https://orcid.org/0000-0002-9879-7497
This scientific commentary refers to ‘Three-dimensional mapping of neurofibrillary tangle burden in the human medial temporal lobe’, by Yushkevich et al. (doi:).
Characterizing the progression of Alzheimer’s disease pathology is critical for early diagnosis and assessment of response to treatment. Accumulation of tau neurofibrillary tangles (NFTs) is one of the pathological hallmarks of Alzheimer’s disease and highly related to neurodegeneration and cognitive decline. Histological studies over the past 30 years have demonstrated the characteristic spread of NFTs in ageing and Alzheimer’s disease. The transentorhinal region of the medial temporal lobe (MTL) is thought to be one of the first sites of tau accumulation, along with the locus coeruleus and dorsal raphe nucleus. Tau then spreads from the MTL to limbic and isocortical areas.
Although these histological studies have led to significant advances in biomarker and drug development, they are intrinsically limited and may not actually reveal a complete picture of the underlying NFT progression because of limited sampling of the brain; ‘the sheer size of the cerebral cortex, the enormous amount of tissue to be processed, and the complexity of the many regions to be analyzed have virtually precluded a fully comprehensive neuropathological characterization of Alzheimer’s disease’.1 In this issue of Brain, Yushkevich and colleagues2 perform the most comprehensive characterization of NFT burden to date, using the latest advances in neuroimaging, machine learning, and histological reconstruction to densely quantify 3D maps of NFT burden in the MTL of 18 individuals.
In their study, the authors made use of brain samples from two different research centres, from donors aged 45 to 93 with no known neurological disease (n = 12) or who were previously subjects in a dementia imaging study (n = 6). While one hemisphere was reserved for diagnostic pathology, tissue from the other underwent a series of imaging and cutting steps to ensure accurate alignment could be maintained or recovered to automatically reconstruct histological slides in 3D. This included overnight MRI at 9.4 T to acquire 200-μm resolution anatomical images, lower resolution 7 T MRI to correct distortions, sectioning using a 3D printed mould and cutting guide of the tissue designed from the 7 T scan, and frozen sectioning into 50-μm sections with block-face photographs after each cut.
Sections were stained for neuronal cell bodies (Nissl, every 0.5 mm) and phosphorylated tau (AT8, every 1 mm), then digitized at 20× resolution. Since manually annotating tau pathologies across the entire set of slides would be infeasible, the authors developed a custom web-based slide annotation tool to crowdsource over 11 000 example annotations from 12 raters. They then used these annotations to train an artificial neural network in a weakly supervised approach, whereby pixel-wise segmentation is performed using image-level labels, in order to automatically annotate all the slides. The authors employed the WildCat algorithm, which consists of a fully convolutional neural network with a ResNet-101 architecture followed by a multilevel transfer layer and spatial pooling, to classify histology image patches into tangle versus non-tangle classes. These segmentations were transformed to a reconstructed 3D space through a series of largely automated image processing and registration steps that align the MRI volumes, block-face images, and tissue (including torn fragments) from histology slides. Tau burden heat maps were generated for each subject and also registered to a standard in vivo MRI template, where MTL regions of interest along with whole-brain regions of interest were used to provide summary measures. An illustration of the main imaging, histological processing, reconstruction, and quantification stages is shown in Fig. 1A.
Mapping tau NFT deposition. (A) Overview of the main steps towards generating dense histological tau NFT burden maps in Yushkevich et al.2 (B) Comparison of techniques for mapping tau deposition: histological sampling post-mortem, in vivo mapping with tau PET, and dense 3D histology.
The authors found that the transentorhinal cortex was consistently an area of high tangle burden. However, unlike the classical observations suggesting that early stage tangles are confined to this region,3 this work demonstrated similar and sometimes greater burden in surrounding MTL structures, including the amygdala, anterior temporal lobe, and anterior portions of the CA1 and subiculum hippocampal subfields. The semi-supervised WildCat approach was able to classify NFT burden with a high degree of accuracy, with especially strong performance discriminating between the lowest classes of NFT burden. This demonstrates the potential for mapping the earliest stages of tau pathology. Greater NFT burden in anterior structures over posterior, or an anterior-posterior gradient, in the MTL was also generally observed; seemingly contrary to expectations given that progression in later stages of typical Alzheimer’s disease dementia follows a posterior trajectory. It should be noted however that the majority of cases included in the study did not present with high amyloid-β burden, which may have limited the representation of more Alzheimer’s disease-typical progression patterns of tau.
These findings by Yushkevich et al.,2 when considering the evidence that the subjects were at an early stage of tau deposition, may suggest that early-stage progression has a more nuanced pattern than previously described, with a greater effect on anterior temporal regions. Anterior and posterior MTL have differential involvement in brain networks subserving memory, with an anterior temporal system involved in object memory, and a posterior medial system involved in scene and spatial memory.4 Recent imaging studies have shown that in the earliest stage of preclinical Alzheimer’s disease, functional changes occur in the anterior temporal network, which precede the posterior medial network changes that are accompanied by structural changes.5 Early NFT deposition preferentially affecting anterior temporal network regions could potentially explain this finding.
Three-dimensional reconstruction of histology is a challenging endeavour, and demands expertise and innovation in histological processing, MRI, image processing, and machine learning in order to minimize distortion, align tissue, and automate quantification effectively. Previous efforts to quantify tau protein tangles in 3D histopathology of post-mortem human tissue have been limited to only one or two specimens at most,6 whereas Yushkevich et al.2 mapped 18 specimens. This major advancement in the sample size was made possible through improved automation in nearly all stages. Careful experimental protocol design also played a role, for example, tissue cutting aided by a 3D printed guide constrained and simplified the subsequent task of image registration. The net effect of these physical and computational processing steps was validated by the authors, demonstrating alignment error of anatomy on the order of <0.4 mm, sufficient for accurate 3D reconstruction. Importantly, Yushkevich et al.2 have released the software and scripts used to perform image registration and quantification, opening up the possibility for similar histological reconstruction studies at other sites as well.
Recent advances in tissue clearing techniques, such as CLARITY, CUBIC and iDISCO, coupled with light-sheet imaging and machine learning-based segmentation, offer an alternative and unique opportunity for fast and precise mapping of 3D proteinopathy. Employing these techniques would sidestep many of the challenges of serial histology, and potentially enable the investigation of a larger number of specimens. These techniques have already been applied to visualize and map Alzheimer’s disease pathology in animal models7 and thin sections (500 μm) from human specimens.8 While these novel mapping techniques offer great potential for studying 3D protein burden and propagation, technical limitations exist that preclude clearing, staining and imaging larger post-mortem human specimens.
One might question the necessity of histological studies to investigate tau deposition, since we now have the ability to map progression of tau in vivo with PET imaging and tau-specific ligands. PET imaging also offers the ability to perform longitudinal studies, and the potential to provide individualized biomarkers. First-generation tracers, such as 18F-AV-1451 (also known as 18F-flortaucipir) have been used extensively in research studies, including large studies of subjects with cognitively normal ageing, mild cognitive impairment, and Alzheimer’s disease.9 Although PET imaging studies mostly recapitulate histopathologically derived staging schemes, it is critical to note that specificity is still a challenge.
First, increasing evidence indicates that tau ligands can potentially bind to other, non-tau protein deposits including transactive response DNA-binding protein 43 (TDP-43) and α-synuclein.10 Second, the characterization of tau-PET binding patterns in Alzheimer’s disease and other tauopathies has been hampered (especially in the hippocampus and MTL regions) by off-target signal of the basal ganglia and choroid plexus, particularly across first-generation tau ligands. Furthermore, while first-generation tau ligands show high affinity to the isoform combinations that are typical for Alzheimer’s disease (i.e. 3R/4R paired helical filament tau), they may also target non-Alzheimer tauopathies. Thus, histopathology still remains the gold standard reference for detecting and quantifying tau deposition, with 3D reconstructed histology especially important for developing and validating neuroimaging biomarkers (as summarized in Fig. 1B). Furthermore, this method is currently the only option available when investigating tauopathies where PET tracers have not yet been developed, such as those involving TDP-43 or α-synuclein.
In conclusion, Yushkevich et al.2 have developed a framework to enable 3D quantification of NFT burden, and applied this framework to investigate early tau progression in the region where it first appears in Alzheimer’s disease (MTL). This work and potential future applications to other proteinopathies will likely serve as important references for in vivo neuroimaging studies.
Competing interests
The authors report no competing interests.
