Self-supervised deep learning for H&E-stained histopathology in a spatial omics context

Spatial omics is an emerging discipline that integrates molecular profiling with the spatial localization of physical tissues. A key task in spatial omics data analysis is precise cell segmentation, which is essential for accurately mapping spatial omics data onto segmented cells. This step enables the incorporation of functional diversity information within tissues, facilitating deeper biological insights. This thesis presents a novel self-supervised learning (SSL) framework for cell segmentation in histopathology images. The proposed approach addresses key challenges in digital pathology, particularly the lack of extensive manual annotations, by leveraging a dense Siamese network architecture. The model optimizes global, dense, and level-wise embeddings, enabling the learning of pixel-level representations that capture both high-level contextual information and fine-grained spatial details. This enhances the accuracy of downstream cell segmentation tasks. The effectiveness of SSL in learning domain-specific representations without the need for extensive labeled data is demonstrated. The proposed framework enhances model robustness against variable staining protocols and common artifacts in histological datasets. Furthermore, the importance of reliable ground truth for segmentation evaluation and the careful selection of image transformations to generate robust embeddings is highlighted. To further contextualize the segmentation results, a spatial data analysis framework has been developed to integrate histopathological image morphology with spatial transcriptomics data from spatial omics datasets. This integration bridges self-supervised cell segmentation with spatial omics data, facilitating the discovery of biologically meaningful insights. Overall, this work presents an innovative approach to cell segmentation, establishing a foundation for future advancements in self-supervised learning for histopathology. By linking morphological features with spatial data, this method has the potential to drive progress in spatial omics research and precision medicine.