Doctoral thesis defence: 'Representation Learning for Computational Pathology and Spatial Omics'

Welcome to a defense of Eduard Chelebian's doctoral thesis in computerized image processing. The defense will be held in English.

Abstract: Artificial intelligence (AI) advancements have enhanced the analysis and interpretation of computational pathology. Through representation learning, deep learning models can automatically identify complex patterns and extract meaningful features from raw data, revealing subtle spatial relationships. Spatial omics, which captures spatially resolved molecular data, naturally aligns with these approaches, enabling a deeper examination of tissue architecture and cellular heterogeneity. However, early spatial omics methods often overlooked the morphological and spatial context inherent in tissues.

The integration of spatial omics with imaging AI and representation learning provides a comprehensive view for understanding complex tissue environments, providing deeper insights into disease mechanisms and molecular landscapes. This thesis investigates how deep learning-derived representations from biological images can be utilized in the context of spatial omics and disease processes.

Key contributions of this work include: (i) investigating the correlation between representations learned from models trained on hematoxylin-eosin (H&E)-stained images and underlying gene expression profiles; (ii) applying self-supervised learning to identify genetically relevant patterns across H&E and DAPI staining; and (iii) developing a framework that leverages self-supervised representations to refine cell-type assignments obtained from spatial transcriptomics deconvolution methods. As a culmination of this part of the thesis, this research introduces (iv) a conceptual framework for understanding representations within spatial omics and provides a survey of the current literature through this lens.

The thesis further includes practical applications such as (v) developing a tool for annotation of whole-slide images (WSI) using self-supervised representations and (vi) exploring the use of weakly-supervised learning to identify early tumor-indicating morphological changes in benign prostate biopsies.

Link to Diva