Undersampled MRI reconstruction with patch-based directional wavelets
Xiaobo Qu1,2, Di Guo2, Bende Ning1, Yingkun Hou3, Yulan Lin1, Shuhui Cai1, Zhong Chen1,*
1Department of Electronic Science, Fujian Provincial Key Laboratory of Plasma and Magnetic Resonance, Xiamen University, Xiamen, 361005, China
2Department of Communication Engineering, Xiamen University, Xiamen, 361005, China
3School of Information Science and Technology, Taishan University, Taian 271021, China
Citations: Xiaobo Qu, Di Guo, Bende Ning, Yingkun Hou, Yulan Lin, Shuhui Cai, Zhong Chen. Undersampled MRI reconstruction with patch-based directional wavelets, Magnetic Resonance Imaging, in press, 2012.
Download the paper here
[Possible changes could be made in the published version. Please visit journal website for the final version]
Download the code: Toolbox_sparseMRI_PBDW
Compressed sensing has shown great potential in reducing data acquisition time in magnetic resonance imaging (MRI). In traditional compressed sensing MRI methods, an image is reconstructed by enforcing its sparse representation with respect to a preconstructed basis or dictionary. In this paper, patch-based directional wavelets are proposed to reconstruct images from undersampled k-space data. A parameter of patch-based directional wavelets, indicating the geometric direction of each patch, is trained from the reconstructed image using conventional compressed sensing MRI methods and incorporated into the sparsifying transform to provide the sparse representation for the image to be reconstructed. A reconstruction formulation is proposed and solved via an efficient alternating direction algorithm. Simulation results on phantom and in vivo data indicate that the proposed method outperforms conventional compressed sensing MRI methods in preserving the edges and suppressing the noise. Besides, the proposed method is not sensitive to the initial image when training directions.
Keywords: MRI; Fast imaging; Compressed sensing; Wavelet; Directional transform
In PBDW, the image is divided into patches, and the geometric direction is selected out among the candidate directions to provide the sparsest representation for each patch. Then, pixels in each patch are arranged parallel to the geometric direction and formed into a column vector. Finally, wavelet transform is performed on the column vector of each patch. The flowchart of PBDW is illustrated in Fig.1, and an example shown in Fig. 2 further illustrates the process.
Fig. 1. Flowchart of the patch-based directional wavelets. The process of geometric direction estimation in a patch is shown in the dashed rectangle.
Fig. 2. Enhanced sparsity by rearranging the pixels parallel to optimal geometric direction. (A) A patch with all candidate directions; (B) projecting a pixel to the axis which is orthogonal to a given direction; (C) the indexes of the arranged pixels; (D) a patch marked with two candidate directions; (E) rearranged pixels (1D vectors) parallel to the two directions, and the horizontal axis is the index of the pixel in the new vector; (F) sorted magnitude of 1D Haar wavelet coefficients and (G) approximation error by preserving S largest coefficients.
Fig. 5. Flowchart of the proposed method.
1. Reconstruction with phantom
A simulated phantom with nonpiecewise constant image features is adopted in simulation [26]. The sampling pattern with sampling rate 0.30 is shown in Fig. 6A.
Fig. 6. Reconstructed phantom images. (A) The Cartesian sampling pattern with a sampling rate 0.30; (B) fully sampled image; (C–H) reconstructed images using ODWT+TV, TV, PBW, SIDWT, PBDW with directions estimated from the SIDWT-based reconstruction and PBDW with directions estimated from the fully sampled image.
2. Reconstruction with in vivo data
Fig. 7. Reconstructed lemon images. (A) The Cartesian sampling pattern with a sampling rate 0.40; (B) fully sampled image; (C–H) reconstructed images using ODWT+TV, TV, PBW, SIDWT, PBDW with directions estimated from the SIDWT-based reconstruction and PBDW with directions estimated from the fully sampled image.
Fig. 8. Reconstructed brain images. (A) The Cartesian sampling pattern with a sampling rate 0.45; (B) fully sampled image; (C)-(H) reconstructed images using ODWT+TV, TV, PBW, SIDWT, PBDW with directions estimated from the SIDWT-based reconstruction, PBDW with directions estimated from the fully sampled image.
Fig. 10. Reconstructed lemon images with added noise. (A) The Cartesian sampling pattern with a sampling rate of 0.50; (B) fully sampled image; (C–H) reconstructed images using ODWT+TV, TV, PBW, SIDWT, PBDW with directions estimated from the SIDWT-based reconstruction and PBDW with directions estimated from the fully sampled image.
Fig. 11. Reconstructed brain images with added noise. (A) The Cartesian sampling pattern with a sampling rate of 0.60; (B) fully sampled image; (C–H) reconstructed images using ODWT+TV, TV, PBW, PBDW with directions estimated from the SIDWT-based reconstruction and PBDW with directions estimated from the fully sampled image.
Fig. 12. Effect of the initial reconstruction in PBDW for brain image. (A) RLNE versus times of estimation directions at a sampling rate of 0.45, (B) RLNE versus the sampling rates when the times of estimation directions are two. Note: The “0” in the horizontal axis in panel (A) corresponds to the first image used to train the directions. This image is obtained using conventional CS-MRI methods.
In this paper, PBDW-based undersampled MRI reconstruction is proposed. Geometric directions in PBDW are trained from the reconstructed image using conventional CSMRI methods and integrated into the final undersampled MRI reconstruction. Simulation results on phantom and in vivo data indicate that the proposed method can better preserve the edges and suppress the noise than traditional CS-MRI methods do. The PBDW is suggested to follow conventional CS-MRI reconstructions to improve the reconstruction if computational burden is acceptable for the applications. The limitation of PBDW is that edges are blurred at low sampling rates. Further improvement on the reconstruction is possible if better direction estimation is obtained. Finding a reasonable initial image at low sampling rate is still open. In order to avoid introducing an explicit initial image to train the directions, one may incorporate the direction training in the reconstruction formulation so that geometric directions totally rely on the undersampled data, rather than the image reconstructed using conventional CS-MRI methods. Extension of this work with the weighted ℓ 1 norm optimization [28] and nonconvex optimization [29] is being considered to further improve the reconstruction.
[1] Finn JP, Nael K, Deshpande V, Ratib O, Laub G. Cardiac MR imaging: state of the technology. Radiology 2006;241:338–54.
[2] Yang RK, Roth CG, Ward RJ, deJesus JO, Mitchell DG. Optimizing abdominal MR imaging: approaches to common problems. Radiographics 2010;30:185–U210.
[3] Hu YL, Glover GH. Increasing spatial coverage for high-resolution functional MRI. Magn Reson Med 2009;61:716–22.
[4] Lustig M, Donoho D, Pauly JM. Sparse MRI: the application of compressed sensing for rapid MR imaging. Magn Reson Med 2007;58:1182–95.
[5] Le Pennec E, Mallat S. Sparse geometric image representations with bandelets. IEEE Trans Image Process 2005;14:423–38.
[6] Knoll F, Bredies K, Pock T, Stollberger R. Second order total generalized variation (TGV) for MRI. Magn Reson Med 2011;65:480–91.
[7] Do MN, Vetterli M. The contourlet transform: an efficient directional multiresolution image representation. IEEE Trans Image Process 2005;14:2091–106.
[8] Qu XB, Zhang WR, Guo D, Cai CB, Cai SH, Chen Z. Iterative thresholding compressed sensing MRI based on contourlet transform. Inverse Probl Sci En 2010;18:737–58.
[9] Gho SM, Nam Y, Zho SY, Kim EY, Kim DH. Three dimension double inversion recovery gray matter imaging using compressed sensing. Magn Reson Imaging 2010;28:1395–402.
[10] Qu X, Cao X, Guo D, Hu C, Chen Z. Combined sparsifying transforms for compressed sensing MRI. Electron Lett 2010;46:121–2.
[11] Chen YM, Ye XJ, Huang F. A novel method and fast algorithm for MR image reconstruction with significantly under-sampled data. Inverse Probl Imag 2010;4:223–40.
[12] Ravishankar S, Bresler Y. MR image reconstruction from highly undersampled k-space data by dictionary learning. IEEE Trans Med Imaging 2011;30:1028–41.
[13] Aharon M, Elad M, Bruckstein A. K-SVD: an algorithm for designing overcomplete dictionaries for sparse representation. IEEE Trans Signal Process 2006;54:4311–22.
[14] Peyré G, Le Pennecc E, Dossal C, Mallat S. Geometrical image estimation with orthogonal bandelet bases. In: Van De Ville D, Goyal K, Papadakis M, editors. SPIE Wavelets XII, San Diego CA, USA. Bellingham WA: SPIE; 2007. p. 1–10.
[15] Le Pennec E, Mallat S. Bandelet image approximation and compression. Multiscale Model Sim 2005;4:992–1039.
[16] Peyré G, Mallat S. Surface compression with geometric bandelets. Acm Trans Graphic 2005;24:601–8.
[17] Trzasko J, Manduca A. Highly undersampled magnetic resonance image reconstruction via homotopic l0-minimization. IEEE Trans Med Imaging 2009;28:106–21.
[18] Baraniuk R, Choi H, Neelamani R, Ribeiro V, Romberg J, Guo H, et al. Rice wavelet toolbox. http://dsp.rice.edu/software/rice-wavelet-toolbox. 2009.
[19] Yang J, Zhang Y, Yin W. A fast TVL1-L2 minimization algorithm for signal reconstruction from partial Fourier data. Technical report, Rice University. ftp://ftp.math.ucla.edu/pub/camreport/cam09-24.pdf. 2009.
[20] Baker CA, King K, Liang D, Ying L. Translational-invariant dictionaries for compressed sensing in magnetic resonance imaging. In: Wernick M, Tian J, editors. 8th IEEE International Symposium on Biomedical Imaging-ISBI'11, Chicago, IL, USA. Piscataway, NJ:IEEE; 2011. p. 1602–5.
[21] Vasanawala S, Murphy M, Alley M, Lai P, Keutzer K, Pauly J, et al. Practical parallel imaging compressed sensing MRI: summary of two years of experience in accelerating body MRI of pediatric patients. In: Wernick M, Tian J, editors. 8th IEEE International Symposium on Biomedical Imaging-ISBI'11, Chicago, IL, USA. Piscataway, NJ: IEEE; 2011. p. 1039–43.
[22] Yang JF, Zhang Y, Yin WT. A fast alternating direction method for TV l1-l2 signal reconstruction from partial fourier data. IEEE J Sel Topics Signal Process 2010;4:288–97.
[23] Block KT, Uecker M, Frahm J. Undersampled radial MRI with multiple coils. Iterative image reconstruction using a total variation constraint. Magn Reson Med 2007;57:1086–98.
[24] Song JY, Liu YH, Gewalt SL, Cofer G, Johnson GA, Liu QH. Leastsquare NUFFT methods applied to 2-D and 3-D radially encoded MR image reconstruction. IEEE Trans Biomed En 2009;56:1134–42.
[25] Kashyap S, Yang ZL, Jacob M. Non-iterative regularized reconstruction algorithm for non-Cartesian MRI: NIRVANA. Magn Reson Imaging 2011;29:222–9.
[26] Smith DS, Welch EB. Non-sparse phantom for compressed sensing MRI reconstruction. International Society for Magnetic Resonance in Medicine 19th Scientific Meeting-ISMRM'11. Montreal, Canada; 2011. p. 2845.
[27] Lustig M. Sparse MRI toolbox. http://www.stanford.edu/~mlustig/ SparseMRI.html. 2011.
[28] Hu C, Qu X, Guo D, Bao L, Chen Z. Wavelet-based edge correlation incorporated iterative reconstruction for undersampled MRI. Magn Reson Imaging 2011;29:907–15.
[29] Majumdar A, Ward RK. An algorithm for sparse MRI reconstruction by Schatten p-norm minimization. Magn Reson Imaging 2011;29:408–17.