CS++: Compressed Sensing & Beyond
Tuesday 21 April 2009
Room 313BC 16:00-18:00


Pablo Irarrazaval and Krishna S. Nayak

16:00  377. Accelerating SENSE Using Distributed Compressed Sensing
    Dong Liang1, Kevin f. King2, Bo Liu3, Leslie Ying1
Dept. of Electrical Engineering and Computer Science, Univ. of Wisconsin-Milwaukee, Milwaukee, WI, USA; 2Global Applied Science Lab, GE Healthcare, Waukesha, WI, USA; 3MR Engineering, GE Healthcare, Waukesha, WI, USA
    Most existing methods apply compressed sensing (CS) to parallel MRI as a regularized SENSE reconstruction, where the regularization function is the L1 norm of the sparse representation. However, the CS conditions such as incoherence are not necessarily satisfied. To address the issue, a method is proposed which first reconstructs a set of aliased images from all channels simultaneously using distributed CS (DCS), and then the final image using Cartesian SENSE. The results on a set of eight-channel data show that the proposed method is able to achieve a higher reduction factor than the existing methods.
16:12 378. Distributed Compressed Sensing for Accelerated MRI
    Ricardo Otazo1, Daniel K. Sodickson1
Center for Biomedical Imaging, NYU School of Medicine, New York, NY, USA
    A framework for combining parallel imaging with compressed sensing is presented using the theory of distributed compressed sensing which extends compressed sensing to multiple sensors using the principle of joint sparsity. We present a greedy reconstruction algorithm named JOMP (Joint Orthogonal Matching Pursuit) that uses intra- and inter-coil correlations to jointly sparsify the multi-coil image instead of sparsifying the individual images. We show that for a sufficient number of coils, the number of measurements required by JOMP-PMRI to reconstruct a truly sparse image is very close to the image sparsity level. The performance of JOMP-PMRI with compressible images is assessed with a simulated brain image to show feasibility of higher accelerations with increasing number of coils.
16:24 379. L1 SPIR-IT: Autocalibrating Parallel Imaging Compressed Sensing
    Michael Lustig1, Marcus Alley2, Shreyas Vasanawala2, David L. Donoho3, John Mark Pauly1
Electrical Engineering, Stanford University, Stanford, CA, USA; 2Radiology, Stanford University; 3Statistics, Stanford University
    A detailed approach of combining auto-calibrating parallel imaging (acPI) with compressed sensing (CS) is presented. The acquisition and the reconstruction are carefully optimized to meet the requirements of both methods in order to achieve highly accelerated robust reconstructions. Poisson-disc sampling distribution is used to achieve the required incoherency for CS and uniform density for acPI. A novel L1-wavelet penalized, iterative reconstruction (L1 SPIR-iT) is used to enforce consistency with the calibration and data acquisition, and in addition, joint sparsity of the reconstructed coil images. High quality in vivo, 5-fold accelerated reconstruction using only 4 coils is demonstrated.
16:36 380. L1-Norm Regularization of Coil Sensitivities in Non-Linear Parallel Imaging Reconstruction
    Carlos Fernández-Granda1,2, Julien Sénégas3
École des Mines, Paris, France; 2Universidad Politécnica de Madrid, Spain; 3Philips Research Europe, Hamburg, Germany
    Joint estimation of the coil sensitivities and the image in parallel imaging can suppress aliasing more effectively than methods based on low-resolution sensitivity estimates. We propose a joint estimation approach related to Compressed Sensing that exploits the sparsity of the coil sensitivities in k-space and in a base of Chebyshev polynomials within a greedy scheme to solve the ill-posed reconstruction problem. In vivo data reconstructions are presented and compared to results obtained with Generalized SENSE and Joint SENSE.
16:48 381. SPArse Reconstruction Using a ColLEction of Bases (SPARCLE)
    Ali Bilgin1,2, Onur Guleryuz3, Theodore P. Trouard2,4, Maria I. Altbach2
Electrical and Computer Engineering, University of Arizona, Tucson, AZ, USA; 2Dept. of Radiology, University of Arizona, Tucson, AZ, USA; 3Dept. of Electrical Engineering, Polytechnic Institute of NYU, Brooklyn, NY, USA; 4Biomedical Engineering, University of Arizona, Tucson, AZ, USA
    We introduce a new sparse reconstruction framework where sparsity is enforced in a collection of bases rather than a single one. Results indicate that this new framework yields significantly improved reconstruction quality.
17:00 382. Ultra-High Resolution 3D Upper Airway MRI with Compressed Sensing and Parallel Imaging
    Yoon-Chul Kim1, Shrikanth S. Narayanan1, Krishna S. Nayak1
Ming Hsieh Department of Electrical Engineering, University of Southern California, Los Angeles, CA, USA
    Ultra-high resolution 3D imaging of the vocal tract can provide insight into the shaping that occurs during complex speech articulation. The combined use of compressed sensing (CS) and parallel imaging is investigated to maximize spatial resolution while maintaining scan-time appropriate for a single sound production task (~7 seconds). Compared to conventional reconstructions, boundary depiction was improved by using high-resolution phase constraints, sensitivity encoding, and regularization based on total variation and the l1-norm of the wavelet transform. Eight-fold acceleration was achieved leading to 1.33x1.33x1.33 mm3 resolution and 7-second scan time.


17:12 383. Highly-Accelerated First-Pass Cardiac Perfusion MRI Using Compressed Sensing and Parallel Imaging
    Ricardo Otazo1, Daniel Kim1, Daniel K. Sodickson1
Center for Biomedical Imaging, NYU School of Medicine, New York, NY, USA
    Compressed sensing and parallel imaging are combined into a single joint reconstruction paradigm named k-t Parallel-Sparse for highly accelerated first pass cardiac perfusion imaging. The method exploits the joint sparsity in the sensitivity-encoded images to achieve higher accelerations than for coil-by-coil sparsity alone, and it does not require dynamic training data. We demonstrate the feasibility of high in vivo acceleration factors of 8 and 12 and assess the effect of respiratory motion.
17:24 384. Motion Estimated and Compensated Compressive Sensing Dynamic MRI Under Field Inhomogeneity
    Hong Jung1, Jaeseok Park2, Jong Chul Ye3
KAIST, Daejon, Korea; 2Yonsei Univ. medical center, Korea; 3KAIST, Korea
    Recently, we proposed a compressed sensing dynamic MR technique called k-t FOCUSS that extends the conventional k-t BLAST/SNESE by exploiting the sparsity of x-f signal. Especially, we found that when a fully sampled reference frame is available more sophisticated prediction methods such as RIGR and motion estimation and compensation (ME/MC) can significantly sparsify the residual and improve the overall reconstruction quality. Among these, ME/MC is especially useful since it can be used for arbitrary trajectories such as radial and spiral. However, our extensive experiments with non-cartesian trajectory have demonstrated that there exist technical issues in applying the ME/MC to non-cartesian trajectory due to the field inhomogeneities. This paper showed that if the ME/MC is done in magnitude image domain and the lost phase is compensated from the current frame estimate, the field inhomogeneity problem can be significantly alleviated. Furthermore, we showed that the introduction of half-pel ME/MC and intra block mode within the estimation loop can improve the overall reconstruction quality of compressed sensing dynamic MRI.
17:36 385. Fast Relaxation Parameter Mapping from Undersampled Data
    Mariya Doneva1, Christian Stehning2, Peter Börnert2, Holger Eggers2, Alfred Mertins1
University of Luebeck, Luebeck, Germany; 2Philips Research Europe, Hamburg, Germany
    The quantitative assessment of MR parameters like T1, T2, ADC, etc. requires the acquisition of multiple images of the same anatomy, which results in long scan times. However, these data can be described by a model with only a few parameters and in that sense they are highly compressible. Thus, Compressed Sensing (CS) could be applied to accelerate the data acquisition. In this work we introduce a model-based reconstruction from undersampled data, which performs simultaneous image reconstruction and parameter mapping and demonstrate it for the example of T1 mapping.
17:48 386.

Quality Index for Detecting Reconstruction Errors Without Knowing the Signal in L0-Norm Compressed Sensing

    Carlos A. Sing-Long1,2, Cristian A. Tejos1,2, Pablo Irarrazaval1,2
Departamento de Ingenieria Electrica, Pontificia Universidad Catolica de Chile, Santiago, R.M., Chile; 2Biomedical Imaging Center, Pontificia Universidad Catolica de Chile, Santiago, R.M., Chile
    Compressed Sensing allows reconstructing signals, if they are sparse in some representation, from some of its Fourier coefficients. The reconstruction conditions are stated in terms of the support size of the signal. Since it is generally unknown, it is impossible to determine if there are reconstruction errors due to high undersampling rates. Our work introduces a modified fixed-point solver for a continuous approximation of the l0-norm and an index which shows high correlation with the reconstruction error. This index does not need any a priori information and may be used to determine if the undersampling rate needs to be reduced.