Purchase this article with an account.
Eamon Caddigan, Dirk Walther, Li Fei-Fei, Diane Beck; Decoding of natural scene categories from transformed images using distributed patterns of fMRI activity. Journal of Vision 2008;8(6):330. doi: https://doi.org/10.1167/8.6.330.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Human observers are able to quickly and efficiently extract information, such as the “gist,” from images of natural scenes (Potter & Levy, 1969). Previous studies have identified brain regions that respond selectively to images of natural scenes, including the parahippocampal place area (PPA; Epstein & Kanwisher, 1998) and retrosplenial cortex (RSC; O'Craven & Kanwisher, 2000). However, it is not known to what extent these place-selective regions participate in the categorization of natural scenes. As a means of testing for the presence of scene-category information in these regions, we used fMRI and statistical pattern recognition algorithms (Cox & Savoy, 2003) to identify distributed patterns of activity associated with natural scene categories (beaches, mountains, forests, tall buildings, highways, and industrial scenes). In our first experiment, fMRI data was aquired while subjects passively viewed 60 images from each of six categories, in 6 blocks of 10 images each of the same category, organized into 12 runs during which images were displayed upright or inverted on alternating runs. In a leave-one-run-out (LORO) cross-validation procedure, we found that statistical pattern recognition algorithms were able to predict the categories of the scene viewed by the participants at rates significantly above chance using voxels in retinotopic cortex or the PPA. Using a more sensitive classifier than reported last year, we found a significant inversion effect in both PPA and retinotopic cortex: i.e. training on upright and testing on inverted resulted in a decrement in performance relative to testing on new upright images. A subsequent experiment used a similar design, with alternating runs consisting of sequences of large and small images. Above-chance classification rates were obtained in the PPA, with a decrease in accuracy for small scene images vs. large images. These results suggest that the PPA is sensitive to changes in image size and orientation.
This PDF is available to Subscribers Only