25 August 2020 in IEEE Transactions on Computational Imaging
For compressed sensing magnetic resonance imaging (CS-MRI) that utilize sparse representations, the regularization parameter establishes a trade-off between sparsity and data fidelity. While convergence to the desired solution is slow for mean squared error (MSE) optimal constant regularization, continuation using decreasing parameter values enables faster convergence. To derive an explicit rule for continuation, we propose an intermediate step optimization that involves maximization of the l2-norm of the gradient descent update. This is achieved by inclusion of an extra prior to the CS-MRI cost function. The solution is obtained using an alternating minimization approach in which the first sub-problem deals with the sparse regularization using the previously computed parameter value, and the second sub-problem aims at estimation of the parameter value to be used in the succeeding iteration. The solution to the second sub-problem is computed using standard root finding methods. Irrespective of the initial choice of the regularization parameter, we show that application of this continuation based proximal approach enables faster convergence to the desired solution.
2 July 2024 in European Journal of Radiology
7 June 2017 in Magnetic Resonance in Medicine
28 December 2017 in IEEE Transactions on Computational Imaging
Low sampling density and reduced signal-to-noise ratio resulting from accelerated magnetic resonance data acquisition calls for sparsity promoting regularization for improved reconstruction. The simplest optimization strategy relies on thresholded Landweber algorithm that results in reconstructions with slow rate of convergence. A varying threshold can accelerate the rate of convergence, but does not guarantee low steady-state errors. In this paper, we propose a statistical approach to optimize the tradeoff between steady-state reconstruction error and speed of convergence, by iteratively adapting the regularization parameter. Analogous to the generalized discrepancy principle that implicates the nonnegativity of discrepancy level (absolute difference between l2-norms of consistency and perturbation errors) as provided by regularized reconstruction, the proposed method of adaptation entails reduction of the discrepancy level in l1-norm of the consistency and sparse approximation errors. In discrete form, this is deduced from the convergence properties of Landweber updates soft thresholded using monotonically decreasing thresholds. With very large number of iterations, the thresholds approach a lower bound determined by the noise level and adaptation function. Further improvements in image quality can be achieved by separately adapting the thresholds in each wavelet subband. Application of adaptive regularization combined with over-relaxation to parallel MRI shows significant improvements in both speed and image quality.
7 March 2018 in Magnetic Resonance in Medicine
1 November 2015 in Concepts in Magnetic Resonance Part A
SMSreekanth Madhusoodhanan