Highly accelerated free-breathing real-time myocardial tagging for exercise cardiovascular magnetic resonance

We developed the real-time tagging sequence by adding a bSSFP-based highly accelerated Cartesian imaging block after tagging preparation (Fig. 1). In this technique, in addition to acceleration by undersampling, further acceleration was achieved by omitting or truncating the high-frequency k-space regions along the phase encoding (PE) direction. Thus, the tagging gradients were applied perpendicular to the PE direction, preserving spatial harmonics solely along the readout direction. Furthermore, bSSFP ramp-up magnetization catalyzing pulses were added before imaging to reduce off-resonance artifacts during the approach to a steady state. When excess fat around the heart was a concern, a fat saturation pulse was added before tagging preparation to avoid chemical shift artifacts that may interfere with the visualization of the deformed tagging pattern and complicate image analysis.

Fat saturation was carried out upon detection of the cardiac trigger by applying a spectrally selective inversion recovery pulse with width = 5.1 ms and flip angle = 110^°. Five binomially distributed (1-3-4-3-1) radiofrequency pulses were applied to create a tagging line pattern with 8 mm spacing. Applying five linearly increasing ramp-up pulses before the onset of imaging was sufficient to reduce artifacts while minimizing the time delay after applying the tags, which was important during post-exercise due to elevated heart rates. The total imaging delay after cardiac trigger detection was 38 ms.

The bSSFP-based imaging sequence was implemented with twofold outer k-space truncation along the PE direction and incoherent random sixfold undersampling. This implementation resulted in a combined 12-fold acceleration rate. Two-chamber and short-axis tagging images were collected at the basal, mid-ventricular, and apical levels with the following imaging parameters: matrix size = 176 × 176, reconstructed spatial resolution = 2.0 × 2.0 mm², slice thickness = 8 mm, distance factor = 150%, bandwidth = 1235 Hz/Px, TE = 1.1 ms, TR = 2.62 ms, temporal resolution = 28.8 ms, asymmetric echo and flip angle = 30°. The acquisition window for two-chamber images was 1.2 s, while for each short-axis slice was 2 s to allow magnetization recovery before acquisition of the next slice.

We implemented an additional bSSFP-based breath-hold ECG-segmented sequence for comparison. Imaging was divided over eight heartbeats to enable an implementation whose imaging parameters were identical to those used for real-time tagging. An \(\alpha\)/2 pulse was added TR/2 after imaging was completed for a given heartbeat to store the steady-state magnetization and minimize ghosting artifacts resulting from the interruption of the steady state by the tagging block. Images were acquired with GeneRalized Autocalibrating Partial Parallel Acquisition (GRAPPA) rate 2, and retrospective ECG gating with 25 calculated cardiac phases.

Two prospective studies were conducted to assess the proposed technique (Table 1). All subjects provided informed consent approved by the Institutional Review Board for using their medical information in research. All patient data were handled in accordance with the Health Insurance Portability and Accountability Act. In the first study (Fig. 3), we focused on evaluating the technique exclusively at rest by recruiting 27 patients (19 male, 54 ± 16 years) and 19 healthy subjects (7 male, 26 ± 4 years). Patients were selected from those undergoing clinical CMR and willing to undergo an additional 5–10 min of scanning. All patients were eligible except for those with cardiovascular implantable electronic devices. CMR technologists approached patients based on time and staffing availability for consent. In the second study, we assessed the feasibility of real-time tagging imaging under physiological stress by recruiting 26 patients (18 male, 55 ± 15 years) and 23 healthy subjects (10 male, 40 ± 14 years) for an exercise-CMR study. The patient cohort included individuals with various heart failure types and LV ejection fraction levels, as well as those with shortness of breath under evaluation. Patients with significant valvular disease or cardiovascular implants were excluded.

Imaging was performed at 3 T using a commercial 18-channel cardiac coil and a 12-channel spine array (MAGNETOM Vida Siemens Healthineers, Erlangen, Germany). In the first study, we collected real-time tagging images in all subjects. To serve as comparison, ECG-segmented images were also collected in all patients (27 of 27) and a subset of healthy subjects (14 of 19). To assess repeatability, an extra set of real-time tagging images was collected in a subset of patients (6 of 27) and all healthy subjects (19 of 19) (Fig. 4A). In the Ex-CMR study, real-time tagging images were collected in all subjects before and after supine bike exercise (Fig. 4B). Subjects were exercised outside the scanner bore using a CMR-compatible cycle ergometer (Lode, Groningen, The Netherlands) attached to the scanner table. The work rate was started at Ω and increased by ∆Ω = + Ω every 2 min while subjects maintained a constant pedaling speed of 75 rpm. Subjects exercised with Ω = 10–25W resistance protocols. Heart rate during exercise was measured using a pulse oximeter at the end of each incremental step. Immediately after reaching the target heart rate ([220 − age] × 0.85), exhaustion, or after 10 min, subjects were returned to the scanner bore for post-exercise stress imaging. There was 4–8 s gap between the end of exercise and stress imaging. Further, cine images were collected prior to tagging to evaluate left and right ventricular function immediately post-exercise but they are not reported as part of this study. Therefore, this resulted in an additional delay of 91 ± 27 s prior to tagging imaging.

In the first study, we aimed to investigate the feasibility of quantifying myocardial deformation using real-time tagging images. We employed an open-source DL model for landmark tracking to generate measures of mid-wall global circumferential strain (GCS) [44]. The DL model has been previously trained for short-axis tracking. Therefore, deformation was only quantified in short-axis images. Each slice was analyzed separately, with a global value obtained by averaging strain across slices. The strain was derived from ECG-segmented and real-time images at rest to assess agreement and from repeated real-time acquisitions to assess repeatability.

In the Ex-CMR study, we aimed to assess the feasibility of real-time tagging imaging during physiological stress using real-time tagging images collected at rest and after supine exercise. Two experienced CMR readers (F.G: 5 years, S.N: 14 years) independently evaluated the quality of 2-chamber and short-axis images. The quality of all real-time images was evaluated, including those collected at rest in the first study. The readers were blinded to the timing of acquisition and were presented with the images separately and in random order. Each reader independently rated image quality on a 4-point Likert scale based on tagline quality and artifact level. Tagline quality was graded as follows: 1-excellent, with well-defined tagline patterns and clear differentiation between dark and bright lines; 2-good, with well-defined patterns but with partial signal intensity in saturated lines; 3-moderate, with distinguishable but not well-defined patterns and partial signal intensity in saturated lines that appeared grey. Taglines also showed minor motion-induced blurring; 4-poor, with no well-defined patterns, significant brightness in saturated regions, and significant respiratory and cardiac motion-induced blurring. Artifact level was graded as follows: 1-none, with no artifacts present in the image; 2-minimal, with slight artifacts present that did not significantly impact image quality; 3-moderate, with noticeable artifacts present that somewhat degrade image quality; 4-significant, with severe artifacts rendering the image unusable for diagnostic purposes. Lastly, the feasibility of quantifying myocardial deformation after physiological exercise was assessed by evaluating mid-wall GCS in real-time tagging images post-exercise.

The SciPy (stats, v1.7.1) package was used for statistical analysis [46]. We assessed agreement between ECG-segmented and real-time measures of mid-wall GCS using linear regression, constraining the intercept to zero, and expressed it as slope [95% confidence interval (CI)]. The correlation was evaluated using the Pearson correlation coefficient, r, with the following criteria: r values above 0.8 indicated a very strong correlation, above 0.7 good, above 0.6 moderate, above 0.3 fair, and below 0.3 poor [47]. Scan/re-scan reproducibility of real-time tagging mid-wall GCS measures was assessed using the coefficient of variation (CoV) and intraclass correlation coefficient (ICC). Based on the ICC [95% CI] estimate, values less than 0.5, between 0.5 and 0.75, between 0.75 and 0.9, and greater than 0.90 were considered poor, moderate, good, and excellent reliability, respectively. The minimal detectable change was defined as standard measurement error × \(\sqrt{2}\times 1.96\), where the standard measurement error is given by standard deviation × \(\sqrt{1-\text{ICC}}\) and the standard deviation is based on the measurements from the first scan. Differences in GCS were also visualized using Bland–Altman plots and expressed as mean difference [mean difference ± 1.96 × standard deviation of the difference].