Spatial omics technologies enable high-resolution molecular mapping of tissues, but their limited throughput, high cost, and lack of standardization constrain population-scale studies and clinical translation. In parallel, cell type annotation of whole-slide H&E histology remains labor-intensive, limiting scalability and reproducibility. Here we present MeowCat (Multi-resolution Omics-informed Whole-slide Cell Annotation Tool), a computational framework that generates single-cell-resolution spatial cell type maps from routine H&E histology, unifying information from fragmented multi-resolution, multi-slide spatial omics data across platforms.Trained and validated on lung adenocarcinoma datasets comprising 238,488 spots and 2,047,381 cells, MeowCat achieved an AUROC of ~0.95 for major cell type prediction and generalized robustly across cancerous and non-cancerous tissues profiled by diverse spatial platforms. Applied to three lung cancer cohorts spanning over 1.5 billion cells, MeowCat enabled population-scale analyses of tumor progression, multi-scale immune niche organization, 3D tissue reconstruction, and automated lymphocyte spectrum profiling.MeowCat further demonstrated prognostic utility through interpretable graph neural network modeling, revealing that lymphocyte spatial organization and connectivity with proximal B-cell and myeloid populations are associated with favorable clinical outcomes. Together, MeowCat bridges boutique spatial omics studies with population-scale pathology and advances translational AI for precision medicine.

Learning Objectives:

1. Explain how multi-resolution spatial omics data can train deep learning models to predict cell types from H&E histology
2. Assess how H&E-predicted cell type maps enable population-scale tumor immune niche analysis to inform diagnosis and prognosis
3. Identify how H&E-predicted cell type maps enable full-spectrum lymphocyte characterization associated with clinical outcomes