Self-supervised learning for characterising histomorphological diversity and spatial RNA expression prediction across...

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in histology

Background:

• Digitization of histological archives necessitates automated methods for analyzing tissue substructures and pathology.
• Current methods require extensive manual annotation, limiting large-scale analysis.
• Integrating morphological data with molecular information remains a challenge.

Purpose of the Study:

• To develop self-supervised learning models for automated analysis of histology images.
• To segment tissues, quantify pathologies, and link histology to gene expression.
• To enable large-scale, data-driven research in digital pathology.

Main Methods:

• Trained a Vision Transformer on 1.7 million histology images from the Genotype Tissue Expression (GTEx) consortium.
• Utilized self-supervised learning for tissue segmentation and pathology detection.
• Developed RNAPath models to predict spatial RNA expression from H&E histology.

Main Results:

• Achieved superior tissue segmentation performance compared to other self-supervised methods (43% increase in silhouette score).
• Accurately detected and quantified histological pathologies like arterial calcification (AUROC=0.93).
• RNAPath successfully predicted and localized RNA expression, validated with immunohistochemistry.

Conclusions:

• Self-supervised machine learning applied to histology archives enables automated analysis of tissue pathology and organization.
• This approach facilitates the study of the interplay between tissue morphology and gene expression.
• The developed methods offer powerful tools for large-scale digital pathology research.