TauRUS Tau-PET Atlas (MNI space, 1 mm)

Included is an atlas of brain regions representing the typical spatial patterns of tau-PET signal distribution across individuals spanning the Alzheimer's disease spectrum. The regions were created using hypothesis-free, data-driven methods, and are designed to be tau-PET biomarkers used for summarizing tau-PET signal in the brain. A full atlas is included, as well as each ROI separately, and a set of masks of the hippocampus. These are divided into winner-takes-all and cluster-core masks (see below). All images are in MNI space at 1 mm resolution. In addition, atlases at three different resolutions are included to support different types of analyses.

METHODS:

The participant sample included 123 individuals with [¹⁸F]AV1451-PET from the BioFINDER cohort (Hansson et al., 2016), including 31 amyloid-negative healthy controls, 24 amyloid+ healthy controls, 21 amyloid-positive patients with mild cognitive impairment, and 47 amyloid-positive patients with Alzheimer's disease dementia.

Cross-subject [¹⁸F]AV1451-PET covariance networks were derived using an open-source unsupervised consensus-clustering algorithm called Bootstrap Analysis of Stable Clusters (BASC). BASC was originally designed to extract multi-resolution network parcellations from resting-state functional MRI data, where it builds consensus between clustering solutions across within- and between-subject stability matrices (Bellec et al., 2010). The algorithm was adapted to 3D [¹⁸F]AV1451 data by stacking all 123 BioFINDER [¹⁸F]AV1451 images along a fourth (subject) dimension, creating a single 4D image to be submitted as input. BASC first reduces the dimensions of the data with a previously described region-growing algorithm (Bellec et al., 2006), which was set to extract spatially constrained atoms (small regions of redundant signal) with a size threshold of 1000mm³. In order to reduce computational demands, the Desikan-Killainy atlas (Desikan et al., 2006) was used as a prior for region constraint, and the data was masked with a liberal gray matter mask, which included the subcortex but had the cerebellum manually removed (since this was used as the reference region for [¹⁸F]AV1451 images). The region-growing algorithm resulted in a total of 730 atoms, which were included in the BASC algorithm.

BASC next performs recursive k-means clustering on bootstrapped samples of the input data. After each clustering iteration, information about cluster membership is stored as a binarized adjacency matrix. The adjacency matrices are averaged resulting in a stability matrix representing probabilities of each pair of atoms clustering together (Figure 1). Finally, hierarchical agglomerative clustering with Ward criterion is applied to the stability matrix, resulting in the final clustering solution. The process is repeated over several clustering solutions (in this case, between 1 and 50), and the M-STEPs method (Bellec, 2013) was implemented to find the most stable clustering solutions. Briefly, M-STEPS identifies stable clustering solutions that demonstrate the best linear approximation of all solutions across a given subset. Therefore, M-STEPS identifies multiple optimal clustering solutions at different resolutions. In order to maintain relative similarity to Braak neuropathological staging (i.e. six regions-of-interest), we chose the lowest resolution solution for subsequent analysis. However, we have also provided the other higher-resolution solutions with this dataset. Note that no size constraints were imposed on clustering solutions.

BASC includes an option for outputting cluster “cores”, representing the portions of within-cluster peak stability for each cluster. Given that our aim was to produce covariance networks that would be generalizable across samples, we assumed that cluster cores would be more reliable than using clusters in their entirety, and therefore we used cluster cores in all subsequent analyses. Consequently, voxels were only included in a cluster when cluster probability membership exceed 0.5 (BASC default setting), eliminating unstable voxels from analysis (Bellec et al., 2010; Garcia-Garcia et al., 2017). This ensures voxels are only including if they fell within the same cluster around > 50% of bootstrap samples. Both winner-takes-all and cluster-cores are included in this dataset.

RESULTS:

The M-STEPS algorithm identified five-, nine- and 32-cluster solutions as optimal solutions. The five-cluster solution was selected for further analysis. The clusters were interpreted and named as follows: “1: Subcortical”, “2: Frontal”, “3: Medial/Anterior/Inferior Temporal”, “4: Temporo-parietal” and “5: Unimodal Sensory”. Cluster 3 bore resemblance to regions often involved in early tau aggregation and atrophy (Braak and Braak, 1991), while Cluster 4 also appeared similar to regions commonly associated with neurodegeneration in AD (Dickerson et al., 2011). Of note, the hippocampus was largely unrepresented in any of the cluster-cores, though some voxels in the head of the hippocampus were included in Cluster 3, and a few distributed voxels were included in Cluster 1 (Subcortex). However, using a winner-takes-all clustering approach, voxels in the hippocampus were distributed between Clusters 1 and 3.

The clusters showed some similarity to pathological Braak stages (Break et al., 1991), and outperformed other tau-PET ROIs in describing cognitive data in a separate cohort (see Vogel et al., 2019 Human Brain Mapping).