Self-Supervised Learning for Genetically Relevant Domain Identification in Morphological Images

Spatial-omics techniques are used for profiling the local gene expression of a sample, while imaging is typically used for capturing the phenotype, in turn determined by the gene expression. However, the correlation between them may not necessarily be direct. Tissue morphology can also reflect genes expressed during development, or gene expression may not yet have resulted in tissue restructuring. Thus, recent efforts have shown that integration of spatial-omics with imaging provides additional information of biological relevance. In this work, we show that morphological feature extraction on H&E images using self-supervised learning is more versatile for capturing relevant tissue domains compared with the previously proposed approaches. Furthermore, self-supervised learning allows the novel use of DAPI for domain identification, which enables morphological integration also for experiments where H&E staining is not available.