Combined segmentation and classification of Alzheimer’s disease pathology distribution within human hippocampal subregions

Terri‐Leigh Stephen,Laura Korobkova,Kymry T Jones,Ryan P Cabeen,Maricarmen Pachicano,Debra Hawes,Carol A Miller,Michael S Bienkowski

DOI: https://doi.org/10.1002/alz.066689

2022-12-01

Abstract:Alzheimer’s disease (AD) is a complex neurodegenerative disease characterized by progressive amyloid and tau pathology. Quantifying and classifying AD pathology across the human brain is challenging due to its large size and expertise needed for critical evaluation of the morphological variety in structure types (e.g. diffuse vs dense plaques, neurofibrillary tangles, neuropil threads, etc.). High‐throughput microscopy imaging and artificial intelligence allow for large‐scale, quantitative ‘digital pathology’ surveys. Due to their structural differences, distinct pathological aggregates in microscopy images can be identified and quantified by trained artificial neural networks (Tang et al 2019). However, current approaches do not incorporate anatomical segmentation to map the distribution of AD pathology across whole tissue sections. Understanding where and when AD pathology is distributed across different brain regions is critical to defining cell type susceptibility to neurodegeneration. This project’s goal was to create a combined neural network classification and segmentation quantification approach to generate anatomically‐accurate pathology maps of the human hippocampus and investigate the spatial distribution of distinct types of pathology. Whole slide images of 4G8 and Gallyas stained coronal tissue sections from the USC ADRC Neuropathology Core were imaged using a slide scanning microscope (29 patients with low, intermediate, or high AD stages). 4G8 sections were manually annotated and segmented using the Allen Human Brain Atlas into thirteen layered hippocampal subregions and delineated using Quantitative Imaging Toolkit (QIT). The machine learning toolkit Ilastik was utilized to quantify and classify 4G8 amyloid staining (dense vs diffuse plaques) and Gallyas (neurofibrillary tangles) within each hippocampal subregion and overlaid to segmentation maps. In the hippocampus, we found diffuse amyloid pathology increased more rapidly across progressive AD stages compared to dense amyloid plaques or tau neurofibrillary tangles. The ratio of diffuse/dense labeling was highest in the presubiculum and parasubiculum in high AD patients. Patient APOE4 genotype also significantly increased the fraction of diffuse amyloid labeling in high AD patients. Integrating machine learning quantification and classification approaches with anatomical segmentation maps allows for rapid characterization of AD pathology across hippocampal subregions and provides new insights into AD progression across the brain.

clinical neurology