Fig. 1
From: Image guided construction of a common coordinate framework for spatial transcriptome data
Schema representation of STaCker. The workflow (a) takes as inputs the tissue images from a pair of reference and moving spatial transcriptome slices, combined with the contour maps generated upon the gene expression profiles from the respective slices. The resulting composite images are subsequently aligned through a deep neural network-based registration module. The registration module outputs the inferred deformation field to align the spatial coordinates of the spots/cells in the moving slice. The architecture of the registration module (b) takes a four-level contracting path and a four-level expanding path with skip connections at all levels. The final layer of the decoder is further convoluted to generate the spatial velocity field followed by a vector integration to output the deformation field for the alignment. Synthetic images with segmentation label maps (Methods) are used to train the module. The moved label map, after applying the deformation field to the moving label map, is compared to the reference label map. The difference constitutes the key component in the loss function.
