Abstract
Visual processing in the human ventral cortex entails extraction of features from retinal images that mediate perception. In the human ventral cortex, early and late visual areas have been implicated in the analysis of simple and complex features respectively. If we view this processing pathway as a sequence of decision stages, each extracting progressively more abstract and invariant features from the output of preceding stages, then by considering the relationship between signal uncertainty and the slope of a psychometric function, we can show that the extent to which the output of a decision stage is perturbed by noise added to the visual stimulus will be more threshold-like (steeper log-log slope) for a decision stage further down the decision cascade. To test this theory, we used images of scenes and added visual noise that matched the signal's spatial-frequency power spectrum. The resulting images were rescaled to maintain a constant mean luminance and rms contrast across all noise levels. We localized individually in each observer the retinotopic regions and the LOC, and measured event-related BOLD response in these regions during a scene discrimination task performed at 4 noise levels. Behavioral performance increased with increasing signal-to-noise ratio. We found that log %BOLD signal change from fixation baseline vs. log SNR is well-described by a straight line for all visual areas. The regression slope increased monotonically from early to late areas along the ventral stream. A factor of 8 change in SNR produced little change to the BOLD response in V1/V2, but resulted in progressively larger changes in V4v, posterior (LO), and anterior (pFs) subregions of the LOC. In accordance with our theory on noise perturbation, the results suggest approximately ordered decision stages in the ventral pathway.