<p>Spatial transcriptomics (ST) has redefined our exploration of tissue-level cellular heterogeneity and spatial architecture; yet accurately pinpointing functional regions within complex, high-dimensional datasets remains a pressing hurdle. To tackle this, we introduce DSSMST, a Deterministic State Space Model for Spatial Transcriptomics, as a self-supervised learning (SSL) framework that integrates a Deterministic State Space Model (DSSM), graph neural networks (GNNs), and contrastive learning. Central to DSSMST is the DSSM module, whose robust dynamic modeling capacity enables it to capture spatial gradient variations and continuous dependencies in ST data—overcoming the limitations of static graph-based approaches—thereby establishing a solid basis for precise spatial domain identification. Complementing this, a customized self-supervised contrastive learning mechanism refines the latent embedding space, empowering the model to better distinguish between subtly differing spatial domain features. This integration effectively elevates the overall accuracy of spatial domain delineation. We assessed DSSMST on multiple representative ST datasets using diverse metrics. Experimental findings reveal that DSSMST achieves leading spatial domain identification accuracy and maintains competitive robustness and generalization across multiple datasets, underscoring its strong potential for advancing ST research. The source code, tutorials, and reproducibility instructions are publicly available at <a href="https://github.com/JiruiZhang/DSSMST">https://github.com/JiruiZhang/DSSMST</a>.</p>