Spatial resolution, signal-to-noise ratio, and smoothing in multi-subject functional MRI studies
Introduction
Acquisition resolutions in typical functional MR imaging experiments today are approaching 3 × 3 × 3 mm3 voxel sizes. As more centers move to higher field strength (>1.5 T), one of the goals appears to be to increase this spatial resolution further. Higher spatial resolution can provide better localization, down to the level of the cortical columns in the human visual cortex (Menon et al., 1997) and in animals (Kim and Duong, 2002), and in a layer-specific manner in other cortical regions (Silva and Koretsky, 2002). The majority of human cognitive neuroscience studies are performed on group data, and this places certain demands and constraints on the data. Group analyses, and in many cases single-subject analyses, are performed using SPM (Friston et al., 1995) or other statistical packages that smooth the functional data using a typical filter with a FWHM of 8 mm. A wide range of motivations for smoothing have been given including to match the spatial scale of hemodynamic responses, to normalize the error distribution (by the Central Limit Theorem) to improve the validity of inferences based on parametric tests, and, finally, in the context of inter-subject averaging, it has been shown that it is necessary to smooth to project the data down to a scale where homologies in functional anatomy are expressed across subjects. In particular, recent work demonstrates that, using a wide spectrum of state-of-the-art registration algorithms, the misalignment in anatomic registration alone (without considering functional organizational variability), across multiple subjects, is of the order of 9 mm due to individual anatomic variability (Hellier et al., 2001, Hellier et al., 2003). Recent developments in non-linear spatial normalization might improve image registration accuracy, but, even if perfect alignment could be achieved based on anatomy (which is often impossible due to fundamental anatomic differences), there may still be additional variance introduced by differences in functional organization present in cortical regions.
Thus, if one accepts the premise that the data must be smoothed to accommodate the anatomic and functional variations, or for the statistical reasons given above, then the physics of MR suggest that acquiring at progressively higher resolutions will result in an unnecessary and irrecoverable loss of signal-to-noise ratio (SNR). In this paper, we sought to demonstrate that, if smoothing is to be employed, the data should be acquired at lower resolutions in order to maximize SNR since smoothing high-resolution data to a lower resolution does not recover the SNR loss that occurs when the data are acquired in high resolution. In addition, we investigated the increased signal loss from partial volume effects associated with moving to a larger voxel size and examined the impact of single-subject SNR improvement in multi-subject studies.
Section snippets
Theory
The equation for signal to noise ratio in a conventional 2D echo planar imaging pulse sequence can be written as follows:
where Δx, Δy, and Δz are the voxel dimensions, Nx and Ny represent the acquisition matrix size in the x and y directions respectively, Nave represents the number of averages, and υ represents the bandwidth. Assume that the last 4 terms are held constant, and define this constant as κ, such that the equation becomes
This equation
Materials and methods
Images were acquired on a 3 T Siemens Trio whole body imaging system with a quadrature head coil. All volunteers (N = 17:10 males, 7 females) gave informed consent, and this study was approved by the Yale IRB.
Results
Data from some subjects were discarded due to incomplete motor cortex coverage and/or audio equipment malfunction, resulting in useable motor data from N = 11 subjects, auditory data from N = 10 subjects, and visual data from all N = 17 subjects. Mean t values for individual subjects and z scores of mean percent signal changes across subjects were obtained within specific volumes of interest (VOIs), namely, those comprising the motor cortex, right and left auditory cortices, and the visual
Discussion
Noise in functional MRI can be divided into at least four categories: Johnson or thermal noise, physiological noise, system noise, and inter-subject variability (Desmond and Glover, 2002). The first noise source, the thermal noise, is directly impacted by the choice of spatial resolution, flip angle, TE, TR, acquisition matrix, and bandwidth. Physiological noise arises from fluctuations in basal cerebral metabolism, pulsatile blood flow, and motion, and as such it is signal-dependent (Krueger
Conclusions
In agreement with theory, the current study revealed that, for single-subject fMRI studies, there is a clear SNR advantage in acquiring data at a lower resolution whether or not spatial smoothing is to be employed. The LR benefit is further increased when one capitalizes on the possibility of acquiring more images per slice by using the optimal TR for the lower resolution. The advantage of the LR acquisition in single subjects is hampered in part by the fact that spatial resolution only impacts
Acknowledgments
Special thanks for support from NIH NS40497, NS38467, and EB00473.
References (15)
- et al.
Estimating sample size in functional MRI (fMRI) neuroimaging studies: statistical power analysis
J. Neurosci. Methods
(2002) - et al.
Mapping cortical columnar structures using fMRI
Physiol. Behav.
(2002) - et al.
Reducing inhomogeneity artifacts in functional MRI of human brain activation—Thin sections vs. gradient compensation
J. Magn. Reson.
(2000) - et al.
Comparison of physiological noise at 1.5 T, 3 T and 7 T and optimization of fMRI acquisition parameters
NeuroImage
(2005) - et al.
Functional MRI of auditory verbal working memory: long-term reproducibility analysis
NeuroImage
(2004) - et al.
Repetition time in echo planar functional MR imaging
Magn. Reson. Med.
(2001) - et al.
Hum. Brain Mapp.
(1995)
Cited by (40)
Reproducing the effect of hippocampal network-targeted transcranial magnetic stimulation on episodic memory
2022, Behavioural Brain ResearchAccuracy and precision in super-resolution MRI: Enabling spherical tensor diffusion encoding at ultra-high b-values and high resolution
2021, NeuroImageCitation Excerpt :Notably, both regularization and smoothing always lead to an effective resolution that is lower than the original or nominal resolution. As such, sampling directly at the effective resolution would increase both SNFR (Fig. 2) and SNR efficiency (Scouten et al., 2006) compared to sampling at the original resolution and performing post-hoc smoothing. For an evaluation of effective spatial resolution and therefore better assessment of SRR, suitable phantoms could be used in future work (Fellner et al., 2001; Price et al., 1990).
Reduced regional volumes associated with total psychopathy scores in an adult population with childhood lead exposure
2018, NeuroToxicologyCitation Excerpt :Normalized images were modulated by multiplying voxels by non-linear components, correcting for differences in subject brain sizes. An 8 mm full-width half-maximum (FWHM) Gaussian Kernel was used to smooth modulated gray and white matter images to improve the signal to noise ratio, help normalize the distribution of the data, and minimize the impact of registration errors (Scouten et al., 2006, Kiebel et al., 1999). Our statistical analyses were conducted in multiple steps.