Spatial transcriptomics enables high-resolution mRNA quantification while preserving cellular spatial context. Accurate spatial domain identification is crucial for downstream analysis, yet current methods often fail to effectively integrate gene expression with spatial information or leverage multi-scale contrastive learning. We present MCST, a novel deep learning framework that combines graph convolutional networks with multi-scale contrastive learning to learn informative latent embeddings by jointly modeling spatial and transcriptional data. MCST constructs spatial neighborhood graphs, integrates gene expression through GCNs, and employs multi-scale contrastive learning to refine node representations. Comprehensive evaluations across clustering, trajectory inference, and SVG detection tasks demonstrate MCST’s superior accuracy and robustness in spatial domain identification, establishing it as a powerful tool for spatial transcriptomics analysis.