Abstract
Multivariate machine learning algorithms were applied to BOLD fMRI data obtained from human subjects for decoding the orientation of gratings, with voxels larger than the width of orientation columns. Contributions to this successful decoding using low-resolution BOLD fMRI can potentially be made by 1) functionally selective large blood vessels, 2) orientation bias in large-scale organization, and 3) local orientation irregularities. In order to examine this issue, we re-analyzed cerebral blood volume-weighted fMRI data from cat visual cortex (Fukuda et al., J. of Neurosci, 2006). To remove large vessel contributions, ferrous iron oxide contrast agent was injected into blood. The functional data were obtained with 0.156 x 0.156 x 1 mm3. Then, high-resolution data were down-sampled to low-resolution up to 3 mm (the average orientation cycle is 1.0-1.4 mm). Linear support vector machine analysis showed that the presented orientation can be predicted above the chance level, even at 3-mm voxel resolution. To separate contributions from local orientation irregularities and from large-scale organizations, data were band-pass filtered with center frequency of 0.4 cycles/mm (frequency range of local irregularities) and 0.1 cycles/mm (low frequency). In both conditions, the presented orientation can be predicted above the chance level, with slightly better accuracy for the spatial filter of higher frequencies. Our analysis indicates that 1) large vessel contribution is not essential, and 2) local orientation irregularities can contribute for decoding of orientations in low-resolution fMRI data.
Meeting abstract presented at VSS 2013