Introduction

Alzheimer’s disease (AD) pathology is characterized by phosphorylated-tau in the form of neurofibrillary tangles (NFTs) and extra-cellular deposits of amyloid-beta (Aβ) [1,2,3], which are thought to lead to neurodegeneration and cognitive decline. Measures of Aβ, tau and neurodegeneration (A/T/N) can be obtained using positron emission tomography (PET), cerebrospinal fluid (CSF) or structural magnetic resonance imaging (MRI), and are now being used to biologically define AD within a research framework [4]. Compared to Aβ, the accumulation of NFTs in particular is strongly correlated with neurodegeneration and cognitive decline [5,9, 10]. Eventually, the NFTs spread to other cortical regions of the brain. Because of the MTL’s early involvement, its subregions are expected to be among the most sensitive brain regions to early pathological changes in AD.

A growing number of studies have used in vivo structural MRI to derive measurements of subtle volumetric changes in the MTL caused by neurodegeneration [11, 24]. Most recently, Wisse et al. conducted a postmortem study which investigated the association between the thickness of MTL subregions in ex vivo MRI and semi-quantitative ratings of different neurodegenerative pathologies derived from histology samples taken from the contralateral hemisphere [

Methods

Specimen preparation and MRI

Brain hemispheres were obtained from twenty-nine donors; one specimen from the brain bank operated by the National Disease Research Interchange (NDRI), sixteen specimens from autopsies performed at the University of Pennsylvania Center for Neurodegenerative Disease Research (CNDR) and twelve specimens from the University of Castilla-La Mancha (UCLM) Human Neuroanatomy Laboratory (HNL) in Spain. Human brain specimens were obtained in accordance with the University of Pennsylvania Institutional Review Board guidelines, and the Ethical Committee of UCLM. Where possible, pre-consent during life and, in all cases, next-of-kin consent at death was given. CNDR hemispheres were fixed in 10% formalin solution for at least 30 days before extracting intact MTL blocks. HNL cases were fixed by perfusion with 4% paraformaldehyde through both carotid arteries. The blocks were then imaged on a Varian 9.4 T animal scanner at a 200 × 200 × 200 µm3 resolution. Details of the imaging protocol are provided in Additional file 1: Section 1.1 and Additional file 1: Fig. S7.

Histological processing and immunohistochemistry

Following MRI scanning, each of the specimens underwent histological processing for neuropathological examination and neuroanatomical analysis. In twenty-seven specimens, dense serial histology was performed. The specimens were cut into 2 cm blocks using custom molds that were 3D printed to fit each MTL specimen, cryoprotected and sectioned at 50 μm intervals in a sliding microtome coupled to a freezing unit (Microm, Heidelberg). For neuropathological diagnosis, we were interested in examining the anterior ERC at the mid-level of the amygdala, and the dentate gyrus (DG) and CA subfields at the mid-level of the body of the hippocampus. Therefore, for the remaining two specimens that did not undergo serial histology, two tissue blocks were cut at the level of the amygdala and hippocampal body, cryoprotected and sectioned into 50 μm sections. For each of the specimens, two adjacent sections were sampled at the mid-level of the amygdala and the hippocampal body for immunohistochemistry. In each case, the two sections were immunostained for tau and TDP-43 using anti-human PHF-Tau (monoclonal antibody (mAb), mouse, Thermo Scientific, Product Number MN1020, 1:500) and 66318-1-Ig Anti-phospho (409/410) TDP-43 (mAb, mouse, Proteintech, 1:350) respectively. Sections were then mounted on 7.5 cm × 5 cm slides and digitally scanned at 20× resolution.

Semi-quantitative neuropathology ratings

Regional thickness analysis was performed using histopathology ratings of tau and TDP-43 pathologies derived from the MTL both ipsilateral and contralateral to the one that was scanned. Semi-quantitative ratings for tau and TDP-43 pathology were obtained for the contralateral hemisphere by obtaining tissue samples at the time of autopsy and are available for a subset of twenty-eight donors (CNDR and HNL specimens) in three MTL regions routinely examined in the CNDR neuropathology evaluations [29].

For each of the specimens, the boundaries between MTL subregions cornu ammonis (CA) 1, CA2, CA3, DG, SUB, presubiculum, parasubiculum, hippocampal amygdala transition area (HATA), SRLM, ERC, BA35, BA36, area TE and the parahippocampal cortex (areas TF and TH) were identified on the basis of cytoarchitectural features in the Nissl stained sections following the anatomical rules presented in the Atlas of the Human Brain by Mai et al. [30]. Annotations were performed on each histology slice by the team of neuroanatomists at UCLM (the hippocampal subfields and ERC were annotated by M.M.A, E.A.P, M.M.R, M.M.L, C.R.P, S.C.S, J.C.D, M.C.P and F.M.R., supervised by R.In; R.In annotated the boundaries of the temporopolar cortex, PRC and PHC, and revised the annotations of the group). The neuroanatomists viewed scanned Nissl slides in an open-source web-based system (https://github.com/pyushkevich/histoannot) and used line drawings and text labels to annotate boundaries between adjacent anatomical regions. Following histology reconstruction and registration to MRI, the boundary annotations were overlaid on the co-registered MRI and histology images. Additionally, the manual segmentations of the whole MTL cortex and SRLM that were used to guide groupwise registration were registered to the histology images to inform the outer MTL boundary. The MRI, histology and MTL segmentation images were displayed side-by-side in ITK-SNAP [31] to facilitate manual tracing of the subfield segmentations in 3D MRI space, which was performed by S.L. with supervision from L.E.M.W and R.In (Additional file 1: Fig. S9).

For each of the eleven specimens, the completed segmentations were then mapped to the MRI atlas using the deformable transformations generated by the groupwise registration pipeline. Note that for each specimen, small gaps in the segmentation may exist between blocks. A consensus segmentation of the MRI atlas was obtained by application of voxel-wise majority voting among the eleven segmentations with slight regularization by a Markov Random Field prior. More details on this approach are provided in the supplemental information, Section S1.4.5 of Adler et al. [1: Section 1.4). To test the effects of tau pathology on regional thickness, we fit a general linear model (GLM) at each vertex on both the MTL and SRLM surfaces with the average rating of tau pathology as the independent variable, thickness as the dependent variable, and age and TDP-43 rating as covariates. More details on the statistical analysis are provided in Additional file 1: Section 1.5.

Results

Demographics

The specimens included in this study contain varying neuropathological diagnoses, including AD neuropathologic change and neuropathological diagnoses such as argyrophilic grain disease (AGD), frontotemporal lobar degeneration (FTLD) with TDP-43 inclusions, cerebrovascular disease and Lewy body disease. Figure 1 summarizes the demographic and neuropathology data for this brain donor cohort. The average ipsilateral tau and TDP-43 ratings across the twenty-nine specimens are 1.50 ± 0.97 and 0.54 ± 0.82 respectively. The average age is 74 years (range 44–93 years). Additional file 1: Table S2 provides more detailed demographic data and pathology information for each specimen.

Fig. 1
figure 1

Demographic and diagnostic summaries for the twenty-nine brain donors. The tau and TDP-43 pathology ratings refer to the average rating computed from measurements sampled at three medial temporal lobe locations (entorhinal cortex at the mid-level of the amygdala and subiculum/cornu ammonis and dentate gyrus at the mid-level of the hippocampus). Dashed lines are used to indicate the mean value across specimens

Full size image

Computational atlas of the medial temporal lobe

Figure 2 shows the MRI atlas of the MTL constructed from twenty-nine ex vivo specimens as a synthetic “average” MR image and a consensus MTL subregion segmentation derived from serial histology in eleven specimens. The atlas construction pipeline achieves excellent groupwise alignment between ex vivo MRI scans and captures the average shape of the MTL (Additional file 1: Section 2.1, Additional file 1: Fig. S2, Additional file 1: Fig. S3 and Additional file 1: Fig. S4). Following groupwise registration, each specimen has a pointwise spatial correspondence to this atlas. This correspondence is limited to the region of the MTL cortex which was semi-automatically segmented in each specimen. Figure 3 provides a visualization of the quality of the registration between individual specimen images and the final atlas. In some specimens, the PRC, which includes BA35, was particularly challenging to register due to significant anatomical variability in cortical folding and branching patterns [32, 33]. Overall, the warped specimens look similar to each other following registration, although some minor mis-registrations remain. Quantitative and visual evaluation of atlas quality at different stages of atlas construction, and comparisons with an alternative atlas-building strategy are presented in Additional file 1: Section 2.1.

Fig. 2
figure 2

Computational atlas of the medial temporal lobe (MTL) constructed by groupwise registration of the magnetic resonance image (MRI) scans of twenty-nine ex vivo specimens. Three coronal sections are shown ordered from anterior (ant) to posterior (post), indicated as I, II and III, as well as a sagittal and axial section through the MTL. For each section, the “average” MRI is shown with and without the consensus MTL subregion segmentation derived from serial histology in eleven specimens. In the top right, a 3D reconstruction of the MTL atlas is shown along with a 3D brain rendering indicating the location of the MTL within the brain. (med medial, lat lateral, sup superior, inf inferior, SUB subiculum, SRLM stratum radiatum lacunosum moleculare, CA cornu ammonis, DG dentate gyrus, HATA hippocampal amygdala transition area, ERC entorhinal cortex, BA Brodmann Area)

Full size image
Fig. 3
figure 3

Coronal view of the MRI scans of each of the twenty-nine specimens warped into the space of the MRI atlas. The corresponding atlas image is outlined in blue, in the bottom-right corner. The more similar the warped images are, the better the atlas quality. The dashed blue circles point out examples where the perirhinal cortex (region surrounding the collateral sulcus) was particularly challenging to register due to significant variability in cortical folding patterns. The color bar at the bottom of each image indicates the average neurofibrillary tangle (NFT) rating for that specimen. Yellow represents a rating of 0 (no/rare pathology) and red represents a rating of 3 (severe Alzheimer’s disease)

Full size image

Effects of tau pathology on regional thickness

Figure 4 shows the results of the pointwise thickness analysis performed on the MTL and SRLM examining the regional effects of tau pathology on cortical thickness using pathology ratings derived from the MTL ipsilateral to the thickness measures. Cases CNDR12 (44 y.o.) and HNL01 (45 y.o.) were excluded from the thickness analyses since these younger cases are outliers in terms of age and including them skews age effects, thereby dampening the associations between pathology and thickness. In this work we only consider the effects of tau and TDP-43 pathology on MTL structure since recent work studying the contribution of mixed pathology to MTL atrophy in AD showed no clear relationship between neurodegeneration and either Aβ or \(\alpha\)-synuclein [34, 9, 29]. These NFT density maps are generated using a weakly supervised deep learning algorithm, trained to specifically detect tangles and pre-tangles on AT8-stained sections and have been shown to correlate strongly with manual NFT counts [29]. Three cases with a primary FTLD-Tau or AGD diagnosis (CNDR01, CNDR06 and CNDR07) were excluded from the analysis since they likely contain 4R-tau inclusions which can be difficult to distinguish from AD-related NFTs. In the remaining twelve cases, we examined the correlation between the average NFT burden computed within each hotspot and the median thickness of each hotspot, and observe very strong correlations, although the hotspot in the ERC does not reach significance (Additional file 1: Fig. S12). These results are weakened when we include age in the model. This is likely due to the small sample size and the fact that age and tau are significantly correlated (average R = 0.61 across the three hotspots). Despite the small dataset, these results are encouraging and suggest that NFTs are indeed playing a role in driving neurodegeneration in the ERC, SUB and SRLM hotspots. As we expand our dataset, collect more quantitative histology data and develop automated methods to extract quantitative measures of the different types of tau and TDP-43 pathology, in future work we will be able to better overcome the above-mentioned limitations and validate the relationship between NFTs and thickness.

Overall, our findings provide a more refined understanding of how tau pathology is associated with cortical thinning within the MTL and motivate further characterization of the MTL in AD using detailed ex vivo MRI analysis. The clusters identified from the tau thickness analyses indicate granular MTL regions where in vivo measures of neurodegeneration are expected to be strongly associated with tau pathology. In an exploratory analysis, we attempted to use the ex vivo derived, AD-specific hotspots (Fig. 4B) as biomarkers in a longitudinal analysis of ADNI data (not shown). While the hotspots did not show statistical effects in a group comparison of Aβ negative, tau negative (A−T−) and Aβ positive, tau positive (A+T+) patient groups with mild cognitive impairment, BA35 achieved the strongest group discrimination. The lack of statistically significant group differences with the hotspots may be due to the small ex vivo sample size, clinical heterogeneity in our dataset and difficulties in accurately aligning BA35 in the atlas.

Constructing a probabilistic atlas of the MTL has far-reaching applications in AD research beyond the work shown here. Perhaps future analyses leveraging this technology in a patient cohort more consistent with the one we would expect to encounter in an AD clinical trial would result in hotspots that are more sensitive to longitudinal change in the presence of NFT pathology, potentially enabling the development of neurodegeneration biomarkers which are more effective during early AD clinical trials. Furthermore, in future work, 3D quantitative maps of NFT density derived from serial histology imaging will be mapped into ex vivo atlas space to generate a comprehensive probabilistic description of the progression of NFT pathology at each Braak stage [29, 48, 49]. This will allow us to describe NFT topography during the different stages of AD in more detail than current descriptions, which are in 2D and based on selective sampling of the MTL [1, 2]. Additionally, in a future version of the atlas, cytoarchitecture-guided anatomical labels of MTL subregions will be included from a larger number of specimens. Such an atlas would reflect anatomical ground truth and can be used to inform in vivo MRI protocols for segmentation of MTL subregions, thereby improving the accuracy of MRI biomarkers derived from these subregions.