Medial Temporal Lobe Networks in Alzheimer's Disease: Structural and Molecular Vulnerabilities

Abstract

The medial temporal lobe (MTL) is connected to the rest of the brain through two main networks: the anterior-temporal (AT) and the posterior-medial (PM) systems. Given the crucial role of the MTL and networks in the physiopathology of Alzheimer's disease (AD), the present study aimed at (1) investigating whether MTL atrophy propagates specifically within the AT and PM networks, and (2) evaluating the vulnerability of these networks to AD proteinopathies. To do that, we used neuroimaging data acquired in human male and female in three distinct cohorts: (1) resting-state functional MRI (rs-fMRI) from the aging brain cohort (ABC) to define the AT and PM networks (n = 68); (2) longitudinal structural MRI from Alzheimer's disease neuroimaging initiative (ADNI)GO/2 to highlight structural covariance patterns (n = 349); and (3) positron emission tomography (PET) data from ADNI3 to evaluate the networks' vulnerability to amyloid and tau (n = 186). Our results suggest that the atrophy of distinct MTL subregions propagates within the AT and PM networks in a dissociable manner. Brodmann area (BA)35 structurally covaried within the AT network while the parahippocampal cortex (PHC) covaried within the PM network. In addition, these networks are differentially associated with relative tau and amyloid burden, with higher tau levels in AT than in PM and higher amyloid levels in PM than in AT. Our results also suggest differences in the relative burden of tau species. The current results provide further support for the notion that two distinct MTL networks display differential alterations in the context of AD. These findings have important implications for disease spread and the cognitive manifestations of AD.

SIGNIFICANCE STATEMENT The current study provides further support for the notion that two distinct medial temporal lobe (MTL) networks, i.e., anterior-temporal (AT) and the posterior-medial (PM), display differential alterations in the context of Alzheimer's disease (AD). Importantly, neurodegeneration appears to occur within these networks in a dissociable manner marked by their covariance patterns. In addition, the AT and PM networks are also differentially associated with relative tau and amyloid burden, and perhaps differences in the relative burden of tau species [e.g., neurofibriliary tangles (NFTs) vs tau in neuritic plaques]. These findings, in the context of a growing literature consistent with the present results, have important implications for disease spread and the cognitive manifestations of AD in light of the differential cognitive processes ascribed to them.

Introduction

The medial temporal lobe (MTL) is the earliest region to display atrophy over the course of Alzheimer's disease (AD), consistent with early neurofibriliary tangle (NFT) involvement. This is followed by spread of NFTs and concomitant atrophy, in a stereotypical manner, reaching limbic regions and ultimately the neocortex (Braak and Braak, 1991; Pini et al., 2016). This topographical overlap suggests a direct relationship between tau pathology and atrophy, which has been confirmed in studies associating structural MRI with pathology (Whitwell et al., 2008; de Flores et al., 2020), CSF (de Souza et al., 2012; Tardif et al., 2018), or positron emission tomography (PET; Das et al., 2018; Xie et al., 2018; La Joie et al., 2020) measures of tau pathology. Importantly, tau pathology and subsequent atrophy spread across anatomically interconnected regions, which has led to the theory of a “prion-like” propagation, where pathology spreads within specific networks (Peng et al., 2020). Recent imaging investigations support this idea, showing a close relationship between functional connectivity and tau/neurodegeneration topography and propagation (Seeley et al., 2009; Ossenkoppele et al., 2019; Franzmeier et al., 2020a,b).

The MTL is composed of several subregions including the hippocampus as well as entorhinal cortex (ERC), perirhinal cortex [PRC; including Brodmann areas (BA)35 and BA36], and parahippocampal cortex (PHC). These different regions do not function in isolation but are connected to the rest of the brain through two specific networks that converge in the hippocampus. The anterior-temporal (AT) network, composed of the PRC, anterior hippocampus (aH), lateral ERC, amygdala, and temporopolar and lateral orbitofrontal cortices, is thought to support object and semantic memory, object perception and affective/emotion processing. The posterior-medial (PM) network, composed of PHC, posterior hippocampus (pH), medial ERC, retrosplenial cortex, precuneus, angular gyrus, and medial prefrontal cortex, is preferentially involved in episodic memory, scene perception and social cognition (Ranganath and Ritchey, 2012). Importantly, available data from the literature have demonstrated a functional vulnerability of these specific systems in the context of normal aging and early AD. However, findings are partly inconsistent, showing reduced connectivity within the PM network and either reduced or increased AT network connectivity (Das et al., 2015; Liu et al., 2016; Salami et al., 2016; Berron et al., 2020). In addition, recent publications suggest a specific sensitivity of the AT and PM systems to AD pathology, where tau preferentially deposits in the AT network while amyloid pathology was preferentially found in PM regions (Adams et al., 2019; Maass et al., 2019).

Given the crucial role of the MTL and brain networks in the physiopathology of AD, the present study aimed at investigating the structural and molecular vulnerabilities of the AT and PM systems over the course of the disease. To do that, we used a combination of resting-state functional MRI (rs-fMRI), longitudinal structural MRI as well as tau and amyloid PET data acquired in three distinct cohorts covering the full AD spectrum. In particular, we wanted to determine whether longitudinal atrophy occurred conjointly within or across these networks. If the former, this would support the notion of neurodegeneration occurring along networks, potentially because of transsynaptic spread of pathology. To address this, we evaluated whether longitudinal atrophy in BA35 and PHC, two MTL nodes of the AT versus PM systems, respectively, structurally covaried with distinct brain regions reflecting the AT versus PM networks, which were defined with resting state functional connectivity data from an independent older adult cohort. In addition, we aimed to further our understanding regarding the relative vulnerability of these two systems to amyloid and tau pathologies. Based on prior observations, we hypothesized that tau-PET uptake was relatively higher in the AT network while amyloid uptake was higher in the PM network. To go further, we investigated whether the relationship between both pathologies is different within the AT and PM networks.

Materials and Methods

Participants

Alzheimer's disease neuroimaging initiative (ADNI)

Part of the data used in the preparation of this article was obtained from the ADNI database (https://adni.loni.usc.edu/). The ADNI was launched in 2003 as a public–private partnership, led by Principal Investigator Michael W. Weiner. The primary goal of ADNI has been to test whether serial MRI, PET, other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of MCI and early AD. For up-to-date information, see www.adni-info.org.

Participants in ADNI GO and ADNI 2 studies who had longitudinal T1-weighted (T1w) MRI scans and β-amyloid PET scans available were considered for this study. All available scans within 1.2- to 2.5-year follow-up were analyzed for each participant. A summary standardized uptake value ratio (SUVR) derived from Florbetapir PET (available publicly in the processed data on the ADNI website) was used to determine the Aβ status of each participant by applying a threshold of SUVR ≥ 1.11 (Landau et al., 2013). In total, 349 participants were selected and grouped into Aβ negative (Aβ–) cognitively normal (CN) controls, Aβ positive (Aβ+) CN, Aβ+ mild cognitive impairment (MCI), and Aβ+ AD patients.

Participants in ADNI 3 with available β-amyloid and tau PET scans were also included in this study. As described previously, a threshold of SUVR ≥ 1.11 was used to determine the Aβ status. In total, 186 participants were selected and grouped into Aβ + CN, Aβ+ MCI, and Aβ+ AD patients.

Aging brain cohort (ABC)

CN older adults enrolled in the ABC who had T1w and resting-state fMRI scans available were considered for this study. These participants were recruited from the Clinical Core of the Alzheimer's Disease Core Center (ADCC) at the University of Pennsylvania. These individuals are evaluated on an annual basis, including standard psychometric testing as prescribed by the Uniform Data Set 3.0 (UDS 3.0; Weintraub et al., 2018) and given an annual consensus diagnostic designation. This study was approved by the Institutional Review Board of the University of Pennsylvania and informed consent was provided by all subjects. In total, 68 CN participants were selected.

The demographic characteristics of the samples are summarized in Table 1. See Figure 1 for flow-charts showing the process of arriving at these cohorts.

Table 1.

Demographics of the samples

Structural covariance analyses: ADNI GO and ADNI 2
	CN Aβ– (n = 117)	CN Aβ+ (n = 65)	MCI Aβ+ (n = 148)	AD Aβ+ (n = 19)
Age (SD), year	71.9 (6.1)	75.3 (5.8)	72.9 (6.6)	75.6 (8.7)
Sex (F, M)	57,60	43,22	65,83	11,8
Education (SD), year	17.1 (2.3)	16.0 (2.8)	16.3 (2.7)	15.9 (2.7)
MMSE	29.1 (1.3)	29.1 (0.9)	27.7 (1.8)	22.7 (9.7)
Number of scans (SD)	2.4 (0.5)	2.4 (0.5)	2.5 (0.5)	2.3 (0.5)
Follow-up time (SD), d	655.6 (174.5)	669.8 (164.3)	602.0 (186.3)	540.2 (167.6)

PET analyses: ADNI 3
	CN Aβ+ (n = 67)	MCI Aβ+ (n = 71)	AD Aβ+ (n = 48)
Age (SD), year	76.9 (5.8)	76.2 (6.8)	78.0 (9.3)
Sex (F, M)	23,44	40,31	25,23
Education (SD), year	16.5 (2.3)	15.8 (2.6)	15.3 (2.4)
MMSE (SD)	28.7 (1.6)	26.9 (2.2)	22.1 (4.4)

Functional connectivity analyses: ABC
	CN (n = 68)
Age (SD), year	71.6 (6.5)
Sex (F, M)	44,24
Education (SD), year	15.8 (2.6)
MMSE (SD)	29.2 (0.9)

CN, cognitively normal; MCI, mild cognitive impairment; Aβ, amyloid-β; MMSE, Mini-Mental State Examination.

Figure 1.

Flow-charts showing cohorts selection.

Neuroimaging data acquisition

ADNI imaging protocol

The MRI imaging protocols of the ADNI study that were used to acquire the T1w MRI scans were previously described in Jack et al. (2008) and Leow et al. (2006). Amyloid PET imaging consisted of a continuous 200min brain scan (four frames of 5-min duration) 50 min following ∼10 mCi of ¹⁸F-florbetapir injection. Tau PET imaging consisted of a continuous 30-min brain scan (six frames of 5-min duration) 75 min following ∼10 mCi of ¹⁸F-flortaucipir injection. Amyloid and tau PET images were downloaded from the ADNI data archive in the most fully postprocessed format with the image description “Coreg, Avg, Std Img and Vox Siz, Uniform Resolution.” Up-to-date information about ADNI imaging protocols can be found at adni.loni.usc.edu/methods/mri-tool/mri-analysis and http://adni.loni.usc.edu/methods/pet-analysis-method/pet-analysis.

ABC imaging protocol

Each subject of the ABC underwent a MR scan at the University of Pennsylvania using a 3T Siemens MAGNETOM Prisma MRI scanner. First, T1w structural images were acquired [repetition time (TR) = 2400 ms; echo time (TE) = 2.24 ms; inversion time (TI) = 1060 ms; flip angle = 8°; 208 axial slices; slice thickness = 0.8 mm; field of view = 240 × 256 mm²; matrix = 300 × 320; in-plane resolution = 0.8 × 0.8 mm²; acquisition time = 6.38 min]. Second, resting-state functional acquisitions were obtained using a gradient-echo echoplanar (EPI) sequence (TR = 720 ms; TE = 37 ms; flip angle = 52°; 72 axial slices; slice thickness = 2 mm; field of view = 208 × 208 mm²; matrix = 104 × 104; in-plane resolution = 2 × 2 mm²; multiband acceleration factor = 8; 420 volumes; acquisition time = 5.12 min). Finally, B0 maps were acquired (TR = 580 ms; TE1 = 4.12 ms; TE2 = 6.58 ms; flip angle = 45°; 60 axial slices; slice thickness = 3 mm; field of view = 240 × 240 mm²; matrix = 80 × 80; in-plane resolution = 3 × 3 mm²; acquisition time = 1.34 min).

Image processing

Longitudinal structural measures of MTL subregions (ADNI GO/2)

The longitudinal MRI scans from ADNI were processed using a tailored pipeline that accounts for common confounds of conventional approaches. A multiatlas automatic segmentation algorithm “ASHS-T1” (Xie et al., 2019, 2020; de Flores et al., 2020) was used to label the anterior and pH, ERC, PRC (subdivided in BA35 and BA36), and PHC in each baseline MRI scan. The relative annualized volume atrophy rate (in %) was computed for each region using a deformation-based morphometry pipeline and a comprehensive quality control was performed as described in a previous publication (Xie et al., 2020). More precisely, the quality control included three levels: the MRI scans, the ASHS-T1 automatic segmentation, and the longitudinal estimation. In the first step, all the baseline MRI scans were checked visually and those with severe motion artifact or blurring were excluded. Then, baseline ASHS-T1 segmentations were visually checked for all subjects. All subjects included in the present study had good quality segmentations for BA35 and PHC, as these regions are of specific interest in our analyses. Lastly, the quality of longitudinal estimation was assessed. Scans of follow-up timepoints that were poorly registered to the baseline were excluded from the atrophy rate estimation. However, because of the very large number of registration pairs, it was not feasible to visually check all of them. Instead, we checked a subset of registrations selected using the following method: (1) we computed the normalized cross-correlation (NCC) of the registered scans, which is a good indicator of registration quality, and then selected pairs of registrations that yielded NCC that are 1.5 SD below the mean. (2) We also randomly selected 5% of the remaining pairs to confirm that they had high registration quality. Bilateral measurements of each subregion were averaged.

Whole-brain longitudinal mapping of cortical thickness (ADNI GO/2)

The ANTs longitudinal cortical thickness pipeline was used to generate whole brain atrophy maps (Tustison et al., 2019). First, a group template and corresponding prior probability images were created using an independent group of 52 CN ADNI-1 subjects, as previously described (Tustison et al., 2019). Then, an unbiased single-subject template (SST) was created for each subject of the present study. The ANTs cross-sectional cortical thickness pipeline (Tustison et al., 2014) was applied to the SST (with the group template and priors as input). SST prior probability maps were created and the ANTs cross-sectional cortical thickness pipeline, with the SST as the reference template, was applied to each individual time point image. Least squares regression models were applied to estimate annualized change based on each time point cortical thickness maps for each subject. Each resulting whole brain atrophy maps, warped to the template space, was visually checked and no failure was detected.

rs-fMRI (ABC)

First, data were checked for the lack of artifact or of abnormal variance distribution through the application of the TSDiffana routine (https://imaging.mrc-cbu.cam.ac.uk/imaging/DataDiagnostics). B0 maps were then used to perform susceptibility distortion correction. Next, six-parameter rigid body motion correction was implemented to account for head motion-related artifacts (Friston et al., 1995). All included subjects experienced minimal head motion (<2-mm translation and <2° rotation in any direction) at all times during acquisition. Following motion correction, the fMRI data were co-registered to the T1 structural scan. A 4-mm full-width at half-maximal (FWHM) smoothing was applied to the rs-fMRI scan together with a bandpass filter (0.008–0.08 Hz) to the time series to reduce low-frequency drift and high-frequency noise. Physiologic noise was eliminated via linear regression to factor out the global signal and mean signals from white matter and CSF regions. The Atropos method was used for three-tissue segmentation (Avants et al., 2011). We also regressed out the six parameters of head motion (obtained from motion correction) and their six temporal derivatives to minimize motion-induced signal variation.

Anatomical regions from ASHS-T1 segmentations were used as regions of interest (ROIs). More precisely, the union of the aH and BA35 was used as the seed region for the AT system, and the union of the pH and PHC was used as the seed for the PM system, in accordance to Ranganath and Ritchey (2012). Segmentations were carefully visually checked manually edited when needed. Note that the ERC was not used as a seed since this region is thought be involved in both the AT and PM systems (Maass et al., 2015). Since time series were extracted after a 4 mm smoothing, several slices perpendicular to the long axis of the hippocampus were removed from the aH and pH ROIs to avoid signal contamination from one ROI to the other (see Fig. 2). In order to remove these voxels, an intermediate step was necessary. High-resolution T2 images (0.4 × 0.4 × 1.2 mm³) acquired perpendicular to the long axis of the hippocampus were available for each subject. aH and pH were first transformed to T2 space. Then, the two most posterior slices of the aH ROI and the three most anterior slices of the pH ROI were isolated to generate a 6-mm exclusion mask. This mask was transformed to T1 space and each voxel were removed from the original aH and pH ROIs. Thus, the ROIs were not adjacent and signal from one ROI was unlikely to contaminate the other. Linear Pearson correlations between the average residual time-series signals for each seed were used to generate functional connectivity maps for each subject (two maps per subject). The resulting connectivity maps were Fisher r-z transformed for variance stabilization, warped to template space and a 6-mm FWHM smoothing was applied. Unsmoothed warped connectivity maps were also averaged across all subjects, resulting in an average connectivity map for each of the two seeds.

Figure 2.

Example of ASHS-T1 hippocampal ROIs used as seeds for functional connectivity analyses. A 6-mm gap was generated between the two ROIs to avoid signal contamination from one ROI to the other. Red, aH; blue, pH.

Amyloid and Tau PET (ADNI 3)

The postprocessed ¹⁸F-florbetapir and ¹⁸F-flortaucipir PET images were generated by averaging coregistered individual frames, re-oriented in a standardized image space such that the anterior-posterior axis of the subject is parallel to the AC-PC line, followed by scanner-specific filtering to generate an image with a uniform isotropic resolution of 8 mm full width half maximum. More details about the preprocessing used is described at http://adni.loni.usc.edu/methods/pet-analysis-method/pet-analysis/. Postprocessed PET images were then registered to subject's T1w structural MRI using ANTs. The MRI scan was parcellated into cerebellar, cortical, and subcortical ROIs using a multiatlas segmentation method (Wang et al., 2011). The atlas set and the resulting segments are described in the study by Landman and Warfield (2012). Mean PET tracer uptake in the cerebellar gray matter and gray+white matter was used as reference regions for ¹⁸F-Flortaucipir and ¹⁸F-Florbetapir, respectively, and a SUVR map was generated for each participant for each PET modality. Each SUVR map was then moved to template space using ANTs and masked to only include gray matter voxels. Since amyloid and tau co-localize into neuritic plaques, a subset of amyloid plaques defined by the presence of hyperphosphorylated tau protein (Jellinger, 2020), we generated tau to amyloid ratio (tau/amyloid) maps to better highlight tau pathology that may be relatively more or less independent of amyloid pathology. To do that, the warped SUVR tau map was divided by the warped SUVR amyloid map (at the voxel level) for each subject. Tau, amyloid and tau/amyloid scans were averaged across the 186 subjects for uptake patterns visualization.

Experimental design and statistical analyses

Defining AT and PM networks

Functional connectivity maps generated in the CN group, with combinations of aH/BA35 versus pH/PHC as seeds, were used to define the AT and PM networks, respectively. In order to highlight brain regions that specifically belong to the AT versus PM systems and to avoid networks overlap, the two sets of 68 connectivity maps were compared using a “paired t test” model in the SPM12. A threshold of P_{voxel(uncorr)} < 0.01 with a cluster extent determined by an iterative procedure for estimating cluster-level false-positive rates (Monte Carlo simulation) with 10,000 iterations was used to achieve a cluster corrected statistical significance of P_{cluster(corr)} < 0.05 (AFNI_19.2.26, 3dClustSim program). Note that two separate one sample t tests were also performed to evaluate the extent that the networks overlap.

Structural covariance analyses

Correlations between summary atrophy rates in BA35 and PHC and voxel-wise atrophy maps in the whole brain were performed in SPM12 to evaluate structural covariance patterns. In order to highlight brain regions that specifically structurally covary with BA35 versus PHC, both BA35 and PHC atrophy rates were entered in a single multiple regression model. A threshold of P_{voxel(uncorr)} < 0.01 with a cluster extent determined by an iterative procedure for estimating cluster-level false-positive rates (Monte Carlo simulation) with 10,000 iterations was used to achieve a corrected statistical significance of P_{cluster(corr)} < 0.05 (AFNI_19.2.26, 3dClustSim program).

Relationship between structural covariance and functional connectivity

Three complementary analyses were performed to evaluate the relationship between MTL structural covariance and functional connectivity. First, the spatial overlap between structural covariance patterns (thresholded maps) and the AT and PM networks (thresholded maps) was visually inspected. Then, goodness of fit (GOF) indices between structural covariance maps (nonthresholded T-maps) and AT and PM networks were calculated. GOF is expressed as the difference between the mean T-score of all voxels that fall within the network of interest (T_inside) and the mean T-score of gray matter voxels outside the network of interest (T_outside), thus T_inside – T_outside (Ossenkoppele et al., 2019). Thus, four indices were calculated (two structural covariance maps × two networks). Note that regions corresponding to BA35 and PHC were removed from the network masks to avoid autocorrelation effects. Lastly, correlation analyses were performed between averaged functional connectivity maps (z values) and structural covariance maps (T-values), for each gray matter voxel (n = 375,222). A total of four correlation analyses were performed (two structural covariance maps × two averaged functional connectivity maps).

Vulnerability to AD proteinopathies of the AT and PM networks

In order to assess the relative vulnerability of the AT versus PM systems to pathologies, masks derived from functional connectivity analyses were used to extract mean SUVR values from amyloid and tau scans, as well as from the tau to amyloid ratio maps. Note that hippocampal ROIs were removed from the AT and PM masks to avoid off target binding contamination from the choroid plexus. For each modality (amyloid, tau, tau to amyloid ratio), a paired t test model was employed to compare signal extracted within both networks. In addition, we investigated whether the relationship between amyloid and tau is different within the AT versus the PM networks. This analysis was performed at the voxel level, for each gray matter voxel included in the AT and PM networks. An ANOVA was used with tau SUVR (from the averaged SUVR tau map) as the dependent variable and amyloid SUVR (from the averaged SUVR amyloid map) and network (AT or PM) as the independent variables. The amyloid × network interaction was tested.

Results

Defining the AT and PM networks

MTL networks highlighted in the CN population from the ABC dataset are shown in Figure 3A (continuous maps are illustrated in Fig. 4). These functional connectivity analyses revealed two distinct networks, corresponding to the AT and PM systems. The AT network, defined by regions more functionally connected to aH/BA35 than pH/PHC, encompasses the amygdala and the temporopolar cortex extending posteriorly to inferior and middle temporal gyri. In contrast, the PM network, defined by regions more functionally connected to pH/PHC than aH/BA35, encompasses the posterior cingulate cortex, the precuneus, the cuneus, the angular and occipital gyri and the right insula. Results from one sample t tests are illustrated in Figure 5. The two networks largely overlap, where nonoverlapping regions include the temporal pole for AT and cingulate, occipital and occipitotemporal regions for PM. In addition, analyses were repeated using more stringent thresholds (P_{voxel(uncorr)} < 0.005, P_{cluster(corr)} < 0.05 and P_{voxel(uncorr)} < 0.001, P_{cluster(corr)} < 0.05). As illustrated in Figure 6, results were overall unchanged.

Figure 3.

Functional connectivity and structural covariance of MTL subregions. A, AT and PM functional networks defined using seed-based correlations (seed: aH and BA35 for AT; pH and PHC for PM) in the CN population from the ABC. Connectivity maps were compared using a paired t test (P_{cluster(corr)} < 0.05, P_{voxel(uncorr)} < 0.01) to highlight specific networks. B, Whole-brain longitudinal structural covariance of BA35 and the PHC cortices (P_{cluster(corr)} < 0.05, P_{voxel(uncorr)} < 0.01) generated in the ADNI GO/2 population (spanning from Aβ– CN to Aβ+ AD patients). C, Overlap between networks and BA35 covariance. D, Overlap between networks and PHC covariance.

Figure 4.

Functional connectivity and structural covariance of MTL subregions, continuous T-value maps. A, AT and PM functional networks defined using seed-based correlations (seed: aH and BA35 for AT; pH and PHC for PM) in the CN population from the ABC. Connectivity maps were compared using a paired t test. B, Whole-brain longitudinal structural covariance of BA35 and the PHC cortices generated in the ADNI GO/2 population (spanning from Aβ– CN to Aβ+ AD patients).

Figure 5.

AT and PM functional networks defined using two separate one sample t tests. The networks were defined using seed-based correlations (seed: aH and BA35 for AT; pH and PHC for PM) in the CN population from the ABC. Two separate one sample t tests with a threshold of P_{voxel(uncorr)} < 0.001; P_{cluster(corr)} < 0.05 were used.

Figure 6.

Functional connectivity and structural covariance of MTL subregions using different thresholds. A, AT and PM functional networks defined using seed-based correlations (seed: aH and BA35 for AT; pH and PHC for PM) in the CN population from the ABC. Connectivity maps were compared using a paired t test (P_{cluster(corr)} < 0.05, P_{voxel(uncorr)} < 0.01/P_{cluster(corr)} < 0.05, P_{voxel(uncorr)} < 0.005/P_{cluster(corr)} < 0.05, P_{voxel(uncorr)} < 0.001) to highlight specific networks. B, Whole-brain longitudinal structural covariance of BA35 and the PHC cortices (P_{cluster(corr)} < 0.05, P_{voxel(uncorr)} < 0.01/P_{cluster(corr)} < 0.05, P_{voxel(uncorr)} < 0.005/P_{cluster(corr)} < 0.05, P_{voxel(uncorr)} < 0.001) generated in the ADNI GO/2 population (spanning from Aβ– CN to Aβ+ AD patients).

Structural covariance analyses

Longitudinal structural covariance patterns for BA35 and PHC, highlighted in the ADNI GO/2 population (spanning from Aβ– CN to Aβ+ AD patients), are shown in Figure 3B (continuous maps are illustrated in Fig. 4). Overall, BA35 atrophy covaried with atrophy of the amygdala and the left temporal cortex (just inferior temporal on the right), while PHC atrophy structurally covaried with longitudinal atrophy of the posterior cingulate cortex, the precuneus, the cuneus and the angular and occipital gyri. Analyses were repeated using more stringent thresholds (P_{voxel(uncorr)} < 0.005, P_{cluster(corr)} < 0.05 and P_{voxel(uncorr)} < 0.001, P_{cluster(corr)} < 0.05). As illustrated in Figure 6, structural covariance pattern of PHC was overall unchanged. For the structural covariance of BA35, results partially hold using more stringent thresholds, with patterns including the MTL and the posterior portion of the lateral temporal cortex at P_{voxel(uncorr)} < 0.005, P_{cluster(corr)} < 0.05 while BA35 structurally covaried within the MTL at P_{voxel(uncorr)} < 0.001, P_{cluster(corr)} < 0.05.

Relationship between structural covariance and functional connectivity

The relationship between MTL structural covariance and functional connectivity was assessed using three different and complementary approaches. First, the spatial overlap between structural covariance patterns and the AT and PM networks is illustrated in Figure 3C,D. Overall, BA35 structural covariance overlapped with the AT network while PHC structural covariance overlapped with the PM network with little structural covariance outside of the respective networks. Note that the same conclusion applies when networks derived from one sample t tests are considered. Second, we generated GOF indices between structural covariance maps and AT and PM networks. BA35 covariance showed better GOF with the AT system (1.10 vs −0.50 for PM). In contrast, PHC covariance showed better GOF with the PM system (0.92 vs −0.33 for AT). Third, each voxel of the averaged connectivity maps generated in the CN population from the ABC were correlated with each voxel of structural covariance maps generated in the ADNI GO/2 population. As illustrated in Figure 7, BA35 structural covariance demonstrated better correlation with AT functional connectivity (r = 0.24, p < 0.001 vs r = 0.04, p < 0.001 for PM functional connectivity), while PHC structural covariance displayed better correlation with PM functional connectivity (r = 0.18, p < 0.001 vs r = −0.03, p < 0.001 for AT functional connectivity). To confirm these observations, correlation coefficients were statistically compared using the R package “cocor” (Diedenhofen and Musch, 2015). As reported in Table 2, the results of all tests lead to the convergent conclusion that BA35 structural covariance is significantly more associated with AT functional connectivity than PM functional connectivity, while PHC structural covariance is significantly more associated with PM functional connectivity than AT functional connectivity.

Figure 7.

Correlations between functional connectivity and structural covariance. Each voxel of the averaged connectivity maps (z values) generated in the CN population from the ABC were correlated with each voxel of structural covariance maps (T-values) generated in the ADNI GO/2 population.

Table 2.

Correlation coefficients comparisons

	BA35 structural covariance ∼ AT functional connectivity vs BA35 structural covariance ∼ PM functional connectivity	PHC structural covariance ∼ AT functional connectivity vs PHC structural covariance ∼ PM functional connectivity
Pearson and Filon's z (1898)	z = 144.0206, p-value = 0.0000	z = −150.4716, p-value = 0.0000
Hotelling's t (1940)	t = 146.3813, df = 375219, p-value = 0.0000	t = −152.9268, df = 375219, p-value = 0.0000
Williams' t (1959)	t = 146.3212, df = 375219, p-value = 0.0000	t = −152.9084, df = 375219, p-value = 0.0000
Olkin's z (1967)	z = 144.0206, p-value = 0.0000	z = −150.4716, p-value = 0.0000
Dunn and Clark's z (1969)	z = 143.3883, p-value = 0.0000	z = −149.5178, p-value = 0.0000
Stanley and Hills' (1970)	t = 146.3813, df = 375219, p-value = 0.0000	t = −152.9268, df = 375219, p-value = 0.0000
Steiger's (1980)	z = 142.8124, p-value = 0.0000	z = −148.8888, p-value = 0.0000
Meng, Rosenthal, and Rubin's z (1992)	z = 142.2512, p-value = 0.0000	z = −148.2490, p-value = 0.0000
Hittner, May, and Silver's (2003)	z = 142.7896, p-value = 0.0000	z = −148.8813, p-value = 0.0000
Zou's (2007) confidence interval	0.1914 0.1967	−0.2066 −0.2013

Statistical comparisons were performed using the R package “cocor” (Diedenhofen and Musch, 2015).

Vulnerability to AD proteinopathies of the AT and PM networks

Average uptake patterns for tau, amyloid and tau/amyloid PET are illustrated in Figure 8. Overall, tau PET uptake was particularly high in temporal (medial and lateral), parietal (medial and inferior) and orbitofrontal regions, while amyloid uptake was more diffuse and involved almost all brain regions (except the MTL). Tau/amyloid signal was high in temporal regions, particularly in the MTL, as well as in the lateral parietal cortex. Note that the signal was relatively low in midline parietal structures. When comparing signal within the AT versus PM systems (Fig. 8), tau uptake was significantly higher in the AT network than in the PM network (t₍₁₈₅₎ = 3.52, p = 5.40e⁻⁰⁴), while amyloid uptake was higher in the PM network than in the AT network (t₍₁₈₅₎ = −30.25, p < 2.2e⁻¹⁶). Tau/amyloid signal was significantly higher in the AT network than in the PM network (t₍₁₈₅₎ = 18.08, p < 2.2e⁻¹⁶). These analyses were repeated by including additional 195 Aβ– CN individuals. Results were unchanged, although more significant, with higher tau uptake and tau/amyloid ratio in AT than in PM (t₍₃₈₀₎ = 4.91, p = 1.36e⁻⁰⁶; t₍₃₈₀₎ = 32.37, p < 2.2e⁻¹⁶, respectively) and higher amyloid uptake in PM than in AT (t₍₃₈₀₎ = −45.47, p < 2.2e⁻¹⁶). Lastly, the analysis performed at the voxel level showed a significant interaction of the effects of amyloid and network on tau (F_(1,116663) = 6224.3, p < 2.2e⁻¹⁶). As illustrated in Figure 9, the slope of the association between tau and amyloid SUVRs was steeper within the AT network than in the PM network.

Figure 8.

Tau and amyloid accumulation within the AT and PM networks. A, Tau, amyloid, and tau/amyloid ratio scans were averaged across 186 subjects for uptake patterns visualization. B, Mean SUVR values in the AT and PM networks were estimated from amyloid, tau, and tau/amyloid ratio maps. For each modality, a paired t test model was used to compare signal extracted within both networks.

Figure 9.

Association between tau and amyloid within the AT versus PM networks. Each gray matter voxels from the averaged SUVR tau and amyloid maps and included in the AT and PM masks (n = 116,667) were used to investigated whether the relationship between amyloid and tau is different within the AT versus the PM networks.

Discussion

The aim of the present study was to evaluate the structural and molecular vulnerabilities of two specific MTL networks, namely, the AT and PM networks, over the course of AD. In particular, we aimed to determine whether these networks displayed evidence of dissociable neurodegenerative change and the degree to which they differed in underlying molecular pathology, which would have implications for the pathophysiologic cascade of AD and the relative vulnerability of these networks. To that end, we used a combination of rs-fMRI, longitudinal structural MRI, as well as tau and amyloid PET data acquired in three distinct cohorts covering the full AD spectrum.

Consistent with prior literature, our functional connectivity analyses performed in a group of 68 CN individuals revealed two specific MTL networks. Overall, the amygdala and the temporopolar cortex (extending posteriorly to inferior and middle temporal gyri) were more functionally connected to a seed composed of the aH and BA35 than the pH and PHC. In contrast, the posterior cingulate cortex, the precuneus, the cuneus, the angular and occipital gyri and the right insula were more functionally connected to pH/PHC than aH/BA35. As expected, these specific networks are overall in line with the two systems proposed by Ranganath and Ritchey (2012) in their review based on anatomic and functional studies performed in humans, monkeys and rats. Recently, the AT and PM networks have been investigated in vivo in humans using neuroimaging techniques. Studies have consistently reported two distinct MTL systems, although subtle differences can be noticed, likely because different approaches were employed. For example, the AT and PM networks were highlighted using seeds made of PRC versus PHC (Kahn et al., 2008; Libby et al., 2012; Zhuo et al., 2016), aH versus pH (La Joie et al., 2014; Das et al., 2015; Vogel et al., 2020), or anterolateral ERC versus posteromedial ERC (Adams et al., 2019). To our knowledge, the present study is the first to use combinations of hippocampal and extrahippocampal regions (aH/BA35 vs pH/PHC) as seeds. Note that it might appear more appropriate to use BA35 alone as a seed to parallel the structural covariance analyses. However, as illustrated in Figure 10, a subpart of the AT network was obtained, since BA35 is only a portion of PRC. Further, BA35 is a very small region located in a part of the brain which is prone to distortion in fMRI scans, and might not be ideal to be used as a seed by itself. Thus, combining MTL nodes belonging to the same network appeared as a reasonable approach and allowed to highlight specific networks corresponding to the AT and PM systems.

Figure 10.

Functional connectivity of MTL subregions, alternative seeds. AT and PM functional networks defined using seed-based correlations (seed: BA35 for AT; PHC for PM) in the CN population from the ABC. Connectivity maps were compared using a paired t test (P_{cluster(corr)} < 0.05, P_{voxel(uncorr)} < 0.01/P_{cluster(corr)} < 0.05, P_{voxel(uncorr)} < 0.005/P_{cluster(corr)} < 0.05, P_{voxel(uncorr)} < 0.001) to highlight specific networks.

Using longitudinal structural MRI data in a large population from ADNI GO/2 covering the AD spectrum from CN to mild AD-related dementia, our analyses revealed distinct MTL structural covariance patterns. BA35 atrophy covaried with atrophy of the amygdala and the left temporal cortex, while PHC atrophy structurally covaried with longitudinal atrophy of the posterior cingulate cortex, the precuneus, the cuneus and the angular and occipital gyri. Over the past few years, several studies have been performed to evaluate MTL structural covariance with the rest of the brain in the context of normal and pathologic aging (Bohbot et al., 2007; Olman et al., 2009; Persson et al., 2014; Montembeault et al., 2016; Stening et al., 2017; Chang et al., 2018; Li et al., 2018, 2019; Nordin et al., 2018; Wang et al., 2018; Plachti et al., 2020; Vogel et al., 2020). Previous investigations demonstrated that structural covariance differs across the longitudinal axis of the hippocampus, with aH overall covarying with temporal areas while pH covaried with parietal and occipital regions (Persson et al., 2014; Stening et al., 2017; Nordin et al., 2018; Li et al., 2019; Vogel et al., 2020). These results are consistent with our observations. However, it is important to note that these previous works were performed using cross-sectional data only, while our analyses were based on a longitudinal design, which is a more suitable approach to evaluate more pure effects of active neurodegeneration and disease progression topography, given that cross-sectional analyses may be influenced by developmental covariance. In addition, the present study is the first, to our knowledge, to contrast the structural covariance of two extrahippocampal regions (BA35 and PHC) across the AD spectrum. Importantly, BA35 and PHC structural covariance patterns overlapped with the AT and PM networks, respectively, supporting the notion that not only are these regions important nodes of functionally dissociable networks, but that neurodegeneration occurs in a dissociable fashion between them in the context of AD. Our analyses also revealed specific correlations between structural covariance and functional connectivity. These observations are in line with several neuroimaging works showing a relationship between functional connectivity and tau/neurodegeneration topography and propagation (Seeley et al., 2009; Zhou et al., 2012; Mutlu et al., 2017; Bischof et al., 2019; Ossenkoppele et al., 2019; Franzmeier et al., 2020a,b). Altogether, these observations performed in vivo in humans support the “prion-like” propagation theory which proposed that AD pathology, more particularly tau pathology and subsequent atrophy, spreads across anatomically interconnected brain regions (Brettschneider et al., 2015; Peng et al., 2020). In addition, the MTL structural covariance patterns highlighted here overall recapitulate the global AD atrophy pattern, classically described in the literature, where temporal and parietal regions are among the most affected brain areas (Dickerson et al., 2009; La Joie et al., 2012; Pini et al., 2016). This strengthens the view that the MTL is an epicenter of AD neurodegeneration but also suggests that pathology propagates to the neocortex through two distinct networks.

Lastly, we used PET data from ADNI 3, acquired in Aβ+ CN, MCI and AD patients, to evaluate the relative vulnerability of the AT and PM networks to amyloid and tau pathologies. Our analyses showed that tau uptake was relatively higher in the AT network than in the PM network, while amyloid uptake was relatively higher in the PM network than in the AT network. Similar to tau uptake, tau to amyloid ratio signal was more prominent in the AT network than in the PM network. Note that this ratio signal was particularly high in the MTL, suggesting that tau pathology in this region is relatively independent of amyloid. In addition, we found a significant amyloid × network interaction, where the slope of the association between tau and amyloid SUVRs was steeper within the AT network than in the PM network. This indicates that for a given level of amyloid, tau level is higher in AT than in PM. These results may suggest that more of the tau in the PM system is associated with neuritic tau given the higher amyloid plaque burden while we might expect more tangles in the absence of neuritic plaques in AT. However, this hypothesis is purely speculative. While tau levels are lower in the PM network, there is still significant tau pathology present in these regions, at least in the majority of patients studied here. Lastly, local relationships between amyloid and tau were explored in the present study, but nonlocal interactions between the two pathologies may also exist.

Overall, our observations are consistent with recent reports investigating AD pathology within MTL networks (Adams et al., 2019; Maass et al., 2019). More precisely, Adams et al. (2019) showed that in CN older adults, tau pathology is preferentially deposited in regions that are functionally connected to the anterolateral ERC (i.e., the AT system), with stronger connectivity being associated with increased tau deposition. Maass et al. (2019) also described early tau deposition in the AT system, which was associated with deficits in mnemonic object discrimination, while the presence of amyloid deposits was preferentially found in PM regions. Thus, early AD pathology seems to target both systems and might explain the functional disruptions observed in patients in previous investigations (Das et al., 2015; Chen et al., 2016; Liu et al., 2016; Berron et al., 2020), but does so with differential pathologic underpinnings. Altogether, these data suggest a predominant role of the AT and PM networks in the pathophysiology of AD.

Our study has some limitations. We used data from three different datasets: resting-state functional connectivity data were acquired at University of Pennsylvania, longitudinal structural MRI data came from ADNI GO/2, and amyloid and tau PET data came from ADNI 3. Thereby, we could not assess the direct relationships between functional connectivity, the presence of pathologies and structural covariance. Having all such data acquired in the same participants would have allowed us to explore pathologic mechanisms more deeply. However, this can also be seen as a strength, since this indicates that our results are potentially generalizable. Further studies performed on a unique sample (with longitudinal imaging in all modalities) and including more advanced patients are thus needed to fully understand the involvement of the AT and PM networks in the pathophysiology of AD across the disease spectrum.