Large scale, voxel‐to‐voxel correlation analysis of tau positron emission tomography signal to histological pathology in non‐Alzheimer tauopathies

Background Tau pathology strongly correlates with neuronal loss and clinical decline across tauopathies, emphasizing the need to detect and monitor tau pathology in living patients. Positron Emission Tomography (PET) using radiotracers like Flortaucipir ([18F]AV‐1451) enables good visualization of AD tau pathology. However, the sensitivity and specificity of this tracer to visualize tau deposits in non‐AD tauopathies remain poorly understood. PET‐to‐autopsy validation has been challenging because of issues with tissue deformation and large‐scale pathological analysis. Here, we present i) an interdisciplinary approach combining immunohistochemistry, computer vision, and convolution neural network (CNN) to 3D‐reconstructed, billion‐pixel digital pathology tau images of human brain hemispheres, co‐register them to PET and ii) preliminary PET‐to‐pathology voxel‐to‐voxel correlation analyses. Method We employed a pipeline developed in‐house combining whole‐brain processing, free‐floating immunohistochemistry and computer‐based algorithms for 2D and 3D registration. SlideNet, our CNN for segmenting tau inclusions in digital images (Figure 1A), aims at generating quantitative, 3D‐tau‐maps. We applied SlideNet to images of whole hemisphere sections in case #1 with Cortico‐Basal‐Degeneration and case #6 with Progressive‐Supranuclear‐Palsy labeled by immunohistochemistry (CP‐13 antibody, Ser202) (Figure 2A), and the demographics of cases are shown in Table 1. We registered the resulting heatmaps (Figure 1B) to MRI and PET coordinates (Figure 3), enabling voxel‐to‐voxel comparisons. Result The CNN output probability map containing three categories: tau in neurons, tau in glia/processes, and background (Figure 2B). After precision‐recall statistical thresholding at 0.7 (Figure 4A), the resulting tau segmentation shows in Figure 2C. The area‐under‐curve on receiver‐operating‐characteristic graph was 0.84 in testing and validation datasets from the two cases we processed (Figure 4B), suggesting that our CNN selects tau inclusions with strong class separation. We registered the histology‐based, quantitative 3D‐tau‐map and PET scan onto the 3D MRI space (Figures 1C‐D). Preliminary correlation analysis a coefficient of 0.182 (p<0.001) between low resolution tau map and PET scan (Figure 5). Conclusion CNN segmentation of pathology combined with whole‐brain processing and high‐performance computer vision enables voxel‐to‐voxel correlations analysis between histological tau and tau PET tracer binding, a keystep toward early detection of tauopathies. Ongoing regional analysis will provide more granular correlations, especially in areas known for off‐target binding, and we are currently finalizing the result in another 6 cases (Table 1).