Multiple neural networks supporting a semantic task: An fMRI study using independent component analysis
Introduction
Semantic access, the process of extracting meanings from different forms such as sounds or visually displayed objects/words, plays an important role in comprehending the corresponding sources. It is critical for interpreting and interacting with the environment and its importance is also evidenced by a number of publications in the neuroscience literature investigating its various aspects (Bookheimer, 2002, Jobard et al., 2003, Price, 2000, Shelton and Caramazza, 1999; etc). Taking a visual word as an example, there are several processing stages during the semantic task. Generally, the word or letter string is first analyzed in a sensory process. It is then mapped to a word lexicon to transform the word form to morphemes. The semantic integration is finally completed through semantic access (Marinkovic et al., 2003, Rayner and Posnansky, 1978).
The visual word semantic access has been the focus of many studies ranging from lesion-deficit cases (Ladavas et al., 1997, Shelton and Caramazza, 1999, Warrington and Shallice, 1979) to behavior (Rayner and Posnansky, 1978, Zhou and Marslen-Wilson, 1999) and the use of electrophysiological (Bentin, 1987, Hauk et al., 2006) or neuroimaging techniques (Drzezga et al., 2008, Jobard et al., 2003, Small et al., 2000). Evidence from neuroimaging studies suggests that many brain regions are involved in semantic access of a visual word form at various stages. They are mainly located at the occipital, occipitotemporal, left temproparietal, and frontal cortices. In the first stage, the visual stimuli evoke the activation of bilateral occipital cortex, where the primary encoding of the visual stimuli is processed. As part of the encoding, the bilateral fusiform gyri, especially the left middle ventral fusiform gyrus, are responsible for the local feature processing of the input stimuli while the bilateral lingual gyri for the global processing (Fink et al., 1996). Noting its important role in this encoding stage, some researchers referred to the left middle ventral fusiform gyrus as visual word form area (Cohen et al., 2000, Cohen et al., 2004, Gaillard et al., 2006, McCandliss et al., 2003) though others have questioned that this area is the only one responsible to processing visual word form (Drzezga et al., 2008, Hillis et al., 2005, Starrfelt and Gerlach, 2007). In the second stage, the mapping of visual words to a lexicon evokes parts of left temproparietal and parietal cortices. Previous studies have found that the left posterior superior temporal cortex is engaged in translating orthography to phonology (Paulesu et al., 2000, Small et al., 2000) and have identified the supramarginal gyrus as the phonological store area (Baddeley, 1992, Paulesu et al., 1993). The angular gyrus, implicated in semantic association tasks (Binder et al., 1996, Bright et al., 2004, Buchel et al., 1998, Davis et al., 2004), was presumed a role in mapping visually presented inputs to linguistic representations (Grady et al., 2001). At the final stage, the semantic access and integration are completed in regions such as left temporal, parietal and frontal cortices. Among them, the left middle temporal gyrus (MTG) is involved in semantic processing, visual perception, and multimodal sensory integration (Ben-Shachar et al., 2004, Bottini et al., 1994, Kansaku et al., 2000, Kuperberg et al., 2000, Luke et al., 2002, Vingerhoets et al., 2003, von Kriegstein et al., 2003). Moreover, the left MTG was considered to play an important role in processing phonetic and lexical–semantic information (Muller et al., 1997). Meanwhile, it has been observed that the angular gyrus was also involved in semantic access in some studies (Brunswick et al., 1999, Drzezga et al., 2008, Gorno-Tempini and Price, 2001, Vandenberghe et al., 1996) and that the left inferior frontal cortices (including Broca's area) participated in semantic retrieval, semantic categorization and other semantic associated conditions (Daselaar et al., 2002, Demb et al., 1995, Demonet et al., 1992, Golby et al., 2005, Koch et al., 2002, Petersen et al., 1988, Small et al., 2000, Thompson-Schill, 2003). Furthermore, the left inferior frontal gyrus (IFG) was claimed to be divided into two functional areas: 1) the posterior and dorsal part, involved in phonological processing, and 2) the anterior and ventral part, involved in semantic processing (Adams and Janata, 2002, Binder et al., 2003, Buchanan et al., 2000, Herholz et al., 2002, Jennings et al., 1998, Noesselt et al., 2003, Noppeney and Price, 2004, Poldrack et al., 1999, Vigneau et al., 2006). Another study has suggested that this division specialization can be extended to the middle frontal gyrus (MFG) for Chinese word processing (Booth et al., 2006).
The findings from the previous studies above were mostly based on a univariate subtraction approach examining each of the brain regions separately (i.e., contrasting between the well-controlled conditions on regional basis with univariate analysis). However, an essential property of functional brain organization is that the complex cognition depends on the integrated activity of spatially distributed brain regions (Friston, 2002). Especially for language processing, the results from a meta-analysis have suggested that there are large-scale architecture networks rather than modular organization of language in the left hemisphere (Vigneau et al., 2006). Therefore, it is essential to investigate the neural networks underpinning word semantic access using functional connectivity analysis. A multivariate view over the whole brain may re-confirm results based on the univariate subtractive methods. Furthermore, the inter-relationships among various brain regions and their involvements at different stages prior to and during the semantic access can be further clarified.
Localized to a limited number of brain regions and with a relatively simple correlational analysis method, functional connectivity of visual word semantic access has been investigated and reported in the literature (Hampson et al., 2006, He et al., 2003, Horwitz et al., 1998). Most of these studies have focused on investigating the local functional connectivity between Broca's area, Wernicke's area and the angular gyrus and have found different functional connectivity for distinct conditions or participant groups. The study of He et al. (2003) showed the strength change of functional connectivity between Broca's area and Wernicke's area with varying levels of semantic retrieval effort. Another positron emission tomography (PET) study reported the functional links of angular gyrus with posterior language areas in normal participants and the functional disconnection of the left angular gyrus with these regions in dyslexia patients (Grady et al., 2001). A study investigating the variation of the functional connectivity between the left BA39 and Broca's area with reading ability demonstrated that good readers showed stronger connectivity while the poor ones relied more on the orthographic to phonological decoding (Hampson et al., 2006).
The region of interest (ROI) approach of correlation analysis depends on measuring the correlation of the time courses between limited numbers of pre-selected brain regions. It is noted that region-based analyses not only assess the correlation/interaction between pre-selected regions but they are also used to assess the correlation between pre-selected regions and the whole-brain (e.g. Bokde et al., 2001, Hampson et al., 2006). However, it failed to reveal all the connected areas over the whole-brain that are integrated in one functional network because this approach could only reveal the relationship between pre-selected regions. Currently, several studies have indicated that the completion of one cognitive task may recruit several cognitive components each implicated with a different functional network (Calhoun et al., 2002b, Schmithorst and Brown, 2004). Indeed, several studies reported multi-functional sub-networks corresponding to different cognitive components of a cognitive task such as simulated driving (Calhoun et al., 2002b), visual perception (Calhoun et al., 2001c), math processing (Schmithorst and Brown, 2004), music processing (Schmithorst, 2005) and auditory narrative comprehension (Karunanayaka et al., 2007, Schmithorst et al., 2006). Thus, for studying cognitive tasks such as complicated as visual word semantic access, the system-wise multivariate approach is a preferred method.
The current study used independent component analysis (ICA) as an integrated multivariate approach to examine the brain volume as a whole (see below for more detailed description). As one of several approaches for functional connectivity analysis over the whole brain, ICA has been used to separate fMRI data into spatial independent components together with their associated time courses (McKeown and Sejnowski, 1998). ICA does not require accurate modeling of the hemodynamic response function (HRF), which depends upon participants, cognitive conditions and brain locations (Calhoun et al., 2001b). The activated brain areas detected in one independent component (IC) interact as one functional network driven by the same time course of the corresponding component. It has been suggested that ICA extracts functional networks that reflect the global integrated rather local/segregated characters (Margulies et al., 2007, van de Ven et al., 2004). GroupICA is extended for multi-subject analyses from individual ICA and can be used to generate random-effect statistical inferences across subjects (Calhoun et al., 2001b). For both simulated data and human fMRI data, GroupICA can yield results largely similar to those of individual ICA. Furthermore, GroupICA incurs the least amount of both computation time and arbitrary interpretation based on manual operations (Calhoun et al., 2001b).
The purpose of our study is first to explore systematic characteristics of functional brain connectivity involved in semantic access for words presented visually (visual word). We hypothesized that there are several functional networks underpinning this process, one of them being the sensory stimulus/input-related component and others responsible for different cognitive components. Because semantic access is the focus of the current study, we used a semantic judgment task for visually presented words. In order to investigate the multiple aspects involved, the task was not specifically designed to target only one component but various activated brain regions as extensively as possible. We are also interested in generalizing our hypothesis to other sensory modality of input stimuli. Therefore, a semantic judgment task based on auditory words was used secondarily to examine this possible generalization. We expected that while the input networks activated by different sensory stimuli would vary, the networks for the subsequent semantic processing stages would show a substantial degree of similarity.
Section snippets
Participants
Fourteen healthy volunteers [8 males and 6 females, ages between 19 and 26 years (Mean ± SD: 21.1 ± 3.74 years), right-handed] participated. Handedness was determined by the Edinburgh Inventory (Oldfield, 1971). All subjects are native Chinese (Mandarin) speakers, and have no history of psychiatric or neurological abnormalities. All participants had normal or corrected to normal vision through the use of MRI-compatible lenses. The purpose of the study was explained to the participants and each of
The functional networks of WJ
We found that there were three task-related ICs, IC1, IC2 and IC3, separated from GroupICA analysis of WJ paradigm (Fig. 2). The correlation coefficients between the 6-sec delayed ON–OFF paradigm and the time course of each of the three networks were, respectively, 0.88 (4.53 × 10− 43), 0.52 (5.05 × 10− 10) and 0.788 (3.81 × 10− 38) (Panels A, B and C respectively in Fig. 3).
Our results showed that the IC1 involved two clusters (Panel A of Fig. 2): the posterior cluster in the bilateral middle occipital
Discussion
The ICA method can be used to identify spatially independent sources from the linearly mixed fMRI data (McKeown et al., 1998). The brain regions identified in one component were driven by the same time course and constituted a specific functional network (Calhoun et al., 2002a). The current results demonstrated three WJ task-related ICs separated by the GroupICA. As shown in Fig. 3, the time courses of the three functional networks were all significantly correlated with the paradigm of
Conclusion
Taken together, our results strongly support that there are multiple brain networks underpinning either language task (WJ or AJ) and that these two tasks may share a common functional network responsible for word semantic access regardless their sensory input modalities.
Acknowledgments
This work is supported by the Natural Science Key Foundation of Beijing (4061004), National Natural Science Foundation of China (60628101), and National High-Tech R&D Program (863 Program, 2006AA01Z132).
References (86)
- et al.
A comparison of neural circuits underlying auditory and visual object categorization
NeuroImage
(2002) - et al.
Neural correlates of syntactic movement: converging evidence from two fMRI experiments
NeuroImage
(2004) Event-related potentials, semantic processes, and expectancy factors in word recognition
Brain Lang.
(1987)- et al.
Differential prefrontal-temporal neural correlates of semantic processing in children
Brain Lang.
(2006) - et al.
Functional Interactions of the Inferior Frontal Cortex during the Processing of Words and Word-like Stimuli
Neuron
(2001) - et al.
Specialization of phonological and semantic processing in Chinese word reading
Brain Res.
(2006) - et al.
Unitary vs. multiple semantics: PET studies of word and picture processing
Brain Lang.
(2004) - et al.
Recognition of emotional prosody and verbal components of spoken language: an fMRI study
Cogn. Brain Res
(2000) - et al.
fMRI activation in a visual-perception task: network of areas detected using the general linear model and independent components analysis
NeuroImage
(2001) - et al.
Distinct unimodal and multimodal regions for word processing in the left temporal cortex
NeuroImage
(2004)