Single-shot GE-EPI is widely used in fMRI. However, it is susceptible to field inhomogeneity induced geometric distortions, therefore retrospectively unwarping of the single-shot GE-EPI data is important for accurately co-registering functional images with structural images. A commonly used unwarping technique is based on field map of image, which is often acquired at the beginning of experiments. However its reliability may be compromised by motions and physiological activities of subjects during experiments, and thus the field map is desired at each time point of a dynamic fMRI measurement series. The field map in distorted coordinates can be obtained from the phase data by a proper phase unwrapping procedure ¹. Alternatively, it can also be calculated by integrating the off-resonance gradient maps which can be obtained from analyzing the k-space echo shifting of each voxel ^2-3, or from calculating the phase differences of two adjacent voxels ³. The aim of this work is to qualitatively and quantitatively compare the performance of the above-mentioned three reference-free unwarping methods on human brain data.

The data processing schemes of the three unwarping methods are shown in Fig.1.

Human brain data of one healthy volunteer were acquired on a Siemens 3T Prisma scanner with informed consent, using a 20-channel head coil and the built-in BOLD sequence. Three orthogonal orientations were scanned with fat suppression, 24 slices for each. FOV=220×220mm² with slice thickness=5mm, TR/TE=3000ms/30ms, acquisition matrix=64×64, SW=2440Hz/voxel, echo spacing=490μs. Referential TSE images were scanned with TR/TE=5000ms/104ms and acquisition matrix=256×256.

The unwarping results of representative slices are shown in Fig.2. Two slices are selected for each orientation, one with relatively large deviation in the estimated off-resonance frequency values (Δf) obtained from the three field map calculation procedures, and the other with minor difference. Visual inspection suggests that most of the distortions can be corrected by all three methods, and generally k-space filtering method produced relatively more stable unwarping results than the other two methods.

Furthermore, quantitative investigation were conducted.

First, the SNR of selected areas were calculated. Figure 3 shows that unwarping procedures led to the reduction of SNR and apart from some specific slices, the SNR obtained from three unwarping methods were very close.

Secondly, the disparity of the field maps obtained from three methods were evaluated by calculating the root-mean-squares deviation of the estimated Δf between each two methods in the labeled areas of the previous study, and the results are shown in Fig.4. For traversal orientation, since vertical stretching existed for most of the slices and k-space filtering method generally achieved better unwarping performance, Δf obtained by it presented larger deviation from the values obtained by the other two methods. For coronal and sagittal orientations, since distortions were regional, three methods achieved similar unwarping quality for most of the slices except for some small differences at regional profiles. The estimated Δf were also very close, especially for what was obtained by k-space filtering and phase difference methods because they shared similar gradient map integration strategy.

Thirdly, the variation tendency of the estimated off-resonance frequency values from the data acquired by different coils was measured. Based on Fig. 4, for each orientation one slice with relatively small deviation between 3 field maps was selected to study. Two areas were measured on them, one was relatively homogeneous and immune to distortions, the other with relatively large field gradients and susceptible to distortions. Figure 5 shows that the homogeneous regions resulted in smaller variations of Δf compared with the inhomogeneous regions. The k-space filtering and phase unwrapping methods gave rise to larger variations than phase difference method because both the determination of accurate k-space shifts and phase values depend more closely on the SNR of acquired data.

Both qualitative and quantitative experimental results demonstrate that k-space filtering method can provide more stable unwarping performance and SNR level than the other two methods attributed to two reasons. First, the off-resonance gradient map obtained is of higher SNR than phase difference method; secondly, the underlying local linear fitting involved can help provide more accurate field map than phase unwrapping method, especially at regions with large field gradients. Because the estimated off-resonance frequency values vary with coil, combining the field maps obtained from different coils with corresponding sensitivity distribution improved the accuracy of the estimated field map by making it less dependent to the selection of initial seed in related region-growing process ⁴, which would facilitate the automatic processing of large amounts of data in fMRI.