Aliasing Artifact Reduction in Spiral Real-Time MRI

Abstract

Purpose

To mitigate a common artifact in spiral real-time magnetic resonance imaging (RT-MRI), caused by aliasing of signal outside the desired FOV. This artifact frequently occurs in mid-sagittal speech RT-MRI.

Methods

Simulations were performed to determine the likely origin of the artifact. Two methods to mitigate the artifact are proposed. The first approach, denoted “large FOV” (LF), keeps an FOV that is large enough to include the artifact signal source during reconstruction. The second approach, denoted “estimation-subtraction” (ES), estimates the artifact signal source then subtracts a synthetic signal representing that source in multi-coil k-space raw data. Twenty-five mid-sagittal speech production RT-MRI datasets were used to evaluate both proposed methods. Reconstructions without and with corrections were evaluated by two expert readers using a 5-level Likert scale assessing artifact severity. Reconstruction time was also compared.

Results

The origin of the artifact was found to be a combination of gradient nonlinearity and imperfect anti-aliasing in spiral sampling. The LF and ES methods were both able to substantially reduce the artifact, with an averaged qualitative score improvement of 1.25 and 1.35 Likert levels for LF correction and ES correction, respectively. Average reconstruction time without correction, with LF correction, and with ES correction were 160.69±1.56, 526.43±5.17, and 171.47±1.71 ms/frame.

Conclusion

Both proposed methods were able to reduce the spiral aliasing artifacts, with the ES reduction method being more effective and more time efficient.

Introduction

Real-time MRI is an established tool to study the dynamics of vocal tract shaping during human speech production. Compared to other imaging modalities – such as computer tomography (1), electromagnetic articulography (2), and ultrasound (3) – MRI (4–9) involves no ionizing radiation, can provide arbitrary imaging planes, and provides excellent tissue boundary delineation. RT-MRI for speech production is usually designed to achieve high enough temporal resolution to reveal fast movement of articulators and constriction events. Spiral sampling trajectory is commonly used due to its high sampling efficiency (10–12). This, combined with advanced reconstruction has made it possible to perform single-plane (13), multi-planar (14), and even 3D (15,16) visualization of speech articulators. Recently, real-time tagging (17,18) has also made it possible to visualize internal tongue mechanics.

One common artifact, aliasing from outside the desired FOV, occurs in all of these approaches that involve spiral sampling in the mid-sagittal vocal tract plane. The artifact is a result of gradient nonlinearity, spiral sampling, and subject anatomy. Gradient nonlinearity can cause strong image warping and intensity enhancement (19,20), especially at locations that are farther away from the magnet iso-center. Furthermore, spiral trajectories rotate, and therefore the readout anti-aliasing low-pass filter (LPF) does not provide FOV restriction (21,22). These two effects combine to cause a high signal area outside the FOV. This convolved with the spiral point spread function (PSF) can cause aliasing artifact within the FOV.

In a recent data collection effort leveraging spiral RT-MRI to study interactions of speech articulators and speech signal properties across talkers using the MRI protocol (23), 79% of scans suffered from different levels of the spiral aliasing artifact. These artifacts can overlap with important articulators, negatively impact the visualization of speech production, and also cause problems for automatic post-processing such as boundary detection (24,25). Previous solutions have included manually adjusting the imaging plane or adjusting the FOV such that the artifact minimally disrupts the targeted articulators under investigation. This manual approach limits the flexibility of the prescription, does not fully resolve the issue, and is time consuming. A general solution to address the spiral aliasing artifact is needed.

Some of the streak artifacts that are common in radial sampling (26–30) and quantitative susceptibility mapping (31,32) have a similar origin as the spiral aliasing artifact, arising from a high signal. Previous efforts to solve these artifacts belong to two types. One leverages parallel imaging, using either coil weighting (26,27,29,30) or coil sensitivities (28) to suppress the source of the aliasing. These methods can be limited when the coil array has a limited number of channels and can lead to signal reduction. Another type of artifact reduction performs a two-step reconstruction (31,32), where in the first step the high signal is estimated and subtracted, then allowing the low-signal area to be reconstructed without aliasing. Similar to the two-step method, partial FOV reconstruction has also been proposed to improve image quality (33) or to achieve a higher temporal resolution (34).

In this present study, we propose and evaluate two methods to suppress the spiral aliasing artifact, termed “large FOV” (LF) method and “estimation-subtraction” (ES) method. The LF method keeps a large FOV including the aliasing source during reconstruction. The ES method estimates the aliasing source signal from temporally combined image, synthesizes its k-space signal, and subtracts this signal in the acquired multi-coil k-space. The processed k-space is then used for image reconstruction. We evaluate both methods in the context of speech production RT-MRI, complemented with qualitative evaluation by two expert readers. Reconstruction time without and with both corrections is also compared.

Method

Speech RT-MRI Methods

Data acquisition in this work is based on the speech production RT-MRI system described in Lingala et al (13) and using a protocol described in Lingala et al (23). Image acquisition was performed on a 1.5T Signa Excite scanner (GE Healthcare, Waukesha, Wisconsin), with a custom 8-channel upper airway receive coil array, and a real-time interactive imaging platform RT-Hawk (Heart Vista, Los Altos, California) (35) to control the scan. A 13-interleaf spiral-out trajectory supporting a 20×20 cm² FOV with a bit-reversed temporal sampling order (36) was used to acquire k-space data. General scan parameters for the 2D spiral gradient-echo sequence were: TR/TE=6.0/0.8 ms, voxel size = 2.4×2.4×6 mm³, receiver bandwidth = ±125 kHz, flip angle = 15°. The system also records audio synchronized with MRI data and the details can be found in Lingala et al. (13).

A temporally constrained reconstruction (37) was used to reconstruct the dynamic images. The following cost function was solved using a conjugate gradient algorithm with line search:

$${\text{argmin}}_m\Vert FSm-d{\Vert }_2^2+{\lambda }_t\Vert \sqrt{{\left({\nabla }_{t}m\right)}^2+ϵ}{\Vert }_1$$ (1)

where S is the coil sensitivity map, F is the non-uniform fast Fourier transform (NUFFT), m is the dynamic images to be reconstructed, d is the acquired multi-coil k-space data, λ_t is the regularization parameter, ∇_t is the temporal finite difference operator, ϵ is a small positive constant to avoid singularity, and $\Vert \cdot {\Vert }_2^2$ and ‖·‖₁ are l₂ and l₁ norms, respectively. The algorithm stops at 150 iterations or when the line search finds a step size < 1e-5. Sensitivity maps were estimated from temporally combined coil images using the Walsh method (38). The regularization parameter was chosen to be 0.08C based on previous experience (13), where C is the highest pixel intensity in the adjoint NUFFT reconstructed images (F^T d). Dynamic images were reconstructed with 2 spiral arms/frame (12 ms/frame, 83.28 frames per second). Images were processed with 1.25×FOV (105×105 matrix size) during the iterative reconstruction and then cropped to the desired FOV (84×84 matrix size).

Simulation

To demonstrate the origin of the artifact, we performed simulations using a mid-sagittal slice of the XCAT phantom (39) covering from the head to mid-chest. We used Biot-Savart law to simulate coil sensitivities of the 8-channel custom receiver coil. A gradient profile of the Zoom coil (GE Healthcare, Waukesha, Wisconsin) in the form of spherical harmonic coefficients was used to approximate the gradient nonlinearity (40). Fully sampled multi-coil spiral k-space data was estimated with the 13-interleaf spiral trajectory, supporting a 20 cm circular FOV. We simulated a dwell time of 1 μs, and then sub-sampled the k-space to a dwell time of 4 μs, which was the dwell time used in in-vivo acquisitions. This allowed for simulation of the readout LPF, and its imperfections.

Artifact Reduction Methods

We propose two methods to correct for the aliasing artifacts, as illustrated in Figure 1. The LF approach keeps a large FOV including the artifact source during the reconstruction. The ES approach estimates the aliasing source then subtracts a synthetic signal representing that source from multi-coil k-space raw data.

Figure 1. Flowchart of the artifact reduction methods.

The raw k-space data was temporally combined and reconstructed to 2.5×FOV to obtain an image for each coil (coil images). Then the coil images were root sum-of-squares (SoS) combined and smoothed with a Gaussian filter. From the smoothed image, a mask (M₁) was automatically detected around the largest pixel intensity outside the $\sqrt{2}\times \text{circular}$ FOV by thresholding. The threshold was set to be 0.5× the highest pixel intensity outside the $\sqrt{2}\times \text{circular}$ FOV. From there, the large FOV reconstruction can be performed, by applying two masks including M₂ ($\sqrt{2}\times \text{circular}$ FOV) and M₁ during reconstruction. For the ES method, M₁ is applied on the coil images, and the masked coil images were used to estimate a k-space for the aliasing source. This estimated k-space was subtracted from the raw k-space for further processing.

The first step in both methods is to identify the aliasing source (M₁). A 2.5×FOV adjoint NUFFT reconstruction is performed on temporally combined k-space data d(k, c) to get coil images m(x, c). The coil images are root sum-of-squares combined, and the resulting image is smoothed by a 2D Gaussian filter to get m(x). k, x and c stand for k-space location, spatial location, and coil. The pixel with the highest intensity outside the $\sqrt{2}\times \text{circular}$ FOV (${\text{SI}}_{\text{max}}^{\text{out}}$) is identified, and its surrounding pixels with intensities $\ge 0.5\times {\text{SI}}_{\text{max}}^{\text{out}}$ compose M₁. The mask is then extended two pixels to smooth the boundary and fill holes. A ratio $r_{\text{out}/\text{in}}={\text{SI}}_{\text{max}}^{\text{out}}/{\text{SI}}_{\text{max}}^{\text{in}}$ is calculated to help indicating the level of aliasing artifact, where ${\text{SI}}_{\text{max}}^{\text{in}}$ represents the highest signal intensity (SI) within the $\sqrt{2}\times \text{circular}$ FOV.

Large FOV (LF) method

Aliasing is caused by signal leakage into undesired locations. If the aliasing source can be faithfully reconstructed, its aliasing can be reduced. Based on this, we propose to reconstruct the image with a large FOV including the aliasing source. However, reconstructing the image directly with a larger FOV (e.g., 2.5×FOV) causes blurring artifact, since the trajectory only supports a circular 1×FOV when fully sampled. Supporting Information Figure S1 illustrates the PSF analysis and an in-vivo example of the blurring. To solve this, two masks (M₁ and M₂) are applied during the reconstruction to keep only the signal within the masks, where M₂ is a $\sqrt{2}\times \text{circular}$ FOV, to maintain signal within a 1×square FOV when cropped. The masks can constrain the signal leakage and remove the blurring as illustrated in Supporting Information Figure S1.

Estimation-subtraction (ES) method

After the image space aliasing source M₁m(x, c) is identified, we estimate the k-space signal of the aliasing source $\widehat{d}\left(k,t,c\right)$ and subtract it form the acquired k-space d(k, t, c) to obtain aliasing-free k-space: $\tilde{d}\left(k,t,c\right)=d\left(k,t,c\right)-\widehat{d}\left(k,t,c\right)$. $\widehat{d}\left(k,t,c\right)$ is estimated by applying adjoint NUFFT to M₁m(x, c). Note that a scale factor was implicitly included in the adjoint NUFFT to match the intensities between $\widehat{d}\left(k,t,c\right)$ and d(k, t, c). Specifically, the scale factor a was calculated for each dataset as: $a=\underset{a}{\text{min}}{\Vert M_1{\sum }_t{F}^{T}d\left(k,t,c\right)-a{M}_1{\sum }_t{F}^{T}F\left[m{\left(x,c\right)}_{\times t}\right]\Vert }_2^2$, and $\widehat{d}\left(k,t,c\right)=aF\left[M_{1}m{\left(x,c\right)}_{\times t}\right]$, where ×t denotes that the matrix is replicated t times. Further reconstruction is done by replacing k-space data d(k, t, c) in equation (1) by $\tilde{d}\left(k,t,c\right)$.

Evaluation

We used a dataset resulted from a recent speech study data collection effort using the protocol (23) to evaluate the aliasing reduction methods. By the time of this study, data for 58 volunteers were prepared for use. Each subject was scanned while producing multiple speech stimuli, and we retrospectively chose the data corresponding to the stimuli provided in Appendix. USC Institutional Review Board approved the study and consent was obtained from each volunteer. The 58 datasets were reconstructed with three following methods: no correction, the ES correction, and the LF correction. Initial evaluation of these reconstructions was done by an MRI physicist to identify a threshold of r_out/in for apparent aliasing artifact.

Artifact level was qualitatively evaluated by two speech experts that have 18 years and 28 years of experience in reading speech RT-MRI, respectively. Artifact level was scored using a 5-level Likert scale, where the raters were instructed that a score of “1” indicated the “least severe artifact” and “5” indicated the “most severe artifact.” Before undertaking the rating task, two samples selected by an MRI physicist were given to the raters as representative of the two scale endpoints as defined above. All videos were evaluated without audio to eliminate potential bias due to audio quality. In the first rating round, 54 datasets (4 datasets were excluded by initial evaluation as explained later in Results) without correction were rated by both experts. From these results, 5 datasets were randomly picked from each of the 5 levels at which the two experts agreed, in order to balance the rating tasks of the corrected videos. If a certain level did not have enough (≥5) videos, we randomly filled the level with datasets where at least one expert had given that grade. Three datasets for level 3 were graded 3/2 and two datasets for level 4 were graded 5/4 by expert 1/expert 2. The resulting 25 datasets in three forms viz., no correction, ES correction, and LF correction were presented in a second rating task in a randomized order. These 75 videos were scored again by the original raters using the same scale.

The proposed methods were implemented in MATAB 2019b (MathWorks, Natick, Massachusetts), and executed on a Xeon 2.4GHz CPU (Intel Corporation, Santa Cara, California), with the iterative reconstruction portion (solving Equation (1)) running on a Tesla P100 GPU (Nvidia Corporation, Santa Clara, California). Reconstruction time without correction, with LF correction, and with ES correction was recorded.

Results

Figure 2 shows the simulation results. Figure 2(a) illustrates the extracted slice from the XCAT phantom (39). Figures 2(b–c) show the root sum-of-squares combined coil sensitivity map and the phantom image with distortion. Even though the coil has low sensitivity outside the desired FOV, the accumulation of signal due to gradient nonlinearity still produces a strong signal. Figures 2(d–e) show the limited effectiveness of readout LPF in suppressing the signal outside the FOV, which is consistent with prior reports that document the effects of LPF on spiral sampling (21,22). Figure 2(f) shows a representative in-vivo acquisition with the spiral aliasing artifact. The simulated artifact closely matches this in-vivo artifact, suggesting that the origin of the aliasing is largely due to gradient nonlinearity and spiral sampling. Supporting Information Video S1 shows a video reconstructed with 2.5×FOV and with fully sampled data (13 arms/frame). We observe a pulsatile signal change near the hotspot, which is consistent with cardiac pulsation.

Figure 2. Illustration of spiral aliasing artifacts in a numeric phantom and in-vivo.

The numeric phantom is based on XCAT (39) with spoiled gradient echo recalled contrast (a), using simulated coil sensitivities (root sum-of-squares (SoS) shown in (b)), and gradient non-linearity maps (not shown) of the Zoom gradient coil. Simulated images with nonlinear gradients (c) illustrate the geometric distortion and the hotspot outside the FOV (green arrow). Simulated spiral reconstructions without (d) and with (e) readout LPF both contain spiral aliasing artifacts appear within the FOV (blue arrows). This behavior closely matches in-vivo data, with a representative example shown in (f).

Figure 3 shows a representative example of aliasing reduction results. Two time frames are shown without correction, with LF correction, and with ES correction (Figure 3 a–b). Time-intensity profiles are shown in Figure 3(c). Both correction methods were able to reduce the aliasing artifacts identified by orange arrows, with the LF methods having residual artifacts also pointed by orange arrows. Supporting Information Video S2 shows the corresponding video.

Figure 3. Artifact reduction results.

This figure shows an example of the artifact reduction results. This subject was scored 5/4, 2/2, and 2/2 by expert 1/expert 2 for no correction, LF correction and ES correction, respectively. (a-b) show two time frames from the entire dynamic image series. The orange arrows point to spiral aliasing artifacts passing through the vocal tract. There are remining mild artifacts in the LF correction, and no apparent artifacts in the ES correction. (c) shows temporal line profiles of the three methods of the blue line marked in (a). The position of the line profile goes through an area affected by the artifact, pointed by orange arrows. Corresponding time frames shown in (a-b) are marked with dashed lines. The entire stimuli contain 5 sentences, with a total duration of 21.6 seconds. Supporting Information Video S2 shows the corresponding video.

The initial evaluation by the MRI physicist found one dataset had no apparent aliasing artifact, however, application of the ES and LF methods caused additional artifacts as shown in Supporting Information Figure S2. This dataset had an r_out/in ≤ 0.4; this threshold also identified three other datasets that had no apparent aliasing, and we excluded these 4 datasets in the following evaluation.

From the first-round speech expert evaluation, 46 out of 54 datasets were considered to have aliasing artifact by at least one expert (Likert score ≥ 2), resulting in 79% of the acquired datasets (46 out of 58) having artifact. Table 1 lists the average scores of the second-round rating. For datasets graded ≥2, the average score was improved by 1.1 or 1.4 with LF correction and by 1.4 or 1.3 with ES correction, according to the two raters. The average score for datasets without artifact (score = 1) was 1 for both LF and ES corrections as scored by rater 1 and was 1.4 for LF correction and 1.2 for ES correction as scored by rater 2. Figure 4 shows the second-round evaluation results in sanky diagrams.

Table 1. Evaluation results (2 raters) and reconstruction time of artifact resection methods.

Artifact severity level was scored by two expert raters using a 5-level scale, where the raters were instructed that a score of “1” indicated the “least severe artifact” and “5” indicated the “most severe artifact.” The listed scores were averaged based on the second rating task results. The “w/o artifact” indicates datasets that have videos without correction rated to “1”, and “w/ artifact” indicates the rest datasets. The reconstruction time was calculated based on time per time frame and averaged across all 58 datasets.

	No Correction	Method 1: Large FOV	Method 2: Estimate & Subtract
Expert 1 w/o Artifact	1.0	1.0	1.0
Expert 2 w/o Artifact	1.0	1.4	1.2
Expert 1 w/ Artifact	3.5	2.4	2.1
Expert 2 w/ Artifact	3.3	1.9	2.0
Reconstruction Time (ms/frame)	160.69 ± 1.56	526.43 ± 5.17	171.47 ± 1.71

Figure 4. Expert qualitative evaluation of artifact correction methods.

The results are shown in the sanky diagram separately for the two expert raters. In each diagram, the center column represents reconstructions without correction, and the left and right column represents reconstructions applied with LF and ES corrections, respectively. The number of datasets graded for each score is shown. The curves have a general trend of decreasing to the left (LF) and right (ES) sides, which means the artifact reduction methods have decreased the artifact severity.

The average reconstruction time was 160.69±1.56, 526.43±5.17 and 171.47±1.71 ms/frame for no correction, LF correction, and ES correction, respectively, as reported in Table 1. Overall, the results suggest that while both methods produce artifact reduction, the ES method is more effective and time efficient than the LF method in suppressing the artifact.

Discussion

We have identified the source of a common artifact observed in mid-sagittal spiral speech RT-MRI and have demonstrated two effective methods for mitigating this artifact. The source is a spurious signal hotspot caused by gradient nonlinearity and ineffective anti-aliasing filtering during spiral readouts. The two proposed solutions, termed LF (“large FOV”) and ES (“estimation-subtraction”), are both capable of mitigating the issues. The ES method proved to be faster and more effective than the LF method in the context of speech RT-MRI sampled with bit-reversed spiral trajectory.

We used a sagittal slice from the XCAT phantom and the spherical harmonic coefficients of the Zoom coil to demonstrate the spiral aliasing artifact. The spherical harmonic model is an approximation of the actual gradient map, especially at locations that are far away from the magnet iso-center. In practice, we often see the hotspot appears as an oval; however, a triangular hotspot occurred in our simulation. This is likely due to a gradually flattened gradient occurred in practice but not reproduced by the spherical harmonic model.

We have proposed two methods to reduce the spiral aliasing artifacts for spiral RT-MRI. The LF method requires a larger memory and a longer reconstruction time than the ES method. However, when the spiral trajectory is golden-angle (GA) or pseudo GA (PGA) rotated, the supported FOV increases with the number of unique rotations (41). When the supported FOV is large enough, the LF method can be simply applied without using mask and without introducing blurring. Supporting Information Video S3 shows an example of applying a 2.5×FOV reconstruction on a 34-interleaf PGA sampled dataset (14), where aliasing was reduced without blurring. In speech production RT-MRI studies in which the speech audio is recorded simultaneously with the image acquisition, an improved noise cancellation is achieved when using a periodic gradient oscillation (42). This improved noise cancellation method does not favor GA sampling or PGA sampling with a large cycle time.

The success of ES correction relies on a faithful reconstruction of the hotspot. In speech RT-MRI, vocal tract motion for speech production occurs in the region(s) of speech articulators and is not directly at the hotspot, which is substantially inferior to the vocal tract in this imaging protocol. Thus, it is possible to temporally combine the dynamic data to obtain a good estimation of the hotspot. Even though the prescribed FOV does not support the hotspot, the aliasing from signal within the FOV to the hotspot area is small, leading to a faithful reconstruction of the hotspot. When the hotspot moves, it may be possible to obtain a faithful reconstruction of the hotspot by using the sliding-window method (also known as view sharing) (43). If the motion of the hotspot is restricted within a small region, the LF method may be performed by using a static mask that includes all the range of movement. Both correction methods are not restricted to spiral sampling, and the ES method is not restricted to reducing artifacts arising from outside the FOV.

The proposed methods will require re-evaluation if applied to different body regions, datasets, resolution and/or FOV settings. Parameters such as the reconstructed FOV when identifying the aliasing source (currently at 2.5×FOV) should be adjusted based on the location of the aliasing source. The automatic detection of the aliasing source may need to be redesigned if there is a major change in the location, shape, and number of sources.

There are several limitations in this study. First, although we found r_out/in ≤ 0.4 can identify aliasing-free datasets, this threshold only selects a limited number of aliasing-free datasets (here e.g., 4 out of 12). An improved method to identify aliasing-free datasets can further avoid unnecessary application of the aliasing reduction methods. Another limitation is that we have not used quantitative metrics to evaluate the two proposed methods. In fact, we have applied several candidate metrics, including a recent proposed edge sharpness score (44). However, the unique feature of the artifact – usually a sharp arch across the FOV – may result in improved edge sharpness, leading to biased results. On the other hand, qualitative scoring by experts is subject to intra and inter-reader variability. That said, with regard to replication variability, Supporting Information Figure S3 shows the confusion matrixes of the two rating rounds for the no-correction reconstruction with qualitative scores being relatively repeatable, with only one case (1 reader out of 2, 1 dataset out of 25) where the two reads differed by more than 1 grade.

Conclusion

Gradient nonlinearity and spiral sampling can lead to unique aliasing artifact in mid-sagittal speech RT-MRI. Two proposed artifact reduction methods – a large-FOV method and an estimation-subtraction method – can substantially reduce these artifacts. The estimation-subtraction method is faster and more effective in the context of mid-sagittal speech RT-MRI with a bit-reversed spiral order. These methods may be generally applied to spiral imaging where the desired FOV is smaller than the signal producing region.

Supplementary Material

SUP

Supporting Information Figure S1. Point spread function analysis for the large FOV reconstruction. Top row shows PSFs of reconstructions using 1.25×FOV, 2.5×FOV, and 2.5xFOV with a $\sqrt{2}\times \text{circular}$ FOV mask (M₂). Trajectory used to generate these PSFs was a 13-interleaf spiral trajectory supporting an 84×84 matrix size, which is the same trajectory as used for in-vivo acquisition in this study. The PSFs of 1.25×FOV and 2.5×FOV with mask are similar, however the PSF of 2.5×FOV is wider as denoted by a black arrow, causing blurring in the reconstructed image. The bottom row shows a time frame of reconstructions with 1.25×FOV, 2.5×FOV, and 2.5xFOV with a $\sqrt{2}\times \text{circular}$ FOV mask (M₂), where the blurring can be seen on the 2.5×FOV reconstruction. Note that this example is used to illustrate the impact of M₂, thus we picked a dataset without aliasing artifact and did not use M₁ during reconstruction.

Supporting Information Figure S2. Artifacts caused by an oversized mask in ES and LF reconstructions. This figure shows an example of additional artifact caused by the spiral aliasing reduction methods, where there was no apparent aliasing artifact before correction. The automatic masking (a) was able to identify a hotspot outside the FOV; however, due to its low signal intensity, the mask includes a large area of only aliasing (or signal leakage) from the signal within the FOV. When this signal leakage is subtracted from the k-space as in the ES method or reconstructed as in the LF method, the signal within the FOV is affected, causing increased artifacts as denoted by yellow arrows (b). However, only 1 out of 58 datasets suffered from additional artifact, and the threshold r_out/in ≤ 0.4 was able to identify it as an aliasing-free dataset, preventing it from being processed.

Supporting Information Figure S3. Confusion matrixes of the two-rounds of evaluation. The figure shows confusion matrix data for the two rounds of evaluation in which raters score the no-correction reconstruction videos. The bias is mostly with ±1 grade, with only one instance having ±3 deviation.

V1. Supporting Information Video S1. Dynamic illustration of the spiral aliasing artifact.

This video was reconstructed to 2.5×FOV using NUFFT with 13 spiral arms per time frame to reduce the impact of regularization. The sagittal imaging plane goes through the main artery and heart of the subject. The aliasing artifact is nearly static; however, a pulsatile signal change close to the hotspot is seen, which is likely due to heartbeat.

V3. Supporting Information Video S3. Aliasing artifact reduction in PGA sampling.

This example shows a 34-interleaved PGA sampled acquisition reconstructed without correction, with the ES method, and with the LF method. The LF reconstruction was performed directly with 2.5×FOV without any masks. The LF reconstruction has little blurring artifact and reduced aliasing since a 34-interleaved PGA trajectory supports a large FOV, so direct reconstruction with a 2.5×FOV can reconstruct signal that includes the hotspot. The horizontal dark line going through the tongue is due to an interleaved multi-slice acquisition strategy with orthogonal slices. This results in additional saturation of locations where the slices intersect.

V2. Supporting Information Video S2. Video corresponding to Figure 3.

This example shows the reduced spiral aliasing artifact by both LF and ES methods. The video is provided with noise-cancelled speech audio.

Appendix

Read speech stimuli: “She had your dark suit in greasy wash water all year. Don’t ask me to carry an oily rag like that. The girl was thirsty and drank some juice followed by a coke. Your good pants look great however your ripped pants look like a cheap version of a k-mart special. Is that an oil stain on them?”. This stimuli is known as the phonetically rich sentences in the linguistics literature (23).