Feasibility Study for Dual Phase Free Breathing 3D bSSFP MRA With Preserved Eddy‐Current Tolerance With Adaptive k‐Space Lengths (PETAL) Acquisition

ABSTRACT

Purpose

To demonstrate the feasibility of a novel dual‐phase self‐navigated, free‐breathing balanced steady‐state free‐precession whole‐heart MRA acquisition for aortic annulus size measurement using a Cartesian trajectory that mitigates eddy‐current artifacts by minimizing k‐space jumps between consecutive readouts with predictable scan times.

Methods

The proposed method uses a trajectory with Preserved Eddy‐current Tolerance with Adaptive k‐space Lengths (PETAL) that enables the acquisition of multiple navigators per heartbeat to address respiratory motion throughout a 330–510 ms acquisition window across different cardiac phases. Imaging was performed in 15 healthy volunteers on a 3 T MR system using a sequence compiled in Pulseq. Images were reconstructed in systole and diastole. Aortic annulus perimeter and area were measured for both cardiac phases. A clinical MRA and a breath‐held 2D bSSFP cine planned at the valve plane were used to statistically compare metrics in diastole and systole, respectively.

Results

In diastole, the PETAL sequence yielded a high correlation with the clinical reference for aortic annulus area ($\mathrm{ICC}=0.987$) and perimeter ($\mathrm{ICC}=0.996$) and low bias (0.008 cm² and −0.018 cm, respectively), with no statistically significant difference. In systole, the proposed method presented a high correlation with the reference for aortic annulus area ($\mathrm{ICC}=0.989$) and perimeter ($\mathrm{ICC}=0.986$), bias of 0.02 cm² and 0.073 cm, respectively, no statistically significant difference for aortic annulus area and a statistically significant difference for perimeter ($p=0.029$).

Conclusion

The proposed approach enables dual‐phase, contrast‐free, free‐breathing acquisition of isotropic 3D whole‐heart images in predictable 10 min scan times for aortic size measurements in systole and diastole.

1. Introduction

Cardiovascular magnetic resonance imaging (CMR) has emerged as a clinically important technique for the assessment of cardiovascular disease (CVD), a leading cause of morbidity and mortality worldwide [1]. For specific applications, such as transcatheter aortic valve replacement (TAVR) planning, precise measurement of the aortic annulus (AA) is essential to ensure accurate prosthesis sizing and avoid complications like valve dislocation or annular rupture [2]. Traditionally, computed tomography (CT) has been the gold standard for acquiring 3D aortic annulus measurements [2, 3]. However, CT is not always an option, especially for patients with contraindications to contrast agents, such as those with renal insufficiency, who face increased risks of contrast‐induced nephropathy [4]. In these cases, CMR serves as a valuable alternative due to its ability to visualize and measure valvular anatomy without exogenous contrast agents [5].

Traditional 2D CMR sequences, often used in cardiac imaging, require predefined acquisition planes, which reduces their flexibility. They also rely on patient cooperation for breath‐holding, which is challenging for many patient populations and can compromise image quality due to inconsistent chest wall positioning [6]. In contrast, three‐dimensional CMR presents several advantages, including simplified planning and potentially higher signal‐to‐noise ratio (SNR) [7]. It also allows for random undersampling in the 2D phase‐encoding plane with high incoherence, enabling exploitation of 2D sparsity for parallel imaging reconstruction [8]. The 3D images also offer greater flexibility, enabling reformatting into any desired plane after acquisition when isotropic data is acquired [6]. However, physiological motion remains a challenge in 3D CMR, and breath‐held Cartesian acquisition is prohibitively long.

To address respiratory motion, diaphragm navigators are commonly used to allow free breathing in 3D acquisitions. For example, the non‐contrast navigator‐assisted 3D whole‐heart cardiac magnetic resonance angiography acquisition (clinical MRA) is a free‐breathing sequence that provides the required anatomical assessment to ensure precise prosthesis sizing [9]. But, prospective gating is often hindered by navigator echo positioning, arrhythmia, and breathing pattern variability, leading to unpredictable scan times [10, 11]. In the worst case, scans are aborted due to low efficiency for patients with highly irregular breathing patterns [12, 13, 14]. In addition, clinical MRA provides data from only a diastolic phase, limiting its ability to account for dynamic changes in the aortic shape throughout the cardiac cycle [3, 15].

To allow free‐breathing acquisition while improving scan time predictability, self‐navigated trajectories can be employed, often acquiring a 1D respiratory signal from a superior–inferior projection [16]. However, as these techniques require the periodic acquisition of the center of k‐space (navigators), they also introduce the risk of eddy‐current distortions, as demonstrated in previous studies [17, 18]. These artifacts are more pronounced at higher magnetic fields (such as 3 T) and in fast sequences like balanced steady‐state free precession (bSSFP), which are commonly used in CMR due to their high signal‐to‐noise ratio and blood‐to‐muscle contrast [19], leading to lower image quality [17, 20, 21]. Image‐based navigators [22, 23] have also been proposed to estimate and correct for respiratory motion as an alternative to 1D navigators. While these techniques improve motion robustness, the acquisition of image‐based navigators typically requires additional time (exceeding 70 ms) compared to the 1D navigators, making them less suitable for applications demanding longer acquisition windows [22].

To reduce eddy‐current artifacts, self‐navigated non‐Cartesian trajectories have been proposed in previous studies [24, 25, 26, 27]. However, non‐Cartesian acquisitions inherently increase computational complexity during image reconstruction [28]. As a result, several approaches have been developed to reduce reconstruction time while mitigating eddy‐current artifacts using non‐linear k‐space trajectories on the Cartesian grid. An early implementation by Prieto et al. [12] proposed a single‐phase acquisition with a Cartesian spiral trajectory, reducing computational complexity and ensuring predictable scan times using a diaphragm navigator for respiratory motion correction. Subsequent work demonstrated the feasibility of ungated Cartesian spiral acquisitions, leveraging an external ECG signal to sort data into 16 cardiac phases [17]. Although promising, this approach showed suboptimal contrast in volunteers without the administration of gadolinium‐based contrast agents. Additionally, the spiral trajectory used did not incorporate variable density sampling, and therefore did not optimally leverage compressed sensing to mitigate undersampling artifacts [29]. A variable density spiral acquisition was later incorporated, but only provided a single‐phase diastolic acquisition [30, 31, 32].

To address these challenges, we propose a novel, self‐navigated, free‐breathing, bSSFP‐based dual‐phase 3D whole‐heart MRA acquisition. The method uses a hybrid Cartesian trajectory designed to mitigate eddy‐current artifacts by minimizing the k‐space jumps between consecutive readouts. T2 preparations [33, 34] are applied to improve the contrast between the blood pool and the myocardium without the need for exogenous contrast agents. The proposed trajectory naturally returns to the center of k‐space and provides variable density k‐space coverage. This approach allows for the acquisition of multiple navigators per heartbeat while minimizing transient artifacts, enabling us to address respiratory motion throughout long acquisition windows and across different cardiac phases. This enables dual‐phase imaging at 20% and 60% of the RR interval, when the largest and smallest aortic annulus are commonly observed [35]. Moreover, this new trajectory improves the predictability of scan times when compared to clinical MRA, while continuing to eliminate the need for breath‐holding and reducing the reliance on patient cooperation. The proposed method also takes advantage of an isotropic acquisition, offering the ability to retrospectively reformat images into various viewing planes.

2. Methods

2.1. Sampling Trajectory

An intrinsically self‐navigated Cartesian‐like trajectory has been developed to provide variable density k‐space coverage while minimizing the distance between adjacent readouts to mitigate eddy‐current artifacts. The chosen pattern is based on the polar equation of the rhodonea (rose) curve, encoded in $k_x$ and $k_y$ with $k_z$ readout, defined by the equation below:

$$r=a·\sin (\mathit{k}\theta )$$

where a is the amplitude, k is the angular frequency of the sinusoid, and r is the radial coordinate for a given polar angle θ. For this trajectory we selected an angular frequency of 2 to minimize the k‐space jumps between readouts. A petal shaped trajectory is determined for 0 ≤ θ≤πk, as shown in Figure 1A. Sampled points along the trajectory are defined by linear increments of πk/11, corresponding to the 11 readout acquisitions between consecutive self‐navigation signals. Sampled points are converted from the polar‐based trajectory to the nearest point on the Cartesian grid as defined by Euclidean distance ($x=r\cos \theta$ and $y=r\sin \theta$) to simplify the image reconstruction process. A combination of one petal and its 180° rotated pair is acquired, as shown in Figure 1B. This combination is referred to as a PETAL. The center of k‐space is acquired, followed by 11 acquisitions along the curve, naturally returning to the center of the k‐space, acquiring the same number of points along the rotated curve, and returning to the center of k‐space again. Therefore, two navigators are naturally acquired per PETAL and variable density sampling is achieved with increased density near the center of k‐space.

FIGURE 1.

Trajectory design: Schematic of the trajectory designed to mitigate eddy‐current artifacts caused by rapidly varying magnetic fields during the encoding of each readout. Panel (A) illustrates the k‐space sampling trajectory pattern, where the first lobe is acquired by starting at the center of k‐space, followed by 11 acquisitions along the curve, and then naturally returning to the center. Panel (B) shows the 180°‐rotated pair of the trajectory. Panel (C) depicts the subsequent PETALs, each rotated by the golden angle, with the amplitude adjusted to increase k‐space coverage. The first 15 PETALs are plotted to demonstrate this sampling strategy.

To increase k‐space coverage and optimize sampling when data is separated into distinct cardiac phases [36], the amplitude $a$ varies for different PETALs and each PETAL is rotated by a golden angle from the previous one, as shown in Figure 1C.

2.2. Sequence

The proposed sequence employs an ECG‐gated, free‐breathing, 50 ms T2‐prepared [34], fat‐saturated [37], 3D balanced steady‐state free precession (bSSFP) acquisition compiled in Pulseq [38], utilizing 10 dummy pulses with a linear flip angle ramp‐up [39] to achieve steady‐state conditions (Figure 2). The imaging parameters include a repetition time (TR) of 3.5–3.8 ms, echo time (TE) of 1.75–1.9 ms, and a flip angle between 40° and 45°, depending on specific absorption rate (SAR). A non‐selective RF pulse with a time‐bandwidth‐product of 2.0 and a duration of 1.3 ms is used. The field of view is 307 × 307 × 614 mm, and the matrix is set to 192 × 192 × 384. The acquired isotropic voxel size is 1.6 mm. Notably, the number of PETALS per heartbeat, and consequently the acquisition window, varies with the patient's heart rate (330–510 ms). For example, for a heart rate of 60 beats‐per‐minute, we selected an acquisition window of ˜500 ms, acquiring 6 PETALS and resulting in the acquisition of 12 navigators throughout the acquisition window; however, the overall scan time is consistently maintained at 10 min.

FIGURE 2.

Pulse sequence diagram: Schematic of the pulse sequence used for the ECG‐triggered, 3D bSSFP acquisition. In each heartbeat, a T2 preparation sequence and a spectrally selective fat saturation pulse (Fat‐Sat) are played out, followed by a linear ramp‐up of 10 dummy pulses to achieve steady‐state. The number of navigators and petals acquired per heartbeat varies based on volunteer heart rate.

2.3. Experiments

All imaging was performed on a 3 T MR system (MAGNETOM Cima.X, Siemens Healthineers, Forchheim, Germany) equipped with an ultra‐high performance gradient system achieving up to 200 mT/m maximum gradient strength. An 18‐channel body coil and an integrated spine matrix coil were used.

2.3.1. Phantom

Phantom imaging was performed to evaluate the impact of the k‐space trajectory on eddy current‐related artifacts. Using the proposed sequence, images were acquired using three different trajectories: (1) a fully sampled Cartesian acquisition (reference), (2) the proposed PETAL trajectory, and (3) a fully sampled random trajectory. To characterize the properties of each trajectory, we calculated the Euclidean distances between adjacent k‐space readouts and plotted the distribution of these distances as a histogram. To assess the degree of artifact in each image, we computed Pearson product–moment correlation coefficients, Structural Similarity Index Measure (SSIM) and Root Mean Square Error (RMSE) between the Cartesian reference and the reconstructed volumes of each acquired trajectory. To investigate the effect of the proposed PETAL trajectory's multi‐navigator strategy to address respiratory motion, we used the respiratory signal extracted from a volunteer scan to retrospectively bin the data based on navigator timing. We compared two binning strategies: (1) using a single navigator per heartbeat, where the entire acquisition window is rejected if the navigator is flagged as an outlier; and (2) using multiple navigators per heartbeat, allowing for partial rejection of only the data temporally closest to the outlier navigator. The same metrics were used to assess the performance of each respiratory‐resolved image.

2.3.2. In Vivo

Fifteen healthy volunteers with no known history of cardiovascular disease (8 females, average BMI: 21.3 ± 1.5 kg/m², average age: 27.1 ± 7 years old, average heart rate: 63.8 ± 10.9 bpm) were scanned using the proposed sequence under a research protocol approved by the Institutional Review Board of the Cleveland Clinic, Cleveland, OH. For comparison, a clinical reference free‐breathing, ECG‐gated 3D bSSFP navigator‐gated non‐contrast MRA (Clinical MRA) was acquired in diastolic phase with a 3 mm acceptance window, SPAIR, and 50 ms T2prep. For the clinical MRA sequence, the image parameters were: FOV = 307 × 307 mm², 128 slabs, resolution = 1.6 × 1.6 × 1.6 mm³, TR = 2.86–3.54 ms (TE = TR/2), flip angle = 38°–45°, iPAT = 2, partial Fourier = 6/8, acquisition windows = 353–380 ms. There is currently no golden standard for clinical 3D MRA in systole [5, 40, 41]. The acquisition of a second whole‐heart sequence as a systolic phase control was not feasible due to SAR limitations and time constraints. Therefore, to assess the conformational changes throughout the cardiac cycle observed in the dual‐phase PETAL acquisition, a breath‐held 2D bSSFP CINE was planned at the valve plane based on a coronal localizer and an LVOT CINE. The parameters for the 2D CINE at the valve plane were: FOV = 340 × 264 mm², resolution = 1.5 × 1.5 mm², 6 slices of 6 mm thickness with zero gap, echo spacing = 1.56 ms, flip angle = 30°, acceleration factor of 3 with GRAPPA, and Cartesian acquisition. The scan duration was recorded for all volunteers. Scan efficiency was calculated by determining the fraction of the data acquired that was used for the final reconstruction [17] for the proposed method and the clinical MRA.

2.4. Reconstruction

The images were reconstructed using region optimized virtual (ROVir) coil combination [42] using a GRE based scan with a large field of view (500 mm³, 100 × 100 × 100 matrix) to suppress information from undesired spatial regions and address wrap that could arise from the reduced encoding FOV. For respiratory motion correction (Figure 3A), the respiratory signal was derived using Principal Component Analysis (PCA) applied to a matrix of Fourier transformed navigator signals (6–12 navigators/heartbeat sampled every 11*TR) from each receiver coil [16]. Because the ECG triggering used in the PETAL acquisition results in unevenly sampled navigators in time, Lomb‐Scargle periodogram analysis [43] was used to analyze the frequency components of the resulting principal components. The component with the highest peak in the respiratory frequency range (0.1–0.5 Hz) was then selected as the 1D respiratory signal for motion correction. Outliers were determined using a thresholding technique where respiratory signal points deviating more than 1 standard deviation from the mode were flagged as outliers. The data points acquired closest in time to the points flagged as outliers were not used in the final reconstruction. The use of multiple navigators allowed us to reject or accept partially the acquisition window, reducing the amount of undersampling without introducing more motion as described in Figure 3B. The motion resolved data was then separated into two distinct cardiac phases based on the ECG gating (systole, diastole). The acquisition window used for each cardiac phase was chosen to provide a temporal resolution similar to CT scans (75–85 ms) [41]. The corresponding undersampling factor varied per cardiac phase and volunteer, resulting in an average accelerator factor of 4. Duplicate phase encoding lines within a cardiac phase were averaged. Each cardiac phase was reconstructed independently with SENSE (Sensitivity Encoding) using a total variation (TV) constraint in the phase encoding spatial dimensions [44] to address undersampling artifacts. The regularization weight was selected empirically and fixed across all volunteers ($\lambda =1\times {10}^{-4}$). JSENSE [45] was used to compute the sensitivity maps. All reconstructions were done off‐line with a non‐optimized Python based implementation using a Nvidia A100 GPU, and the average reconstruction time for both cardiac phases was 13 ± 2.7 min.

FIGURE 3.

Proposed motion resolved reconstruction. (A) PCA is applied to a matrix of the Fourier transform of the navigator signal from each receiver coil, and the principal component with the highest peak in the respiratory frequency range (0.1–0.5 Hz) using Lomb‐Scargle periodogram analysis is selected as the 1D respiratory signal for motion correction. Respiratory signal points within 1 standard deviation from the mode value are kept (green) and those with values outside are rejected (red). The data points acquired closest in time to the respiratory signal points flagged as outliers are not used in the final reconstruction. The respiratory motion resolved data is then further separated into distinct cardiac phases based on the ECG gating. Compressed sensing with a 2D spatial total variation constraint is used to reconstruct the undersampled data. (B) Schematic showing an example of using 4 navigators per heartbeat to track the respiratory signal. Each navigator is accepted (light green region) or rejected (light red region) based on the respiratory signal, and image signals are rejected according to the nearest navigator, allowing partial rejection of the acquisition window within a heartbeat.

2.5. Statistical Analysis

The aortic annulus area (cm²) and perimeter (cm) were measured using the closed curve tool in 3D Slicer [46] defined by 10 points. The plane used for the measurement was manually selected for the proposed sequence and the clinical MRA. The plane positioning was done similarly to the prescribed image position for the 2D CINE acquisition. Statistical analysis was performed using GraphPad Prism (v9.5.1, GraphPad Software, Boston, MA). The correlation between the aortic annulus metrics for diastole and systole was plotted comparing the proposed sequence and the corresponding frame of the clinical MRA and CINE, respectively. The correlation between measurements was assessed using the intra‐class correlation coefficient (ICC) assuming fixed raters rate each target [47]. The analysis was conducted in Python using the Pingouin [48] statistical package. Bland–Altman analysis was performed between the proposed method and the references to evaluate bias. Additionally, a non‐parametric alternative to the paired t‐test (Wilcoxon matched‐pairs) was used to test for significant differences (p < 0.05) between the metrics obtained from the proposed method and the references. GraphPad Prism was also used to evaluate the mean and variance of the scan duration between the proposed method and the clinical MRA. Additional statistical analysis comparing the results of the diastolic frame of the CINE to the clinical MRA and to the proposed method are provided in Figures S1 and S2, respectively.

3. Results

3.1. Phantom Results

In Figure 4A–C, we evaluated the impact of trajectory design on eddy current artifacts. Despite being fully sampled, the random trajectory exhibited visibly more artifacts compared to PETAL, indicating that these artifacts are not related to undersampling but rather to abrupt k‐space transitions. This was quantified using a correlation matrix, SSIM and RMSE, where PETAL showed a higher correlation coefficient and SSIM and lower RMSE with the Cartesian reference ($R_{\text{cartesian}\times \text{PETAL}}=0.794,\text{SSIM}=0.93,\text{RMSE}=5.87\times {10}^{-5}$) compared to the random trajectory ($R_{\text{cartesian}\times \text{Random}}=0.509,\text{SSIM}=0.83,\text{RMSE}=1.01\times {10}^{-4}$), suggesting reduced eddy current‐induced distortion. Additionally, the histogram analysis of inter‐readout distances revealed that PETAL resulted in smaller k‐space jumps, with a lower mean (27.6 mm⁻¹) and maximum (129.3 mm⁻¹) distance compared to the random trajectory (mean: 95.6 mm⁻¹, max: 261.6 mm⁻¹), further supporting its robustness against system‐related artifacts. Figure 4D illustrates the benefit of using multiple navigators per heartbeat to mitigate respiratory motion. For the PETAL trajectory, this approach reduced undersampling, providing 48% k‐space coverage compared to 40% when using a single navigator, and improving scan efficiency (60% and 54% respectively). Correspondingly, the image reconstructed using multiple navigators showed a higher correlation coefficient and SSIM and lower RMSE with the average PETAL image ($R_{\text{PETAL}\times \text{multiple navigators}}=0.90,\text{SSIM}=0.96,\text{RMSE}=3.71\times {10}^{-5}$) than the single‐navigator reconstruction (${\mathrm{R}}_{\text{PETAL}\times \text{single navigator}}=0.87,\text{SSIM}=0.84,\text{RMSE}=1.55\times {10}^{-4}$).

FIGURE 4.

Phantom validation of PETAL trajectory and motion binning: (A) Comparison of average (no respiratory or cardiac motion binning) reconstructions on phantom data acquired using Cartesian (reference), PETAL, and random trajectories. While eddy‐current artifacts are prominent in the images acquired with the random trajectory, they are visibly reduced with the proposed PETAL approach. (B) The matrix of Pearson product–moment correlation coefficients for Cartesian, PETAL, and Random acquisitions show increased agreement between PETAL and the Cartesian reference compared to Random. (C) Histogram comparing the distribution of k‐space jumps from the proposed trajectory (blue) to a fully random trajectory (green) demonstrating decreased Euclidean distance between acquired lines of k‐space for the PETAL distribution. (D) Comparison of single navigator per heartbeat versus multi‐navigators per heartbeat respiratory binning strategy on the undersampling pattern and image reconstruction for PETAL. The multiple navigator approach presented decreased undersampling and increased image correlation with the average PETAL reconstruction.

3.2. In Vivo Results

Figure 5A–C shows example clinical MRA and PETAL images of a representative healthy subject in the three anatomical planes. In Figure 5D, reconstructed images for diastolic and systolic phases were reformatted into the valve plane to demonstrate where the aortic annulus metrics were acquired. Additionally, representative aortic annulus measurements are shown in Figure 6 for the reconstructed volumes (in end‐systole and diastole) and their respective clinical references.

FIGURE 5.

Multi‐view comparison of PETAL versus clinical MRA: Sagittal (A), axial (B), and coronal (C) views of healthy volunteer images acquired using clinical MRA (diastole), and the two volumes (systole and diastole) acquired using PETAL. The aortic annulus perimeter and area were measured parallel to the valve plane, along the blue line, and are shown in the short axis of the aortic annulus (D).

FIGURE 6.

Aortic annulus measurements: PETAL versus reference methods: Example aortic annulus measurements in systole for PETAL (A) and the reference 2D CINE at the valve plane (B). Similarly, the measurements in diastole for the proposed method (C) and the reference whole‐heart Clinical MRA (D).

3.2.1. Diastole Statistical Analysis

To evaluate the performance of the PETAL trajectory in vivo, we compared aortic annulus area and perimeter during diastole against reference values obtained from clinical MRA. As shown in Figure 7A, PETAL measurements of the AA area during diastole demonstrated a strong correlation with clinical MRA ($\mathrm{ICC}=0.987$, n = 15). Bland–Altman analysis (Figure 7B) showed a mean difference of 0.008 cm², with 95% limits of agreement ranging from −0.315 to 0.331 cm². The distribution of differences was relatively scattered and exhibited no clear bias, indicating consistent agreement across measurement ranges. A Wilcoxon matched pairs test (Figure 7C) revealed no statistically significant difference between methods (p = 0.916). Similar results are observed when comparing AA area in diastole measured using two commercially available sequences (clinical MRA and the CINE), as shown in Figure S1A–C. Similarly, PETAL‐based AA perimeter measurements in diastole were highly correlated with references (Figure 7D; $\mathrm{ICC}=0.996$). Bland–Altman analysis (Figure 7E) reported a bias of −0.018 cm, with 95% limits of agreement between −0.152 and 0.116 cm. No systematic trend was observed in the differences. The Wilcoxon matched pairs test (Figure 7F) confirmed that this difference was not statistically significant (p = 0.301). Similar results are observed when comparing the AA perimeter in diastole measured with Clinical MRA and CINE, as shown in Figure S1D–F.

FIGURE 7.

PETAL versus Clinical MRA—Diastolic aortic annulus analysis: Statistical analysis of aortic annulus metrics in diastole for 15 healthy volunteers obtained using PETAL acquisition and the clinical MRA. Aortic annulus area is compared using correlation (A), Bland–Altman analysis (B), and a Wilcoxon matched pairs test (C). Aortic annulus perimeter is also compared using correlation (D), Bland–Altman analysis (E), and a Wilcoxon matched pairs test (F).

3.2.2. Systole Statistical Analysis

To further assess the performance of PETAL‐based measurements, we compared systolic measurements of the AA area and perimeter against reference values obtained from 2D CINE at the aortic valve plane.

As shown in Figure 8A, PETAL‐based measurements of the AA area in systole were highly correlated with those from the CINE control ($\mathrm{ICC}=0.989$, n = 15). Bland–Altman analysis in Figure 8B revealed a bias of 0.02 cm², with 95% limits of agreement between −0.324 and 0.364 cm². The data points were relatively dispersed without a clear trend, indicating consistent agreement across the range of values. A Wilcoxon matched pairs test (Figure 8C) showed no statistically significant difference between the two methods (p = 0.609).

FIGURE 8.

PETAL versus CINE—Systolic aortic annulus analysis: Statistical analysis of aortic annulus metrics in systole for 15 healthy volunteers obtained from PETAL acquisitions and CINE at the valve plane. Aortic annulus area is compared using correlation (A), Bland–Altman analysis (B), and a Wilcoxon matched pairs test (C). Aortic annulus perimeter is also compared using correlation (D), Bland–Altman analysis (E), and a Wilcoxon matched pairs test (F).

Measurements of the AA perimeter in systole also showed strong correlation between PETAL and CINE (Figure 8D; $\mathrm{ICC}=0.986$). Bland–Altman analysis (Figure 8E) indicated a mean difference of 0.073 cm, with 95% limits of agreement ranging from −0.203 to 0.349 cm. As with the area, differences suggested no systemic bias. However, the Wilcoxon matched pairs test comparing systolic AA perimeter in Figure 8F identified a small but statistically significant difference between methods (p = 0.029).

3.2.3. Contrast Difference Between Systole and Diastole

The time between T2‐preparation pulses and acquisition of the diastolic volume is longer than the time to the systolic volume acquisition, as preparation pulses are applied only at the beginning of each cardiac phase. Therefore, a difference in myocardium/blood pool contrast was observed between the two cardiac phases. This difference was measured for a representative volunteer by selecting regions of interest (ROIs) in the blood pool and myocardium (septum) for the phases shown in Figure 9. It was found that the blood pool‐to‐myocardium contrast was 16% lower in the volume reconstructed from data acquired further from the T2prep application (diastole), compared to the volume acquired closer to preparation pulses (systole).

FIGURE 9.

Systole versus diastole image contrast: Sagittal (A), axial (B), and coronal (C) views of images from a representative healthy volunteer acquired using clinical MRA, and the two volumes acquired using PETAL, representing diastole and systole. The volumes were reformatted along the valve plane (D). The blood pool‐to‐myocardium contrast was 16% lower in the PETAL diastolic volume (reconstructed from data acquired further from the T2prep application) compared to the PETAL systolic volume (acquired closer to preparation pulses).

3.2.4. Scan Duration and Efficiency

Scan duration was recorded for the proposed method and the clinical MRA for the 15 healthy volunteers. The resulting box‐plot analysis is shown in Figure 10A. The average scan time (mean ± standard deviation) for the proposed method was 10.13 ± 0.35 min, while the average scan time for the clinical MRA was 14.20 ± 5.38 min. Additionally, we calculated scan efficiency for both methods, as shown in Figure 10B. The average efficiency (mean ± standard deviation) of the proposed method was 62% ± 2.5% compared to 25% ± 10% for the clinical MRA.

FIGURE 10.

Scan time and efficiency comparison: (A) Box plot comparing scan duration for the proposed sequence (PETAL) and the clinical MRA for 15 volunteers. The scan duration (mean ± standard deviation) was 10.13 ± 0.35 min for PETAL versus 14.20 ± 5.38 min (25% percentile of 9 min and 75% percentile of 16 min) for clinical MRA. (B) Box plot comparing scan efficiency. The average efficiency (mean ± standard deviation) of the proposed method was 62% ± 2.5% compared to 25% ± 10% for the clinical MRA.

4. Discussion

The proposed method achieved a contrast‐free dual‐phase free‐breathing bSSFP‐based 3D whole‐heart MRA acquisition, implementing a new trajectory that mitigates eddy currents at 3 T while providing predictable scan times of 10 min (10.13 ± 0.35 min) compared to clinical MRA (14.20 ± 5.38 min). The acquisition process is simplified as no navigator echo positioning or prospective image planning is required. Reconstructed volumes can be reformatted to the user's preferred orientation due to the isotropic spatial resolution. Sampling on the Cartesian grid enables fast image reconstruction in approximately 15 min. The use of the petal‐like trajectory that naturally returns to the center of k‐space allows the lower frequencies to be better sampled while minimizing the distance between adjacent readouts, mitigating aliasing and eddy‐current artifacts, which are particularly important for higher magnetic fields and fast sequences. The acquisition of multiple navigators per heartbeat allows respiratory motion tracking for longer acquisition windows, allowing imaging in systole and diastole.

From the phantom experiment, we observed that the PETAL acquisition yielded a higher correlation to the Cartesian reference image when compared to a fully sampled random trajectory. The higher correlation was indicative of reduced eddy current‐related artifacts, as the trajectory was the only difference between images. The inter‐readout jumps are a known contributor to eddy current artifacts [17, 21], and the reduced level of artifacts was attributed to the more compact distance distribution.

In addition to reduced eddy‐current artifacts, the proposed trajectory design presents two other main benefits: (1) higher sampling density of lower frequencies of k‐space and (2) the acquisition of multiple navigators per heartbeat without introducing large k‐space jumps. (1) The proposed trajectory naturally mimics a Gaussian distribution, and the resulting variable density pattern benefits compressed sensing reconstruction [26, 49]. The higher sampling density of the center of k‐space provided by the proposed method reduces aliasing artifacts. This was demonstrated in Figures 4 and S3, where phantom images acquired with PETAL and random trajectories are compared to Cartesian images, repectively. As the center of the k‐space is not fully sampled for the random trajectory, undersampling artifacts are more pronounced, reducing image quality (correlation coefficient = 0.67 with respect to Cartesian reference) when compared to the same experiment using the PETAL trajectory (correlation coefficient = 0.90). The improved performance of variable density acquisitions has previously been demonstrated for free‐breathing diastolic 3D MRA [30, 31, 32] using image‐based navigators [22] to address respiratory motion. However, the time required to acquire these navigators prevents acquisition in different cardiac phases. (2) To address this limitation, the proposed method allows the acquisition of multiple 1D self‐navigating lines per heartbeat, without introducing large k‐space jumps, allowing for motion tracking throughout longer acquisition windows. The acquisition window can be binned into shorter segments associated with the respiratory signal obtained from principal component analysis of the navigator signals. This can reduce undersampling as less data is rejected when a navigator is determined as an outlier. The use of multiple self‐navigators is similar to the whole‐heart CINE previously proposed by Usman et al. [17], where 1D navigators are acquired periodically throughout the whole scan duration, enabling the retention of the respiratory component by filtering in the respiratory frequency range (0.1–0.5 Hz); however, gadolinium‐based contrast agents were needed for an ideal contrast between myocardium and blood pool. To improve the image contrast without exogenous contrast agents while allowing the visualization of AA in two phases, we instead proposed the use of T2‐preparation pulses [33, 34] every heartbeat. We used Lomb‐Scargle periodogram analysis to define the principal component with the highest peak in the respiratory frequency range from unevenly sampled navigators.

From the in vivo experiments, a high correlation between the proposed method and the clinical references of the aortic annulus perimeter and area for both systole and diastole was observed. Bland–Altman analysis showed a small bias and no discernible trend, which demonstrates the feasibility of the proposed method as an alternative to the conventional MRA. In the Wilcoxon matched pairs test, the diastolic AA measurements were not significantly different from those acquired through conventional MRA. Although systolic measurements exhibited a high correlation with clinical values, a small but statistically significant difference between PETAL and CINE measurements was observed between the two methods. This discrepancy is expected due to the differences in sequence parameters, including slice position, potential conformational changes occurring off‐plane, slice thickness, and resolution. It is important to note that similar correlation coefficient and bias were observed between the diastolic phase CINE and clinical MRA AA measurements (Figure S1), demonstrating that the observed differences are within the expected limits when comparing two different clinically validated acquisition methods.

The scan times for the proposed method were, on average, 4.07 min faster than the clinical MRA with a significantly smaller standard deviation, demonstrating the improved predictability of acquisition time. As MRA is an intrinsically time‐consuming procedure when compared to CTA, faster and more predictable scan durations can facilitate scheduling logistics and, combined with the simplified acquisition process, can potentially increase MRA utility [50, 51].

While yielding the faster scan times, the proposed method provides an additional volume enabling the evaluation of the aortic anatomy in the end‐systolic phase while clinical MRA provides volume only at diastole. This additional information in systole is important as the aortic annulus is subjected to conformational changes throughout the cardiac cycle. Typically, the annulus area and perimeter are larger in the systolic phase than in the diastolic phase [52]. The dynamics of the AA have significant implications for accurate prosthesis sizing, resulting in potential under sizing if only diastolic measurements are used [53], which has been associated with longer‐term mortality [15]. Therefore, procedure guidelines recommend using the information from multiple cardiac phases to properly select prosthesis sizes [3, 52, 54], which is made possible with the proposed method. We acquire data at 20% and 60% of the RR interval, when the largest and smallest aortic annulus are commonly observed [54]. Although this study targeted the end‐systolic phase, the initiation of image acquisition can be adjusted, if necessary, to accommodate preparatory modules such as T2 preparation, fat saturation, and linear ramp‐up, requiring approximately 100 ms, thereby enabling imaging closer to the peak‐systolic phase as well.

4.1. Limitations

Due to SAR constraints, the application of two T2‐preparation pulses per heartbeat was not feasible. As the timing between the T2‐preparation pulses and image acquisition can impact the contrast between myocardium and blood pool, the image acquired further from the preparation pulses (diastole) shows inferior blood‐to‐myocardium contrast when compared to the systolic volume. Even though this difference did not significantly impact the analysis of the aortic annulus in this study, further investigation is necessary to verify the impact on CVD patients. In addition, the fat saturation method chosen in this study to allow image acquisition in end‐systole is more susceptible to field inhomogeneities and did not suppress epicardial fat as well as the method used in the clinical reference (SPAIR). Future work will focus on optimizing the fat saturation method within the constraints of the proposed method.

Furthermore, an outlier rejection method was employed to mitigate respiratory motion, aiming to reduce computational complexity and reconstruction time in this study. This approach yielded a scan efficiency of 62% on average. While this represents an improvement over conventional clinical MRA, alternative techniques, such as those incorporating 1D translational [17] or 3D affine [12] motion correction, have been shown to achieve 100% scan efficiency. However, reconstruction times for these techniques are longer than 40 minutes. Future work could be done to incorporate motion correction techniques to improve efficiency if reconstruction time is not a constraint. Additionally, compressed sensing with a TV constraint in space was used in this study. However, as data is acquired throughout a longer acquisition window covering multiple cardiac frames, it is also possible to exploit sparsity across the temporal dimension using techniques such as k‐t SENSE [55] and k‐t SPARSE‐SENSE [56]. These approaches, combined with respiratory motion correction methods, could be explored in future work with the potential to further reduce scan times without compromising image quality and to provide multiple cardiac phases.

Despite the promising results of the proposed method, this study was performed on a small healthy volunteer cohort with normal heart rates (63.8 ± 10.9 bpm) and smaller body habitus (BMI 21.3 ± 1.5) which does not reflect the anatomy or physiology often seen in CVD patients. Therefore, future work should focus on including patients with indication for TAVR to evaluate the performance of the proposed sequence and the diagnostic value when compared to clinical protocols. Additional validation is also required to confirm the method's ability to capture the peak aortic annulus. Furthermore, due to SAR limitations and time constraints, a 2D CINE was used as a reference for the systolic AA measurements. While the primary purpose of the 2D control was to assess conformational changes of the aortic annulus, the use of a 2D CINE introduces several challenges that may have contributed to the differences observed in the systolic aortic annulus perimeter measurement despite being highly correlated. These include difficulties in accurately matching measurement positions and differences in image resolution, slice thickness, and temporal resolution. Future work should aim to further validate AA measurements in systole.

5. Conclusion

The proposed self‐navigated, free‐breathing dual‐phase bSSFP‐based whole‐heart MRI acquisition is a potential alternative to the conventional contrast‐free clinical MRA that provides additional end‐systolic information without additional scan time. By using a petal‐like trajectory that naturally returns to the center of k‐space, multiple navigators can be acquired per heartbeat without introducing large gradient variation, effectively addressing the challenges of respiratory motion and eddy‐current artifacts, resulting in more predictable scan times. The use of this trajectory leverages the advantages of isotropic 3D CMR acquisitions, including the ability to retrospectively reformat images into any desired plane, providing greater flexibility when compared to traditional 2D sequences. The high correlation and low bias of measurements with clinical sequences demonstrate the feasibility of this non‐contrast, free‐breathing approach.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

FIGURE S1: CINE versus Clinical MRA – Diastolic aortic annulus analysis: Statistical analysis of aortic annulus metrics in diastole for 15 healthy volunteers from CINE and the clinical MRA. Aortic annulus area is compared using correlation (A), Bland–Altman analysis (B), and Wilcoxon matched pairs test (C). Aortic annulus perimeter is also compared using correlation (D), Bland–Altman analysis (E), and Wilcoxon matched pairs test (F).

FIGURE S2: PETAL versus CINE – Diastolic aortic annulus analysis: Statistical analysis of aortic annulus metrics in diastole for 15 healthy volunteers obtained using PETAL acquisition and CINE. Aortic annulus area is compared using correlation (A), Bland–Altman analysis (B), and Wilcoxon matched pairs test (C). Aortic annulus perimeter is also compared using correlation (D), Bland–Altman analysis (E), and Wilcoxon matched pairs test (F).

FIGURE S3: Phantom experiment – Random trajectory: Average reconstruction of a fully sampled random acquisition. B) Comparison of respiratory binning strategies (single navigator per heartbeat versus multi‐navigators per heartbeat) on the undersampling pattern and image reconstruction for PETAL and Random acquisition. C) Correlation matrix between average reconstruction and the undersampled motion corrected images. As the center of k‐space is not fully sampled, significant image degradation is observed when compared to the average image (R _{Random×multiple} navigators = 0.69, SSIM = 0.85, RMSE = 2.7×10⁻⁴ and R _{Randomxsingle} navigator = 0.58, SSIM = 0.80, RMSE = 5.2×10⁻⁴).

Ribeiro Salles Moura T., Mai D., Wexler E., et al., “Feasibility Study for Dual Phase Free Breathing 3D bSSFP MRA With Preserved Eddy‐Current Tolerance With Adaptive k‐Space Lengths (PETAL) Acquisition,” Magnetic Resonance in Medicine 95, no. 4 (2026): 2078–2091, 10.1002/mrm.70192.