AI‐based histological analysis in Alzheimer’s disease : the point of view of the biologist
Benoit Delatour- Psychiatry and Mental health
- Cellular and Molecular Neuroscience
- Geriatrics and Gerontology
- Neurology (clinical)
- Developmental Neuroscience
- Health Policy
- Epidemiology
Abstract
Background
Tauopathies in Alzheimer’s disease are characterized by occurrence of neurofibrillary tangles (NFTs) and neuritic plaques (NPs) that are closely associated to cognitive decline and are key elements for neuropathological diagnosis and staging. The automated detection and segmentation of NFTs and NPs has not been fully implemented but thanks to deep learning (DL) based approaches could be assessed today.
Method
Whole‐slide images of phosphotau‐immunostained brain sections (frontal cortex) from AD patients presenting discrete to florid lesion loads were generated. Using a commercial software (©Visiopharm) we developed an iterative “human‐in‐the‐loop” pipeline to improve quality of annotations of NFTs and NPs and create an exhaustive and accurate ground truth database that was used to train DL algorithms for the different objects morphologies to be detected and segmented in brain parenchyma.
Result
A U‐Net based CNN was trained, and the obtained algorithm was precise in NP detection (F1 score: 0.86) and segmentation (Dice score: 0.77). A second algorithm was precise in NFT detection (F1 score: 0.85) and segmentation (Dice score: 0.82).
Conclusion
Two main objectives were reached : 1) optimization of human annotations thanks to iterative AI‐expert feedbacks, 2) implementation of methods to automatically accelerate, refine and robustify histological analysis of discrete objects. Such approaches may help to uncover and quantify subtle morphological differences between AD variants that otherwise could not be neuropathologically segregated. The biological relevance of dichotomizing NFTs and NPs will be discussed in terms of clinico‐pathological correlations and stratification needs. Also, the scope and limits of user‐friendly and commercially available DL‐learning image analysis softwares, directly accessible to researchers and pathologists with limited IT skills, will be reviewed.