Correlation between in vivo ¹⁸F‐flortaucipir PET and whole‐brain postmortem histological tau signals in Alzheimer's disease

Abstract

INTRODUCTION

Clinicopathological studies offer crucial interpretations of ¹⁸F‐flortaucipir (FTP) tau positron emission tomography (PET) signal but are limited by methods. We leveraged whole‐brain, quantitative immunohistochemical (IHC) Alzheimer's Disease (AD) tau density maps to comprehensively evaluate the FTP tracer.

METHODS

We generated IHC maps for three AD‐tau antibodies—AT8, AT100, and MC1—using two human brains histologically staged at Braak IV and VI. FTP‐PET scans were acquired 6 and 10 weeks prior to death. Using region‐wise and voxelwise methods, we correlated FTP‐PET with IHC signals.

RESULTS

Only AT8 (p‐tau Ser202/Thr205; neuronal and neuritic tau pathology) showed a notable correlation with FTP standardized uptake value ratios (SUVRs) in the Braak VI case (Spearman's rank correlation coefficient [r_s] = 0.461, p < 0.001). FTP SUVRs failed to capture medial temporal lobe (MTL) tau burden, whereas neocortical regions showed lower IHC burden but more variability in FTP uptake.

DISCUSSION

Although FTP signals correlate well with AT8‐positive tau in the more advanced case, they underestimate the severity of MTL burden, potentially confounding assessments of tau‐targeted therapies.

Highlights

• Whole‐brain tau histology was correlated with ante‐mortem ^{1 8}F‐flortaucipir positron emission tomography (PET)

• AT8‐positive tau showed the strongest correlation with flortaucipir PET signal

• Flortaucipir PET underestimated tau burden in medial temporal lobe regions

• Tau–PET relationships varied by brain region and disease stage in Alzheimer's disease

1. BACKGROUND

More than 30 million people are estimated to have symptomatic Alzheimer's disease (AD) worldwide, and possibly 10 times more in preclinical stages. ¹ Although a definitive diagnosis still relies on postmortem brain tissue examination, advances in biofluid‐based and neuroimage‐based biomarkers enable detection of AD pathology in living patients, especially when brain pathology reaches moderate to severe levels. ² , ³

AD is characterized by the accumulation of amyloid beta (Aβ) plaques and neurofibrillary tangles (NFTs) formed by misfolded proteins in the brain. Plaques first appear in the neocortex, then the allocortex, basal ganglia, midbrain, and cerebellum, whereas NFTs appear first in the neuromodulatory subcortical system, followed by the allocortex and neocortex. ⁴ , ⁵ , ⁶ This relatively stereotyped topographical progression of NFTs forms the basis for Braak staging of neurofibrillary pathology. ⁷ In vivo detection of regional tau involvement is therefore an appealing biomarker for disease staging and evaluating therapeutic efficacy.

Positron emission tomography (PET) ligands have enabled in vivo visualization of AD‐tau pathology. Specifically, [¹⁸F] Flortaucipir (FTP), a widely used PET ligand, selectively binds to paired helical filaments (PHF‐tau) in vitro. ⁸ , ⁹ In vivo, FTP signals correlate with longitudinal atrophy ¹⁰ , ¹¹ and cognitive decline. ¹² Several studies have explored correlations between FTP‐PET signals and postmortem immunohistochemistry (IHC), with important insights. However, such studies faced limitations like long intervals between PET and autopsy, use of semi‐quantitative approaches or restricted regions of interest (ROIs), and lack of precise spatial mapping. ¹³ , ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ Key questions remain: which AD‐tau species best correlates with FTP in vivo, whether regional sensitivity differs, and whether the IHC–FTP relationship is linear across tau burden.

To enhance the clinical interpretation of the FTP‐PET tracer, we conducted a quantitative autopsy‐PET correlation study to investigate the relationship between FTP‐PET standardized uptake value ratios (SUVRs) and histological measures of AD‐tau density. We processed two human brains, histologically staged at Braak stages IV and VI, from individuals who underwent FTP‐PET, 6 and 10 weeks before death, using a novel semi‐automatic high‐performance computing pipeline. ²⁰ We generated three‐dimensional (3D) IHC maps of tau using three antibodies that label overlapping but non‐identical aspects of pathological tau: AT8 (phospho‐tau Ser202/Thr205), which robustly labels neuronal and neuritic tau pathology across disease stages ²¹ , ²² ; AT100 (phospho‐tau Thr212/Ser214), which labels tangles and neuropil threads in a phosphorylation‐dependent PHF‐like epitope ²³ ; and MC1, which recognizes a disease‐associated misfolded tau conformational epitope (with relatively greater sensitivity to somatodendritic/NFT‐like pathology than neuropil threads) ²² , ²⁴ (Figure S1). IHC heatmaps and FTP‐PET scans were co‐registered to each subject's native T1 magnetic resonance imaging (MRI) obtained in cranio postmortem, enabling direct comparisons between histology and PET using ROI or whole‐brain voxelwise methods. The primary aims of this study were to: (1) understand relationships between FTP SUVRs and IHC signals in various brain regions; (2) determine which AD‐tau conformation best correlates with FTP uptake to inform the biological basis of the signal; and (3) assess the relation between FTP and IHC signals across varying histological tau burdens, especially when the burden is low. We predicted a significant, positive correlation between FTP SUVRs and IHC tau signals in the case with more extensive tau pathology (Braak VI), aligning with previous studies, but with region‐specific variation in correlation strengths. ¹³ , ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ We also hypothesized that MC1 would correlate worse with FTP SUVRs than AT8 and AT100 due to its lower sensitivity to diffuse neuritic pathology.

2. METHODS

We conducted a cross‐sectional clinicopathological correlation study in which each participant contributed a single ante‐mortem ^{1 8}F‐flortaucipir PET scan paired with a single postmortem quantitative neuropathological assessment. Participants were recruited between November 2016 and June 2017 as part of a clinical trial (ClinicalTrials.gov NCT02350634). Ante‐mortem tau PET imaging was acquired within 10 weeks of death (Case 1: 6 weeks; Case 2: 10 weeks), followed by postmortem MRI and brain autopsy. Two participants met all predefined inclusion criteria during the recruitment period, including availability of high‐quality ante‐mortem flortaucipir PET imaging within 10 weeks of death and complete postmortem tissue suitable for whole‐brain quantitative histological mapping. The final sample includes all eligible participants during recruitment, reflecting the rarity of cases meeting inclusion criteria. Although the number of subjects was small, the whole‐brain design enabled thousands of voxelwise PET–histology comparison points across cortical and subcortical regions, providing substantial analytic depth for within‐brain association analyses. Subjects’ consent was obtained according to the Declaration of Helsinki and the study has been approved by the ethical committee of the University of California San Francisco.

2.1. Image acquisition

Both FTP‐PET scans were done on a PET/CT GE Discovery 690 scanner. Both participants received an intravenous bolus injection of 370 MBq (10 mCi) of ¹⁸F‐Flortaucipir followed by 20 min of continuous FTP‐PET imaging beginning 80 min post‐injection. Four 5‐min frames were acquired and reconstructed with the Vue Point HD algorithm with 3 iterations and 24 subsets. Finally, a 3.0 mm post reconstruction smoothing filter was applied, achieving a final 6 mm full width at half maximum (FWHM) spatial resolution. The time delay between PET acquisition and time of death was 6 weeks for Case 1 and 10 weeks for Case 2. Both participants underwent postmortem T1‐weighted 3‐Tesla (3T) structural MRI (sMRI) acquisition within 10 hours of death, followed by brain procurement (autopsy). Postmortem sMRI for Case 1 was done on a 3T Siemens Skyra MRI system using a 3D magnetization‐prepared rapid gradient echo  T1‐weighted sequence: repetition time (TR) = 2300 ms, echo time (TE) = 2.98 ms, inversion time (TI) = 900 ms, 176 sagittal slices, within plane field of view (FOV) = 256 × 240 mm², voxel size = 1 × 1 × 1 mm³, flip angle = 9°, bandwidth = 240 Hz/pix. Postmortem sMRI for Case 2 was acquired on a GE Discovery 3T MR750 using a 3D spoiled gradient echo T1‐weighted sequence: TI = 400 ms, 200 sagittal slices, within plane FOV = 256 × 256 mm², voxel size = 1 × 1 × 1 mm³, flip angle = 11°, bandwidth = 31.25 Hz/pix.

RESEARCH IN CONTEXT

1. Systematic review: We reviewed prior clinicopathological studies comparing ^{1 8}F‐Flortaucipir (FTP) positron emission tomography (PET) with postmortem tau pathology, focusing on studies indexed in PubMed and reference lists of key reviews. Existing literature shows variable correlations between FTP PET and tau pathology, often limited by long PET–autopsy intervals, semi‐quantitative histology, restricted regional sampling, and limited assessment of tau conformations.

2. Interpretation: By integrating short‐interval ante‐mortem FTP PET with whole‐brain, quantitative, three‐dimensional tau histology using multiple antibodies, this study demonstrates that FTP PET preferentially reflects AT8‐positive tau pathology and systematically underestimates tau burden in medial temporal lobe regions, particularly at higher pathological loads. These findings refine the biological interpretation and limitations of FTP PET.

3. Future directions: Future studies should examine larger, diverse cohorts, evaluate newer tau PET tracers across tau conformations, and clarify regional and nonlinear PET–pathology relationships to improve biomarker accuracy for disease staging and therapeutic monitoring.

2.2. MRI and PET processing

Postmortem sMRI studies were segmented and parcellated into ROIs with FreeSurfer v.7.1.1 based on the Desikan–Killiany atlas. ²⁵ FTP‐PET scans were co‐registered to their corresponding postmortem sMRI with Statistical Parametric Mapping 12 (SPM12). Inferior cerebellar gray matter masks were created by reverse normalizing the SUIT cerebellar template to each subject's native MRI space with SPM12. SUVRs were calculated for FTP‐PET scans by referencing the mean tracer uptake in the inferior cerebellar gray matter. ²⁶ SUVRs from non‐zero PET voxels were averaged within Desikan–Killiany atlas ROIs to create ROI‐wise quantifications or mapped for whole‐brain voxelwise analyses. We additionally corrected inferior cerebellar gray normalized FTP scans for partial‐volume effect using a Geometric Transfer Matrix approach for ROI‐wise quantifications. ²⁶ , ²⁷ Partial volume corrected (PVC) SUVRs were used in secondary analyses.

2.3. Histological processing and tau heatmap generation

We generated 3D IHC tau density heatmaps from postmortem tissue specimens and co‐registered them to corresponding native postmortem T1‐weighted sMRI space. This pipeline was described in detail previously. ²⁰ In sum, each whole postmortem brain was fixed in buffered formalin, embedded in celloidin, and sectioned into sets of four 160‐µm‐thick serial full coronal slides. A few tissue samples were cut off from the brain slides for neuropathological assessment of AD neuropathologic changes (ADNCs), synucleinopathies, TAR DNA binding protein 43 (TDP‐43) proteinopathies, primary tauopathies, and cerebrovascular disease, using the UCSF Neurodegenerative Disease Brain Bank protocol. Braak neurofibrillary tangle stage was determined in the diagnostic neuropathological assessment using AT8 immunohistochemistry (pSer202/pThr205) following standard Braak criteria. ⁷ A deep learning neural network, IHCNet, was used to segment tau signals from the background in all digitized immune‐stained slides. IHCNet generated probability maps of tau exclusions and was trained independently for each antibody. The performance of each antibody‐specific neural network was validated and tested with the receiver‐operating characteristic (ROC) curve and area under the curve (AUC). ²⁰ Immunostained slides were partitioned into 1 µm² units. We calculated the percent of the area with above‐threshold (probability >0.5) histological signal, with 100% being complete coverage. Thus, the value within each unit represented the area occupied by IHC signal rather than the intensity of signal. This approach avoids artifacts introduced by variable chromogen exposure times (e.g., DAB), which could impact signal intensity. Heatmaps were then normalized to mitigate suboptimal staining and batch variations, rescaled to 10% resolution, stacked, and co‐registered to the reconstructed 3D postmortem brains (blockface images). The blockface images were co‐registered to the corresponding postmortem MRI volumes with the Advanced Normalization Tools (ANTs), first through an affine transformation and then the Symmetric Normalization (SyN) model for diffeomorphic transformation. The same transformation matrices were applied to the 3D IHC heatmaps. We computed Dice coefficients with lateral ventricle masks to evaluate the goodness of co‐registration. For this study, IHC heatmaps were additionally smoothed with a 5.81 mm isotropic gaussian kernel to achieve a final resolution of ≈6 mm FWHM, matching that of FTP‐PET images and allowing for better comparison between these two different imaging modalities. After co‐registration and smoothing, IHC heatmaps, postmortem sMRI studies, and FTP‐PET scans all had a voxel size of 1 mm³. ROI‐wise histological signals were calculated for FreeSurfer‐defined ROIs by summing the signals in each ROI and then normalizing by the volume of that ROI. Voxelwise values were taken as is at each voxel because voxel sizes and coordinates matched those of FTP‐PET scans after co‐registration to postmortem sMRI studies. After interpolations and smoothing, voxelwise histological values no longer precisely represent percentage of area covered; however, because all heatmaps are processed in a uniform way, the relative burden is still comparable between heatmaps.

2.4. Regions of interest and lobar groupings

Twenty‐nine FreeSurfer‐defined ROIs were excluded for both cases due to their irrelevance to the study (i.e., brainstem due to the limited spatial resolution of PET; white matter because they lack tau deposition in AD; and ventricles). Moreover, 10 ROIs in the left and right basal ganglia and choroid plexus were excluded from both cases for being known off‐target binding regions for FTP. ²⁸ Additional ROIs were excluded on a case‐ and antibody‐specific basis. Because Case 1 is a Braak IV case lacking tau deposits in most of the neocortex as confirmed by microscopy, only 368 of 837 brain slides were fully stained at an interval of 16 (Slides 280 to 648). As a result, 21 ROIs in Case 1, mainly in the left and right frontal and occipital regions, had a histological signal of 0 for AT8, AT100, and/or MC1; these regions were excluded from the study. In the end, 51 ROIs remained for Case 1. For Case 2, immunostaining quality issues and severe atrophy caused 4 ROIs to be excluded from the AT8 and MC1 analyses, and 10 ROIs for AT100 analyses. For Case 2, 68 ROIs were left for AT8 and MC1, and 62 ROIs for AT100. Lobar groupings were created based on Desikan et al. ²⁵ We further divided the medial temporal lobe (MTL) into cortical regions (entorhinal cortex, fusiform gyrus, and temporal pole) and limbic regions (hippocampus, amygdala, and parahippocampus) to discriminate between different structures in the MTL. A complete list of ROIs and their lobar assignments included in analyses can be found in Table S1.

2.5. Statistical analysis

All analyses were done in R (version 4.5.0). Two‐way analyses of variance (ANOVAs) were used to determine if ROI‐wise histological signals and FTP SUVRs differed between cases and lobar regions. The p‐values were two‐tailed, and p < 0.05 was deemed significant. Post hoc pairwise comparisons were conducted using estimated marginal means, and p‐values were adjusted using the Tukey method. For ROI‐wise analyses, we fitted Gaussian Mixture Models (GMMs) using the R Mclust library on each IHC data distribution to derive antibody‐specific IHC thresholds and identify sub‐threshold ROIs. We then computed Spearman's rank correlation coefficients between averaged IHC and FTP signals in above‐threshold ROIs. The p‐values < 0.05 were considered statistically significant. Similarly, for voxelwise analyses, we correlated FTP and IHC signals between non‐zero voxels within selected MTL and neocortical ROIs using Spearman's rank correlation. Hexagonal binned plots were used to visualize the joint distribution of FTP SUVRs and IHC signals for AT8, AT100, and MC1. We additionally fitted a non‐linear model to voxelwise signals with a self‐starting, three‐parameter logistic function (SSlogis in R) to evaluate the relationship between FTP and IHC signals across varying histological burden. We selected four bilateral, FreeSurfer‐defined MTL ROIs that are sites of early tau accumulation and often used in studying preclinical AD: entorhinal cortex, hippocampus, parahippocampal gyrus, and amygdala. Note that FreeSurfer‐defined parahippocampus is slightly different from the anatomic label defined through microscopic examination of the brain tissue. However, we used the FreeSurfer‐defined parahippocampal gyrus as it is the convention for brain image processing. We additionally chose four bilateral, FreeSurfer‐defined neocortical ROIs that start accumulating tau later as AD progresses for Case 2 (Braak VI): the inferior temporal gyrus, the posterior cingulate cortex, the lateral occipital cortex, and the precentral gyrus. Potential sources of bias included off‐target binding, partial‐volume effects, and registration error; we mitigated these via predefined ROI exclusions, partial‐volume correction sensitivity analyses, and voxelwise comparisons.

3. RESULTS

3.1. Participant characteristics

Table 1 summarizes the demographics, clinical and neuropathological diagnoses, existence of comorbid pathology, and volume‐weighted FTP‐PET quantifications in the bilateral entorhinal cortex and temporal metaROI ²⁹ for each participant. In brief, Case 1 was an 88‐year‐old cognitively normal donor. In vivo FTP‐PET quantification in the temporal metaROI was 1.213 SUVR. Postmortem examination of the brain showed moderate AD pathology; the brain was staged at Braak IV with intermediate ADNC (A2B2C1). The second brain was from a 76‐year‐old donor diagnosed with dementia with severe AD pathology postmortem, staged at Braak VI with a high ADNC (A3B3C3). FTP quantification in the temporal metaROI was 2.058 SUVR.

TABLE 1.

Brain donor characteristics and PET quantification. Demographic information, clinical diagnosis, neuropathological staging, comorbid pathology, and ^{1 8}F‐flortaucipir PET SUVRs in entorhinal cortex and temporal metaROI for both cases.

	Case 1	Case 2
Age, years	88	76
Sex	F	M
Clinical diagnosis	Cognitively normal	AD dementia
CDR	0	3
Interval PET to autopsy	6 weeks	10 weeks
Postmortem evaluation
Amyloid status	Positive	Positive
Braak stage	4	6
AD neuropathological change	A2B2C1	A3B3C3
TDP‐43 proteinopathy	Absent	Absent
Alpha‐synuclein pathology	Absent	Absent
PET quantification
Entorhinal cortex SUVR	1.184	1.717
Temporal MetaROI SUVR	1.213	2.058

Abbreviations: AD, Alzheimer's disease; CDR, Clinical Dementia Rating; F, female; M, male; SUVR, standardized uptake value ratio; PET, positron emission tomography.

3.2. Neuropathology and image processing

The two brains were processed using our previously published histological and computational pipeline to create 3D whole‐brain tau density heatmaps. ²⁰ For each brain, three tau density heatmaps were generated from postmortem slices immunostained for different forms of AD‐tau: p‐tau Ser202/Thr205 (AT8), p‐tau Th212/Ser214 (AT100), and conformational tau (MC1). The ROC‐AUC evaluations for the deep learning neural networks (IHCNets) have been published previously for each antibody. ²⁰ For each subject, heatmap‐to‐MRI co‐registration accuracy was evaluated by Dice coefficients of lateral ventricle masks, ranging from 0.8185 (± 0.1070) to 0.9112 (± 0.0405). ²⁰ Co‐registration results for postmortem MRI, FTP‐PET, and histological tau heatmaps were visualized in Figure 1.

FIGURE 1.

Visualization of image co‐registration. (A) Co‐registration of postmortem magnetic resonance imaging with ^{1 8}F‐flortaucipir (FTP) positron emission tomography and immunohistochemical (IHC) tau heatmaps in Case 1 (Braak IV) at three representative voxel locations. (B) Corresponding overlays for Case 2 (Braak VI). FTP standardized uptake value ratios (SUVRs) are scaled from 0–2.7 and IHC heatmaps from 0–1.7 to facilitate visual comparison. Case 2 shows higher signal intensity and more widespread neocortical involvement than Case 1.

3.3. ROI‐wise histological and FTP signal distributions

IHC signals and inferior cerebellar normalized FTP SUVRs were averaged within each MRI‐defined Desikan–Killiany ROI (Figure 2). Table S1 provides a full list of ROIs analyzed in Case 1 (n = 51) and Case 2 (n = 68 for AT8 and MCI; n = 62 for AT100). Two‐way ANOVAs was used to determine differences in FTP and IHC signal intensities between cases and lobar regions as defined by Desikan et al. ²⁵ For all 3 IHC signals and FTP SUVRs, ANOVAs showed significant main effect of case (AT8: F(1,103) = 95.2; AT100: F(1,97) = 69.6; MC1: F(1,103) = 24.4; FTP: F(1,107) = 210.2; p < 0.001 for all) and of lobar regions (AT8: F(7,103) = 18.7; AT100: F(7,97) = 10.8; MC1: F(7,103) = 19.4; FTP: F(7,107) = 13.1; p < 0.001 for all). Post hoc pairwise comparisons suggested that Case 2 had higher IHC signals and FTP SUVRs than Case 1; limbic regions of the MTL had the highest IHC signals in all three stains, but occipital regions had the highest FTP SUVRs. Significant interaction effect of case and region was also found for AT8, MC1, and FTP signals (AT8: F(7,103) = 2.37, p = 0.027; MC1: F(7,103) = 3.35, p = 0.003; FTP: F(7,107) = 4.09, p < 0.001), but not AT100 (F(7,97) = 1.67). For AT8, Case 2 had significantly higher signals than Case 1 in all regions except for the MTL; for MC1 signals, Case 1 and Case 2 significantly differed in frontal, cingulate, and insular cortices; and for FTP SUVRs, only the insula cortex lacked significant difference between Cases 1 and 2. FTP SUVR distributions corrected for partial‐volume effects are available in Figure S2.

FIGURE 2.

Distributions of ¹⁸F‐Flortaucipir (FTP) positron emission tomography signal and histological tau burden across regions. Mirrored grouped bar plots show region‐wise FTP standardized uptake value ratios and immunohistochemistry tau burden (AT8, AT100, MC1) for the left and right hemispheres. Case 1 (Braak IV) shows elevated tau mainly in medial temporal lobe regions, whereas Case 2 (Braak VI) shows widespread neocortical involvement. Two‐way analyses of variance (ANOVAs) revealed significant effects of case and region for all measures.

3.4. ROI‐wise correlation between histological signal and FTP SUVR

As expected for a Braak stage IV case, histological tau burden outside the MTL was minimal in Case 1 (Figure 3A). However, many non‐MTL ROIs showed variable FTP‐PET signal (see left side of graphs in Figure 3A) that was not associated with IHC signal. Because this was likely due to the known off‐target binding of FTP‐PET, ²⁸ , ³⁰ , ³¹ , ³² we examined only ROIs with higher IHC binding by fitting GMMs to IHC signal distributions and calculating IHC thresholds using mean + 2.5 SD of the lowest gaussian in a two‐peak solution with unequal variance. Resulting thresholds were: 0.094, 0.053, and 0.051 for AT8, AT100, and MC1, respectively. Subthreshold ROIs were excluded from analysis in Case 1. The remaining ROIs were almost entirely within the MTL, with a few ROIs in cingulate and lateral temporal lobe. The range of FTP SUVRs was narrow (1.007 ≤ FTP SUVRs ≤ 1.257). For Case 1, there were no significant correlations between IHC signal and FTP SUVRs for any AD‐tau antibodies (Spearman's rank correlations AT8: n = 19 ROIs, r_s  = −0.154, p = 0.527; AT100: n = 16 ROIs, r_s  = 0.082, p = 0.763; MC1: n = 15 ROIs, r_s  = −0.186, p = 0.507). Results were similar with PVC (Figure S3A). No additional Case 2 ROIs were excluded due to GMM thresholding, as the best fit GMMs, determined by Bayesian Information Criterion (BIC), were mostly single‐peak solutions. AT8 signal showed a significant direct correlation with FTP SUVRs (n = 68 ROIs, r_s = 0.424, p < 0.001). AT100‐FTP and MC1‐FTP correlations both remained non‐significant (AT100: n = 62 ROIs r_s = 0.163, p = 0.204; MC1: n = 68 ROIs r_s = −0.195, p = 0.111). After correcting FTP SUVR for partial‐volume effects, the correlation between MC1 signal and PVC‐corrected FTP SUVR became significant (r_s = −0.271, p = 0.025), with a negative coefficient. The findings for AT8 and AT100 remained consistent between PVC‐corrected and uncorrected analyses (Figure S3B).

FIGURE 3.

Region of intrest (ROI)‐wise correlations between immunohistochemistry (IHC) tau burden and 18F‐Flortaucipir (FTP) positron emission tomography signal. Scatterplots show relationships between averaged IHC tau signal and FTP standardized uptake value ratios (SUVRs) across ROIs. (A) Case 1 (Braak IV), after exclusion of subthreshold ROIs using Gaussian mixture model (GMM) thresholds, showed no significant correlations. (B) Case 2 (Braak VI) demonstrated a significant positive correlation between AT8 signal and FTP SUVRs. LTL, lateral temporal lobe; MTL, medial temporal lobe.

To compare how the relationships between FTP and IHC signals differ by lobar region, we plotted Case 2 ROIs by region (Figure S4). As expected, relationships between IHC signal and FTP SUVRs varied among these regions. Limbic MTL ROIs had the highest histological burden but low FTP SUVRs, whereas frontal ROIs showed stronger alignment. In contrast, other neocortical regions had variable PET SUVRs without a corresponding increase in histological burden.

3.5. Within ROI voxelwise correlation for select ROIs

Voxels within an ROI may exhibit a gradient of IHC and PET signals. As a result, the influence of small areas with high correlation within an ROI may be diluted when averaged with regions of low correlation across all voxels. By examining relationships on a voxel‐by‐voxel basis we explored the distribution and correlation of FTP SUVRs and IHC signals within ROIs. For Case 1, we selected four bilateral FreeSurfer‐defined ROIs in the MTL that were early sites of telencephalic tau accumulation in AD: the entorhinal cortex (Braak I region), hippocampus (Braak stage II region), parahippocampal gyrus (Braak stage III/IV region), and amygdala (Braak stage III region). ⁷ Figure 4 visualizes the voxelwise joint distributions of FTP SUVRs and IHC signals in these ROIs using hexagonal binned plots. Spearman's rank correlations between IHC signal and FTP SUVRs were significant (p < 1e‐10 for all) for all three antibodies in the four selected ROIs, and Spearman's rank correlation coefficients ranged from −0.260 to 0.599. Although we hypothesized that the relationship between IHC and FTP signals would be noisy at low histological tau burden and increase linearly with increasing histological tau burden, we found instead a ceiling effect in FTP SUVRs. FTP SUVRs plateaued at high histological burden. This suggests a limited range for PET to reflect the extent of IHC signal. In regions that exhibited this ceiling effect, we fitted nonlinear least squares with self‐starting logistics models. Asymptotes for the fitted curves were shown in Figure 4 and ranged from 1.14 to 1.76 FTP SUVRs.

FIGURE 4.

Voxelwise relationships between 18F‐Flortaucipir (FTP) positron emission tomography and immunohistochemistry (IHC) tau signals in Case 1. Hexagonal binned plots show voxelwise distributions of FTP standardized uptake value ratios (SUVRs) and IHC tau signal within four bilateral medial temporal lobe ROIs. Logistic models illustrate ceiling effects in FTP SUVRsat higher histological burden. Darker colors indicate higher voxel density.

For Case 2 (Braak VI), in addition to the four early MTL bilateral ROIs, we added four neocortical bilateral ROIs that accumulate tau only at later Braak stages (Figure 5B): the inferior temporal gyrus (Braak stage IV), posterior cingulate cortex (Braak stage IV), lateral occipital cortex (Braak stage V), and precentral gyrus (Braak stage VI). All IHC signals in all eight selected Case 2 ROIs showed significant correlations with FTP SUVRs (p < 1e⁻¹⁰), except for AT100 in the hippocampus. Spearman rank's correlation coefficients ranged from −0.157 to 0.751. In contrast to Case 1, Case 2 showed negative correlations in the hippocampus for all three antibodies, and noticeably, MC1 had a narrower range of histological signal than AT8 and AT100 in all ROIs except for the hippocampus and lateral occipital cortex where the ranges were similar. Like Case 1, voxels in the Case 2 entorhinal cortex and parahippocampal gyrus showed a ceiling effect for all antibodies. Fitted logistic models demonstrated that FTP SUVRs first increased in an almost linear fashion (most voxels have at least moderate tau burden) but plateaued around 1.40−2.05 FTP SUVRs as histological tau burden continued to increase. Hexagonal binned plots showed noisy IHC‐FTP joint distributions for all three antibodies in the hippocampus, with no discernible ceiling effects, and the IHC‐FTP relationship in the amygdala appeared more linear for AT8 and MC1. In neocortical ROIs, fitted logistic models in the neocortical ROIs generally had higher asymptotes (1.72–3.16 FTP SUVRs) than those in the MTL ROIs.

FIGURE 5.

Voxelwise relationships between 18F‐Flortaucipir (FTP) standardized uptake value ratios and immunohistochemistry (IHC) tau burden in medial temporal and neocortical regions of interest (ROIs)FF in Case 2. (A) Voxelwise distributions for four medial temporal lobe ROIs (Braak I–IV). (B) Voxelwise distributions for four neocortical ROIs (Braak IV–VI). Logistic fits demonstrate regional differences and plateau effects in FTP signal. All correlations were significant except AT100–FTP in the hippocampus.

4. DISCUSSION

We investigated how in vivo FTP signal relates to neuropathologically confirmed AD‐tau lesions in two human brains histologically staged at Braak stages IV and VI, scanned shortly before death. Our approach leveraged a novel deep‐learning pipeline to generate 3D quantitative density heatmaps of AD‐tau burden stained with AT8, AT100, and MC1. Each antibody targeted different conformations and phosphorylation states of pathological tau. By co‐registering these 3D IHC maps with FTP‐PET scans and high‐resolution post‐mortem T1‐weighted structural MRI studies, we achieved both ROI‐ and voxelwise comparisons that go beyond most prior work limited to smaller, fewer regions or semi‐quantitative assessments. ¹³ , ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹

A primary observation was the contrasting correlations between IHC signal and FTP SUVRs in the two cases with significantly different levels of AD burden. In Case 1 (Braak stage IV), numerous non‐MTL ROIs lacked histological tau but showed variability in FTP SUVRs (left side of each plot in Figure 3A) that remained after partial‐volume correction (Figure S3A), suggesting that SUVR values below around 1.3 reflect non‐tau, off‐target binding. These values align with reported FTP positivity thresholds of 1.2–1.4 SUVRs. ³³ , ³⁴ , ³⁵ Thus, most Case 1 ROIs would be considered FTP‐negative despite measurable IHC signal. Here, “IHC signal” refers to our quantitative, downsampled heatmap value (percent area occupied by above‐threshold AT8/AT100/MC1 immunoreactive structures), which aggregates neuronal (pretangle/tangle‐like somatodendritic staining) and neuritic/neuropil thread pathology and does not explicitly separate these lesion types, as described in Methods. Applying histological thresholds to exclude these noisy ROIs left mainly MTL regions, with IHC signals ranging from 0.1 to 1.2 unit. FTP SUVRs in these ROIs remained between 1.0 and 1.3 SUVR, failing to capture the tenfold variability in histological tau burden. The concomitant lack of significant correlation between FTP SUVRs and any of the three IHC tau signals is consistent with reports that FTP‐PET lacks sensitivity to detect tau and in early to mid‐Braak stages in MTL. ¹⁴ , ¹⁶ , ¹⁷ , ¹⁹

In Case 2 (Braak stage VI), the overall FTP‐IHC relationship was stronger. ROI‐wise analyses revealed a significant positive correlation with AT8 (r _s = 0.424, p < 0.001), whereas AT100 and MC1 correlations were nonsignificant (p = 0.204 and p = 0.111, respectively). However, even when ROI‐wise correlation was significant, regional variability remained evident (Figure S4).

ROI‐wise and voxelwise analyses converged on a regional dissociation between histological tau burden and FTP uptake. ROI‐wise distributions showed that limbic MTL regions had the highest histological tau burden across antibodies, whereas occipital regions had the highest FTP SUVRs (Figure 2); notably, MTL structures exhibited high AT8 and AT100 immunoreactivity but comparatively modest FTP SUVRs, suggesting that FTP may underestimate pathological load in heavily affected MTL regions, consistent with saturation/ceiling behavior and compounded by partial‐volume limitations in small, atrophic structures. This pattern aligns with prior reports indicating that FTP behaves differently in MTL than neocortical regions in AD, ¹⁷ potentially reflecting regional differences in sensitivity, binding substrates, or tau aggregate characteristics. Regional variation in tau maturity, post‐translational modification profiles, and conformational/isoform heterogeneity may further influence tracer correspondence. ²² , ³⁶ Conversely, across neocortical regions, FTP SUVRs have been demonstrated to track with co‐localized tau pathology at autopsy. ¹⁶ , ¹⁷ , ¹⁸ Within neocortical regions with low tau pathology, however, FTP SUVRs may present some variability, implying a proportionally larger influence of non‐tau/background signal. Consistent with this framework, an exploratory voxelwise analysis by Chen et al. ³⁷  demonstrated moderate FTP to IHC p‐tau correspondence, suggesting other factors may influence binding even though tau pathology is a major contributor.

Voxelwise analyses replicated this discordance with finer granularity, showing that regional averaging can obscure localized PET–histology associations (e.g., Case 1 MTL), yet also revealing that in nearly all MTL ROIs for both cases FTP SUVRs plateaued at moderate‐to‐high histological burden (Figures 4 and 5A), underscoring a limited PET dynamic range in highly affected regions. The highest FTP SUVRs were ≈1.2 in Case 1 MTL and generally 2–3 in Case 2 neocortex (slightly lower in MTL), and the consistently higher FTP SUVRs in Case 2 than Case 1 at comparable histological burden (Figure 3) further supports a role for stage‐ and region‐dependent influences beyond resolution alone.

The differential performance of the three tau antibodies (AT8, AT100, MC1) provides complementary windows into pathological tau and may help contextualize the biological determinants of FTP signal. Within this framework, the closer correspondence between FTP and AT8 than AT100 and MC1 is not inconsistent; rather, at PET resolution the PET–histology relationship is strongly influenced by lesion morphology and spatial distribution after downsampling and smoothing. Because AT8 captures abundant neuritic/thread‐like pathology and neuropil labeling from early disease stages, it may better track PET‐scale variance in many regions. In contrast, AT100 shows a slightly lower abundance and dynamic range, reducing correlations even when AT100‐positive pathology is present. Representative staining patterns for AT8, AT100, and MC1 in our tissue are provided in Figure S1, and full 3D reconstructions of each antibody map for each case, including videos are available in a previous article. ²⁰ An important strength of our study is the comprehensive and quantitative nature of our histological mapping and the short gap between PET scan and autopsy. Rather than sparse or semi‐quantitative sampling, we processed nearly the entire brain at high resolution (1 µm²) measures of tau pathology. We then downsampled and smoothed these data to match the resolution of PET. This allowed spatially matched, voxel‐by‐voxel comparisons across the whole brain, which is a methodological advance over prior studies limited to a few dissected blocks or small regions of interest. Our pipeline also employed advanced co‐registration steps with postmortem MRI to ensure that the IHC heatmaps aligned as precisely as possible with in vivo imaging data. ²⁰ Moreover, by examining both ROI‐wise summaries and voxelwise distributions, we captured broad regional patterns as well as local heterogeneities in IHC and FTP signal.

Nevertheless, several limitations should be acknowledged. First, only two AD cases were examined, which does not capture the full phenotypic or genetic heterogeneity of the disease. Second, tissue distortion during autopsy and sectioning, together with repeated co‐registration transformations, likely introduced some spatial misalignment, particularly in highly folded regions such as the hippocampus. Third, although histological data were downsampled to approximate PET resolution, partial‐volume effects and residual smoothing remain potential biases. Fourth, although multiple tau post‐translational modifications are described in AD, ³⁶ we assessed only three. We also did not include β‐sheet–binding dyes such as Thioflavin as its relative insensitivity to diffuse thread pathology makes it less optimal for our voxelwise PET‐scale comparisons. Finally, despite efforts to mitigate nonspecific binding and atrophy‐related effects, off‐target flortaucipir binding to other tissue components cannot be fully excluded. We did not quantify candidate non‐tau contributors to FTP retention, which will be important to incorporate in future end‐of‐life PET–autopsy studies. Because the study includes all eligible cases accrued during recruitment (N = 2), findings provide high‐resolution mechanistic insight rather than population‐level generalizability. We intentionally did not generate matched ROI‐/voxelwise Aβ maps in the same framework as the tau heatmaps because prior validation work does not support Aβ as a meaningful direct source of FTP signal. ⁹ , ¹⁸ Nevertheless, future studies may incorporate regionally matched Aβ mapping to formally test for any residual influence on FTP signal, particularly in low‐tau regions.

In sum, our results show that FTP tends to correlate with AT8‐positive tau in regions of moderate to high disease severity, indicating that the tracer has utility in identifying areas of substantial AD‐related tau. However, at very high levels of pathology, FTP may plateau and consequently underrepresent high tau burden. At lower tau levels, nonspecific or off‐target uptake can confound signal interpretation. These observations highlight the importance of regional context, disease stage, and tau conformation when interpreting FTP signals, particularly in clinical trials where anti‐tau therapies, aimed to reduce protein aggregates, use change in FTP SUVRs as a measure of that reduction. The distinct patterns revealed by the three tau antibodies also emphasize the heterogeneity of tau pathology itself, which may require newer PET tracers or complementary markers to fully capture disease progression in both early and late stages. By integrating multiple staining methods and varying image scales, future research can further clarify the sensitivity and specificity of FTP for distinct tau species, enabling more accurate staging of AD and better monitoring of disease progression or therapeutic interventions.

AUTHOR CONTRIBUTIONS

Yishu Chao performed data analyses, created figures, and drafted the manuscript. Yuheng Chen performed data analyses, contributed to figure design, provided critical input, and contributed to method development. Trevor A. Chadwick contributed to data interpretation and provided critical review of the manuscript. Theresa M. Harrison supported data analyses and provided critical review of the manuscript. Helmut Heinsen led method development, offered critical insights on study design revisions and provided critical feedback. Daniela Ushizima led method development, performed analyses, and provided critical feedback. Duygu Tosun conceived the study, obtained funding, led method development, analyzed data, and offered critical input on the manuscript. William J. Jagust provided conceptual guidance and critical feedback on data interpretation and drafted the manuscript. Lea T. Grinberg conceived the study, obtained funding, oversaw and led method development and data analyses, and drafted the manuscript. All authors read and approved the final manuscript.

CONFLICT OF INTEREST STATEMENT

Leonardo Iaccarino is a full‐time employee and shareholder of Eli Lilly and Company. Lea Grinberg and Duygu Tosun collected data for Clinical trial NCT02350634 and received a research grant from Avid Radiopharmaceuticals that funded data acquisition used in this study. The other authors declare no conflicts of interests. Author disclosures are available in the supporting information. Author disclosures are available in the supporting information. Author disclosures are available in the supporting information.

CONSENT STATEMENT

Consent was obtained as part of the clinical trial NCT02350634

Supporting information

Supporting Information

Supporting Information

DATA AVAILABILITY STATEMENT

Data and code used in this study can be made available upon request.