Adaptive Adjustment of Relaxation Parameters for Algebraic Reconstruction Technique and its Possible Application to Sparsity Prior X-ray CT Reconstruction

TL;DR

This paper evaluates adaptive relaxation parameters in iterative algorithms for sparsity-prior X-ray CT reconstruction, demonstrating improved convergence and image quality in compressed sensing environments, especially with noise.

Contribution

It introduces a data-driven relaxation adjustment method for ART in sparsity-prior CT, analyzing its performance against traditional methods in simulated and real noisy environments.

Findings

Data-driven relaxation improves reconstruction quality over traditional methods.

Thresholding is necessary for convergence when data are noisy.

Adaptive relaxation accelerates convergence in compressed sensing CT reconstruction.

Abstract

In this paper, we systematically evaluate the performance of adaptive adjustment of the relaxation parameters of various iterative algorithms for X-ray CT reconstruction relying on sparsity priors. Sparsity prior has been found to be an efficient strategy in CT reconstruction where significantly fewer attenuation measurements are available. Sparsity prior CT reconstruction relies on iterative algorithms such as the algebraic reconstruction technique (ART) to produce a crude reconstruction based on which a sparse approximation is performed. Data driven adjustment of relaxation has been found to ensure better convergence than traditional relaxation for ART. In this paper, we study the performance of such data driven relaxation on a (CS) compressed sensing environment. State-of-the-art algorithms are implemented and their performance analyzed in regard to conventional and data-driven relaxation. Experiments are performed both on simulated and real environments. For the simulated case, experiments are conducted with and without the presence of noise. Correlation coefficients, root mean square error, structural similarity index and perceptual dissimilarity metric were used for the quantitative comparisons of the results. Experiments reveal that data driven relaxation also ensures overall better quality reconstruction in a CS environment compared to traditional relaxation. However, when the data are corrupted by noise, inconsistencies emerge in the convergence unless a threshold is imposed on the maximum amendments. Data driven relaxation seems a logical choice to more rapidly reach the solution. In a compressed sensing environment, especially when the data are corrupted by noise, a threshold to specify the maximum amendments needs to be specified. Experimentally, we have set the threshold as 20% of the previous value and thus have ensured more consistency in the convergence.