Validation of machine learning models to detect amyloid pathologies across institutions

Area of Science:

• Neuropathology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Semi-quantitative scoring, such as Consortium to Establish a Registry for Alzheimer's Disease (CERAD), is standard in Alzheimer's disease (AD) neuropathology.
• Machine learning (ML) models, particularly convolutional neural networks (CNNs), show promise in generating quantitative scores from whole slide images (WSIs) that correlate with human assessments.

Purpose of the Study:

• To validate previously published CNN algorithms for AD pathology assessment in a new cohort.
• To assess the robustness of CNN-derived measures against pathological heterogeneity and varying diagnoses.
• To compare computational scoring of gray matter vs. whole tissue amyloid burden.

Main Methods:

• Evaluation of 40 brain bank cases with diverse neuropathological diagnoses (AD, AD with Lewy body disease, TDP-43 inclusions) using established CNN algorithms.
• Comparison of quantitative CNN scores for gray matter and whole tissue amyloid burden against CERAD-like scores.
• Stratification of CNN scores based on AD pathology, co-pathologies, and NIA Reagan criteria.

Main Results:

• CNN-derived quantitative scores significantly correlated with human-derived CERAD-like scores for both gray matter and whole tissue amyloid burden.
• CNN scores effectively stratified cases based on neurodegenerative status and NIA Reagan criteria.
• The AD + TDP-43 group showed significantly higher CNN scores for cored plaques compared to the AD-only group.
• Whole tissue computational scores demonstrated a stronger correlation with CERAD-like categories than region-specific analyses.

Conclusions:

• CNN-based computational pathology models are robust and validated across different cohorts, demonstrating resilience to pathological heterogeneity.
• These ML algorithms provide reliable quantitative measures for AD neuropathology, supporting their integration into routine practice.
• Whole tissue analysis using CNNs offers a more comprehensive and correlated assessment compared to traditional focused field-of-view methods.