Johannes Breitling1, Steffen Goerke1, Mark E. Ladd1, Peter Bachert1, and Andreas Korzowski1
1German Cancer Research Center (DKFZ), Heidelberg, Germany
Synopsis
In this study a novel method for the denoising of
CEST MRI data is presented, combining the formation of subsets of similar
spectra and the subsequent application of a principal component analysis. Exploiting
only the subtle spectral differences of these reduced datasets – as opposed to
using all spectra for the analysis – allows for a better identification and
isolation of the obscured underlying spectral features. The proposed denoising
resulted in an SNR gain by approximately a factor of four compared to the
noisy initial data and an additional 14% compared to the
conventional principal component analysis denoising.
Introduction
Recently, we developed an adaptive denoising algorithm
for CEST spectra, combining the identification of spectral redundancies by
means of a principal component analysis (PCA) with an appropriate data-driven
truncation criterion to separate the relevant spectral features from noise-like
characteristics.1 However, an even better denoising performance for CEST data is
desirable, for example, to generate high-SNR ground truths for the reliable
training of neural networks. In this study a novel method is presented,
extending the previous approach by a second stage of denoising2,3 (Step 3 in methods section). To this end, the result of the previously
established denoising approach is used to define for each spectrum the subset
of its most similar spectra. Subsequent application of a PCA to these reduced
datasets – exhibiting only subtle spectral differences – allows for a better identification
and isolation of the obscured underlying spectral features.Methods
The proposed denoising algorithm comprises the
following steps (Fig. 1):
Step 1: The acquired CEST data of size $$${u}\times{v}\times{y}\times{n}$$$ (one 3D image for each of the $$$n$$$ saturation frequency offsets) is reformatted
into a Casorati matrix of size $$${m}\times{n}$$$ (with $$$m\leq{u}\cdot{v}\cdot{y}$$$ being the number of remaining voxels after
skull stripping). Subsequently, a variance-stabilizing transformation (VST) is
applied to the matrix to correct for the Rician noise induced bias.4
Step 2: In the first denoising stage, all noisy spectra are
denoised using the previously established adaptive PCA, whereby the optimal
truncation is determined by the Nelson criterion.1,5
Step 3: In the newly introduced second denoising stage, each
resulting prefiltered spectrum is used to define the group of its $$$K$$$ (in this study set to $$$10\cdot{n}$$$) most similar spectra (Step 3a),3 whereby the similarity between two
spectra $$$\mathbf{s_i}$$$ and $$$\mathbf{s_j}$$$ is determined by the cosine similarity:
$$\textrm{sim}(\mathbf{s_i},\mathbf{s_j})=\frac{\mathbf{s_i}\cdot\mathbf{s_j}}{||\mathbf{s_i}||\cdot||\mathbf{s_j}||}$$
The corresponding groups of noisy spectra are each
denoised individually using again the adaptive PCA (Step 3b and c). As a
result, each group yields an estimate for all included spectra.
Step 4: The final denoised spectra are calculated by
averaging the corresponding estimates from all groups.
Step 5: Application of the inverse VST and reformatting to an
image series yields the denoised dataset.
In vivo 3D CEST MRI
was performed on a 7T whole-body scanner (Siemens Healthineers, Germany) using
the snapshot-CEST approach6 with a matrix size of $$${128}\times{104}\times{12}$$$ and a resolution of $$${1.7}\times{1.7}\times{3}~{mm}^3$$$. Presaturation was obtained by 140
Gaussian-shaped pulses $$$(t_p=15~ms,~{duty}~{cycle}=60\%,~{t_{sat}}=3.5~{s},~\textrm{mean}~{B_1}=0.7\mu{T})$$$ applied at 56 unevenly distributed
offsets. Low-SNR data as input for the denoising were acquired with bandwidth $$$(BW)=1560~Hz/pixel$$$ and flip angle $$$(FA)=1^\circ$$$ and high-SNR data with $$$BW=560~Hz/pixel$$$ and $$$FA=4^\circ$$$. All acquired image data were motion corrected using an
automatic multimodal rigid registration algorithm in MITK,7 corrected for $$$B_0$$$ inhomogeneities and
manually skull stripped. A ground truth was obtained by averaging five high-SNR
datasets and application of the first stage (i.e. single-stage) denoising. To
obtain a reference for the denoising performance, the conventional single-stage
denoising was also applied to the low-SNR data. Image contrasts were calculated
by the asymmetric magnetization transfer ratio $$$(MTR_{asym}=(M(-\Delta\omega)-M(\Delta\omega))/M_0)$$$ at $$$\Delta\omega=3.5~ppm$$$.Results
Application of the proposed two-stage denoising
algorithm to the low-SNR data results in an image quality comparable with the
ground truth (see last paragraph in Methods for definition of the ground truth)
and slightly improved compared to using only the conventional denoising approach
(Fig. 2). In the corresponding Bland-Altman plots (Fig. 3), this translates into
a reduced spread of the limits of agreement – and thereby gain in SNR – by a
factor of approximately four compared to the noisy initial data and 14% compared to the single-stage denoising. A small bias of the proposed
denoising is observable (i.e. difference of the mean from zero in the Bland-Altman
plot), which can however in large parts be assigned to the different BW and FA
of the low-SNR data acquisition with respect to the ground truth as it is already
present in the noisy data.Discussion
The proposed method extends the previously presented denoising
algorithm1 by a second denoising stage allowing, improved image
quality. Moreover, the selection and subsequent denoising of similar subsets of
spectra reduces the risk of spectral features being undesirably removed or
significantly blurred. This is especially important for rare spectral features only present in a small region of the acquisition volume (e.g. in the tumor) containing
the important clinical information. However, introduction of the second stage
comes at the cost of increased computation time, as the spectral matching and
subsequent denoising are repeated for each voxel. In a next step, a thorough
investigation on realistic simulations will be necessary to be able to compare
the denoising results with an unbiased ground truth, thereby allowing quantification
of the SNR gain across multiple noise levels, as well as investigation of the
effect of the denoising on rare spectral features.Conclusion
In this study, a novel denoising algorithm for CEST MRI is presented that exploits in a second stage the
denoising of spectrally matched groups. The resulting improved
image and spectral quality might in the future allow for the development of
faster CEST imaging sequences, application of metrics and methods especially
prone to noise, and generation of high-SNR ground truths for the reliable
training of neural networks.Acknowledgements
JB acknowledges the financial support of the
International Max Planck Research School for Quantum Dynamics in Physics,
Chemistry and Biology.References
1. Breitling J,
Deshmane A, Goerke S, et al. Adaptive Denoising For Chemical Exchange
Saturation Transfer MR Imaging. NMR Biomed. 2019;32(11).
2. Zhang X, Peng
J, Xu M, et al. Denoise Diffusion-Weighted Images Using Higher-Order Singular
Value Decomposition. Neuroimage. 2017;156:128-145.
3. Zhang L, Dong
W, Zhang D, Shi G. Two-Stage Image Denoising By Principal Component Analysis
With Local Pixel Grouping. Pattern Recognit. 2010;43(4):1531-1549.
4. Foi A. Noise
Estimation And Removal In MR Imaging: The Variance-Stabilization Approach. In: 2011
IEEE International Symposium on Biomedical Imaging: From Nano to Macro.
IEEE; 2011:1809-1814.
5. Nelson LR.
Some Observations On The Scree Test, And On Coefficient Alpha. J Educ Res
Meas. 2005;3(1):1-17.
6. Zaiss M, Ehses
P, Scheffler K. Snapshot-CEST: Optimizing Spiral-Centric-Reordered Gradient
Echo Acquisition For Fast And Robust 3D CEST MRI At 9.4 T. NMR Biomed.
2018;31(4):e3879.
7. Floca R.
Matchpoint: On Bridging The Innovation Gap Between Algorithmic Research And
Clinical Use In Image Registration. In: IFMBE Proceedings. Vol 25.
Springer, Berlin, Heidelberg; 2009:1105-1108.