On representation and reproducibility.

Abstract

Two papers were selected this year for their implementation of novel analytical approaches on archived data provided by the Dartmouth fMRI Data Center. Coincidentally, the two independent studies, conducted in different parts of the world, tested their models on the same data set, originally collected by Ishai, Ungerleider, Martin, Schouten, and Haxby (1999) and Ishai, Ungerleider, Martin, and Haxby (2000). Reanalyzing the same data set, however, was not the only common aspect of these two studies. Both have essentially replicated the original results, and more importantly, both have provided new and complementary tools that can be used in conjunction with the classical statistical parametric maps (SPM): a dynamic causal model (DCM) to estimate the connectivity between cortical regions (Mechelli, Price, Noppeney, & Friston, 2003), and an empirical Bayes method for estimating the reproducibility of activation within voxels (Liou, Su, Lee, & Cheng, 2003). The representation of faces and objects in the human brain has been a topic of an ongoing debate. Functional MRI studies have identified category-selective patches, or ‘‘modules,’’ in the ventral occipito-temporal cortex that respond preferentially to faces (Kanwisher, McDermott, & Chun, 1997), places (Epstein & Kanwisher, 1998), and human body parts (Downing, Jiang, Shuman, & Kanwisher, 2001). However, other fMRI studies have shown that any given category, regardless of its biological significance (e.g., faces, houses, animals, tools, or even chairs, bottles, and shoes), evoked a widely distributed and overlapping pattern of activation, which included regions responding maximally to that category, as well as suboptimal responses in adjacent regions. Moreover, a variety of stimulus format (written words, photographs, line drawings) and cognitive tasks (passive viewing, naming, delayed matching, N-back, etc.,) evoked virtually identical patterns of activation in the ventral stream (Haxby et al., 1999, 2001; Ishai et al., 1999, 2000; Chao, Haxby, & Martin, 1999). The functional architecture of the ventral pathway thus seems to consist of continuous, topologically arranged visual attributes, the nature of which is currently unknown. The category-related activation in the ventral stream, however, is not the mere product of the hierarchical, bottom–up, ‘‘feature’’ analysis suggested by early fMRI studies (e.g., Malach et al., 1995). Selective attention (e.g., O’Craven, Downing & Kanwisher, 1999) and mental imagery studies (Ishai, Ungerleider, & Haxby, 2000; Ishai, Haxby, & Ungerleider, 2002; O’Craven & Kanwisher, 2000) have indicated that face and object perception is also modulated by top–down effects, likely originating in the parietal and frontal cortex. In their article ‘‘A Dynamic Causal Modelling Study on Category Effects: Bottom–Up or Top–Down Mediation?’’ Mechelli and colleagues investigated the extent to which category-related responses in the ventral occipito-temporal cortex are mediated by inputs from low-level visual areas and regions in the parietal cortex. To that end, they combined fMRI with a novel neural modeling technique. First, the original fMRI dataset was reanalyzed, using the classical SPM approach, to identify the visual response (i.e., regions that responded more to faces, houses, and chairs than to scrambled pictures), and the category-responsive regions (i.e., regions that responded maximally to one category relative to the other two). Then, for each individual subject, DCMs were constructed for the face-, houseand chair-responsive regions in the ventral stream, as well as two visually responsive, but not category-specific, dorsal regions, namely, V3 and the parietal cortex. DCM estimates the neuronal interactions between cortical regions within a Bayesian framework. In contrast with other models of cortical connectivity, the DCM uses designed perturbations with different inputs and assumes that the hemodynamic response is modulated by experimental parameters and not intrinsic noise (see Friston, Harrison, & Penny, 2003). Interestingly, in all subjects, the category-related responses were mediated by input from V3, but not from the parietal cortex. For example, when subjects were viewing chairs, the intrinsic effective connectivity from V3 to the chair-responsive region in the occipital cortex was stronger than the connectivity to the faceor houseresponsive regions. These findings are in support of bottom–up processing of visual attributes. The lack of differential top–down effects from the parietal cortex is perhaps not surprising, as an imagery study has shown that the generation of mental images of faces, houses, and chairs evoked category-related responses in the extrastriate cortex, and nondifferential activation in the parietal (IPS and precuneus) and frontal regions (Ishai et al., 2000). Because the same network of parietal and frontal regions is activated in many attention, retrieval from memory, and visual imagery tasks, it is National Institute of Mental Health