A method to obtain a high-resolution 2D image from a series of subpixel-encoded low-resolution 2D scans is presented. The phaseless character of this encoding allows unknown signal phase fluctuations to be easily discarded and makes the method a promising alternative to k-space segmenting in motion-sensitive experiments such as diffusion MRI. Based on an analogy with the structured illumination method used in superresolution optics, important improvements have been introduced that reduce artefacts caused by k-space truncation and magnetic field inhomogeneity. The utility of the method is demonstrated by a 3-fold resolution enhancement of diffusion-weighted EPI of human head.
A single-shot imaging sequence such as EPI is preceded by a tagging module consisting of two RF pulses separated by a gradient pulse of integral $$$ k’ $$$ (Fig.1). Low resolution single shot images are reconstructed using a k-space window $$$ W(k) $$$ to limit the Gibbs ringing, and transformed back to k-space following the magnitude calculation to remove the random phase. The contributions from the k-space bands shifted by $$$ \pm k’ $$$, introduced by the tagging, are separated from the central band based on a phase cycle of the RF pulses (which shifts the pattern in space), shifted to their proper positions, Fourier transformed and summed up. The result is equal to the FT of the true k-space data multiplied by three shifted copies of the low-resolution k-space window:
$$ S_{\text{final}}(k) = S_{\text{true}}(k) \times \left[ W(k+k’) + W(k) + W(k-k’) \right] $$
The resolution gain is thus maximally threefold when W is rectangular and k’ is set to its width. This, however, leaves Gibbs ringing on the low-resolution images, which propagates to the final SR image. Results are better with a sine-trapezoidal window and the three bands properly overlapping to produce a flat top of the final filter. The trade-off is a slightly reduced resolution gain. Due to magnetic field inhomogeneity, the sinusoidal tagging pattern is distorted with a local phase shift equal to the spin phase accrual between the two RF pulses. This phase shift propagates to the reconstructed k-space sidebands and causes ringing artefacts. However, the shift can be calculated based on a magnetic field map and taken into account during the reconstruction. For that purpose, the phase of partial images reconstructed from each band is corrected by the calculated position-dependent shift (with opposite sign for opposite-side bands) prior to the addition of the images.
The advantage of partial band overlapping is demonstrated in Fig.2. With adjacent bands (and maximum resolution enhancement) the low resolution images must be reconstructed without apodization and therefore show Gibbs ringing, which propagates to the final SR image. An attempt to reduce ringing by k-space apodization (hamming window) makes the image even worse due to periodicity of the effective filter caused by the replication in Eq. 1. However, with partial band overlapping (and a slightly reduced resolution gain), a filter can be applied that combines to a flat-top widow upon replication and effectively reduces the ringing phenomenon. The artefact caused by field-inhomogeneity and the related tagging distortion, as well as the advantage of the field map based reconstruction is shown in Fig. 3. Without correction, ringing appears in off-resonance regions. The correction is effective everywhere except for regions where local field gradients are strong enough to cause intra-voxel dephasing. The utility of phaseless subpixel encoding in demonstrated by a diffusion-weighted (b=1000 s/mm2) EPI acquisition carried out on a 3T whole body system (Philips Achieva) equipped with 30 mT/m gradients (Fig. 4). A healthy male adult volunteered for the study in compliance with the institution guidelines. Single-shot images with resolution 3.2 mm (readout direction, horizontal) and 1.2 mm (phase encoding, vertical) were required with three shifts of the tagging pattern and used to reconstruct a super-resolved image with isotropic resolution of 1.2mm. This reconstruction is artefact-free despite apparent phase fluctuations between the shots. With classic segmented (interleaved) EPI the same phase fluctuation pattern would cause significant ghosting.