Figure 1.

Single-image activation of FFA and PPA discriminates preferred from nonpreferred images. The graphs show the 96 object images ranked by the activation they elicited in each ROI. Each bar represents activation to one of the 96 object images in percent signal change averaged across four subjects. Each image is placed exactly below the bar that reflects its activation, so that the images are ordered from left to right (i.e., only the x-coordinate is meaningful). The leftmost image activated the region most strongly, the rightmost image activated the region most weakly. The highest- and lowest- ranked images are enlarged to give a first impression of the region's response preference. The bars are color-coded for category to give an overall impression of category selectivity without having to inspect all single images. Insets show ROC curves and associated AUCs, indicating performance for discriminating faces from nonfaces (red) and places from nonplaces (blue). We used a two-sided label-randomization test to determine whether discrimination performance was significantly different from chance (H0: AUC = 0.5). Since we tested discrimination performance at five different ROI sizes for each region, we corrected p values for multiple (five) comparisons using Bonferroni correction. Error bars indicate SE of the activation estimates, averaged across four subjects. FFA and PPA were each defined at 128 voxels in each hemisphere, based on an independent block-localizer experiment. Note that the smooth falloff is a necessary consequence of the rank ordering of the activation profile. Therefore further analyses are required to test for preference inversions (Figs. 3, 4), gradedness (Figs. 5, 6), and a categorical step (Fig. 6).