MicNet: integrating spatially resolved transcriptomes and pathology images by contrastive deep neural network

Recent breakthroughs in spatially resolved transcriptomic technologies have enabled molecular characterization of cells while preserving spatial and morphological contexts. However, integrating transcriptomic profiles and pathology images remains a challenge. Here, we developed a novel unsupervised representation learning method, MicNet, to project pathology image and transcriptomic data onto a shared representative domain for biological interpretation. MicNet maximizes the correlation between image and molecular features from the same sample while minimizing it for different samples. MicNet outperformed existing approaches in multiple analysis tasks, including spatial domain detection, spatially variable gene identification, and spatial organization visualization.