Increasing the scan‐efficiency of pulmonary imaging at 0.55 T using iterative concomitant field and motion‐corrected reconstruction

To develop an iterative concomitant field and motion corrected (iCoMoCo) reconstruction for isotropic high‐resolution UTE pulmonary imaging at 0.55 T.


INTRODUCTION
Contemporary mid-field MRI systems have demonstrated high-quality lung imaging. 1,2These 0.55 T systems offer improved B 0 field homogeneity, longer T * 2 relaxation times, and reduced susceptibility gradients near air-tissue interfaces. 3In the lung, T * 2 at 0.55 T is ∼10 ms 1,4 compared to 1 to 3 ms at clinical field strengths of 1.5 and 3 T. Longer T * 2 relaxation times imply slower signal decay and can enable improved sampling efficiency (e.g., using spiral readouts) for increased SNR and longer TE to lower hardware and pulse-sequence design constraints.
We recently demonstrated a method for high-resolution (1.75 mm isotropic) pulmonary imaging on a prototype 0.55 T scanner using a self-gated stack-of-spirals UTE sequence, with an 8.5-min acquisition. 5We showed the necessity for concomitant field correction at 0.55 T for high-resolution spiral imaging and presented a robust online reconstruction that included trajectory and concomitant field corrections.Our low-latency image reconstruction required <5 min to produce high-quality pulmonary images. 5However, our method had low scan efficiency because it used only 40% of the acquired data, binned from a stable respiratory phase.
Recently, Zhu et al 6 demonstrated an iterative motion-corrected (iMoCo) constrained reconstruction for pulmonary imaging using 3D radial trajectories.This method used 100% of the acquired data to reconstruct a high-quality pulmonary image at a single respiratory phase.The iMoCo reconstruction could significantly improve the scan efficiency of our previous stable-binned stack-of-spirals method by enabling shorter scan times or providing significant improvements in SNR.However, the iMoCo reconstruction does not correct for concomitant fields, which cause blurring and signal shading, and are especially challenging at lower field strengths coupled with high gradient amplitudes. 1,4,5inally, the iMoCo technique used long reconstruction times that do not provide images "online" (i.e., during the patient's exam), which limits adoption in clinical settings.
In this work, we present two improvements to iMoCo: (1) we incorporate concomitant field corrections to enable the application of this reconstruction at lower field strengths; and (2) we optimize the reconstruction on graphics processing units (GPUs) to enable low-latency online reconstruction during the scan session.We demonstrate our iterative concomitant field and motion corrected (iCoMoCo) reconstruction for a prototype 0.55 T high gradient performance scanner and a commercial 0.55 T scanner, with lower gradient specifications, in both patients and healthy volunteers.

Image acquisition
This study was approved by our local Institutional Review Board and informed written consent was obtained from all subjects (ClinicalTrails.govidentifier NCT03331380).A total of 12 human subjects were scanned for this study, eight healthy volunteers, one patient with an active coronavirus disease 2019 (COVID-19) infection, one patient with lung nodules, one patient with chronic granulomatous disease (CGD), and one patient with lymphangioleiomyomatosis (LAM).Subjects were scanned on two systems for this study: a prototype 0.55 T system (MAGNETOM Aera, Siemens Healthcare) with gradients peak slew rate: 180 mT/m/s, and max amplitude (Gmax): 45 mT/m, and a commercial 0.55 T (MAGNETOM Free.Max, Siemens Healthcare) with gradient peak slew rate: 40 mT/m/s and Gmax: 25 mT/m.All 12 subjects were scanned on the prototype Aera and three of the healthy volunteers were scanned on the commercial Free.Max for comparison on the same day.Data were acquired on the Aera using three gradient settings (Gmax: 6.2 mT/m, Gmax: 16.2 mT/m, and Gmax: 28.9 mT/m), adjusted during sequence programming, to compare the images for different gradient performance.Data were acquired on the Free.Max using a Gmax of 6.2 mT/m.Subjects were scanned using a research application of self-gated stack-of-spirals 3D UTE spoiled gradient echo sequence with the following imaging parameters: TE/TR = 0.5/7.7 ms, flip angle (FA) = 5 • , readout duration = 5.4 ms, FOV = 450 × 450 × 224 mm 3 , matrix size = 256 × 256 × 112 (1.75 mm isotropic resolution), slice-oversampling = 14.3%, total encoded matrix size = 256 × 256 × 128, total acquisition time = 8.5 min, coronal acquisition.To determine the minimum required acquisition time, we retrospectively clipped the data from the end of the scan to simulate reduced scan times (2, 4, 6, and 8 min).The spiral trajectories were designed for each gradient configuration using 78 shots (Gmax: 6.2 mT/m, Free.Max, and Aera), 33 shots (Gmax: 16.2 mT/m, Aera) or 19 shots (Gmax: 28.9 mT/m, Aera).The total number of spiral rotations was 444 for the 8.5-min acquisition and was the same for all gradient settings.All scans were performed with slice encoding as the inner-loop (i.e., all slice-encodes were acquired for one spiral interleave before the next spiral rotation).

iCoMoCo image reconstruction
Superior-inferior (SI) navigator readouts were acquired approximately every 200 ms to estimate the respiratory signal.Five SI navigators are acquired for every stack of slice-encodes.The number of navigators per stack determines the temporal resolution of the navigator signal and it can be tailored for pediatric or tachypneic patients.Respiratory signal is extracted from the navigator readouts using previously published methods. 5Briefly, the navigator readouts were sorted based on the rotation angle of the spiral interleave and a high-pass Kaiser-window filter was used to filter trajectory-related signal fluctuations.A Kaiser-window band-pass filter was used to isolate respiratory frequencies (0.1-0.7 Hz).Singular value decomposition was used to estimate the lung motion because of respiration by selecting the largest singular value for each coil channel.Singular value decomposition was used to estimate the eigenvectors of the filtered data and the eigenvector with the largest singular value was selected for each coil-channel.Coil clustering was used to select channels that best represent the respiratory motion. 7This step was implemented on GPU to minimize the computation time required to extract the respiratory signal from navigator readouts.
Equal amounts of acquired data were binned into different respiratory motion states.Binning was performed based on respiratory position and direction of motion in the respiratory cycle (i.e., during inspiration vs. during expiration) and determined from the respiratory signal as illustrated in Figure 1.The respiratory motion resolved reconstruction was performed on this binned data.We divided the data into 12 motion states to minimize motion-related blurring. 6The number of bins was heuristically determined using data from five subjects.Unlike the original iMoCo implementation, where the respiratory resolved reconstruction was performed at coarser-resolution, we performed a full-resolution respiratory motion resolved reconstruction.The respiratory-resolved reconstruction was performed with spatial and temporal total variation constraints ( s = 0.05,  t = 0.05).The hyperparameters for the reconstruction were chosen based on parameter sweeps shown in Video S1 and Figure S1.To improve reconstruction speed, we reconstructed images for each respiratory bin on a separate GPU to parallelize the reconstruction task.The respiratory motion-resolved images were registered to a reference image (typically end-expiration) to estimate the deformation fields (M k ) for the final reconstruction.A 3D optical flow-based registration method was used for image registration 8-10 using a coarse-to-fine multi-resolution approach, which iterated the optical flow algorithm on an image pyramid with up to four levels up to the original image resolution.
The iterative motion and concomitant field-corrected reconstruction were performed to estimate a single respiratory phase using 100% of the acquired data.Concomitant field correction and deformation fields for respiratory motion states were incorporated into one cost function for image reconstruction.Concomitant field correction used a multi-frequency interpolation (MFI) 11 ; MFI weights and demodulation frequencies were calculated using previously published methods. 12 where d nk is motion-resolved multi-channel data, X is the end-expiration high-resolution image to be reconstructed, M k is estimated motion field, S n is coil-sensitivity of the nth channel, C p is the MFI weight for concomitant field correction, F is the non-uniform Fourier transform, D p is the demodulation operator at a given frequency for MFI, W k is the density compensation weight for each motion state, and TV s is the spatial total variation term ( s = 0.05).N c represents total number of channels, K is the total number of respiratory states, and P is the total number of demodulation frequencies.The optimization problem is solved by using the split Bregman method. 13Each respiratory state is reconstructed on a separate GPU to minimize reconstruction time, like the respiratory-resolved reconstruction.If the number of GPUs (N) is less than the motion states (B) then the reconstruction parallelizes over the GPUs such that each GPU reconstructs B/N motion states with any remainder going to the first few GPUs.
For comparison, data were also reconstructed using our previously proposed concomitant field-corrected conjugate gradient sensitivity encoding (CG-SENSE) reconstruction using 40% of the data binned to one stable respiratory phase. 5Spatial TV ( s = 0.05) was added for denoising to make fair head-to-head comparisons between CG-SENSE and iCoMoCo.

Low-latency online reconstruction implementation
Low-latency inline image reconstruction was implemented using Gadgetron 14 and deployed on a system equipped with Dual AMD EPYC processors (64 cores each), 1 TB RAM, 4× Nvidia A100 80 GB GPUs.The trajectories and/or gradient waveforms were streamed along with the data from the scanner 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 because of hardware imperfections, which can cause image artifacts. 15Inaccuracies The difference in respiratory binning based on just respiratory position versus respiratory position and direction in the respiratory cycle (i.e., during inspiration and during expiration).(A) Shows a representative respiratory signal extracted from the superior-inferior navigator signal.(B-left) Sorted respiratory signal based on respiratory position from lowest (end-expiration) to highest (end-inspiration).(B-right) The respiratory signal is sorted based on respiratory position and respiration direction.To account for the direction of respiration, the respiratory waveform is multiplied by its gradient before sorting.This sorted signal is then divided into N equal-sized respiratory bins to ensure each bin has approximately the same amount of data as represented by the dotted lines on the plot.The green + yellow points and black + orange points in (A) correspond to the same position but different directions in the respiratory cycle.These translate to the same bin, which are points next to each other when sorted based on position (B-left) and to separate bins when sorted based on position and direction of respiration (B-right).(C) The images reconstructed using the same number of bins demonstrate the improvement in sharpness of the lung nodule (red arrow), in the image reconstructed using positional and directional binning.AU; arbitary units.

F I G U R E 2
Flowchart of image reconstruction.To generate an image (X) data from each channel (n) and motion state (k) is concomitant field corrected using multi-frequency interpolation (MFI).In the MFI operator, data is demodulated at P frequencies and non-uniform Fourier transformed to produce P images, which are multiplied by MFI weights and summed to produce a concomitant field corrected image.These images go through coil combination and are deformed with a graphics processing units (GPU) b-spline interpolation using the kth motion field and averaged to produce image X.This process is reversed in the adjoint operation for the forward pass.*Represents conjugation.iCoMoCo , iterative concomitant field and motion corrected; GPUs, graphical processing units; Im, images; Δf , demodulation frequency.in spiral trajectories were corrected using gradient system impulse response functions. 15,16Density compensation weights were estimated using an iterative method that helps minimize artifacts because of non-uniform sampling in k-space. 17Coil compression was used to conserve GPU memory and to improve reconstruction speed.Data were compressed down to eight channels for both systems; from 18 channels for the Aera and from 15 channels for the Free.Max.The number of compressed channels was heuristically determined to minimize degradation in image quality because of coil compression.Coil sensitivity maps were estimated from non-binned data using the Walsh method. 18ur goal was to shorten the reconstruction time and limit the memory needs for clinical workflow adoption, and therefore, we optimized both each individual reconstruction step as well as the buffering of the data and task scheduling of the reconstruction workflow.For instance, gradient impulse response function trajectory corrections were applied to trajectories line-by-line during image acquisition as the data accumulated into the reconstruction buffer to reduce reconstruction time.
Second, iterative density compensation, coil sensitivity map estimation, extraction of the respiratory signal from superior-inferior navigators, split-Bregman solver, and image warping based on deformation fields were all implemented and performed on GPUs.Image registration used a fast implementation, but was done on the central processing unit.

Image comparisons
Image quality was quantitatively and qualitatively assessed using the following comparisons: 1. Apparent signal-to-noise ratio (aSNR) were calculated in five healthy volunteers for the 2-, 4-, 6-, and 8-min simulated scan times for the iCoMoCo and CG-SENSE reconstruction methods.aSNR was measured as the ratio of mean signal in a region of interest (lung parenchyma) divided by SD of noise-only region (the airway).
2. iMoCo and iCoMoCo images were compared to demonstrate the value of concomitant field correction at 0.55 T using a gradient setting with peak gradient amplitudes of 6.2, 16.2, and 28.9 mT/m.
3. Image sharpness was assessed in five healthy volunteers using relative maximum derivative, 5,6 which is calculated as the ratio of the maximum change between the liver-lung interface divided by the maximum intensity in the liver.A higher relative maximum derivative signifies sharper diaphragm boundary.
4. Images from three volunteers who were scanned on the same day on both the prototype Aera and commercial Free.Max scanners were compared in terms of aSNR and visible undersampling artifacts.This comparison was done using data acquired on the Free.Max and Aera with a Gmax of 6.2 mT/m.Overall, the diaphragm sharpness was comparable between the two methods, as measured by the maximum derivative (Figure 4B), and vessel sharpness was comparable based on qualitative visual assessment.The sharpness of iCoMoCo was comparable to CG-SENSE with improved visualization of certain features because of improvements in SNR. Figure 5 shows the comparison in image sharpness between non-binned free-breathing images, CG-SENSE, and iCoMoCo images, for a scan time of 6, 8, and 6 min, respectively.This highlights the robustness of respiratory binning and motion correction used for iCoMoCo.

Concomitant field correction
Figure 6 demonstrates the concomitant field correction at 0.55 T and the dependence of blurring artifact severity on peak gradient amplitude.This figure shows that for the commercial Free.Max with peak gradient amplitude of 6.2 mT/m there is almost no visible blurring because of concomitant fields and correction does not offer significant improvements in image quality.However, using peak gradient amplitudes of 16.2 and 28.9 mT/m concomitant field correction with iCoMoCo provides visible improvement in image quality especially toward the apex of the lungs, as indicated by the red arrows.

F I G U R E 3
Image quality comparison between concomitant field corrected CG-SENSE reconstruction and iterative concomitant field and motion corrected (iCoMoCo) for scan times of 2, 4, 6, and 8 min.Representative images from one healthy volunteer are shown in this figure .A representative single slice is shown in the top two rows, whereas the bottom two rows show a maximum intensity projection over 15 slices.Respiratory binning and image reconstruction was repeated for each scan time.Green arrows point to vessels in the lungs, which are more visible in images reconstructed using iCoMoCo with scan times >4 min.These vessels are invisible with CG-SENSE when scan time is <8 min.

F I G U R E 4
Quantitative assessment of image quality for CG-SENSE and iterative concomitant field and motion corrected (iCoMoCo).(A) Apparent signal-to-noise ratio (aSNR) measured in five healthy volunteers for CG-SENSE (blue) and iCoMoCo (orange) for acquisition times of 2, 4, 6, and 8 min.Error bars represent SD across the five healthy volunteers.(B) Maximum derivative measured at the diaphragm boundary in a central slice for five healthy volunteers with the 8-min acquisition for CG-SENSE and 6 min acquisition for iCoMoCo.A higher maximum derivative indicates a sharper diaphragm boundary.

Image reconstruction time
For a 6-min acquisition, the iCoMoCo reconstruction took <13 min for data acquired on the prototype Aera with Gmax of 28.9 mT/m, < 5 min for data acquired on the prototype Aera with Gmax of 16.2 mT/m, and <3.5 min for the data acquired on the Free.Max with Gmax of 6.2 mT/m.The difference in reconstruction time is attributed to the number of frequency bins required for concomitant field correction (2, 8 and 21 for 6.2, 16.2, 28.9 mT/m).Table 1 shows the data sizes and reconstruction times for each simulated scan time, the reconstruction times have been subdivided to show the time required for each reconstruction step.For comparison, CG-SENSE reconstruction with TV regularization took <90 s.The image reconstruction code is open-source and available at https://github.com/NHLBI-MR/icomocoand a docker image, which can be used to directly run the code without compiling from scratch, is available via dockerhub at: https://hub.docker.com/r/gadgetronnhlbi/ubuntu_2004_cuda117_public_icomoco/tags along with a test data set which can be found on zenodo: Image quality comparison between non-binned, respiratory binned CG-SENSE, and iterative concomitant field and motion corrected (iCoMoCo) image reconstructions in healthy-volunteers.Free-breathing and iCoMoCo image reconstructions used 6 min of acquired data, whereas CG-SENSE images were reconstructed using 8 min of data.Red-arrows highlights areas of visual improvement in image sharpness in both CG-SENSE and iCoMoCo image reconstruction compared to Free-breathing reconstructions with no motion compensation.Overall image sharpness is comparable between CG-SENSE and iCoMoCo.Green-arrows show vessels that are better delineated with iCoMoCo compared to CG-SENSE.https://doi.org/10.5281/zenodo.10456573.Instructions for how to run the code can be found on the github repository.

Patient imaging
Figure 7 shows representative examples of single-slice images from each patient group reconstructed from the clipped 6 min acquisition at 1.75 mm isotropic resolution.
The figure highlights mosaic attenuation in the patient with CGD, air-filled cysts in the patient with LAM, a patient with a large lung nodule, and ground glass opacities in the patient with active COVID-19 infection.Images are shown in all three orientations (axial, sagittal, and coronal) to highlight the benefits of isotropic imaging.For comparison, Figure S2 shows images in the same patients reconstructed using CG-SENSE.The improved SNR and visualization of some pathologies with iCoMoCo reconstruction can be easily appreciated for scan times of 6 min.

Free.Max versus Aera
Figure 8 shows the image quality comparison between the commercial Free.Max and the prototype Aera 0.55 T systems.The image quality on the Free.Max is comparable to the ramped-down 0.55 T Aera.Parenchymal aSNR was also similar for the two scanners 7.55 ± 0.73 and 6.23 ± 0.43 (n = 3 healthy volunteers) for Free.Max and Aera, respectively.We observed undersampling artifacts in one of the three volunteers for Free.Max and Aera, which may suggest the need for slightly longer scan times with Gmax of 6.2 mT/m to counteract reduced k-space coverage.The trajectory design for Gmax = 6.2 mT/m required 78 evenly sampled shots to be fully sampled, but only ≤26 spiral arms were present in each bin, which could result in undersampling artifacts depending on the distribution of the arms acquired using golden angle view ordering.

DISCUSSION
In this work, we extended the iMoCo image reconstruction 6 to include concomitant field correction, called iCo-MoCo, and we reimplemented the reconstruction to significantly increase reconstruction speed.This image reconstruction technique uses 100% of the acquired data to generate high-quality images with a shorter scan time than our previously published method that used a CG-SENSE reconstruction.The reconstruction pipeline was implemented in Gadgetron and deployed inline on the MRI system such that images returned to the scanner within 5 min 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 to achieve sufficient SNR and longer reconstruction times.We chose to use a 6-min acquisition in this work because it provided robust image quality in all the subjects.We observed undersampling artifacts in two of the 12 subjects when using the 4-min acquisition.These artifacts can potentially be resolved by improving the reconstruction.The 4-min acquisition still provides adequate image Comparison of image quality between iterative motion corrected (iMoCo) and iterative concomitant field and motion corrected (iCoMoCo) (i.e., with and without concomitant field correction) for data acquired with different peak gradient amplitudes.Single slice images (left) and maximum intensity projection across 10 slices (right) are shown.(Top) Data acquired on the commercial 0.55 T system (MAGNETOM Free.Max) with a peak gradient amplitude of 6.2 mT/m, (middle) data acquired on the prototype 0.55 T system (MAGNETOM Aera) with a peak gradient amplitude of 16.2 mT/m, and (bottom) data acquired on the Aera with peak gradient amplitude of 28.9 mT/m.Red arrows highlight features that were blurred because of concomitant fields and are significantly improved after concomitant field correction.The images in this figure are from three different volunteers.Gmax, maximum gradient amplitude.
quality for clinical applications and may even be preferable for patient comfort in clinical settings.We also did not observe the square root of 2.5 improvement in aSNR between CG-SENSE and iCoMoCo, which is likely because of the variation in denoising effect of the spatial and temporal total variation.We added concomitant field correction to the image reconstruction, which is important for high-resolution imaging at lower fields with high peak gradient amplitudes.However, at lower peak gradient amplitudes, the need for concomitant field correction is reduced and the problem collapses to an iterative motion-corrected reconstruction.The reconstruction speeds are significantly faster with no concomitant field correction and otherwise inversely proportional to the number of frequency bins required for concomitant field correction.The iCo-MoCo implementation is generic and will work for 3D radial imaging and other systems and field strengths as well.Additionally, we found that using a peak gradient amplitude >16.2 mT/m induced significant gradient heating, which causes B 0 frequency drift.B 0 frequency drift causes image blurring and requires additional corrections.

T A B L E 1
Reconstruction time and aSNR versus acquisition time.We did not perform this correction in this work, but it can be explored in future work.

Reconstruction time aSNR
In this work, we binned our data based on respiratory position and the direction of respiration (i.e., inspiration and expiration are treated separately).In our comparisons, we found this to produce sharper images compared to using just respiratory position.Previous studies used 6 to 8 respiratory positions, 6 but using our respiratory binning method, we divided the data into 2× more bins, to keep the same positional variation in each bin.Increasing the number of bins, increases the undersampling factor of the respiratory resolved reconstruction.

F I G U R E 8
Image quality on data acquired using the commercial Free.Max 0.55 T scanner and the prototype Aera 0.55 T scanner in three healthy volunteers.Data on the both systems were acquired using gradient amplitudes of 6.2 mT/m.Images were reconstructed using 6 min of acquired data.Green box highlights undersampling artifacts visible likely because of the lower gradient performance.
We heuristically determined that 12 bins produced sharp images and provided enough data in each bin for consistent reconstruction performance.Data could be divided into more bins, but this would likely require a lower-resolution respiratory resolved reconstruction and then upsampling the deformation fields to match the resolution of the final iCoMoCo reconstruction, like the original iMoCo paper.The reconstruction can produce an iCoMoCo image at any respiratory bin 1 to 12.In our experience, end-inspiratory and end-expiratory bins produce the sharpest images.We chose end-expiration for figures in this work to enable comparisons with CG-SENSE reconstruction, which used 40% of data binned to the most stable respiratory phase, which is end-expiration.
Unlike the original iMoCo paper, we were able to use a full resolution for the intermediate respiratory resolved reconstruction because of the higher sampling efficiency of spiral trajectories.However, the trade-off between image SNR and image registration quality was not evaluated in this work, but could be relevant for future work.Overall, we found this approach produced images with comparable sharpness to binned CG-SENSE reconstructions.
The reconstruction method presented in this work was optimized for speed and we used the split-Bregman solver because it is a fast solver for problems involving total-variation regularization.There may be potential opportunities to use a better solver that can perform better in cases with higher undersampling, especially using trajectories with lower sampling efficiency than spiral trajectories.However, for lower-field gradient-echo images, we still recommend using spiral imaging to improve sampling and scan efficiency to maximize SNR.We used coil compression and found that using eight virtual channels for both prototype Aera and Free.Max did not result in a visual loss in image quality.However, coils with more elements and different geometries would change the optimal number of virtual channels.Increasing the number of channels can influence reconstruction speed, and memory usage and may require further memory optimization to enable fast reconstruction.
Our work also had a few limitations.First, we did not detect bulk motion and did not try to reject or correct the data with bulk motion.Bulk motion can significantly degrade image quality if the patient moves during the scan, especially because these scans are still several minutes long.In future work, we will explore image-based navigators or pilot tone-based methods to detect bulk motion.The data with bulk motion either can then be rejected or corrected using additional motion correction-based reconstructions.Second, residual blurring because of off-resonance was not corrected with our existing method.We limited our spiral readout duration to ∼5 ms in the lung because of off-resonance blurring.][21] Third, to improve the reconstruction speeds in this work we used expensive hardware (i.e., GPUs with 80 GB of memory and processors with 128+ cores).As a rule, relative to the data size after coil compression, we needed ∼5× the memory capacity on the primary GPU and about 1× the memory on the GPUs used for parallelization.This allowed us to demonstrate online reconstructions with the potential for integration into clinical workflow.However, the availability of such hardware is limited.A cloud-based implementation could overcome the hardware inaccessibility and could be tuned to trade-off reconstruction speed for cost.Finally, our resolution was 1.75 mm 3 , which is lower than recent publications that demonstrate up to 0.9 mm 3 with a scan time of ∼8.5 min using the 3D radial bSSFP technique bSTAR. 2

CONCLUSION
In this work, we presented an iterative and motion-corrected image reconstruction that includes concomitant field correction for efficient pulmonary imaging on mid-field 0.55 T scanners.We demonstrated this technique, which uses ∼100% of the acquired data and implemented a low-latency (<5 min) reconstructions to provide diagnostic-quality images integrated into clinical workflows.We also showed that it could be used to reconstruct high-quality images on a commercially available 0.55 T scanner with lower gradient specifications.We validated and evaluated our technique in healthy volunteers and patients with known lung disease.

Figure 2 .
shows the forward and backward model of the reconstruction used to solve the following cost function,

F I G U R E 7
Representative image quality for 6 min clipped scan duration in cases of patients with (A) chronic granulomatous disease (CGD), (B) lymphangioleiomyomatosis (LAM), (C) lung nodules, and (D) ground glass opacities.Coronal images, axial images, and sagittal images are shown for each patient.Mosaic attenuation (green box) and air trapping (red box) are shown for the CGD patient, some air-filled cysts are shown for the LAM patient (red arrows), and two lung nodules are shown in the patient with lung nodules (green box).Ground glass opacities are shown in the patient with active coronavirus disease 2019 infection in both the right and left lungs (green box).
Figure3shows representative images using iCoMoCo reconstruction with 100% of the data, compared to CG-SENSE reconstruction with only 40% of the data from a stable respiratory phase, each for scan times of 2, 4, 6, and 8 min.iCoMoCo provides visibly higher SNR, with no evidence of respiratory motion blurring.iCoMoCo reconstruction with a simulated scan time of 4 to 6 min does not produce significant degradation in image quality (i.e., significant loss in signal or reduced visibility of anatomy), whereas with CG-SENSE reconstruction, significant noise amplification occurs for scan times <8 min.
Times are reported for data acquired using Gmax of 16.2 and 28.9 mT/m.Times for the two Gmax 16.2 and 28.9 mT/m only differed at the final stage of the reconstruction (iCoMoCo) and are reported as time 16.2 mT/m |time 28.9 mT/m.Values are reported for reconstruction times rounded to the nearest 1 s.aSNRs are reported as mean ± SD across five healthy volunteers.Data size is reported before coil compression.Abbreviations: aSNR, apparent signal-to-noise ratio; iCoMoCo, iterative concomitant field and motion corrected; DCF, density compensation function; CG-SENSE, conjugate gradient senstivity encoding; Resp, respiratory; GB, gigabyte. Note: