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

Deposits of abnormal Tau inclusions in the brain are a well-known Alzheimer9s disease (AD) pathological feature and are the best predictor for neuronal loss and clinical decline. As such, Tau is a potential in-vivo imaging biomarker that could leverage earlier AD diagnosis. In fact, there are several research initiatives to develop PET tracers for imaging Tau in the human brain. Validation of such studies, however, is only reliably performed using histological data, where Tau inclusion can be accurately located. Manually locating and segmenting Tau in such images is unfeasible since a whole human brain slice, at the adequate resolution, spans several Gigabytes of data and carries thousands of inclusions. In this study, we present our preliminary results on using convolutional neural networks (CNN) to automatically locate and segment Tau in whole human brain histological slides. CNNs are multi-layered representation-learning methods capable of automatically discover unknown patterns that best characterize a raw data set. Moreover, they model data in a bottom-up approach, where data is characterized by small low-level features, such as edges and connected objects, in lower layers and increases in abstraction and complexity in the following layers. Our dataset is composed of histological slices stained for AT100, AT8. Each slide was imaged in our in-house built slide scanner at 1.22um resolution. Each image is about 82000x37000 pixel wide. On average, a whole human brain dataset has 16Tb of data. Images were preprocessed and stitched using the UCSF Wynton cluster. A UNet variant was used for Tau localization and segmentation on the full-resolution images, which was performed using an NVIDIA Titan X 12Gb GPU. We then computed Tau 9heatmaps9 based on the segmented images that, in turn, were registered to the MRI.