As both single-cell and spatial transcriptome techniques have certain shortcomings, techniques integrating single-cell transcriptome and spatial transcriptome were developed. In order to improve the similarity of mapping single-cell matrix to spatial matrix and reduce the loss function value, the RK-Tangram algorithm was developed by improving the objective function of the Tangram algorithm for deep learning and also correcting the gradient value of the optimization algorithm Adam, inspired by the Runge-Kutta method. Applying it to three sets of simulated data and real data from mouse brain cortical, motor and visual regions, compared with the original Tangram algorithm, the results showed that the RK-Tangram algorithm not only improved the similarity of the mapping and reduced the loss function value, but also extended the genome-wide mapping of the spatial transcriptome and corrected spatial measurements with low-quality. In addition, deconvoluting the spatial transcriptome’s data to single cell’s provided a higher resolution mapping of tissue types.