Co-occurrence of different neurodegenerative diseases is increasingly common with age and acts as a confounding factor in the development of disease-specific biomarkers. Yet, even by the gold standard of evaluating immunostaining for aggregated proteins in autopsy brains, pathologic complexity makes it impossible to reliably quantify the mixture of diseases by visual inspection, especially when coexistent disorders both feature the same aggregated protein, albeit in different disease-specific patterns. Here, we hypothesize that recent advances in deep learning can identify the distinctive patterns of Alzheimer disease (AD) and progressive supranuclear palsy (PSP) neuropathology, thereby allowing us to de-convolve their individual contributions from phospho-tau immunostaining of mixed pathologies. We will tackle this problem in three steps. First, in order to incorporate biological knowledge and enable interpretability of our disease predictions, we will develop a set of deep learning classifiers to identify disease relevant ?features? in virtual whole slide images. These features will include different types of cells (e.g. neurons, astrocytes), aggregates (e.g. tufted astrocytes and senile plaques that are enriched in PSP and AD, respectively) and tissue regions (gray vs. white matter, which differ in pattern of involvement in these diseases). Second, based on the assumption that comorbid pathologies exhibit a mixture of pure disease features, we will build disease classifiers from pure AD and pure PSP cases. Given a local patch of tau-stained tissue, these classifiers will return their confidence that tissue exhibited either of these diseases. We will evaluate two approaches, one building on the ?features? identified above and the other a more traditional black- box deep learning approach working purely off of image patches. Finally, we will evaluate our pure disease classifiers on cases with mixed pathologies based on pathologist review and concordance with antibodies to tau isoforms whose individual histomorphologies help to distinguish between AD and PSP. As they will identify established neuropathology features demonstrated by the widely-used AT8 phospho-tau and 3R and 4R tau isoform immunostaining, our classifiers will be a valuable resource for future digital imaging based studies in neuropathology. Our framework for de-convolving comorbidities from autopsy samples can be extended to other diseases, thus enabling better integration with clinical and biomarker data, and ultimately, improved antemortem diagnosis and therapy.
Co-occurrence of different neurodegenerative pathologies is increasingly common with age. Here, we aim to use deep learning to distinguish the relative contributions of individual comorbid diseases from tau-stained images of autopsy brains affected by two comorbid conditions, Alzheimer disease and progressive supranuclear palsy. Technologies developed during this proposal will provide powerful machine learning tools for neuropathology, while its successful completion will facilitate better integration with clinical and biomarker data, and ultimately, improved antemortem diagnosis and therapy.
