Abstract
Coronary MRI faces significant challenges, including lengthy acquisition times and limited spatial resolution. While parallel imaging can accelerate scans, image quality often degrades at acceleration rates beyond 2x. This study introduces a highly under-sampled coronary MRI method based on Compressed Sensing (CS) that utilizes transform domain dependencies to reduce blurring and artifacts, enabling acceleration factors up to 8x.
Introduction
Coronary Magnetic Resonance Imaging (MRI) is often considered the "holy grail" of cardiac imaging, yet it is plagued by a fundamental trade-off: achieving high-resolution images of small, moving arteries typically requires prohibitively long scan times. These lengthy acquisitions lead to motion artifacts and patient discomfort. While parallel imaging (using multiple receiver coils) provides some speed-up, it typically fails at acceleration factors greater than two, resulting in unusable, noisy images.
Compressed Sensing (CS) has emerged as a mathematical framework to reconstruct images from "under-sampled" data. However, standard CS methods often treat image components, such as wavelet coefficients, as independent variables. In reality, anatomical features like edges are highly correlated with their neighbors. This study proposes a novel CS approach that models these dependencies to preserve image integrity at high speeds.
Methodology
The researchers developed a reconstruction algorithm that moves beyond simple sparsity by incorporating a Bayesian Least-Squares approach with a Gaussian Scale Mixture (GSM) model.
The BLS-GSM Innovation
Standard CS methods typically use $l_1$ norm minimization to find the "sparsest" solution in a transform domain. The proposed BLS-GSM method instead captures the local correlations between a wavelet coefficient and its spatial neighbors. The algorithm iteratively ensures the reconstructed image is consistent with the raw MRI data while using the GSM model to "denoise" the image, effectively distinguishing between actual anatomy and random noise.
Experimental Validation
The feasibility of the method was tested using in vivo human data from a 1.5T Philips Achieva system. A 3D, free-breathing, ECG-triggered SSFP sequence was used to image the right coronary artery. Fully sampled "gold standard" data was retrospectively under-sampled by factors of 2, 4, 6, and 8 to compare the performance of the BLS-GSM algorithm against conventional $l_1$ norm minimization.
Results and Discussion
The primary finding of this study is that incorporating transform-domain dependencies significantly reduces the blurring and "blocky" artifacts commonly associated with high-acceleration CS.
As illustrated in the figure above, the BLS-GSM method maintains superior visual fidelity even at 8x acceleration. While the standard $l_1$ reconstruction becomes increasingly blurred and loses structural detail at factors of 6x and 8x, the BLS-GSM approach preserves the sharpness of the right coronary artery. This breakthrough suggests that a scan that previously took 10 minutes could potentially be completed in under two minutes without a loss in diagnostic quality.
Conclusion
This feasibility study demonstrates that modeling neighborhood dependencies in the transform domain is a powerful strategy for highly under-sampled coronary MRI. By breaking the traditional 2x acceleration barrier, this method offers a path toward faster, more reliable cardiac diagnostics. Future research will focus on optimizing the computational efficiency of the algorithm for real-time clinical use and validating the technique in larger patient cohorts.
Summary updated