Table 1.
Gated scans acquired 9–18 slices per heartbeat, and ungated scans acquired 6–9 slices (~20–33 images per second) at 30 rays/frame. The SMS sets order also indicates the acquisition order after the saturation pulse.
Fig 1.
Pulse sequences used in this study.
The gated sequence applies an SR pulse (composed of 6 pulses) after an ECG trigger followed by acquisition of an SMS set. 3–6 SMS sets were acquired depended on the heartrate. The ungated sequences continuously acquire 2 (sequence 1) or 3 (sequence 2) SMS sets, with one SR pulse applied for each SMS set (sequence 1) or one SR pulse applied for all SMS sets (sequence 2).
Fig 2.
Flowchart describing the pre-processing and image reconstruction.
Fig 3.
Iterative reconstruction using SMS GROG.
The acquired radial k-space was first separated into different phase modulations, then within each phase modulation a set of GROG operators was trained and applied to interpolate the radial k-space onto a Cartesian grid. In the iteration loop, the Cartesian k-space was demodulated and modulated back and forth to calculate the data fidelity term.
Fig 4.
Self-gating steps for ungated acquisitions.
Summation of the first 10 pre-contrast frames was threshold to keep low intensity areas to create mask 1, and maximum intensity map was threshold to keep high intensity areas to be mask 2. Then these two masks were combined and smoothed by a Gaussian filter to identify blood pools in all the simultaneously excited slices. The signals in the blood pool masks of all simultaneously acquired slices were summed together to give a 1D signal that was normalized to be between 0 and 1. Near-systolic and near-diastolic frames were selected from the normalized signal.
Fig 5.
Illustration of the respiratory binning process.
An average intensity map was first calculated by averaging pixel intensities along the time dimension using mean-filter-smoothed dynamic images. From this map, the half of the pixels with the highest intensity were kept and then normalized by each pixel’s highest intensity along time. Another average intensity map was calculated using the normalized data and half of the highest intensity pixels in this map were selected in order to mainly preserve the chest wall signal. PCA was applied on the selected pixels. The PCA component that has highest intensity in frequency range 0.1–0.5 Hz (within the first 10 principal components) was selected and normalized to be the respiratory gating signal. All time frames were binned into 4 respiratory bins using the normalized signal.
Fig 6.
Illustration of pixel tracking temporal regularization.
(a) The right image shows the signal intensity line profile over time of the red line in the left image. Standard temporal total variation calculates total variation along the blue line, however pixel tracking temporal total variation calculates total variation along the yellow line. (b) The blue line shows estimated motion track of one pixel at the edge of blood pool. (c) Pixel tracking regularization is applied on both the entire image sequence (shown below the time axis), as well as each motion bin (shown above the time axis).
Fig 7.
Results from an ECG-gated stress acquisition with a small perfusion deficit (subject 6).
6 SA slices and 3 2CH LA slices were acquired within each heartbeat. (a) shows all slices near peak myocardial enhancement. (b) shows a SA slice and a 2CH LA slice closely matching the corresponding LGE images are shown. A small area of deficit can be seen on both slices pointed by the arrows. This is more clearly seen in the time curves shown in (c), (d) for the corresponding ROI. Model-based registration [34] was performed before image segmentation. The time curves were normalized by proton density images and the pre-contrast baseline signal was subtracted.
Fig 8.
Illustration of an 18 slice acquisition at rest in subject 2.
12 SA, 3 2CH LA and 3 4CH LA slices were all acquired within one heartbeat for every heartbeat. For the SA slices, each column was acquired as an SMS set, and the 2CH LA slices or 4CH LA slices were each acquired as separate SMS set.
Fig 9.
Results from an ungated acquisition (subject 10).
Stress and rest images for systolic and diastolic phases are shown. In each image three slices in a row were acquired simultaneously. Time frames were picked to show peak myocardial enhancement.
Fig 10.
Tissue time intensity curves from an ungated stress acquisition (subject 12).
The left image shows the selected slice with six cardiac segments. The right image shows the tissue intensity curves of the corresponding segments.
Fig 11.
Dark band artifacts using overlapping slice positions.
Faint dark bands (pointed by blue arrows) can be seen on all of the simultaneously acquired 2CH LA slices at slice intersections with previously acquire SA slices. Images were from subject 11.
Fig 12.
Comparison of PT-STCR and STCR in subject 4.
3 SA and 1 2CH LA slices were used in the SNR/CNR analysis are shown (9 slices were acquired). Vertical red lines on the images are used for comparison of signal intensity line profiles over time. PT-STCR reconstruction preserves the myocardial border better than STCR and has fewer artifacts. Images showing the signal intensity line profiles over time show that the PT-STCR reconstruction is sharper than the STCR reconstruction.
Fig 13.
The results shown here are averaged across all segments of SA, 2CH LA and/or 4CH LA slices from one Gadolinium injection. For ungated acquisitions, near-systolic and near-diastolic images were included in averaging. PT-STCR reconstructions have higher SNR and CNR than STCR reconstructions overall.
Fig 14.
Comparison of undersampled reconstructions in a 72 rays/frame dataset (subject 8).
Simulated undersampling was done by keeping only the first 30 rays. The 72 rays reconstruction is slightly better than the 30 rays reconstruction, however both reconstructions have high quality and contrast.