Alzheimer’s Disease Progressively Alters the Face-Evoked Visual-Processing Network

Abstract

Background:

Postmortem studies of Alzheimer’s disease (AD) brains not only find amyloid-β (Aβ) and neurofibrillary tangles (NFT) in the primary and associative visual cortical areas, but also reveal a temporally successive sequence of AD pathology beginning in higher-order visual association areas, followed by involvement of lower-order visual processing regions with disease progression, and extending to primary visual cortex in late-stage disease. These findings suggest that neuronal loss associated with Aβ and NFT aggregation in these areas may alter not only the local neuronal activation but also visual neural network activity.

Objective:

Applying a novel method to identify the visual functional network and investigate the association of the network changes with disease progression.

Methods:

To investigate the effect of AD on the face-evoked visual-processing network, 8 severe AD (SAD) patients, 11 mild/moderate AD (MAD), and 26 healthy senior (HS) controls undertook a task-fMRI study of viewing face photos.

Results:

For the HS, the identified group-mean visual-processing network in the ventral pathway started from V1 and ended within the fusiform gyrus. In contrast, this network was disrupted and reduced in the AD patients in a disease-severity dependent manner: for the MAD patients, the network was disrupted and reduced mainly in the higher-order visual association areas; for the SAD patients, the network was nearly absent in the higher-order association areas, and disrupted and reduced in the lower-order areas.

Conclusion:

This finding is consistent with the current canonical view of the temporally successive sequence of AD pathology through visual cortical areas.

Keywords

Alzheimer’s disease FAUPA functional areas of unitary pooled activity task-fMRI visual functional network

INTRODUCTION

Alzheimer’s disease (AD) is characterized by two hallmark neuropathology findings: extraneuronal plaques consisting primarily of aggregated amyloid-β peptide (Aβ) and intraneuronal neurofibrillary tangles (NFT) composed of aggregates of hyperphosphorylated tau protein [1, 2]. Aβ and NFT activate astrocytes and microglia, inducing production of proinflammatory factors such as interleukins and nitric oxide, and ultimately resulting in excessive neuroinflammation, oxidative stress, neuronal damage, and cell death [2 –5]. Aβ plaque deposition is associated with synaptic network dysfunction and longitudinal cognitive decline, and Aβ and tau pathology correlates with cognition in mild cognitive impairment [6, 7]. Postmortem studies of AD brains show the presence of Aβ and NFT in the primary and associative visual cortical areas [8, 9]. A morphologically heterogeneous population of NFT-bearing neurons can be found in primary visual cortex (Brodmann Area 17) along with a considerable number of threads representing degenerating axons in the underlying white matter [8]. In comparison to the mean number of NFT in area 17, NFT is observed to increase 20-fold in the immediately adjacent visual association cortex of area 18, and shows a further doubling in area 20, the higher-order visual cortex of the inferior temporal gyrus [9]. These observations suggest that AD may alter visual processing function. Indeed, neuropsychological and neuroimaging findings have demonstrated that functions of both the dorsal and ventral visual streams are impaired during the progression of AD [10 –14]. Aβ and NFT aggregations may alter not only the local neuronal activation but also associated functional network activity.

The non-invasive neuroimaging technique of blood oxygenation level dependent (BOLD) functional magnetic resonance imaging (fMRI) is an effective tool for studying large-scale systems-level functional organization of neural activity in the human brain [15 –17]. In a typical task-fMRI study, the subjects perform tasks following a task paradigm. The most popular statistical approach of identifying the task-evoked activation map is the general linear model (GLM) [18, 19]. In this model, an expected ideal response is generated first based on the task paradigm, and then fit to the BOLD signal time course on a voxel-by-voxel basis, yielding a map of t or z statistics. Then, a threshold with a chosen significance level is used to identify the activated voxels, producing an activation map. This GLM assumes that task-induced BOLD signals behave similarly to the expected ideal response, and any deviations from the ideal response can be attributed to noise. This assumption may be valid for healthy subjects, but could be invalid for patients who have difficulty performing the task properly. Thus, more appropriate methods need to be developed for better analyzing the image data of task-fMRI studies.

The discovery of functional areas of unitary pooled activity (FAUPAs) with fMRI is reported in a recent study [20]. A FAUPA is defined as an area in which the temporal variation of the activity is the same across the entire area, and new techniques are used to identify FAUPAs that involve an iterative aggregation of voxels dependent upon their intercorrelation [21]. FAUPAs are determined objectively and automatically, based on the assumption that the temporal variation of the activity is the same across the entire area within a FAUPA. The determination requires no a priori knowledge of the activity-induced ideal response signal time course. This method enables us to identify FAUPAs that are associated with a specific task [20]. Using the signal time course of a task-associated FAUPA may identify the functional network specific for that task, and comparing these task-evoked networks between healthy controls and those with neurologic diseases, such as AD, may reveal the relationship between the functional network changes and the disease.

In this study, 19 AD patients and 26 age-matched healthy controls undertook a task-fMRI with the task of viewing face photos. We conducted three analyses: 1) we validated the existence of FAUPAs for these elderly subjects to replicate the findings of a previous study with young and middle-aged adults [20]; 2) we investigated the effect of AD on the face-evoked visual-processing network. To do so, we first identified a task-associated FAUPA in area V1 and then used the signal time course of this FAUPA, instead of using an expected ideal response of the GLM, to determine the visual-processing network; and 3) we compared the signal time course of the FAUPA with the expected ideal response to test the validation of the GLM assumption for AD patients.

METHODS AND MATERIALS

Participants

Eight severe AD (SAD) patients [6 female, ages from 55 to 85 years old with mean (MN)±standard deviation (SD) = 71.3±13.2], 11 mild/moderate AD (MAD) (8 female, ages from 67 to 86 with MN±SD=77.6±6.2), and 26 healthy senior (HS) controls (18 female, ages from 55 to 89 with MN±SD=74.0±6.2) participated in this study. (One HS control’s fMRI images were disrupted and unanalyzable, and consequently this HS subject’s data were removed from further analysis.) All participants took part in a larger behavioral study of pain responses in AD [22]. HS subjects had no subjective memory complaints and were recruited via newsletters and AD support groups. AD subjects were patients in the Cognitive & Geriatric Neurology clinic at Michigan State University. Diagnosis of probable AD was made by a geriatric neurologist (AB) based on the fourth edition of the diagnostic and statistical manual of mental disorders (DSM-IV) [23] and the National Institute of Neurological and Communicative Diseases and Stroke/Alzheimer’s Disease and Related Disorders Association (NINCDS-ADRDA) [24] criteria and included serum testing to rule out of infectious and metabolic causes of memory loss, neuropsychological testing to rule out depressive pseudodementia, and structural MRI to rule out vascular or other structural cause of dementia. We excluded anyone with: Type II diabetes, history of stroke or transient ischemic attack, central or peripheral neuropathy, and diagnosis of additional neurological (e.g., epilepsy) or psychiatric disorders (e.g. major depression, schizophrenia) other than AD. Acetylcholinesterase inhibitors and anti-depressant medication use was not exclusionary.

Neuropsychological tests

After screening, neuropsychological testing took place, including completion of Mini-Mental State Examination (MMSE) [25], and Cornell Scale for Depression in Dementia (CSDD) [26]. No subjects had a CSDD score indicative of probable depression (>12) [26]. As in prior studies by the authors [22, 27], we defined MAD as MMSE 11-23 and SAD as MMSE≤10. For details please refer to our previous studies [22, 27].

Testing took place in accordance with the Declaration of Helsinki and all protocols were approved by the Michigan State University Internal Review Board. Written informed consent was obtained personally or via named guardians or health care power of attorney for all HS as well as AD subjects. Verbal assent was obtained from participants before all testing procedures.

Task paradigm

Each participant undertook a task-fMRI including three 6 min runs. The task paradigm consisted of a total of 9 trials divided into 3 different conditions: viewing unfamiliar faces (UFs), familiar faces (FFs), and recent self-photos (SPs) taken from different angles (Fig. 1, top panel). Each trial was comprised of a 25 s task period followed by a 15 s rest period. During the task period, one face photo was presented for 5 s with a total of 5 photos presented, and participants passively viewed each presented photo. The order of trial presentation was pseudo-randomized for the three runs (Fig. 1). During the 15 s rest period, participants were asked to focus their eyes on a fixation cross mark at the center of the blank screen and were asked to try not to think of anything in particular.

Fig. 1

Illustration of the task paradigm and the three 6 min task-fMRI runs. The task paradigm consists of nine 25 s task-on periods interleaved with 15 s task-off periods. During a task-on period, a face photo was displayed for 5 s and a total of five photos were displayed (top panel). During the 15 s task-off period, the subjects were asked to focus their eyes on a fixation cross mark at the center of the blank screen and were asked to try not to think of anything in particular. For illustrative purpose, the three sample face photos shown are from an online database of face photographs [28]. SP, self-photos taken from five different angles; UF, unfamiliar face; FF, familiar face.

Photograph stimuli

Recent SPs of each subject were all obtained by the experimenter. Five different photos were obtained via electronic camera under mildly varied angles (direct front view, two parasagittal views of the left and right face, and photos looking slightly up and down at the subject). Five FFs were chosen amongst recent photos provided by the subject or family members. Such faces included spouses and immediate family members such as siblings, adult children, and grandchildren. Any individuals who could be photographed by the experimenter were photographed as such. However, many photos utilized in the study were electronic or scanned portrait photographs in the possession of the patient, spouse, or other family members. Utilized photographs were those that provided a clear, unobstructed and direct view of the face with good resolution. For each subject, five UFs were obtained from an online database of face photographs [28]. Utilized photographs were of similarly aged (within five years), gendered, and raced individuals. For all photos, portions of the body not including the head/neck were cropped out and the background was removed in Adobe Photoshop and replaced with a white background. All photos had coloration neutralized with black & white filtering and were additionally modified to maintain a consistent brightness and contrast in Adobe Photoshop.

Image acquisition

Functional brain images were acquired on a GE 3T Signa® HDx MR scanner (GE Healthcare, Waukesha, WI) with an 8-channel head coil using a gradient echo Echo-Planar-Imaging pulse sequence with these parameters: 38 contiguous axial slices, slice thickness = 3 mm, TE (time of echo)/TR (time of repetition) = 28/2500 ms, flip angle 80°, FOV (field of view) = 220 mm, matrix size = 64×64, and first 4 time points discarded. A MR-compatible Hyperion digital projection system (Psychology Software Tools, Inc., Pittsburgh, PA), placed at the back of the magnet room, was used to present visual stimuli, and the stimuli were projected onto a vertical screen with a 23×30 degree of visual angle placed inside the magnet bore. The stimulation presentation was controlled by a PC equipped with E-Prime (Psychology Software Tools, Inc., Pittsburgh, PA), and triggered by the very first RF pulse to be synchronized with the image acquisition. The participants viewed the screen through a mirror mounted on top of the head coil. For the participants who needed vision correction, MR-compatible lenses were used. Head movement was minimized by restraint using a piece of tape and cushions. Each participant undertook three 6 min task-fMRI scans, and each scan yielded a total of 144 volume images (total time points N = 144). Prior to each run participants were instructed to pay attention to the photo when a photo was presented on the screen and to the fixation cross mark at the screen center when no photos were presented. A live-view eye camera was used to ensure participants stayed awake and appeared to attend to visual stimuli, and those who started to drowse were instructed to remain awake and view the photos via the intercom microphone on the scanner operator’s console. (Subject eye-movements were not recorded during the scanning.) To identify anatomical regions, 180 sagittal T₁-weighted 1 mm³ isotropic volumetric inversion recovery fast spoiled gradient-recalled images (10 min scan time), with cerebrospinal fluid (CSF) suppressed, were obtained to cover the whole brain with the following parameters: TE/TR = 3.8/8.6 ms, time of inversion = 831 ms, TR of inversion = 2332 ms, flip angle = 8°, FOV = 25.6 cm×25.6 cm, matrix size = 256×256, slice thickness = 1 mm, receiver bandwidth =±20.8 kHz).

Image preprocessing

A standard image preprocessing stream was performed using AFNI [20, 29]. It included removing spikes, slice-timing correction, motion correction, spatial filtering with a Gaussian kernel with a full-width-half-maximum of 4.0 mm, computing the mean volume image, bandpassing the signal intensity time course to the range of 0.009 Hz–0.08 Hz, and computing the relative signal change ΔS (%) of the bandpassed signal intensity time course for each of the three runs of each subject. ΔS = [S (t) - S₀] × 100/S₀ (%), where S(t) is the signal time course and S₀ is the mean of S(t). (As eye-movements were not recorded during the scanning, any potential eye-blink induced artifacts were not removed.) After these preprocessing steps, the three signal time courses from the three runs were sorted according to the three face categories, and a concatenated signal time course was reconstructed for each category. Further image analysis was carried out using in-house developed Matlab-based software algorithms.

FAUPA determination

A statistical model and algorithms were developed, implemented in Matlab and fully evaluated to identify FAUPA [20, 21]. The determination of a FAUPA involved an iterative aggregation of voxels dependent upon their intercorrelation. This determination consisted of two major procedures. 1) The algorithm, with a first statistical criterion, identified a stable region-of-interest (ROI) in which the signal time courses of all voxels showed a similar temporal behavior; and 2) the algorithm, with a second statistical criterion, determined whether this stable ROI satisfied the condition of being a FAUPA by comparing the temporal behavior of signal time course of the voxels within the FAUPA with those bordering the FAUPA. For further details, please see reference [20].

Identification of task-associated FAUPAs and the task-evoked visual-processing network

For each participant and each face category, we identified a task-associated FAUPA in the putative primary visual area (V1), and used its signal time course as a reference function to compute the Pearson correlation coefficient (R) map in the original space. Then, using AFNI we converted all R maps to a standard template space (icbm452, an averaged volume of 452 normal brains) for group analysis. Atrophy is a concern for older brains, particularly brains for SAD patients, and a brain structure difference between two subject groups could lead to a significant increase in detected functional differences if this structure difference is not handled properly when converting the R maps to the standard space [30]. To avoid smoothing-reduced R-values in these atrophy-affected areas during the warping, we selected the resampling mode of nearest neighbor, which ensured that the local R value remains almost unaffected. For each group and each face category, we thresholded the mean R map at R > 0.7 (N = 144, p < 4.5×10^–17) to yield a face-evoked visual-processing network for the group and the category. For each face category, to compare the visual-processing network among the three groups, we generated three group-specific masks. A SAD mask composed the visual-processing network of the SAD patients alone; a MAD mask composed the visual-processing network of the MAD patients excluding the SAD mask; and a HS mask composed the visual-processing network of the HS excluding both MAD and SAD masks. We use these masks to investigate the association of network disruption with AD severity. For group statistical tests, the R values were converted to Z values through Fisher’s Z transformation to improve the normality of the distribution, and then Bonferroni correction was applied to multiple comparisons [31].

Validation of the GLM assumption for AD patients

The GLM [18, 19] assumes that task-induced BOLD signals behave similarly to the expected ideal response, and any deviations from the ideal response can be attributed to noise. Letting n(t) denote task-evoked underlying neural activity, the task-induced BOLD signal s(t) is the convolution of n(t) with the hemodynamic response function (HRF) that reflects the neurovascular coupling mechanism [32 –34]. The fMRI measured BOLD response y(t) = s(t) +δ(t), where δ(t) represents the unavoidable physiological and measurement noise unrelated to n(t). To determine whether a voxel is activated by the task, the GLM first generates an expected ideal response x(t) that is the best estimate of s(t), and then fits it with the BOLD response y(t) using the equation y(t) =β x(t) + e(t), where β is the parameter estimate for x(t) and e(t) is the error term that accounts for the residual error between βx(t) and y(t). When βx(t) = s(t), the residual error e(t) =δ(t). In order to get the best possible fit of the model to the BOLD response y(t), the task paradigm is used to convolve it with the HRF to generate the ideal response x(t), assuming n(t) follows the task paradigm design. Accordingly, the GLM assumes that x(t) behaves similarly to s(t). When x(t) behaves similarly to s(t), the assumption is valid. When x(t) substantially deviates from s(t), the assumption is not valid. To test the validity of this assumption for AD patients, we generated an ideal response signal time course based on the task paradigm, using the 3dDeconvolve program in AFNI with the model convolution kernel BLOCK (http://afni.nimh.nih.gov/afni). We then computed the relative signal change of this ideal response signal time course by subtracting the signal time course with its mean and dividing that change with the mean. Then, for each face category and each participant we computed the temporal correlation R of this ideal response with the BOLD signal time course of the selected task-associated FAUPA in area V1. For each face category, we compared the R between the HS and AD groups to test whether there were any significant differences in the R between the two groups. (The AD group consisted of the MAD and SAD patients together.) The GLM assumption was deemed valid if there were no significant difference between the two groups; otherwise, it was deemed invalid. An invalid assumption would render the GLM analysis in error.

RESULTS

Neuropsychological tests

There was no age or educational attainment difference between HS and AD groups (Table 1). The AD subjects showed a significant decrease in the MMSE score compared to HS (two-tailed t-test, p = 2.2×10^–11), and a significant increase in the CSDD (two-tailed t-test, p = 1.5×10^–10). When subdivided, MAD and SAD sub-groups showed no statistical difference in the CSDD, but a significant difference in the MMSE score (two-tailed t-test, p = 2.7×10^–7).

Table 1

Subject demographics (mean±standard deviation)

	Participants (N)	Age (y)	Education (y)	MMSE	CSDD
	HS (25)	74.0±6.2	17.2±2.9	29.1±1.1	1.1±1.3
	AD (19)	74.9±9.9	16.3±2.4	14.1±8.3	8.5±4.1
	MAD (11)	77.6±6.2	15.5±2.6	20.2±3.8	8.5±4.2
	SAD (8)	71.3±13.2	17.3±1.8	5.6±3.9	8.6±4.3
p (two-tailed t-test)	HS versus AD	0.71	0.28	2.2 × 10^–11	1.5 × 10^–10
	HS versus MAD	0.12	0.12	9.3 × 10^–13	2.4 × 10^–9
	HS versus SAD	0.42	0.94	2.2 × 10^–23	1.0 × 10^–8
	MAD versus SAD	0.17	0.13	2.7 × 10^–7	0.96

HS, healthy senior controls; AD, Alzheimer’s disease; MAD, mild/moderate AD (MMSE 23-11); SAD, severe AD (MMSE≤10); MMSE, Mini Mental State Examination; and CSDD, Cornell Scale for Depression in Dementia (normal range 0–12). Bold indicates the p-value of a significant group difference.

FAUPA identification

For each participant, FAUPAs were detected for each of the three reconstructed signal time courses. Although the total number of FAUPAs varied substantially from participant to participant, the group-mean number of FAUPAs showed a similar distribution for all three face categories (Fig. 2a). Each FAUPA’s signal time course was computed as the mean signal time course of all voxels within the FAUPA. Within each FAUPA, the overall correlation between all voxels is characterized by the mean of the Rs between the FAUPA’s signal time course and all voxels’ signal time courses. For each participant, the mean R-value of all FAUPAs was remarkably similar for all three face categories, and the group-mean of these mean R values was almost identical for all three face categories and all three subject groups (Fig. 2b). The total number of voxels in a FAUPA varied from 3 to 29. For the UF category, a similar distribution of FAUPAs as a function of the number of voxels was observed for all three subject groups (Fig. 3a). The other two face categories of FF and SP also showed a similar distribution for all three subject groups as that of the UF category (Supplementary Figures 1a and 2a). The group-mean of the mean number of voxels per FAUPA also showed a similar value for all three face categories and all three subject groups (Fig. 2c).

Fig. 2

Comparison of the FAUPAs detected for the three face categories and three participant groups. (a) The group-mean number of the FAUPAs; (b) The group-mean R value; and (c) The group-mean number of voxels per FAUPA. HS: healthy senior; MAD: mild/moderate AD; SAD: severe AD; UF: unfamiliar face; FF: familiar face; and SP: self-photo. The error bars indicate the corresponding standard deviations.

Testing the separation of FAUPAs from their adjacent areas

The separation of a FAUPA from its adjacent area was tested by comparing the R-values of all voxels within the FAUPA with those of the voxels bordering the FAUPA (one-tailed t-test) [20]. (The R was computed between the FAUPA’s signal time course and each voxel’s signal time course.) The FAUPA was considered to be separated significantly from its adjacent area if its corresponding p < 0.05 (multiple comparisons uncorrected). For the UF category, Fig. 3b shows a similar histogram of the group-mean p values for all three subject groups. The other two face categories of FF and SP also showed a similar histogram for all three subject groups as that of the UF category (Supplementary Figures 1b and 2b). Among all those having p values≥0.05, the largest percentage value was 1.49%, showing that more than 98.5% of them were separated significantly from their adjacent areas.

Fig. 3

Comparison of the histograms of the FAUPAs detected for the unfamiliar face (UF) category and three participant groups. (a) The histogram of group-mean FAUPAs as a function of the total number of voxels contained in a FAUPA; (b) The histogram of group-mean p values as a function of p-value range. HS, healthy senior; MAD, mild/moderate AD; SAD, severe AD. The error bars indicate the corresponding standard deviations.

Identification of task-associated FAUPAs

Individual-level, task-associated FAUPAs were identified in area V1 (Fig. 4). For each participant, the three selected FAUPAs (one for each of the three face categories) had an overlapped area as illustrated in the top three panels in Fig. 4. The task-induced BOLD signal changes were conspicuous for each of the three face categories, with variation from trial to trial, from category to category, and from subject to subject. For each face category, the group-mean signal time course reflected the task paradigm and showed similar behavior for the three groups of subjects (Fig. 4, bottom three panels).

Fig. 4

Illustration of the selected task-associated FAUPAs for a representative SAD patient and the group-mean signal time courses of the selected FAUPAs for the three participant groups. Top three panels: the red clusters in the left images represent the selected task-associated FAUPAs in the putative V1 for the three face categories, respectively. These three FAUPAs have an overlapped area that contained five voxels as shown in (c). For each FAUPA, the right plot shows the signal time courses of these five voxels and their mean signal time course as well. For each time course, the magnitude represents the percentage signal change relative to the mean signal of the time course. The dashed-lines represent the ideal response. (The ideal response was scaled by a factor of 2 for a better visual illustration.) Vox: voxel. Bottom three panels: for each of the three face categories, the plot shows the group-mean signal time courses of the selected FAUPAs for the three participant groups, respectively. HS, healthy senior; MAD, mild/moderate AD; SAD, severe AD.

Visual-processing network disruption of the AD

For each face category, Fig. 5 illustrates the identified face-evoked visual-processing networks for the HS, MAD, and SAD patients, respectively. Within each participant group, the visual-processing network was almost identical across the three face categories, demonstrating the independence of the network from the face category. For the HS, in the ventral pathway the visual-processing network started from area V1 and ended within the fusiform gyrus. However, the extent of this visual-processing network was disrupted and reduced in the MAD patients, and nearly absent in the SAD patients as clearly illustrated in the left three columns in Fig. 5. This gradual disruption to the visual-processing network from the HS to the SAD patients was also reflected in the three group-specific masks (Fig. 5, right column). The SAD mask covered an area that was mainly within the visual cortex. The MAD mask, excluding the area covered by the SAD mask, covered an area that extended both within and outside the visual cortex, but barely reached the fusiform gyrus. The HS mask, excluding the areas covered by the MAD and SAD masks, covered an area that encompassed much of the fusiform gyrus.

Fig. 5

Illustration of the face-evoked visual-processing networks in the ventral pathway (left three columns) identified for the three face categories and three subject groups. For each subject and each face category, a Pearson correlation coefficient (R) map was first computed using the signal time course of the selected task-associated FAUPA as the reference function, and then, for each group, in the standard space the group-mean R map was computed and thresholded with R > 0.7. The face-evoked visual-processing network was confined within the visual cortex for the severe AD, but extended to the fusiform gyrus (FG) for the healthy senior controls with that in between for the mild/moderate AD. For each face category, a group-specific mask was generated for each participant group as shown in the right column. UF, unfamiliar face; FF, familiar face; and SP, self-photo.

Association of visual-processing network disruption with AD severity

For each face category, we assessed the integrity of the visual-processing network by computing the mean R of the network using the three masks for each group (Fig. 6). For the SAD mask, the mean R showed a similar value for all three subject groups, regardless of the face category. For the MAD mask, the mean R showed a similar value for both the HS and MAD patients, but a significantly reduced value for the SAD patients, again regardless of the face category (Bonferroni t-test, the overall p < 0.005 for the 3 comparisons), demonstrating a disrupted and reduced network from the MAD to SAD patients. For the HS mask, the mean R was significantly reduced not only for the SAD group but also for the MAD group, again regardless of the face category (Bonferroni t-test, p < 0.02), demonstrating a disrupted and reduced network from the HS to MAD patients. Overall, AD disrupted and reduced the face-evoked visual-processing network, and the more severe the disease, the greater the disruption and reduction.

Fig. 6

Group comparisons of the three visual-processing networks for the three face categories. For the SAD mask, the group-mean R showed no statistical difference among the three groups of the healthy senior (HS), mild/moderate AD (MAD) and severe AD (SAD) subjects, regardless the face category. For the MAD mask, the R was significantly reduced for the SAD compared to either the MAD or the HS (Bonferroni t-test, max p < 0.005), but showed no statistical difference between the HS and MAD, demonstrating a significantly disrupted and reduced visual-processing network from MAD to SAD. Again, this result is independent of the face category. For the HS mask, the R showed a significant reduction for both the SAD and MAD compared to the HS (Bonferroni t-test, max p < 0.02), further demonstrating a disrupted and reduced visual-processing network from HS to MAD. This result is also independent of the face category. UF, unfamiliar face; FF, familiar face; SP, self-photo. The error bars indicate the corresponding standard deviations.

Validation of the GLM assumption for AD patients

For each face category, the R between the ideal response and the signal time course of the selected FAUPA in area V1 showed a significant difference between the HS and AD groups (Student’s t-test, p = 0.032 for unfamiliar faces, 0.010 for familiar faces, and 0.005 self-photos, respectively) (Fig. 7a). This result demonstrates that the task-induced signal changes in area V1 of the AD patients did not behave like that of the HS participants, showing that the assumption of the GLM is invalid in area V1 for the AD patients. To further test the validity of the GLM for the AD patients, we computed the MN and SD of the R-values in the 25 subject HS group for each of the three face categories, and used them to establish a threshold TH = MN–1.645×SD (one-tail probability < 5%) for the category. Then, we used the TH to divide the AD group into two subgroups: one group with R > TH (>95% probability of the value occurring by chance) and the other with R < TH (<5% probability of the value occurring by chance) (Fig. 7b). Each participant has three R-values for the three face categories. If a participant had two or three R-values less than the corresponding TH values, the participant was sorted to the 5% chance group; otherwise the participant was sorted to the 95% chance group. If the assumption of the GLM were valid for the AD patients, i.e., the photo-induced BOLD response in area V1 is similar to the ideal response, then we would expect to have about 95% of the AD patients with R > TH and about 5% of them with R < TH. However, we found that 57.9% of the AD patients were sorted to the group with R > TH (8 MAD + 3 SAD) and 42.1% to the group with R < TH (3 MAD + 5 SAD), providing further evidence of the invalid assumption of the GLM for the AD patients. Note that, with the same criterion, one HS was removed from the HS group, consistent with the 95% chance for the twenty-five HS participants to remain in the group. For each face category, the R between the two AD subgroups was significantly different (Bonferroni t-test, max p = 0.0008), but no statistical difference existed between the HS and AD groups with R > TH (Fig. 7b). (Note that the AD group is independent of the HS group, so in principle the R between the two AD subgroups is not assured to be significantly different, depending on the R distribution of these two subgroups.) For the UF category, Fig. 7c illustrates the group-mean signal time course of the selected FAUPA in area V1 for these three subgroups. For the HS and AD with R > TH, the group-mean signal time courses were similar to each other and matched well with the ideal response. However, for the AD with R < TH, the signal time course deviated substantially from the ideal response. Similar behaviors were observed for both FF and SP categories (Supplementary Figure 3). These results show that the assumption of the GLM in area V1 is valid for both subgroups of HS and AD with R > TH, but invalid for the subgroup of AD with R < TH.

Fig. 7

Group comparisons of the R between the ideal response and the task-induced BOLD responses in area V1 for the HS and AD patients. (a) For each face category, the comparison of the R showed a significant difference between the HS (N = 25) and the AD (N = 19) (Student’s t-test, max p = 0.032); (b) For each face category, the group-mean R showed no statistical difference between the HS subgroup (R > TH, N = 24) and the AD subgroup (R > TH, N = 11), but a significant difference between the AD subgroup (R > TH, N = 11) and the AD subgroup (R < TH, N = 8) (Bonferroni t-test, max p = 0.0008); and (c) Illustration of the group-mean signal time course of the selected FAUPA in the V1 for the three subgroups. UF, unfamiliar face; FF, familiar face; SP, self-photo. The error bars indicate the corresponding standard deviations.

We further compared the visual-processing network for the three subgroups: 1) HS with R > TH, 2) AD with R > TH, and 3) AD with R < TH. Fig. 8a illustrates the three visual-processing networks for these three subgroups for the UF category, and similar visual-processing networks were observed for the other two face categories (Supplementary Figure 4). The visual-processing network of the HS with R > TH (N = 24) is almost the same as that of the entire group (N = 25) (Fig. 5). The network of the AD with R > TH is similar to that of the MAD subgroup alone. The network of the AD with R < TH is similar to that of the SAD subgroup alone. For the AD mask, i.e., the visual-processing network of the AD with R > TH, for each face category the R of the AD with R > TH showed no statistical difference compared to that of the HS with R > TH, but the R of the AD with R < TH was significantly reduced compared to that of the AD with R > TH (Bonferroni t-test, max p = 0.048) (Fig. 8b). For the HS mask, i.e., the visual-processing network of the HS with R > TH excluding the AD mask, for each face category the R of the AD with R > TH was significantly reduced compared to that of the HS with R > TH (Bonferroni t-test, max p = 0.002), providing further evidence of a disrupted and reduced visual-processing network from HS to AD.

Fig. 8

Comparisons of the group-mean visual-processing network for the three subgroups of HS with R > TH, AD with R > TH and AD with R < TH. (a) The three images in the left three columns illustrate the visual-processing network for the subgroup of HS with R > TH (top panel), AD with R > TH (middle panel), and AD with R < TH (bottom panel), respectively. The red cluster in the right image in the middle panel represents the mask of the visual-processing network for the subgroup of AD with R > TH. The red cluster in the right image in the top panel represents the mask of the visual-processing network for the subgroup of HS with R > TH, excluding the mask for the AD with R > TH; and (b) For the AD mask, the R showed no statistical difference between the subgroups of AD with R > TH and HS with R > TH, but significantly reduced for the AD with R < TH for each face category (Bonferroni t-test, max p = 0.048). For the HS mask, the R showed a significant reduction for the AD with R > TH compared to that for the HS with R > TH (Bonferroni t-test, max p = 0.0020), consistent with the disrupted visual-processing network illustrated in the middle panel in (a). UF, unfamiliar face; FF, familiar face; SP, self-photo. The error bars indicate the corresponding standard deviations.

Association of the disrupted visual-processing network with MMSE

For each face category, we generated a mask composed of the HS visual-processing excluding that of the SAD (i.e., the largest disrupted part of this network), and computed a mask-mean R for each subject. For each subject, we then computed a mean R averaged over the three face categories to quantify this network disruption. This R was significantly reduced from HS to MAD, and further significantly reduced from MAD to SAD (Fig. 9, left), consistent with their MMSE pattern (Fig. 9, middle). This R was also significantly correlated with the MMSE (Fig. 9, right), suggesting the potential of using the network disruption as a predictor or biomarker of AD progression.

Fig. 9

Association of the disrupted network with the disease severity. Left: group comparison of R of the disrupted network for the three subject groups. R was significantly reduced from HS to MAD (one-tailed t-test) and further significantly reduced from MAD to SAD (one-tailed t-test); Middle: group comparison of MMSE for the three subject groups (Table 1); and Right: scatter plot of R versus MMSE. MMSE, Mini-Mental State Examination; and r, the correlation of R with MMSE over all subjects (N = 44).

DISCUSSION

FAUPAs were detected for each face category and each individual participant in older adults, replicating the previous findings in young and middle-aged adults [20]. They were determined objectively and automatically. The FAUPA determination assumes that the temporal variation of the underlying pooled activity is the same across the entire area. So, the determination requires no a priori knowledge of the activity-induced ideal response signal time course. This assumption implies a perfect temporal correlation everywhere, i.e., R = 1, within the FAUPA. The observed group-mean R of 0.96 is almost identical for all three face categories and participant groups (Fig. 2b), consistent with the definition of FAUPAs and with previous findings [20]. The group-mean total number of voxels per FAUPA is about the same for all three participant groups, across all three face categories (Fig. 2c), again consistent with previous findings [20]. The total number of FAUPAs, however, varies substantially from participant to participant, as reflected in the large standard deviation (Fig. 2a), also consistent with previous findings [20]. Our previous study with young and middle-aged adults showed that the total number of FAUPAs for the resting state was significantly larger than that for the task state, indicating the dependence of these FAUPAs on the brain functional state [20]. Even if all subjects were instructed to perform the same task, their brain activity may vary from subject to subject. A large variation in number of FAUPAs across different subjects was observed in that study, which may reflect individually different brain activity. In this study, the observed large variability in the number of FAUPAs may also reflect the individuality of brain activity of each individual even if the same task was performed. How this individuality is related with the disease remains to be investigated.

The histogram of the FAUPAs as a function of the total number of voxels contained in a FAUPA shows a similar distribution for all three face categories and participant groups (Fig. 3a, Supplementary Figures 1a and 2a), and is very similar to that reported previously [20]. The histogram of the p values that test the separation of the FAUPAs from their adjacent areas also shows a similar distribution for all three face categories and participant groups (Fig. 3b, Supplementary Figures 1b and 2b), and more than 98.5% of them are separated significantly from their adjacent areas. This histogram is also very similar to that reported previously [20]. Overall, the findings of this independent study with both healthy elderly controls and AD patients are very similar to those reported previously, providing further evidence for the existence of FAUPAs and also demonstrating the independence of FAUPAs with age [20].

A task-associated FAUPA is defined as one that is activated when the task is performed [20]. In this study, the task was to view the visually presented face photos. Accordingly, we identified a task-associated FAUPA in area V1 for each face category and each participant. Figure 4a–c demonstrates that for each face category, the signal time courses of the five overlapped voxels are almost identical to each other, showing that the pooled activity within the FAUPA is a dynamically unitary activity across the entire area. This suggests that the FAUPA plays the role of a functional unit for a particular neural computation of the task. These signal time courses also show that the task-induced BOLD signal varies from trial to trial and from participant to participant, and these variations may provide a measure to characterize an individual’s response for each task trial. Using the signal time course of a task-associated FAUPA as the reference function, a correlation analysis with all voxels across the brain may yield a functional network specific for the task, i.e., a task-processing network. Using the signal time course of a task-associated FAUPA to identify the task-processing network may have advantages over using an expected ideal response in the GLM [18, 19] because it takes into account response variations from trial to trial and from participant to participant, potentially yielding a more objectively identified task-processing network for each individual. This may be particularly important for patients with difficulty in performing a task properly. The GLM assumes that task-induced BOLD signals behave similarly to an expected ideal response, and any deviations from the ideal response can be attributed to noise. The visual stimulation-induced BOLD signals in area V1 substantially deviated from the ideal response in 42.1% of the AD patients (Fig. 7c and Supplementary Figure 3), and it would be incorrect to simply attribute these deviations to noise. These substantially deviated BOLD responses in area V1 were not due to closing eyes as we did not notice any prolonged eye closure during any fMRI scans. When the task-induced BOLD signals deviate from the expected ideal response dramatically, it invalidates the assumption and renders the activation map identified with the GLM in error.

The face-evoked visual-processing network in the ventral pathway started from area V1 and ended within the fusiform gyrus for the HS (Fig. 5). AD disrupted and reduced this visual-processing network in a severity dependent manner; the more severe the cognitive impairment, the greater the disruption and reduction of the network, independent of the face category (Figs. 5 and 6). For the MAD patients, the network was disrupted and reduced mainly in the higher-order visual association areas. For the SAD patients, the network was nearly absent in the higher-order visual association areas, and was disrupted and reduced in the lower-order visual association areas. This finding is consistent with the current canonical view of the temporally successive sequence of AD pathological deposition from the higher-order visual association areas to the lower-order visual association areas and then to the primary visual cortex [9 , 35–39]. The significant correlation of the disrupted network R with MMSE provides further evidence to show an association of network disruption with AD severity (Fig. 9). These findings suggest the potential of using functional network changes in the visual system from normal aging to AD progression as a predictor or biomarker of AD progression.

It is well documented that the fusiform gyrus is an essential extra-striate location involved in face perception and recognition [40]. The face-evoked visual-processing network from V1 to fusiform gyrus should be responsible for these early face feature-related processes, but most likely not those face perception-related processes that are critical for face recognition. Nevertheless, these early face feature-related processes are important for face perception and deficiencies in this processing likely affect face perception directly. The AD-associated progressive disruption and reduction in this visual-processing network is consistent with the progressive disintegration of overall cognitive processing in later-stage AD [41, 42]. These results demonstrate the potential of using task-processing network disruption analyses to characterize progressive brain activation reductions in AD. A task-associated FAUPA is identified for each individual participant, which takes into account any potential variation in individual task response (Fig. 4a–c), and then its signal time course is used to determine the task-processing network; thus individually determined task-processing networks may provide a more reliable measure of objectively assessing disease-specific and clinically relevant network variations.

An AD-disrupted visual-processing network is further confirmed by comparing the network between the two subgroups of HS and AD with R > TH (Fig. 8 and Supplementary Figure 4). As an R map is determined by the reference function, different reference functions yield different R maps. For each face category, the R between the ideal response and the signal time course of the selected task-associated FAUPA in area V1 showed a significant difference between the HS and AD groups (Fig. 7a). This implies a significant difference in this signal time course between the two subject groups. Accordingly, we should expect different visual-processing networks identified with these different signal time courses, yielding a potential confounding effect on the observed AD-disrupted visual-processing network. Nevertheless, the visual stimulation-induced BOLD signals in area V1 are very similar to the ideal response in the two subgroups of HS and AD with R > TH, and the signal time courses are also very similar to each other between these two subgroups (Fig. 7c and Supplementary Fig. 3). If the visual-processing network were not affected by AD, it would remain intact regardless of whether it was identified with the signal time course of V1 task-associated FAUPA or the ideal response time course because all those time courses were very similar to each other for these two subgroups. The fact that the visual-processing network was substantially reduced in the subgroup of AD with R > TH, compared to that of HS with R > TH, rules out the potential confounding effect of the different signal time courses on the identified visual-processing networks, confirming an AD-disrupted visual-processing network.

As previously mentioned, all findings reported here were independent of the face category. A limitation of this study relates to the task paradigm. For each participant, a total of only five self-photos, five familiar faces and five unfamiliar faces were used. For each face category, all five photos were presented during one trial period (25 s task on), and each trial occurred nine times in total. Accordingly, unfamiliar faces would have become somewhat more familiar to participants across the study period, particularly the HS subjects. This could have potentially contributed to a lack of differences in the network activation between categories. However, this lack of differentiation is not unexpected. Prior work indicates that stimuli related to categories of ‘self,’ ‘familiar,’ and ‘other’ are differentiated outside the fusiform area, primarily among cortical midline (medial prefrontal, anterior, and posterior cingulate) and paralimbic (insula, frontal operculum) structures [43]. Examination of activity in each of these regions is beyond the scope of this study.

Another limitation of this study relates to the lack of a more effectively controlled regulation of attention during the task. Individuals who appeared to look away from the stimuli or drowse during scanning were asked to open their eyes and focus via a speaker in the control room, but no overt response was asked of subjects during the scanning phase, thus “open-eyed inattention” during the task is possible, and may even be severity-specific, as later-stage AD patients are more attentially impaired than those early in their course. Thus, we cannot exclude differential attentional effects as part of the explanation for our findings, though in real life, such differences may also underlie facial recognition deficits in this population. A small number of SAD patients and heterogeneous sample size are also a limitation of this study.

It is worth mentioning the lack of eye tracking in this study as another limitation. Had we done so, the recorded eye movements during the face presentation period could provide the information that might be utilized to interpret BOLD activity variations from trial to trial (Fig. 4a–c). Our recent study reports a novel method to identify task-specific networks [44]. That study demonstrates a one-to-one relation between the network activity and the task performance from trial to trial, offering a means of testing the causal relationship between neural network activity and human task performance by systematically manipulating task performance and measuring corresponding network activity change. For the task-associated FAUPAs in V1, trial-by-trial activity changes are conspicuous for each run (Fig. 4a–c). The information of eye movements during each task may enable us to relate the movements with the task-evoked activation from trial to trial, but such a study is beyond the scope of this study.

Data availability

Both the original and processed fMRI images plus final research data related to this publication will be available to share upon request with a legitimate reason such as to validate the reported findings or to conduct a new analysis.

Footnotes

ACKNOWLEDGMENTS

The authors thank scanning support from MR technologist Scarlett Doyle. This work was supported by the Michigan State University (MSU) Radiology Pilot Scan Program. We also thank funding sources, including the MSU Deptartment of Family Medicine Pearl Aldrich Graduate Student Fellowship (Grant: RT083166-F5015) and the Blue Cross Blue Shield of Michigan Foundation (Grant: 1981.SAP).

Authors’ disclosures available online ().

The supplementary material is available in the electronic version of this article: .

References

Bloom

(2014) Amyloid-beta and tau: The trigger and bullet in Alzheimer disease pathogenesis. JAMA Neurol 71, 505–508.

Cerquera-Jaramillo

, Nava-Mesa

, Gonzalez-Reyes

, Tellez-Conti

, de-la-Torre (2018) Visual features in Alzheimer’s disease: From basic mechanisms to clinical overview. Neural Plast 2018, 2941783.

Gonzalez-Reyes

, Nava-Mesa

, Vargas-Sanchez

, Ariza-Salamanca

, Mora-Munoz

(2017) Involvement of astrocytes in Alzheimer’s disease from a neuroinflammatory and oxidative stress perspective. Front Mol Neurosci 10, 427.

Hou

, Liu

, Wang

, Ma

, He

, Zhang

, Yu

, Guan

, Ma

(2011) The effects of amyloid-beta42 oligomer on the proliferation and activation of astrocytes in vitro. In Vitro Cell Dev Biol Anim 47, 573–580.

Lindberg

, Selenica

, Westlind-Danielsson

, Schultzberg

(2005) Beta-amyloid protein structure determines the nature of cytokine release from rat microglia. J Mol Neurosci 27, 1–12.

Jagust

(2016) Is amyloid-beta harmful to the brain? Insights from human imaging studies. Brain 139, 23–30.

Malpas

, Saling

, Velakoulis

, Desmond

, O’Brien

; Alzheimer’s Disease Neuroimaging Initiative (2015) Tau and amyloid-beta cerebrospinal fluid biomarkers have differential relationships with cognition in mild cognitive impairment. J Alzheimers Dis 47, 965–975.

Leuba

, Saini

(1995) Pathology of subcortical visual centres in relation to cortical degeneration in Alzheimer’s disease. Neuropathol Appl Neurobiol 21, 410–422.

Lewis

, Campbell

, Terry

, Morrison

(1987) Laminar and regional distributions of neurofibrillary tangles and neuritic plaques in Alzheimer’s disease: A quantitative study of visual and auditory cortices. J Neurosci 7, 1799–1808.

10.

Adlington

, Laws

, Gale

(2009) Visual processing in Alzheimer’s disease: Surface detail and colour fail to aid object identification. Neuropsychologia 47, 2574–2583.

11.

Mandal

, Joshi

, Saharan

(2012) Visuospatial perception: An emerging biomarker for Alzheimer’s disease. J Alzheimers Dis 31(Suppl 3), S117–S135.

12.

Kavcic

, Vaughn

, Duffy

(2011) Distinct visual motion processing impairments in aging and Alzheimer’s disease. Vision Res 51, 386–395.

13.

Deng

, Shi

, Lei

, Liang

, Li

, Chu

, Wang

; Alzheimer’s Disease Neuroimaging Initiative (2016) Mapping the “what” and “where” visual cortices and their atrophy in Alzheimer’s disease: Combined activation likelihood estimation with voxel-based morphometry. Front Hum Neurosci 10, 333.

14.

Delbeuck

, Van der Linden

, Collette

(2003) Alzheimer’s disease as a disconnection syndrome? Neuropsychol Rev 13, 79–92.

15.

Ogawa

, Lee

, Kay

, Tank

(1990) Brain magnetic resonance imaging with contrast dependent on blood oxygenation. Proc Natl Acad Sci U S A 87, 9868–9872.

16.

Kwong

, Belliveau

, Chesler

, Goldberg

, Weisskoff

, Poncelet

, Kennedy

, Cohen

, Turner

(1992) Dynamic magnetic resonance imaging of human brain activity during primary sensory stimulation. Proc Natl Acad Sci U S A 89, 5675–5679.

17.

Logothetis

(2008) What we can do and what we cannot do with fMRI. Nature 453, 869–878.

18.

Friston

, Holmes

, Worsley

, Poline

, Frith

, Frackowiak

(1995) Statistical parametric maps in functional imaging: A general linear approach. Hum Brain Mapp 2, 189–210.

19.

Worsley

(2001) Statistical analysis of activation images. In Functional MRI: An Introduction to Methods, Jezzard

, Matthews

, Smith

, eds. Oxford University Press, New York, pp. 251–270.

20.

Huang

(2018) Human brain functional areas of unitary pooled activity discovered with fMRI. Sci Rep 8, 2388.

21.

Huang

(2018) Method and system for determining brain-state dependent functional areas of unitary pooled activity and associated dynamic networks with functional magnetic resonance imaging. United States Patent and Trademark Office, PCT Application (PCT/US2018/019819), filing data: February 27, 2018.

22.

Beach

, Huck

, Miranda

, Bozoki

(2015) Autonomic, behavioral, and subjective pain responses in Alzheimer’s disease. Pain Med 16, 1930–1942.

23.

American Psychiatric Association (2013) Diagnostic and Statistical Manual of Mental Disorders, 5th Edn. American Psychiatric Association, Washington, DC.

24.

McKhann

, Drachman

, Folstein

, Katzman

, Price

, Stadlan

(1984) Clinical diagnosis of Alzheimer’s disease: Report of the NINCDS-ADRDA Work Group under the auspices of Department of Health and Human Services Task Force on Alzheimer’s disease. Neurology 34, 939–944.

25.

Folstein

, Folstein

, McHugh

(1975) “Mini-mental state”: A practical method for grading the cognitive state of patients for the clinician. J Psychiatr Res 12, 189–198.

26.

Alexopoulos

, Abrams

, Young

, Shamoian

(1988) Cornell scale for depression in dementia. Biol Psychiatry 23, 271–284.

27.

Beach

, Huck

, Miranda

, Foley

, Bozoki

(2016) Effects of Alzheimer Disease on the facial expression of pain. Clin J Pain 32, 478–487.

28.

Minear

, Park

(2004) A lifespan database of adult facial stimuli. Behav Res Methods Instrum Comput 36, 630–633.

29.

Cox

(1996) AFNI: Software for analysis and visualization of functional magnetic resonance neuroimages. Comput Biomed Res 29, 162–173.

30.

Dukart

, Bertolino

(2014) When structure affects function–the need for partial volume effect correction in functional and resting state magnetic resonance imaging studies. PLoS One 9, e114227.

31.

Glantz

(1997) Primer of Bio-statistics, Fourth ed. McGraw-Hill Health Professions Division, New York.

32.

Boynton

, Engel

, Glover

, Heeger

(1996) Linear systems analysis of functional magnetic resonance imaging in human V1. J Neurosci 16, 4207–4221.

33.

Friston

, Fletcher

, Josephs

, Holmes

, Rugg

, Turner

(1998) Event-related fMRI: Characterizing differential responses. Neuroimage 7, 30–40.

34.

Glover

(1999) Deconvolution of impulse response in event-related BOLD fMRI. Neuroimage 9, 416–429.

35.

Brun

, Gustafson

(1976) Distribution of cerebral degeneration in Alzheimer’s disease. A clinico-pathological study. Arch Psychiatr Nervenkr (1970) 223, 15–33.

36.

Arnold

, Hyman

, Flory

, Damasio

, Van Hoesen

(1991) The topographical and neuroanatomical distribution of neurofibrillary tangles and neuritic plaques in the cerebral cortex of patients with Alzheimer’s disease. Cereb Cortex 1, 103–116.

37.

Braak

, Braak

(1995) Staging of Alzheimer’s disease-related neurofibrillary changes. Neurobiol Aging 16, 271–278.

38.

Albers

, Gilmore

, Kaye

, Murphy

, Wingfield

, Bennett

, Boxer

, Buchman

, Cruickshanks

, Devanand

, Duffy

, Gall

, Gates

, Granholm

, Hensch

, Holtzer

, Hyman

, Lin

, McKee

, Morris

, Petersen

, Silbert

, Struble

, Trojanowski

, Verghese

, Wilson

, Xu

, Zhang

(2015) At the interface of sensory and motor dysfunctions and Alzheimer’s disease. Alzheimers Dement 11, 70–98.

39.

Cho

, Choi

, Hwang

, Kim

, Lee

, Ryu

, Lee

, Lyoo

(2016) In vivo cortical spreading pattern of tau and amyloid in the Alzheimer disease spectrum. Ann Neurol 80, 247–258.

40.

Kanwisher

, McDermott

, Chun

(1997) The fusiform face area: A module in human extrastriate cortex specialized for face perception. J Neurosci 17, 4302–4311.

41.

Mograbi

, Brown

, Morris

(2009) Anosognosia in Alzheimer’s disease–the petrified self. Conscious Cogn 18, 989–1003.

42.

Weiler

, Northoff

, Damasceno

, Balthazar

MLF

(2016) Self, cortical midline structures and the resting state: Implications for Alzheimer’s disease. Neurosci Biobehav Rev 68, 245–255.

43.

Northoff

, Heinzel

, de Greck

, Bermpohl

, Dobrowolny

, Panksepp

(2006) Self-referential processing in our brain–a meta-analysis of imaging studies on the self. Neuroimage 31, 440–457.

44.

Huang

(2020) Dynamic activity of human brain task-specific networks. Sci Rep 10, 7851.