Qingyong Zhu1, Yuanyuan Liu2, Zhuo-Xu Cui1, Ziwen Ke1, and Dong Liang1,2
1Research Center for Medical AI, SIAT, Chinese Academy of Sciences, Shenzhen, China, 2Paul C. Lauterbur Research Center for Biomedical Imaging, SIAT, Chinese Academy of Sciences, Shenzhen, China
Synopsis
We propose a novel DenOising induCed iTerative recOnstRuction framework (DOCTOR) to realize fast $$$T_{1\rho}$$$ parameter mapping from under-sampled k-space measurements. The proposed formulation constrains simultaneously intensity-based and orientation-based similarity between the reconstructed images and denoised prior images. Two state-of-art 3D denoising technologies are utilized including NLM3D and BM4D. The reconstruction alternates between two steps of denoising and a quadratic programming attacked by non-linear conjugate gradient method. The parameter maps are created from the reconstructed images using conventional fitting with an established relaxometry model. Through experiments in-vivo $$$T_{1\rho}$$$-weighted brain MRI datasets, we can observe superior image-quality of the proposed DOCTOR scheme.
Introduction
The observation forward model for accelerated scanning can eventually be
approximated as a discretized linear system:$$Y=\mathcal{A}X+N$$where $$$X\in \mathbb{C}^{U\times W}$$$ denotes the desired image series
to be reconstructed. $$$W$$$ is the total number of spin-lock
time (TSL). The observed k-space data
is given as $$$Y\in \mathbb{C}^{V\times W}$$$. The
linear operator $$$\mathcal{A}:\mathbb{C}^{U\times W}\rightarrow\mathbb{C}^{V\times W}$$$ formulates the observation forward model which
performs a multiplication by coil sensitivities followed by an undersampled
Fourier transform in each column. The backward program of reconstructing $$$X$$$ is an ill-posed problem by the reason that
the problem is underdetermined with $$$V \ll U$$$ . By adding penalty function to a cost function, the
variational formulation for this inverse problem can be written as: $$\min_{X}\frac{1}{2}\parallel\mathcal{A}X-Y\parallel^{2}_{2}+\lambda\mathcal{H}(X)$$where the penalty function $$$\mathcal{H}(X)$$$ enforces some a priori knowledge about the image series $$$X$$$, which usually acts as a regularization depended on transform-coefficient
sparsity$$$^{1,2}$$$, low-rank nature$$$^{3,4}$$$or the combinations$$$^{5,6}$$$. Method
In the paper, we propose a novel DenOisinginduCed iTerative recOnstRuction framework (DOCTOR)
to accelerate $$$T_{1\rho}$$$ parameter mapping from
under-sampled k-space measurements. The
parameter maps will be created subsequently from the reconstructed images using conventional fitting
with an established relaxometry model. We specify the main iteration of our reconstruction
algorithm as follows:$$\boldsymbol{Step~1}:X^{\mathcal{D}}=\mathcal{D}enoising(X,\sigma);\\~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\boldsymbol{Step~2}:\min_{X}\frac{1}{2}\parallel\mathcal{A}X-Y\parallel^{2}_{2}+\lambda\sum_{i}^{fov}(\parallel X_{i}-X^{\mathcal{D}}_{(i)}\parallel_{2}^{2}+\alpha\parallel (\mathcal{G}X)_{i}\times(\mathcal{G}X^{\mathcal{D}})_{(i)}\parallel_{2}^{2})$$where $$$\boldsymbol{Step~1}$$$ denotes a denoising step for 3D MR
images with multiple TSLs. The step aims to progressively induce $$$X^{\mathcal{D}}$$$ as prior-image series, which have
underlying image structures enforced by the powerful denoising model. $$$\sigma$$$ is the strength-parameter of
denoising, which is varied to gradually uncover image structures. In $$$\boldsymbol{Step~2}$$$, a quadric optimization with observation fidelity can be directly attacked
using non-linear conjugate gradient (NLCG) algorithm. The two $$$l_{2}$$$-norm based regularizing terms enforce the intensity-based and
orientation-based similarity between the reconstructed image and denoised prior
images in each iterative process. $$$\lambda$$$ is the trade-off parameter. $$$\alpha$$$ is used to control the contribution of orientation-based similarity prior. $$$\mathcal{G}$$$ represents the 3D gradient operator. A multi-scale gradients $$$^{7}$$$ in the
horizontal and vertical directions are given as: $$\mathcal{G}_{h}x=\frac{1}{J}\sum_{j=1}^{J}\frac{x(s+j,t)-x(s,t)}{\sqrt{j}};\\\mathcal{G}_{v}x=\frac{1}{J}\sum_{j=1}^{J}\frac{x(s,t+j)-x(s,t)}{\sqrt{j}}$$where $$$J$$$ is the scale number. $$$s$$$ and $$$t$$$ are the row and column index. When $$$J$$$ is large, the response of gradient on edges is
large and the tiny difference of weak edges in denoised prior images can be
effectively captured. Moreover, the symbol $$$\times$$$ represents the cross product, and the cross
product of vectors $$$u$$$ and $$$ v$$$ is defined as a determinant: $$u\times v=\begin{vmatrix} \boldsymbol{i} & \boldsymbol{j} & \boldsymbol{k}\\ u_{1} & u_{2} & u_{3}\\ v_{1} & v_{2} & v_{3}\end{vmatrix}=(u_{2}v_{3}-u_{3}v_{2})\boldsymbol{i}-(u_{1}v_{3}-u_{3}v_{1})\boldsymbol{j}+(u_{1}v_{2}-u_{2}v_{1})\boldsymbol{k}$$The
determinant has magnitude zero when the vectors are parallel.Experiment
The
experiments are carried out on $$$T_{1\rho}$$$-weighted brain MR images (TE = 7ms, TR = 3000ms,
excitation pulse flip angle =90°, refocusing pulse flip angle =180°, echo train
length =16, pixel size =0.6 × 0.6 mm2, matrix size =768368, slice
thickness =5mm, and TSLs =1, 20, 40, 60, and 80ms). 1D variable density random
under-sampling pattern is adopted to generate sampled k-space data. We compare our
reconstructions using the L+S $$$^{8}$$$ and Hankel-HOSVD (H2OSVD). Two
state-of-art 3D MR image denoising technologies based on block matching and
filtering are adopted including NLM3D $$$^{9}$$$ and BM4D $$$^{10}$$$. All reconstructions were performed in
Matlab R2017a on a standard laptop (Windows 10, 64 bit operation system,
Intel(R) Core(TM) i7-9700 CPU, 3 GHz, 32 GB RAMS). The reconstruction quality
is quantified with the relative error (RE), which are defined as RE$$$=\frac{\parallel X-\hat{X}\parallel_{2}}{\parallel X\parallel_{2}}\times 100\%$$$, where $$$X$$$ and $$$\hat{X}$$$ denote the fully-sampled original
image series and the reconstructed image series, respectively.Results
Figure 1 shows the reconstruction results of all comparison
methods at TSL=1ms and TSL= 80ms with sampling ratio (SR) of 25.00$$$\%$$$. Among
them, the returned images of the H2OSVD method has unacceptable artifacts (indicated by the blue arrow), and the
L+S method has a relative good result than H2OSVD. However, L+S still produces
some visible blurs. As expected, the DOCTOR based reconstructions have
superior performances: the DOCTOR-NLM3D and DOCTOR-BM4D methods both further
sharpen the images and suppress the artifacts. It is not difficult to see from
the error maps that the DOCTOR-BM4D method can define more fine details (indicated by the red arrow). Figure
2 shows that the DOCTOR methods achieve the parameter maps with higher quality
compared with the L+S and H2OSVD methods at SR=16.47$$$\%$$$, 25.00$$$\%$$$ and 33.15$$$\%$$$. The RE values of the reconstruction images and
parameter maps using all methods at different SRs are listed in the Figure.3. The
DOCTOR methods, especially DOCTOR-BM4D, have lower RE values than other
methods.Conclusion
In the paper, we propose a DOCTOR framework to accelerate $$$T_{1\rho}$$$ parameter mapping. Experimental
results prove the effectiveness and superiority of the proposed DOCTOR methods. Acknowledgements
This work was supported in part by the National Key R&D Program of China (2017YFC0108802 and 2017YFC0112903); National Natural Science Foundation of China (61771463, 81830056, U1805261, 81971611, 61871373, 81729003, 81901736); Natural Science Foundation of Guangdong Province (2018A0303130132); Key Laboratory for Magnetic Resonance and Multimodality Imaging of Guangdong Province; Shenzhen Peacock Plan Team Program (KQTD20180413181834876); Innovation and Technology Commission of the government of Hong Kong SAR (MRP/001/18X); Strategic Priority Research Program of Chinese Academy of Sciences (XDB25000000).References
- Doneve M, Brnert P, Eggers H, et al. Compressed
sensing reconstruction for magnetic resonance parameter mapping. Magnetic
Resonance in Medicine. 2010; 64(4): 1114-1120.
- Velikina
J V, Alexander A L, Samsonov A. Accelerating MR parameter mapping using sparsity-promoting
regularization in parametric dimension. Magnetic Resonance in Medicine. 2013;
70(5): 1263-1273.
- Zhang
T, Pauly J M, Levesque I R. Accelerating parameter mapping with a locally low
rank constraint. Magnetic Resonance in Medicine. 2015; 73 (2): 655-661.
- Park S H, Jin K, Hwan K, et al. Acceleration
of MR parameter mapping using annihilating filter based low rank hankel matrix
(ALOHA). Magnetic Resonance in Medicine. 2016; 76(6):
1848-1864.
- Zhao B, Lu W, Hitchens T K, et al. Accelerated MR
parameter mapping with low-rank and sparsity constraints. Magnetic Resonance in Medicine.
2015; 74 (2): 489-498.
- Zhu Y J, Liu Y, Ying
Leslie Y, et al. SCOPE: signal compensation for low-rank plus sparse matrix
decomposition for fast parameter mapping. Physics in Medicine & Biology.
2018; 63(18):185009(16pp).
- Zhu Q, Wang
W, Cheng J, et al. Incorporating
reference guided priors into calibrationless parallel imaging reconstruction.
Magnetic Resonance Imaging.
2019; 57:347-358.
- Otazo R, Emmanuel C, and Daniel K S. Low-rank plus sparse matrix decomposition
for accelerated dynamic MRI with separation of background and dynamic
components. Magnetic Resonance in Medicine. 2015; 73(3): 1125-1136.
- Manjon J V, Coupe P, Bonmati L M, et al. Adaptive
non-local means denoising of MR images with spatially varying noise levels. Journal
of Magnetic Resonance Imaging. 2010; 31(1): 192-203.
- Maggioni M, Katkovnik V, Egiazarian K, et al. Nonlocal transform-domain filter for volumetric data denoising
and reconstruction. IEEE Transaction on Image
Processing. 2013; 22(1): 119-133.