Computer Science · Artificial Intelligence
P4‐338: MACHINE LEARNING‐BASED HISTOPATHOLOGICAL APPROACH TO TAU PATHOLOGY
Maxim Signaevski, Kurt Farrell, Nabil Tabish, Elena Baldwin, Marcel Prastawa, John Koll, Gerardo Fernández, Jack Zeineh, Dushyant P. Purohit, Russell W. Hanson, John F. Crary
See on the mapAbstract
Accumulation of abnormal tau in neurofibrillary and glial tangles occurs in Alzheimer disease (AD) and a spectrum of amyloid-independent primary tauopathies. These tauopathies have diverse and overlapping morphological phenotypes that obscure diagnostic classification and quantitative assessments. There is a critical need to augment our ability to recognize and quantify tau pathology in post-mortem human brain tissues to facilitate clinical, genetic, and biomarker studies. Recently, powerful machine learning-based approaches have emerged that allow recognition and quantification of pathological changes from digital images. We are applying deep learning algorithms to augment the neuropathological assessment and create quantitative data for further clinicopathological correlations, as well as molecular and genomic studies of AD and neurodegenerative disease. Histopathological material was derived from brain samples from the Mount Sinai Alzheimer's Disease Research Center and the Mount Sinai/JJ Peters VA Medical Center Brain Bank (total n=1,700). We used a custom informatics platform to manage slide annotations and full slide images. Slides were imaged using the Philips Ultra Fast Scanner integrated with our in-house information management system. Human expert annotators labeled digital histology slides containing a range of pathological changes (e.g., tau tangles, diffuse plaques, neuritic plaques, and background elements). We utilized annotated whole slide regions to train a deep learning convolutional neural network (CNN) implemented in PyTorch against the human gold standard annotations. Our preliminary results show that the deep learning framework is capable of identifying and quantifying some pathological and diagnostic traits, which we further integrate with other metadata. We achieved >99% validation accuracy and <1% validation loss correctly distinguishing multiple degenerative features over 50 epochs in our framework consisting of 16,812,353 parameters of which 9,177,089 are trainable. Machine learning is a potentially useful tool for complex pathological assessment of AD and other tauopathies. Using our deep learning classifiers, we have the potential to integrate cell- and region-specific annotations with genetic, molecular and proteomic data. Such quantitative analysis provides unbiased data for clinicopathological correlations and enhances our knowledge of the disease.