Spatial transcriptomics data analysis integrates gene expression profiles with their corresponding spatial locations to identify spatial domains, infer cell-type dynamics, and detect gene expression patterns within tissues. However, the current spatial transcriptomics analysis neglects the multiscale cell-cell interactions that are crucial in biology. To fill this gap, we propose multiscale cell-cell interactive spatial transcriptomics (MCIST) analysis. MCIST combines the advantages of an ensemble of multiscale topological representations of cell-cell interactions in the gene expression space with those of cutting edge spatial deep learning techniques. We validate MCIST by a comparison of 14 cutting edge methods on a huge collection of 37 benchmark spatial transcriptomics datasets. We demonstrate that MCIST yields superior performance in spatial domain detection. It achieves the best clustering score on 23/37 datasets and is among the top three methods on 33/37 datasets, whereas the second best method scored only 6/37 and 17/37 on these measures, respectively. In terms of overall performance with regards to a quantitative metric, MCIST offers over an 11\% improvement to the previous state-of-the-art in spatial domain detection. Additionally, MCIST offers multiscale insights with respect to trajectory inference, differentially expressed gene detection, and signaling pathway enrichment analysis. Our MCIST sheds valuable light on the multiscale perspective in spatial transcriptomics.