An ROI-analysis of Activation in FG2, Amygdala lb and dlPFC : How are they Functionally Organized in a Face Working Memory task

University of Linköping | Department of Behavioural Sciences and Learning Master of Science in Psychology

Spring term 2020

An ROI-analysis of Activation in FG2,

Amygdala lb and dlPFC - How are they

Functionally Organized in a Face Working

Memory task?

Jonathan Mira

Kalle Österman

The Psychologist Programme consists of 300 academic credits taken over the course of five years. The programme has been offered at Linköping University since 1995. The curriculum is designed so that the studies focus on applied psychology and its problems and possibilities from the very beginning. The coursework is meant to be as similar to the work situation of a practicing psychologist as possible. The programme includes two placement periods, totaling 16 weeks of full time practice. Studies are based upon Problem Based Learning (PBL) and are organized in themes, Introduction 7,5 credits, Cognitive psychology and the biological bases of behavior, 37,5 credits; Developmental and educational psychology, 52,5 credits; Society, organizational and group psychology, 60 credits; Personality theory and

psychotherapy, 67 credits; Research methods and degree paper 47,5 credits.

This report is a psychology degree paper, worth 30 credits, spring semester 2020. Main supervisor Emil Holmer and associate supervisor Josefine Andin.

Department of Behavioral Sciences and Learning Linköping University

581 83 Linköping

Telephone +46 (0)13-28 10 00 Fax +46 (0)13-28 21 45

Department of Behavioural Sciences and Learning 581 83 Linköping

SWEDEN

2020-05-14

Language Report category ISBN

Swedish X English

Licentiate dissertation

Degree project ISRN LIU-IBL/PY-D—20/508—SE

Bachelor thesis

X Master thesis Title of series, numbering ISSN

Other report

URL Title

An ROI-analysis of Activation in FG2, Amygdala lb and dlPFC

- How are they Functionally Organized in a Face Working Memory task?

Authors

Jonathan Mira and Kalle Österman

Abstract

Working memory (WM) for facial identity and WM for facial expressions of emotions is important in everyday functioning and seems to have different neurobiological correlates. We investigated the level of neural activation in three regions of interest (ROI): the fusiform face area (FFA), dorsolateral prefrontal cortex (dlPFC), and amygdala; and how they are related to behavioral performance during an n-back task involving face stimuli with a complex background figure within an fMRI-paradigm. Participants performed three different 2-back tasks, one for facial expressions of emotions (EMO), one for the facial identity (ID), and one for a background figure presented behind the face (FIG). We hypothesized that the FFA would activate more in ID, the amygdala would activate more during EMO, and that the dlPFC would activate in all n-back tasks. An ROI analysis was done to extract mean activation values from the participants (N = 32) in the fusiform gyrus area 2 (FG2), the laterobasal amygdala (amygdala lb), and dlPFC in the different tasks. A one way repeated measures ANOVA revealed a similar activation in FG2 and

amygdala lb in both ID and EMO. During the FIG task higher activation in FG2 was shown in comparison with ID and EMO, and lower activation in amygdala lb was shown in comparison to ID. dlPFC was activated in all tasks. Furthermore, there was a negative correlation between amygdala lb activation and reaction time in the FIG task, where an abstract figure was kept in WM and facial information was to be ignored. These results indicate that the activation in FG2 and amygdala lb might not differ between WM for facial identity and WM for facial expressions of emotions, which is unexpected in comparison to perception studies where a difference in these nodes has been

reported for processing these two different types of information. This might suggest that the role of these neural nodes differ depending on WM load and task irrelevant features.

Abstract

Working memory (WM) for facial identity and WM for facial expressions of emotions is important in everyday functioning and seems to have different neurobiological correlates. We investigated the level of neural activation in three regions of interest (ROI): the fusiform face area (FFA), dorsolateral prefrontal cortex (dlPFC), and amygdala; and how they are related to behavioral performance during an n-back task involving face stimuli with a complex

background figure within an fMRI-paradigm. Participants performed three different 2-back tasks, one for facial expressions of emotions (EMO), one for the facial identity (ID), and one for a background figure presented behind the face (FIG). We hypothesized that the FFA would activate more in ID, the amygdala would activate more during EMO, and that the dlPFC would activate in all n-back tasks. An ROI analysis was done to extract mean activation values from the participants (N = 32) in the fusiform gyrus area 2 (FG2), the laterobasal amygdala (amygdala lb), and dlPFC in the different tasks. A one way repeated measures ANOVA revealed a similar activation in FG2 and amygdala lb in both ID and EMO. During the FIG task higher activation in FG2 was shown in comparison with ID and EMO, and lower activation in amygdala lb was shown in comparison to ID. dlPFC was activated in all tasks. Furthermore, there was a negative correlation between amygdala lb activation and reaction time in the FIG task, where an abstract figure was kept in WM and facial information was to be ignored. These results indicate that the activation in FG2 and amygdala lb might not differ between WM for facial identity and WM for facial expressions of emotions, which is unexpected in comparison to perception studies where a difference in these nodes has been reported for processing these two different types of information. This might suggest that the role of these neural nodes differ depending on WM load and task irrelevant features.

Keywords

Acknowledgments

First and foremost we would like to thank our supervisor Emil “Godér” Holmer and co-supervisor Josefine “Huettel” Andin. Thank you for your knowledge and patience, we are forever thankful for all the help and time you invested in us and this project. We had a blast.

We would also like to thank Peter ”Emergency Radio” Ekberg, Erik “Lander” Dehapapi Dehababa Wirén Ghasemi, and Elias “Nocco-loco” Elserud for the most wonderful, innovative and fruitful lunches and discussions. You kept us afloat during this deeply unsocial time. Thank you!

Contents

Introduction 8

Background 8

History of brain mapping 8

Functional magnetic resonance imaging - fMRI 9

Face processing 10

The core- and extended system 11

Working memory 12

WM in fMRI 13

Face perception and WM - how do they function together? 14

Aims 15 Hypothesis 15 Method 16 Participants 16 Ethical considerations 17 Stimuli 18 Conditions 18 Procedure 19 Experimental design 21 Image acquisition 21 Preprocessing 21 First-level analysis 23

Regions-of-interest (ROI) definitions 23

Data acquisition 23

Statistical analysis 25

Results 26

fMRI results 26

Emotion vs identity vs background figure 26

Behavioral measurements results 29

Accuracy 29

Reaction time 29

Behavioral performance correlations with relative activation across ROIs 30

Discussion 32

Activity in FG2 32

Activity in dlPFC 35

Behavioral performance 36

Future Directions and Limitations 37

Conclusions 38

References 40

Introduction

Face processing and its effect on human functioning is a truly complex process that gives rise to a series of questions. For example, how are we cognitively able to communicate and navigate within a social context by using facial expressions? Why is it that some individuals struggle with it while others strive? How are different brain lesions affecting the ability to process facial information?

The ability to process facial information might be one of the most highly developed visual perceptual functions, and it plays a vital role in social interaction (Haxby et al., 2000). In everyday life, social interactions are almost inevitable and effective interaction is

facilitated by the ability to accurately recognize whom you are dealing with as well as by correctly attributing meaning to the individual's expressions. Thus, the face provides a lot of information in social communication, for instance, identity, intention, and mood, which might affect how we interpret interactions (Jack & Schyns, 2017).

The aim of this study is to examine if working memory (WM) for facial identity and WM for facial expressions of emotions differs, as reflected in BOLD responses, across three regions of interest; fusiform face area (FFA), amygdala, and dorsolateral prefrontal cortex (dlPFC). We will investigate the neural activity by using an fMRI paradigm. Below we will start by giving a short background to brain mapping in order to introduce the field of brain studies to the reader, before describing fMRI as a measurement technique. Thereafter we will describe the cognitive and neurological aspects of face processing and WM.

Background History of brain mapping

In modern society, early research regarding the brain's anatomy and function is often defined by the case of Phineas. P. Gage, a young man who had his frontal lobe pierced by an iron rod (Harlow, 1848). Gage survived, but his personality changed drastically after the accident, from an inspiring and charismatic to a mean and impulsive one. The drastic change indicated that damage to the frontal lobe played an important part. Early theories about the brain's function and anatomy were based on a few cases of damages and lesions on brains, and 20 years after the case of Phineas Gage, Broca (1861) came in contact with a man named Leborgne who had a progressive loss of speech (Huettel, 2009). During Leborgne’s autopsy, Broca found a lesion in the left hemisphere of the frontal lobe. The conclusion he came to was that this area (today known as Broca’s area) is involved in speech production. Later on, Brodmann (1909) defined different areas in the cerebral cortex depending on the cellular

composition, such as cortical thickness and cell density (Huettel et al., 2009). The difference in cellular composition makes up the cytoarchitecture.

To exclusively study the brain's anatomy and functions through individuals with lesions has its obvious disadvantages. It involves fewer participants, is invasive (and

therefore potentially dangerous), and it is difficult to use in order to draw conclusions about possible functional brain networks (Huettel et al., 2009). More widely used methods apply measurement techniques that use different types of equipment that measure brain function. For example electroencephalography (EEG), magnetoencephalography (MEG), computed tomography (CT), positron emission tomography (PET), and magnetic resonance

imaging/functional magnetic resonance imaging (MRI/fMRI; Huettel et al.,2009). Experiments are conducted in a way that will create a physiological change in the brain through behavioral or cognitive tasks.

Measurement techniques are based on neuroimaging that depends on two variables: the spatial- and temporal resolution. Spatial resolution refers to the smallest object that can be resolved by a sensor. In practical terms this means that more numerous small sized voxels give more information than fewer and bigger sized voxels (Baars & Gage, 2013). Temporal resolution refers to in what time one is able to notice possible physiological changes. The most widely used equipment in cognitive neuroscience is MRI/fMRI. This is likely due to the delicate balance between spatial and temporal resolution, non-invasiveness and signal fidelity that MRI/fMRI has in comparison to other techniques (Bandettini, 2009).

Functional magnetic resonance imaging - fMRI

fMRI is based on the Blood-oxygenated-level-dependent (BOLD) signal (Huettel et al., 2009). In this way the measurement technique takes advantage of the hemodynamic response in strong magnetic fields (Ogawa & Lee, 1990). The hemodynamic response refers to the fast flow of blood to the blood vessel where the brain has been activated and neuronal metabolism has occurred. The hemoglobin molecule becomes deoxygenated after neuronal metabolism, but the blood flow supplies the neurons with a lot of new oxygenated hemoglobin (Glover, 2011). The hemoglobin molecules have different magnetic properties depending on if it is oxygenated or deoxygenated (Huettel et al., 2009). The magnetic properties of the

deoxygenated hemoglobin molecule create inhomogeneities, noise to the signal in the

magnetic field for a short time. MR can identify the inhomogeneities and produce images that take this noise into consideration, denoted as T2* (Huettel et al., 2009), and these are the images used in fMRI. A hemodynamic response can be produced by for example cognitive

tasks, and can therefore be studied in fMRI-experiments (Glover, 2011). However, it is worth noticing that there is always a difficulty with making psychological inferences via

manipulations in experimental designs (Huettel et al., 2009). This problem is related to reverse inferences, as we cannot be sure that the neuronal metabolism necessarily is connected to the cognitive task that we want to examine due to e.g. experimental

shortcomings, participants engaging in other thoughts than those we want to induce, physical movement, physiological noise or perhaps even neural metabolism that is not connected to the brain’s function. Researchers can try to control for this by experimental design, so that the BOLD-response has a more reliable connection to the cognitive task. They can also

preprocess the data in order to standardize and clean up the data before a statistical analysis.

Face processing

Face processing has been studied for many years related to, for example: prosopagnosia (face blindness; Hecaen & Angelergues, 1962), face recognition (Baron, 1979, 1981), and has been described in several theoretical models of face recognition (Bruce & Young, 1986; Valentine 1991; Valentine et al., 2016). The perception of faces has also been shown to differ from other types of perception, for instance, object perception with different categories e.g houses, cats, scissors, or shoes (Haxby et al., 2000, 2001).

The variety of information that can be communicated through the face implies a complex neuronal process (Haxby et al, 2000). There is robust evidence regarding that the lateral part of the fusiform gyrus, often described as the fusiform face area, (Kanwisher et al., 1997; Kanwisher & Yovel, 2006) is involved in face perception and recognition of the identity of the face (Haxby et al., 2000, 2002; Ishai, 2008). In other words, the FFA seems to have an involvement when processing invariant features of the face. Furthermore, the

fusiform gyrus is active in other cognitive tasks that involve face stimuli, e.g. WM

(Schweizer et al., 2019). Although there exists a lot of evidence of FFA:s involvement in face processing, there is still an ongoing discussion whether the region is a face-specific brain module (Kanwisher & Yovel, 2006), or if it also taps onto a visual expertise module that activates stronger during individuation of categories and in processing of complex stimuli for experts (Bilalic, 2016). According to the theory about FFA as a visual expertise module, humans are experts on faces, which would explain a stronger activation in FFA for faces than other visual stimuli.

Although the lateral fusiform gyrus is the most extensively mentioned brain area concerning face processing, it is not believed to be the only cortical or subcortical area

involved in face processing (Ishai, 2008). Face processing is instead often discussed in terms of a network model (Haxby & Gobbini, 2011).

The core- and extended system

Haxby et al. (2000) suggests that there is a core system consisting of occipitotemporal regions in the extrastriate visual cortex that is activated during face processing. Haxby & Gobbini (2011) also proposes that there is an extended system that helps in the processing of face stimuli. In the extrastriate visual cortex, the specific regions involved in the core system are the inferior occipital gyrus - the occipital face area (OFA), the FFA, and the posterior superior temporal sulcus (pSTS) (Haxby & Gobbini, 2011). Adjacent areas to these three are often significantly activated during this type of processing. The neuronal processes in the core system could be divided into two, where the first neuronal process is related to invariant face recognition which includes the comprehension of facial structure and identity of a person (Haxby et al., 2000, 2002). The second neuronal process is connected to the processing of changeable aspects of faces, as for example facial expressions, lip movement, and eye gaze.

In invariant face recognition, there is greater activation in the FFA and for changeable aspects, the pSTS has been shown to be more activated. The activation is not dichotomous as there seems to be a variation in activation pattern in both areas depending on, for example, eye gaze and facial expression (Calder & Young, 2005). The pSTS might even play more of a role in face perception integration than simply being responsible for changeable aspects. In other words, it is possible that pSTS is integrating information from the changeable aspects with the invariant face information (Calder & Young, 2005).

The extended system is not part of the core system but it affects and modulates the core system in the extrastriate visual cortex through a process called “top-down modulatory feedback” (Haxby & Gobbini, 2011, s.105). Factors that might recruit the extended system are facial expressions of emotions, motor simulation, and person knowledge. In the extended system the amygdala has consistently been shown to be recruited for facial expressions of emotions (Haxby & Gobbini, 2011). For example, human lesion studies have found that damage to the amygdala impairs the recognition of facial expressions of emotions (Adolphs, 2002). Studies also suggest that the amygdala receives visual information used to recognize facial expressions via two classes of input mechanisms: a direct subcortical process from the thalamus (an area thought to function as a relay station, by projecting all sensory information (except olfactory) to the cortex; Carlson, 2017), and through a cortical process via the visual cortex (Adolphs, 2002). For example Anders et al. (2004) have shown that the amygdala is involved in the processing of visual information of emotions projected directly from the

thalamus (in the absence of prior processing in the visual association cortex). In the study the researchers found that people with a phenomenon called affective blindsight, could still recognize emotions in facial expressions without the conscious awareness of seeing a person's face. Patients with affective blindsight have a deficit in visual abilities associated with

damage in the primary visual cortex (V1), but consequently do not seem to lack the ability to differentially categorize and respond to emotionally salient stimuli, which might depend on some mediation by the amygdala (Hamm et al., 2003). Given the current evidence, the amygdala seems to play a crucial role in processing facial expressions of emotions even though it is not part of the core system of face processing.

The neuronal processing of invariant face features and facial expressions of emotions explains some aspect of the neurocognitive processing that is involved during face

perception. In summary in the core system the FFA is the area that has shown the most consistent activation while processing invariant features of a face (Haxby & Gobbini, 2011). When processing facial expressions of emotions, the amygdala, connected to the extended system, consistently shows activation (Haxby & Gobbini, 2011).

Working memory

Everyday tasks such as filling up the coffee machine with the right amount of coffee powder, planning the workweek and reading course literature, requires the capacity to temporarily keep multiple steps and their intermediate consequences in mind. Likewise, social interaction, and in particular face processing, involves the temporary maintenance and manipulation of complex percepts, associated with multi-layered semantic representations (Schweizer et al., 2019).

In cognitive psychology, WM is regarded as a system that enables us to functionally solve multi-step tasks (Miyake & Shah, 1999). There are several models conceptualizing WM, but Baddeley and Hitch (1974) defined the field as they presented their

multi-component model of WM (Miyake & Shah, 1999). The model comprises multiple specialized components of cognition and includes a supervisory system; the central executive, and three specialized temporary memory systems; a visuo-spatial sketchpad, a phonological loop, and an episodic buffer. The central executive controls and regulates the informational flow to the temporary memory systems (the visuospatial sketchpad, the phonological loop, and the episodic buffer), as it is involved in the process of focusing and switching attention and in activating representations from long-term memory (Baddeley & Logie, 1999). The temporary memory systems keep memory traces active in mind, since they aid perceptual processing by

allowing the system to compare what is experienced now with what was seen or heard

moments ago. The visuo-spatial sketchpad maintains and manipulates visual information, and the phonological loop keeps auditory information active. The episodic buffer can hold

multidimensional episodes or chunks and is multimodal in that it is able to manipulate information from different perceptual modalities (Baddeley, 2007).

WM in fMRI

The mapping of WM in the brain has been of interest to researchers for decades. A brain area first established through lesion studies, playing a central role for WM related processing, is the prefrontal cortex (PFC) (Khan & Muly, 2011). Among others, Jacobsen and Nissen (1937) studied cerebral functions in primates, during delayed response tasks that required temporary memory processing of spatial location, where the subjects watched the

experimenter hide a raisin under one of two objects. The delay was constructed by lowering a curtain for a brief period of time (varying from a few seconds to minutes) in front of the objects, which required that the primate kept the spatial location of the raisin in mind before choosing whether the raisin was under the left or the right object. Jacobsen and Nissen (1973) found that primates with damaged PFCs performed worse compared to primates with

functioning PFCs, even after shorter delays. Since then, the PFC has been thought to be one of the brain regions primarily related to what we today refer to as WM (Aben et al., 2012; Lezak, 2012). The PFC has also been connected with other cognitive processes, including cognitive control (Miller, 2000), and stimulus related abilities such as encoding of stimuli and sustained attention to stimuli (D'Esposito et al., 1998). These PFC related cognitive

components could be seen as related to the theoretical concept of the central executive in Baddeley's multicomponent model (Baddeley & Logie, 1999; Kahn & Muley, 2011). In a study by Wilson et al. (1993), the role of the PFC as a form of general WM was

questioned as they found that different parts of the PFC are specialized for different kinds of tasks. However, LeDoux (1998) argued that these findings did not rule out the existence of a system in the brain, consisting of a general-purpose workspace and a set of executive

functions, that coordinates the activity of the temporary memory systems (the visuospatial sketchpad, phonological loop, and episodic buffer) in WM. But there are still difficulties that arise when trying to map executive and attentional functions onto a brain structure like the PFC, as the region is large and heterogeneous, both structurally and functionally (Kane & Engle, 2002; Khan & Muly, 2011).

An area within the PFC strongly related to general WM functions is the dorsolateral PFC (dlPFC; Aben et al., 2012; Curtis & D’Esposito, 2003; Kane & Engle, 2002; Owen et

al., 2005). The dlPFC has been frequently identified in research concerning WM processing. In lesion studies of the dlPFC, researchers have found general impairments of the monitoring and manipulation of information in WM (Petrides, 2000). In n-back tasks, one of the most widely used WM tasks in fMRI research, there is also support for dlPFC involvement in WM processing (Lezak et al., 2012; Mencarelli et al., 2019; Owen et al., 2005). Furthermore, the

n-back task has good face validity as a WM task since it involves the maintenance,

continuous updating, and processing of information (Gajewski et al., 2018). In the n-back task, the subject is asked to report when the same, common component, stimulus item appears “n” steps back from the item at hand. And as n increases, WM load is increased neurobiologically (Wang et al., 2019). For example, Wang et al. (2019) found that, when comparing different n-load in the n-back task (1-back vs 2-back) an increased activation in dlPFC was shown with a higher “n”. This finding lends further support to the role of dlPFC in WM, as the dlPFC activity seems to have a positive correlation with higher WM load.

Face perception and WM - how do they function together?

Face perception and WM play an important part in everyday functioning (Haxby et al., 2000; Baddeley & Hitch, 1974). But how do they function together? The studies that have mainly guided the aim of this master thesis is the experimental fMRI study by Neta and Whalen (2011), concerning WM for facial expressions of emotions and facial identity, and an fMRI meta-analytic review regarding how affective versus neutral information influence WM, by Schweizer et al. (2019). Both of these studies examine face processing in the context of WM. Neta and Whalen (2011) found differences in accuracy and reaction time when participants were asked to remember the identity or the facial expression of emotion in a 2-back task. Participants answered more accurately and faster in the identity task. Neta and Whalen (2011) also reported consistent activation in the dlPFC in both conditions. Moreover, task specific neural activation was also reported. More specifically, when participants remembered facial identity compared to facial expressions of emotions more activity was registered in the fusiform gyrus. Further, when facial expressions of emotions were stored in WM, more activity was registered in the amygdala.

Schweizer et al. (2019) meta-analytic review focused on WM studies including facial stimuli, and presented a comparison between neutral or affective tasks concerning this type of stimulus. Schweizer et al. (2019) results of relevance for this master thesis was, in particular, that they found negligible negative behavioral effects as a result of affective versus neutral tasks for healthy individuals. Although the effects were negligible, participants answered

more accurately and faster in identity (neutral) tasks than in tasks concerning emotional expressions (affective tasks) (Schweizer et al., 2019), which supports Neta and Whalens (2011) findings. Schweizer et al. (2019) highlighted some large scale neural networks involved in WM processing of affective or neutral stimuli. More specifically these networks are: the dorsal frontoparietal control network, which shows more involvement while

processing neutral stimuli, and the ventral salience network, which shows higher involvement while processing affective stimuli (Schweizer et al., 2019). An important node for WM-processing in the frontoparietal control network seems to be the dlPFC, and the amygdala seems to play an important role in emotion recognition and processing in the salience network.

To expand on this research we will try to further examine how the FFA, amygdala, and dlPFC activates during WM processing of face stimuli and how level of activation relates to behavioral performance.

Aims

The aim of this master thesis is to examine if the neural activity, as reflected in BOLD responses, differs between WM for facial identity or WM for facial expressions of emotion across three regions of interest (ROI); FFA, amygdala, and dlPFC. We will also examine if the relative activation in the chosen ROIs is related to WM performance. In addition, we will explore if the facial stimuli affects the activity in the ROIs, when facial information is

irrelevant for solving the task and is to be suppressed. Participants will perform three different 2-back tasks, one for facial expressions of emotions (EMO), one for the facial identity (ID), and one for a background figure presented behind the face (FIG). The ROIs in the present study were chosen in accordance with previous work. More specifically, the FFA has been shown to play a vital role in face perception (Haxby et al., 2000, 2002; Ishia, 2008). The amygdala has consistently been activated while looking at affective stimuli and is proposed to be involved in the processing of facial expression of emotion (e.g., Adolphs, 2002; Neta & Whalen, 2011; Sabatinelli et al., 2011, Schweizer et al.,2019). The dlPFC has consistently been shown to be involved in WM tasks (Owen et al.,2005; Schweizer et al., 2019; Wang et al., 2019).

Hypothesis

Does the activation across FFA, amygdala, and dlPFC differ between WM for facial identity or facial expressions of emotion? We hypothesise 1) that there will be a stronger activation for FFA in WM for facial identity compared to WM for facial expression of emotions (Haxby, 2000, 2002; Neta & Whalen, 2011), and 2) stronger activation for amygdala in WM for facial expressions of emotions compared to WM for facial identity (Neta & Whalen, 2011; Schweizer et al., 2019). We also predict 3) that dlPFC will be activated during all WM tasks (Owen et al., 2005; Schweizer et al., 2019; Wang et al., 2019).

Our second question regards the behavioral performance as we want to investigate if it differs between WM for facial expressions of emotions and WM for facial identity and if it is related to the activation in the ROIs? We hypothesise 1) better performance (higher

accuracy and faster reaction time (RT)) in WM for facial identity compared to facial expression of emotion (Neta & Whalen, 2011; Schweizer et al., 2019). Furthermore we hypothesise 2) that higher activation in the amygdala and dlPFC will have a negative correlation with accuracy for facial expressions of emotions (Neta & Whalen, 2011) and 3) higher activation in FFA and dlPFC will have a negative correlation with RT for facial identity (Neta & Whalen, 2011). Finally, we will explore the activity in each ROI in a visual WM task, when facial information isirrelevant for solving the task and is to be suppressed. We will also examine how the behavioral performance in this task is related to the level of activation.

Method Participants

Participants were recruited through informational flyers and personal communication. 42 individuals reported interest to participate, and after email contact with the administrator all individuals needed to answer a form with screening questions. These questions allowed administrators to exclude participants that did not fulfill inclusion criterias. The screening regarded disabilities, visual impairments, handedness, age, native language and

claustrophobia.

There is some evidence of psychiatric conditions affecting the neuronal activation pattern, for example some studies with WM tasks have indicated that patients with major depression disorder (MDD) are processing valenced stimuli differently than healthy controls (Levens & Gotlib, 2010). Bipolar disorder seems to be related to a reduction in activation in the dlPFC and might also compensate with other areas (Townsend et al., 2010) and some studies about autism and face perception also indicated a different activation in the amygdala (Perlman et al., 2011). Normal or corrected to normal vision is needed to be able to respond

correctly to the stimuli, and left handedness may imply difficulties in analysis of data due to possible different activation patterns in the brain (Kolb & Whishaw, 2015). An age restriction was used due to a difference in BOLD-activation that has been shown in older adults

compared to younger adults in n-back tasks (45 years and older compared to 17-30 years; Wang et al., 2019). A total of nine individuals were excluded from the study due to two reporting neuropsychiatric conditions, seven reporting left handedness. One participant was excluded from the analysis after the experiment due to an incidental finding.

That left 32 healthy adults participants for analysis(18 female), with a mean age of 23.0 (SD = 2.92; ranging between 19 and 30 years old). Participants performed in the average range on standardized tasks of both non-verbal cognitive ability; the Visual puzzles subtest from WAIS-IV (measured in scale points; M = 11.4, SD = 2.5), and verbal WM; the Letter-Number subtest from WAIS-IV (M = 9.7, SD = 1.6). The participants had normal or corrected-to-normal vision, were right handed, native Swedish speakers, and did not report any ongoing history of significant neurological or psychiatric conditions

Ethical considerations

The above mentioned recruitment and screening process of participants was applied to ensure the quality of the data. In an ethical point of view this is motivated since the time using the MRI scanner is valuable, both in an economical- and health care perspective, which implies that we as scientists should take the necessary measures to more effectively utilize the time that was allocated.

Screening and inclusion procedures were also executed to guarantee the participants safety during the experiment. Even though MRI is regarded as a relatively safe and

noninvasive imaging technique, the magnetic properties of an MRI scanner is strong enough to damage implanted metal objects, such as aneurysm clips or pacemakers, in participants (Huettel et al., 2009). The static magnetic field in the scanner is able to pull objects containing for example iron, nickel and cobalt, toward its bore with a violent force, and interactions between implanted metal objects and the magnetic field could potentially cause participants severe tissue heating and burns. Thus, one safety measure was to screen

participants so that no one with potentially harmful medical implants/devices or metal in their body entered the machine.

The ethical problem in fMRI research regarding that the image acquisition might reveal abnormalities in the brain that might impose an health issue, is something that researchers need to consider before running the experiment. Therefore in our study,

participants were informed, and gave consent to the possibility of being contacted by health professionals in case of incidental findings. Images were then examined by clinical health professionals; medical consultants administered the first screening process and prominent images were sent to local clinicians who made the final decision whether to disclose the information to the participant and further examine the participant, or not.

Written informed consent was obtained from all participants as well as an ethical clearance from the local ethical review board in Linköping, Sweden. Participants were compensated with 1000 SEK for their participation.

Stimuli

The face stimuli used in the images were taken from the “Radboud Faces Database” (Langner et al., 2010). The “Radboud Faces Database” is a set of face stimulus items intended to be used in fMRI studies on emotions (Langer et al., 2010). The set of faces in the database was validated in a study by Langner et al. (2010), who showed an overall 82% agreement rate between the intended and the chosen expression. The selected image stimuli consisted of four individual female faces displaying four different emotional expressions (angry, happy,

fearful, and sad), which made up a total of 4 x 4 different stimuli images. Only images of women were used during the experiment as we wanted to avoid any, for the study,

uninteresting variance associated with differences in the item images. The choice to use only items depicting emotional expressions from women was due to research showing that women tend to show greater expression of emotions overall (Chaplin, 2015), which could help the participants to recognize the correct emotion and hopefully contribute to a more reliable activation during the fMRI. Furthermore all images had background figures behind the face stimulus that were visible on both sides of the face (see Fig. 1.). The perceptual control condition images had the same structure as the other images, with the exception that some of the background figures were coloured black instead of grey.

Conditions

The conditions in the study were framed within a n-back paradigm, or more specifically a 2-back paradigm. The choice to use this paradigm was due to it being used in previous

neuroimaging research of both face processing and WM (Neta & Whalen, 2011; Schweizer et al., 2019). In the n-back task, the subject is, as earlier explained, asked to report when the same common component stimulus item appears as “n” (number of steps) back from the item at hand. To exemplify this, the 2-back task in our experiment demands that the subject

answers “yes” when the current image shows the same emotion/identity/figure stimuli as in the image that was presented 2 steps back. The task process is regarded as a useful task investigating WM, as it fits the theoretical assumptions made about a multicomponent WM-model, as defined by Baddeley (Baddeley & Logie, 1999).

To specify the conditions further, the participants completed four different types of tasks in the study (see Fig. 1.). Each 2-back condition contained the same stimuli but the task differed. The four tasks were emotional expression (“is it the same expressed emotion as 2 steps back?”), facial identity (“is it the same individual as 2 steps back?”), background figure (“is it the same background figure as 2 steps back?”), and a visual control task (“is the figure in the background black?”). Correct responses on the tasks in the trials were 31 % “yes”-answers and 69 % “no”-“yes”-answers, which were distributed randomly across the conditions.

Figure 1

Illustration of the different tasks including the control task and the 2-back conditions, from the left; CONTROL, EMO, ID and FIG

Procedure

All participants took part in a behavioural testing session before the fMRI session, and were tested concerning non-verbal cognitive ability. Before entering the scanner, participants were instructed to respond as accurately and as quickly as possible during the presentation of each trial, this by pressing one of two buttons using their right hand. By using their index finger they answered “YES” on the left button and by using their middle finger they answered “NO” on the right button. Participants also practiced the four conditions (EMO, ID, FIG, and

CONTROL) in one block before running the experiment. If the participant failed to answer the trials in this block, or for some other reason did not answer, the test administrator started the block over again until all practice trials were answered.

When participants were installed in the scanner, instructions were repeated. In the scanner, participants viewed the images through goggles that were attached on their heads while in the machine. Stimuli were presented using the Presentation software (Presentation version 10.2, Neurobehavioral Systems Inc., Albany, CA). Each trial started with a 5000 ms period during which a cue displayed on the screen indicated which task was to be performed next. The cues (translated from swedish) were: “2-back, emotion” (for the emotional

expression task), “2-back, identity” (for the facial identity task) “2-back, background figure” (for the background figure task), and “black background figure?” (for the control task). After the cue, the stimulus was displayed for 1500 ms while the participant responded, which was followed by a fixation cross (used as a jitter) that was presented five times for either 500 (n = 10), or 950 ms (n = 5) (jittered with a half TR= 450 ms, see e.g. Perini et al. (2018)).Task presentation was blocked, and there were 16 trials per block. Thus, each block lasted for 34 000 ms. Between blocks, there was a 12 000 ms break and a fixation cross was presented. Participants were instructed to move as little as possible. In total, there were 4 runs with 8 alternating blocks in each (see Fig. 2.).

Figure 2

Example of order between conditions within and across runs for one participant.

Block 1 Block 2 Block 3 Block 4 Block 5 Block 6 Block 7 Block 8

Run 1 2-back, ID + 2-back, EMO + 2-back, FIG + Control task + 2-back, FIG + Control task + 2-back, EMO + 2-back, ID Run 2 2-back, EMO + Control task + 2-back, ID + 2-back, FIG + 2-back, ID + 2-back, FIG + Control task + 2-back, EMO Run 3 2-back, FIG + 2-back, ID + Control task + 2-back, EMO + Control task + 2-back, EMO + 2-back, ID + 2-back, FIG Run 4 Control task + 2-back, FIG + 2-back, EMO + 2-back, ID + 2-back, EMO + 2-back, ID + 2-back, FIG + Control task

Experimental design

A design was used as an experimental design in the present study. The blocked-design is a way to separate the experimental conditions into distinct blocks which enables the measuring of the BOLD signal caused by neuronal activity, assumed to reflect distinct states of brain functioning such as WM (Huettel et al., 2009). The choice to add jittering to the experimental design was due to the fact that these five fixations in each block enabled the possibility to analyze the data through a mixed block/event-related design, something that was not used for the purposes of the present master thesis. The order of conditions was counterbalanced within participants across runs, and between participants. This was done as we wanted to observe the specific signal for each condition, which is possible by this semi-randomization as it contributes with balancing the results and reducing the risk of any effects that depend on the order of presentation of condition (Huettel et al., 2009).

Image acquisition

All subjects were scanned on a Siemens 3T Prisma scanner, 64 channel head-coil. Visual stimuli were presented with a PC running the software Presentation® (Presentation version 20.0, Neurobehavioral Systems Inc., Albany, CA). The participants viewed the images through VisuaStim Transducer® goggles (Resonance Technology Company, Inc.). A buttonbox (Lumina LS-RH®, from Cedrus) was used to record participants' behavioral responses. Cushions minimized head movement.

T1-weighted images were collected using a high-resolution 3D magnetization-prepared rapid gradient echo sequence with voxel size: 0.9*0.9*0.9mm, echo time [TE]= 2.36, [TR] = 2300ms, 208 slices, field of view [FOV]= 288x288, inversion time [TI] = 900ms, flip angle = 8°.

Functional images were acquired using functional T2*-weighted imaging sequence: 3*3*3mm, TR = 901 ms, 45 slices, 457 volumes per run (interleaved/simultaneous acq.), sensitive to blood-oxygenation-level-dependent (BOLD) contrast. AC-PC aligned, with no gap (spacing 3mm = no interslice gap; [TE] = 30ms, [FOV]= 204x204, flip angle = 59°).

Preprocessing

The goal of the preprocessing of data in fMRI studies is to remove uninteresting variability from the data and to prepare it for statistical analysis (Huettel et al., 2009). The preprocessing was done using Statistical Parametric Mapping Software 12 (SPM12; Wellcome Department

of Cognitive Neurology, London, UK) implemented in MATLAB (Mathworks, Natick, MA, USA).

A slice time correction was done to shift every voxel's time series to modulate the sampling to act like it was done simultaneously, since the two-dimensional MRI acquisition process only allows data to be acquired one slice at a time. Furthermore we controlled for small head movement by realignment. This was done to be able to make the assumption that every voxel is an image of the same region at every time point measured in the machine. By realigning, the images are corrected in the time series to a single reference image and resliced in order to create realigned versions of the original data. In this way realignment contributes to avoid a possible mismatch of the location of subsequent images in the time series as an effect of head motion (so called bulk motion).

Functional images were then coregistered with the MRI images, collected in the beginning of the session, as a way to allow for visualization of single-subject task activation added upon the individuals anatomical information. This simplified later transformation of the fMRI images into a standard coordinate system, which in turn facilitates anatomical identification of block-related activation.

Segmentation of brain tissue in the anatomical images was applied as a means to help determine the tissue class of a voxel. In SPM 12, segmentation is combined with spatial normalization (further described below) and bias field correction, so that the prior probability that any voxel contains gray or white matter can be determined using a probabilistic atlas of tissue types. This probability is then combined with the data from the image to help

determine and differentiate the tissue class of one voxel from another with the same voxel intensity (Poldrack et al., 2011).

Normalization was applied as a way to transform MRI data from an individual subject to match the spatial properties of a standardized image (Huettel et al., 2009). The goal of spatial normalization in fMRI is to compensate for shape differences between subjects by mathematically molding the images of each brain so that they are the same as those of every other brain using standardized coordinates.

Finally, the images were smoothed using an isotropic 6-mm full-width-half maximum gaussian kernel. By smoothing the fMRI data the signal-to-noise ratio increases by reducing the signal detection of non-systematic high-frequency spatial noise. The smoothing also helps validate the statistical assumptions and remove artefacts.

First-level analysis

The preprocessed data were applied in the first level analysis using SPM 12, in order to further analyze the fMRI data. In a first level analysis a design matrix is specified since the matrix defines the task design that the imaging data will be modeled to. In the first level analysis of this study (the within subject level) three contrasts were modeled, contrasting each experimental condition to the control condition (which was used as a baseline), for every participant. The contrasts, which consist of differences in BOLD activation, were modelled as following: emotional expression > control (EMO > control), facial identity > (ID > control), and background figure > control (FIG > control). These contrast files were later used to extract the ROI values from the data in Anatomy toolbox (version 2.2b) and MarsBar (version 0.44).

Regions-of-interest (ROI) definitions

The choice to apply a region of interest (ROI) analysis was a result of the a priori predictions of the outcome, as the method is often used to better correct for the differences in anatomy between different subjects and also maintaining statistical power (Saxe et al., 2006). In an ROI analysis, the fMRI signal used for statistical tests corresponds to the mean activation within the predefined region, instead of voxel by voxel as in the whole-brain approach (Poldrack, 2007).

ROI analyses were conducted using the programs: Anatomy toolbox for amygdala and fusiform gyrus; and MarsBaR for dlPFC, both running under SPM 12 (Wellcome Department of Imaging Neuroscience, London, United Kingdom). Anatomy toolbox uses probabilistic maps to define the ROI:s for amygdala and fusiform gyrus, whereas dlPFC is defined in MarsBar by using specified coordinates.

Data acquisition

Due to a large number of missing data (“not a number”, NaN) values in the ROI analyses of the data in two (the centromedial nuclei group-, and the superficial nuclei group of the amygdala) of the four possible amygdaloid complex ROIs available in Anatomy toolbox, we narrowed down our analysis of the ROI comprising the amygdala to the laterobasal nuclei (Amygdala lb; Amunts et al., 2005; see Fig. 3.). This decision was due to 1) time restrictions precluding us from solving this problem by replacing the NaN values in a new preprocessing of the data, and 2) the available data revealed that there was a correlation between the four amygdala regions of r ~ .70, indicating similar level of activation across sub-regions. The amygdala data that were the most robust came from the amygdalostriatal transition zone

(amygdala asTR) and the laterobasal nuclei amygdala (amygdala lb), where the amygdala lb had the least missing data, only lacking values from one participant. This was handled by excluding this participant in later statistical analysis of the amygdala lb. In addition to this, one previous study showed that the amygdala lb seems to be a site of integration for sensory information, reported to have axonal connections with sensory areas, such as the visual and auditory cortex as well as the thalamus (Bzdok et al., 2013).

The anatomical area fusiform gyrus 2 in Anatomy toolbox (FG2; Caspers et al., 2013; see Fig. 3.) was chosen as our ROI since Caspers et al. (2013) and Zhang et al. (2016)

reported evidence indicating that the FG2 comprises the FFA. The region selected to

anatomically represent dlPFC was not available in Anatomy toolbox, thus it was based on the WM load sensitive mask defined by the “2-back > 1-back” used in Wang et al. (2019; see Fig. 3; Appendix 1 for coordinates).

Mean values from all the three chosen ROIs for each participant were then extracted using Anatomy toolbox for amygdala lb and FG2, and MarsBar for dlPFC. These values were then imported into SPSS statistics 25 (IBM Corp, SPSS Statistics, Version 25.0, IBM

Corporation, Armonk, NY) for statistical analysing.

Figure 3

The chosen regions of interest in the study. (A) Anterior view of the brain depicting the FG2 (violet), the

amygdala lb (red), and dlPFC (blue); (B) The three regions viewed through a Sagittal (left)-, Coronal (middle)-, and axial (right) plane; (C) multislice presentation of the regions

Statistical analysis

The statistical analysis of the contrast values from the ROI-analysis, as well as the accuracy and RT for each condition were performed in SPSS statistics 25 (IBM Corp, SPSS Statistics, Version 25.0, IBM Corporation, Armonk, NY). A one way repeated measures analysis of variance (ANOVA) were used to examine if there was significant difference in relative activation between conditions: relative activation in FG2 (EMO vs ID vs FIG); relative activation in amygdala lb (EMO vs ID vs FIG); and relative activation in dlPFC (EMO vs ID vs FIG); accuracy (EMO vs ID vs FIG); RT (EMO vs ID vs FIG). Pairwise comparisons with Bonferroni corrections were performed to investigate where the differences were. Bonferroni corrections were used due to the explorative nature of the FIG-condition. Pearson's

correlations were performed within conditions (EMO, ID & FIG) for each behavioral measure respectively (accuracy & RT) and the relative activation in each ROI (FG2, amygdala lb & dlPFC).

Shapiro-Wilk test as well as histograms were used to check if the data was normally distributed. If an outlier was extreme (3 IQR and above) or if a mild outlier (1.5-3 IQR) created a non-normally distributed data it was excluded from the analysis. Mauchly’s test was performed to check if the assumption of sphericity was violated in the one way repeated measures ANOVA. Mauchly’s test indicated that the assumption of sphericity was violated in relative activation FG2 and dlPFC ANOVAs (FG2 χ2_{(2) =10.64, p = .005 (ε =.80), dlPFC,}

χ2(2) =17.45, p = < .001 (ε =.69)). Huynh-Feldt correction was used if the epsilon value was > .75. Greenhouse-Geisser correction was used if the epsilon value was < .75. In order to

receive an confidence interval for the correlations in SPSS bootstrapping was conducted on the 95th percentile (1000 samples) in all correlations.

Results fMRI results

Emotion vs identity vs background figure

The one way repeated measures ANOVA with Huynh-Feldt correction, revealed a

statistically significant effect of condition in activity for FG2, F(1.61, 47.74) = 8.18, p = .002, η2 _{= .21. Pairwise comparisons using Bonferroni correction revealed a significant difference}

between condition FIG towards EMO, MD = 1.29, SE = .42, p = .013 and FIG towards ID,

MD = 1.15, SE = .36, p = .011 (see Table 1 & Fig. 4.). No significant difference between ID

and EMO MD = .14, SE = .25 p = 1.000. The results were not in line with our hypothesis for FG2. The confidence interval does not contain 0 in any condition, this revealed that the activation in FG2 in all conditions is significantly higher than the baseline activation (see Fig. 4.).

Table 1

Mean relative activation in FG2 in the different conditions and pairwise comparisons of the mean differences in relative activation in FG2 between conditions

Figure 4

Mean relative activation in FG2 in the different conditions

The one way repeated measures ANOVA for amygdala lb revealed a statistical significant effect of condition on activation, F(2, 58) = 7.84, p = .001, η2 _{=.21. Pairwise comparisons}

with Bonferroni corrections revealed a significant difference between the FIG and ID condition, MD = -.28, SE = .08, p = .005 (see Table 2 & Fig. 5.). No significant difference between ID and EMO MD = .13, SE = .06 p = .09 or between FIG and EMO MD = -.15, SE = .07 p = 0.15. The results were not in line with our hypothesis for amygdala lb. The

confidence interval does not contain 0 in any condition, this revealed that the activation in amygdala lb in all conditions is significantly lower than the baseline activation (see Fig. 5.).

Table 2

Mean relative activation in amygdala lb in the different conditions and pairwise comparisons of the mean differences in relative activation in amygdala lb between conditions

Figure 5

Mean relative activation in amygdala lb in the different conditions

The one way repeated measures ANOVA for dlPFC with Greenhouse-geisser correction revealed a statistical significant effect of condition on activation in dlPFC, F(1.39, 43.02) = 8.30, p = .003, η2_{=.21. Pairwise comparisons with Bonferroni corrections revealed a}

significant difference between the FIG- and EMO conditions, MD = -1.30, SE = .40, p = .008 and EMO and ID, MD = -.64, SE = .20, p = .010 (see Table 3 & Fig. 6.). No significant difference between ID and FIG, MD = .65, SE = .33, p = 0.162 (see Table 3 & Fig. 6.). That dlPFC had a relative activation that was higher in all conditions, in comparison to baseline, was in line with our hypothesis. The confidence interval does not contain 0 in any condition, this revealed that the activation in dlPFC in all conditions is significantly higher than the baseline activation (see Fig. 6.). This was in line with our hypothesis.

Table 3

Mean relative activation in dlPFC in the different conditions and pairwise comparisons of the mean differences in relative activation in dlPFC between conditions

Figure 6

Relative activation in dlPFC in different conditions

Behavioral measurements results

Accuracy

Performance (percentage of correct trials) was calculated separately for EMO, ID and FIG for the subjects in all trials. The one way repeated measures ANOVA revealed a significant main effect of condition, F(2, 60) = 44.90, p<.001, η2 = .60 and pairwise comparisons with

Bonferroni corrections revealed that participants were more accurate in the ID condition than the EMO condition and the FIG condition (p<.001; p<.001; see Fig. 7.), as well as more accurate in the EMO condition compared to the FIG condition p = .003 (mean ± standard error: EMO = 80.5% ± 1.2, ID = 86.6% ± 1.1, FIG = 76.6% ±1.0; see Fig. 7.).

Reaction time

RT was calculated separately for the three conditions. A condition (EMO, ID, FIG) one way repeated measures ANOVA for RT revealed a significant effect of condition, F(2,58) = 50.90, p<.001, η2 = .61 and pairwise comparisons with Bonferroni corrections revealed that RTs were significantly longer for the EMO and FIG condition than the ID condition (p<.001;

p<.001; mean ± standard error: EMO = 843 ms ± 10, ID = 730 ms ± 14, FIG = 831 ms ± 18;

Figure 7

Percentual accuracy across conditions

Figure 8

Reaction time across conditions

Behavioral performance correlations with relative activation across ROIs

One significant negative correlation was found between behavioral performance and relative activation in the FIG condition, between RT and amygdala lb, r(30) = -.45 CI [-.729, -.023], p = .013 (see Table 6 & Fig. 9). No significant correlation in the EMO or ID condition was found (see Table 4 & 5). The results were not in line with our hypothesis.

Table 4

Correlations for the EMO condition

Table 5

Correlations for the ID condition

Table 6

Correlations for the FIG condition

Figure 9

Discussion

The aim of this study is to examine if the neural activity differs between WM for facial identity or WM for facial expressions of emotion across three ROIs; FFA, amygdala, and dlPFC. In addition, we investigate if the relative activation in the chosen ROIs is correlated to WM performance. Lastly, we explore if task irrelevance has an effect on activation. At first, we will discuss the relative activation in each ROI and the difference between conditions. Thereafter, a discussion of the behavioral data and its correlations with relative activation will be made. At last a conclusion will be made on how this information could be integrated into a theoretical framework of WM for facial identity or WM of facial expressions of emotions.

Activity in FG2

That FG2 shows activation in all conditions was expected since all items contain face stimuli with invariant features (Haxby & Gobbini, 2011). In FG2 we expected a higher activation in the ID condition, because it is hypothesized to activate WM for facial identity, in contrast to the EMO condition, which is hypothesized to activate WM for facial expressions of emotions (Neta & Whalen, 2011). However, we found no support for this hypotheses in our study. Instead there is a significant higher activation in the FIG condition in comparison to the EMO-condition and the ID-condition. The difference can perhaps be explained by the still quite unexplored functional properties of FG2 (Winawer et al.,2010; Caspers et al., 2013). Although the FG2 is believed to contain FFA (Caspers et al., 2013; Zhang et al., 2016), and earlier empirical research have shown FFA to be involved in face recognition and face perception for invariant structures in faces (Haxby & Gobbini, 2011), the FFA has also been found to contain clusters of neurons that are distinctly linked to non-facial stimuli (Grill-Spector et al., 2006; Bilalic, 2016). Furthermore, Chang et al. (2015) report that the FG2 activation also correlates with higher frequency of letter stimuli on a horizontal meridian than on a vertical meridian. The heterogeneity in function for the FG2 between individuals in general terms is believed to be due to differences in cytoarchitectonic structure in the area, where different fiber connections relate to somewhat different functions (Chang et al., 2015).

In some studies of other WM tasks than n-back, differences in brain activation is also thought to be dependent on whether the object that is being manipulated is task relevant or task irrelevant (Schweizer et al., 2019). In the present experiment this could be translated into that, although all participants look at the same stimuli in all conditions, the relative activation in different areas will vary depending on how stimuli is represented in WM. The level of representation in WM is thought to be regulated by the conditions that guide the participant to

what part of the image that is task relevant or not. Since FG2 has been shown to be activated while looking at both face related and non-face related stimuli, the higher activity in FG2 for the FIG-condition might be explained by non-face related clusters of neurons that were activated in order to process the background figure behind the face when the background figure was task relevant. The task relevant non-face related cluster activation in the FIG-condition might have a greater activation than the task related face-selective clusters activation in the ID-condition in FG2. It is also possible that both the non-face-related clusters and the face-selective clusters are activated at the same time due to the image

structure, but that the task relevance factor moderates the activation more for the background figure than the facial identity. The horizontal meridian placement of the background figure might also increase the activation in FG2, as the faces are placed vertically, which might have resulted in less activation in the ID condition (Chang et al., 2015).

That there is no significant difference in FG2 between the ID and EMO condition is not easily explained. Perhaps although facial identity is task irrelevant in the EMO-condition, it still activates FG2 similarly as if it would be task relevant. These findings are not in line with previous studies, and therefore we conclude that it is possible that the background figure has some impact in this case, since the background figure is not part of the face stimuli in Neta and Whalens (2011) study. On a representational level in WM, attentional resources (i.e common pool resources; Lavie et al., 2004; Park et al., 2007) could be devoted to

differentiating between the face- and the background figure representation, instead of solely focusing on differentiating between the representation of facial identity and the representation of facial expressions of emotions. As a result there might not be enough common pool

resources to contribute to a statistical significant difference in FG2 activation between ID and EMO.

In summary, our findings in FG2 indicate that this region has other functional properties than just the perception of facial identity. Our findings also support the claim that task relevance might affect the activation in FG2. Finally, the finding that WM for facial identity, and WM for facial expressions of emotions show no significant difference in FG2 could perhaps be explained by the background figures impact on attentional resources in WM.

Activity in amygdala lb

The results revealed a lower activation in amygdala lb in all conditions in comparison to the control task. The relatively lower activation in amygdala lb could possibly be explained by

the fact that more complex tasks direct the neuronal activation towards the goal directed processes and away from other processing, such as the processing of emotions (Sörqvist et al., 2016). In this case, all conditions in comparison to the control task involve WM

processing as the goal directed process whereas the control task, the baseline, only involves perceptual processing. Therefore the activation in amygdala lb might be lowered in all conditions compared to the control task due to the ongoing WM processing.

The results also suggest a significant effect of condition in amygdala lb. But the difference in relative activation between conditions was not in line with our hypothesis since the EMO-condition did not represent the highest activation in amygdala lb. Instead, there was no significant difference between the EMO condition and ID condition, but a significant higher activation in ID in comparison to FIG. The EMO-condition should, according to earlier research (Neta & Whalen, 2011; Schweizer et al., 2019), more reliably activate the amygdala lb than the ID condition. This is due to the EMO condition involving the task relevant WM for facial expressions of emotions and the ID condition involving the task relevant WM for facial identity (Neta & Whalen, 2011; Schweizer et al., 2019). Along the lines of the discussion about FG2 and the similar activation in both conditions the same conclusion could perhaps be made here. It is possible that a similar amount of resources in WM are put into processing facial expressions of emotions in amygdala lb in both task relevant (EMO) and task irrelevant (ID), if the participant needs to differentiate the face stimuli from the background figure.

The results also revealed that the ID condition activates the amygdala lb more than the FIG condition. This could possibly be explained by two main reasons. Firstly, it might be easier to suppress task irrelevant facial expressions of emotions, when trying to process the task relevant background figure in comparison to the task relevant facial identity. Perhaps it is more difficult to differentiate between facial identity and facial expressions of emotions because they are included within the same object, whereas the background figure is separate from the face object. Secondly, the significant difference in activation in amygdala lb

between ID and FIG could also be explained by the difference in how cognitively demanding the tasks are (Sörqvist et al., 2016). That the participants show a lower accuracy as well as slower RT in the FIG condition compared to the ID condition implies that it is more difficult and cognitively demanding. According to this reasoning the FIG-condition should have the lowest activation in amygdala lb due to the higher cognitive demand, which is also the case according to the results. The thesis about cognitive demand should be regarded with some

caution since the EMO condition also shows significantly lower accuracy than the ID condition but no difference in amygdala activation was found.

In summary, our findings in amygdala lb support the claim by Sörqvist et al. (2016) that activation in amygdala is lowered when the cognitive demand increases. The finding that WM for facial expressions of emotions and WM for facial identity show no difference in amygdala lb could perhaps indicate that the background figures impact attentional resources in WM.

Activity in dlPFC

Consistent with our hypothesis, the results show that all the conditions activated the dlPFC in comparison to the control task. This activation was in line with results from similar

experimental studies using the n-back paradigm (e.g., Owen et al., 2005; Neta & Whalen, 2011; Schweizer et al., 2019; Wang et al., 2019). Therefore, our findings lend further support to theories connecting the dlPFC with WM as we show that these results are also replicable in tasks requiring processing in WM for facial identity and WM for facial expressions of

emotions. Our study also sheds new light to the field since the baseline measure used here (the control task) provides new information in comparison to Neta and Whalen (2011) who only used a fixation cross baseline. We show that the higher activity in dlPFC in the

conditions is not a result of perceptual brain activity but rather reflects WM processing. The results showing that WM for facial expressions of emotions activates dlPFC significantly more than both WM for facial identity and a WM task in which facial

information was task irrelevant (the FIG condition), was not expected. However, WM studies have shown some evidence pointing towards that neutral task distractors, in comparison to affective distractors, will activate the dlPFC more due to the neutral value of the distractor (Schweizer et al.,2019). Although there is not any distractor in our experimental design, the task irrelevant features might have a distracting effect. A possible explanation for the higher activation in dlPFC in the EMO condition, could be attributed to more neutral task irrelevant features in the EMO condition than in the ID and FIG conditions. This should be regarded with caution in relation to our design, because Schweizer et al. (2019) also found that task relevant neutral features activates the dlPFC more than task relevant affective features. This would mean that task irrelevant neutral features activate dlPFC more than task relevant neutral features, but the investigation of this hypothesis is outside the scope of our study.

It is also possible that the higher activation in dlPFC could be related to Okon-Singer et al. (2015) who points out that when emotional information is processed in WM, it is poised

to hijack endogenous attention and other kinds of top–down control mechanisms. These mechanisms could be linked to the process of soft prioritization (Pessoa, 2009), a process where low level arousal has shown to enhance WM processing. In this case, low level of arousal from the task relevant facial expressions of emotions could be thought to enhance the activity in dlPFC.

In summary, our results support the view of dlPFC being involved in WM processing independent of stimuli. It is not totally clear why WM for facial expressions of emotions shows a higher activation in dlPFC, but it could perhaps be related to the neutral task irrelevant features or the process of soft prioritization.

Behavioral performance

The results show that accuracy in the ID condition is higher than in the EMO condition, which in turn is higher than FIG. For response time (RT), participants are faster in the ID condition than both the EMO and FIG condition. No difference is found between EMO and FIG for RT. This is in line with the previous study by Neta and Whalen (2011) and Schweizer et al (2019), and in accordance with our hypothesis. The higher accuracy and faster RT in the ID condition imply that the ID condition was easier to perform in than other conditions. We suggest that WM processing might be more difficult and take longer time for participants when they need to process the facial expressions of emotions if it is task relevant than it would take to suppress it if it is task irrelevant. This could theoretically be related to how the face perception system is activated in the WM task (Haxby & Gobbini, 2011). The ID condition activates the core system and only needs to suppress the facial expressions of emotions in the extended system, whereas the EMO condition involves the extended system but suppresses the core system. Maybe the level of processing in the core system affects both the accuracy and RT positively in comparisons to more processing in the extended system.

That participants have lower accuracy in the FIG condition in comparison to the ID and EMO could be explained by it potentially being more cognitively demanding. To be able to accurately answer the tasks in the FIG condition, the participant needs to hold the complex background figure, whose forms and shapes are more novel to the participants in contrast to faces and facial expressions of emotions. This should logically be more difficult than holding invariant face features and facial expressions of emotions in WM, since this is something participants practice in their everyday life.

We did not find support for our predicted correlations between relative activation and behavioral performance. Instead, one negative correlation with performance was found