Abstract
The analysis of functional neuroimaging data relies primarily on inferential statistics such as analysis of variance to detect differences in brain voxel activations as a function of experimental condition. These analyses, however, do not answer an important question: “How reliably do patterns of brain activation indicate the task in which the brain is engaged or the stimulus being processed by the subject?” Recent work shows the utility of refocusing the analysis on patterns of activation (Carlson et al., 2001). We applied classification analyses to fMRI data from Haxby et al., (2001) to measure the discriminability of brain activation patterns resulting from viewing different kinds of objects (e.g., faces, houses). Principal component analysis was used to derive a basis set from a subset of the brain scans, using ventral temporal voxels that differed significantly for objects in the inferential analyses. Individual scans were coded via their projections in the derived space. The utility of individual principal components (PCs) for separating object classes was assessed by discriminant analysis of the individual scan projections. All analyses used generalized classification tests in which odd/even runs of scans served alternatively as training/test scans. We computed d's for discriminating face versus object scans from 7 categories. We found that single PCs discriminated between faces and the object categories, with d's for the best PCs ranging from .99 to 2.46. The second best predictor yielded d's varying from .91 to 1.64. An interesting aspect of this finding is that orthogonal patterns of activation discriminated face versus object categories with high accuracy. Thus, more than one kind of pattern can discriminate between categories. An advantage of this analysis is that the patterns of activity most useful for separating the object categories can be mapped back onto structural scans and interpreted as brain images.