Self-gated 3D stack-of-spirals UTE pulmonary imaging at 0.55T

Abstract

Purpose:

To develop an isotropic high-resolution stack-of-spirals UTE sequence for pulmonary imaging at 0.55 Tesla by leveraging a combination of robust respiratory-binning, trajectory correction, and concomitant-field corrections.

Methods:

A stack-of-spirals golden-angle UTE sequence was used to continuously acquire data for 15.5 minutes. The data was binned to a stable respiratory phase based on superoinferior readout self-navigator signals. Corrections for trajectory errors and concomitant field artifacts, along with image reconstruction with conjugate gradient SENSE, were performed inline within the Gadgetron framework. Finally, data were retrospectively reconstructed to simulate scan times of 5, 8.5, and 12 minutes. Image quality was assessed using signal-to-noise, image sharpness, and qualitative reader scores. The technique was evaluated in healthy volunteers, patients with coronavirus disease 2019 infection, and patients with lung nodules.

Results:

The technique provided diagnostic quality images with parenchymal lung SNR of 3.18 ± 0.0.60, 4.57 ± 0.87, 5.45 ± 1.02, and 5.89 ± 1.28 for scan times of 5, 8.5, 12, and 15.5 minutes, respectively. The respiratory binning technique resulted in significantly sharper images (p < 0.001) as measured with relative maximum derivative at the diaphragm. Concomitant field corrections visibly improved sharpness of anatomical structures away from iso-center. The image quality was maintained with a slight loss in SNR for simulated scan times down to 8.5 minutes. Inline image reconstruction and artifact correction were achieved in <5 minutes.

Conclusion:

The proposed pulmonary imaging technique combined efficient stack-of-spirals imaging with robust respiratory binning, concomitant field correction, and trajectory correction to generate diagnostic quality images with 1.75 mm isotropic resolution in 8.5 minutes on a high-performance 0.55 Tesla system.

1 |. INTRODUCTION

Lower-field MRI systems equipped with high-performance hardware can potentially provide safe and comprehensive assessment of lung disease. Lower-field systems are well suited for structural and functional pulmonary imaging due to the reduced susceptibility and prolonged ${\text{T}}_2^{*}$ times.^1,2 In the lung, susceptibility gradients at air–tissue interfaces cause blurring and/or signal loss at 1.5T or 3T, and because susceptibility is directly proportional to field strength (B₀), these artifacts are significantly reduced at lower fields.

UTE sequences have found utility at 1.5 Tesla (T) and 3T for high-resolution structural lung imaging,^3–5 oxygen-enhanced imaging,^6,7 contrast-enhanced perfusion,⁸ and ${\text{T}}_2^{*}$ mapping.⁹ Most implementations have used 2D radial sampling, and more recently, 3D free-breathing radial,⁴ Fermat looped, orthogonally encoded trajectories,¹⁰ and yarnball¹¹ UTE sequences. Radial sampling is compatible with very short ${\text{T}}_2^{*}$ times but uses relatively low-sampling duty cycle. 3D UTE sequences are preferred to 2D UTE approaches because the 3D RF pulses are less sensitive to delays unlike the half-RF pulses used for 2D UTE imaging. Volumetric excitation is also more SNR-efficient; can achieve isotropic resolution; and when combined with 3D non-Cartesian trajectories, for example, 3D radial trajectories, offers reduced vulnerability to motion artifacts.¹²

Here we demonstrate the optimization of a free-breathing UTE 3D stack-of-spirals imaging sequence on a contemporary 0.55T scanner, which takes advantage of the reduced susceptibility and longer ${\text{T}}_2^{*}$ times at 0.55T for high-resolution imaging. ${\text{T}}_2^{*}$ in the lungs is approximately 10 ms at 0.55T,¹³ whereas it is 0.5–2.2 ms at 1.5T and 3T,^14,15 which may allow for longer signal readouts. Longer spiral readouts improve the efficiency of imaging sequences and are particularly attractive for 0.55T to combat SNR loss from moving to lower field strengths. We have previously demonstrated a breath-held lower anisotropic resolution stack-of-spirals acquisition for oxygen-enhanced MRI.¹⁶ However, to our knowledge, a free-breathing 3D stack-of-spirals UTE sequence has not previously been optimized for high-resolution lung imaging at 0.55T.

Even at lower field strengths, spiral imaging has its unique challenges. In general, spiral readouts are susceptible to blurring due to off-resonance and concomitant fields. Off-resonance blurring scales with B₀ and is reduced at lower field strengths; however, blurring due to concomitant fields scales inversely with B₀ and quadratically with gradient amplitude.¹⁷ Therefore, concomitant fields can be a significant source of blurring when using a high-performance gradient system at lower field strengths.

Free-breathing acquisitions for lung imaging require monitoring of respiratory motion. Johnson et al used adaptive respiratory gating with respiratory bellows for 3D radial UTE imaging.⁴ However, the use of respiratory bellows affects workflow and can be uncomfortable and unreliable. Alternatively, navigator pulses¹⁸ can be used for respiratory triggering, but these approaches affect lung signal and are not well suited for lung imaging. k-space center (DC) based navigation approaches have also been demonstrated and can be used for trajectories that sample the center of k-space frequently. However, these approaches require frequent sampling of the center of k-space and may not be compatible with all 3D trajectories. In this work, we employ a superoinferior (SI) readout for self-navigator signal, which has been demonstrated previously for cardiac imaging and high-resolution coronary artery imaging^19,20 and is potentially compatible with any 3D trajectory.

The purpose of this work is to demonstrate a robust method for free-breathing isotropic high-resolution lung imaging at 0.55T. Specifically, we optimized a 3D self-gated stack-of-spirals UTE sequence by leveraging corrections for trajectory errors and concomitant fields, with corrections and reconstructions implemented in-line for rapid image reconstruction using Gadgetron (v4.1).²¹ We demonstrate this method in healthy volunteers, patients with lung nodules, and patients with coronavirus disease 2019 (COVID-19) infection.

2 |. METHODS

2.1 |. Human subjects imaging

All subjects were scanned on a high-performance 0.55T system (prototype Magnetom Aera, Siemens Healthcare, Erlangen, Germany) under a National Heart, Lung, and Blood Institute Institutional Review Board-approved research protocol. Informed written consent was obtained from all subjects. Subjects were scanned using phased-array receiver coils including an 18-channel spine array and 6-channel body array. A total of 19 human subjects were scanned for this study: 10 healthy volunteers, 3 patients with active COVID-19 infection, 4 patients who recovered from COVID-19, and 2 patients with lung nodules.

All subjects were scanned using a free-breathing 1.75 mm isotropic-resolution stack-of-spirals imaging protocol with the following imaging parameters: TE/TR = 0.5/7.7 ms, flip angle = 5°, readout duration = 5.4 ms, FOV = 450 × 450 × 224 mm³, matrix size = 256 × 256 × 112, peak gradient amplitude = 26.7 mT/m, acquisition bandwidth = 500 KHz, total acquisition time = 15 minutes 30 seconds, coronal acquisition. The spatial resolution and imaging parameters used in this study were selected to detect lesions larger than 4 mm with adequate contrast and SNR. The spiral readout duration (5.4 ms) was heuristically determined in 4 healthy volunteers who were scanned with readout durations of 2.8, 5.4, and 7.2 ms (Supporting Information Figure S1). To explore the minimum required acquisition time, we also retrospectively clipped the data from the end of the scan to simulate reduced scan times (5 minutes, 8.5 minutes, and 12 minutes) and the effect on image quality.

2.2 |. Pulse sequence

We modified a prototype self-gated stack-of-spirals 3D UTE spoiled gradient echo sequence (Siemens Heathcare, Erlangen, Germany) for the purposes of this study. A slab-selective minimum-phase RF pulse with a time-bandwidth of 5 and duration of 0.35 ms was used for both navigator and data acquisitions. Data sampling was performed with a spiral-out trajectory at a fixed TE. We used slice oversampling factor of 15% to compensate for the obtuse pulse profile. Slice-phase encoding was the inner loop of the acquisition, and the interleave loop was the outer loop; that is, for every interleave, all slice-phase encodes were acquired before the next interleave.

The spiral waveforms for the stack-of-spirals readouts were adaptively designed in the pulse sequence using the slew-rate-limited approach.²² In this approach, the k-space trajectory is calculated based on a fixed duration for each spiral arm with constraints on slew rate and peak gradient amplitude, allowing for control of off-resonance–related blurring (dependent on readout duration) and concomitant field-related blurring (dependent on readout duration and peak gradient amplitude). In this work, the spiral trajectory was designed to fully sample k-space with 19 shots, and each interleave was rotated with golden angle (180 (3 − √5) degrees) to efficiently cover k-space. A total of 911 shots were acquired in 15 minutes and 30 seconds to match the SNR of previously published works at higher field strengths.^4,23

2.3 |. Embedded navigator gating

SI navigator readouts were used to estimate respiratory motion. The readout direction for the navigator pulses was fixed in the SI direction to maximize sensitivity to physiological motion during respiration. A navigator readout was acquired every 200 ms, with readout duration = 3 ms and TR = 7.7 ms. The sampling interval for the navigator readout was determined heuristically to balance scan efficiency and accurate respiratory motion signal measurement.

The respiratory motion signal was extracted from the SI navigator readouts using principle component analysis,²⁰ as shown in Figure 1:

1. The navigator readouts were acquired at regular intervals during the 3D stack-of-spirals acquisitions, which makes them sensitive to a trajectory-dependent signal modulation. The navigator readouts were sorted by the azimuthal angle of the preceding spiral readout and reshaped into a matrix of size (N_c × N_r) × N_int × N_nav, where N_c is the number of channels; N_r is the number of samples in the SI readout; N_int is the number of interleaves (azimuthal angles) for golden-angle view ordering; and N_nav is the number of navigator readouts per interleave. (In this study, 5 navigators were acquired per stack-of-slices at a given azimuthal angle).

2. Angular filtration was used to remove artifactual trajectory-dependent signal modulation from the navigator readouts.²⁰ The sorted navigator readouts were filtered along the angular direction (N_int) using a Kaiser window high-pass filter with a pass band of 0.04 rad⁻¹. The filtered readouts were resorted into the original temporal order for further processing.

3. Data after angular filtration was sorted into the matrix of size (N_c × N_r) × N_t, after computing the Fourier transform along the readout dimension, where N_t is the total number of navigator readouts during the image acquisition. Navigator readouts were then filtered using a Kaiser window bandpass filter (pass band 0.1 Hz–0.7 Hz and stop band <0.05 Hz and >0.8 Hz) along the time dimension to isolate relevant respiratory frequencies.

4. Singular value decomposition was used to estimate the eigen vectors of the filtered data. The eigen vector with the largest singular value was selected for each channel N_c.

5. Previously published coil-clustering method was used for choosing the channel that best represents respiratory motion signal.²⁴

FIGURE 1.

(A) Flowchart for extraction of respiratory signal from SI navigator readouts. The navigator readouts are reshaped into a matrix of shape (N_c × N_r) × N_int × N_nav before angular filtration along the interleave dimension. The navigator readouts after angular filtration are reshaped into a matrix of shape N_c × N_r × N_t. They are then filtered using a bandpass filter along the time dimension and reshaped into a matrix of shape N_r × N_t × N_c. Singular value decomposition is used to select the eigen vector with the largest singular value from each of the shape N_c channels. This now represents the motion signal in the range of respiratory frequencies for each channel. The motion signals are clustered using previously published methods to get the best representation of respiratory motion signal. (B) The Kaiser window high-pass filter along the angular dimension used in step 2 and Kaiser window bandpass filter used in step 3. The gray dotted lines represent the transition regions from the stop to pass band (and vice versa) in the filter response plots. N_c, number of channels; N_int, number of interleaves; N_r, number of samples; N_t, total number of navigator readouts; SI, superoinferior

The respiratory navigator signal was used to sort all of the interleaves into motion states. A bin that contains 40% of the data in the most stable respiratory phase was used for image reconstruction. The effect of angular filtration on respiratory motion signal and subsequent image quality was assessed in 6 healthy volunteers. In these volunteers, respiratory motion signal and images were reconstructed without and with angular filtration (steps 1 and 2 above).

2.4 |. Concomitant field correction

Concomitant fields are nuisance nonlinear spatially dependent magnetic fields that are produced whenever gradient field coils are activated. Fortunately, these fields can be predicted using Maxwell’s equations. Assuming symmetric gradient coils, typical for high-performance and clinical systems, the phase generated by these fields to the lowest order is given by:

$$\begin{gathered} {\varphi }_c(x,y,z,t)=\gamma {\int }_{-\infty }^t\frac{g_z^2\left(t^{\prime }\right)}{8B_0}\left(x^2+y^2\right)+\left(\frac{g_x^2\left(t^{\prime }\right)+g_y^2\left(t^{\prime }\right)}{2B_0}\right)z^2 \\ -\left(\frac{g_x{g}_z}{2B_0}\right)xz-\left(\frac{g_y{g}_z}{2B_0}\right)yzd{t}^{\prime }, \end{gathered}$$

where B₀ is the main magnetic field; and g_x, g_y, and g_z are gradient waveforms along the physical x, y, and z axes. The image distortion caused by concomitant fields is inversely proportional to B₀ and maximum gradient amplitude. For spiral imaging, this image distortion appears as a spatial blur that worsens away from iso-center, with larger gradient amplitudes and longer readouts. In this work, we estimated the concomitant fields using previously described methods^17,25 and corrected for them using multi-frequency interpolation (MFI).²⁶ Briefly, spiral data was demodulated at 24–29 frequencies, and pixel-wise weighted sum of the demodulated images based on the estimated local frequency was used to reconstruct the final corrected image. Weights for combining the images reconstructed at various frequencies for MFI were estimated using least squares with QR algorithm.

2.5 |. Image reconstruction

The image reconstruction workflow is shown in Figure 2. All image reconstructions were developed in MatLab (R2021a, Mathworks, Natick, MA) and then translated to Gadgetron²¹ for fast inline image reconstruction. Image reconstructions were performed on a system equipped with Dual Intel Xeon processors (Intel Corporation, Santa Clara, CA), 512 GB RAM, 3x Nvidia Quadro RTX 8000 (NVIDIA Corporation, Santa Clara, CA) graphical processing units (GPU). To improve the speed of reconstruction, GPU implementations were used for both density compensation and reconstruction.

FIGURE 2.

Image reconstruction flowchart. Data and gradient waveforms were streamed from the scanner to Gadgetron. The gradient waveforms were converted to trajectories and attached to the corresponding data for reconstruction. The entire dataset up to 30 GB was used to generate the coil sensitivity maps, which were used for CG-SENSE reconstruction. The weights for the binned and non-binned data were estimated using iterative density compensation. The navigator data was used to extract the respiratory signal, which was used to bin 40% of the data to the most stable respiratory phase. The binned acquisitions (data and trajectories) were passed to the reconstruction. The trajectories were converted to gradients and demodulation frequencies Δf_i, and combination weights C_i for demodulated images were estimated. The data was demodulated at each frequency, and NUFFT was used to perform the gridding reconstruction. The corrected image was generated using the following equation $I={\sum }_{i=0}^N{C}_{i}I{m}_i$. The corrected image was then passed into the CG-SENSE reconstruction to get the final reconstructed images. CG, conjugate gradient; NUFFT, nonuniform fast Fourier transform transform

The gradient waveforms were streamed along with the data from the vendor reconstruction software to Gadgetron to be used for image reconstruction. Gradient waveforms prescribed by the pulse sequence are distorted when they are played out on the scanner due to hardware imperfections, which can cause image artifacts.²⁷ The gradient impulse response function was used to characterize the gradient system,^27,28 and the gradient impulse response function measurements were applied to correct trajectory imperfections during image reconstruction. An iterative technique was used to estimate weights for density compensation²⁹ with a kernel width of 5.5, oversampling factor of 2.1 (matched with nonuniform fast Fourier transform transform), and 15 iterations.

Coil compression was used to conserve GPU memory and improve reconstruction speed. Data from 18 channels was compressed down to 8 channels. The number of compressed channels was heuristically determined to minimize degradation in image quality due to coil compression. Coil sensitivity maps (CSM) were estimated from non-binned data using the Walsh method.³⁰ The data for CSM was limited to 15 GB (total data size is 40 GB, and 22 GB without and with coil compression) to allow GPU-based reconstruction and CSM estimation. We did this by triggering the CSM estimation once 15 GB of coil-compressed data had been collected. We found this to be sufficient for estimating high-quality coil sensitivity maps even for isotropic imaging. The respiratory-binned data was reconstructed using conjugate gradient (CG) SENSE reconstruction to reduced artifacts from nonuniform undersampling in k-space.³¹ A tolerance of 1e-3, max. iterations 3, and no regularization were used for CG-SENSE reconstruction.

Concomitant field correction was done prior to CG-SENSE reconstruction. Spiral data was demodulated at frequencies estimated for MFI, and non-uniform fast Fourier transform was used to generate demodulated images with undersampling artifacts. These images were combined using the MFI weights to generate a concomitant field-corrected image, which was used as the input to a CG-SENSE solver as shown in Figure 2.

2.6 |. Image analysis

Respiratory binning and image reconstruction were repeated separately for the full 15.5- minutes acquisition, as well as the clipped 5 minutes, 8.5 minutes, and 12 minutes acquisitions. To quantitatively assess image quality, several metrics were computed.

1. The improvement in image sharpness from robust binning was assessed using relative maximum derivative, as previously demonstrated.³² Relative maximum derivative measures the sharpness of diaphragm and was calculated for images reconstructed 1) with binning without angular filtration, and 2) with binning and angular filtration. Paired t test was used to compare the difference in relative maximum derivative without and with angular filtration from the 6 healthy volunteers. A higher relative maximum derivative signifies sharper diaphragm boundary and improved binning performance.

2. SNR with pseudo-replica method³³ and apparent SNR (aSNR) were calculated in 6 healthy volunteers for the 4 scan times. aSNR was measured as the ratio of mean signal in a region of interest divided by SD of noise-only region (the airway). Both SNR and aSNR were measured using real-valued images to avoid Rician noise distribution, with regions of interest in the lung parenchyma, aortic arch, and an airway as shown in Figure 6A. aSNR was calculated to perform comparisons with previously published work at higher fields.

3. The image quality and artifacts were qualitatively evaluated by a senior radiologist with 15 years of experience. Images were scored using a 5-level Likert scale for image quality (5 = excellent image quality with high SNR; 1 = very poor, nondiagnostic image quality) and for artifact level (5 = no artifacts; 1 = severe artifacts due to undersampling and/or cardiorespiratory motion). Images from 6 healthy volunteers reconstructed for 5 minutes, 8.5 minutes, and 12 minutes, and 15.5 minutes scan times were scored, which resulted in 24 data sets. The images were randomized before scoring to enable double blind analysis.

FIGURE 6.

SNR for acquisition times of 5, 8.5, 12, and 15.5 minutes. (A) Shows the regions of interests where SNR was measured (blue) airway, (orange) lung parenchyma, and (green) aortic arch. (B) SNR measured using the pseudo-replica method for acquisition times of 5, 8.5, 12, and 15.5 minutes

3 |. RESULTS

3.1 |. Image reconstruction

For the full 15.5-minute acquisition, the image reconstruction, including concomitant field correction, required <4 minutes after the end of the scan. Data sizes and reconstruction times are provided in Table 1 for reduced scan times. The reconstruction time was divided as follows: data processing and binning 20 seconds, CSM estimation 40 seconds, CG-SENSE with concomitant field correction approximately 180 seconds. Image reconstruction code is open source and available at: https://github.com/NHLBI-MR/selfgated_noncartesian_reconstruction. Iterative density compensation helped reduce artifacts from nonuniform sampling in k-space, which is common for binned golden-angle reconstructions. An example of this is shown in Supporting Information Figure S2.

TABLE 1.

Image reconstruction times and image quality measurements

Acquisition time	Data size	Reconstruction time	Median image quality score	Median artifact quality score	SNR			aSNR
					airway	Lung	Aortic arch	airway	Lung	Aortic arch
5 minutes	13.3 GB	2.5 minutes	3.5	3.5	0.7 ± 0.4	3.2 ± 0.6	11.2 ± 3.7	0.6 ± 0.4	2.9 ± 0.6	10.2 ± 2.7
8.5 minutes	21.3 GB	2.8 minutes	4.5	4.0	0.9 ± 0.5	4.6 ± 0.9	14.8 ± 4.9	0.8 ± 0.4	3.9 ± 0.5	12.7 ± 3.3
12 minutes	32.0 GB	3.3 minutes	5	4.0	1.1 ± 0.5	5.5 ± 1.0	17.1 ± 5.3	0.9 ± 0.5	4.5 ± 0.8	14.5 ± 4.5
15.5 minutes	40.0 GB	4.0 minutes	5	4.0	1.2 ± 0.5	5.9 ± 1.3	23.3 ± 6.4	1.0 ± 0.4	4.8 ± 1.1	15.7 ± 5.8

Image quality and artifact score are based on 5-level Likert scale for image quality (5 = excellent image quality with high SNR; 1 = very poor, non diagnostic image quality) and for artifact level (5 = no artifacts; 1 = severe artifacts due to undersampling and/or cardiorespiratory motion). Image SNR and aSNR is given as mean ± SD.

Abbreviation: aSNR, apparent SNR.

3.2 |. Respiratory motion signal

Figure 3 demonstrates the improvement in accuracy of the extracted respiratory motion signal when using angular filtration. Figure 3A) shows the extracted respiratory motion signal and the frequency spectrum of the respiratory motion signal with and without angular filtration. The 2 artifactual peaks in the spectrum are indicated by the red arrows. These peaks are likely caused by the trajectory-dependent modulation of the navigator readouts. Figure 3B) shows examples of images that were produced from data binned without and with angular filtration. Red arrows point to artifacts in images that were likely due to trajectory dependence of respiratory motion signal. Green arrows highlight improvements in image sharpness due to improvement in accuracy of the motion signal with angular filtration. Figure 4C shows the improvement in sharpness of the diaphragm measured using relative maximum derivative, which was used as a metric for evaluating the performance of respiratory binning. The relative maximum derivative with angular filtration was significantly (p < 0.001) higher than relative maximum derivative without angular filtration.

FIGURE 3.

Effects of angular filtration on respiratory signal and images reconstructed from the binned data. (A) Plots of respiratory signals with angular filtration are shown in blue, without angular filtration are shown in orange. Left: Respiratory waveform extracted from the navigator data both with and without angular filtration, shown for a 50 seconds window. The red arrows show the suspected artifactual peaks in the respiratory waveforms. Right: Frequency spectrum of the respiratory waveform with and without angular filtration. The frequency spectrum was calculated by taking the centered Fourier transform of the respiratory waveform. Red arrows point to artifactual peaks in the spectrum due to trajectory dependent modulation. These peaks were present in all subjects acquired with this protocol. (B) Shows the effect of angular filtration on image quality in binned data for a representative example. The red arrows point to trajectory-dependent artifacts, and green arrows highlight more subtle improvements in image sharpness due to the extraction of the correct respiratory signal with angular filtration. (C) Relative maximum derivative calculated to estimate the sharpness of the diaphragm without and with angular filtration. ** represents statistically significant difference with p < 0.001 (paired t test)

FIGURE 4.

Coronal 3D UTE maximum intensity projection images in a healthy volunteer and patient with lung nodules (A,D) before, (B,E) after retrospective binning, and (C,F) following concomitant field corrections. (A-F, top) shows full lung FOV and (bottom) shows zoomed lung FOV around the apex of the lung to highlight improvements from concomitant field corrections. Blue boxes highlight the region around the diaphragm that shows improvement in image sharpness from binning. Green boxes highlight the region, which shows significant improvement in image sharpness due to concomitant field correction. Yellow circles show the lung nodule, which was blurred due to respiratory motion and concomitant fields. Significant improvement in pulmonary vessel delineation in the lungs is shown with concomitant field correction in both subjects. The maximum intensity projections were generated with 15 central slices in the healthy volunteer and 7 slices around the lung nodule in the patient with lung nodule

3.3 |. Respiratory motion and concomitant field artifact correction

Improvement in image quality from robust respiratory binning and in-line concomitant field corrections are demonstrated in Figure 4 for a healthy volunteer and a patient with lung nodules using maximum intensity projections of 15 slices for the healthy volunteer and 7 slices for the patient with lung nodules. The image sharpness around the vessels and diaphragm is visibly improved with respiratory binning (blue boxes). Concomitant fields are directly proportional to the square of distance from isocenter in the SI direction and to distance from isocenter in left–right and anteroposterior directions. Therefore, blurring and signal loss are worse at the lung apex, where improvements in sharpness of vessels/nodule are most evident (green boxes). The yellow dotted circle highlights the lung nodule, which is better delineated in the respiratory-binned and concomitant field-corrected image.

3.4 |. Scan duration comparisons

Figure 5 shows images reconstructed for scan times of 5, 8.5, 12, and 15.5 minutes in 3 subjects. Overall, we did not observe significant deterioration in image quality (i.e., significant loss in signal, or reduced visibility of anatomy) with simulated reduction in scan time down to 8.5 minutes which shows that, potentially, an 8.5 minutes scan is sufficient to achieve diagnostic quality images with our technique. We also show the measured reduction in SNR and aSNR from shorter scan times in Figure 6B and Table 1. The average SNR (mean ± SD) in 6 healthy volunteers was 3.18 ± 0.0.60, 4.57 ± 0.87, 5.45 ± 1.02, and 5.89 ± 1.28 in the lung parenchyma; and the results of qualitative image scoring were (mean ± SD) of 3.67 ± 0.82, 4.50 ± 0.54, 4.67 ± 0.51, and 4.67 ± 0.51 for scan durations of 5, 8.5, 12, and 15.5 minutes. Supporting Information Figure S3 shows the apparent SNR to serve as comparison to previously published work. We did observe some signal in the airway as shown in Figure 6B and Table 1, which was likely due to partial volume effects from nearby tissues. The SNR in the background was measured to be <0.3 for all scan times.

FIGURE 5.

Effect of simulated scan times on image quality in 3 representative cases. MIPs over 15 central slices and 1 central slice are shown for all 3 cases for scan times of 5, 8.5, 12, and 15.5 minutes. MIP, maximum intensity projection

3.5 |. Patient imaging

Figure 7 shows representative examples of MIPs and single-slice images from each subject group reconstructed from the 8.5 minutes acquisition and includes a healthy volunteer, a patient with lung nodules, a patient with acute COVID-19 infection, and a patient who had recovered following COVID-19 infection. The healthy volunteer had scoliosis but was otherwise healthy. The patient with lung nodule has 2 nodules: one 7 mm and one 26 mm; the slices with the 7 mm nodule are shown. The 7 mm nodule was confirmed with CT and is clearly visible and well delineated in the MR images. The patient with active COVID-19 infection had ground glass opacities in the right lung, which were confirmed with CT and can be clearly seen in MR images. The patient recovered following COVID-19 did not have significant pulmonary infiltrates present. The other 5 patients with a history of COVID-19 infection showed no ground glass opacities or infiltrates. Supporting Information Video S1 shows all individual slices through the lung in all 4 subjects from Figure 7 and Supporting Information Video S2 shows all individual slices through the lung in the 5 patients with known disease which are not shown in Figures 1–7.

FIGURE 7.

Representative image quality for 8.5 minutes clipped scan duration in cases of healthy volunteers, patients with lung nodules, patient with active COVID-19 infection, and a patient after recovery from COVID-19. (top) Coronal maximum intensity projections of 15 slices are shown; (center) single central slice from the stack used to generate the MIP is shown; and (bottom) an axial slice is shown. A 7 mm lung nodule is shown in the patient with lung nodules in the upper right lung (yellow circle). Ground glass opacities are shown in the patient with active COVID-19 infection in both the upper right and upper left lungs

4 |. DISCUSSION

In this work, we demonstrate a robust high-resolution pulmonary imaging technique for 0.55T systems, which can achieve 1.75 mm isotropic resolution in 8.5 minutes. We did this by combining robust self-gating using an SI readout, UTE stack-of-spirals acquisition, trajectory corrections, and concomitant field corrections. Respiratory gating, image reconstruction, and artifact correction were implemented in Gadgetron and deployed inline on the MRI system such that images return to the scanner within 5 minutes of the acquisition for integration into the clinical workflow. We demonstrated high-quality imaging in healthy volunteers and in patients with known lung disease. We anticipate that higher spatial resolution can be achieved with this sequence but may require longer scan times than 8.5 minutes to achieve sufficient SNR.

The use of low-field MRI for high-resolution pulmonary imaging has previously been explored at 0.2T.^2,34,35 These previous studies used permanent magnets and/or open magnet designs,² which offer inferior field homogeneity and low-performance gradients but were cheaper and more patient-friendly. Studies with these MRI systems had low SNR and field inhomogeneity-related artifacts. SNR reduction is one of the key disadvantages of moving to lower fields, and minimizing SNR loss is especially important in pulmonary MRI where proton density is low. Efficient trajectories (e.g., spirals) can improve pulse-sequence efficiency and allow recovery of SNR.³⁶ Our 0.55T system was equipped with a contemporary superconducting magnet, which offers a homogeneous magnetic field to minimize image blurring and other susceptibility artifacts and used a high-performance gradient system (maximum gradient = 45 mT/m, maximum slew rate = 200 mT/m/ms), which enable efficient pulse-sequence design.

In this work, we used stack-of-spirals for 3D imaging instead of a Cartesian readout. 3D non-Cartesian image reconstruction adds substantial complexity. However, stack-of-spirals acquisition enabled longer readouts and more efficient sampling of k-space, resulting in improved SNR, as demonstrated in Supporting Information Figure S4. We show that, compared to Cartesian imaging, a stack-of-spirals trajectory with a 5.4 ms readout increases the efficiency of the sequence by >30%. Stack-of-spirals imaging also enabled shorter TEs, 0.52 ms compared to 1.2 ms for Cartesian imaging, which led to a further improvement in SNR of approximately 7%. Overall, the stack-of-spirals imaging with a 5.4 ms readout lead to a >40% improvement in SNR compared with 3D Cartesian imaging. Supporting Information Figure S5 shows that, for the same data sampling time per acquisition, the total time required for the 1.75 mm isotropic volume is 14 minutes 54 seconds for Cartesian versus 8.5 minutes for stack-of-spirals acquisition. It also shows the reduction in total scan time with readout duration for spiral imaging. Finally, using stack-of-spirals trajectories enabled the use of golden angle view-ordering for retrospective binning of respiratory navigator-gated acquisition.

The implementation presented in this work gives comparable image quality to recent pulmonary imaging techniques presented on 1.5T and 3T systems. The recent work presented on 1.5T scanners achieved 1.3 mm isotropic resolution in 16 ± 1 minutes to achieve diagnostic image quality.³⁷ On 3T scanners, building on the work by Johnson et al, who achieved 1.25 mm isotropic resolution with diagnostic image quality in 9.5 minutes,⁴ Zhu et al recently demonstrated an imaging technique that leverages iterative motion-corrected reconstruction to achieve 1.25 mm isotropic resolution in 5.5 minutes at 3T.³² They were able to do so by using all of the data (all respiratory phases) to reconstruct a single respiratory phase. The iterative motion correction method achieved an average aSNR of approximately 3.5 in the lung parenchyma and 17.5 in the aortic arch. Comparatively, our method without iterative motion correction reconstruction and with only 40% of the acquired data was able to achieve diagnostic image quality with 1.75 mm isotropic resolution in 8.5 minutes with aSNR of 6.09 ± 0.70, and 20.91 ± 5.60 in the lung parenchyma and aortic arch, respectively.

There are several ways to implement concomitant field correction. Two prominent approaches are: 1) MFI with demodulation in frequency domain, and 2) image convolution method with a nonstationary kernel.³⁸ In this work, we used the MFI-based method in which the correction was performed before the CG-SENSE reconstruction. Alternatively, the concomitant field correction could be implemented as part of the encoding operator of the CG-SENSE reconstruction, but we did not do this to keep the reconstruction times clinically feasible with the available hardware. Overall, the reconstruction time for CG-SENSE could be further reduced with more GPUs; however, these require sufficient capacity (>22 GB) to perform gridding of large datasets. In general, we believe that to make this correction compatible with a range of constrained reconstructions, it should be implemented as part of the encoding operator or using the image convolution approach. This will be explored in future work.

Our methods also had some limitations. We used self-navigators for respiratory motion binning but did not detect or reject bulk motion, which can significantly impact image quality if the patient moves during the scan. In future work, we can explore the use of image-based navigators, which can be used to detect bulk motion. Data affected by bulk motion can either be rejected or corrected using motion correction based reconstructions. We binned to only 1 respiratory motion state and rejected approximately 60% of the data. Additionally, 40% of data binned for all subjects can also result in blurring due to residual respiratory motion. Iterative motion correction–based approach³² might allow further improvements to the efficiency of our technique by using the entire dataset for reconstruction and reduce blurring due to residual respiratory motion from stable binning a fixed amount of data. Motion correction–based approaches may also help reduce scan duration, which will significantly reduce the risk of respiratory drift and bulk motion. Residual blurring due to off-resonance was present in a small number of cases and was not corrected with our existing method. We will explore ways to efficiently incorporate recently published off-resonance correction methods into our pipeline in future work.^39–41 The qualitative scoring in this work was only done by 1 reader, which is subject to variability that was not measured. However, the purpose of scoring here was to evaluate image quality with simulated reduction in scan time, which was sufficiently accomplished with just 1 reader. We used a slab-selective excitation and a stack-of-spirals trajectory, which requires phase encoding in the slice direction and led to a significantly longer TE of 0.52 ms versus 0.18 ms had we used a 3D trajectory such as 3D cones. Using a longer TE reduced the efficiency of the sequence and caused an SNR loss of approximately 6%.

5 |. CONCLUSION

In this work, we describe an optimized approach for free-breathing high-resolution UTE pulmonary imaging on a high-performance 0.55T system. We implemented in-line artifact corrections and rapid image reconstruction within the open-source Gadgetron framework to enable diagnostic quality images in clinical settings. We validated and evaluated our technique in healthy volunteers, patients with lung nodules, and patients with COVID-19 infection.

Supplementary Material

Supporting figures

FIGURE S1 Comparison of image-quality for spiral readout durations of 2.8, 5.4, and 7.2 ms in four healthy volunteers. Increasing the readout duration from 2.8 ms to 5.4 ms leads to a substantial improvement in image quality due to improved SNR. However, further increase in readout duration to 7.2 ms offers negligible additional SNR. The figure also shows that using a readout of 7.2 ms increases the susceptibility to blurring due to off resonance as pointed to by the red arrows. Readout duration of 5.4 ms presents a good balance of improvement in SNR and susceptibility to off-resonance related blurring and was, therefore, chosen for this study

FIGURE S2 Effect of iterative density compensation to improve image quality and reduce artifacts from non-uniform sampling. The figure shows binned images reconstructed without iterative density compensation, with iterative density compensation and difference images. Hoge’s method (Hoge RD et al MRM 2005) was used to estimate weights in images reconstructed without iterative density compensation. The red arrows show artifacts due to non-uniform sampling in images reconstructed without iterative density compensation

FIGURE S3 Apparent signal-to-noise ratio (SNR) for acquisition times of 5, 8.5, 12, and 15.5 minutes

FIGURE S4 Improvement in data sampling efficiency and SNR with increased readout duration. (A) Efficiency vs. readout duration (B) relative SNR vs. readout duration compared to a Cartesian readout (red circle). Data sampling efficiency was calculated as readout duration/TR × 100. Relative SNR was calculated using √[(spiral sampling efficiency/Cartesian sampling efficiency)] × (exp(−0.52 ms/T₂)/exp(−1.2 ms/T₂)). For spiral imaging, TE = 0.52 ms and for Cartesian imaging TE = 1.2 ms, based on sequence simulations. These plots assume TR = 3 ms and readout duration = 1.2 ms for Cartesian imaging, and TR = readout duration + 2.3 ms for stack-of-spirals imaging. We also assumed that 1) the total acquisition time for both Cartesian and stack-of-spiral acquisitions was fixed at 8.5 minutes, 2) the resolution for both trajectories was the same, 3) minimum TR was used for both acquisitions, and 4) effects of TR on GRE signal were ignored. These simulations estimate an SNR increase of 40% compared to Cartesian readout for the parameters used in this study (spiral readout = 5.4 ms, TE = 0.52 ms)

FIGURE S5 Total scan time vs. readout duration comparing our spiral readout (blue arrow) and a simulated 3D Cartesian trajectory (red arrow). For comparison, we used a fixed spatial resolution and fixed data sampling duration matched to our spiral acquisition (8.5 minutes total scan time × 70% sampling efficiency = 357 s total sampling), and we assumed the same data sampling duration (357 s) for the simulated Cartesian acquisition. These two sampling strategies should generate matched SNR, ignoring TE differences and under-sampling differences between the two sampling patterns, which both favor stack-of-spirals. Under-sampling is difficult to predict in our simulations because they are dependent on respiratory pattern. These plots assume TR = 3 ms and readout duration = 1.2 ms for Cartesian imaging, and TR = readout duration + 2.3 ms for stack-of-spirals imaging. This figure demonstrates that using Cartesian sampling we need to scan for 14 minutes 54 seconds to achieve the same SNR as achieved by our stack-of-spiral acquisition with a 5.4ms readout in 8 minutes 30 seconds

Video S1

VIDEO S1 All 112 slices through the lungs of four subjects shown in Figure 7. Legend: Healthy volunteer, patient with lung nodules, patient with COVID-19, and patient who recovered from COVID-19

Video S2

VIDEO S2 All slices through the lungs of five patients not shown in figures. Legend: (1,2) patients with COVID-19, (3,4,5) patients who recovered from COVID-19

DATA AVAILABILITY STATEMENT

To support the findings of our paper, we make the reconstruction source code available at https://github.com/NHLBI-MR/selfgated_noncartesian_reconstruction. Some of the data that supports the findings of this study is available upon request, especially to test the reconstruction algorithms.