Alternatives to the use of the DFT in MRI and spectroscopic reconstructions

Abstract

Standard and functional magnetic resonance imaging (MRI and f MRI) make of use the two-dimensional (2D) discrete Fourier transform (DFT). Many MR spectroscopic techniques use the 1D DFT. Experimental or time constraints frequently require that the DFT be applied to finite-length (truncated) data sequences. Truncation is essentially a windowing of the data and introduces artifacts and resolution loss in images or spectra. A number of alternative reconstruction algorithms have been proposed to counteract these problems. These algorithms attempt to model the known data and use the modeling information to implicitly or explicitly extrapolate the data to overcome the windowing. One modeling approach, the Transient Error Reconstruction Algorithm (TERA), uses an autoregressive moving average method to recover the missing data. In this article, we briefly discuss variants of the TERA algorithm and development of neural networks to take better account of the differing data properties of MR data sets. Our success with neural networks in fMRI reconstruction has led us to challenge some of the standard approaches to validating MR algorithms and develop our own. These new approaches include k-space phantom generation and automated computer observer (ROC analysis) to evaluate algorithms in terms of their clinical relevance. We have also developed an upgraded image quality measure based on Daly's Visual Differences Predictor. This models the ability of the human visual system to detect significant differences between images produced by different MRI reconstruction algorithms. We also present a protocol for generating Shepp-Logan phantoms which avoids the introduction of the high-frequency k-space data distortion present in the existing approach. © 1997 John Wiley & Sons, Inc. Int J Imaging Syst Technol, 8, 558–564, 1997