C-arm cone beam CT perfusion imaging using the SMART-RECON algorithm to improve temporal sampling density and temporal resolution

Abstract

In this work, a newly developed reconstruction algorithm, Synchronized MultiArtifact Reduction with Tomographic RECONstruction (SMART-RECON), was applied to C-arm cone beam CT perfusion (CBCTP) imaging. This algorithm contains a special rank regularizer, designed to reduce limited-view artifacts associated with super-short scan reconstructions. As a result, high temporal sampling and temporal resolution image reconstructions were achieved using an interventional C-arm x-ray system. The algorithm was evaluated in terms of the fidelity of the dynamic contrast update curves and the accuracy of perfusion parameters through numerical simulation studies. Results shows that, not only were the dynamic curves accurately recovered (relative root mean square error ∈ [3%, 5%] compared with [13%, 22%] for FBP), but also the noise in the final perfusion maps was dramatically reduced. Compared with filtered backprojection, SMART-RECON generated CBCTP maps with much improved capability in differentiating lesions with perfusion deficits from the surrounding healthy brain tissues.

1. INTRODUCTION

Stroke is a leading health problem and major cause of death in United States. Ischemic stroke is the most common form of stroke, representing around 80% of the entire stroke population.¹ For intra-arterial stroke therapy that has demonstrated its clinical benefits in a series of recent trials,^2–6 the timing from stroke onset to the initiation of treatment is critically important, as every delay of 30 minutes may be associated with a 10% decrease in the chance of achieving good treatment outcomes.⁷ Many efforts have been taken to shorten door-to-treatment time. One potential strategy is to triage those acute ischemic stroke patients mostly likely to benefit from intra-arterial therapy (e.g., those who are likely to have large vessel occlusions) directly to the interventional suite. By bypassing the MDCT suite or MRI suite, this strategy may potentially save at least one hour of transfer time per patient.

To bypass the current MDCT- or MRI-based stroke imaging paradigms, the imaging capability of the C-arm x-ray imaging system equipped in the interventional suite need to be renovated, such that it can reliably achieve parenchyma imaging to exclude hemorrhage, vessel/angiographic imaging to identify the site of occlusion and the status of collateral blood supply, and perfusion imaging to assess the extent of nonsalvageable and salvageable ischemic tissues. This work focuses on the feasibility of cone beam CT perfusion (CBCTP) imaging using a C-arm interventional x-ray system.

Compared with MDCT perfusion imaging, C-arm CBCTP imaging has its advantages such as whole-brain coverage and isotropic spatial resolution, but it also has some major technical challenges: Currently, it takes about 3–6 seconds for a state-of-the-art C-arm system to perform a single cone beam CT scan over an angular range of about 200 to 260 degrees. Unlike MDCT that can rotates continuously, C-arm system needs a pause time of about 1 second to reverse the direction of gantry rotation before another cone beam CT scan can be performed. Within a total time window of 1 minutes for perfusion imaging, the total number of time frames is about 7 to 9 in CBCTP imaging, compared with 30 to 60 times frames easily achievable by MDCT perfusion. The low temporal resolution and temporal sampling density caused by the slow gantry rotation speed of C-arm system are the two major challenges of CBCTP imaging, and a factor of 3 to 4 improvement in temporal resolution and sampling density is desirable to achieve high quality and reliable CBCTP imaging.

In this work, a newly developed reconstruction algorithm, Synchronized MultiArtifact Reduction with Tomographic RECONstruction (SMART-RECON)⁸ was applied to CBCTP imaging. This algorithm enabled the reconstruction of three to five time-resolved cone-beam CT image volumes from each short CBCT scan. A good balance between limited-view artifacts and temporal resolution was achieved, and the possibility of CBCTP imaging was demonstrated through numerical simulation studies.

2. METHODS

The SMART-RECON algorithm may be described mathematically as follows:

$$\overrightarrow{X^{\ast }}=\underset{\overrightarrow{\mathbf{X}}}{\arg \min }\frac{1}{2}\left[{(\overrightarrow{\mathcal{Y}}-\mathcal{A}\overrightarrow{\mathbf{X}})}^{T}D(\overrightarrow{\mathcal{Y}}-\mathcal{A}\overrightarrow{\mathbf{X}})+\lambda {\Vert {\mathbf{X}}_A\Vert }_{\ast }\right],$$ (1)

where $\overrightarrow{\mathcal{Y}}$ denotes the vectorized projection data, $\overrightarrow{\mathbf{X}}$ is the vectorized CT images to be solved, D is a statistical weighting matrix commonly used in statistical iterative reconstruction, X_A is a prior image-augmented spatial-temporal image matrix, in which a prior image without limited-view angle artifacts and also without temporal information was used to augment X. Minimizing the nuclear norm of this augmented matrix is used to regularize the CT reconstruction. In the case of CBCP imaging, the first column of X_A is given by the vectorized prior image, and each of the other columns corresponds to the vectorized images of each time point. For example, if three super-short scans need to be reconstruct from each of the seven cone beam CT sweeps, the number of columns of X_A is 1 + 7 × 3 = 22. Details of the implementation methods of SMART-RECON can be found in Refs. ⁸ and ⁹.

The feasibility of CBCTP imaging using SMART-RECON was validated through numerical simulation studies. An anthropomorphic digital brain perfusion phantom developed by Manhart et al. was used.¹⁰ Its perfusion properties were varied across voxels to reduce the sparsity of its CT images, so that this phantom does not strongly favor some compressed sensing-based reconstruction algorithms. The dynamic contrast uptake curve for each voxel was given by convolving a known arterial input function (AIF) with a residue function, which is calculated from the known perfusion parameters based on the indicator-dilution theory. A relatively small circular ischemic lesion with severely reduced cerebral blood volume (CBV) was simulated in the right hemisphere of the brain, which was surrounded by a much larger ischemic lesion with reduced cerebral blood flow (CBF).

The simulated CBCTP acquisitions of this numerical phantom contained nine sweeps, each covering 260 degrees over 5.2 seconds. Two of the nine sweeps were acquired prior to the wash-in of contrast bolus and were used as mask (or baseline) images, while the other seven sweeps were acquired with contrast update. The pause time between two consecutive acquisitions was set to 1.2 seconds, which was similar to the pause time of typical C-arm interventional systems. Noise was added to the simulated CBCTP data in the sinogram domain; the magnitude of the noise was adjusted so that the noise standard deviation of the CBCT images reconstructed by filtered backprojection (FBP) was 4 HU.

Data acquired from each sweep were reconstructed using both SMART-RECON and FBP. For SMART-RECON, five time-resolved image frames were reconstructed from each 260 degree short scan. As a result, each reconstructed image only used data from 260°/5=52° angular range without angular overlap. Such a data partition scheme corresponds to a temporal window of about one second for each angular sub-sector. After reconstruction, three regions of interest (ROIs) were placed in the CBCTP source images to quantify reconstruction accuracy: one ROI was placed in an major feeding artery, the second one was placed in the white matter, and the third one was placed in the issue with reduced CBF. These ROIs were used to quantify the accuracy of the measured contrast update curves and noise level of CBCTP source images. The second and third ROIs were also placed in the final perfusion maps to evaluate the quantitative accuracy of parametric perfusion measurements.

3. RESULTS

Figures 1 and 2 show the arterial input function (AIF) and tissue attenuation curve given by the ground truth, FBP reconstruction, and SMART-RECON. The AIF and TAC generated by SMART-RECON were much consistent with the ground truth. The relative root mean square error (rRMSE) of the temporal curve between SMART-RECON and ground truth was no greater than 5%, as opposed to no less that 14% generated by FBP (Table 3).

Figure 1.

Comparison of the arterial input function given by FBP and SMART-RECON with the ground truth.

Figure 2.

Comparison of the tissue attenuation curve given by FBP and SMART-RECON with the ground truth.

Figure 3 shows representative source images reconstructed by FBP and SMART-RECON. Compared with FBP, SMART-RECON effectively reduced image noise thus improved the visibility of the circular ischemic lesion with reduced CBV (close-ups in Fig. 3). The capability of source image noise reduction by SMART-RECON was also evidenced by the measured noise standard deviation values shown in Fig. 4 and Table 3: SMART-RECON reduced the noise magnitude by approximately an order of magnitude.

Figure 3.

Comparison of the source images generated by FBP and SMART-RECON.

Figure 4.

Comparison of noise standard deviation (STD) of source images generated by FBP and SMART-RECON.

Figures 5–7 compare cerebral blood flow (CBF), cerebral blood volume (CBV), and mean transit time (MTT) maps generated by FBP and SMART-RECON with the ground truth. Compared with perfusion maps generated by FBP, those produced from SMART-RECON demonstrated significantly reduced noise and improved visibility of the perfusion deficits in the digital phantom. In particular, the mismatch in the volume of ischemic lesions between CBV and CBF (or between CBV and MTT) was easily visualizable in the SMART-RECON images but barely noticeable in the FBP images.

Figure 5.

Comparisons of cerebral blood flow (CBF) maps generated by FBP and SMART-RECON with the ground truth.

Figure 7.

Comparisons of mean transit time (MTT) maps generated by FBP and SMART-RECON with the ground truth.

4. CONCLUSIONS

In conclusion, a newly developed reconstruction algorithm–Synchronized MultiArtifact Reduction with Tomographic RECONstruction (SMART-RECON)–was applied to cone beam CT perfusion imaging to address the technical challenges associated with the hardware limitations of the C-arm interventional system, namely low temporal resolution and temporal sampling density. With SMART-RECON, the temporal sampling rate and temporal resolution can be improved from 4–5 seconds to 1–2 seconds, and parametric perfusion maps with improved image quality and quantitative accuracy can be generated. Further clinical evaluations of SMART-RECON-based CBCTP imaging will be performed in future work to explore the robustness and clinical value of this method.

Figure 6.

Comparisons of cerebral blood volume (CBV) maps generated by FBP and SMART-RECON with the ground truth.

Table 1.

Relative root mean square error (rRMSE) and noise standard deviation values measured in the three ROIs in the digital phantom.

ROI region	rRMSE		Noise level (HU)
	SMART	FBP	SMART	FBP
ROI 1 (normal tissue)	0.05	0.22	0.5 ± 0.04	4.0 ± 0.3
ROI 2 (ischemic tissue)	0.04	0.14	0.4 ± 0.04	4.0±0.5

Table 2.

Mean and ±standard deviation values of perfusion parameters measured in the normal brain tissue (ROI 2) and ischemic tissue with reduced blood flow (ROI 3).

	CBF(ml/100g/min)			CBV(ml/100g)			MTT(second)
	Truth	SMART	FBP	Truth	SMART	FBP	Truth	SMART	FBP
ROI 1	15 ± 1.8	18 ± 4.0	16 ± 9.7	1.4 ± 0.2	1.5 ± 0.3	1.2 ± 1	5.5 ± 0.3	5.3 ± 0.9	NA
ROI 2	8 ± 4.3	8 ± 1.8	10 ± 3.6	1.0 ± 0.2	1.2 ± 0.3	0.7 ± 1	12.6 ± 2	8.8 ± 2.1	NA