High-throughput microscopy has outpaced analysis; biomarker-optimized CNNs are a generalizable, fast, and interpretable solution.

The most common approach to characterize neuropathology in Alzheimer's disease (AD) involves a manual survey and inspection by an expert neuropathologist of postmortem tissue that has been immunolabeled to visualize the presence of amyloid beta in plaques and around blood vessels and neurofibrillary tangles of the tau protein. In the case of amyloid beta pathology, a semiquantitative score is given that is based on areas of densest pathology. The approach has been well-validated but the process is laborious and time-consuming, and inherently susceptible to intra- and inter-observer variability. Moreover, the tremendous growth in genetic, transcriptomic, and proteomic data from AD patients has created new opportunities to link clinical features of AD to molecular pathogenesis through pathology, but the lack of high throughput quantitative and comprehensive approaches to assess neuropathology limits the associations that can be discovered. To address these limitations, we designed a computational pipeline to analyze postmortem tissue from AD patients in a fully automated, unbiased, and high throughput manner. We used deep learning to train algorithms with a Mask Regional-Convolutional Neural Network to detect and classify different types of amyloid pathology with human-level accuracy. After training on pathology slides from the Mt Sinai cohort, our algorithms identified amyloid pathology in samples made at an independent brain bank and from an unrelated cohort of patients, indicating that the algorithms were detecting reproducible and generalizable pathology features. We designed the pipeline to retain the position of the pathology it detects, making it possible to reconstruct a map of pathology across the entire whole slide image, facilitating neuropathological analyses at multiple scales. Quantitative measurements of amyloid pathology correlated positively and significantly with the severity of AD as measured by standard approaches. We conclude that we have developed a computational pipeline to analyze digitized images of neuropathology in high throughput and algorithms to detect types of amyloid pathology with human-level accuracy that should enable a neuropathological analysis of large tissue collections and integration of those results with orthogonal clinical and multi-omic measurements.

Cell death is an essential process in biology that must be accounted for in live microscopy experiments. Nevertheless, cell death is difficult to detect without perturbing experiments with stains, dyes or biosensors that can bias experimental outcomes, lead to inconsistent results, and reduce the number of processes that can be simultaneously labelled. These additional steps also make live microscopy difficult to scale for high-throughput screening because of the cost, labor, and analysis they entail. We address this fundamental limitation of live microscopy with biomarker-optimized convolutional neural networks (BO-CNN): computer vision models trained with a ground truth biosensor that detect live cells with superhuman, 96% accuracy more than 100 times faster than previous methods. Our models learn to identify important morphological characteristics associated with cell vitality without human input or additional perturbations, and to generalize to other imaging modalities and cell types for which they have no specialized training. We demonstrate that we can interpret decisions from BO-CNN models to gain biological insight into the patterns they use to achieve superhuman accuracy. The BO-CNN approach is broadly useful for live microscopy, and affords a powerful new paradigm for advancing the state of high-throughput imaging in a variety of contexts.