Jingyuan Lyu1, Jiali Zhong2, Yu Ding1, Qi Liu1, Lele Zhao3, Jian Xu1, Weiguo Zhang1, and Ruchen Peng2
1UIH America, Inc., Houston, TX, United States, 2Beijing LuHe Hospital, Capital Medical University, Beijing, China, 3United Imaging Healthcare, Shanghai, China
Synopsis
This abstract presents a new
approach to compressed sensing cardiac MRI, which enables free-breathing whole
heart coverage cine imaging within 30 seconds.
Using a phased array coil,
data was acquired continuously along Cartesian sampling trajectories using a
lookup table without ECG gating. Each slice was continuously sampled for a fixed
period of time, before the slice-selective RF excitation pulse switch to the next
slice. In reconstruction, the approach jointly updates coil sensitivity maps
and images, integrated with compressed sensing. In post-processing, virtual ECG
is calculated based on unsupervised machine learning.
Introduction
To date, compressed sensing magnetic
resonance imaging (CSMRI) using multichannel receiver coils has emerged as an
effective tool to reconstruct images from highly accelerated scan in various
applications. [1-2] However, the issue of accurate estimation of coil
sensitivities in cardiac cine imaging combined with respiration motion has not
been fully addressed, which limits the level of speed enhancement achievable
with the technology.
The self-calibrating (SC)
technique for sensitivity extraction has been well accepted, especially for
dynamic imaging, and complements the common calibration technique that uses a
separate scan. However, the existing method to extract the sensitivity
information from the SC data is not accurate enough in cardiovascular MRI, and
thus erroneous sensitivities affect the reconstruction quality when they are
directly applied to the reconstruction equation.
This abstract presents a new
approach to compressed sensing cardiac cine imaging. Using a phased array coil,
data was acquired continuously along Cartesian sampling trajectories using a
lookup table (pre-calculated based on Latin Hyper Sampling [4]) without ECG
gating. Each slice was continuously sampled for a fixed period of time (roughly
2 seconds) to cover the whole cardiac cycle, before the slice-selective RF
excitation pulse switch to the next slice. In reconstruction, the approach jointly
updates coil sensitivity maps and images, [4, 5] integrated with compressed
sensing. In post-processing, the approach uses unsupervised k-median method to
detect end-diastolic phases without ECG. The whole cardiac cycle images were processed slice-by-slice.Method
Model: The main goal is to design and implement a sampling and
reconstruction strategy that enables full heart coverage from highly accelerated free breathing real-time acquisitions,
with a relatively high spatial resolution (2.5 × 2.5 mm2)
and temporal resolution (40 ms). We adopt the VALAS [4] sampling pattern, and
multi-channel blind deconvolution model [5] to update the coil sensitivities. Because
respiration motion is involved, we assume each time frame / phase has its own
coil sensitivity map.
Problem formulation: Incorporating the above models, the data consistency model can be
formulated as $$d_{t,j}=\Omega_t FS_{t,j} \mathbf x_t$$ where $$$d_{t,j}$$$, $$$S_{t,j}$$$ represent the undersampled k-space data and coil
sensitivity profile from the j-th coil at the t-th time frame,
respectively; $$$F$$$ is the 2-dimensional fast Fourier transform, $$$\Omega_t$$$ represents
the undersampling operator.
Initialization: An
initial coil sensitivity map is estimated from the temporal average k-space
data.
Alternating optimization: The images and coil
sensitivity maps are calculated and updated by the following optimization
function alternatively: $$$\mathbf x_t=\underset{\mathbf x_t}{\arg\min} \frac{1}{2} \sum_{j} ||\Omega_t FS_{t,j} \mathbf x_t-d_{t,j}||^2+\lambda_1||TV_s(\mathbf x_t)||_1+\lambda_2||TV_t(\mathbf x_t)||_1$$$, $$$S_{t,j}=\underset{\mathbf S_{t,j}}{\arg\min}||\Omega_t FS_{t,j} \mathbf x_t-d_{t,j}||^2+\lambda||S_{t,j}F_t||_1$$$. $$$F_t$$$ represents temporal FFT.
Virtual ECG: After the cine
images were acquired, unsupervised k-median method [6] is used to cluster the
cine images to three groups. k-median clustering, originally from signal processing, that is popular
for cluster analysis in data mining. The median image was used as reference
image to calculate the correlation curve for all $$$\mathbf x_t$$$. Virtual ECG is calculated based on the correlation
curve and temporal resolution.Results
The
proposed method was implemented with a bSSFP sequence on a clinical 3T scanner
(uMR 780 United Imaging Healthcare, Shanghai, China) with the approval of local
IRB. Imaging parameters were: imaging matrix: 192(readout) x 165(PEs) x 36(phases)
x11(slices), TR/TE = 2.81/1.40 ms, slice thickness = 8 mm, flip angle =80°,
bandwidth = 1300Hz/pixel, spatial resolution = 1.82x1.82 mm2, 15 PEs/phase, 36 phases/slice, Scan time: 1.7s/slice.
Fig. 1
shows all image frames from one slice; Fig.2 shows end-diastolic phase images
from the proposed method, compared with traditional compressed sensing without
jointly reconstruction; Fig. 3 shows reconstructed end-systolic images from
both methods; Fig.4 shows the spatial-temporal profiles of the myocardium. The
ventricular morphology is better preserved from the proposed reconstruction
method, without any wall motion artifacts and/or bSSFP banding artifacts.
For comparison,
whole heart coverage retro cine was also implemented. Imaging parameters were: TR/TE=3.05/1.4ms,
bandwidth=1000Hz/pixel, Scan time: 14.6s/slice. 11 breath-hold for the whole heart
retro-cine scan. Fig. 5 compares the proposed method with retro-cine. Motion
artifacts were found in retro-cine, when the volunteer cannot hold the breath
well; whereas the reconstructed image from proposed method is in a real-time
fashion, without motion artifacts.Conclusion
We have proposed a novel approach to real-time cardiac
cine without breath hold and ECG gating. The approach effectively combines compressed
sensing, multichannel blind deconvolution, auto phase detection in the same
framework. The proposed method has shown success to recover whole heart cine
images (11 slices) without ECG from highly accelerated free breathing acquisitions.Acknowledgements
No acknowledgement found.References
[1]
Uecker, Martin et al., ESPIRiT--an eigenvalue approach to autocalibrating
parallel MRI: where SENSE meets GRAPPA. Magnetic resonance in medicine, vol. 71,3: 990-1001, doi:10.1002/mrm.24751, 2014.
[2]
Feng L, et al., Highly accelerated real‐time cardiac cine MRI using k–t SPARSE‐SENSE. Magnetic resonance in medicine, vol. 70.1: 64-74, 2013.
[3] Sheng J., Liu B., Ying L., Improved
self-calibrated spiral parallel imaging using JSENSE, Medical
Engineering and Physics, vol. 31, No. 5, pp. 510-514, June 2009.
[4] Lyu J., Ding Y., Zhong J., Zhang Z., Zhao L., Xu J.,
Liu Q., Peng R., and Zhang W., Toward single breath-hold whole-heart
coverage compressed sensing MRI using VAriable spatial-temporal LAtin hypercube
and echo-Sharing (VALAS), no.4752, ISMRM, 2019.
[5] Lyu J., Nakarmi U., Zhou Y., Zhang C.,
Ying L., Calibration-free Parallel Imaging Using Randomly Undersampled
Multichannel Blind Deconvolution (MALBEC), no. 3232, ISMRM, 2016.
[6] Jain A. K. and Dubes R. C., Algorithms for
Clustering Data. Prentice-Hall, 1988.