Statistical Wavelet Structure Based MRI Compressed Sensing Reconstruction Using a Hidden Markov Tree Model
Enhao Gong,Xiao Wang,Kui Ying
2011-01-01
Abstract:Introduction: Compressed Sensing (CS) [1, 2] is an acceleration technique that enables reduction in the sampling and recently applied in MRI [3] based on the sparsity of medical images. Conventional CS recovery techniques are based on simplistic sparsity of signal,which uses uniform L1-norm penalty regardless of whether the coefficients contribute to significant information for pathological diagnosis. This results in reconstruction errors, like blurring edges. In this study, we propose a new algorithm using Hidden Markov Tree model [4] to extract statistical structural information in wavelet domain, such that the sparsity is constrained selectively. In this way important coefficients are enhanced and artifacts are further reduced. Experimental results on phantom simulation and in-vivo show more preserved details and better artifact reduction. Methods: I. Model Based Compressed Sensing: Conventional CS algorithms approximate MRI images based on simplistic sparse assumption and use a uniform L1-norm penalty for all coefficients [1~3] without considering whether they contribute to fine details in image domain or wavelet domain. Several new studies [5, 6] demonstrated that there are significant performance gains by exploiting realistic models beyond simplistic sparsity. II. Modeling Wavelet Structure: MRI images have sparse wavelet expansions and significant wavelet coefficients exhibit properties that can be modeled by wavelet quad-tree [7] (Figs. 1.a, 1.b): Persistence Property (significance persists across the scales), Scale-dependence (persistence becomes stronger to finer scale) and Decaying Magnitudes (magnitude drops exponentially to finer scale). It should be noted that the wavelet domain structure cannot be defined just by values since they obey mixture distribution and threshing cannot differentiate noise with small but important signal. Thus, statistical methods and models are required. III. Hidden Markov Tree (HMT) Model: We used a Hidden Markov Tree to model the probability density function of each wavelet coefficient as a Gaussian mixture density [8] with a hidden binary state , which in our application indicates whether coefficients are ‘Negligible’ ( = ) or ‘Significant’ ( = ). The properties of 2D wavelet quad-tree are captured by a Hidden Markov Quad-Tree Model in which the Gaussian mixture distribution and Transition Matrices between states (See Figs. 1.c, 1.d) are functions of several parameters including the scale J in wavelet-tree. IV. Regulating CS Reconstruction with Model-based Structure: We estimated HMT parameters Θ( ) = , , , , , , , , , and the probability of hidden states using Expectation-Maximization (EM) algorithm [7, 9] and a Tree-Viterbi algorithm, which are efficient and enjoy just linear complexity. Then we employed statistical matrices based on state probabilities of being ‘Significant’ to further regulate iterative optimization and statistically penalize coefficients that are ‘Negligible’ (small or isolated) to wavelet structure. ( : wavelet operator. : partial Fourier transform. : HMT hidden states. Parameters Θ are computed from measured k-space and reconstructed image ). ( ) ∝ ( = | ,Θ ( )) ( ) = ( ) , . . ‖ − ‖ <