The effect of spatial resolution on decoding accuracy in fMRI multivariate pattern analysis
Introduction
Functional magnetic resonance imaging (fMRI) is currently the most popular non-invasive method to investigate human brain structure and function. It indirectly measures neural activity primarily via the blood oxygenation level-dependent (BOLD) effect. Standard univariate statistical analysis (i.e. general linear model (GLM) analysis) of the task-based fMRI data has been utilized to detect voxel-wise differences of BOLD activation levels and, thus, to infer which brain areas are involved in a certain task. In recent years, multivariate pattern analysis (MVPA) has been used in fMRI to extract information from spatially distributed activation patterns, which may go undetected in conventional univariate analysis. Reliable decoding of information from fMRI data acquired at 3 T has been demonstrated from activation patterns in different brain areas (Haxby et al., 2001; Cox and Savoy, 2003; Haynes and Rees, 2005; Kamitani and Tong, 2005; Kriegeskorte and Bandettini, 2007; Formisano et al., 2008). Different biophysical hypotheses have been proposed to explain the ability of MVPA on fMRI data to detect information inaccessible with GLM. It has been suggested that MVPA is sensitive to information encoded at the sub-millimeter scale of neuronal functional columns. Such information, even if sampled at the lower resolution of standard fMRI voxel sizes (e.g. 3 × 3 × 3 mm3), may be accessible by MVPA due to local variations and irregularities in the columnar organization, resulting in weak but consistent biases in fMRI responses of the different voxels (Boynton, 2005; Kamitani and Tong, 2005; Haynes and Rees, 2006; Kamitani and Tong, 2006; Kriegeskorte and Bandettini, 2007); this mechanism is, therefore, named hyperacuity or voxel biased sampling. Alternatively, the transposition from high spatial frequency components of columns preferences to lower spatial frequency of the fMRI signal may be attributed to the cortical vasculature. This hypothesis is based on the fact that, using the standard gradient echo (GE) MRI sequences, the fMRI signal stems mostly from veins draining blood from a given tissue volume (see Uludag et al., 2009). Thus, a specific vein could be more sensitive to one neuronal population than another introducing a spatial bias. Hence, this hypothesis is known as biased draining regions (Kamitani and Tong, 2005; Gardner et al., 2006; Kamitani and Tong, 2006; Kriegeskorte and Bandettini, 2007; Gardner, 2010; Kamitani and Sawahata, 2010; Kriegeskorte et al., 2010; Shmuel et al., 2010).
According to another hypothesis, MVPA may rely on large spatial scale non-columnar organization (Op de Beeck, 2010), such as radial preference maps (Freeman et al., 2011). Since MVPA represents a computational scheme to non-locally average the fMRI signal, in this framework, MVPA would be able to detect low spatial frequency information too weak to be detected with univariate analysis.
Note that these hypotheses are not mutually exclusive (see Shmuel et al., 2010; Swisher et al., 2010). Nevertheless, they do predict testable effects of spatial smoothing on decoding performance. Op de Beeck has shown that spatial smoothing does not deteriorate decoding performance of objects and orientations from activation patterns in lateral occipital cortex and V1, respectively (Op de Beeck, 2010). He interpreted these results as an argument against hyperacuity and in favor of large-scale organization. Further support for this hypothesis comes from the finding that it is possible to decode across experimental sessions performed in different days (Freeman et al., 2011). In contrast, several studies (Swisher et al., 2010; Alink et al., 2013; Misaki et al., 2013) demonstrated that spatial smoothing decreases decoding accuracies for orientation and ocular dominance from V1 data, suggesting relevant information content at the individual voxel level. The few investigations so far on the underlying mechanisms of MVPA on fMRI data and the effect of spatial smoothing have been limited to the early visual cortex. In addition, they have been restricted to a small set of stimuli and decoding tasks (e.g. decoding of orientation, ocular dominance, and direction of motion) and have yielded conflicting evidence.
The main goal of the current study is to investigate how information at different spatial resolutions contributes to MVPA decoding. We employed ultra-high field (7 T) fMRI to acquire high-resolution data (1.1 mm isotropic), which were then reconstructed at different effective spatial resolutions from original k-space data to evaluate the effects of spatial resolution on MVPA decoding performance. Based on an experimental paradigm and on stimuli that were used in a previous fMRI study at 3 T (Formisano et al., 2008), we presented speech stimuli (vowels) from different speakers and considered the single-trial decoding of vowels and speakers from auditory cortical response patterns. Compared to conventional 3 T fMRI, 7 T fMRI presents several advantages, such as higher signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR), and therefore the possibility of higher spatial resolution with lower partial volume effects and greater spatial specificity (Yacoub et al., 2005; Uludag et al., 2009; Polimeni et al., 2010). On the other hand, it presents challenges such as larger distortions, sensitivity to motion, and larger number of voxels to be handled by the decoding algorithm (Formisano and Kriegeskorte, 2012). Therefore, we also investigated the effects of temporal SNR, CNR, and head motion and of typical noise-reduction steps (spatial smoothing) on MVPA performances.
Section snippets
Subjects
Ten healthy volunteers (seven females, age range 25–32) with normal hearing took part in this experiment. Informed consent was obtained from all participants according to the approval by the Ethical Committee of the Faculty of Psychology and Neuroscience, University of Maastricht.
Stimuli and task
We used the same auditory stimuli as in the study by Formisano et al. (2008) consisting of three vowels (/a/, /i/, /u/) spoken by three different speakers (sp1: female, sp2: male, sp3: male). For each of these 9
Univariate analysis
Univariate analysis showed significant BOLD signal activation (Q[FDR] < 0.05 and cluster size threshold of 4 voxels; FDR stands for false discovery rate) in the auditory cortex in response to the stimulus sounds (Fig. 2), similar to Formisano et al. (2008). Outside the temporal lobe, significant activation was observed in the frontal (Broca's area) and parietal lobe. Pair-wise contrast analysis between speakers and/or vowels did not show significant effects for any comparison (data not shown).
Discussion
In this study, we used high-resolution 7 T fMRI data and their reconstruction at different effective spatial resolutions to investigate the relevant spatial scale of information content of the fMRI signal in the auditory cortex for vowel and speaker identity. To the best of our knowledge, this work is the first investigating the relation between spatial resolution and MVPA with fMRI at ultra-high magnetic field strength. Instead of acquiring additional data at lower spatial resolutions, we
Acknowledgments
This work was financially supported by the Marie Curie Initial Training Network grant (PITN-GA-2009-238593) of EU and VIDI grant (KU, 452-11-002) of the Netherlands Organization for Scientific Research (NWO). EF was supported by an NWO VICI grant (453-12-002).
References (36)
- et al.
Functional quantitative susceptibility mapping (fQSM)
NeuroImage
(2014) - et al.
Modeling and analysis of mechanisms underlying fMRI-based decoding of information conveyed in cortical columns
NeuroImage
(2011) - et al.
Functional magnetic resonance imaging (fMRI) “brain reading”: detecting and classifying distributed patterns of fMRI activity in human visual cortex
NeuroImage
(2003) - et al.
Combining multivariate voxel selection and support vector machines for mapping and classification of fMRI spatial patterns
NeuroImage
(2008) - et al.
Multivariate examination of brain abnormality using both structural and functional MRI
NeuroImage
(2007) - et al.
Seeing patterns through the hemodynamic veil—the future of pattern-information fMRI
NeuroImage
(2012) Is cortical vasculature functionally organized?
NeuroImage
(2010)- et al.
Spatial smoothing hurts localization but not information: pitfalls for brain mappers
NeuroImage
(2010) - et al.
Decoding seen and attended motion directions from activity in the human visual cortex
Curr. Biol.
(2006) - et al.
Analyzing for information, not activation, to exploit high-resolution fMRI
NeuroImage
(2007)