Non-circular CBCT orbit design and realization on a clinical robotic C-arm for metal artifact reduction

Abstract

Metal artifacts have been a difficult challenge for cone-beam CT (CBCT), especially for intraoperative imaging. Metal surgical tools and implants are often present in the field of view and can attenuate X-rays so heavily that they essentially create a missing-data problem. Recently, an increasing number of intra-operative imaging systems such as robotic C-arms are capable of non-circular orbits for data acquisition. Such trajectories can potentially improve sampling and the degree of data completeness to solve the metal-induced missing-data problem, thereby reducing or eliminating the associated image artifacts. In this work, we extend our prior theoretical and experimental work and implement non-circular orbits for metal artifact reduction on a clinical robotic C-arm (Siemens Artis zeego). To maximize the potential for clinical translation, we restrict our implementation to standard built-in motion and data collection functions, also available on other zeego systems, and work within the physical constraints and limitations on positioning and motion. Customized software tools for data extraction, processing, calibration, and reconstruction are used. We demonstrate example non-circular orbits and the resulting image quality using a phantom containing pedicle screws for spine fixation. As compared with a standard circular CBCT orbit, these non-circular orbits exhibit significantly reduced metal artifacts. These results suggest a high potential for image quality improvements for intraoperative CBCT imaging when metal tools or implants are present in the field-of-view.

1. INTRODUCTION

Cone-Beam CT (CBCT) has found wide adoption in the clinical workflow of image-guided procedures. However, high density objects such as metallic surgical tools and implants can heavily attenuate the incoming X-rays, resulting in photon starvation and beam hardening. For relatively large objects like implants and tools, metal can eliminate almost all incoming photons, effectively creating a missing data problem. The resulting metal artifacts often appear as streaks in the image volume and can obscure critical anatomical structures, thus complicating treatment. Often, such artifacts occur exactly where high image quality is most desired, such as pedicle screw tips. It is imperative to accurately locate the screw tips during surgery to ensure correct screw placement and avoid adverse complications. To address this problem, many software-based metal artifact reduction (MAR) methods have been proposed¹. Most commonly, MAR relies on algorithmic estimation and interpolation of the missing data based on surrounding good data. However, such data “replacement” methods cannot generally recover the true missing data – particularly for fine features obscured along the ray paths passing through the metal.

To approach this missing data problem, we have demonstrated that non-circular acquisition orbits can increase the sampling completeness around metal objects by imaging in and out of the central plane, effectively looking “over and around” the metal objects^2–4. Modern robotic C-arm systems have enabled the flexibility to potentially realize these noncircular orbits in a clinical setting. Previous investigations have demonstrated the utility of such orbits on experimental test-benches or modified C-arm systems including laborious step-and-shoot procedures with very long acquisition times⁵ or through specialized control systems that are not clinically available⁶. In this work, we present a new implementation of non-circular orbits on a clinical robotic C-arm using standard acquisition functions. We were able to design orbits for metal-tolerant data acquisition and artifact-free reconstruction while maintaining reasonable acquisition times and working within system motion constraints. A vendor-provided tool for access to measurement data was utilized and a custom data processing chain for calibration and reconstruction of data was developed and applied. We demonstrate the performance of two different non-circular orbits in a physical phantom featuring stainless-steel pedicle screws and compare results with a standard circular scan.

2. METHODS

2.1. Orbit design

While various non-circular orbits have been explored for metal tolerance², such trajectories often are typically “over complete” – with more sampling than one would require for an exact reconstruction⁷. In ordinary “no metal” scans, such over-completeness provides more data than required; however, with missing data due to metal, data can still be complete even when some samples are missing. We focus on two such non-circular orbits: 1) The double-circle-plus-arc orbit and 2) a sawtooth orbit (similar to the sinusoidal orbits studied in⁶). Both orbits had a fixed isocenter. The double-circle-plus-arc orbit consisted of two tilted circular scans and an arc. Each of the circles had a fixed ±22° CRAN/CAUD angle throughout a full 360° LAO/RAO rotation, and the additional arc maintained a fixed LAO/RAO angle through a sweep of CRAN/CAUD angles. The sawtooth orbit can be considered a simplified sinusoidal orbit, in which the source oscillated between ±22° CRAN/CAUD for two full cycles while completing a full 360° LAO/RAO rotation. These orbits were chosen for their expected good performance in the presence of metal and relative ease of implementation on the clinical robotic C-arm. A standard circular scan (20s DynaCT Head 70kV) was also acquired as a baseline for comparison. Table 1 and Figure 2 show the properties and diagrams of the achieved orbits.

Table 1.

Properties summary of the achieved orbits.

Scan type	LAO/RAO range	CRAN/CAUD range	Dose (uGy/frame)	Projections	Scan time
Circular	200°	-	1.2	496	20 sec
Sawtooth	360°	±22°	0.36	673	12 min
Double-circle-plus-arc	360° × 2	±22° and 71.5° (arc)	0.36	644 + 133 (arc)	2.5 min

Figure 2.

Diagram of orbits. The location and orientation of the screw model are approximate.

While we succeeded in implementing the designed non-circular orbits with up to 22° in CRAN/CAUD tilt, we achieved this through careful arrangement of the phantom on the patient bed to maximize the range of CRAN/CAUD angles. Such specific preparation may not be possible in a clinical setting, and the range of possible motion will be dependent on many task-specific factors such as location of the target organ, patient posture, and body size. Therefore, it would be necessary to constrain orbit designs with the physical limits of the system. For example, we set up a typical C-spine imaging experiment and identified the maximum range of source positions [Figure 1B]. For this setup, the most restrictive section of the LAO/RAO range limited the maximum negative CRAN/CAUD tilt angle to −14°. This means that the −22° circle in the double-circ-plus-arc orbit was not possible, but, with careful phasing, the sawtooth orbit with ±22° maximum tilt angle was feasible within the constraints [Figure 1B]. Alternately, one may exploit the sampling redundancy in the double-circle-plus-arc orbit and eliminate redundant portions of the circles that exceed the constraints, effectively making a “multi-arc” orbit. For example, if a −22° tilt angle cannot be achieved at −90° LAO/RAO, an equivalent projection can be acquired at +22° CRAN/CAUD and +90° LAO/RAO, which is well within the constraints. In this work, we present results from the full double-circle-plus-arc orbit and the sawtooth orbit for understanding the effect of orbit design. Future orbit designs will incorporate physical constraints.

Figure 1.

(A) Photos of the phantom. The pictured phantom does not exactly match the images due to movement during assembly. (B) Possible source positions for a specific cervical spine imaging setup. Red dots: the maximum CRAN/CAUD tilt from the equator at certain LAO/RAO angles. Blue dots: illustration of a continuous area of possible source positions. Green dots: illustration that the sawtooth orbit fits inside the possible range.

2.2. Test phantom

A test phantom was used that featured three pairs of stainless-steel pedicle screws connected with stainless steel rods, as would be used for a three-level spine fixation procedure, shown in Figure 1A. The screws were placed in a cylindrical plastic container with a 15 cm diameter. 3D-printed radial line-pairs were placed in the axial plane near the middle pair of screws and in the sagittal plane between the screws. The extra space was filled with variably sized plastic spheres for background clutter.

2.3. Zeego implementation

The orbits were performed on a clinical robotic C-arm (Siemens Artis zeego). We opted to implement all orbital motions using standard control functions available on all clinical zeego systems – without requiring the service mode – to demonstrate the potential of this work for clinical translation. A vendor-provided raw data extraction tool and customized data processing, calibration, and reconstruction software was used.

The non-circular orbit projections were acquired with the Digital Radiography (DR) mode at 7.5 frames/sec and 0.36 uGy/frame. The motions were realized differently between the two non-circular orbits. The double-circle-plus-arc orbit was driven manually using the bedside joystick controller. For each circle, the C-arm was first driven to the desired CRAN/CAUD angle and then driven through the LAO/RAO angles at a relatively constant speed. Each circle was broken into three consecutive segments due to a file size limit. An arc was acquired for improved data completeness by fixing the LAO/RAO angle and only tilting in CRAN/CAUD. The sawtooth orbit was realized with Programmed Positions. During the preparation, the C-arm was first driven manually through the desired orbit in segments, and 15 “navigation points” along the trajectory were saved into Programmed Positions. During the acquisition, the C-arm was commanded to move from one to the next navigation point, and the C-arm followed a trajectory calculated by the built-in path-planner at a joystick-controlled speed. However, the path-planner did not always produce a straight line between navigation points. When the flat panel detector got close to the bed due to CRAN/CAUD tilts, the path planner tended to be conservative and calculated a “safer” L-shaped trajectory.

2.4. Geometry calibration

While it would be ideal to reconstruct the images directly using the geometry coordinates recorded by the system, these encoded positions were subject to some error. We conjectured that system vibrations and the relatively complex controls could cause unexpected shifting of the isocenter within the phantom. Thus, for these initial studies, we performed a 3D-2D registration to find a more accurate projection matrix for each frame to demonstrate the best-case performance⁸. First, a conventional circular orbit reconstruction was created using filtered back-projection (FBP). Then, both the reconstruction and the projections were segmented using simple thresholding to retain only the metal regions. For each projection, the system-recorded geometry was used for initialization, and the corrected projection matrix was found iteratively. The metric was the gradient correlation between the segmented projection and the forward projection from the metal volume.

2.5. Reconstruction and MAR

A model-based iterative algorithm was used for reconstruction to accommodate the non-circular geometries. Specifically, we use a Penalized Weighted Least Squares (PWLS) objective with quadratic penalty and ordered subsets⁹. We ran 50 iterations for each reconstruction. The images were reconstructed with 1 mm³ isotropic voxels. We also implemented a simple metal artifact reduction algorithm to further refine the reconstructions¹⁰. The MAR involved the following steps: 1) create an initial reconstruction and segment out the metal volume, 2) threshold metal regions in each projection and forward projection of the previous reconstruction, 3) replace the projection data in metal regions by interpolating from surrounding areas, and 4) create a metal-free reconstruction and add back in the metal volume.

3. RESULTS

3.1. Orbit evaluation

Figure 2 shows diagrams of the orbits implemented. The double circle orbit shows some small discontinuities due to some missed frames and one larger discontinuity in each circle due to a timing mistake during the manual acquisition. The sawtooth orbit overall resembles the planned path. It shows a few downward “CRAN/CAUD spikes”, which occurred when the panel got close to the patient bed. While the programmed positions were technically reachable via a straight line, the path-planner calculated a conservative path that increased the distance between the panel and the bed during motion. Additionally, during the sawtooth acquisition, the path-planner occasionally rotated the source and the panel 180°, which took roughly 20s each time, and the fluoroscopy occasionally stopped mid-segment. Both cases required re-acquiring the current segment. Overall, the sawtooth orbit took roughly 12 minutes to complete, including the time spent on multiple reacquisitions of some segments. The double-circle-plus-arc orbit took 2.5 minutes, including 1 minute for each circle and 0.5 minute for the arc. Redundant frames with similar geometry from overlapping segments were manually excluded.

3.2. Reconstructions

Figure 3 show reconstructions of different orbits with/without applying the MAR. We pick challenging slices in the axial, sagittal, and oblique views to best showcase the capability of non-circular orbits. In the axial view, two screws coincide in the same axial plane with small spheres in between. The sagittal slice lies between the screws, where there are radial line-pairs and no metal but shows strong metal artifacts in the circular scan. The oblique view shows the full profile of the pedicle screws. Both non-circular orbits show significant improvements compared to the circular scan. The without-MAR non-circular reconstructions already reveal most of the spheres and the shape of the screws that are not visible in the circular scan. In the axial plane, the double-circle-plus-arc orbit shows radial line-pairs that are only 7 mm from the body of the screw, which are invisible in the circular scan and barely visible along the edge in the sawtooth scan. The simple MAR further reduced the metal artifacts between and near the screws. The spheres between the screws show more accurate shape and contrast, and the metal artifacts in the sagittal views are virtually eliminated. Comparing the areas near the screw tips, the non-circular scans greatly reduce the metal artifacts and reveal more background features. The MAR then further improves the metal artifact reduction.

Figure 3.

PWLS reconstructions of the images without- and with-MAR correction. The small screws and green lines on the right show the approximate location of the slices shown. The axial views show the middle pair of screws. The oblique views show the side profile of a set of screws. The sagittal views show the central slice between the screws where there should be no metal.

4. DISCUSSION AND CONCLUSION

This work presents an implementation of two non-circular orbits on a clinical robotic C-arm CBCT scanner with a challenging phantom with large metal implant. The results show a significant decrease in metal artifacts with the orbits alone. A simple MAR algorithm can refine the results even further. Both non-circular orbits used much lower dose than the standard head protocol. Furthermore, the acquisition was completed entirely with standard control functions of the Siemens Artis zeego. The double-circle-plus-arc orbit shows the best metal artifact reduction effect, but it may not be usable due to physical constraints. The sawtooth orbit, which was within the physical constraints for a C-spine imaging setup, also delivered promising metal artifact reduction performance. The 12-minute acquisition time for the sawtooth orbit can be shortened by reducing the number of segments and the need for re-acquisitions. While this double-circle-plus-arc may be impractical due to physical constraints, one may exploit the sampling redundancy and eliminate sections outside the physical constraints. Doing so would also reduce the number of frames needed, which would further reduce the already clinically feasible 2.5-minute acquisition time. Results from a similar system showed that sub-minute acquisition time should be possible with a more specialized control system⁶.

We are currently developing a new online geometric calibration method using fiducials of unknown placement, which will eliminate the need for a prior scan and greatly reduce computation time. Other ongoing work includes optimizing the orbit design with the physical constraints, designing a more anthropomorphic phantom, and improving the speed of the workflow.