Abstract
For 3D radial data reconstruction in magnetic resonance imaging (MRI), fast Fourier transform via gridding (gFFT) is widely used for its fast processing and flexibility. In comparison, conventional 3D filtered back projection (cFBP), while more robust against common radial kspace centering errors, suffers from long computation times and is less frequently used. In this study, we revisit another backprojection reconstruction strategy, namely twostep 2D filtered backprojection (tsFBP), as an alternative 3D radial MRI reconstruction method that combines computational efficiency and certain error tolerance. In order to compare the three methods (gFFT, cFBP, and tsFBP), theoretical analysis was performed to evaluate the number of computational steps involved in each method. Actual reconstruction times were also measured and compared using 3D radialMRI data of a phantom and a human brain. Additionally, the sensitivity of tsFBP to artifacts caused by radial kspace centering errors was compared with the other methods. Compared to cFBP, tsFBP dramatically improved the reconstruction speed while retaining the benefit of tolerance to the radial kspace errors. Our study therefore suggests that tsFBP can be a promising alternative to the conventional back projection method for 3D radial MRI reconstruction.
Similar content being viewed by others
Introduction
Radial acquisition (RA), also known as projection acquisition, embodied the first kspace trajectory used in MRI, as proposed by Lauterbur^{1}. Here the kspace is filled with a set of radial spokes that extend out from the origin. The RA scheme has recently gained popularity in the MR community due to the possibility of a shorter minimum echo time (TE) and better tolerance to motion and flow artifacts^{2,3,4} as compared to conventional Cartesian kspace acquisition. It is, however, sensitive to system imperfections such as timing delays between the actual and requested starting points of gradient waveforms^{5,6} because the readout direction in the kspace varies between repetitions. Delays and other forms of imperfections in the gradient system for radial sampling can cause the spokes to miss the center of the kspace, possibly resulting in severe artifacts in the final image. To correct the gradientdelay effects in RA imaging, several methods have been proposed to estimate and compensate for them^{5,6,7}. The RA scheme also computationally burdens the image reconstruction system more than the Cartesian kspace acquisition scheme because the data must lie on a Cartesian grid in order to take advantage of the high computational speed afforded by the fast Fourier Transform (FFT) algorithm.
In terms of image reconstruction approaches, radial data can be reconstructed by two categories of reconstruction methods: 1) FFT via gridding (gFFT) and 2) filtered backprojection (FBP) based on the Radon transform^{8}. The latter exploits the Fourier slice theorem^{9}, which states that the Fourier transform of spokeencoded data is the projection of an object perpendicular to the spoke direction. A major difference between gFFT and FBP is that the former performs image reconstruction in the kspace while the latter does so in the image space. Although both methods have been successfully used, gFFT is often preferred to the FBP approaches for 3D radialdata reconstruction due to its much shorter processing time and flexibility to handle irregular kspace patterns. For conventional 3D FBP (cFBP), a set of planar projections are obtained using the 3D projection slice theorem and then backprojected onto the 3D image space for reconstruction. The most timeconsuming part of cFBP is the 3D volumetric backprojection of every radial spoke, which makes the computational time grow rapidly with the image resolution. Lauterbur and Lai proposed a much faster alternative to this straightforward but timeconsuming cFBP method in 1980^{10,11}. This approach, which will hereafter be called “twostep 2D FBP” (tsFBP), has been used in Xray imaging but has received little attention from the MR community. Considering the growing role of 3D radial scans in MRI, we believe that it is timely to systematically compare the computational speed and error tolerance of the different radial reconstruction methods.
A main goal of this study is to evaluate the computational performance of three different 3D radialscan reconstruction methods: gFFT, cFBP, and tsFBP. Mathematical expressions for the computational requirements of each reconstruction method were formulated by calculating the number of major operations in the reconstruction process. Experiments were also performed to demonstrate the robustness of the FBP approaches against simulated radial offcentering errors of the kspace trajectory, that can be caused by gradient delays, eddy currents, or finite digitization bandwidth, in comparison to gFFT. Recently it has been shown that gFFT can be further accelerated by utilizing algorithms tailored to specific radial trajectories^{12}. In the present work, we considered only the most widely used, conventional gridding algorithm in gFFT for simplicity.
Results
Simulations
Figure 1 shows the ratios between the operational counts (Eqs. 4–6) of tsFBP and the other two methods for varying numbers of receiver channels and image matrix sizes. tsFBP showed a dramatic decrease in computational requirement in comparison to cFBP for all cases (Fig. 1a). Under the channel combination scenario considered (see “Methods”), tsFBP also showed computational speed advantage over gFFT in cases where the matrix size (N) was small, or the number of channels (N_{Ch}) was large (Fig. 1b–d). For example, when the convolutionkernel width for gridding (ω) was two, which was the case for our gFFT reconstruction of the real data, tsFBP outperformed gFFT when N_{Ch} ≥ 16. On the other hand, when N_{Ch} = 1, which corresponds to another likely scenario of channel combination (see “Discussion”), tsFBP took more computational steps than gFFT (by up to × 9) except with the largest ω. Under this scenario, tsFBP still had a large advantage over cFBP, by a factor of > 30.
Table 1 summarizes the theoretically estimated computational burdens of the three reconstruction methods and the actual computation times measured during the real data reconstruction using the same acquisition and reconstruction parameters. While the theoretical estimations of cFBP and gFFT were ~ 220 and ~3 times higher than those of tsFBP, respectively, their actual computational times were ~ 253 and ~ 8 times higher for the acquisition parameters used in the 3 T experiment. With the 9.4 T acquisition setup, the actual reconstruction times of cFBP and gFFT were longer than tsFBP by factors of ~ 143 and ~ 0.8, respectively, while the factors were ~150 and ~ 0.5 in theoretical calculations. The agreement between the actual computational times and the theoretical expectations appears reasonable given the approximate nature of the theoretical estimation.
Experiments
Figure 2 shows the representative axial images of the American College of Radiology (ACR) phantom and the human brain reconstructed by gFFT, tsFBP and cFBP. Image qualities obtained from tsFBP and gFFT were similar when the offcentering of the kspace trajectory was small (Fig. 2a–d). However, in the presence of large kspace offcentering due to artificial echo shifts, gFFT showed contrastvarying artifacts from center to periphery (Fig. 2e–f). This is because gFFT adds an echoshiftdependent phase modulation to the final image, resulting in undesired contrast variations, whereas tsFBP and cFBP are insensitive to phaserelated artifacts thanks to magnitudebased projections. For detailed information of the echopeak distributions of both asacquired and echopeakshifted data, refer to Supplementary Fig. S1.
Figure 3 shows another potential advantage of FBP over gFFT in terms of defining the “true k = 0” in discrete sampling. As shown in Fig. 3a a ringshaped contrast variation is seen at the periphery of the phantom (white arrowheads) when a mismatch between the true k = 0 and the apparent k = 0 exists in gFFT. Figure 3d illustrates an example of two apparent k = 0 positions (k_{0,a}), neither of which exactly corresponds to the true k = 0 (k_{0}). Due to the finite sampling bandwidth, k_{0,a} could be located to the left (k_{0,a} < k_{0}; bluedashed line), or to the right (k_{0,a} > k_{0}; reddashed line) of k_{0}, as shown in the detailed views of Fig. 3e,f, respectively. In our data, the edge artifact for gFFT nearly disappeared (white arrows) when k_{0,a} was shifted by a half of the sampling interval, which implies that such a slight shift could better match k_{0,a} to the true k = 0. In contrast, the image reconstructed by tsFBP showed little boundary artifact without any shift (Fig. 3c) since tsFBP exploits the magnitudebased projections for reconstruction and is thus insensitive to the phaserelated errors as mentioned above. More examples including the human brain as well as the phantom images, which were reconstructed with gFFT while varying the apparent k = 0 position during the reconstruction, are presented in Supplementary Figs. S2–S4.
When the image to be reconstructed has actual phase variation at the time of imaging, magnitudebased projection can suffer from artifacts caused by the loss of phase interference information. This can be alleviated by employing complex projections, namely by retaining the phase in the back projections of tsFBP. This point is illustrated in Fig. 4, where the image phase was varied by varying the echo time in the presence of static field (B_{0}) inhomogeneity. Here an ACR phantom was imaged at 3 T with a twochannel head volume coil at five different echo times (TE = 0.18 to 10 ms) in the presence of susceptibilityinduced B_{0} inhomogeneity as well as channeldependent RF (B_{1}, both transmit and receive) phase. Images were reconstructed from the individual channels and then combined as the sum of the squares. The magnitudeonly projection can be seen to suffer severe loss of image quality at TE > 5 ms, compared to the complex projection. The locations of image blurring and signal dropout in Fig. 4(d,e) corresponded to the locations of strong B_{0} inhomogeneity on the perimeter and near an air bubble, indicated by the arrows in Fig. 4(km). Since B_{0}induced phase is proportional to TE while B_{1} phase is TEindependent, image phase at low TE is dominated by the latter. In our example, this phase varied relatively slowly in space (Fig. 4n,o) and did not appear to degrade the image quality of magnitudeonly projections very much at low TE (< 5 ms).
Discussion
Computation time and channel combination
One of the reasons why gFFT is widely used for 3D radialdata reconstruction is that it takes much shorter processing time than cFBP. In this study, we examined another type of FBP method, tsFBP, and demonstrated that its computational requirement could be dramatically lower than cFBP, and even be comparable to gFFT in some practical situations (Fig. 1 and Table 1).
Among the factors that influence the computational efficiency, Fig. 1 highlighted three, namely the number of receiver channels (N_{Ch}), the final matrix size (N), and the convolutionkernel width (ω). tsFBP showed much better overall computational efficiency than cFBP, which was more pronounced as N_{Ch} decreased or N increased. In contrast, when compared to gFFT, the computational efficiency of tsFBP tended to increase as N_{Ch} increased, N decreased, or ω increased (ω is only relevant to gFFT).
We note that in Eqs. (4–6) and Fig. 1, the signal combination for different receiver channels was assumed to be done before the backprojection for the FBP methods, while in gFFT the channel combination was done in the last step. It is possible for the FBP methods also to first create channelspecific images and then have them combined later. Such approach has several benefits. First, it would allow coil sensitivity profiles to be used for better signaltonoise ratio (SNR) in the final image (Supplementary Fig. S5, demonstrating about 17% SNR difference). Second, coilbycoil reconstruction better maintains kspace center consistency among different radial spokes, which can result in reduced intensity artifacts compared to projectionlevel coil combination (e.g., Fig. S5a compared to Fig. S5b). Finally, it could enable parallel imaging such as radial sensitivity encoding (SENSE) with iterative reconstruction^{13}, provided complex projection is utilized. In such a (channelwise reconstruction) case, the reconstruction time estimates Eqs. (4, 5) should be modified to have one factor of \({N}_{Ch}\) multiplying all the individual steps. \({N}_{Ch}\) will then be common to all three methods so the ratios between computational burden estimates will reduce to those with \({N}_{Ch}\) = 1 in Fig. 1. In this case, tsFBP is still more than 30 times faster than cFBP in our example (Fig. 1a), but its speed comparison with gFFT becomes less favorable (at most faster by factor of 2 and slower in most cases, Fig. 1d, \({N}_{Ch}\) = 1).
We also note that one can consider performing channel combination before gridding in gFFT. This would be possible in case of magnitudebased reconstruction, in which multiple receivers’ projection data are first combined by 1D FFT magnitude summation, and then interpolated to the Cartesian grid for 3D FFT. This will decrease the computation time for gFFT by approximately N_{Ch}. The computational speed comparison will then again reduce to the case of N_{Ch} = 1 in Eqs. (4, 5) and Fig. 1.
Advanced computing hardware such as multiple CPU cores can substantially improve the computational speed independent of the reconstruction method. The fact that the proposed method, tsFBP, can also benefit from multicore computation is illustrated in Table 2, where we listed measured image reconstruction times for the three methods obtained with 1, 3, 6 cores. We observed that the hardwareenabled acceleration was comparable between tsFBP and gFFT (an order of magnitude), while the gain was somewhat weaker for the slowest method, cFBP.
Magnitude versus complex projections
tsFBP with magnitudeonly projection has certain advantages over gFFT in terms of the tolerance to offcentering errors of kspace trajectories that might result from gradient timing delays, eddy currents, or finite digitization bandwidth. In gFFT that is phase sensitive, offcentered kspace data creates unwanted phase modulation in the image space, causing contrastvarying artifacts. Several approaches have been proposed to mitigate the kspace offcentering problem for gFFT reconstruction. For example, the gradient delay could be compensated by realigning the radial data in postprocessing with another calibration scan^{14} or by monitoring the actual gradient strength during the scan to estimate the actual kspace trajectory^{15}. These and most other corrections require additional scans or hardware, and do not guarantee complete removal of offcentering artifacts when the true echo peak is placed between two adjacent samples, which is inherently possible in discrete sampling with a finite sampling bandwidth (Fig. 3). While simple 1D interpolation of the radial data can effectively increase the sampling frequency and reduce the digitization error in gFFT, this will increase the grid size and therefore the overall processing time. In comparison, magnitudebased FBP is inherently more robust against trajectory errors as long as they are purely radial. Obviously, however, loss of phase information in magnitudeonly method precludes applications such as phasebased susceptibility imaging^{16,17} and phasecontrast flow imaging^{18,19,20}.
As was shown in Fig. 4, when significant phase variation exists in the image, such as due to B_{0} inhomogeneity^{21} or RF phase, magnitudeonly FBP will fail to reconstruct the true image magnitude because of the unrecoverable, destructive interference in 1D projections. This drawback will be more severe for higher magnetic fields and longer echo times, larger field of view (FOV), and more localized RF coils. Such problems can be mitigated by employing complex projections (Fig. 4).
kspace trajectory
Another limitation of the proposed tsFBP method comes from the requirement of a specific radial spoke placement which is not spherically symmetric. An immediate consequence of such asymmetry is that the point spread function is anisotropic as illustrated in Supplementary Fig. S6. Furthermore, in order to satisfy the Nyquist criterion to avoid undersampling artifacts, the spokes on the equator should satisfy (we use definition of k without 2\(\uppi\))
where the maximum kspace extent is given by \({k}_{max}=\left(1/2\right) N/FOV\) assuming an isotropic field of view and grid size (\({N}_{x}={N}_{y}={N}_{z}=N)\). In tsFBP, the number of halfecho radial spokes satisfying the above criterion is
This compares with the required number of spokes in uniform spherical sampling,
Therefore, \({N}_{v,tsFBP}\) is \(\uppi /2\approx 1.57\) times larger than \({N}_{v,unif}\), which means that tsFBP takes 1.57 times longer acquisition than gFFT based on uniform radial sampling. When the number of spokes is short of the Nyquist condition, undersampling artifacts appear in tsFBP as illustrated for specific cases in Supplementary Fig. S7. The artifact pattern depends on the angular direction in which undersampling occurs. While the additional scan time by 1.57 for full Nyquist sampling is significant, in our experience tsFBP with reduced spokes, namely \({{N}_{v,unif}\le N}_{v}<{N}_{v,tsFBP}\), incurred relatively minor SNR penalty (2–3% in case of \({{N}_{v}= N}_{v, unif}\) , Supplementary Fig. S8). Theoretically, this is understood by the following formula to account for different trajectories in SNR calculation^{22}:
With D_{1}(k) and D_{2}(k) replaced by the mean kspace density functions for tsFBP and uniform trajectories, we obtain the SNR ratio of 0.977, in close agreement with experimental observation.
Conclusion
In conclusion, the tsFBP method can bring high computational efficiency to classical back projection reconstruction of 3D radial data. With magnitudeonly projection, it is relatively insensitive to radial kspace offcentering errors. For phasesensitive imaging or for iterative reconstruction, tsFBP can still be used with complex projection. We expect that tsFBP could be a useful addition to the existing image reconstruction toolsets for 3D radial scan MRI which is receiving more and more attention for short echo times as well as relative motion insensitivity.
Methods
Comparison of conventional 3D and 2D backprojection methods
For conventional filtered backprojection (cFBP), a set of planar projections are obtained using the 3D projection slice theorem and then backprojected onto the 3D image space for reconstruction. In this reconstruction process, every radial spoke data \(s\left(k,\theta ,\phi \right)\) is backprojected with a filter of \({k}^{2}\mathrm{sin}\theta\) as in the following equation^{23}:
On the other hand, the twostep filtered backprojection (tsFBP) performs image reconstruction using 2D filtered backprojection in two steps in which a ramp filter \(\leftk\right\) is applied to each radial data \(s\left(k,\theta \right)\), as in the following equation^{23}:
Equations (1–2) are generally valid when the image to be reconstructed has spatially dependent phase across it caused by static field inhomogeneity or radio frequency (RF) excitation or reception phase. When such phase variation can be neglected, for example thanks to a very short echo time or relatively uniform RF fields, then the kspace function \(s\) is Hermitian symmetric, and the integral along a full radial line in kspace will be real (up to a constant phase). In such a case, filtered backprojection can be done with magnitude projections only. For example, in 2D, Eq. (2) can be replaced by:
Utilizing ultrashort echotime (UTE) imaging and single or fewchannel volumetric RF coils, we experimentally observed that magnitude projections often yielded acceptable image quality.
Twostep 2D filtered backprojection (tsFBP) method
The schematic diagram of the tsFBP method, including its unique dataacquisition scheme, is illustrated in Fig. 5. In tsFBP, 3D radial data are collected in a series of 2D radial spoke sets, each of which fills the kspace in the form of a vertical disc by incrementing the polar angle θ from 0 ≤ θ < π (full echo sampling) or 0 ≤ θ < 2π (free induction decay or partial echo sampling) for every azimuthal angle φ in the range of 0 ≤ φ < π (Fig. 5a). A specific reconstruction algorithm (outlined below) for this particular scheme of data collection is then applied to accomplish 3D image reconstruction. Note that while we used an ultrashort echotime (UTE) sequence for radial data collection, in principle nonUTE radial acquisition is also compatible with tsFBP. In the presence of static field inhomogeneity, however, nonUTE acquisition would require complex projections to properly handle the image phase.
The 3D image reconstruction is achieved by the following steps. First, a sinogram consisting of a series of 1D projections is obtained from 1D FFT of all of the radial spokes on a vertical disc at a given φ in the kspace (Fig. 5b). Subsequently, 2D projection images are reconstructed from each sinogram using 2D FBP for every φ (1^{st}step 2D FBP; Fig. 5c). Finally, the zcoordinatematched strips taken from the stack of φdependent 2D projection images are 2D backprojected to reconstruct a final 3D image slice by slice (2^{nd}step 2D FBP; Fig. 5d,e). For each step of tsFBP, a rampfilter is applied prior to the backprojection.
Mathematical expressions for computational requirement of each reconstruction method
In order to evaluate the computational requirement of each reconstruction method independent of particular programming environment, we analyzed the number of major operations required for the reconstruction process.
Firstly, tsFBP can be divided into three main processes of 1D FFT, 1^{st}step 2D FBP, and 2^{nd}step 2D FBP. In this case, the number of operations for tsFBP is formulated as follows:
where N_{v} (= N_{v,θ}N_{v,φ}) is the number of the acquired radial spokes; N_{s} is the number of sampling points per spoke; N_{Ch} denotes the number of receiver channels; N_{x}, N_{y}, N_{z} are the final matrix sizes in the x, y, z axes; N_{v,θ} is the number of θ increments in a vertical disc of Fig. 5a; and N_{v,φ} is the number of φ increments. In the first term of Eq. (4), N_{s}log_{2}(N_{s}) represents the number of operations required for 1D FFT of each spoke. Multiplying that by the number of spokes (N_{v}) and the receiver channels (N_{ch}) yields the total number of operations needed for 1D FFT of all radial spokes. In the second term, which corresponds to the 1^{st}step 2D FBP process (Fig. 5c), N_{v,θ}N_{x}N_{z} represents the number of operations required for 2D FBP of 2D radial spokes at a given φ. Here, each of the line projections (N_{v,θ}) is backprojected to a matrix of size N_{x}N_{z}. Its multiplication by the number of φ increments (N_{v,φ}) then gives the total number of operations during the 1^{st}step 2D FBP process. The final, 2^{nd}step FBP term, which is similar to the 1^{st}step FBP, is obtained by multiplying the operational counts (N_{v,φ}N_{x}N_{y}) for the 2D FBP of a zcoordinatematched slice by the number of these slices (N_{z}).
cFBP mainly consists of 1D FFT and 3D FBP, and the number of total operations is given by:
Here, the first term for the 1D FFT process is just the same as that of Eq. (4). In the second term, N_{x}N_{y}N_{z} represents the number of operations for the 3D backprojection performed at an angle of each spoke, and multiplication by the number of spokes (N_{v}) gives the total number of 3D FBP operations.
For gFFT, the main steps consist of gridding, 3D FFT, and intensity correction. The number of operations in total is:
where ω is the convolutionkernel width for gridding and V is the kspace oversampling factor. The first term corresponds to the 3D gridding operation for all the sampling points of the radial kspace data, whose number is given by the product of the acquired radial spokes and the sampling points per spoke (N_{v}N_{s}). The gridding algorithm distributes each sampling point to the Cartesian grid by convolving with a gridding kernel such as the Kaiser–Bessel window. The second term relates to the processing time of 3D FFT with an oversampling factor^{24}. The final, intensity correction term captures the division process by the inverse Fourier transform of the convolution kernel at each voxel of the final image, which is intended to compensate for the interpolation effect due to the convolutionkernel application. That is multiplied by the number of channels (N_{Ch}) since reconstruction is performed separately for each channel. Table 1 summarizes the mathematical expressions for the number of main operations at individual reconstruction stages of the three reconstruction methods.
It is worth mentioning at this point that gFFT can also be done with twostep 2D gridding, rather than 3D gridding, provided that the kspace trajectory is compatible with tsFBP as explained above. While such twostep gFFT was not part of our experimental tests, for completeness we list below the mathematical expression for its operation count.
Compared to Eq. (6), the above expression shows that the gridding and the intensity correction operations now each consist of two steps. For gridding (the first two terms), the interpolation kernel size is reduced from \({\omega }^{3}\) to \({\omega }^{2}\), reflecting 2D gridding. Each of the two gridding operations has roughly the same magnitude as the 3D gridding operation (\({N}_{v}{N}_{s}{\omega }^{3}\) in Eq. 6). Therefore, Eq. (7) is comparable to Eq. (6) when \(\omega =2\), but twostep gFFT will be more favorable for a large interpolation kernel size.
Simulations
To compare the computational requirements, the number of major operations was calculated for each reconstruction method using Eqs. (4–6) as a function of an isotropic matrix size N = N_{x} = N_{y} = N_{z} with a varying number of receiver channels N_{Ch}. Parameter setups for the simulation were: N_{s} = N/2, N_{v} = 2 (π N_{s})^{2} to satisfy the Nyquist criterion, V = 2, and ω = 2–4^{25,26}. Values of the oversampling factor (V) and the convolutionkernel width (ω) were chosen to be typical for conventional gFFT reconstructions. Four simulations were carried out for comparison between tsFBP and cFBP as well as between tsFBP and gFFT, varying ω from two to four (Fig. 1).
Supplementary Fig. S7 shows images reconstructed by tsFBP while varying the number of polar (θ) and azimuthal (φ) angles. For this, we performed simulations using a 3D SheppLogan phantom. 3D radial signals corresponding to a set of polar (θ) and azimuthal angles (φ) were generated. The phantom grid size and the final reconstruction matrix size were both 128 × 128 × 128.
The basic accuracy of tsFBP reconstruction was verified against conventional methods through simulations on the same numerical phantom. Supplementary Figure S9 shows images reconstructed by gFFT, cFBP and tsFBP with different filter and interpolation functions used in the backprojection methods (cFBP, tsFBP). The following simulation parameters were used: (tsFBP)\({N}_{v,\theta }\) = 201, \({N}_{v,\phi }\) = 201; (cFBP) N_{v} = 40,401; and (gFFT) N_{v} = 40,401, V = 2 and ω = 4. The difference images with respect to the ground truth showed that while gFFT was the most accurate, tsFBP performed better than or similarly to cFBP for the simulation parameters used.
Experiments
In order to evaluate the computation times of each reconstruction method on actual scan data and also their robustness against offcentering of the kspace trajectory, 3D radialscan experiments were performed on phantoms and the human brain on preclinical 9.4 T (BioSpec 94/30, Bruker, Ettlingen, Germany) and clinical 3 T (Magnetom Trio, Siemens, Erlangen, Germany) scanners. For the human brain imaging at 3 T, informed written consent was obtained from a healthy volunteer according to a protocol approved by the Sungkyunkwan University ethics committee. All data sets were obtained using the CODE (Concurrent Dephasing and Excitation) pulse sequence^{27}, which is a 3D radial gradientechobased UTE sequence. The sequence generates a partial gradient echo for each radial spoke, sufficient to construct a 1D projection in the corresponding direction. We note that magnitude projections based on partial echoes can suffer from blurring and background brightening due to lost kspace data on the shorttail side. While this can be mitigated by homodyne filtering or more symmetric echoes, the CODE data were sufficient for our main purpose of computation time comparison and echo shift simulation. All the 1D projection data for radial reconstruction were obtained such that the end points of the corresponding radial spokes spanned the whole (not half) kspace sphere.
First, the standard ACR phantom and the human brain scans were performed at 3 T using a fourchannel head coil. Scan parameters were: N_{v} = 64,800, N_{v,φ} = 180, N_{v,θ} = 360, TR = 3 ms, TE = 0.22 ms, flip angle = 5°, FOV = 300 × 300 × 300 mm^{3}, and spatial resolution = 1 × 1 × 1 mm^{3}. Second, another phantom imaging was also performed at 9.4 T using a singlechannel volume coil. Scan parameters were: N_{v} = 64,800, N_{v,φ} = 180, N_{v,θ} = 360, TR = 3 ms, TE = 0.15 ms, flip angle = 5°, FOV = 60 × 60 × 60 mm^{3} , and spatial resolution = 0.47 × 0.47 × 0.47 mm^{3}.
Data obtained at 3 T and 9.4 T were also used to show the robustness of tsFBP to the offcentering errors of the kspace trajectories in comparison to gFFT. For better demonstration of the kspace offcentering effects and tolerance to them, some echo peaks of the 3 T scan data were manually shifted during the reconstruction process. In other words, while the echo peaks were located between the 3rd and 6th points in the acquired data (“asacquired data”), some of the echo peaks were randomly shifted by up to 15 points from their original locations to create a larger deviation of the echopeak positions (for detailed information of the echopeak distributions, refer to Fig. S1 in Supplementary Information).
In addition, another source of misregistration of the center of kspace (digitization error) was investigated using the phantom data at 9.4 T. Due to the finite digitization bandwidth, sometimes the echopeak position of a discretely sampled echo signal does not exactly coincide with the true k = 0 even after correction of the shifts induced by gradient delays and eddy currents. That is, the true k = 0 is placed between two sampling points. It was shown that, in this case, the image quality can be further improved by a manual echo shift smaller than the sampling interval for gFFT (Fig. 3).
Image reconstruction
The three reconstruction methods of gFFT, cFBP, and tsFBP were applied to image reconstruction of 3D radially sampled CODE data, and their computation times were recorded during the reconstruction processes. All calculations were carried out offline using Matlab (R2011a, MathWorks, Natick, MA, USA) on a workstation equipped with an IntelXeon CPU (3.50 GHz). We note that other available packages such as Berkeley Advanced Reconstruction Toolkit (BART)^{28} could be used for reconstruction and could take different computation times. Given the generally good agreement between the mathematical operation counts and our Matlab results, however, we do not believe that using other packages will significantly change our conclusion. For the gFFT reconstruction, the mean value of all echopeak positions was used to define k = 0. The convolutionkernel width (ω) was set as two.
Code and data availability
The datasets generated and analyzed during the current study are available from the corresponding author upon reasonable request. An example to compare the image reconstruction methods proposed in this study is available online. To use example related data and code, please use the link below (https://drive.google.com/drive/folders/12U1ERpsoM2G7Ku5Nnc9drpllPMbixBs?usp=sharing).
References
Lauterbur, P. C. Image formation by induced local interactions: examples employing nuclear magnetic resonance. (1973).
Irarrazabal, P. & Nishimura, D. G. Fast three dimensional magnetic resonance imaging. Magn. Reson. Med. 33, 656–662 (1995).
Bergin, C., Pauly, J. & Macovski, A. Lung parenchyma: projection reconstruction MR imaging. Radiology 179, 777–781 (1991).
Glover, G. H., Pauly, J. M. & Bradshaw, K. M. Boron11 imaging with a threedimensional reconstruction method. J. Magn. Reson. Imaging 2, 47–52 (1992).
Brodsky, E. K., Samsonov, A. A. & Block, W. F. Characterizing and correcting gradient errors in noncartesian imaging: are gradient errors linear timeinvariant (LTI)?. Magn. Reson. Med. 62, 1466–1476 (2009).
Peters, D. C., Derbyshire, J. A. & McVeigh, E. R. Centering the projection reconstruction trajectory: reducing gradient delay errors. Magn. Reson. Med. 50, 1–6 (2003).
Block, K. & Uecker, M. in Proceedings of the 19th Annual Meeting of ISMRM. 2816–2816.
Deans, S. R. The Radon transform and some of its applications. (Courier Corporation, 2007).
Mersereau, R. M. & Oppenheim, A. V. Digital reconstruction of multidimensional signals from their projections. Proc. IEEE 62, 1319–1338 (1974).
Lauterbur, P. G. & Lai, C.M. Zeugmatography by reconstruction from projections. IEEE Trans. Nucl. Sci. 27, 1227–1231 (1980).
Kramer, D., Schneider, J., Rudin, A. & Lauterbur, P. True threedimensional nuclear magnetic resonance zeugmatographic images of a human brain. Neuroradiology 21, 239–244 (1981).
Smith, D. S., Sengupta, S., Smith, S. A. & Brian Welch, E. Trajectory optimized NUFFT: faster nonCartesian MRI reconstruction through prior knowledge and parallel architectures. Magn. Reson. Med. 81, 2064–2071. https://doi.org/10.1002/mrm.27497 (2019).
Qu, P. et al. Iterative backprojection reconstruction for radial SENSE. in Proc. Intl. Soc. Mag. Reson. Med. (2007).
Duyn, J. H., Yang, Y., Frank, J. A. & van der Veen, J. W. Simple correction method for kspace trajectory deviations in MRI. J. Magn. Reson. 132, 150–153 (1998).
Barmet, C., Zanche, N. D. & Pruessmann, K. P. Spatiotemporal magnetic field monitoring for MR. Magn. Reson. Med. 60, 187–197 (2008).
de Rochefort, L. et al. Quantitative susceptibility map reconstruction from MR phase data using bayesian regularization: validation and application to brain imaging. Magn. Reson. Med. 63, 194–206 (2010).
Kim, J. H., Kim, J. H., Lee, S. H., Park, J. & Lee, S. K. Fabrication of a spherical inclusion phantom for validation of magnetic resonancebased magnetic susceptibility imaging. PLoS ONE https://doi.org/10.1371/journal.pone.0220639 (2019).
Markl, M., Frydrychowicz, A., Kozerke, S., Hope, M. & Wieben, O. 4D flow MRI. J. Magn. Reson. Imaging 36, 1015–1036 (2012).
Dyverfeldt, P. et al. 4D flow cardiovascular magnetic resonance consensus statement. J. Cardiovasc. Magn. Reson. 17, 72 (2015).
Markl, M. et al. Advanced flow MRI: emerging techniques and applications. Clin. Radiol. 71, 779–795 (2016).
Lee, S. K., Hwang, S. H., Barg, J. S. & Yeo, S. J. Rapid, theoretically artifactfree calculation of static magnetic field induced by voxelated susceptibility distribution in an arbitrary volume of interest. Magn. Reson. Med. 80, 2109–2121. https://doi.org/10.1002/mrm.27161 (2018).
Liao, J. R., Pauly, J. M., Brosnan, T. J. & Pelc, N. J. Reduction of motion artifacts in cine MRI using variabledensity spiral trajectories. Magn. Reson. Med. 37, 569–575. https://doi.org/10.1002/mrm.1910370416 (1997).
Haacke, E. M. et al. Magnetic resonance imaging: physical principles and sequence design. (1999).
Bracewell, R. N. & Bracewell, R. N. The Fourier Transform and Its Applications. Vol. 31999 (McGrawHill New York, 1986).
Beatty, P. J., Nishimura, D. G. & Pauly, J. M. Rapid gridding reconstruction with a minimal oversampling ratio. IEEE Trans. Med. Imaging 24, 799–808 (2005).
Jackson, J. I., Meyer, C. H., Nishimura, D. G. & Macovski, A. Selection of a convolution function for Fourier inversion using gridding (computerised tomography application). IEEE Trans. Med. Imaging 10, 473–478 (1991).
Park, J. Y. et al. Short echotime 3D radial gradientecho MRI using concurrent dephasing and excitation. Magn. Reson. Med. 67, 428–436 (2012).
Uecker, M. et al. Berkeley advanced reconstruction toolbox. in Proc. Intl. Soc. Mag. Reson. Med(2015).
Acknowledgements
This work was supported by NRF [National Research Foundation of Korea (Grant No. NRF2017R1A2B2004944)] and IBS [Institute for Basic Science (Grant No. IBSR015D1)].
Author information
Authors and Affiliations
Contributions
J.L. (1, Jeongtaek), J.P. and J.Y.P. conceived the research. J.L. (1) made a theoretical analysis. J.L. (1) and J.P. wrote the reconstruction program. J.L. (2, Joonyeol), S.K.L. helped with data analysis. J.L. (1), J.P. and S.K.L. drafted the manuscript. S.K.L. and J.Y.P. edited the manuscript and supervised the revision. J.Y.P. supervised the whole project. All authors participated in discussions, contributed ideas, read and revised the manuscript.
Corresponding authors
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher's note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Park, J., Lee, J., Lee, J. et al. Strategies for rapid reconstruction in 3D MRI with radial data acquisition: 3D fast Fourier transform vs twostep 2D filtered backprojection. Sci Rep 10, 13813 (2020). https://doi.org/10.1038/s41598020706984
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41598020706984
This article is cited by

Convolutional Neural Network in Medical Image Analysis: A Review
Archives of Computational Methods in Engineering (2023)

Line scanbased rapid magnetic resonance imaging of repetitive motion
Scientific Reports (2021)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.