The spatial organization of tissues plays a fundamental role in shaping disease pathology, progression, and therapeutic response. Capturing these spatial relationships at single-cell resolution is essential for uncovering disease mechanisms and advancing diagnostic and treatment strategies. Although single-cell RNA sequencing (scRNA-seq) has transformed our ability to capture cell-type diversity and dynamic cell states, its dissociative nature removes spatial information critical for studying tissue organization and cellular interactions. In recent years, spatially resolved omics technologies have emerged, enabling molecular profiling while preserving the spatial tissue architecture. However, these platforms remain constrained by high costs, limited resolution, and restricted tissue capture areas, hindering detailed cellular-level analysis and inhibiting scalability for large studies. In contrast, hematoxylin and eosin (H&E) stained images offer broad spatial coverage and detailed morphology, are accessible and cost-effective, but lack molecular information, depend on expert interpretation, and often contain regions of variable quality that can impact computational workflows. To address these challenges, this dissertation presents three novel methods that collectively improve the computational utility of histology images and leverages machine learning to integrate information across modalities for enhanced spatial analysis. First, we developed a quality-control workflow that systematically quantifies histological features to detect non-informative and poor-quality regions, improving image preprocessing and enhancing downstream spatial omics and digital pathology tasks. Second, we developed a spatial integration method that overcomes the resolution constraints and capture size limitations of conventional spatial transcriptomics (ST) platforms, enabling high-resolution gene expression prediction and annotation of cellular-level tissue architecture across expansive tissue landscapes. Lastly, we developed a deep learning algorithm that reconstructs the spatial locations of dissociated single cells using learned spatial gene expression patterns from reference ST data. Together, these methods enable integrative spatial analysis of tissue architecture, cell-type distributions, and molecular profiles across a range of disease states, providing scalable and generalizable tools for dissecting tissue heterogeneity in complex biological systems.