Automated segmentation by deep learning of neuritic plaques and neurofibrillary tangles in brain sections of Alzheimer’s Disease Patients

Abstract

Alzheimer’s Disease (AD) is a neurodegenerative disorder with complex neuropathological features, such as phosphorylated tau (p-tau) positive neurofibrillary tangles (NFTs) and neuritic plaques (NPs). The quantitative evaluation of p-tau pathology is a key element for the diagnosis of AD and other tauopathies. Assessment of tauopathies relies on semi-quantitative analysis and does not consider lesions heterogeneity (e.g., load and density of NFTs vs NPs).In this study, we developed a deep learning-based workflow for automated annotation and segmentation of NPs and NFTs from AT8-immunostained whole slide images (WSIs) of AD brain sections. Fifteen WSIs of frontal cortex from four biobanks with different tissue quality, staining intensity and scanning formats were used for the present study. We first applied an artificial intelligence (AI-)-driven iterative procedure to improve the generation of pathologist validated training datasets for NPs and NFTs. This procedure increased the annotation quality by more than 50%, especially for NPs when present in high density. Using this procedure, we obtained an expert validated annotation database with 5013 NPs and 5143 NFTs. As a second step, we trained two U-Net convolutional neural networks (CNNs) for accurate detection and segmentation of NPs or NFTs. The workflow achieved a high accuracy and consistency, with a mean Dice similarity coefficient of 0.81 for NPs and 0.77 for NFTs. The workflow also showed good generalization performance across different patients with different staining and tissue quality. Our study demonstrates that artificial intelligence can be used to correct and enhance annotation quality especially for complex objects, even when intermingled and present in high density, in brain tissue. Furthermore, the expert validated databases allowed to generate highly accurate models for segmenting discrete brain lesions using a commercial software. Our annotation database will be publicly available to facilitate human digital pathology applied to AD.

Learning precise segmentation of neurofibrillary tangles from rapid manual point annotations

“…Studies using deep learning methods to detect NFTs predominantly omit segmentation models due to the massive annotator investment required to collect manual ground truth training masks at the pixel level [9,10,47] despite the reported research advantages of object-and pixel-based NFT counts over typical IHC positive pixel count methods [48]. Indeed, segmentation models unlock deeper neuropathological phenotyping than the standard pathological analyses that compress pathological information into a single category or semi-quantitative grade -examples include nuclei and tissue segmentation as tools to improve cancer grading schemes such as Gleason and C-Path scores [49][50][51].…”

“…In neuropathology, deep learning on digitized whole slide images (WSIs) of brain tissue can automate detecting and quantifying distinct pathological features such as amyloid beta plaques [4][5][6]. This includes recognizing and quantifying the NFTs central to AD diagnosis and staging [7][8][9][10]. However, several challenges persist.…”