Diagnosis of Alzheimer’s disease and tauopathies on whole slide histopathology images using a weakly supervised deep learning algorithm

Abstract Neuropathological assessment at autopsy is the gold standard for diagnosing neurodegenerative disorders. We aimed to develop a pipeline for diagnosing Alzheimer's disease and other tauopathies, including corticobasal degeneration, globular glial tauopathy, Pick’s disease, and progressive supranuclear palsy. We used deep learning (DL)-based approach called clustering-constrained-attention multiple instance learning (CLAM) on whole slide images (WSIs) of tau immunohistochemistry in three brain regions from 120 patients. We also augmented gradient-weighted class activation mapping (Grad-CAM) to the model for visualizing cellular-level evidence in the model’s decisions. The model using the sections of cingulate and superior frontal gyri achieved the highest area under the curve (0.970±0.037) and diagnostic accuracy (0.873±0.087). Grad-CAM showed the highest attention in known pathognomonic tau lesions for each disease (e.g., Pick bodies for Pick’s disease). Our findings supported the feasibility of the DL-based approach for the classification task on WSIs, which encouraged further investigation, especially focusing on clinicopathological correlation studies.