Project Scientist University of California, Los Angeles
Background: Displacement encoding with stimulated echoes (DENSE) MRI is a valuable technique for myocardial strain evaluation1,2. Spiral cine DENSE MRI has demonstrated efficient and robust strain measurements3. However, DENSE has inherently low signal-to-noise ratio (SNR) due to the acquisition of the stimulated echo. Random matrix theory (RMT) denoising is one promising method to improve SNR4,5, which leverages the low-rank property and random Gaussian noise characteristics in images. In this study, we aimed to develop an RMT denoising technique to improve the SNR in 2D spiral DENSE MRI and evaluate it in vivo.
Methods: Eleven subjects (age 44±20 years; 3 healthy, 8 patients with heart disease) were scanned on clinical 3T scanners using a 2D spiral DENSE sequence3. A short-axis slice was acquired with breath-holding. Imaging parameters included FOV=360×360mm2, resolution=2.8×2.8mm2, slice thickness=8mm, TR/TE=16/1.08ms, cardiac phases=22-25, spiral interleaves per image=6, displacement encoding frequency=0.10 cycles/mm, through-plane dephasing frequency=0.08 cycles/mm. We performed denoising for the 5D data (2D images with Ne displacement encodings, Nc coils and Np cardiac phases). A patch size of m by n was selected empirically, and local patches were flattened into 2D (1st dimension: m×n, 2nd dimension: Ne×Nc×Np). After singular value decomposition, noise variance was estimated using RMT, and optimal singular value shrinkage6 was performed to suppress noise. Non-denoised and denoised DENSE images were processed3,7 to generate displacement maps, circumferential strain (Ecc) maps, and strain-time curves in 6 mid-level short-axis LV segments8. Magnitude SNR and phase SNR9 in the myocardium were measured and compared. Bland-Altman analysis was used to compare strain measurements. SNR improvement was evaluated with Wilcoxon signed-rank tests (p< 0.05 as significant).
Results: No signal of any anatomical structure was observed in the signal residual of the coil images (Figure 1a), demonstrating effective noise suppression without removing the desired signal. Coil-combined magnitude and phase images were greatly improved visually after denoising, and random background noise was removed (Figure 1b). The strain-time curves from denoised images showed reduced standard deviations in strain measurements (Figure 2). Both magnitude and phase SNR were significantly improved after denoising (all p< 0.01; Figure 3a-c). Bland-Altman analysis showed the strain measurements from non-denoised and denoised results were consistent (mean difference=0.00; Figure 3d).
Conclusion: We developed a RMT denoising technique to improve magnitude and phase SNR in 2D spiral cine DENSE MRI. Strain maps calculated from denoised images were less noisy while the mean strain measurements in non-denoised and denoised results were consistent. Further validation with different DENSE protocols may enable DENSE imaging with a higher resolution and more robust post-processing.
