Learning precise segmentation of neurofibrillary tangles from rapid manual point annotations.

Area of Science:

• Neuropathology
• Computational Biology
• Artificial Intelligence

Background:

• Neurofibrillary tangles (NFTs) composed of abnormal tau protein are a key pathological feature of Alzheimer disease (AD).
• Accurate quantification of NFT burden is crucial for deep phenotyping and understanding AD's relationship with clinical, demographic, and genetic factors.
• Manual analysis of NFTs in digital whole slide images (WSIs) is labor-intensive, prone to variability, and not scalable for large datasets.

Purpose of the Study:

• To develop and validate a scalable, open-source deep-learning approach for quantifying NFT burden in post-mortem human brain tissue WSIs.
• To enable detailed spatial distribution and morphology analysis of NFTs at a scale unachievable by manual methods.

Main Methods:

• A semantic segmentation model was trained using single-point-per-NFT annotations on regions of interest from temporal cortex WSIs of AD cases across three institutions.
• The model was designed to generate detailed NFT boundaries at the single-pixel level.
• Performance was compared against a deep object detection model, and both were correlated with expert semi-quantitative scores.

Main Results:

• The semantic segmentation model achieved an Area Under the Receiver Operating Characteristic curve of 0.832 and an F1 score of 0.527 on a held-out test set.
• A deep object detection model showed comparable performance, achieving results 60% faster than segmentation.
• Both deep-learning models demonstrated strong correlation with expert semi-quantitative NFT burden scores at the whole-slide level.

Conclusions:

• The developed deep-learning pipeline offers a scalable and accurate method for quantifying NFT burden in Alzheimer disease brain tissue.
• This open-source tool facilitates high-throughput, detailed analysis of NFT spatial distribution and morphology, advancing AD research.
• The approach overcomes limitations of manual assessment, enabling deeper insights into AD pathology and its correlates.