Automated classification of tau neuropathology from histological images using vision-based machine learning

Abstract Primary tauopathies are neurodegenerative diseases characterised by accumulation of misfolded, hyperphosphorylated tau protein in neuronal and glial cells, each defined by hallmark inclusion morphologies. Post-mortem neuropathological examination remains the diagnostic gold standard but relies on scarce specialist expertise, and existing computational approaches are more developed for quantifying overall tau pathology burden in histological sections than for resolving cellular context. We evaluated whether Google Teachable Machine, a freely available no-code computer vision platform, could classify tauopathies and their associated tau inclusions. Two classifiers were trained on phosphorylated tau-immunostained sections from 125 cases: one classifying pancortical pathology patterns across eight distinct tauopathies at 5× magnification, the other classifying nine individual tau inclusion morphologies at 40× magnification. Case-level accuracy was 73.5% and 81.8% for the low- and high-magnification models respectively, with image-level accuracies of 77.4% and 72.2%. Recall reached 90% or higher for Alzheimer's disease, neurofibrillary tangles and granular/fuzzy astrocytes, with favourable area under the curve values across most categories. Performance was reduced in rare, underrepresented and morphologically overlapping categories, and where class boundaries depend on amyloid-β burden not visible on tau immunohistochemistry. These findings demonstrate that general-purpose vision models can be used for pattern recognition and morphology-based classification in neuropathology, providing an accessible benchmark for future approaches. Trained classifiers are provided as a web application.