Generalizable Prediction of Alzheimer Disease Pathologies with a Scalable Annotation Tool and an High-Accuracy Model

Characterizing the cardinal neuropathologies in Alzheimer disease (AD) can be laborious, time consuming, and susceptible to intra- and inter-observer variability. The lack of high throughput unbiased approaches to reliably assess neuropathology hampers efforts to use pathology as a means to link clinical features of AD to molecular pathogenesis in the ever-growing datasets of persons with AD. To remove this roadblock, we designed an annotation tool in addition to a computational pipeline to analyze digital microscopic images of postmortem tissue from persons with AD in a fully automated and unbiased manner in only a fraction of the time taken with conventional approaches and allows neuropathological analyses and lesion quantification at multiple scales. The pipeline includes a Mask Regional-Convolutional Neural Network (Mask R-CNN) we trained to detect, classify, and segment different types of amyloid. To establish ground truth for training and validation, we utilized an existing open source platform, QuPath, and developed a tool to collect consensus annotations of neuropathology experts. The Mask R-CNN identified amyloid pathology in samples (with accuracy: 94.6%, F1: 87.7%, Dice: 81.8%) unrelated to the training dataset, indicating that it detects generalizable pathology features. Its quantitative measurements of amyloid pathology on 298 samples correlated with the severity of AD neuropathology assessed by experts and neuropathologists (CERAD ratings) and estimates of cognitive compromise (Clinical Dementia Ratings (CDR)). Our computational pipeline should enable rapid, unbiased, inexpensive, quantitative, and comprehensive neuropathological analysis of large tissue collections and integration with orthogonal clinical and multi-omic measurements.