH, “3D-2D registration for surgical guidance: effect of projection view angles on registration accuracy.,” Phys. Med. Biol 59(2), 271–287 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
- [11].Ilharreborde B, Steffen JS, Nectoux E, Vital JM, Mazda K, Skalli W and Obeid I, “Angle measurement reproducibility using EOS three-dimensional reconstructions in adolescent idiopathic scoliosis treated by posterior instrumentation.,” Spine (Phila. Pa. 1976) 36(20), E1306–13 (2011). [DOI] [PubMed] [Google Scholar]