Dissociating the roles of the default-mode, dorsal, and ventral networks in episodic memory retrieval

Abstract

Emerging evidence indicates that three canonical brain networks—default-mode, dorsal, and ventral—play critical roles in many high-level cognitive tasks. The goal of the present study was to investigate the three network regions' involvement in episodic memory retrieval. To this end, we performed meta-analyses of prior functional MRI studies using a variant of the Remember–Know paradigm as the behavioral task. The analyses yielded three main findings. First, default-mode network regions, including the anterior and posterior midline cortex, the angular gyrus, and the medial temporal regions, were associated with greater activity during Remember (recollection) than during Know (familiarity) responses. This is consistent with the view that the default-mode network supports self-referential processing. Second, the dorsal network regions, including the dorsal frontal and parietal cortices, were associated with greater activity during Know (weak memory) than during Remember (strong memory) responses. This is consistent with the view that the dorsal network mediates executive control processing. Third, the ventral network regions, including the ventral frontal and parietal cortices, the insular cortex, and the caudate regions, increased activity with increasing familiarity strength. This is consistent with the view that the ventral network supports salience processing. These findings clarify the differential contributions of the default-mode, dorsal, and ventral networks to episodic memory retrieval and also indicate that many episodic retrieval-related activations may actually reflect more general attention/executive operations. More generally, the findings suggest that many activations observed in functional neuroimaging studies are components of networks that respond in concert rather than regions activated in isolation.

Introduction

Lesions to the medial temporal lobe can abolish episodic memory function, whereas lesions outside of this area seldom cause amnesia, indicating that the medial temporal lobe is a key structure for the support of episodic memory. However, functional neuroimaging studies have shown episodic memory to be associated with widespread brain areas. For example, in an attempt to study episodic memory retrieval, which is the focus of the present study, many prior functional magnetic resonance imaging (fMRI) studies have investigated brain regions showing greater activity for old items that were correctly classified as “old” than for new items that were correctly classified as “new”. This “old/new” effect is associated with several brain regions, including the prefrontal, parietal, sensory/perceptual, and, occasionally, the medial temporal regions (for meta-analysis, see Spaniol et al., 2009). These widespread activations are unexpected given the neuropsychological evidence suggesting the central importance of the medial temporal regions, and their functional relevance is still being widely debated (e.g., Herron et al., 2004, Naghavi and Nyberg, 2005, Buckner and Carroll, 2007). Here, we present a global, system-wide model linking common retrieval-related activations to the activity of multiple large-scale brain networks and test three specific predictions of the model, based on meta-analyses of relevant prior fMRI studies.

It is widely accepted that the brain is organized into multiple distributed large-scale networks. Drawing from broad, qualitatively similar distinctions emerging from prior task-based fMRI studies (e.g., Raichle et al., 2001, Corbetta and Shulman, 2002, Wager and Smith, 2003, Dosenbach et al., 2006, Sridharan et al., 2007, Eckert et al., 2009) as well as from resting-state connectivity fMRI studies (e.g., Fox et al., 2005, Fox et al., 2006, Damoiseaux et al., 2006, De Luca et al., 2006, Vincent et al., 2006, Seeley et al., 2009), we found the following three canonical networks are important for attention-demanding cognitive tasks: (i) the default-mode network, which consists of the anterior and posterior midline cortex, the angular gyrus, and the medial temporal regions; (ii) the dorsal network, which includes the dorsal frontal and parietal cortices; and (iii) the ventral network, which includes the ventral frontal and parietal cortices, the insular cortex, and the subcortical regions. The default-mode network regions mainly mediate self-referential processing, or, more generally, internally oriented processing (Binder et al., 1999, Gusnard et al., 2001, Fransson, 2005, Fox et al., 2005, Buckner and Carroll, 2007, Golland et al., 2008), as indicated by high activations of the regions during imagining the future (prospection), mental navigation, conceiving the viewpoint of others (theory of mind), and mind-wandering (for meta-analysis, see Spreng et al., 2009). The dorsal network regions mainly support executive control processes, as indicated by high activity levels when the demand for controlled processing is maximal, such as during interference resolution, working memory, and response selection (e.g., MacDonald et al., 2000, Corbetta and Shulman, 2002, Wager and Smith, 2003, Nee et al., 2007). The ventral network regions mainly support salience processing, as indicated by activity increases in response to increasing degrees of subjective salience, whether perceptual, emotional, or homeostatic (e.g., Phillips et al., 1998, Corbetta and Shulman, 2002, Peyron et al., 2002, Seeley et al., 2007, Sridharan et al., 2007). A broad distinction among the three networks, like the present one, while ultimately should be made more specific, provides useful heuristics at this stage, which are helpful for guiding further research.

One of the major impetuses for distinguishing the dorsal from the ventral network was Corbetta and Shulman's (2002) attention model, in which the dorsal network supports goal-directed, top-down attention, whereas the ventral network mediates stimulus-driven, bottom-up attention. For example, one fMRI study using a search paradigm (Corbetta et al., 2000) found that the dorsal network showed increased activity as soon as the subject was given instructions to search for a target and maintained this increase throughout the search period. However, the activity increased in the ventral network when the subject finally detected the target. Corbetta and Shulman's attention model also presumes an interaction between the dorsal and ventral networks. For example, they discussed the evidence that the salience of objects, which is processed by the ventral network, is strongly influenced by their behavioral relevance and the evidence that the top-down selection of objects, which is mediated by the dorsal network, is modulated by bottom-up distinctiveness of objects in a visual scene. These bidirectional influences between the dorsal and ventral networks indicate that the distinction between the attentional functions of the two networks is not a sharp contrast but a more graded difference (Cabeza, 2008).

There have been two previous lines of research relating one or more of the three networks—default-mode, dorsal, and ventral—to retrieval-related activations. One such line has proposed that episodic remembering is supported mainly by the default-mode network (e.g., Buckner and Carroll, 2007, Schacter and Addis, 2007, Hassabis et al., 2007, Vincent et al., 2006). Empirical support for this view comes mainly from studies of autobiographical memory (for meta-analysis, see Svoboda et al., 2006). However, this view does not fit well with activations observed in most other retrieval paradigms, which usually involve more widespread regions, including the lateral frontal and dorsal parietal areas (e.g., Spaniol et al., 2009). The other line of research is the recently proposed Attention to Memory (AtoM) model (Cabeza, 2008, Cabeza et al., 2008, Ciaramelli et al., 2008). Extending the attention model of Corbetta and Shulman (2002) to the domain of memory, the AtoM model proposes that the dorsal posterior parietal cortex (PPC) supports top-down attention processes that are guided by retrieval goals, whereas the ventral PPC mediates bottom-up attention processes that are captured by salient retrieval output. Despite some empirical support (Ciaramelli et al., 2008), this model is ultimately limited by its restriction to the PPC regions. Moreover, the model may be limited even in its power to account for PPC activity, due to its neglect of the default-mode network. For example, the angular gyrus (the posterior part of the ventral PPC), which is a key component of the default-mode network, shows robust activations in most retrieval tasks (e.g., Shannon and Buckner, 2004, Wagner et al., 2005). In sum, previous studies to relate retrieval activations to multiple networks have been limited in scope due to considering only one or two networks and/or focusing on a small region of the brain.

Our hypotheses regarding roles of the default-mode, dorsal, and ventral networks in episodic retrieval were directed toward activations observed in the Remember–Know paradigm or its variants. One dual-process theory of recognition memory proposes that recognition can be based either on vividly remembering specific contextual details, or recollection, or on a feeling of oldness in the absence of contextual details, or familiarity (e.g., Yonelinas, 2002). Though there is a running debate about whether recollection and familiarity depend on qualitatively similar or different processes (e.g., Yonelinas, 2002, Dunn, 2004, Wais, 2008), the distinction has guided numerous cognitive, neuropsychological, and neuroimaging studies. To distinguish between recollection and familiarity memory, a Remember–Know procedure discriminates between old items that are “remembered” (recollection) and old items that are “known” (familiarity). Although there is no one-to-one direct correspondence, on average, Remember versus Know responses align more with recollection versus familiarity memory, respectively. A critical advantage of the Remember–Know paradigm for the present study was that it gave us the ability to construct three different activation contrasts, targeting the default-mode network, the dorsal network, and the ventral network (see below), within the single paradigm.

We made the following three predictions, based on the models of the default-mode, dorsal, and ventral network functions described above. First, as mentioned above, prior evidence suggests that the default-mode network supports self-referential processing. Self-referential processing is a strong component of recollection, but is at best weak in familiarity memory. Thus, we predicted that the Remember > Know contrasts would mainly activate default-mode network regions. Second, prior evidence indicates that the dorsal network mediates executive control processes. A Remember–Know procedure used to distinguish recollection versus familiarity also tends to separate strong memories from weak (Dunn, 2004, Wais, 2008). Controlled retrieval processes, such as iterative search, post-retrieval monitoring, and response selection, should be strongly engaged when memories are weak (Know), whereas there should be little need for control processing when memory strength is maximal (Remember). Thus, we predicted that the Know > Remember contrasts would mainly activate dorsal network regions. Third, prior evidence indicates that the ventral network supports salience processing. In a modification of the Remember–Know procedure (e.g., Yonelinas et al., 2005, Daselaar et al., 2006), subjects are asked to rate test items in terms of familiarity strength (from “highly unfamiliar” to “highly familiar”) or confidence (from “definitely new” to “definitely old”). These familiarity ratings are likely to track the salience of the retrieved output. Thus, we predicted that parametric contrasts involving increasing familiarity ratings would mainly activate ventral network regions.

To test the three predictions of the present study, we performed meta-analyses of relevant prior fMRI studies. A primary use for meta-analysis in neuroimaging is to identify significant concordances in patterns of brain activity across a set of independent studies using a specific task/paradigm (Wager et al., 2007). There is an empirical basis for proposing that meta-analysis is a more efficient and bias-free means of accomplishing this than a narrative review or even a tabular review would be (Laird et al., 2005b). The present study determined whether the consistently activated regions identified by the meta-analyses were consistent with our three predictions. As Wager et al. (2009) noted, a group of consistently co-activated regions across several independent studies is a potential candidate for a spatially distributed functional network in the brain. Thus, meta-analyses of neuroimaging studies offer one method for characterizing functional networks (Toro et al., 2008). The present meta-analyses involved fMRI studies using the Remember–Know paradigm or one of its variants as the behavioral task. There have been two prior reviews of this literature, one based on a tabular method (Skinner and Fernandes, 2007) and the other based on a meta-analytic method, which, however, was confined to the Remember > Know contrasts (Spaniol et al., 2009). The tabular review suggested relatively complex patterns of dissociations between recollection and familiarity, involving the frontal, parietal, sensory, and medial-temporal cortices. The meta-analytic review focused on comparisons of subjective versus objective recollection (e.g., source memory retrieval), and one of the main conclusions was that subjective, but not objective, recollection was associated with activation of the hippocampus. Thus, the goals and/or methods of prior reviews differ from the present study, and the critical contrasts investigated are also different.