Structural and molecular determinants of medial temporal lobe network vulnerability in aging and Alzheimer’s disease Authors: Elise Saul, Léa Chauveau, Brigitte Landeau, Jade Lasserve, Blandine Montagne, Emilie Foyard, Géraldine Poisnel, Gaël Chételat, Robin de Flores Published: 2026 ABSTRACT BackgroundThe medial temporal lobe is organized into two memory-critical networks: the anterior–temporal and posterior–medial systems. While these systems show selective vulnerability to aging and Alzheimer’s disease, the underlying structural and molecular determinants of this susceptibility remain unclear. We aimed to characterize the specific white matter pathways supporting these networks, investigate how amyloid-β and neuroinflammation interact to impact their integrity, and determine if structural changes relate to functional connectivity alterations.MethodsIn 88 cognitively unimpaired (CU) older adults (≥ 65 years) from the Age-Well cohort (NCT02977819), we combined longitudinal diffusion MRI tractography, resting-state fMRI, amyloid-β PET (Florbetapir), and plasma glial fibrillary acidic protein (GFAP). We reconstructed fiber pathways of the perirhinal (anterior-temporal system hub) and parahippocampal (posterior–medial system hub) cortices and quantified their microstructural integrity, network connectivity, and relationships to pathology.ResultsThe anterior–temporal and posterior–medial systems relied on partly distinct structural pathways. Both involved the inferior longitudinal and cingulum bundles, but the anterior–temporal system was specifically associated to their inferior portions, as well as thalamic radiations, and callosal fibers, whereas the posterior–medial system relied more on their superior portions and the inferior fronto-occipital fasciculus. We identified a non-linear, inverted U-shaped association between pathway integrity and amyloid-β burden, suggesting dynamic structural changes across early pathological stages. Crucially, plasma GFAP moderated this relationship for the posterior-medial pathway: the negative impact of amyloid-β on structural integrity was exacerbated in individuals with higher astroglial reactivity, highlighting a synergistic pathological effect. Finally, in high amyloid-β individuals, increased anterior-temporal functional connectivity correlated with lower anterior-temporal pathway integrity, suggesting that network higher FC may represent a maladaptive response to early amyloid-β deposition leading to white matter integrity loss.ConclusionsTogether, our findings reveal distinct mechanisms of vulnerability within medial temporal networks: while the posterior-medial system is primarily sensitive to the synergistic effects of amyloid-β and astroglial reactivity, the anterior-temporal system shows lower structural integrity linked to higher FC in the presence of amyloid. By providing a mechanistic framework for these early disruptions, this study advances the understanding of preclinical Alzheimer’s disease and identifies specific structural–functional signatures that could serve as sensitive biomarkers for targeted interventions.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s13195-026-02125-1. Background Alzheimer’s disease (AD) is the most common cause of dementia, affecting more than 55 million people worldwide, yet its mechanisms remain incompletely understood [1]. The medial temporal lobe (MTL) is among the earliest regions affected in AD, marked by neurofibrillary tangle (NFT) accumulation, while amyloid-β (Aβ) deposits together with NFTs constitute the pathological hallmarks of the disease [2, 3]. The MTL is embedded in two large-scale functional networks with distinct cognitive roles: the anterior-temporal (AT) network, centered on the perirhinal cortex (PRC) and supporting semantic memory, object recognition, and affective processes; and the posterior-medial (PM) network, centered on the parahippocampal cortex (PHC) and involved in episodic memory, spatial navigation, and social cognition [4]. Pathology distributes unevenly across these systems: tau preferentially accumulates in AT regions, while Aβ is more prominent in PM regions [5–8]. These networks are also functionally impaired in both AD and aging, contributing to cognitive decline [9–15]. We recently extended these findings, showing that higher AT connectivity is linked to AD progression, whereas lower PM connectivity characterizes normal aging [16]. Despite this evidence, little is known about the structural connectivity (SC) that supports these networks. The MTL is characterized by a sophisticated microstructural organization, with distinct cytoarchitectural properties across its subregions – the hippocampus, entorhinal, perirhinal and parahippocampal cortices – that underlie their differential connectivity profiles [17–20]. The AT system is primarily linked to anterior temporal regions (e.g., amygdala, perirhinal cortex), while the PM system connects to posterior and medial regions, including the posterior cingulate cortex and precuneus [4]. To our knowledge, only one diffusion imaging study has delineated their fiber pathways, implicating distinct segments of the inferior longitudinal fasciculus (ILF), inferior fronto-occipital fasciculus (IFOF), and cingulum bundle [21]. Clarifying these structural circuits is essential, as the MTL’s unique microstructure renders it highly susceptible to earliest stages of AD. The main white matter (WM) tracts connecting MTL subregions to the rest of the brain – including the fornix, the cingulum bundle, the ILF, and the IFOF – have been among the earliest detectable sites of microstructural alterations in the AD continuum [22–25]. These microstructural modifications are increasingly recognized as early markers closely linked to the initial accumulation of Aβ and tau pathology [26, 27]. Indeed, WM integrity is highly sensitive to AD-related processes and may underlie the differential vulnerability of AT and PM systems. Yet, studies linking Aβ to WM changes have produced inconsistent results—some reporting strong microstructural alterations, others only weak or null effects [22, 27–30]. One explanation is that Aβ and WM integrity are related in a non-linear fashion, with early Aβ deposition associated with transient increases in diffusion metrics, followed by degeneration at higher levels [26, 30, 31]. This early increase in thought to reflect a neuroinflammatory « swelling» effect: Aβ plaques trigger a glial response (microglial and astroglial activation) characterized by cellular hypertrophy and expansion into the extracellular space [32–34]. This process can temporarily increase diffusion restriction or alter water displacement before overt axonal loss occurs. Astroglial reactivity can be captured by blood-based glial fibrillary acidic protein (GFAP), now included in the NIA-AA (National Institute on Aging and Alzheimer’s Association) biomarker framework as an indicator of neuroinflammatory activity [35]. GFAP may thus serve as a proxy for early inflammatory phase that masks or precedes classical neurodegenerative signals in diffusion imaging [36]. Finally, although alterations in WM integrity may contribute to AT and PM dysfunction, structural changes alone cannot fully account for the functional disruptions observed in aging and AD. Investigating how SC relates to functional connectivity (FC) in these networks therefore offers crucial insight into whether pathology drives parallel structural and functional decline, or whether functional changes reflect reorganization beyond structural constraints. More specifically, in higher-order association areas such as the MTL, functional interactions often extend beyond direct anatomical links, reflecting the flexibility of these networks [37–40]. Disentangling SC–FC coupling in AT and PM systems may therefore shed light on how AD and aging disrupt brain organization. In this study, we sought to: (i) map the WM pathways linking PRC and PHC to the AT and PM networks, (ii) examine how Aβ load and astroglial reactivity (GFAP) affect the integrity of these circuits, and (iii) assess the relationship between fiber integrity and FC within these networks. We hypothesized that fiber integrity would show a non-linear association with Aβ. Furthermore, based in reported AT higher FC and PM lower FC in AD, we hypothesized opposite SC-FC relationships: specifically, that reduced structural integrity would be associated with higher AT FC but lower PM FC. Methods Participants All participants were included in the Age-Well randomized clinical trial of the Medit-Ageing European Project [41], sponsored by the French National Institute of Health and Medical Research (INSERM). The protocol involved an 18-month non-pharmacological intervention addressing both emotional and cognitive aspects of aging. At baseline, participants underwent a battery of assessments, including neuroimaging, cognitive tests, blood sampling and lifestyle questionnaires. Then, they were randomly divided into three groups (ratio of 1:1:1): one group received meditation training, one a non-native language (i.e., English) training (active controls), and the last group no intervention (passive controls). At the end of the intervention, the same assessments conducted at baseline (neuroimaging, cognitive tests, and lifestyle questionnaires) were repeated. In the present study, data from the baseline and follow-up visits were used to perform longitudinal analyses. Participants were recruited from the general population, they were older than 65 years, native French speakers, retired for at least 1 year, had at least 7 years of education and had performed within the normal range on standardized cognitive tests, including the Mini-Mental State Examination (MMSE) for global cognition, the RL-RI16 for verbal episodic memory, and the Modified Card Sorting Test for executive functioning [41]. They did not show evidence of a major neurological or psychiatric disorder (including alcohol or drug abuse), history of cerebrovascular disease, chronic disease or acute unstable illness, and current or recent medication usage that may interfere with cognitive functioning. Only participants with available structural and resting-sate functional MRI (rs-fMRI), diffusion kurtosis imaging, and 18F-Florbetapir (18F-AV-45) PET were included. A total of 88 participants met all the inclusion criteria, demographic characteristics of the sample are reported in Table 1 (see Supplementary Figure S1 for the flowchart of the inclusion process). Baseline data were collected from November 2016 to April 2018, and follow-up data from September 2018 to November 2019 (mean delay between baseline and follow-up of 21.5 ± 0.25 months). The mean interval between MRI and PET assessments was 20.67 days (± 22.36) at baseline and 20.75 days (± 16.85) at follow-up (see Supplementary Fig S2 for assessment timeline of MRI-PET acquisition). All participants provided written informed consent prior to participation in accordance with the Declaration of Helsinki. The Age-Well trial was approved by the Nord Ouest II ethics committee (Caen, France) (Clinical Trial Number NCT02977819||https://www.clinicaltrials.gov/; trial registration number: EudraCT: 2016–002441–36; IDRCB: 2016-A01767-44; registration date: 2016–11–25).Table 1Demographic characteristics of the Age-Well and IMAP cohortsBaselineFollow-upp-valueAge-WellAge, years68.98 ± 3.91[65—83]70.3 ± 3.53[67—85]-Female, N (%)51 (58)--Education, years13.03 ± 3.26[7–22]--Intervention group, C/E/Med (%)25/35/26(28.4/38/33.6)--APOE4 carriers, N (%)20 (22.7)--Aβ, centiloid8.23 ± 24.21[−23.4 – 101.3]10.37 ± 26.99[−23.79; 101.82] < 0.05 *GFAP, pg/mL155.96 ± 54.73[58.6–353.4]169.05 ± 70.85[62.94; 496.42] < 0.001***BMI25.83 ± 3.74[18.1; 37.15]25.94 ± 4.09[17.42; 37.27]0.556GFR, pg/mL80.48 ± 10.91[47; 105]79.86 ± 10.72[47; 99]0.119IMAPAge, years70.9 ± 6.95[60—84]73.07 ± 6.65[61—86]-Female, N (%)21 (50)--Education, years11.97 ± 3.88[7–20]--Aβ, centiloid3.19 ± 20.11[−24.43—58.1]3.47 ± 22.92[−24.98—69.58]0.615Descriptive statistics are reported at baseline and follow-up for each cohort separately. Continuous variables are expressed as mean ± standard deviation [range]. Categorical variables are expressed as count and percentage. A dash (−) indicates that the variable was not reassessed at follow-up. For the Age-Well cohort, intervention groups are reported as passive control (C), active control (E), and meditation training (Med). Amyloid burden was measured using 18F-Florbetapir PET and expressed in centiloid. Longitudinal changes in continuous variables were assessed using paired Wilcoxon signed-rank tests (baseline vs. follow-up)Abbreviations: Aβ amyloid-β, GFAP glial fibrillary acidic protein, BMI body mass index, GFR glomerular filtration rate, APOE4 apolipoprotein E ε4 allele, C passive control, E active control, Med meditation trainingSignificance levels: *p< 0.05; **p< 0.01; ***p< 0.001 Table 1Demographic characteristics of the Age-Well and IMAP cohortsBaselineFollow-upp-valueAge-WellAge, years68.98 ± 3.91[65—83]70.3 ± 3.53[67—85]-Female, N (%)51 (58)--Education, years13.03 ± 3.26[7–22]--Intervention group, C/E/Med (%)25/35/26(28.4/38/33.6)--APOE4 carriers, N (%)20 (22.7)--Aβ, centiloid8.23 ± 24.21[−23.4 – 101.3]10.37 ± 26.99[−23.79; 101.82] < 0.05 *GFAP, pg/mL155.96 ± 54.73[58.6–353.4]169.05 ± 70.85[62.94; 496.42] < 0.001***BMI25.83 ± 3.74[18.1; 37.15]25.94 ± 4.09[17.42; 37.27]0.556GFR, pg/mL80.48 ± 10.91[47; 105]79.86 ± 10.72[47; 99]0.119IMAPAge, years70.9 ± 6.95[60—84]73.07 ± 6.65[61—86]-Female, N (%)21 (50)--Education, years11.97 ± 3.88[7–20]--Aβ, centiloid3.19 ± 20.11[−24.43—58.1]3.47 ± 22.92[−24.98—69.58]0.615Descriptive statistics are reported at baseline and follow-up for each cohort separately. Continuous variables are expressed as mean ± standard deviation [range]. Categorical variables are expressed as count and percentage. A dash (−) indicates that the variable was not reassessed at follow-up. For the Age-Well cohort, intervention groups are reported as passive control (C), active control (E), and meditation training (Med). Amyloid burden was measured using 18F-Florbetapir PET and expressed in centiloid. Longitudinal changes in continuous variables were assessed using paired Wilcoxon signed-rank tests (baseline vs. follow-up)Abbreviations: Aβ amyloid-β, GFAP glial fibrillary acidic protein, BMI body mass index, GFR glomerular filtration rate, APOE4 apolipoprotein E ε4 allele, C passive control, E active control, Med meditation trainingSignificance levels: *p< 0.05; **p< 0.01; ***p< 0.001 Demographic characteristics of the Age-Well and IMAP cohorts 68.98 ± 3.91 [65—83] 70.3 ± 3.53 [67—85] 13.03 ± 3.26 [7–22] 25/35/26 (28.4/38/33.6) 8.23 ± 24.21 [−23.4 – 101.3] 10.37 ± 26.99 [−23.79; 101.82] 155.96 ± 54.73 [58.6–353.4] 169.05 ± 70.85 [62.94; 496.42] 25.83 ± 3.74 [18.1; 37.15] 25.94 ± 4.09 [17.42; 37.27] 80.48 ± 10.91 [47; 105] 79.86 ± 10.72 [47; 99] 70.9 ± 6.95 [60—84] 73.07 ± 6.65 [61—86] 11.97 ± 3.88 [7–20] 3.19 ± 20.11 [−24.43—58.1] 3.47 ± 22.92 [−24.98—69.58] Descriptive statistics are reported at baseline and follow-up for each cohort separately. Continuous variables are expressed as mean ± standard deviation [range]. Categorical variables are expressed as count and percentage. A dash (−) indicates that the variable was not reassessed at follow-up. For the Age-Well cohort, intervention groups are reported as passive control (C), active control (E), and meditation training (Med). Amyloid burden was measured using 18F-Florbetapir PET and expressed in centiloid. Longitudinal changes in continuous variables were assessed using paired Wilcoxon signed-rank tests (baseline vs. follow-up) Abbreviations: Aβ amyloid-β, GFAP glial fibrillary acidic protein, BMI body mass index, GFR glomerular filtration rate, APOE4 apolipoprotein E ε4 allele, C passive control, E active control, Med meditation training Significance levels: *p< 0.05; **p< 0.01; ***p< 0.001 To replicate our findings in an independent sample, we additionally included CU adults from the IMAP study (Caen, France), a multimodal neuroimaging cohort enrolled between 2008 and 2016 (ClinicalTrials.gov: NCT01638949). From the full IMAP cohort, we selected CU participants aged over 60 years old with available structural MRI and 18F-AV-45 PET at the first two acquisition timepoints (18 months apart). Participants with a history of neurological or psychiatric disorder, alcohol use disorder, drug abuse, or head trauma were excluded. A total of 41 participants met all inclusion criteria, demographic characteristics of the sample are reported in Table 1 alongside those of the Age-Well cohort. Of note, the MTL and tis subregions have been characterized in this cohort by previous publications from our lab [13, 42, 43]. The IMAP + study was approved by the Comité de Protection des Personnes Nord-Ouest III, and all participants provided written informed consent prior to participation. Further details on the cohort and inclusion criteria are provided in Supplementary Methods. Neuroimaging data acquisition Each subject underwent the same baseline and follow-up MR and PET exams at the CYCERON center (Caen, France) using a 3 T Philips (Eindhoven, The Netherlands) scanner and a GE Healthcare PET-CT (Discovery RX VCT 64; Milwaukee, Wis, United States of America). First, T1-weighted structural images were acquired (Repetition Time (TR) = 7.1 ms; Echo Time (TE) = 3.3 ms; flip angle = 6°; 180 slices; slice thickness = 1 mm; Field of View (FoV) = 256 × 256 mm2; matrix = 256 × 256; in-plane resolution = 1 × 1 mm2; acquisition time = 5 min16). Secondly, resting-state functional MRI scans were obtained using an interleaved 2D T2* SENSE EPI sequence designed to reduce geometric distortions (2D-T2*-FFE- EPI axial, SENSE = 2.5; TR = 2400 ms; TE = 30 ms; flip angle = 85°; 44 slices with no gap; slice thickness = 2.8 mm; FoV = 200 × 200 mm2; in plane resolution = 2.5 × 2.5 mm2; 200 volumes, acquisition time = 8.19 min). Then, a high-resolution T2-weighted spin echo anatomical acquisition (3D-T2-SE sagittal; SENSE factor = 2; TR = 2500 ms; TE = 236 ms; flip angle = 90°; 180 slices with no gap; slice thickness = 1 mm; FoV = 250 × 250 mm2; in-plane resolution = 0.98 × 0.98 mm2) and a non-EPI T2* volumes (2D-T2*-FFE axial, SENSE = 2, TR = 600 ms, TE = 16 ms, flip angle = 20°, 70 slices with no gap, slice thickness = 2 mm, FoV = 256 × 256 mm2, in-plane resolution = 2 × 2 mm2) were additionally acquired for fMRI pre-processing steps. Lastly, an echoplanar imaging/spin echo diffusion-weighted sequence (diffusion kurtosis imaging [DKI]) was performed at multiple shell: 3b values (0, 1000, 2000s/mm2) (axial; 30 directions for each non-zero b-value, TR = 6,100 ms, TE = 101 ms, 48 slices, FoV = 216 × 216 mm2, voxel size = 2.7 × 2.7 × 2.7 mm3) and additional blips images with b = 0 s/mm2 (number of signal averages 9) were acquired in reverse phase encoding direction for distortion correction. Finally, 18F-AV-45 PET scans were acquired (resolution = 3.76 × 3.76 × 4.9 mm; FoV = 157 mm; voxel size = 1.95 × 1.95 × 3.27 mm). Before the PET acquisition, a transmission scan was performed for attenuation correction. For 18F-AV-45 PET, each participant underwent a 10-min PET scan beginning 50 min after the injection of:4 MBq/Kg of 18F-AV-45. The preprocessing procedures of neuroimaging data were described in Supplementary material, (see ‘Methods’ section). IMAP participants were repeatedly scanned at the CYCERON center (Caen, France) using a 3 T Philips Achieva scanner. T1-weigthed anatomical images were acquired using a three-dimensional fast-filed echo sequence (3D-T1-FEE sagittal; SENSE factor = 2; TR = 20 ms; TE = 4.6 ms; flip angle = 10°; 180 slices with no gap; slice thickness = 1 mm; FoV = 256 × 256 mm2; in-plane resolution = 1 × 1 mm2). Resting‐state functional volumes were obtained using an interleaved2D T2* SENSE EPI sequence designed to reduce geometric distortions (2D-T2*-FFE- EPI axial, SENSE = 2; TR = 2382 ms; TE = 30 ms; flip angle = 80°; 42 slices with no gap; slice thickness = 2.8 mm; FoV = 224 × 224 mm2; in plane resolution = 2.8 × 2.8 mm2; 280 volumes, acquisition time = 11.5 min). Medial temporal lobe segmentation The perirhinal cortex (PRC), composed of Brodmann areas 35 and 36, and parahippocampal cortex (PHC) were automatically segmented from each participant’s T1-weighted image using the Automated Segmentation for Hippocampal Subfields (ASHS) software with the ashsT1_atlas_upennpmc_07202018 atlas [42, 44]. Each segmentation was visually inspected. Failed segmentations were discarded. Left and right structures were merged to limit the number of analyses. The same segmentation procedure was applied in the IMAP cohort, using identical software and atlas parameters to ensure comparability across cohort [42]. Structural connectivity assessment DKI data were preprocessed using FSL (v5.0.9) to correct for susceptibility-induced geometric distortions (topup), eddy currents, and head motion [34]. The diffusion kurtosis tensor was estimated using the Diffusional Kurtosis Estimator (DKE v2.6) in MATLAB R2012b [35], followed by spatial smoothing (3.375 mm FWHM Gaussian filter). Derived-maps – including Mean (MK) and Kurtosis Fractional Anisotropy (KFA) – were then co-registered to each subject’s T1-weighted image. These two specific DKI metrics were selected to provide a multifaceted biological characterization of WM microstructure: MK summarizes the average degree of diffusion non-Gaussianity across all directions and is often used as a relatively robust global marker of tissue microstructural heterogeneity/restriction. KFA quantifies the directional anisotropy of kurtosis and is sensitive to white matter organization, including fiber coherence and axonal architecture. Together, these metrics offer superior sensitivity over traditional DTI markers for detecting subtle microstructural alterations in complex environments like the MTL. The procedure used to identify fibers belonging to the AT and PM networks—defined as those passing through the PRC and PHC—is illustrated in Fig. 1. The PRC and PHC regions of interest (ROIs) were first rescaled to DKI space and used as seeds for tractography analyses based on diffusion orientation distribution function (dODF) maps. The fiber trajectories were determined using a probabilistic algorithm, modelling the most probable trajectory, consistent with the literature [45]. The tractography parameters were selected based on the literature and optimized empirically [46, 47]. Specifically, the minimum intensity threshold for direction vectors was set at 0.08, the calculation step size was 0.27 mm (1/10 of the voxel size), and the opening angle cone was 40°. Only fibers with lengths between 30 and 1000 mm were considered. One million iterations were performed with starting points randomly distributed in both white and gray matter, and cerebrospinal fluid (CSF) was defined as a termination region [45].Fig. 1Methodological steps to highlight the PRC and PHC pathways (A) as well as the AT and PM functional connectivity (B). Adapted from Chauveau et al., 2025 [16]. FWE, family wise error; TFCE, threshold-free cluster enhancement Methodological steps to highlight the PRC and PHC pathways (A) as well as the AT and PM functional connectivity (B). Adapted from Chauveau et al., 2025 [16]. FWE, family wise error; TFCE, threshold-free cluster enhancement Whole-brain connectomes were filtered by the previously defined ROIs (PRC/PHC) to retain only fibers passing through them. Track Density Imaging (TDI) maps (2 maps for each subject at each timepoint) were then extracted from the filtered tractographies. These maps represent the number of fibers passing through each voxel, with signal intensity proportional to the number of fibers [48]. TDI images from each subject were coregistered to T1-weighted MRI, spatially normalized in MNI space by applying the deformation parameters from the corresponding T1-weighted MRI segmentation, for inter-individual comparison and masked by the group's WM. To mitigate the influence of outliers, voxel intensities were quantitatively winsorized by applying an upper threshold defined as Q3 + 1.5 × (Q3 – Q1), where Q1 and Q3 represent the first and third quartiles based on the distribution of non-zero voxel intensities. Intensities above this threshold were limited to the maximum value observed below it. To highlight the fibers that were specifically passing through PRC versus PHC (and vice versa), the baseline TDI images were smoothed to 2 mm and then compared using paired permutation t-tests (FSL randomise command, with TFCE [Threshold-Free Cluster Enhancement]). The resulting statistical maps were then thresholded at p < 0.05 with a Family Wise Error (FWE) correction, equating to a 95% confidence interval with no false positives, and subsequently binarized to obtain specific masks of fibers passing through either PRC or PHC. SC measures were estimated by averaging signal intensities within PRC/PHC fiber pathway masks extracted from the MK and KFA maps. As diffusion-weighted imaging was not available in the IMAP cohort, we adopted an alternative approach to estimate WM integrity, following the methodology described in Bejanin et al. [49]. Specifically, the PRC and PHC fiber pathway masks defined in the Age-Well cohort (see above) were applied to modulated white matter probability maps generated from SPM12 segmentation of each IMAP participant’s T1-weighted images. For each participant and each timepoint, WM density was estimated by averaging the signal within the PRC and PHC masks. Functional connectivity assessment Rs-fMRI data were first visually inspected for artifacts, lesions, or signal loss. Artifact detection was further enhanced using the TSDiffana routine. Participants with significant head motion (> 2 mm translation or 2° rotation) were excluded. The preprocessing pipeline, conducted via SPM12 and FSL v6.0.5, included slice-timing correction and realignment to the first volume. To ensure high-quality denoising, an Independent Component Analysis (ICA)-based procedure was performed using FSL-MELODIC. Noise-related components (e.g., motion, vascular, or checkerboard artifacts) were manually identified and regressed out for each subject. For spatial accuracy, a multi-step rigid coregistration was employed to align EPI data onto T1-weighted images by leveraging intermediate scans: (1) EPI onto non-EPI T2*, (2) non-EPI T2* onto T2, and (3) T2 onto T1. To minimize geometric distortions, the EPI images were warped to the non-EPI T2* volumes. Finally, the resulting distortion-corrected volumes were normalized to the MNI template using parameters derived from T1 segmentation, smoothed with a 4 mm FWHM Gaussian kernel, and band-pass filtered (0.01–0.08 Hz) using AFNI. The procedure used to estimate FC in the AT and PM networks is the same as in Chauveau et al. (2025) [16]. Following the scheme illustrated in Fig. 1, the PRC (Brodmann areas 35/36) and PHC —normalized to MNI space—served as seeds for the AT and PM networks, respectively. Nuisance regressors comprised low-frequency drifts, mean signals from WM, CSF, and whole brain (plus their first temporal derivatives), and six rigid-body motion parameters. For each subject, voxel-wise Pearson correlations were computed between the mean time series of each ASHS-T1–derived seed and every gray-matter voxel, producing two connectivity maps per scan (PRC → whole brain; PHC → whole brain). To mitigate susceptibility-induced signal loss, voxels exhibiting substantial dropout were excluded via a combined T1/T2*–derived gray-matter mask thresholded at 0.25. Resulting maps were Fisher z-transformed and smoothed with a 6.3 mm FWHM (Full Width at Half Maximum) Gaussian kernel to perform permutation test for the second-level analysis. Indeed, network-specific regions were identified via permutation-based inference (FSL randomize with TFCE; FWE-corrected p < 0.05). One-sample t-tests were performed on both perirhinal and parahippocampal maps to retain only voxels significantly connected to each seed, and paired t-tests (perirhinal > parahippocampal, and vice versa) were conducted to isolate regions preferentially connected to each seed. The AT mask consisted of the conjunction between the perirhinal one-sample and perirhinal-to-parahippocampal paired results; the PM mask analogously combined parahippocampal one-sample and paired outputs. Subject-specific AT and PM masks were then generated by intersecting group-level masks with individual gray-matter segmentations (threshold 0.5), excluding seed voxels to prevent autocorrelation. Within these masks, mean positive Fisher-z values from unsmoothed maps were extracted as indices of global AT and PM connectivity. The same pipeline was used in both cohorts. Blood sampling Blood samples were collected in the morning and after an overnight fast. 18 tubes (68 mL) of blood were collected at baseline and 16 tubes (62 mL) at follow-up. Blood glial fibrillary acid protein (GFAP) levels were measured using the SIMOA (Single Molecule Array) technology with commercial Quanterix kits (Neurology 2-plexB #103520) on a HD-X analyzer. Blood samples were detected and 99.3% displayed values above the lower limit of quantification reported for the assay (LLOQ = 16.6 pg/mL). Three control samples were run at the beginning and at the end of the plate to assess variability. Intra- and inter-assays coefficients of variation (CV) were of 7.4% and 9.7% for GFAP. Amyloid load assessment 18F-AV-45 PET data were co-registered to their corresponding T1-weighted MRI and normalized to the MNI template using deformation parameters from the T1w segmentation procedure. Images were then quantitively normalized using the cerebellar gray matter as the reference region, resulting a global cortical standardized uptake value ratio (SUVRs). Amyloid uptake was extracted and averaged across AD-sensitive regions, including the frontal, temporal and parietal cortices, the precuneus, the anterior striatum, and the insular cortex [50]. SUVR values were converted to centiloid following Klunk et al., (2015) [51] processing method adapted for Florbetapir SUVRs [52]. This procedure was previously validated on data from the Global Alzheimer’s Association Interactive Network (GAAIN). Participants with a centiloid value superior to 12 were considered Aβ positive and the others negative [53–55]. The same pipeline was used in both cohorts. Cognitive composite scores Four cognitive composite scores were used to capture distinct cognitive domains. All composites were constructed by standardizing constituent test scores using the mean and standard deviation of the full sample at baseline (z-scores), and were subsequently restandardized to facilitate longitudinal interpretation, such that changes reflect shifts in standard deviation units. The Preclinical Alzheimer's Cognitive Composite (PACC-5) is a global cognitive measure sensitive to preclinical AD-related decline. It was derived from five neuropsychological measures: the Dementia Rating Scale-2 (total score), the Coding subtest (raw score) of the Wechsler Adult Intelligence Scale (WAIS)-IV, the California Verbal Learning Test – Second Edition (CVLT-II) (delayed recall), the Logical Memory Story B (delayed recall) of the Wechsler Memory Scale (WMS)-IV, and the Category Fluency (GREFEX battery; number of animals produced in 2 min). The unweighted average of the five resulting z-scores yielded the PACC-5. The Verbal Episodic Memory Composite was computed from two episodic memory tests, each providing multiple scores: the California Verbal Learning Test – Second Edition (CVLT-II) and the Logical Memory Story B of the WMS-IV — the same tests used in the PACC-5. Constituent scores included CVLT-II trials 1–5 sum, immediate free recall, and delayed free recall, as well as Logical Memory immediate and delayed free recall. The unweighted average of these five z-scores constituted the composite. Two additional composites captured Executive Functioning and Attention/Processing Speed. The Executive Functioning Composite was derived from the Digit Span Backward subtest (raw score) of the WAIS-IV, the Trail Making Test Part B (GREFEX battery; response time; TMT-B), the Stroop (GREFEX battery; interference index = response time for interfering – neutral items), and Letter Fluency (GREFEX battery; number of correct words starting with "p" produced in 2 min). The Attention/Processing Speed Composite was computed from the Trail Making Test Part A (GREFEX battery; response time), the Stroop (GREFEX battery; naming condition; response time), the Digit Span Forward subtest (raw score) of the WAIS-IV, and the Coding subtest (raw score) of the WAIS-IV. For both composites, Trail Making Test and Stroop scores were reversed prior to averaging so that higher composite scores consistently indicate better performance. For both composites, Trail Making Test and Stroop scores were reversed prior to averaging so that higher composite scores consistently indicate better performance. Statistical analyses Statistical analyses were conducted in R version 4.2.2, using the following packages: tidyverse, rstatix, lme4 and emmeans. The ggplot2 R package was used for figure creation. All analyses were performed on longitudinal data (baseline and follow-up) using linear mixed-effects models, which are well-suited for managing repeated measures within subjects. The models were estimated using restricted maximum likelihood (REML) and incorporated random intercepts for each participant. Covariates included sex, age, years of education, time, intervention group, and body mass index (BMI) and glomerular filtration rate (GFR° for the analysis with GFAP level. Fixed effects were tested for statistical significance using Type II or Type III—for interactions—F-tests, with degrees of freedom approximated via Satterthwaite’s method. We assessed the associations between Aβ load and the integrity of PRC and PHC fiber pathways using polynomial models (second order). A total of four models were estimated (2 SC measures for PRC and 2 SC measures for PHC). We further investigated how GFAP levels modulated these relationships by conducting interaction analyses (SC ~ Aβ * GFAP + covariates). Additionally, we investigated the associations between SC and FC, again performing four separate analyses: two measures for PRC fiber pathway integrity in relation to AT FC, and four measures for PHC fiber pathway integrity in relation to PM FC. To further explore the moderating role of Aβ load, interaction analyses (SC ~ FC * Aβ + covariates) were conducted for each of the four models. Finally, p-values across all primary analyses were corrected for multiple comparisons using the False Discovery Rate (FDR) method as described by Benjamini and Hochberg [56]. Results Identification of PRC and PHC fiber pathways First, we wanted to highlight the WM pathways associated with PRC and PHC, core regions of the AT and PM networks, respectively. We found that the PRC and PHC share common fasciculus, including the hippocampal cingulum and the inferior longitudinal fasciculus (ILF). Note that given the statical design we used (paired t-tests), we also could highlight that the inferior parts of the hippocampal cingulum and ILF were specifically connected to PRC while the superior parts of these tracts were connected to PHC. In addition, the PRC fiber pathway specifically included the thalamic radiations and the corpus callosum, while the PHC fiber pathway specifically included the cingulum and the inferior fronto-occipital fasciculus (IFOF) (see Fig. 2).Fig. 2Spatial distribution of PRC and PHC fiber pathways. White matter fiber pathways specifically associated with the PRC (red) and the PHC (blue) are displayed on a standard MNI brain template. Abbreviation: MNI, Montreal Neurological Institute; PHC, parahippocampal cortex; PRC perirhinal cortex Spatial distribution of PRC and PHC fiber pathways. White matter fiber pathways specifically associated with the PRC (red) and the PHC (blue) are displayed on a standard MNI brain template. Abbreviation: MNI, Montreal Neurological Institute; PHC, parahippocampal cortex; PRC perirhinal cortex Relationship between Aβ load, fibers integrity and neuroinflammation biomarker As reported in Table 2, Table S1 and illustrated in Fig. 3, linear mixed-effects models captured a significant non-linear association between Aβ load (centiloid) and WM microstructure within PRC/PHC pathways. This inverted U-shaped relationship – whereby increasing Aβ load was initially associated with enhanced fiber integrity before declining—was observed across all DKI-derived metrics in the Age-Well cohort and WM integrity measure in the IMAP cohort. To further validate the non-linear nature of this association, we additionally modelled this relationship using Generalized Additive Mixed Models (GAMMs), which allow the shape of the association to be estimated flexibly from the data. GAMMs consistently yielded convergent results (see Table S2).Table 2Relationship between fibers integrity measures and amyloid loadSC ~ Aβ + Aβ 2 + covariatesIntegrity measuresPRC SCPHC SCF-valuep uncorrectedpFDRF-valuep uncorrectedpFDRAgeWellMKAβ6.3< 0.05*< 0.05*7.38< 0.01**< 0.01**Aβ 210.1< 0.01**< 0.01**9.07< 0.01**< 0.01**KFAAβ4.24< 0.05*< 0.05*15.4< 0.001***< 0.001***Aβ 28.57< 0.01**< 0.01**15.6< 0.001***< 0.001***IMAPAβ7.72< 0.01**< 0.05*6.55< 0.05*< 0.05*Aβ 22.13< 0.05*< 0.05*7.02< 0.05*< 0.05*Results from linear mixed-effects models of the form SC ~ Aβ + Aβ2 + covariates (age, sex, education level, intervention group (for Age-Well cohort only), visit) are reported for the PRC and PHC pathwaysAbbreviations: Aβ amyloid-β, FDR false discovery rate, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortexSignificance levels: *p< 0.05; **p< 0.01; ***p< 0.00Fig. 3Non-linear association between amyloid burden and fiber integrity in the PRC and PHC pathways across cohorts. Scatter plots illustrating the quadratic relationship between Aβ (centiloid) and fiber integrity for the PRC (red, top row) and PHC (blue, bottom row) pathways. Panel A shows results from the Age-Well cohort. Panel B shows results from the IMAP cohort. Each dot represents one observation, with grey lines connecting repeated measures from the same individual. Abbreviations: Aβ, amyloid-β; KFA, kurtosis fractional anisotropy, MK, mean kurtosis; PHC, parahippocampal cortex; PRC, perirhinal cortex Table 2Relationship between fibers integrity measures and amyloid loadSC ~ Aβ + Aβ 2 + covariatesIntegrity measuresPRC SCPHC SCF-valuep uncorrectedpFDRF-valuep uncorrectedpFDRAgeWellMKAβ6.3< 0.05*< 0.05*7.38< 0.01**< 0.01**Aβ 210.1< 0.01**< 0.01**9.07< 0.01**< 0.01**KFAAβ4.24< 0.05*< 0.05*15.4< 0.001***< 0.001***Aβ 28.57< 0.01**< 0.01**15.6< 0.001***< 0.001***IMAPAβ7.72< 0.01**< 0.05*6.55< 0.05*< 0.05*Aβ 22.13< 0.05*< 0.05*7.02< 0.05*< 0.05*Results from linear mixed-effects models of the form SC ~ Aβ + Aβ2 + covariates (age, sex, education level, intervention group (for Age-Well cohort only), visit) are reported for the PRC and PHC pathwaysAbbreviations: Aβ amyloid-β, FDR false discovery rate, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortexSignificance levels: *p< 0.05; **p< 0.01; ***p< 0.00 Relationship between fibers integrity measures and amyloid load Results from linear mixed-effects models of the form SC ~ Aβ + Aβ2 + covariates (age, sex, education level, intervention group (for Age-Well cohort only), visit) are reported for the PRC and PHC pathways Abbreviations: Aβ amyloid-β, FDR false discovery rate, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortex Significance levels: *p< 0.05; **p< 0.01; ***p< 0.00 Non-linear association between amyloid burden and fiber integrity in the PRC and PHC pathways across cohorts. Scatter plots illustrating the quadratic relationship between Aβ (centiloid) and fiber integrity for the PRC (red, top row) and PHC (blue, bottom row) pathways. Panel A shows results from the Age-Well cohort. Panel B shows results from the IMAP cohort. Each dot represents one observation, with grey lines connecting repeated measures from the same individual. Abbreviations: Aβ, amyloid-β; KFA, kurtosis fractional anisotropy, MK, mean kurtosis; PHC, parahippocampal cortex; PRC, perirhinal cortex To further characterize this relationship, we identified the tipping point at which the association reverses for each pathway and cohort (see Table 3). In Age-Well, tipping points ranged from 26.8 to 30.5 centiloid for the PRC pathway and from 35 to 38.1 centiloid for the PHC pathway, with only 10–18% of participants exceeding these thresholds. Comparable tipping points were observed in IMAP (39.5 and 22.3 centiloid for PRC and PHC pathways respectively), with similarly small proportions of individuals in the descending phase of the curve (10.3% and 17.6%).Table 3Tipping points of the inverted U-shaped association between amyloid burden and WM integritySC ~ Aβ + Aβ 2 + covariatesIntegrity measuresPRC SCPHC SCTipping pointN subjects > TP% of sample > TPTipping pointN subjects > TP% of sample > TPAge-WellMK30.51415.9351011.4KFA26.81618.238.11011.4IMAP39.5710.322.31217.6For each cohort, pathway and fiber integrity measures, the tipping point corresponds to the centiloid value at which the association between Aβ load and fiber integrity reverses from positive to negative. N subjects > TP and % of sample > TP indicate the number and proportion of participants whose amyloid load exceeded the tipping point, i.e., those in the descending phase of the curveAβ amyloid- Aβ, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortex, SC structural connectivity, TP tipping point Table 3Tipping points of the inverted U-shaped association between amyloid burden and WM integritySC ~ Aβ + Aβ 2 + covariatesIntegrity measuresPRC SCPHC SCTipping pointN subjects > TP% of sample > TPTipping pointN subjects > TP% of sample > TPAge-WellMK30.51415.9351011.4KFA26.81618.238.11011.4IMAP39.5710.322.31217.6For each cohort, pathway and fiber integrity measures, the tipping point corresponds to the centiloid value at which the association between Aβ load and fiber integrity reverses from positive to negative. N subjects > TP and % of sample > TP indicate the number and proportion of participants whose amyloid load exceeded the tipping point, i.e., those in the descending phase of the curveAβ amyloid- Aβ, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortex, SC structural connectivity, TP tipping point Tipping points of the inverted U-shaped association between amyloid burden and WM integrity For each cohort, pathway and fiber integrity measures, the tipping point corresponds to the centiloid value at which the association between Aβ load and fiber integrity reverses from positive to negative. N subjects > TP and % of sample > TP indicate the number and proportion of participants whose amyloid load exceeded the tipping point, i.e., those in the descending phase of the curve Aβ amyloid- Aβ, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortex, SC structural connectivity, TP tipping point Linear models were tested and none of them showed a significant association between SC measures and Aβ load (see Table S3). In addition, interaction analyses showed that plasma GFAP levels moderated the Aβ – structure relationship for the PHC pathway (see Table 4): among participants with low GFAP levels, an increase of Aβ load was associated with a better integrity of PHC fiber pathway. Conversely, in individuals with high GFAP levels, higher Aβ load was associated with poorer fiber integrity for all the PHC fiber pathway integrity measures (see Fig. 4 and Table S4 and S5). We also observed a positive association between Aβ load and GFAP level (see Figure S3).Table 4Interaction between amyloid burden and GFAP on fiber integritySC ~ Aβ *GFAP + covariatesIntegrity measuresPRC SCPHC SCF-valueβp-uncorrectedpFDRF-valueβp-uncorrectedpFDRMK0.77−0.1250.380.5513.23−0.212< 0.001***< 0.01**KFA0.39−0.0930.530.595.46−0.163< 0.05*< 0.05*Results from linear mixed-effects models of the form SC ~ Aβ * GFAP + covariates (age, sex, education level, intervention group, visit, BMI, GFR) are reported for the PRC and PHC pathways, for each fiber integrity measures (MK and KFA)Abbreviations: Aβ amyloid-β, BMI body mass index, FDR false discovery rate, GFAP glial fibrillary acidic protein, GFR glomerular filtration rate, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortex, SC structural connectivitySignificance levels: *p< 0.05; **p< 0.01; ***p< 0.001Fig. 4Moderation of GFAP levels on the amyloid burden – PHC fiber integrity association. Scatter plots illustrating the interaction between Aβ (centiloid) and GFAP levels on PHC fiber integrity (MK and KFA). Each dot represents one observation, color-coded according to the individual’s GFAP level (green: + 1SD; blue: mean; red: −1SD), with grey lines connecting repeated measures from the same individual. Abbreviations: Aβ, amyloid-β; GFAP, glial fibrillary acidic protein; KFA, kurtosis fractional anisotropy; MK, mean kurtosis; PHC, parahippocampal cortex; PRC, perirhinal cortex; SC, structural connectivity; SD, standard deviation Table 4Interaction between amyloid burden and GFAP on fiber integritySC ~ Aβ *GFAP + covariatesIntegrity measuresPRC SCPHC SCF-valueβp-uncorrectedpFDRF-valueβp-uncorrectedpFDRMK0.77−0.1250.380.5513.23−0.212< 0.001***< 0.01**KFA0.39−0.0930.530.595.46−0.163< 0.05*< 0.05*Results from linear mixed-effects models of the form SC ~ Aβ * GFAP + covariates (age, sex, education level, intervention group, visit, BMI, GFR) are reported for the PRC and PHC pathways, for each fiber integrity measures (MK and KFA)Abbreviations: Aβ amyloid-β, BMI body mass index, FDR false discovery rate, GFAP glial fibrillary acidic protein, GFR glomerular filtration rate, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortex, SC structural connectivitySignificance levels: *p< 0.05; **p< 0.01; ***p< 0.001 Interaction between amyloid burden and GFAP on fiber integrity Results from linear mixed-effects models of the form SC ~ Aβ * GFAP + covariates (age, sex, education level, intervention group, visit, BMI, GFR) are reported for the PRC and PHC pathways, for each fiber integrity measures (MK and KFA) Abbreviations: Aβ amyloid-β, BMI body mass index, FDR false discovery rate, GFAP glial fibrillary acidic protein, GFR glomerular filtration rate, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortex, SC structural connectivity Significance levels: *p< 0.05; **p< 0.01; ***p< 0.001 Moderation of GFAP levels on the amyloid burden – PHC fiber integrity association. Scatter plots illustrating the interaction between Aβ (centiloid) and GFAP levels on PHC fiber integrity (MK and KFA). Each dot represents one observation, color-coded according to the individual’s GFAP level (green: + 1SD; blue: mean; red: −1SD), with grey lines connecting repeated measures from the same individual. Abbreviations: Aβ, amyloid-β; GFAP, glial fibrillary acidic protein; KFA, kurtosis fractional anisotropy; MK, mean kurtosis; PHC, parahippocampal cortex; PRC, perirhinal cortex; SC, structural connectivity; SD, standard deviation Associations between fibers integrity and functional connectivity of MTL subregions, and their interactions with Aβ load We aimed to test whether the integrity of PRC/PHC fibers was associated with AT/PM functional connectivity and how Aβ load might moderate these associations. As reported in Table 5, the associations between FC and fiber integrity did not reach statistical significance in either cohort, for any of the networks studied (see Table S6).Table 5Associations between fiber integrity and FCFC ~ SC + covariatesIntegrity measuresPRC SC/AT FCPHC SC/PM FCF-valuep-uncorrectedp-FDRF-valuep-uncorrectedp-FDRAge-WellMK0.030.850.850.860.350.53KFA0.040.850.851.650.20.36IMAP0.880.380.730.390.540.80Results from linear mixed-effects models of the form FC ~ SC + covariates (age, sex, education level, intervention group, visit) are reported for PRC and PHC pathwaysAbbreviations: FC functional connectivity, FDR false discovery rate, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortex, SC structural connectivity Table 5Associations between fiber integrity and FCFC ~ SC + covariatesIntegrity measuresPRC SC/AT FCPHC SC/PM FCF-valuep-uncorrectedp-FDRF-valuep-uncorrectedp-FDRAge-WellMK0.030.850.850.860.350.53KFA0.040.850.851.650.20.36IMAP0.880.380.730.390.540.80Results from linear mixed-effects models of the form FC ~ SC + covariates (age, sex, education level, intervention group, visit) are reported for PRC and PHC pathwaysAbbreviations: FC functional connectivity, FDR false discovery rate, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortex, SC structural connectivity Associations between fiber integrity and FC Results from linear mixed-effects models of the form FC ~ SC + covariates (age, sex, education level, intervention group, visit) are reported for PRC and PHC pathways Abbreviations: FC functional connectivity, FDR false discovery rate, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortex, SC structural connectivity However, significant interactions between FC in the AT network and Aβ load on PRC SC were found, and this result was consistent across both cohorts. These results suggested that Aβ load modulates the effect of FC on integrity of PRC fiber pathway. Post-hoc analyses revealed that higher AT FC was associated with greater PRC fiber integrity in individuals with lower Aβ load, but with poorer fiber integrity in individuals with higher Aβ load (see Table 6; Fig. 5; and Table S7 and S8).Table 6Interaction between FC and amyloid burden on fiber integritySC ~ FC *Aβ + covariatesIntegrity measuresPRC SCPHC SCF-valueβp-uncorrectedpFDRF-valueβp-uncorrectedpFDRAge-WellMK6.08−0.12< 0.05*< 0.05*1.630.120.200.41KFA0.06−0.010.810.810.500.040.480.48IMAP5.33−0.20< 0.05*0.070.78−0.110.380.61Results from linear mixed-effects models of the form SC ~ FC * Aβ + covariates (age, sex, education level, intervention group, visit) are reported for the PRC and PHC pathways, for each fiber integrity measures (MK and KFA)Abbreviations: Aβ amyloid-β, FC functional connectivity, FDR false discovery rate, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortex, SC structural connectivitySignificance levels: *p< 0.05; **p< 0.01; ***p< 0.001Fig. 5Moderation of amyloid burden on the PRC fiber integrity – AT FC association. Scatter plots illustrating the interaction between AT FC and Aβ levels on PRC fiber integrity (MK). Panel A shows results from the Age-Well cohort. Panel B shows results from the IMAP cohort. Each dot represents one observation, color-coded according to the individual’s Aβ level (green: + 1SD; blue: mean; red: −1SD), with grey lines connecting repeated measures from the same individual. Abbreviations: Aβ, amyloid-β; FC, functional connectivity; MK, mean kurtosis; PRC, perirhinal cortex; SC, structural connectivity; SD, standard deviation Table 6Interaction between FC and amyloid burden on fiber integritySC ~ FC *Aβ + covariatesIntegrity measuresPRC SCPHC SCF-valueβp-uncorrectedpFDRF-valueβp-uncorrectedpFDRAge-WellMK6.08−0.12< 0.05*< 0.05*1.630.120.200.41KFA0.06−0.010.810.810.500.040.480.48IMAP5.33−0.20< 0.05*0.070.78−0.110.380.61Results from linear mixed-effects models of the form SC ~ FC * Aβ + covariates (age, sex, education level, intervention group, visit) are reported for the PRC and PHC pathways, for each fiber integrity measures (MK and KFA)Abbreviations: Aβ amyloid-β, FC functional connectivity, FDR false discovery rate, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortex, SC structural connectivitySignificance levels: *p< 0.05; **p< 0.01; ***p< 0.001 Interaction between FC and amyloid burden on fiber integrity Results from linear mixed-effects models of the form SC ~ FC * Aβ + covariates (age, sex, education level, intervention group, visit) are reported for the PRC and PHC pathways, for each fiber integrity measures (MK and KFA) Abbreviations: Aβ amyloid-β, FC functional connectivity, FDR false discovery rate, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortex, SC structural connectivity Significance levels: *p< 0.05; **p< 0.01; ***p< 0.001 Moderation of amyloid burden on the PRC fiber integrity – AT FC association. Scatter plots illustrating the interaction between AT FC and Aβ levels on PRC fiber integrity (MK). Panel A shows results from the Age-Well cohort. Panel B shows results from the IMAP cohort. Each dot represents one observation, color-coded according to the individual’s Aβ level (green: + 1SD; blue: mean; red: −1SD), with grey lines connecting repeated measures from the same individual. Abbreviations: Aβ, amyloid-β; FC, functional connectivity; MK, mean kurtosis; PRC, perirhinal cortex; SC, structural connectivity; SD, standard deviation No modulating effect of Aβ load on FC—SC was found (see Table S9). Associations between fibers integrity and cognitive performances. Finally, we aimed to go further testing whether the integrity of PRC/PHC fibers was associated with cognitive performances. As reported in Table 7, only the associations between MK (in both PRC and PHC pathways) and executive functioning composite score did reach statistical significance (Fig. 6).Table 7Associations between fiber integrity and cognitive performancesSC ~ cognition composite score + covariatesComposite scoresIntegrity measuresPRC SCPHC SCF-valuep-uncorrectedpFDRF-valuep-uncorrectedpFDRAttention/SpeedMK0.980.320.461.300.26KFA0.540.460.462.160.140.26Episodic memoryMK2.260.130.270.120.730.73KFA0.0010.980.980.190.660.73Executive functioningMK4.74< 0.05*0.066.88< 0.01**< 0.05*KFA0.150.700.700.800.370.37PACC-5MK0.710.40.570.090.770.77KFA0.320.570.570.180.670.77Results from linear mixed-effects models of the form SC ~ cog + covariates (age, sex, education level, intervention group, visit) are reported for PRC and PHC pathwaysAbbreviations: FDR false discovery rate, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortex, SC structural connectivityFig. 6Association of fiber integrity and executive functioning composite score. Scatter plots illustrating the association between PRC fiber integrity (MK) and executive functioning composite score. Each dot represents one observation with grey lines connecting repeated measures from the same individual. Abbreviations: MK, mean kurtosis; PRC, perirhinal cortex; SC, structural connectivity Table 7Associations between fiber integrity and cognitive performancesSC ~ cognition composite score + covariatesComposite scoresIntegrity measuresPRC SCPHC SCF-valuep-uncorrectedpFDRF-valuep-uncorrectedpFDRAttention/SpeedMK0.980.320.461.300.26KFA0.540.460.462.160.140.26Episodic memoryMK2.260.130.270.120.730.73KFA0.0010.980.980.190.660.73Executive functioningMK4.74< 0.05*0.066.88< 0.01**< 0.05*KFA0.150.700.700.800.370.37PACC-5MK0.710.40.570.090.770.77KFA0.320.570.570.180.670.77Results from linear mixed-effects models of the form SC ~ cog + covariates (age, sex, education level, intervention group, visit) are reported for PRC and PHC pathwaysAbbreviations: FDR false discovery rate, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortex, SC structural connectivity Associations between fiber integrity and cognitive performances Results from linear mixed-effects models of the form SC ~ cog + covariates (age, sex, education level, intervention group, visit) are reported for PRC and PHC pathways Abbreviations: FDR false discovery rate, KFA kurtosis fractional anisotropy, MK mean kurtosis, PHC parahippocampal cortex, PRC perirhinal cortex, SC structural connectivity Association of fiber integrity and executive functioning composite score. Scatter plots illustrating the association between PRC fiber integrity (MK) and executive functioning composite score. Each dot represents one observation with grey lines connecting repeated measures from the same individual. Abbreviations: MK, mean kurtosis; PRC, perirhinal cortex; SC, structural connectivity Discussion The AT and PM networks are functionally affected in AD and in aging [16], but structural pathways supporting these functional networks are not fully understood. In order to unravel the complex relationship between SC and FC, in particular in the MTL, this study aimed to i) identify the WM tracts supporting the AT and PM functional networks, ii) investigate how Aβ load and astroglial reactivity affect the integrity of these tracts, and iii) to explore the relationship between structural integrity and functional connectivity within these networks. First, we found that the PRC fiber pathway included the inferior parts of the hippocampal cingulum and the ILF as well as the thalamic radiation. In contrast, the PHC fiber pathway included the superior parts of the hippocampal cingulum and the ILF as well as the cingulum and the IFOF. While these results are generally consistent with the study by Chen et al. [21], which aimed to identify the fibers underlying the AT and PM functional networks, we note some differences. In their study, IFOF was identified as supporting the PM network in addition to the AT network. This discrepancy can be attributed to methodological differences, such as the choice of ROIs. Here, we selected a single ROI per network to avoid introducing bias from our hypotheses and to ensure that network definitions were not dependent on other predefined regions. In contrast, Chen et al. defined their networks using multiple ROIs, which constrained their analyses to regions they had predefined as part of the networks. Additionally, we confirmed the specific belonging of fibers to each network by conducting paired-sample t-tests, allowing us to precisely identify the fibers underlying the AT and PM functional networks. Moreover, our methodological framework enables us to distinguish clearly between the anterior and posterior regions, based on the definition of the PRC and PHC as ROIs, rather than using the anterior or posterior hippocampus as seeds, as is usually done. Second, our analyses confirmed an inverted U-shaped relationship between fiber integrity and Aβ burden, consistent with previous reports [26, 30, 31], and replicated here in an independent cohort, suggesting a tipping point where this association reverses. This tipping point estimated in Age-Well ranged from 26.8 to 30.5 centiloid for the PRC pathway and from 35 to 38.1 centiloid for the PHC pathway, with 10 to 18% of participants in the descending phase of the curve. Comparable values were observed in IMAP, supporting the reproducibility of these estimates across cohorts. These thresholds fall within the range commonly associated with early amyloid positivity (typically ~ 20–30 centiloid [53, 57]), suggesting that the reversal occurs at a relatively early and clinically relevant stage od amyloid accumulation. One explanation is neuroinflammation (reflected partially by astroglial reactivity) accompanying early amyloid accumulation: Aβ plaques trigger microglial and astroglial activation [34], which can increase diffusion restriction by expanding into extracellular space. This is supported by evidence linking neuroinflammation to increased gray matter volume in early AD [58, 59] and amyloid-related hypertrophy [60, 61]. However, we found that GFAP moderated this association only in PHC pathways, with high GFAP linked to decreased fiber integrity at high Aβ, which challenges this hypothesis. In this context, it is important to note that the astroglial response, as indicated by GFAP levels, occurs secondary to the microglial response, which represents the primary inflammatory mechanism [62, 63]. Thus, measuring the microglial response through sTREM2 (soluble Triggering Receptor Expressed on Myeloid Cells 2) levels, for example, might reveal a moderating effect that helps explain the phase during which SC is positively associated with Aβ burden. Alternatively, a recent study by Cai et al., (2025) [64] showed the presence of Aβ aggregates at the myelin-axon interface in rodent models, highlighting this region as a critical site for protein aggregation. We can hypothesize that the accumulation of Aβ deposits at this interface may increase the volume occupied by fiber tracts, thereby exacerbating physical constraints within the WM. Based on prior work, we hypothesized opposite SC–FC relationships across networks: a negative association between PRC fiber integrity and AT FC, and a positive association for the PM network, given reports of AT higher FC and PM lower FC in AD [16]. While no direct association emerged overall, we found a significant Aβ-dependent effect in the PRC pathway: in low Aβ individuals, higher AT FC related to greater PRC SC, whereas in high Aβ individuals, higher AT FC related to lower PRC SC. This pattern suggests that FC levels are associated with PRC fiber integrity in a Aβ-dependent manner. Notably, this relationship was asymmetric: whereas SC was significantly modulated by FC and Aβ load, the reverse model – FC predicted by SC and Aβ – did not reach significance, providing indirect evidence that FC levels may influence SC rather than merely reflecting a compensatory response to fiber loss. This interpretation is consistent with the AT network’s known vulnerability to early tau pathology and its proposed activity-dependent spread [65, 66]. Critically, the Aβ-dependent modulation of the SC-FC relationship in the PRC pathway was replicated in an independent cohort, strengthening confidence in this finding despite the modest sample size of both cohorts. In contrast, inflammatory effects — indexed by GFAP moderation — were specific to the PHC pathway, aligning with its sensitivity to aging and Aβ pathology [7, 67, 68]. This pattern highlights two potentially distinct mechanisms of vulnerability: PRC fibers could be specifically linked to functional connectivity levels and associated tau propagation dynamics, while PHC fibers might be more susceptible to amyloid-associated astroglial reactivity. Finally, among all cognitive domains examines, only executive functioning showed a significant association with MTL fiber integrity, specifically with MK in PHC pathways and as a trend in the PRC pathway. This finding is somewhat unexpected given the classical role of the MTL in episodic memory, yet converges with evidence that the PHC pathway – encompassing the posterior cingulum and the IFOF – anatomically bridges medial temporal, parietal and frontal regions known to support cognitive control. Indeed, the cingulum harbors prominent medial and dorsal prefrontal connections and its microstructural integrity has been linked to cognitive control processes, with parahippocampal connections showing independent contributions to executive performance in individuals with mild cognitive impairment [69]. Similarly, IFOF integrity has been associated with set-shifting performance as assessed by the TMT-B in healthy adults across the lifespan, consistent with the composition of our executive composite score [70, 71]. The absence of significant associations with episodic memory, attention or the PACC-5 may reflect the relatively preserved mnemonic functioning of this cognitively unimpaired sample. Our study has certain limitations, such as the particularly high level of education in the cohort, with an average of one year of post-secondary education, which is notably high for older adults. Thus, the study’s findings may not be generalizable to all populations. As education level is considered as a protective factor, including populations with lower education levels might allow us to capture even more pronounced effects. Another limitation of this study is the relatively small sample size, which may affect the generalizability of the findings: while our longitudinal design provided 176 observations, the sample size (n = 88) remains relatively modest for complex moderation analyses. Furthermore, while all analyses were conducted within a longitudinal mixed-effects framework incorporating both timepoints, our design included only two observations per participants, precluding the estimation of individual random slopes and therefore the explicit modelling rates of change [72]. As a result, conclusions regarding temporal dynamics and directionality remain inferential in nature. It is well established that underpowered interaction analyses can sometimes favor cross-over patterns rather than merely attenuating effects, and while our results align with the theoretical framework, they should be interpreted with caution and require replication in larger independent cohorts. Importantly, however, key findings – namely the non-linear SC-Aβ relationship and the Aβ-dependent SC-FC association – were successfully replicated in an independent cohort, providing initial evidence for their generalizability. Moreover, the absence of tau measures also limits the interpretation of the underlying pathological mechanisms, particularly regarding the potential contribution of tau pathology to longitudinal brain changes. Finally, although the data were collected as part of a randomized clinical trial, the present analyses were observational in nature and therefore do not allow causal conclusions to be drawn. However, a key strength is that all data were acquired at a single center, ensuring consistency across modalities. This is particularly valuable for studies integrating multimodal imaging techniques—such as rs-fMRI, DKI, and Aβ PET—with blood biomarkers, as it minimizes variability and enhances data quality. Conclusions In summary, our work provides a detailed characterization of MTL tracts within key large-scale networks implicated in AD and aging, elucidating both anatomical and functional underpinnings underlying their vulnerability. Our findings support the hypothesized non-linear association between Aβ load and structural integrity. Furthermore, while we expected direct SC-FC relationships, our results reveal a more complex Aβ-dependent modulation: in the AT systems, higher functional connectivity is associated with reduced structural integrity, specifically in individuals with high amyloid burden. In contrast, the PHC pathway appears more susceptible to astroglial reactivity influences, as indexed by GFAP. By delineating how specific alterations in structural and functional connectivity emerge in the preclinical stage, our findings advance the understanding of AD pathophysiology by providing insights into the complex associations between amyloid, neuroinflammation (reflected partially by astroglial reactivity) and connectivity. While these results offer a framework for understanding early disease progression, future studies incorporating tau measures and experimental designs are needed to establish the causal nature of these relationships. Supplementary Information Supplementary Material 1. Supplementary Material 1.