stDGCN: A dual-augmentation graph convolutional network for identifying spatial domains with attention mechanism

Spatial transcriptomics provides high resolution insights into gene expression within tissue architecture. However, accurate spatial domain identification remains challenging due to data sparsity and noise. In this study, a dual-augmentation graph convolutional network (stDGCN) is proposed for robust spatial domain identification. The framework introduces a dual-augmentation strategy that combines spatial neighborhood enhancement and Gaussian perturbation to generate diverse views of gene expression profiles. The original, spatially augmented, and Gaussian noise augmented views are encoded by a shared graph convolutional network, while an attention-based fusion module adaptively integrates the original and spatially enhanced embeddings. To improve representation stability, consistency regularization is employed to align embeddings derived from the original and augmented views. Furthermore, a zero-inflated negative binomial decoder is incorporated to reconstruct gene expression profiles while accounting for the sparsity and over dispersion inherent in spatial transcriptomics data. Extensive experiments across six spatial transcriptomics datasets, including human brain, breast cancer, mouse visual cortex, mouse brain and embryo data, demonstrate that stDGCN consistently outperforms seven state-of-the art methods in spatial domain identification, achieving superior clustering accuracy and robustness. Overall, stDGCN provides a robust and interpretable framework for spatial domain analysis, with potential applications in downstream biological and clinical studies.