Age-informed, attention-based weakly supervised learning for neuropathological image assessment.

Area of Science:

• Neuropathology
• Digital Pathology
• Artificial Intelligence in Medicine

Background:

• Diagnosing neurodegenerative disorders (NDs) like Chronic Traumatic Encephalopathy (CTE) is challenging due to subtle pathological changes.
• Manual histopathological analysis is time-consuming, variable, and may miss early signs of disease.

Purpose of the Study:

• To develop an automated, age-informed computational pipeline for predicting tau pathology in NDs.
• To enhance diagnostic accuracy and identify subtle structural alterations indicative of neurodegeneration.

Main Methods:

• An attention-based multiple instance learning (MIL) pipeline was developed using Luxol Fast Blue and Hematoxylin & Eosin (LH&E) stained whole-slide images (WSIs).
• The model predicts AT8 density (a marker of p-tau aggregation) and incorporates patient age for improved accuracy.
• Quantitative evaluation procedures for foundation models (FMs) were established, assessing attention map properties and robustness.

Main Results:

• The age-informed MIL pipeline accurately identified critical pathological regions and predicted AT8 density.
• Interpretable attention maps highlighted structural changes associated with tau pathology.
• Developed benchmarks demonstrated the model's robustness to variations in staining and noise.

Conclusions:

• The developed pipeline enables scalable, automated WSI analysis for ND diagnosis.
• This approach supports earlier and more precise detection of CTE and other NDs.
• The study provides tools for evaluating and optimizing foundation models in digital neuropathology.