Deep Learning Based Large-scale Histological Tau Protein Mapping for Neuroimaging Biomarker Validation in Alzheimer’s Disease

Abstract

Deposits of abnormal tau protein inclusions in the brain are a pathological hallmark of Alzheimers disease (AD), and are the best predictor of neuronal loss and clinical decline, but have been limited to postmortem assessment. Imaging-based biomarkers to detect tau deposits in vivo could leverage AD diagnosis and monitoring beginning in pre-symptomatic disease stages. Several PET tau tracers are available for research studies, but validation of such tracers against direct detection of tau deposits in brain tissue remains incomplete because of methodological limitations. Confirmation of the biological basis of PET binding requires large-scale voxel-to-voxel correlation has been challenging because of the dimensionality of the whole human brain histology data, deformation caused by tissue processing that precludes registration, and the need to process terabytes of information to cover the whole human brain volume at microscopic resolution. In this study, we created a computational pipeline for segmenting tau inclusions in billion-pixel digital pathology images of whole human brains, aiming at generating quantitative, tridimensional tau density maps that can be used to decipher the distribution of tau inclusions along AD progression and validate PET tau tracers. Our pipeline comprises several pre- and post-processing steps developed to handle the high complexity of these brain digital pathology images. SlideNet, a convolutional neural network designed to process our large datasets to locate and segment tau inclusions, is at the core of the pipeline. Using our novel method, we have successfully processed over 500 slides from two whole human brains, immunostained for two phospho-tau antibodies (AT100 and AT8) spanning several Gigabytes of images. Our artificial neural network estimated strong tau inclusion from image segmentation, which performs with ROC AUC of 0.89 and 0.85 for AT100 and AT8, respectively. Introspection studies further assessed the ability of our trained model to learn tau-related features. Furthermore, our pipeline successfully created 3D tau inclusion density maps that were co-registered to the histology 3D maps.