Known‐component 3D image reconstruction for improved intraoperative imaging in spine surgery: A clinical pilot study

Abstract

Purpose

Intraoperative imaging plays an increased role in support of surgical guidance and quality assurance for interventional approaches. However, image quality sufficient to detect complications and provide quantitative assessment of the surgical product is often confounded by image noise and artifacts. In this work, we translated a three‐dimensional model‐based image reconstruction (referred to as “Known‐Component Reconstruction,” KC‐Recon) for the first time to clinical studies with the aim of resolving both limitations.

Methods

KC‐Recon builds upon a penalized weighted least‐squares (PWLS) method by incorporating models of surgical instrumentation (“known components”) within a joint image registration–reconstruction process to improve image quality. Under IRB approval, a clinical pilot study was conducted with 17 spine surgery patients imaged under informed consent using the O‐arm cone‐beam CT system (Medtronic, Littleton MA) before and after spinal instrumentation. Volumetric images were generated for each patient using KC‐Recon in comparison to conventional filtered backprojection (FBP). Imaging performance prior to instrumentation (“preinstrumentation”) was evaluated in terms of soft‐tissue contrast‐to‐noise ratio (CNR) and spatial resolution. The quality of images obtained after the instrumentation (“postinstrumentation”) was assessed by quantifying the magnitude of metal artifacts (blooming and streaks) arising from pedicle screws. The potential low‐dose advantages of the algorithm were tested by simulating low‐dose data (down to one‐tenth of the dose of standard protocols) from images acquired at normal dose.

Results

Preinstrumentation images (at normal clinical dose and matched resolution) exhibited an average 24.0% increase in soft‐tissue CNR with KC‐Recon compared to FBP (N = 16, P = 0.02), improving visualization of paraspinal muscles, major vessels, and other soft‐tissues about the spine and abdomen. For a total of 72 screws in postinstrumentation images, KC‐Recon yielded a significant reduction in metal artifacts: 66.3% reduction in overestimation of screw shaft width due to blooming (P < 0.0001) and reduction in streaks at the screw tip (65.8% increase in attenuation accuracy, P < 0.0001), enabling clearer depiction of the screw within the pedicle and vertebral body for an assessment of breach. Depending on the imaging task, dose reduction up to an order of magnitude appeared feasible while maintaining soft‐tissue visibility and metal artifact reduction.

Conclusions

KC‐Recon offers a promising means to improve visualization in the presence of surgical instrumentation and reduce patient dose in image‐guided procedures. The improved soft‐tissue visibility could facilitate the use of cone‐beam CT to soft‐tissue surgeries, and the ability to precisely quantify and visualize instrument placement could provide a valuable check against complications in the operating room (cf., postoperative CT).

1. Introduction

Intraoperative imaging systems with fluoroscopic and three‐dimensional (3D) imaging capability have become increasingly prevalent as a means of image guidance and navigation. Cone‐beam computed tomography (CBCT) systems such as the O‐arm [Fig. 1(a), Medtronic, Littleton MA] and mobile C‐arms (e.g., RFD 3D, Ziehm, Nuremberg, Germany; and Cios Spin, Siemens, Erlangen, Germany) have seen principal applications in cranial and spinal neurosurgery and orthopedic surgery, where up‐to‐date depiction of patient anatomy is desired to enable navigation prior to surgical intervention and to assess the delivery of the surgical product.^{1, 2, 3} For procedures involving the placement of surgical tools and implants, imaging near the end of the operation aids in verifying the proper placement of instrumentation and the detection of potential soft‐tissue disruption or hemorrhage, permitting timely revision during the procedure if necessary. While intraoperative CBCT is generally considered sufficient with respect to visualization of bone, the image quality can be improved and extended to more challenging imaging tasks by addressing two main limitations: (a) image noise confounding soft‐tissue visibility, as illustrated in Fig. 1(b); and (b) metal artifacts about surgical instrumentation that obscure nearby anatomical structures, as illustrated in Fig. 1(c).

Figure 1.

(a) The O‐arm system in the clinical setting. (b) Illustration of three‐dimensional image quality using filtered backprojection, showing fairly clear depiction of bone morphology but a level of image noise that limits soft‐tissue visibility (Patient 2 in Table 1). (c) Illustration of metal artifact about a pedicle screw, challenging the assessment of screw placement within the narrow corridor of the spinal pedicle (Patient 6 in Table 1). The images are displayed in arbitrary units (AU) of image intensity. [Color figure can be viewed at wileyonlinelibrary.com]

An additional point of motivation is reduction of radiation dose. While exposure to intraoperative staff from CBCT is minimized with proper workflow and shielding during scans, dose to the patient remains an area for improvement, especially in pediatrics and scenarios that would benefit from multiple scans. Conventional image reconstruction using 3D‐filtered backprojection (FBP) is limited in its dose reduction capability due to noise and artifacts — a point motivating the development of model‐based image reconstruction (MBIR) approaches. Such methods have facilitated widespread dose reduction strategies in diagnostic CT^{4, 5, 6} and offer similar opportunity to enhance image quality and reduce dose in intraoperative CBCT. By modeling system physics and measurement noise as well as incorporating advanced prior knowledge,^{7, 8, 9, 10, 11, 12} MBIR has demonstrated improved CBCT performance for low‐dose and/or sparse‐view acquisition techniques.^{13, 14}

Metal artifact reduction (MAR) methods have been widely demonstrated in diagnostic CT^{15, 16, 17, 18} and implemented in CBCT in applications such as dental and maxillofacial imaging, head and neck surgery, and image‐guided radiation therapy.^{19, 20, 21} A common approach in MAR involves identification of regions corresponding to metal structures and correction of projection data therein, but accurate delineation of such structures in projection data is challenged by overlying anatomical structures, and delineation in the image reconstruction is challenged by the very artifacts that the methods aim to reduce. An approach to mitigate such challenges is to invoke prior information regarding the shape of interventional devices known to be in the field of view (FOV), defined either by manufacturer‐specific device models (e.g., a particular model of pedicle screw) or by a parametric model (e.g., a generic model of spinal fixation rod). Precise localization of such “known components” (KC) can be achieved by 3D–2D registration of the 3D device model and 2D projection data as in Uneri et al.^{22, 23} Combining KC registration with correction of the 2D projection yields a “KC‐MAR” approach that has shown substantial improvement in visualization of anatomical structures adjacent to metal instrumentation.^{24, 25}

MBIR also offers new capabilities for artifact reduction, since the underlying system model can be adapted to account for measurement biases that give rise to artifacts while maintaining the advantages in noise performance associated with a statistical model‐based approach, as mentioned above. For instance, constraints can be imposed on the objective function^{26, 27, 28} to accommodate nonlinear effects induced by known metal objects once their poses are identified. Alternatively, the Known‐Component Reconstruction (KC‐Recon) method provides a joint framework for pose estimation of devices and penalized likelihood (PL) or penalized weighted least‐squares (PWLS) reconstruction of the surrounding anatomy.²⁹ KC‐Recon has been subsequently extended to support deformable or approximate KC models,^{30, 31, 32, 33} integration of a polyenergetic forward model,³⁴ and estimation of the energy‐dependent properties of the component via a spectral transfer function (STF).³⁵ In simulation and phantom studies, KC‐Recon demonstrated combined benefits of metal artifact reduction, improved noise–resolution characteristics, and the potential for strong dose reduction.

The studies below report a modification of the KC‐Recon algorithm to better account for components with more than one material type and translate the method for the first time to clinical studies of patients receiving spine surgery. Studies were conducted on a clinical system for intraoperative CBCT (O‐arm, Medtronic, Littleton MA), including 17 patients treated with a variety of pedicle screw instrumentation. Pedicle screw component models were provided by manufacturers, Medtronic (Littleton MA) and Globus Medical (Audubon PA). The KC‐Recon algorithm was adapted to the O‐arm system and described below in general terms by which the algorithm simplifies to PWLS reconstruction in the absence of instrumentation. A modification to previous implementation of KC‐Recon concerns the STF, which eases the need for precisely knowing the component content and x‐ray beam quality. In this work, we further generalized the spectral model to accommodate beam hardening induced by components comprising different metal types. Building from recent technical assessment of potential O‐arm imaging performance with MBIR,³⁶ we applied the KC‐Recon method to clinical images to investigate the gains in both soft‐tissue contrast resolution, potential dose reduction, and mitigation of metal artifacts in spine surgery.

2. Materials and Methods

2.A. Intraoperative CBCT on the O‐arm

2.A.1. The O‐arm imaging system

The clinical study utilized the O‐arm system deployed at our institution for intraoperative CBCT, as shown in Fig. 1(a). The system features a breakable O‐shaped gantry ring with a 96.5‐cm diameter inner bore. The rotating anode x‐ray tube (Model B100 with A132 insert, Varian Medical Imaging, Salt Lake City, UT) is powered by a 32‐kW generator and has inherent and added filtration equivalent to 0.7 and 4 mm Al, respectively. The flat‐panel detector (PaxScan 4030D, Varian Medical Imaging, Palo Alto, CA) employs a detector matrix of 1536 × 2048 pixel at 194‐µm pixel pitch with a CsI:Tl scintillator. System geometry entails a source–axis distance of 64.8 cm and source–detector distance of 116.8 cm. Standard clinical technique protocols involve ~391 projections (or ~745 projections for high definition) over 360° acquired at 120 kV with 186–596 mAs (16.3–52.4 mGy) using small focal spot and detector readout in dual‐gain mode with 2 × 4 pixel binning.

2.A.2. Clinical study

The study was conducted under Institutional Review Board (IRB) approval involving 17 participants enrolled with informed consent from the population of patients receiving spine surgery at our institution, as shown in Table 1. The age of participants ranged from 24 to 82 yr old, with a mean age of 55.3 yr old; seven participants were men and ten were women. Images were collected using standard clinical protocols for each participant, including one CBCT image acquired near the beginning of the case prior to the placement of surgical implants for purposes of navigation [referred to as “preinstrumentation,” Fig. 1(b)] and another acquired near the end of the case following device placement for purposes of quality assurance [referred to as “postinstrumentation,” Fig. 1(c)].

Table 1.

Case collection for the clinical study of KC‐Recon imaging performance.

Patient #	Preinstrumentation	Postinstrumentation	Treated levels	Spinal systems	Pedicle screws
					Diameter (mm)	Length (mm)	# screws
1	✓	–	L2–5	–	–	–	–
2	✓	–	L4–5	–	–	–	–
3	✓	–	L4–5	–	–	–	–
4	✓	–	L4–5	–	–	–	–
5	✓	–	L5, S1	–	–	–	–
6	✓	✓	T11–12	Solera	4.0	40	2
			L2–3			45	5
7	✓	✓	T4–6	Solera	5.0	35	4
						40	2
8	–	✓	T10–11	Solera	5.5	40	2
						45	2
			L1–2		6.5	40	4
9	✓	✓	L4–5	Voyager	7.5	50	4
10	✓	✓	T7–8	CREO	5.0	40	8
			T10–11
11	✓	✓	T7–11	CREO	5.0	35	3
						40	2
					5.5	35	3
						40	1
12	✓	✓	L2	CREO	5.5	45	2
13	✓	✓	L5, S1	CREO	5.5	35	1
						40	1
						45	2
14	✓	✓	L4–5	CREO	6.5	50	2
						55	2
15	✓	✓	L3–S1	CREO	6.5	35	1
						40	3
						45	2
						50	2
16	✓	✓	L4–S1	CREO	6.5	40	1
						45	1
						50	4
17	✓	✓	T12–L2	CREO MIS	5.5	45	3
						50	3
Total							72

As summarized in Table 1, the postinstrumentation images captured a total of 72 pedicle screws, varying from 4.0 to 7.5 mm in diameter and 35 to 55 mm in length. The screws are designed for open or percutaneous procedures and are from four different spinal systems produced by two manufacturers. The screw dimensions, therefore, reflect a representative range in common clinical use. Models of screw components were obtained from the manufacturers.

All images were reconstructed using KC‐Recon and compared to images reconstructed by the modified 3D FBP algorithm detailed in previous work.³⁶ Preinstrumentation images (for which KC‐Recon reduces simply to PWLS, detailed in Section 2.B.2) were evaluated with respect to soft‐tissue visibility, and postinstrumentation images were evaluated with respect to localization and visualization of surgical instrumentation (viz., spinal pedicle screws) and surrounding tissues. For spine discectomy or decompression procedures not involving instrumentation (Patients 1–5 in Table 1), only “preinstrumentation” scans were acquired. For Patient 8, the preinstrumentation image was lost to offline analysis by a data export error, with no effect on the patient or surgery. One case was excluded from the study, because the O‐arm was operated in a dual isocenter, expanded FOV mode not currently implemented for the KC‐Recon system model.

2.B. Known‐component reconstruction

The KC‐Recon algorithm is illustrated in Fig. 2. The main stages of the algorithm include estimation of component poses and spectral coefficients and reconstruction of the anatomical background, with the PWLS approach^{11, 37} accompanied by ordered subsets and separable quadratic surrogates (OS‐SQS)³⁸ serving as the basis for 3D image reconstruction.

Figure 2.

Flowchart for the KC‐Recon algorithm. For cases involving known instrumentation, the algorithm estimates pose and spectral coefficients of (typically metal) components. For cases without known instrumentation, the method reduces simply to penalized weighted least‐squares reconstruction. [Color figure can be viewed at wileyonlinelibrary.com]

2.B.1. The KC‐Recon framework

If known a priori, the geometric shape and material composition of metal instrumentation can be incorporated into the reconstruction algorithm via a decomposed object model, as in previous work:²⁹

$$\mu =\left(\prod _{n=1}^N\mathbf{D}\left\{T\left({\lambda }^{\left(n\right)}\right)s^{\left(n\right)}\right\}\right){\mu }_{\ast }+\sum _{n=1}^{N}T\left({\lambda }^{\left(n\right)}\right){\mu }_I^{\left(n\right)}$$ (1)

where ${\mu }_{\ast }$ is the unknown background anatomy and $N$ is the number of components with known attenuation distributions ${\mu }_I^{\left(n\right)}$. A parameter vector ${\lambda }^{\left(n\right)}$ denotes the pose of the $n\text{th}$ component, and the transformation $T\left(·\right)$ represents the registration of the component within the image volume. The binary masks $s^{\left(n\right)}$ are set to zero in regions containing the components and are applied to the background anatomy to rule out their effects on the reconstructed image. The operator $\mathbf{D}\left\{·\right\}$ is used to form a diagonal matrix from a vector. For simplicity, the background anatomy excluding the components is denoted:

$${\mu }_{\ast }\left(\mathrm{\Lambda }\right)=\left(\prod _{n=1}^N\mathbf{D}\left\{T\left({\lambda }^{\left(n\right)}\right)s^{\left(n\right)}\right\}\right){\mu }_{\ast }$$ (2)

where $\mathrm{\Lambda }={\left\{{\lambda }^{\left(n\right)}\right\}}_{n=1}^N$ represents the set of parameter vectors for all components. The object decomposition in Eq. (1) permits a more versatile forward model in which propagation of x rays through the anatomical background and through the known components can be modeled separately, yielding:

$$\overline{y}=\mathbf{D}\left\{g\right\}exp\left(-\mathbf{A}{\mu }_{\ast }\left(\mathrm{\Lambda }\right)\right)f\left(\sum _{n=1}^{N}T\left({\lambda }^{\left(n\right)}\right){\mu }_I^{\left(n\right)}\right)$$ (3)

where $\overline{y}$ are the mean projection measurements, $g$ is the unattenuated x‐ray fluence and detector gain, and $\mathbf{A}$ is the system matrix. Due to the polyenergetic nature of the x‐ray beam, a monoenergetic model (Beer's law) is often insufficient to describe spectral effects of high‐density materials such as metal. Using this model alone will therefore induce measurement bias, giving rise to beam hardening and streak artifacts in the reconstruction.³⁹ The factorized forward model in Eq. (3) accommodates this situation by utilizing a monoenergetic model for the background anatomy ${\mu }_{\ast }\left(\mathrm{\Lambda }\right)$, for which beam‐hardening effects are assumed minimal, and an STF denoted $f\left(·\right)$ for the known components ${\mu }_I^{\left(n\right)}$, for which spectral effects are much stronger.

The STF described in Xu et al.³⁵ does not require explicit knowledge or estimation of the x‐ray source spectrum, and it does not require a priori knowledge of component composition. Together with estimation of the background anatomy, the method was previously shown to outperform precalibration approaches in artifact mitigation by accounting for additional physical effects (e.g., beam hardening caused by anatomical structures). In this work, it was extended to estimate unknown material compositions while modeling beam hardening effects that arise from multipart components of different material composition.⁴⁰ In this study, the component refers to the pedicle screw, which generally contains a titanium shaft and a polyaxial cap composed of a different metal (e.g., cobalt chrome or stainless steel). As a result, the STF was modified to account for the spectral effect of $M$ number of materials, where $M=2$ for the current case:

$$f\left(p_m;{\kappa }_m\right)=exp\left(\sum _{i=1}^K{\tilde{\kappa }}_i{\left(\tilde{p}\right)}^i+\sum _{m=1}^{M=2}\sum _{i=1}^K{\kappa }_{m,i}{\left(p_m\left(\mathrm{\Lambda }\right)-\tilde{p}\right)}^i\right)$$ (4)

$$p_m\left(\mathrm{\Lambda }\right)=\sum _{n=1}^N\mathbf{A}T\left({\lambda }^{\left(n\right)}\right)b_m^{\left(n\right)},\tilde{p}=p_1\left(\mathrm{\Lambda }\right)\cap p_2\left(\mathrm{\Lambda }\right)$$ (5)

where the path length $p_m\left(\mathrm{\Lambda }\right)$ spans all component parts made up of material $m$ and is partially modified by the spectral coefficients ${\kappa }_{m,i}$, and $K$ is the order of the polynomial. The $b_m^{\left(n\right)}$ term represents a binary mask that sets regions containing the material in $n\text{th}$ component to a value of one, such that$\sum _m{b}_m^{\left(n\right)}$ is the complement of $s^{\left(n\right)}$. The $\tilde{p}$ term denotes the overlapping region (intersection) between the path lengths for materials 1 and 2. This overlapping region is removed from each path length such that the spectral coefficients ${\kappa }_{m,i}$ only modify the portion of the path length that spans material $m$. A separate set of spectral coefficients was used for $\tilde{p}$, conveniently modeling x‐ray propagation through two materials. When the component contains parts made of only one material ($M=1$, $\tilde{p}=0$), the STF in Eq. (4) reduces to the form in previous work:³⁵

$$f\left(p_m;{\kappa }_m\right)=exp\left(\sum _{i=1}^K{\kappa }_{m,i}{\left(p_m\left(\mathrm{\Lambda }\right)\right)}^i\right)$$ (6)

Additionally, when $K$ is set to 1, the function further reduces to a monoenergetic model, yielding ${\kappa }_1$ equal to the negative attenuation value of material $m$.

With the forward model in Eqs. (3) and (4), the objective function for reconstruction that simultaneously estimates the background anatomy ${\widehat{\mu }}_{\ast }$, spectral coefficients $\widehat{\kappa }$, and registration parameters $\widehat{\mathrm{\Lambda }}$ is defined:

$$\begin{array}{cc}\hfill \left\{{\widehat{\mu }}_{\ast },\widehat{\kappa },\widehat{\mathrm{\Lambda }}\right\}& =argmin\mathrm{\Phi }\left({\mu }_{\ast },\kappa ,\mathrm{\Lambda };y\right)\hfill \\ & =argminL\left({\mu }_{\ast },\kappa ,\mathrm{\Lambda };y\right)-\beta R\left({\mu }_{\ast }\right)\hfill \end{array}$$ (7)

$$\begin{array}{c}L\left({\mu }_{\ast },\kappa ,\mathrm{\Lambda };y\right)=∥\mathbf{A}{\mu }_{\ast }\left(\mathrm{\Lambda }\right)-\sum _{i=1}^K{\tilde{\kappa }}_i{\left(\tilde{p}\right)}^i\\ -\sum _{m=1}^{M=2}\sum _{i=1}^K{\kappa }_{m,i}{\left(p_m\left(\mathrm{\Lambda }\right)-\tilde{p}\right)}^i\\ {-log\left(\mathbf{D}\left\{g\right\}{\overline{y}}^{-1}\right)∥}_W^2\\ \end{array}$$ (8)

where $L$ denotes the data fidelity term involving a weighted L² norm relating the forward projection of the background anatomy $\mathbf{A}{\mu }_{\ast }\left(\mathrm{\Lambda }\right)$ and the spectrally modified path lengths to the line integrals through the object $log\left(\mathbf{D}\left\{g\right\}{\overline{y}}^{-1}\right)$. An edge‐preserving roughness penalty was chosen for $R\left({\mu }_{\ast }\right)$ using a Huber function of pairwise voxel differences over a first‐order neighborhood with a penalty strength adjusted by the parameter $\beta$.⁴¹

As in previous work,³⁵ to ease the computational load, the registration parameters were solved prior to estimation of the spectral coefficients and reconstruction of the background anatomy, as illustrated in Fig. 2. Component poses were estimated by maximizing gradient correlation (GC) between projections and the path length of the registered material volume $p_m\left(\mathrm{\Lambda }\right)$ computed at the same $\theta$ gantry angles:^{22, 23}

$$\widehat{\mathrm{\Lambda }}=argmax\sum _{\theta }\text{GC}\left(y_{\theta },p_{m,\theta }\left(\mathrm{\Lambda }\right)\right)$$ (9)

The solution was initialized from screw trajectory plans drawn on the preinstrumentation image and iteratively solved using the covariance matrix adaptation evolution strategy (CMA‐ES) optimizer. While the orientation (pose) of metal screw shafts and/or caps may vary significantly, the registration method has been shown to capture a wide range of variations in component pose, shape, and even deformation.^{23, 42} Leveraging the estimated registration parameters $\widehat{\mathrm{\Lambda }}$, the spectral coefficients $\widehat{\kappa }$ were solved via iterative coordinate descent (ICD) after every iteration of the reconstruction.

2.B.2. KC‐Recon in the absence of instrumentation

As suggested in Fig. 2, the KC‐Recon framework is general to imaging with or without instruments in the sense that it reduces to PWLS for N = 0 (i.e., scans without known components). For preinstrumentation images, the framework thereby offers improvement to soft‐tissue image quality and/or reduction of radiation dose.¹⁴ For N = 0 components, parameterization in Eq. (1) simplifies to an object model $\mu$ composed entirely of background anatomy (i.e., $\mu ={\mu }^{\ast }$, and ${\mu }_I=0$). Therefore, terms involving the component spectral coefficients and registration parameters are removed, and the data fidelity term $L$ in Eq. (7) reduces to the PWLS estimator:

$$\widehat{\mu }=\arg \min \mathrm{\Phi }\left(\mu ;y\right)=argminL\left(\mu ;y\right)-\beta R\left(\mu \right)$$ (10)

$$L\left(\mu ;y\right)={∥\mathbf{A}\mu -log\left(\mathbf{D}\left\{g\right\}{\overline{y}}^{-1}\right)∥}_W^2$$ (11)

For both pre‐ and postinstrumentation images, the reconstruction process was initialized with an FBP image (with no MAR performed on the known components) and iteratively solved via OS‐SQS³⁸ using 50 iterations of 20 subsets followed by 50 iterations of ten subsets. The separable footprint projector was used for the forward‐ and backprojection.⁴³ Lateral truncation resulting from large patient size and off‐center positioning was mitigated using an elliptical cylinder support of constant attenuation coefficient (water) with center and axes estimated from the projection data.⁴⁴

2.B.3. FBP and basic MAR

FBP reconstruction was based on the Feldkamp–Davis–Kress algorithm⁴⁵ using the same elliptical cylinder support to extrapolate projection data outside the FOV and reduce truncation artifacts.¹⁰ A 2D Hann filter with adjustable cutoff frequency $f_{\text{c}}$ in both $u$ and $v$ directions was implemented as in Uneri et al.³⁶ to give more isotropic noise and resolution characteristics.

For postinstrumentation images, a basic MAR approach based on FBP (denoted “basic MAR”) was also implemented for comparison to KC‐Recon. Metal screws in an initial FBP reconstruction were segmented using intensity thresholding and reprojected to localize metal in the sinogram space. The sinogram data were then replaced by synthesized projections generated by 2D interpolation of pixel values from neighboring regions, and reconstruction by FBP was performed as described above.

For all reconstruction methods, images were reconstructed on a volumetric grid covering (21.2 × 21.2 × 16.0) cm³ (not counting the cylinder support) with isotropic voxel size of (0.4 × 0.4 × 0.4) mm³ using GPU‐accelerated projection, filtering, and regularization operations. Additional speedup for iterative reconstruction was achieved via a multiresolution scheme⁴⁶ in which the peripheral regions of the extended FOV were reconstructed at a coarser voxel size of (1.6 × 1.6 × 1.6) mm³. The same parameter settings were used in pre‐ and postinstrumentation cases.

2.C. Low‐dose simulation

The performance of KC‐Recon was further assessed in low‐dose (LD) imaging scenarios. Using images acquired at normal dose in the clinical study, LD images were simulated using the method reported by Wang et al.,⁴⁷ where quantum noise, system blur (correlation), and electronic noise were modeled according to cascaded linear systems analysis and injected to the projection data to yield image quality that realistically reflected image quality at reduced dose. The LD simulation method was validated using O‐arm images acquired at various dose levels (6.5–41.2 mGy) for a body phantom and cadaver torso. FBP and KC‐Recon images were reconstructed from the simulated LD projections and compared to corresponding reconstructions of real images in terms of noise and contrast characteristics. Details of LD simulation method are in Wang et al.,⁴⁷ extended to the O‐arm system in Zhang et al.⁴⁸

2.D. Image quality evaluation

For preinstrumentation images, the spatial resolution and soft‐tissue contrast‐to‐noise ratio (CNR) were evaluated in KC‐Recon (i.e., PWLS for N = 0 components) and FBP. The CNR was also evaluated as a function of dose using LD images simulated from clinical data. Postinstrumentation images were assessed in terms of the magnitude of blooming and shading artifacts caused by implanted pedicle screws, and LD simulation was employed to investigate potential pitfalls of the method in noisy data. Two‐tailed paired t tests were used in all comparisons to determine whether there was significant difference, and a P < 0.05 was considered statistically significant.

2.D.1. Preinstrumentation images

Spatial resolution

Spatial resolution was evaluated with respect to the edge of homogeneous soft‐tissue structures measured at the same location in all reconstructions for a given patient. Soft‐tissue structures included the paraspinal muscles and low‐contrast solid organs (e.g., liver and kidney) with interstitial fat used as background. As shown in Figs. 3(a) and 3(b), the tissue edge was obtained by Canny edge detection, and the spatial resolution was defined as the width of the edge spread function (ESF) fitted to measurements from line profiles perpendicular to the edge:

$$f\left(x\right)=a-\frac{c}{2}\text{erf}\left(\frac{x-r}{\sqrt{2}{\sigma }_{\text{ESF}}}\right)$$ (12)

where ${\sigma }_{\text{ESF}}$ denotes the ESF width, $r$ represents the distance to the tissue–background edge, $c$ is the tissue contrast, and $a$ approximates the tissue attenuation value. To enable fair comparison of image quality across reconstruction techniques,^{49, 50} the spatial resolution of each image volume was adjusted by tuning the cutoff frequency $f_{\text{c}}$ for FBP and the penalty strength $\beta$ for KC‐Recon to match ${\sigma }_{\text{ESF}}$. Since spatial resolution was independent of dose, $f_{\text{c}}$ used for reconstructing the full‐dose FBP was propagated to FBP images at all simulated LD levels. Conversely, the value of $\beta$ was individually adjusted for KC‐Recon at each dose level to achieve equivalent ${\sigma }_{\text{ESF}}$ in all cases. While ESF width measured at one tissue edge is not representative of spatial resolution of the entire volume, it provides a reasonable starting point for resolution matching within the area of interest in anatomical images.

Figure 3.

(a) Paraspinal muscle‐fat edge (yellow) and perpendicular line profiles (cyan) for assessment of spatial resolution in soft‐tissue boundaries. (b) Example ESF fitted to line profiles along a muscle‐fat edge, with spatial resolution characterized by the width, ${\sigma }_{\text{ESF}}$. (c) Example regions of interest (ROI) (yellow box) for the measurement of the blooming ratio ${\text{R}}_{\text{bloom}}$ defined as the ratio of (d) ${\text{D}}_{\text{ROI}}$ measured at the mean signal half amplitude range compared to the true diameter of the screw, ${\text{D}}_{\text{true}}$. [Color figure can be viewed at wileyonlinelibrary.com]

2.D.1.1.

Soft‐tissue contrast‐to‐noise ratio

Soft‐tissue CNR was evaluated with respect to the same tissues as used for the soft‐tissue edge spread analysis. Regions of interest (ROIs) covering approximately (10 × 10 × 2) mm³ were selected within homogeneous regions of the soft‐tissue structure and adjacent fat background, and CNR was calculated as follows:⁵¹

$$\text{CNR}=\frac{\left|{\overline{\text{ROI}}}_{\text{tissue}}-{\overline{\text{ROI}}}_{\text{background}}\right|}{{\sigma }_{\text{background}}}$$ (13)

where ${\overline{\text{ROI}}}_x$ is the average voxel value within the ROI of material x (i.e., tissue or background), and ${\sigma }_{\text{background}}$ is the standard deviation of voxel values in the background ROI. The improvement in CNR between reconstruction methods was further summarized in terms of the factor:

$${\text{Q}}_{\text{CNR}}=\frac{{\text{CNR}}_{\text{PWLS}}}{{\text{CNR}}_{\text{FBP}}}$$ (14)

where ${\text{CNR}}_{\text{PWLS}}$ and ${\text{CNR}}_{\text{FBP}}$ denote the CNR measured in the KC‐Recon (i.e., PWLS) and FBP images, respectively, for a given patient. Both CNR and ${\text{Q}}_{\text{CNR}}$ are simple metrics that quantitatively characterize performance with respect to large‐area, low‐contrast soft‐tissue visibility.

2.D.2. Postinstrumentation images

Metal artifact (blooming)

One aspect of metal artifact is an apparent widening (“blooming”) of the screw shaft, as illustrated in Fig. 1(c), hindering the visibility of adjacent anatomical structures, such as the pedicle corridor and spinal cord, which are important for assessment of pedicle breach. Using the screw poses estimated from the registration process described in Section 2.B.1, an ROI of (61 × 61 × 5) voxels centered on the pedicle along the long axis of the screw was selected for each screw in a reconstructed volume, as illustrated in Fig. 3(c). The mean signal profile of the ROI orthogonal to the screw axis was computed, as illustrated in Fig. 3(d), and the full‐width at half amplitude range (HAR) of the ROI mean signal was computed:

$$\text{HAR}=\frac{s_{max}+s_{min}}{2}$$ (15)

where $s_{max}$ and $s_{min}$ denote the maximum and minimal value of the mean signal. The apparent width of the “blooming” region, denoted ${\text{D}}_{\text{ROI}}$, was measured at the HAR and the ratio to the true diameter of the screw, denoted ${\text{D}}_{\text{true}}$, was computed:

$${\text{R}}_{\text{bloom}}=\frac{{\text{D}}_{\text{ROI}}}{{\text{D}}_{\text{true}}}$$ (16)

The blooming ratio, denoted ${\text{R}}_{\text{bloom}}$, is close to 1 for little or no blooming artifact.

Metal artifact (shading/streak)

Similarly, the dark shading/streak artifact at the screw tip can diminish the detection of anterior breach or impingement on vital structures. The relative attenuation change, denoted ${\mathrm{\Delta }}_{\mu }$, was used to determine the percentage increase/decrease of the voxel value near the screw tip, defined as:

$${\mathrm{\Delta }}_{\mu }=\frac{{\mu }_{\text{tip}}-{\mu }_{\text{ref}}}{{\mu }_{\text{ref}}}\times 100\%$$ (17)

where ${\mu }_{\text{tip}}$ denotes the mean signal of a polygonal ROI selected immediately anterior to the screw tip within cancellous bone or fat, and ${\mu }_{\text{ref}}$ was computed in a nearby region containing the same type of tissue but uncontaminated by shading.

3. Results

3.A. Preinstrumentation images

Figure 4 shows representative images acquired prior to instrumentation and reconstructed using FBP and KC‐Recon (i.e., PWLS). For each case, FBP and PWLS reconstructions were adjusted to match spatial resolution at the paraspinal muscle‐fat edge by varying $f_{\text{c}}$ over the range 0.3–0.6 for FBP and setting $log\beta$ = −2.5 for PWLS. Since the locations of the selected muscle‐fat regions affect the measured ESF width (±0.15 mm within the same patient), the distance between the muscle‐fat edge and the spine was fixed in all reconstructions for a given patient, and the same edge was used for measurement. The range of ESF width varied from approximately 0.4 to 0.9 mm (matched for each comparison of FBP and PWLS) over all cases, owing to variation in the muscle‐fat edge location and other factors of fluence and penalty strength. Images in Fig. 4 are sorted left to right in terms of normal clinical dose (D_full), ranging 20.9–52.4 mGy. For Patients 16 and 12 shown in Figs. 4(b) and 4(d), a metallic reference frame placed over the patient (for registration to the surgical navigation system) resulted in artifacts visible as vertical banding in the coronal images. Artifacts from involuntary motion of bowel gas are evident in Fig. 4(c). At the D_full dose levels for these standard clinical protocols, both FBP and PWLS reconstructions gave reasonable visualization of soft‐tissues adjacent to the spine, with the latter showing visibly reduced noise (quantified below) and slightly improved delineation of soft‐tissue boundaries, owing to edge‐preserving Huber regularization.

Figure 4.

Example coronal images of four preinstrumentation images acquired at standard clinical protocols (varying in D_full from 20.9 to 52.4 mGy). Filtered backprojection (FBP) and KC‐Recon [i.e., penalized weighted least‐squares (PWLS)] are shown at the top and bottom for each case, respectively. Spatial resolution was matched for all volumes with respect to the tissue edge marked by the yellow lines. The window and level were separately varied as shown in the colorbar below each column (but fixed for the FBP and PWLS comparison in each case) for visualization of soft‐tissues adjacent to the spine. [Color figure can be viewed at wileyonlinelibrary.com]

Figure 5 shows visual comparison of imaging performance for FBP and PWLS at dose levels lower than D_full, simulated via the LD process described in Section 2.C. Preinstrumentation images acquired at D_full = 26.2 mGy were used as the initial data for simulating scans down to 3.3 mGy. An ESF width of 0.52 mm was achieved at the muscle‐fat edge using $f_{\text{c}}$ = 0.37 for all FBP images and $log\beta$ ranging from −1.8 to −2.5 for PWLS images over the 3.3–26.2 mGy dose range. The FBP images are seen to decline rapidly in soft‐tissue visibility below 16.4 mGy due to increased image noise. PWLS maintained clearer soft‐tissue delineation down to the lowest dose level.

Figure 5.

Comparison of filtered backprojection and simplified KC‐Recon (i.e., penalized weighted least‐squares) images reconstructed from projections at full (D_full) and simulated doses (D_sim). Spatial resolution was matched for all volumes with respect to the tissue edge marked by the yellow lines. [Color figure can be viewed at wileyonlinelibrary.com]

The improved soft‐tissue contrast resolution is quantified in Fig. 6. The CNR improvement (${\text{Q}}_{\text{CNR}}$) measured over 16 patients is shown in Fig. 6(a), suggesting an overall 24% gain in CNR at equivalent dose and resolution (P = 0.02). The improvement was further validated by analyzing the CNR as a function of ${\sigma }_{\text{ESF}}$ for FBP and PWLS as shown in Fig. 6(b), where PWLS exhibits consistently higher CNR than FBP at any level of ESF width. Greater CNR enhancement enabled by PWLS was observed at lower dose levels (simulated 3.3–16.4 mGy). As illustrated in Fig. 6(c), the CNR improvement is >1 for all cases and generally increases for lower dose levels, showing the increased advantage of PWLS at lower dose. For some cases (Patients 9 and 12), the ${\text{Q}}_{\text{CNR}}$ was invariant or even decreased at lower dose (though still >1) due to image artifacts affecting the measurement region.

Figure 6.

Quantitative analysis of CNR performance in preinstrumentation images. (a) Boxplot distribution of ${\text{Q}}_{\text{CNR}}$ values measured over 16 patients in soft‐tissue ROIs at matched spatial resolution and dose levels corresponding to standard clinical protocols, showing a mean value of 1.24 in ${\text{Q}}_{\text{CNR}}$. (b) Noise–resolution performance (CNR vs ${\sigma }_{\text{ESF}}$) evaluated by sweeping FBP cutoff frequency $f_{\text{c}}$ and PWLS penalty strength $\beta$, showing improved CNR for PWLS at any level of spatial resolution. (c) ${\text{Q}}_{\text{CNR}}$ evaluated as a function of dose for five patients. The dose level of the clinical scan (D_full) is marked by solid circles, and simulated low‐dose levels (D_sim) are represented by open circles. [Color figure can be viewed at wileyonlinelibrary.com]

3.B. Postinstrumentation images

Figure 7 compares the performance of FBP, PWLS, basic MAR, and KC‐Recon in the presence of bilaterally placed spine pedicle screws. PWLS showed slight image noise reduction but had little or no effect on reducing metal artifacts, showing similar artifact magnitude to FBP. Basic MAR reduced shading artifacts about the screw cap, but sinogram interpolation led to smearing of the nearby anatomical structures. Furthermore, blooming artifacts present in basic MAR were similar to those in FBP due to the difficulty in accurately delineating screws in the reconstructed volume. By incorporating screw shape models into the reconstruction process, KC‐Recon provided a significant reduction in blooming artifacts at the screw shaft and polyaxial screw cap, permitting improved visualization of surrounding anatomical structures such as the pedicle, inferior articular process, and the facet joint. Shading artifacts appearing at the screw tip and polyaxial cap as well as between two screws were also mitigated by KC‐Recon. Residual artifacts may be due to unmodeled auxiliary tools introduced during the surgery (e.g., wire gauze, surgical coils, spinal fixation rods), which confounded estimation of the spectral coefficients for the screw shaft and cap.

Figure 7.

Performance comparison of filtered backprojection, penalized weighted least‐squares (PWLS), basic metal artifact reduction, and KC‐Recon in postinstrumentation images from Patient 6 (a‐d) and Patient 12 (e‐h). Spatial resolution was individually matched for reconstructions of the same patient with respect to the tissue edge marked by the yellow lines. The same regularization strength ($log\beta$ = −3) was used for PWLS and KC‐Recon. See Table 1 for screw models and characteristics. The wire gauze evident near the spinous process in (e–h) was not modeled. [Color figure can be viewed at wileyonlinelibrary.com]

Figure 8 illustrates the performance of FBP, basic MAR, and KC‐Recon at full and simulated dose. Postinstrumentation images acquired at D_full = 41.2 mGy were used to generate simulated LD scans at 3.3 mGy. Each image was matched in resolution at ${\sigma }_{\text{ESF}}$ = 0.54 mm using $f_{\text{c}}$ = 0.55 for FBP and Basic MAR and $log\beta$ = −3 and −2.3 for full and reduced dose KC‐Recon. Figures 8(a)–8(c) show similar performance as in Fig. 7 at full dose, with both basic MAR and KC‐Recon providing a marked improvement in visibility of anatomical features about the screws. At low dose (D_sim = 3.3 mGy), soft‐tissue visualization in FBP is severely degraded by both metal artifact and quantum noise. Basic MAR shown in Fig. 8(e) reduces the severity of metal artifacts but is strongly obscured in regions adjacent to the screws at low dose. As in Fig. 5, KC‐Recon shown in Fig. 8(f) reduces metal artifact while maintaining the noise–resolution advantages of PWLS at low dose.

Figure 8.

Comparison of metal artifacts in filtered backprojection, basic metal artifact reduction, and KC‐Recon images at various dose levels. Representative images of two pedicle screws at 41.2 mGy (D_full) and a simulated low‐dose level of 3.3 mGy (D_sim). Spatial resolution was individually matched for all reconstructions with respect to the tissue edge marked by the yellow lines. [Color figure can be viewed at wileyonlinelibrary.com]

Quantitative comparison of performance for FBP and KC‐Recon with respect to a reduction in blooming and shading artifacts is shown in Figs. 9(a) and 9(b), respectively. All 72 screws imaged at full dose were analyzed. Overall, KC‐Recon led to a significant reduction in blooming artifacts (P < 0.0001), providing a median reduction in ${\text{R}}_{\text{bloom}}$ to 1.12 (i.e., 12% overestimation of screw shaft width) compared to 1.34 for FBP (34% overestimation). The method also significantly reduced the shading artifacts at the screw tip (P < 0.0001), giving a median $\mathrm{\Delta }\mu$ of −8.7% for KC‐Recon, compared to −25.4% for FBP. The performance of KC‐Recon at full and low dose (3.3 mGy) is quantified in Figs. 9(c) and 9(d), computed from LD simulations using 13 screws from three patients. There was no significant difference in ${\text{R}}_{\text{bloom}}$ or $\mathrm{\Delta }\mu$ between full and low dose (P > 0.05), suggesting that KC‐Recon remained effective in artifact reduction even in noisy data.

Figure 9.

Artifact magnitude (${\text{R}}_{\text{bloom}}$ and ${\mathrm{\Delta }}_{\mu }$) for filtered backprojection (FBP) and KC‐Recon. Boxplots demonstrating (a) ${\text{R}}_{\text{bloom}}$ and (b) ${\mathrm{\Delta }}_{\mu }$ measured from 72 pedicle screws imaged at full dose and reconstructed using FBP and KC‐Recon. Low‐dose images at 3.3 mGy were simulated from corresponding full‐dose scans and reconstructed with KC‐Recon, showing similar (c) ${\text{R}}_{\text{bloom}}$ and (d) ${\mathrm{\Delta }}_{\mu }$ values as full‐dose images. [Color figure can be viewed at wileyonlinelibrary.com]

4. Discussion and Conclusions

This work presents the first application of KC‐Recon in clinical data for improved intraoperative image quality in spine surgery. The model‐based reconstruction technique was found to be well suited to reconstructing both pre‐ and postinstrumentation images, yielding measurable reduction in noise and improved soft‐tissue visibility. By virtue of an edge‐preserving regularizer, KC‐Recon provided increase in soft‐tissue CNR compared to FBP at equivalent spatial resolution. The CNR improvement was consistent for all levels of spatial resolution investigated. Stronger improvement in CNR was typically demonstrated at lower dose levels. Depending on the imaging task, the method could support imaging protocols with dose reduced by a factor of ~2–10. By integrating screw models into the image reconstruction process, KC‐Recon further improved image quality by permitting a significant reduction in blooming and shading artifacts for a broad range of screw models, dimensions, and material composition. For postinstrumentation images, the method yielded equivalent reduction in metal artifact at full dose and low dose, and preserved a similar degree of improvement in soft‐tissue contrast resolution as in the preinstrumentation images.

As shown in previous studies,³⁵ the STF plays an important role in metal artifact reduction by allowing for unknown material composition and x‐ray beam quality. A simple modification of the function was proposed and implemented in this work to further extend its capabilities to account for beam hardening induced by surgical products composed of two metal materials. While blooming artifacts about the screw shaft were almost eliminated, some streaks originating from the polyaxial screw cap persisted. In several cases, surgical instruments without specified models — that is, wire gauze, spinal fixation rods, and surgical retractors — were not accounted for and thus challenged the spectral coefficient estimation and resulted in residual artifacts. For parameterizable components such as wires and rods, deformable known‐component registration and reconstruction (dKC‐Recon) methods have proven to be useful.^{30, 31, 32, 42} Methods to model auxiliary and complex instruments are under investigation.

It is also noted that ESF width evaluated at a given soft‐tissue edge is not representative of the overall spatial resolution. Future work could invoke certainty‐based methods or space‐variant penalties in image reconstruction, which increases spatial uniformity over the entire image volume.^{52, 53} Furthermore, task‐based image quality metrics can be leveraged for more sophisticated image quality analysis.^{54, 55} When the imaging task is well defined, task‐based metrics can also inform image acquisition and reconstruction to achieve better performance and potential dose reduction.^{56, 57}

An important consideration for incorporation in clinical workflow is the extent to which KC‐Recon can be implemented in a form with reduced reconstruction time. In this study, the runtime for FBP was ~1 min. For KC‐Recon, registration runtime was ~30 s per screw model, for which finely voxelized screw model (0.05 mm isotropic) and six projection views were used. Reconstruction of a (530 × 530 × 400) voxel matrix using equiangular sampled projections (720 views for high definition protocols) required ~20 s for a single subset iteration. An additional ~10 s is needed for estimation of the spectral coefficients per iteration per material (three materials in total for each case). Given that 50 iterations of 20 subsets followed by 50 iterations of ten subsets were employed, the total runtime for pre‐ and postinstrumentation images such as those in Figs. 4 and 7 was therefore ~8 to 9 h.

Further acceleration may be achieved by adopting more computationally efficient forward‐ and backprojection operations such as Siddon's method⁵⁸ and Peters' algorithm,⁵⁹ which have demonstrated 5× reduction in reconstruction time.⁶⁰ Another strategy involves multiresolution binning of projection data and automatic definition of fine‐ and coarse‐grid regions with proper regularization parameter selection in each region,⁶¹ which has been shown to yield strong gains in speed (>10×) even when high‐fidelity forward projectors were used. Depending on the clinical task, fine voxel grids might only be needed in reconstructing regions adjacent to the instrumentation. Furthermore, momentum‐based approaches⁶² can be leveraged to reduce reconstruction time by another ~10× via reducing the number of iterations for achieving convergence.⁶⁰ Such acceleration methods suggest that the runtime for iterative reconstruction can be brought to a level consistent with interventional workflow (e.g., the 5 × 10 × 10 × speedup described above potentially reducing the runtime for images in this work to (~8 h/500) ~1 min, recognizing that further research is required to ensure stability and convergence in the reconstruction result.

Ongoing and future work includes observer‐based assessment of clinical utility in the delineation of soft‐tissue structures, the detection of subtle complications (e.g., hemorrhage), and the use of KC‐Recon for quantitative evaluation and quality assurance of the surgical product.

Conflicts of Interest

This research was supported by NIH grant R01‐EB‐017226.