Digital pathology-based quantification of cellular features in neurodegenerative disease

Introduction: Neurodegenerative diseases (NDs) involve complex changes in the brain tissue microenvironment. Immunohistochemistry staining is commonly used to semi-quantify pathology based on standard guidelines, while hematoxylin and eosin (H&E) is used to assess tissue morphology. Studying changes in H&E-stained slides using quantitative methods may provide new insights into NDs.

Methods: We have digitized and create an integrated data mart of 8,567 slides from 350 brain autopsies in the Mesulam Center for Cognitive Neurology & Alzheimer's Disease biobank. Using 71 annotated whole-slide images (WSIs) stratified at into training / validation / testing subsets at the patient level, we develop a tissue segmentation model to delineate gray matter (GM) and white matter (WM) regions. A transformer-based self-supervised (SSL) model was then trained on 1.5 million cropped images of detected cell nuclei from GM to learn representations of nuclear morphology. This SSL model was used to classify neurons, astrocytes, oligodendrocytes, and endothelial cells using 14,603 annotated examples. This classifier was evaluated on 4,335 independent cells from test slides.

Results: Figure 1 illustrates the validation of the gray matter and cell classifier performance. The cell classifier F1-score was 91.35%, with the highest observed performance on neurons (96.18% F1-score) and lowest on astrocytes (74.40% F1-score). SSL further revealed distinct clustering of cell types, particularly neuronal subtypes, in the embedding space.

Discussion: Our findings demonstrate that computational models can measure quantitative features of the brain tissue microenvironment. These tools could improve characterization of pathologic changes in NDs, however further development is required. Future work will explore correlations between quantitative measures of the brain tissue microenvironment and clinical data.

Learning Objectives:

1. Describe a pipeline to quantify brain microenvironment features in neurodegenerative diseases from H&E whole-slide images.
2. Explain how self-supervised learning enables classification of major gray matter brain cell types, from morphology.
3. Interpret quantitative outputs and embeddings to characterize pathologic changes in neurodegenerative diseases.