Spatial transcriptomics enables studying cellular interactions by measuring gene expression in situ while preserving tissue context. Within tissues, distinct cellular niches define microenvironments that influence cell states and function. A fundamental task in spatial transcriptomics is identifying differentially expressed genes within a specific cell type across different niches to quantify context-dependent cell state variation. Despite advances in cell segmentation algorithms, the persisting problem of the wrong assignment of molecules to cells can obscure the analysis by introducing spurious differentially expressed genes that originate from neighboring cells rather than the group of interest. Here, we introduce 2DE, a probabilistic framework designed to refine spatial differential expression analyses by filtering out genes that are overexpressed due to local contamination rather than true cell-intrinsic expression. 2DE operates downstream of any differential expression method, filtering irrelevant genes by considering gene overexpression relative to the expression in the neighborhood and returning marker confidence scores. In a study of human breast cancer, we demonstrate that 2DE improves the precision of the discoveries.