Fig. 1: Overview of the STAIG framework.

a STAIG begins with spatial transcriptomic (ST) data. Each slice includes spots with spatial coordinates, gene data, and optional Hematoxylin and Eosin (HE) stained images. Image patches at these spots undergo noise reduction, including bandpass filtering, before being processed for image embeddings via the BYOL framework. Parallelly, an adjacency matrix is created using the spatial data of the spots. b For multiple slices, the image embeddings from each slice are vertically merged, forming an image embedding space where spots are distributed, with dotted lines indicating their Euclidean distances. c Adjacency matrices from each section are combined diagonally to form an integrated adjacency matrix. This matrix is then used to construct a graph, with gene expression data represented as node information. d For spots connected by edges, distances are calculated in the image embedding space. These distances are then transformed into probabilities of random edge removal using a SoftMax function. The original graph undergoes two rounds of edge random removal based on these probabilities, creating two augmented views. Subsequently, features of nodes in these graph views are randomly masked. e A graph neural network (GNN) processes these augmented views, concentrating on node-level differences. This step is encapsulated within a dotted box, emphasizing the neighbor contrastive strategy. f The derived embeddings from the GNN are then utilized for spatial domain identification and integration.