Evaluation of SLIce Dithered Enhanced Resolution Simultaneous MultiSlice (SLIDER-SMS) for human fMRI
Introduction
Advancements in functional magnetic resonance imaging (fMRI) have pushed the limits of achievable spatial resolution enabling sub-millimeter imaging of human brain function down to the columnar level (Yacoub et al., 2007, Yacoub et al., 2008, De Martino et al., 2015, Nasr et al., 2016, Zimmermann et al., 2011) and laminar level (De Martino et al., 2015, Harel et al., 2006, Muckli et al., 2015, Olman et al., 2012, Polimeni et al., 2010, Zhao et al., 2006, Zimmermann et al., 2011). These studies were all performed at ultra-high field (7T) since sub-millimeter fMRI is generally regarded as not feasible at lower field strengths due to SNR and BOLD CNR constraints. However, given that 7T scanners are relatively few in the human research setting, there is strong motivation to further develop techniques to enable sub-millimeter functional imaging at lower, more readily available field strengths. Of course such developments would also be useful at ultra-high field strengths as well given that the field has yet to achieve routine human fMRI at the spatial scale of more invasive techniques such as optical imaging (Lu et al., 2010).
Super-resolution techniques utilizing sub-voxel spatial shifts in the slice direction (Greenspan et al., 2002; Greenspan 2008; Setsompop et al., 2015) have shown great promise for increasing both resolution and SNR efficiency but are not without their challenges. Super-resolution techniques like SLIce Dithered Enhanced Resolution (SLIDER; Setsompop et al., 2015) improve resolution in two ways. First, SLIDER increases the Nyquist sampling frequency by acquiring S sets of high SNR, low-resolution, thick slice data (e.g. 1×1×Sth mm3) with Sth/S mm shifts between each set along the slice direction (where S is the SLIDER factor and Sth is the thickness of the excited slice). Because slices are spatially encoded, in contrast to the frequency and phase encoded dimensions (Mayer and Vrscay 2007), this form of oversampling resolves and unaliases additional high spatial frequency information that is otherwise lost when using traditional inter-slice spacing equal to the slice thickness. However, this high spatial frequency information is dampened relative to lower spatial frequencies due to the point spread function (PSF) of the thicker slices used. Thus, the second way SLIDER improves resolution is by applying a deblurring algorithm along the slice axis. Given the slice profile (e.g. PSF) is known, the deblurring algorithm can estimate what the underlying data would be had a thinner slice, of thickness Sth/S mm, been used.
Because SLIDER acquires S sets of low-resolution, anisotropic thick slice data to more finely sample the image volume along the slice axis, the total volume acquisition time is S times that of the original low-resolution data. Analogous to conventional thin slice, isotropic acquisitions, imaging volume times can be very long (>> 3000 seconds) given it can take well over 100 slices to cover the brain at sub-millimeter resolution. This is undesirable due to the relatively low SNR efficiency (Feinberg et al., 2010, Smith et al., 2013, Xu et al., 2013) and loss of high temporal frequency BOLD information (Chang et al., 2013, Chen et al., 2015, Vu et al., 2016). Fortunately, recent advancements in Simultaneous Multi-Slice (SMS) imaging (Larkman et al., 2001, Moeller et al., 2010) with blipped-CAIPI (Setsompop et al., 2012), in combination with in-plane under-sampling techniques (Sodickson and Manning, 1997, Pruessmann et al., 1999, Griswold et al., 2002; Griswold et al., 2006; Talagala et al., 2013; Ugurbil et al., 2013; Polimeni et al., 2016) have substantially improved scan time and SNR efficiency (Ugurbil et al., 2013, Vu et al., 2015, Vu et al., In Press). Thus the combination of SLIDER and SMS should provide suitable temporal resolution and SNR efficiency for mesoscale imaging at both 3T and 7T.
Besides long imaging volume times, another common concern regarding super-resolution techniques is the residual blurring and noise amplification resulting from post-acquisition deblurring (Lam, 2003, Otazo et al., 2009). In the absence of noise, the deblurring component of super-resolution techniques can perfectly reconstruct high-resolution images from multiple low-resolution images. However, in practice, noise levels in the images can be quite significant leading to high frequency noise amplification and banding artifacts. To mitigate the effect of noise, prior studies have resorted to regularization of the deblurring algorithm, which effectively trades noise amplification for residual blurring. How this tradeoff manifests in the context of SLIDER fMRI and subsequently generated functional activation maps remains unclear. In some SLIDER fMRI applications, the increase in resolved high spatial frequency information may be sufficient without deblurring and may warrant forgoing the post-acquisition deblurring of individual image volumes to achieve maximal BOLD CNR. More specifically, the benefits of deblurring may be limited in the context of fMRI, particularly for high-resolution sub-mm acquisitions, given that the vascular PSF is estimated to be on the order of ~1.5–4 mm (Engel et al., 1997, Parkes et al., 2005, Shmuel et al., 2007) and thus would dominate blurring effects due to SLIDER. Fortunately, contrast subtraction techniques ubiquitous to analysis of fMRI data are known to serve as specificity or PSF enhancing operations (Yacoub et al., 2007, Yacoub et al., 2008, Muckli et al., 2015). This concept is analogous to super resolution techniques used in optical imaging (e.g. spectral precision distance microscopy (Lemmer et al., 2008)) where detailed structural information can be obtained far beyond the diffraction limit by selectively activating spatially sparse subsets of objects at any one time and precisely locating the positions of these objects. Similarly in fMRI, individual columns or layers can be sparsely activated in space at different time points – allowing for their precise location and organizational structure to be determined even in the presence of relatively broad PSFs.
Here, we evaluated the complimentary combination of slice dithering and SMS for high-resolution whole brain fMRI. Specifically, we compared SLIDER-SMS to SMS alone at low-resolution, thick slice acquisition (1.25×1.25×2.5 mm3) as well as to thin slice, high isotropic resolution acquisition (1.25×1.25×1.25 mm3) using the following metrics: data smoothness, k-space energy at various spatial frequency bins (along the slice axis), tSNR, and BOLD CNR. The effect of the deblurring post-processing step with various levels of regularization was also investigated and revealed that SLIDER without deblurring (which we refer to as SLIDER-XD) may actually be preferred for optimal BOLD CNR while still retaining significantly more high-spatial frequency information than conventional low-resolution, anisotropic acquisitions. Using these gains in CNR, we show it is feasible to perform fMRI experiments using 0.65 mm isotropic nominal resolution data acquired at 3T and 0.45 mm data acquired at 7T.
Section snippets
Methods
Data were acquired from four healthy subjects on a Siemens 3T Trio using the standard 32 ch head coil. During the 2D gradient echo EPI fMRI scans, subjects viewed three 96 s runs of flashing checkerboard stimulus (30 s period) for each scan protocol (SMS-5 and SLIDER-2 SMS-5). Imaging parameters were: 1.25 mm isotropic (nominal; 2.5 mm excitation thickness for SLIDER=2); FOV = 210×210×137.5 mm3; PF=6/8; TE=45 ms; TR=3000 ms (1500 ms per dithered volume); Flip angle=84° (72° for SLIDER=2); PE
Results & discussion
Fig. 2 shows the coronal cross section (of the 1.25 mm axially acquired slices) of a representative subject's low-resolution (thick slice), HR (thin slice), HRb, SLIDER-XD, and SLIDER data deblurred with various λ values. Since HRb was blurred in post processing, the original HR image is recovered perfectly without regularization (λ=0). However, as with most super-resolution techniques in the presence of noise, deblurring SLIDER without regularization results in high spatial frequency noise
Conclusions
We evaluated the synergistic combination of SLIDER and SMS in high-resolution whole brain fMRI and found that SLIDER-SMS can generate very high-resolution, high CNR fMRI data at both 3T and 7T. The BOLD CNR of SLIDER was significantly greater than that of HR and even HRb due to the linear relationship between voxel volume and SNR (as opposed to square-root when blurring in post processing) as well as the greater BOLD sensitivity with thicker slices. We also found it advantageous to forgo
Acknowledgements
NIH BRAIN Initiative grant − 1R24MH106096.
References (49)
- et al.
Whole-head rapid fMRI acquisition using echo-shifted magnetic resonance inverse imaging
Neuroimage
(2013) - et al.
Evaluation of highly accelerated simultaneous multi-slice EPI for fMRI
Neuroimage
(2015) - et al.
Human ocular dominance columns as revealed by high-field functional magnetic resonance imaging
Neuron
(2001) - et al.
Influence of fMRI smoothing procedures on replicability of fine scale motor localization
Neuroimage
(2005) - et al.
MRI inter-slice reconstruction using super-resolution
Magn. Reson. Imaging
(2002) - et al.
Combined imaging-histological study of cortical laminar specificity of fMRI signals
Neuroimage
(2006) - et al.
A motion direction map in macaque V2
Neuron
(2010) - et al.
Measuring information gain for frequency-encoded super-resolution MRI
Magn. Reson. Imaging
(2007) - et al.
Contextual Feedback to Superficial Layers of V1
Curr. Biol.: CB
(2015) - et al.
Superresolution parallel magnetic resonance imaging: application to functional and spectroscopic imaging
Neuroimage
(2009)