Contribution of large scale biases in decoding of direction-of-motion from high-resolution fMRI data in human early visual cortex

Abstract

Previous studies have demonstrated that the perceived direction of motion of a visual stimulus can be decoded from the pattern of functional magnetic resonance imaging (fMRI) responses in occipital cortex using multivariate analysis methods (Kamitani and Tong, 2006). One possible mechanism for this is a difference in the sampling of direction selective cortical columns between voxels, implying that information at a level smaller than the voxel size might be accessible with fMRI. Alternatively, multivariate analysis methods might be driven by the organization of neurons into clusters or even orderly maps at a much larger scale. To assess the possible sources of the direction selectivity observed in fMRI data, we tested how classification accuracy varied across different visual areas and subsets of voxels for classification of motion-direction. To enable high spatial resolution functional MRI measurements (1.5mm isotropic voxels), data were collected at 7T. To test whether information about the direction of motion is represented at the scale of retinotopic maps, we looked at classification performance after combining data across different voxels within visual areas (V1-3 and MT+/V5) before training the multivariate classifier. A recent study has shown that orientation biases in V1 are both necessary and sufficient to explain classification of stimulus orientation (Freeman et al., 2011). Here, we combined voxels with similar visual field preference as determined in separate retinotopy measurements and observed that classification accuracy was preserved when averaging in this 'retinotopically restricted' way, compared to random averaging of voxels. This insensitivity to averaging of voxels (with similar visual angle preference) across substantial distances in cortical space suggests that there are large-scale biases at the level of retinotopic maps underlying our ability to classify direction of motion.

On the Definition of Signal-To-Noise Ratio and Contrast-To-Noise Ratio for fMRI Data

“…Many authors explicitly reported tSNR values ranging from 4.42 to 280, while in a few other cases CNR values were reported that varied from 0.5 to 1.8. Note that one study reported the possibility of a CNR value as low as 0.01, but this was specific to the imaging of orientation columns in the visual cortex and caused by a combination of bias and voxel size [16]. An interesting observation was that Hughes and Beer made an explicit distinction between SNR for active clusters and SNR for non-active clusters [17].…”

Exploring structure and function of sensory cortex with 7 T MRI

“…For example, we have used 3D-EPI to acquire high-resolution fMRI data in visual cortex with volume receive coils (Schluppeck D Merriam E 2010, Goncalves, Ban et al 2015) and are now testing data acquisitions in somatosensory cortex with local surface receive coils (2D-EPI) as shown here. The improved BOLD CNR at 7 T has made measuring weaker signals, such as biases across cortical maps easier in individual participants (Beckett, Peirce et al 2012). At the same time, the sensitivity of functional signals from methods other than gradient-echo EPI are more robust and usable.…”

“…They demonstrated the responses to particular stimuli in voxels in V1 were correlated with the corresponding visual field angle for those voxels. They reported that the large-scale bias they had observed, rather than a smaller-scale bias introduced by the voxel sampling, was sufficient to drive the decoding of orientation information with multivariate methods (see also Beckett, Peirce et al 2012, Wang, Merriam et al 2014). In particular, we were interested to test how much data we would need to replicate the results from Freeman, Brouwer et al (2011) when using high resolution 7T with dynamic shimming methods (described above) -3.375 mm 3 compared to the previous study of 8 mm 3…”

Linking brain imaging signals to visual perception

“…2c ). There is an ongoing debate about the origin of the preferences that are exploited by multivoxel pattern classifi cation analysis (MVPA) approaches, with contributions likely from random spatial irregularities in fi ne columnar maps (Swisher et al, 2010 ) as well as coarser scale signals that refl ect the systematic (i.e., retinotopic) organization of stimulus responses to properties such as orientation and motion direction (Raemaekers et al, 2009 ;Freeman et al, 2011 ;Beckett et al, 2012 ). This debate refl ects the complex correspondence between BOLD and neural selectivity (Kriegeskorte et al, 2010 ) and emphasizes the importance of work comparing directly physiology to imaging signals.…”