Spatial transcriptomics is an innovative technology that simultaneously preserves the spatial information and gene expression profiles within tissue samples, bridging the gap between tissue morphology and molecular biology. However, the low resolution of current high-throughput spatial transcriptome techniques poses limitations for downstream biological analysis, as each sequencing spot may contain mixed gene expression profiles from multiple cell types. Therefore, cell type deconvolution has emerged as a crucial task in spatial transcriptomics, aiming to infer cellular composition and spatial distributions. Deep learning (DL), with its powerful feature representation learning capabilities, has been utilized for deconvolution to address key issues, such as cell type feature extraction, platform effect mitigation, spatial information incorporation, and cell type identification. This review provides an overview of DL-based deconvolution methods, highlighting their core mechanisms. We further discuss current limitations and potential improvements to offer insights into enhancing the robustness and applicability of DL-based deconvolution methods.