seldom used (Friedman, Ween, & Albert, 1993). Because of the varying ...... Pugh, K. R., Shaywitz, B. A., Shaywitz, S. E., Constable, R. T.,. Skudlarski, P.
A Parametric Approach to Orthographic Processing in the Brain: An fMRI Study M. -A. Tagamets University of Maryland School of Medicine and Georgetown University Medical Center
Jared M. Novick, Maria L. Chalmers, and Rhonda B. Friedman Georgetown University Medical Center
Abstract & Brain activation studies of orthographic stimuli typically start with the premise that different types of orthographic strings (e.g., words, pseudowords) differ from each other in discrete ways, which should be reflected in separate and distinct areas of brain activation. The present study starts from a different premise: Words, pseudowords, letterstrings, and false fonts vary systematically across a continuous dimension of familiarity to English readers. Using a one-back matching task to force encoding of the stimuli, the four types of stimuli were
INTRODUCTION In recent years, a number of studies have investigated the processing of orthographic information using functional imaging techniques. The goal of these studies has typically been to locate areas of the brain that subserve various subcomponents of written word processing, e.g., visual, phonological, and semantic. Several of these studies have used string stimuli that include real words ( W ), pseudowords (PW, i.e., pronounceable strings that are similar to real words), unpronounceable consonant letter strings (LS), and false-font strings (FF), in which real letters are replaced by letter-like symbols. Such studies are often predicated on the assumption that comparisons of brain activity that is observed while subjects are viewing these different stimuli will reveal distinct regions of the brain involved, specifically, in the phonologic, orthographic, or semantic processing of words. However, it has proved difficult to correlate these studies with other non-imaging evidence, such as that from lesion and psychophysical studies. Moreover, the results of these imaging studies have not been entirely consistent, even at the level of reading single words. One thing that has become clear is that even subtle variations in condition components and experimental parameters can yield significantly different results (Indefrey et al., 1997; Price et al., 1994), and it has been advised that until these issues are better understood, © 2000 Massachusetts Institute of Technology
visually presented to healthy adult subjects while fMRI activations were obtained. Data analysis focused on parametric comparisons of fMRI activation sites. We did not find any region that was exclusively activated for real words. Rather, differences among these string types were mainly expressed as graded changes in the balance of activations among the regions. Our results suggest that there is a widespread network of brain regions that form a common network for the processing of all orthographic string types. &
care should be taken in attributing distinct cognitive processes to specific areas (Beauregard et al., 1997; Demonet, Fiez, Paulesu, Petersen, & Zatorre, 1996; Price et al., 1994; Demonet, Wise, & Frackowiak, 1993; Sergent, Zuck, Levesque, & MacDonald, 1992). It is apparent that many factors are at play when linguistic stimuli are present, even when simple tasks are being examined. It may not really be possible to completely tease apart the various aspects of word processing. For example, orthographic regularity is correlated with the availability of a phonological code, which, in turn, is correlated with the probability of semantic representations. One cannot deliberately suppress the activation of these different processing types. In fact, processing mechanisms may become activated even when the results of the processing are not useful to the task. Thus, for example, orthographic similarities between a pseudoword and a neighborhood real word may be sufficient to cause some activation of the semantic representations of the similar real word. Furthermore, these cognitive operations are not likely to be all-or-nothing in nature. For instance, one aspect of the four stimulus string types (W, PW, LS, FF) that may be of great importance, but has received little attention, is relative familiarity. Letterstring familiarity is a continuous variable, rather than a discrete one. With regard to the statistical patterns of the strings at the lexical level, words are most familiar, followed by PW and Journal of Cognitive Neuroscience 12:2, pp. 281–297
LS. At the sublexical level, these statistical patterns may be described by measures, such as the frequency counts of the bigraphs and trigraphs that compose the strings. Indeed, even within a string type, relative familiarity can vary, and this may have an effect upon processing. Thus, letter strings with greater ‘‘order of approximation’’ to real English words are more easily processed than those with lower order of approximation to real words (Miller, Bruner, & Postman, 1954). False fonts, of course, are the least familiar of the four string types. These two factors— the automatic activation of ‘‘connected’’ processes, and the continuous nature of most cognitive operations— have important implications for imaging research, particularly with regard to the criteria that one uses to conclude that a region of the brain is specialized for a particular type of processing. First, given that most processing is continuous in nature, the amount of brain activation seen in a given region is likely to be a continuous, rather than a discrete, variable. Therefore, relying upon a binary classification— activated or not activated— may be losing important information about the role that a particular area of the brain plays in a particular type of processing. Second, given that certain stimuli might trigger processing mechanisms that do not contribute to performance of the task on hand, it may be possible that a region participates in a certain type of processing, even if that region shows activation for stimuli that presumably do not rely upon that type of processing. For example, if a particular region of the brain shows activation for both words and pseudowords relative to a control condition, but the subtraction of pseudowords from words does not result in a significant activation in that region, it is not necessary, nor even appropriate, to conclude that the region cannot be involved in semantic processing. The notion of string types varying continuously according to relative degrees of familiarity— be it visual, phonological or semantic— fits well with recent connectionist models of single-word reading (see Seidenberg, 1995 for a review). These models do not posit visual word-form centers that are specific to real words. Rather, they describe networks of interconnected sets of orthographic-, phonologic-, and semantic-processing units, which process not only real, familiar words, but all orthographic input. Rather than there being separate processing mechanisms for different types of orthographic stimuli, there may be one common network that handles all orthographic stimuli, with greater or lesser emphasis allocated to each processing unit. Some previous studies have focused on identifying regions of the brain that are specifically involved in processing word forms in the visual modality. An early PET study examined visual processing of single strings: words, PW, LS, and FF. Passive viewing of each of these strings was contrasted with a resting condition, in which subjects fixated on a blank screen (Petersen, Fox, Snyder, & Raichle, 1990). Activation in the left medial282
Journal of Cognitive Neuroscience
extrastriate area was found for words and pseudowords, but not for letterstrings or false fonts, and the authors suggested that this region is involved in visual ‘‘legitimate word-form analysis.’’ The Petersen study has since been followed by a number of other imaging experiments, most of which were PET studies, aimed at differentiating among components, such as lexical analysis, phonology, and semantics in visual single-word processing (Indefrey et al., 1997; Rumsey et al., 1997; Puce, Allison, Asgeri, Gore, & McCarthy, 1996; Pugh et al., 1996; Price et al., 1994; Price, Wise, & Frackowiak, 1996; Howard et al., 1992). Studies that have used visually presented singleword stimuli have varied to some extent in the specific tasks that were performed. The tasks have included passive viewing (Beauregard et al., 1997; Indefrey et al., 1997; Price et al., 1994; Petersen et al., 1990), visualfeature detection (Price et al., 1996), oral or silent reading (Herbster, Mintun, Nebes, & Beckert, 1997; Rumsey et al., 1997; Bookheimer, Zeffiro, Blaxton, Gaillard, & Theodore, 1995; Howard et al., 1992), and lexical decision (Rumsey et al., 1997). The control conditions for the contrasts also varied. Some studies used a common baseline for all tasks, such as a resting condition (Price et al., 1996; Petersen et al., 1990) or passive viewing of a crosshair (Beauregard et al., 1997; Rumsey et al., 1997). Others made direct contrasts between tasks involving different string types (Herbster et al., 1997; Indefrey et al., 1997; Price et al., 1994; Price et al., 1996; Howard et al., 1992), or between strings and other stimuli, such as pictures (Bookheimer et al., 1995) or single symbols (Indefrey et al., 1997). The results of these studies are somewhat varied. In posterior regions, the inferior/middle-temporal region was reported in most comparisons involving real words versus nonword stimuli (Beauregard et al., 1997; Bookheimer et al., 1995; Price et al., 1994; Price et al., 1996; Howard et al., 1992), but not in any contrasts involving nonword string stimuli versus rest. The locations of frontal activation foci varied significantly among studies, even in the left inferior-frontal gyrus, an area known to be crucial for language processing. This variation quite likely reflects the diversity of cognitive tasks that were examined and the specific contrasts that were used. The frontal cortex is thought to play a role in a number of different cognitive processes, such as semantics (Gabrieli, Poldrack, & Desmond, 1998; Petersen et al., 1990; Posner, Petersen, Fox, & Raichle, 1988), working memory (Courtney, Ungerleider, Keil, & Haxby, 1996; Goldman-Rakic, 1996; Miller, Erickson, & Desimone, 1996; Haxby, Ungerleider, Horwitx, Rapoport, & Grady, 1995; Fuster, 1973), encoding and/or retrieval of memory (Nyberg, Cabeza, & Tulving, 1998), and planning (Milner, 1982). It is probable that both linguistic components and other task-related processes engage various combinations of these functions. Volume 12, Number 2
It has been proposed that the early stages of visual processing of orthographic information may proceed in two distinct manners. The normal manner, in which letters are identified in literate adults, is automatic and parallel. However, an alternative mechanism is said to predominate in the young child who is learning to read: This involves the slow, serial identification of each letter in the left-to-right order. This alternative process is said to remain available to the adult reader, though it is seldom used (Friedman, Ween, & Albert, 1993). Because of the varying processes that may be called upon for these different routes, it is likely that networks that subserve reading also include processing units for elementary visual analysis, spatial arrangements, and so on. Thus, the networks that are involved in reading are likely to be multi-componential. In looking for such networks, it is important that nothing change between conditions other than the actual stimulus presented. All other aspects of the task must be identical in both the string and control conditions. The paradigm that is employed in the current study fulfils this requirement. Our study adopts the viewpoint that the nature of these stimulus types is continuous and parametric. The familiarity of a string is intimately related to its encodability as an entity. By using a task whose success depends on encoding whole strings of various types and comparing each of these types to a common baseline condition, we can look at relative changes in patterns of activation across the different levels of familiarity. This approach will help identify the relative roles of each of the nodes within a network of regions that are involved. The strings used in the study vary in familiarity at both the lexical and sublexical levels (see Methods section). To increase the likelihood of obtaining activation in areas of the brain involved in encoding whole words and word-like stimuli, a one-back matching task was employed in this study. Studies of tasks that demand attention have demonstrated stimulus-specific areas that seem to be enhanced when encoding for the stimulus is required. For example, experiments using delayed matching tasks in animal single-cell recordings have shown the involvement of stimulus-specific frontal neurons (Chelazzi, 1995; Miller & Desimone, 1994; Funahashi, Chafee, & Goldman-Rakic, 1993; Fuster, Bauer, & Jervey, 1982) and modulation of specific extrastriate regions (Miller, Li, & Desimone, 1993; Fuster, 1990; Haenny, Maunsell, & Schiller, 1988). Apparently homologous regions have been found in human selective attention and working-memory studies (Baker, Frith, Frackowiak, & Dolan, 1996; Courtney et al., 1996; Haxby et al., 1994; Haxby et al., 1995; Corbetta, Miezin, Dobmeyer, Shulman, & Petersen, 1991). Thus, it would be expected that activation in areas selective for words and word-like stimuli would be enhanced under such conditions. A potential mechanism for such enhancement in an object-matching task was demonstrated in a large-scale
model of PET (Tagamets & Horwitz, 1998). In this model, the enhanced activity is the result of the interaction of extrastriate regions and working memory in the frontal cortex. In order to minimize effects of general attention, response execution and visual input per se, our control condition uses the same one-back matching task, but with single geometric symbols, rather than orthographic strings. By comparing all four string types to this active baseline control, we would expect to identify differences, as well as commonalities, and to be able to track changes across conditions in selected regions in a parametric manner.
RESULTS Each of four stimulus types— words, pseudowords, letterstrings, and false fonts— was visually presented in a separate series, with alternating blocks of task and control conditions within each series (see Figure 1). While undergoing fMRI, subjects were instructed to press the mouse button whenever two matching stimuli were presented in consecutive order. Contrasts between the task and control across 11 subjects were generated separately for each stimulus type. Statistical parametric mapping (SPM) (z) values for all activation foci that exceeded a corrected threshold of p
