Abstract
111
Objectives: The hyperphosphorylation and abnormal aggregation of tau protein is one of the main hallmarks for Alzheimer’s disease (AD). The post mortem studies of AD revealed that tau spreads in a certain pattern along with the disease progression, namely Braak stages. Tau quantification has not yet reached consensus to be introduced into the biomarker scheme, as recent findings provided conflicting results as opposed to Braak staging. Here, we propose a data-driven approach, without any priori assumption such as Braak model, to reproduce the tau progression in AD and support biomarker scheme of tau PET using variational autoencoder (VAE) and hierarchical agglomerative clustering as a building block. We believe that latent or hidden patterns inherently exist in tau PET images. By clustering such essence of features in latent space, we derived and examined the spatial pattern which is assumed to correspond to a certain stage of the AD progression.
Methods: 1080 pairs of T1 MRI image and AV-1451 PET were recruited in total (78 AD, 483 MCI, 519 CN). PET image were spatially normalized to the Montreal Neurological Institute (MNI) space using statistical parametric mapping (SPM8, www.fil.ion.ucl.ac.uk/spm). Our method consists of two building blocks, VAE and hierarchical clustering. In this work, each encoder and generator was built with five convolutional layers and a latent feature dimension of 100. Before being fed into the network, the original data were divided by mean uptake of the cerebellum and down-sampled to half in each dimension. The hierarchical agglomerative clustering was performed in latent space. The significance of the group difference between clusters was evaluated using the Chi-Squared analysis for the categorical clinical phenotype variables. One-way analysis of variance (ANOVA) was performed for continuous clinical phenotype variables, followed by Tukey’s post hoc pairwise test for multiple comparisons.
Results: t-SNE plot as well as contingency table illustrated the clustering result and the diagnosis of 1080 data. The tau distribution of each cluster was arranged in orders, with help of clinical information such as diagnosis, age, MMSE, and APOE4 and the series of cluster 4, cluster 0, and cluster 3 resembled the Braak stages. Standardized uptake value ratio (SUVr) for each ROI was calculated with cerebellum grey matter as a reference region. Across the majority of regions, cluster 3 presented the highest average SUVr. Amongst AD-signature regions, including amygdala, hippocampus, heschl, fusiform, inferior, middle and superior temporal gyrus, insula, and anterior, middle, and posterior cingulate, the most sensitive region was amygdala.
Conclusions: The data-driven approach, without any priori information on tau progress, corresponded well with the Braak staging. Our findings suggest that amygdala is the most sensitive region across the clusters, indicating the most ideal for the early detection of AD. Figure 1 t-SNE plot with clustering reult (left) and diagnosis (middle), and heatmap of contingency table of clustering result and diagnosis (right). conv (MCI-converter) and nonconv (MCI-nonconverter) Figure 2 Average image of cluster 4,0 and 3 Figure 3 SUVr of ROIs, including temporal_inf, cingulum_post, parahippocampal, hippocampus, and amygdala, between clusters
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.