Rosette Trajectories Enable Ungated, Motion-Robust, Simultaneous Cardiac and Liver T2* Iron Assessment

Abstract

Background:

Quantitative T2* MRI is the standard of care for the assessment of iron overload. However, patient motion corrupts T2* estimates.

Purpose:

To develop and evaluate a motion robust, simultaneous cardiac and liver T2* imaging approach using non-Cartesian, rosette sampling and a model-based reconstruction as compared to clinical-standard, Cartesian MRI.

Study Type:

Prospective

Phantom/Population:

Six ferumoxytol containing phantoms (26 −288 μg/mL). Eight healthy subjects and 18 patients referred for clinically indicated iron overload assessment.

Field Strength/Sequence:

1.5T, 2D Cartesian and rosette gradient echo (GRE)

Assessment:

GRE T2* values were validated in ferumoxytol phantoms. In healthy subjects, test-retest and spatial coefficient of variation (CoV) analysis was performed during three breathing conditions. Cartesian and rosette T2* were compared using correlation and Bland Altman analysis. Images were rated by 3 experienced radiologists on a 5-point scale.

Statistical Test:

Linear regression, ANOVA and paired student t-testing were used to compare reproducibility and variability metrics in Cartesian and rosette scans. Wilcoxon rank test was used to assess reader score comparisons and reader reliability was measured using intraclass correlation analysis.

Results:

Rosette R2* (1/T2*) was linearly correlated to ferumoxytol concentration (r² =1.00) and not significantly different than Cartesian values (p=0.16). During breath-holding, ungated rosette liver and heart T2* had lower spatial CoV (liver: 18.4 ± 9.3% Cartesian, 8.8 ± 3.4% rosette, p=0.02, heart: 37.7 ± 14.3% Cartesian, 13.4 ± 1.7% rosette, p=0.001) and higher quality scores (liver: 3.3 [3.0–3.6] Cartesian, 4.7 [4.1–4.9] rosette, p=0.005, heart: 3.0 [2.3–3] Cartesian, 4.5 [3.8–5.0] rosette, p=0.005) compared to Cartesian values. During free-breathing and failed breath-holding, Cartesian images had very poor to average image quality with significant artifacts whereas rosette remained very good, with minimal artifacts (p =0.001).

Data Conclusion:

Rosette k-sampling with a model-based reconstruction offers a clinically useful motion robust T2* mapping approach for iron quantification.

Introduction

T2* is a clinically useful biomarker for tissue iron quantification^1,2. In vivo, endogenous paramagnetic compounds like hemosiderin and ferritin cause microscopic field inhomogeneity and increase local T2*-related dephasing³; thus quantitative T2* can be used to spatially determine tissue specific iron content.

Iron overload, resulting from excess iron accumulation in hereditary hemochromatosis and transfusion dependent patients, may lead to organ failure and death. Historically, liver biopsy was the gold standard method for body iron quantification and chelation management; however procedural complication risk, sampling errors and low patient tolerance have limited enthusiasm for this method⁴. T2* MRI offers a non-invasive, accurate, repeatable and well-tolerated method to assess iron in various organs, including the liver, heart, kidneys, spleen and pancreas⁵. As a result, iron assessment by MRI has replaced liver biopsy as the clinical standard, guiding patient management and enhancing patient survival; further, it has enabled research into basic iron loading mechanisms⁶. Today, cardiac-iron related mortality has been essentially eliminated in well managed patients⁷.

However, despite the success of MR iron assessment, there are several remaining challenges. First, quantification and image quality of existing T2* techniques are easily corrupted by motion⁴, limiting clinical utility in pediatric patients. Typically, motion artifacts are minimized by breath-holding and/or gating strategies that “freeze” respiratory and cardiac motion. Unfortunately, these strategies are inherently inefficient, prolonging scan times. Particularly in pediatric populations, sedation is common, adding risk and expense⁸. Secondly, comprehensive iron assessment requires both cardiac and liver T2* quantification due to disparate iron loading phenotypes⁹. Although liver iron is an excellent marker of total body iron, cardiac iron is the strongest prognostic marker of mortality and does not correlate well with liver iron^10,11. Furthermore, practical challenges including large scan coverage, receive coil placement and differing motion compensation schemes¹² have prevented simultaneous cardiac/hepatic acquisitions, leading to long clinical exams⁴. Although previous work has sought to address these concerns with the use of novel motion compensation strategies and simultaneous liver and cardiac T2* imaging^13,14, these approaches have yet to become a part of routine clinical iron exams.

One successful motion robust approach entails non-Cartesian data sampling strategies such as radial and spiral, as they yield diffuse aliasing artifacts and their frequent sampling of the center of k-space reduces noise¹⁵. Furthermore, non-Cartesian acquisitions are well suited for compressed sensing and other model-based reconstructions that can further reduce motion artifacts¹⁶. Although non-Cartesian sampling is usually described for single echo scans, non-Cartesian strategies for multi-echo brain fMRI¹⁷ and body fat quantification have been successfully implemented¹⁸. Despite this success, no work to date has explored the utility of non-Cartesian imaging for quantitative body T2* iron assessment, particularly in the heart.

Rosettes are non-Cartesian, flower-like k-space trajectories first described by Noll¹⁹ and categorized by Li et al²⁰. Previous and ongoing in vivo work using rosette sampling has focused on low resolution, spectroscopic imaging due to the off-resonance, deconstructive inference property of rosette trajectory self-crossings²¹. Though the off-resonance sensitivity is advantageous in spectroscopic imaging, in most other applications off-resonance degrades image quality as it leads to both blurring and signal drop-outs, characteristic of rosette imaging¹⁹. Much prior work has focused on off-resonance correction of non-Cartesian trajectories^22,23 and one successful approach involves a segmented k-space reconstruction²³.

Therefore, the goal of this work was to develop and evaluate a motion robust, simultaneous cardiac and liver T2* imaging approach using non-Cartesian, rosette sampling and a model-based reconstruction as compared to clinical-standard, Cartesian MRI.

Materials and Methods

All studies were IRB approved and informed consent and assent were obtained in subjects.

Imaging was performed on a GE Signa 450W MRI system with a 20-channel cardiac coil (GE Healthcare, Waukesha, WI). Cartesian and rosette imaging parameters are found in Table 1. For Cartesian scans, a cardiac-gated, multi-echo gradient echo, (GRE) sequence was used with the 8 echoes, initial echo time 1.1ms, echo spacing of 1.3 ms, and trigger delay of 10ms. Other sequence parameters include flip angle 25 degrees, TR 15.7 ms, 40 cm field of view, 1.5 mm in-plane resolution and 8 mm slice thickness. ECG gating was used in patients whereas peripheral gating was used in healthy volunteer scans. A total of 10–16 views per segment were used to keep the scan time between 15–20 seconds, depending on heart rate.

Table 1.

Imaging parameters used in Cartesian and rosette multi-echo, gradient echo pulse sequences.

Parameter	Cartesian	Rosette
Gating	ECG/PPG Gated	Ungated
Matrix Size	256×256	512×512
FOV (cm)	40	50
Resolution (mm)	1.5	1
Slice thickness (mm)	8	8
Flip Angle (deg)	25	15
Repetition Time (ms)	15.7	18
Echo Times (ms)	1.1, 2.4, 3.7, 5.0, 6.3, 7.6, 8.9, 10.2	0.8, 4.6, 7.6, 10.6, 13.6
Scan Time (s)	15–20	15

For rosette scans, we extended the work of Noll¹⁹ and Li et al²⁰ and defined Class II rosettes as follows:

$$\begin{gathered} \text{if}\text{N}\text{is}\text{odd,}\text{q}=\left\{\frac{\text{N}+2}{\text{N}}+\frac{2(\text{k}-1)}{\text{N}},\text{k}\in {\mathbb{Z}}^{+}\right\},\left\{{\mathbb{Z}}^{+}\right\} \\ \text{if}\text{N}\text{is}\text{even,}\text{q}=\left\{\frac{\text{N}+2}{\text{N}}+\frac{4(\text{k}-1)}{\text{N}},\text{k}\in {\mathbb{Z}}^{+}\right\},\left\{{\mathbb{Z}}^{+}\right\} \end{gathered}$$ 1)

where k is an incrementing parameter, N is the number of petals and q is a shape parameter defined as ω₂/ω₁, where ω₁, ω₂ are rotational frequencies defined in the original rosette formulation¹⁹.

In our approach, a single repetition, or flower, is segmented into multiple echoes, or petals (Figure 1 A, D) defined as the trajectory sampling window centered around each k-space center crossing. Each flower is then rotated by a golden angle (137.5°) (Figure 1 B, E) to ensure adequate sampling and that various motion states are evenly distributed throughout k-space²⁴. By grouping similarly timed petals across shots, a multi-echo data set is achieved (Figure 1F).

Figure 1:

A single, q=2.2 rosette repetition, or flower, where the colors correspond to the effective echo time in k-space (A) and kt-space (D). By rotating the nominal rosette flower by the golden angle (137.5ᵒ), k-space (B) and kt-space (E) are adequately sampled. Separating each rosette flower into individual petals defined by the k-space center crossing allows for a multi-echo data acquisition (F). The corresponding readout gradient and spoiling gradient waveforms are also shown (C).

Imaging Parameters

For the rosette sequence, a 15-degree flip angle, a repetition time of 18 ms, and q value of 2.2 were chosen with a total readout duration of 16ms. The sequence was constrained to a maximum slew rate of 75 mT/m/s and gradient amplitude of 40 mT/m to reduce eddy current and gradient timing related artifacts. Trapezoidal waveforms were used to convert the analytical gradients into physical gradient achievable on the scanner and return to the k-space center at the end of readout for an improved steady state condition. The rosette gradient waveforms are depicted in Figure 1C. Both radiofrequency and gradient spoiling were performed. A total of 800 continuous repetitions were performed, each rotated by 137.5° for a comfortable breath-hold scan time of 15 s. The 800 repetitions were grouped into five unique echo times based on k-space center crossing times, of 0.8, 4.6, 7.6, 10.6 and 13.6 ms. A cardiac self-gated reconstruction with a 300ms window 200ms following peak systole and an ungated time-averaged reconstruction using all temporal cardiac samples were performed²⁵ (Figure 2).

Figure 2:

(A) Representative single echo images at two echo times and T2* maps of the gated Cartesian (TE = 1.3 and 7.6 ms), retrospectively self-gated rosette and time averaged rosette (TE = 0.8 and 7.6 ms) in a mid, short axis slice. Though clear delineation exist between the septum and blood pool, the white arrows show temporal blurring of the papillary muscles in a time-averaged T2* map (B) The rosette trajectory starts from the center of k-space and can be used to create a self-gating magnitude signal that encodes cardiac motion. The open squares correspond to every repetition, red open circles correspond to peak systole and gray filled squares correspond to repetitions used in a self-gated reconstruction. (C) The hepatic (purple) and myocardial septal (red) regions of interests are shown.

Gradient delay correction

Prior to image reconstruction a gradient delay correction was performed on the nominal rosette gradient waveforms to correct for hardware imperfections due to gradient switching⁴. A simple, retrospective, bulk gradient delay time constant was empirically determined for each physical gradient axis²⁶ using a ferumoxytol phantom. An exhaustive search approach was used to determine the ideal physical gradient delay on each gradient axis based on closeness of the Cartesian and rosette T2* values. Patient-specific delay correction was performed by interpolating the nominal gradient waveforms and applying the appropriate patient specific rotation matrix, delaying the physical gradients along x, y and z and rotating back onto the nominal coordinate frame to obtain the delayed k-trajectory.

Image Reconstruction

A parallel imaging reconstruction with locally low-rank regularization was used²⁷. The reconstructed images are produced by iteratively solving the following optimization problem:

$$argmi{n}_x\Vert y-FSx{\Vert }_2^2+\lambda \Vert C(x){\Vert }_{*}$$ 2)

where the reconstructed images, x, were transformed into the raw k-space data (y) using a signal model comprised of coil sensitivity maps (S) and the non-uniform Fourier transform operator (F). Coil sensitivity maps were derived using ESPIRiT (Eigenvalue iTerative Self-consistent Parallel Imaging Reconstruction) from a gridded reconstruction of the first 32 k-space samples of the initial echo²⁸. An empirically determined regularization parameter (λ) of 0.0001 was used. A locally low rank constraint was integrated by minimizing the nuclear norm of a Casorati matrix, C(x), whose columns are comprised of patches from different spatial locations in the image. To reduce the computational complexity of each iteration, raw data was coil compressed using principal component analysis²⁹. All reconstruction code was implemented using the Berkeley Advanced Reconstruction Toolbox (BART)³⁰. Magnitude, voxel-wise mono-exponential fitting was performed using custom Python scripts to calculate spatial T2* maps, and truncation was performed in cases of exceptionally short T2* for improved robustness.³¹

Phantom Imaging

Six phantoms containing distilled water, 2% carrageenan by mass³² and variable concentrations of ferumoxytol³³ (26, 36, 72, 120, 168, 288 μg/mL) were constructed. The T2* phantoms were submerged in a room temperature water bath to limit susceptibility artifacts. Cartesian and rosette multi-echo images were acquired in an axial orientation.

In Vivo Imaging

Reproducibility and motion sensitivity experiments were performed in 8 healthy volunteers (7 male, 26.7 ± 4.2 years) whereas accuracy experiments were performed in the healthy volunteers and 18 additional subjects undergoing clinically indicated T2* iron assessment (17 transfusion induced iron overload, 1 hereditary hemochromatosis, 10 male, 17.7 ± 6.0 years). All in vivo images were acquired in a single slice, mid-ventricle, short-axis orientation.

Image quality was rated by 3 radiologists (*.*, 19 years experience, *,* 8 years experience, *,* 5 years experience) blinded to imaging technique on the following 5 point scale¹²: 1 – Very poor, unusable images; 2 – Average image quality with significant artifacts; 3 – Good image quality with moderate artifacts; 4 – Very good quality with minimal artifacts; 5 – Excellent image quality with no significant artifacts. Image quality assessment of the liver and myocardial septum was performed on all Cartesian and rosette T2* maps. A composite mean and standard deviation score across reviewers was calculated. Manual regions of interest were drawn in the myocardial septum, encompassing epicardial and endocardial borders³⁴ and the dome of the liver, with attention to avoid hepatic vessels (Figure 2).

Reproducibility experiments consisted of three, breath-hold Cartesian and rosette acquisitions separated by less than 20 minutes. Test-retest coefficient of variation (CoV_TRT) was calculated as the mean divided by the standard deviation of the three scans and represented a measure of T2* reproducibility. The spatial coefficient of variation (CoV_s) for each T2* map was also calculated as the mean of voxel measurements in a region of interest divided by the standard deviation and was a measure of spatial uniformity in the T2* maps. The motion related error in T2* was calculated as the difference in free-breathing and incomplete breath-hold T2* from breath-hold T2*.

Motion sensitivity was assessed by comparing free breathing and failed breath-hold T2* scans in the same 8 healthy subjects. During the failed breath-hold, volunteers were instructed to hold their breath for the first 5–10 seconds and then resume, normal tidal breathing for the remainder of the scan. The image quality scores, CoV_s and the absolute, average error between the breathing maneuvers and breath-hold test retest T2* results were calculated as metrics of motion sensitivity.

Statistical Test

The linear correlation r² and limits of agreement of phantom R2* values and ferumoxytol concentration with and without gradient delay correction was measured³³. ANOVAs for T2*, CoV_TRT and CoV_s values were performed across breath-hold Cartesian, gated rosette and time-averaged rosette conditions to measured motion sensitivity. On statistically significant comparisons via ANOVA, multiple student t-tests with Tukey correction were performed. Paired student’s t-test was used to compare Cartesian to time-averaged rosette T2*, T2* error, CoV_TRT and CoV_s values in both free-breathing and incomplete breath-hold conditions. Paired student t-test with Dunnett correction was used to compare breath-hold T2*, CoV_TRT and CoV_s values to free breathing and incomplete breath-hold values in Cartesian and rosette scans. Linear correlation r² and limits of agreement between Cartesian and rosette T2* values were calculated in all breath-hold scans and used as a measure of accuracy. Two- way mixed, consistency intraclass correlation coefficient and the 95% confidence interval was determined for image score reader reliability. Medians and interquartile ranges of the composite mean reader scores were calculated and compared using Wilcoxon rank sum test or Kruskal-Wallis test followed by Steel-Dwass correction for multiple comparisons. A p-value < 0.05 was considered significant. Statistical analysis was performed JMP Pro 14 (SAS, Cary, NC).

Results

Our rosette formalism produced a wide range of echo spacings compared to prior definitions (Supplemental Figure). Higher petal and/or k-values resulted in variable density sampling³⁵ with higher average gradient amplitudes at peripheral, high k-spatial frequencies and lower average gradient amplitudes at central, low k-spatial frequencies.

Phantom Imaging

The carrageen, ferumoxytol phantom produced a stable gelatinous phantom across the clinically relevant range of T2* values. Prior to gradient delay correction, the measured rosette 1/T2* (R2*) was linearly correlated with ferumoxytol concentration up to 288 μg/mL (r²=0.999) but was systematically underestimated relative to the Cartesian T2* values by −13.4± 5.8% (p=0.002). A gradient delay of 2.4 μs in the x and y direction and a 0.6 μs delay in the z direction reduced the observed T2* underestimation to −2.8 ± 4.2% (p=0.16) (Supplemental Figure 2) and was applied retrospectively to all rosette scans. Example images before and after the gradient delay correction are shown in Supplemental Figure 3, demonstrating improved spatial homogeneity of T2* mapping.

In Vivo Imaging

In 8 healthy subjects, the average liver T2* was 24.6 ± 4.2ms, 25.1 ± 3.1ms and 24.0 ± 3.6ms (ANOVA p=0.92) while the average myocardial T2* 34.9 ± 3.3ms, 31.9 ± 2.2ms, and 32.0 ± 2.6ms (ANOVA p=0.09) for the Cartesian, self-gated rosette and time-averaged rosette respectively and were not statistically different.

CoV_TRT, CoV_s and quality scores for all breathing conditions can be found in Table 2. There was no statistical difference in breath-hold T2* CoV_TRT between Cartesian, self-gated and time-averaged measurements in the liver (ANOVA, p = 0.48) nor the myocardial septum (ANOVA, p = 0.51). Between Cartesian, self-gated rosette and time-averaged rosette, there was a statistical difference in liver and heart CoV_s (ANOVA, liver p=0.006, heart p<0.001) and image quality (Kruskal-Wallis, liver p =0.01, cardiac p=0.002). Image quality was rated statistically the highest and CoV_s statistically the lowest in the time-averaged breath-held acquisitions for both the liver and myocardium and details can be found in Table 2. Since the time averaged rosette was non-inferior to gated rosette images in terms of accuracy and demonstrated lower spatial variability, and superior quality scores, only time-averaged rosette reconstructions were analyzed further.

Table 2.

Image variability and quality statistics for gated Cartesian, self-gated rosette and time-averaged rosette during different various breathing conditions.

	Parameter	Cartesian	Self-Gated Rosette	Time-Averaged Rosette
T2* Variability
Liver	CoV TRT Breath-Held	7.8 ± 9.7%	3.9 ± 4.0%	3.6 ± 2.3%
	CoV Spatial Breath-Held	18.4 ± 9.3%	19.9 ± 5.6%	8.8 ± 3.4%**,++
	CoV Spatial Free-Breathing	59.1 ± 73.3%		9.3 ± 3.3%
	CoV Spatial Failed Breath-hold	82.6 ± 78.9%		11.9 ± 5.9%*
	% Error from Free-Breathing	18.4 ± 15.6%		9.9 ± 6.9%
	% Error from Failed Breath-Hold	24.5 ± 28.2%		15.4 ± 10.8%
Cardiac	CoV TRT Breath-Held	4.6 ± 3.1%	4.0 ± 1.7%	3.3 ± 1.2%
	CoV Spatial Breath-Held	37.7 ± 14.3%	21.2 ± 5.6%*	13.4 ± 1.7%**,++
	CoV Spatial Free-Breathing	62.5 ± 24.6%		13.7 ± 3.0%**
	CoV Spatial Failed Breath-hold	82.3 ± 28.0%		14.7 ± 2.4%**
	% Error from Free-Breathing	15.7± 10.2%		3.7 ± 2.3%*
	% Error from Failed Breath-Hold	42.4± 36.8%		7.3 ± 10.6%*
Image Quality
Liver	Breath Held	3.3 [3 – 3.6]	4.2 [3.5 – 4.6]	4.7 [4.1 – 4.9]**
	Free-Breathing	2 [1.7 – 2.3]		4.3 [3.8 – 4.3]**
	Failed Breath-hold	2 [1.3 −2.7]		3.8 [3.0 – 4.2]**
	Inter-reader ICC (%)	91.6 [82.7–96.3]%		90.7 [80.8–95.9]%
Cardiac	Breath-Held	3.0 [2.3 – 3]	4.0 [3.3 – 4.6]*	4.5 [3.8 – 5.0]**
	Free-Breathing	2.0 [1.7 – 2.3]		4.5 [3.6 – 4.7]**
	Failed Breath-hold	1.3 [1.0 – 2.3]		3.8 [2.8 – 4.3]**
	Inter-reader ICC (%)	73.2 [59.1–83.4]%		91.3 [82.7–96.3]%*

^*:statistically different from Cartesian (* p<0.05, **p<0.005).

^+:statistically different from self-gated rosette (+ p<0.05, ++ p<0.005). Intraclass Correlation Coefficient (ICC)

Image quality and variability estimates during free-breathing and failed breath-hold conditions are shown in Figure 3 along with representative T2* maps. In Cartesian T2* maps, the liver CoV_s was statistically larger during free-breathing than breath-hold (p=0.04) and the myocardial CoV_s was statistically larger during free-breathing (p=0.008) and incomplete breath-holding (p=0.002) than breath-hold scans. However, there was no statistical pairwise difference in liver or myocardial CoV_s in free-breathing or incomplete breath-holding compared to breath-hold values (all above p>0.05) (Table 2).

Figure 3:

In two representative subjects, gated Cartesian and time-averaged rosette individual echo (TE=7.6, gray scale) and T2* maps, with image quality scores below each map, are shown under three different breathing conditions: breath-hold, free-breathing and failed breath-holding. Diffuse motion-induced aliasing artifacts in individual echoes (white arrows), produce large data corrupting artifacts in gated Cartesian T2* maps (white arrows). Conversely, rosette reconstructions are largely immune to these artifacts and excellent global image quality and myocardium-blood pool delineation is seen in all images.

Additionally, in the liver, the average error in T2* induced by the breathing motion was not statistically different from breath-holding for free breathing (p=0.20) or for incomplete breath-holding scans (p=0.20). However, in the heart, the average error in T2* from breathing motion was larger in free-breathing Cartesian (p=0.006) and incomplete breath-holding Cartesian (p=0.03) than rosette scans (Table 2).

The median and interquartile ranges of the average image quality scores can be found in Table 2. Across all breathing conditions and anatomical locations rosette T2* map quality was rated better than Cartesian image quality (p < 0.005) (Table 2). In Cartesian acquisitions, T2* map image quality was rated as significantly lower during both free-breathing (liver p=0.021, cardiac p=0.010) and failed breath-holding (liver p=0.002, cardiac p=0.009). Conversely, rosette free-breathing image quality scores were not significantly different from breath-held scans (liver p=0.093, cardiac p=0.67). Although failed breath-holding scores were lower in the liver (liver p=0.036, cardiac p=0.06) they were still rated as “very good” (Table 2). Inter-rater intraclass coefficient reliability in the liver was excellent and statistically similar between Cartesian and rosette rated T2* maps (p = 0.96). However, inter-rater intraclass coefficient reliability in the heart was only acceptable in the Cartesian and remained excellent in rosette rated T2* maps (p=0.01).

Cartesian versus rosette T2* was measured in 26 subjects (8 healthy subjects and 18 patient undergoing clinical iron quantification). Rosette T2* systematically underestimated Cartesian T2* as the average paired difference in rosette and Cartesian liver T2* was −1.3 ± 0.3 ms (p<0.001) and myocardial T2* was −3.5 ± 1.2 ms (p<0.001). Correlation of rosette and Cartesian T2* was excellent in the liver (r²=0.98, p <0.001) and weak but significant in the myocardium (r²=0.33, p=0.006) (Figure 5). Across all subjects, the absolute pair-wise difference in CoV_s was higher in Cartesian compared to rosette images by 6.9± 2.3% (p=0.005) in the liver and by 22.1± 5.1% (p<0.001) in the myocardium.

Figure 5:

Correlation (left) and Bland-Altman (right) plots in the gated Cartesian versus time-averaged rosette T2* quantification in the liver (top) and myocardial septum (bottom) for healthy subjects (open circles) and patients referred for iron assessment (closed circles). Unity lines are shown as the dashed line in correlation plots. Liver results showed high agreement across the entire T2* range, with a mean bias of −5.1% and 95% confidence interval, limits of agreement of 9.3% to −21.9%. The gated Cartesian and time-averaged rosette myocardial septum results were more discordant, with a mean bias of −8.4% and 95% confidence interval, limits of agreement of 21.4% to −32.1%.

Discussion

We introduced a method for T2* quantification using rosette k-space sampling and a model-based reconstruction. This approach produced comparable T2* quantification with superior image quality, higher spatial resolution, fewer motion artifacts and reduced spatial variability without gating in similar scan time as compared to the clinical standard technique. Rosette multi-echo imaging offers several practical advantages over traditional approaches.

Firstly, in equal scan time, rosette T2* imaging had similar accuracy and reproducibility and better base image quality and resolution when compared to traditional methods. Since T2* imaging is used longitudinally to follow chelation efficacy, high reproducibility and accuracy is critical in clinical iron management. Rosette imaging demonstrated comparable precision and accuracy in phantoms. In subjects, the liver and cardiac CoV_TRT trended lower than Cartesian but was not statistically significant and agreed well with previously reported data³⁶. Furthermore, the liver T2* displayed high accuracy across the entire physiologic rate. Though we observed deviations in myocardial T2* measurements relative to Cartesian measurements, our cohort had only one subject with excess myocardial iron, a relatively rare condition. The Bland-Altman agreement in Cartesian and rosette myocardial T2* was comparable to previous reports¹⁴ and the variance is most likely due to the high variability of the clinical standard, Cartesian T2* technique in the healthy T2* range.

Traditional bright blood T2* mapping is known to be variable in healthy subjects³⁷. Furthermore, we measured higher spatial variability, lower reader scores and observed poor myocardium-blood pool differentiation in the Cartesian T2* than rosette T2* maps of this study. The improvements in image quality are clinically important as poor image quality reduces clinical confidence in the quantification, resulting in frequent repeat scans and even sedation. Furthermore, although variability in regional iron deposition has been observed in excised hearts, low spatial resolution, high spatial variability and lower reader confidence in Cartesian T2* maps has made it difficult to examine in vivo³⁸. Therefore, precise, accurate and high-resolution rosette T2* mapping is a promising and powerful clinical and research tool.

Another advantage of rosette T2* imaging is high spatio-temporal reproducibility without cardiac gating. The ungated, time-averaged rosette T2* maps displayed higher image quality and comparable test-retest reproducibility to a cardiac gated rosette, cardiac gated Cartesian and literature T2* maps in both the liver and myocardium³¹. By leveraging the additional data from a continuous acquisition across all cardiac phases, the incoherent aliasing artifacts of non-Cartesian rosette acquisitions and the denoising properties of an iterative locally low rank reconstruction, cardiac gating was unnecessary for the generation of reliable and accurate T2* maps, with superior image quality to the clinical standard approach. In the clinical setting, ungated acquisitions are preferable because patient motion, ECG triggering failure and variable and/or high heart rates often confound imaging protocols and reduce diagnostic confidence.

Finally, respiratory gating was not needed for high image quality and reader confidence in liver and myocardial T2* estimation. Cartesian T2* maps were markedly corrupted by respiratory motion; however, rosette T2* imaging was largely unaffected. Clear motion-induced ghosting can be seen in source Cartesian images at various echoes times, which contributes to large errors in T2* estimates, poor map quality and higher spatial variability in both the liver and the heart. On the other hand, rosette images were largely unaffected by respiratory motion which translated to minimal errors in T2*, negligible loss of map quality and no increase in spatial variability for both the heart and liver. Clinically, failed and incomplete breath-holding are commonly encountered in pediatric patients resulting in poor T2* assessment, sedation, repeat scan and even repeats visits. Therefore, ungated motion robust rosette T2* assessments are not only faster and more reliable but also improve patient tolerance and safety.

This work is an extension of much prior work involving rosette trajectories^19,20, non-Cartesian multi-echo imaging^17,18 and motion robust T2* mapping approaches^12,14. In particular, previous quantitative, motion robust T2* mapping strategies have used single acquisitions and non-rigid motion correction to achieve improved image quality in the presence of motion^12,14. Though successful, our approach has the additional benefit of requiring no patient specific motion correction, being completely ungated and having identical scan time to the clinical standard while achieving higher spatial resolution.

Limitations

Despite the major improvements of ungated, rosette T2* images over the conventional approach, this work is not without limitations. First, our study was conducted in a limited number of subjects which increases the likelihood of type II statistical error, especially with respect to the variability and motion sensitivity parameters. Despite this limitation, we conclusively demonstrate that rosette T2* maps produced quantitatively similar reproducibility and accuracy, lower variability and higher quality images. Next, non-Cartesian sampling trajectories are more susceptible to gradient timing related imperfections that can influence image quality and quantitation. Although we address some of these effects by tempering the slew rate and applying a simple retrospective gradient delay correction, patient specific and prospective methods should be explored in the future¹⁸. Gradient timing errors manifest as both magnitude and phase errors which can lead to T2* estimation errors³⁹. Many strategies have been proposed to correct for gradient timing artifacts including direct gradient measurements^40,41 and retrospective algorithms based on measurements made with custom pulse sequences⁴². For instance, previous non-Cartesian multi-echo imaging has used a prescan to measure gradient delays prior to each scan and demonstrated that gradient timing errors can change even on an individual subject basis¹⁸. Another limitation is that rosette T2* maps are more sensitive to off-resonance related artifacts than typical Cartesian maps which can result in noticeable loss of signal and T2* underestimation. This can be seen most readily in the mid-ventricular free wall due to air tissue susceptibility artifacts from the lung. Though susceptibility artifacts are negligible in the septum, care should be taken when performing regional segmentation and global myocardial analysis. Further, although our ungated approach produced images of high quality, reproducibility and robustness, future efforts should explore incorporating motion directly into the reconstruction model⁴³. We postulate that this would allow T2* to be resolved dynamically across motion states and may prove useful in functional MRI related physiological studies. Lastly, magnitude based T2* measurements are confounded by intravoxel fat and more sophisticated mapping algorithms should be explored in the future⁴⁴.

Conclusion

This work introduces a clinically useful approach for motion-robust iron quantification using rosette k-sampling and a model-based reconstruction that produces better quality images and more robust quantification than conventional methods.

Supplementary Material

Supplemental Fig 1

Supplemental Figure 1: Our rosette formalism in Equation 1 categorizes a large number of class II trajectories. Echo spacing and sampling density depends on the shape parameter, q and the chosen slew rate and max gradient amplitude constraint.

Supplemental Fig 2

Supplemental Figure 2: Prior to gradient delay correction, the measured rosette 1/T2* (i.e R2*) was linearly correlated with ferumoxytol concentration up to 288 μg/mL (r²=0.999) but was systematically underestimated relative to the Cartesian T2* values by −13.4± 5.8% (p=0.002). Correlation and Bland Altman plots demonstrated excellent agreement in phantoms following delay correction and rosette T2* values no longer underestimated Cartesian T2* values, −2.8 ± 4.2% (p=0.16).

Supplemental Fig 3

Supplemental Figure 3: Example of echo images and T2* map before and after gradient delay correction. Following gradient delay correction there are noticeable improvements in the spatial homogeneity in the T2* map (white arrows).

Figure 4:

Representative images from four of the 18 patients undergoing, clinical iron examination with gated Cartesian (top row) and time-averaged rosette (bottom row) echo images (TE =7.6 ms, grayscale) and T2* maps (colormaps) are shown with corresponding image quality scores below.