Diagnosis of Alzheimer’s Disease and Tauopathies on Whole Slide Images using Deep Learning Algorithm (S39.001)

Abstract

<h3>Objective:</h3> To develop a pipeline for diagnosing Alzheimer’s disease (AD), corticobasal degeneration (CBD), globular glial tauopathy (GGT), Pick’s disease (PiD) and progressive supranuclear palsy (PSP), on a single whole slide image (WSI) of tau immunohistochemistry. <h3>Background:</h3> Neuropathological assessment at autopsy is the gold standard for diagnosing neurodegenerative disorders; however, it is time-consuming and depends on the expertise of pathologists. As digital pathology has become widely used, machine learning has been adapted for high-throughput evaluation and diagnosis of histopathologic images. <h3>Design/Methods:</h3> We used clustering-constrained-attention multiple instance learning (CLAM) on WSIs of patients with AD (n=30), CBD (n=20), GGT (n=10), PiD (n=20), PSP (n=20) and non-tauopathy (n=20). Three sections (A: motor cortex; B: cingulate and superior frontal gyri; C: caudate nucleus and putamen) with tau immunohistochemistry were scanned and converted to WSIs. We evaluated the models using 5-fold cross-validation. Attention-based interpretation analysis was performed to understand the morphological features contributing to the diagnosis. Within highly attended regions, we also augmented gradient-weighted class activation mapping (Grad-CAM) to the model to visualize cellular-level evidence of the model’s decisions. <h3>Results:</h3> The model trained in Section B showed the highest area under the curve (0.970 ± 0.037) and diagnostic accuracy (0.873 ± 0.087). The heatmap showed the highest attention in the gray matter of the superior frontal gyrus in AD and the white matter of the cingulate gyrus in CBD. The Grad-CAM showed the highest attention in neurofibrillary tangles in AD, threads in the white matter in CBD, globular astrocytic lesions in GGT, Pick bodies in PiD and coiled bodies in PSP. <h3>Conclusions:</h3> Our diagnostic pipeline could diagnose five tauopathies with high accuracy (approximately 87%). These findings demonstrated the feasibility of CLAM for the classification task on WSIs, which encourages further investigation, focusing on clinicopathological correlation studies. <b>Disclosure:</b> Dr. Koga has nothing to disclose. Ms. Kim has nothing to disclose. Dr. Sekiya has nothing to disclose. Mr. Martin has nothing to disclose. Monica Castanedes-Casey has nothing to disclose. Mr. Yao has nothing to disclose. Dr. Dickson has nothing to disclose. Dr. Hwang has received personal compensation in the range of $100,000-$499,999 for serving as a Consultant for AITRICS. Dr. Hwang has stock in KURE.AI Therapeutics.