Spatial transcriptomics (ST) has revolutionized tissue-level gene expression analysis, but high costs and data scarcity limit its widespread application and necessitate data integration from multiple sources, which contain strong batch effects. We present a novel approach to address these challenges, combining (1) stain normalization and augmentation across multiple color spaces; and (2) incorporation of pathology foundation models pre-trained on diverse histopathological datasets. Our method improves the performance and generalizability of the iStar, which predicts near-single-cell spatial gene expression by integrating ST data with histology images. By leveraging hierarchical image feature extraction using foundation models, our approach captures both local and global tissue structures, mimicking the process of pathological examination. Notably, foundation models trained on large, diverse histology image datasets prove robust against batch effects. We evaluate our method using various ST platforms, demonstrating significant improvements in cell type segmentation and out-of-sample gene expression prediction. Our model addresses the complex nature of batch effects in ST data analysis while offering a cost-effective solution to expand the utility of existing ST datasets. By enhancing tissue characterization accuracy, our AI-driven approach advances precision medicine, potentially improving cancer diagnosis and treatment selection.