INTEGRATION OF HISTOPATHOLOGY AND SPATIAL TRANSCRIPTOMICS FOR CANCER CLASSIFICATION: A SYSTEMATIC REVIEW

Abstract

Background: Histopathology remains central to cancer diagnosis and classification, while spatial transcriptomics enables transcriptome-wide molecular profiling with preservation of tissue architecture. The integration of these modalities has created a new class of computational approaches capable of linking morphology with spatial gene expression, identifying tumor subtypes, delineating spatial domains, predicting molecular phenotypes, and potentially improving cancer classification. Objective: To systematically review studies integrating histopathological images with spatial transcriptomic data for cancer classification, tumor subtyping, spatial-domain identification, and prediction of clinically relevant molecular phenotypes. Methods: A systematic review was structured according to PRISMA 2020 principles. PubMed/MEDLINE, Scopus, Web of Science, Embase, IEEE Xplore, and Google Scholar were searched for studies that integrated histopathology or H&E images with spatial transcriptomic data in human cancer. Eligible studies included computational and translational investigations evaluating cancer classification, molecular subtyping, tumor region identification, spatial-domain detection, histology-to-transcriptome prediction, or clinically relevant classification tasks. Reviews, non-cancer studies, purely transcriptomic studies without histological integration, and studies without an identifiable classification or prediction component were excluded. Owing to substantial heterogeneity in cancer type, spatial-transcriptomic platform, histological representation, artificial-intelligence architecture, and performance metrics, findings were synthesized narratively. Results: A total of 1,126 records were identified. After removal of 287 duplicates, 839 records underwent title and abstract screening. Of these, 724 were excluded and 115 reports were sought for retrieval. Six reports could not be retrieved, leaving 109 full-text reports for eligibility assessment. After exclusion of 82 reports, 27 studies were included in the systematic review. Breast cancer was the most frequently evaluated malignancy, followed by brain tumors, pancreatic cancer, colorectal cancer, prostate cancer, and cutaneous squamous cell carcinoma. Deep learning predominated, including convolutional neural networks, vision transformers, graph neural networks, graph convolutional networks, and hybrid multimodal architectures. Across studies, integration of morphology and spatial gene expression improved spatial-domain delineation, molecular-subtype identification, tissue classification, and prediction of spatially resolved gene-expression patterns. Several studies also demonstrated the ability to infer clinically relevant molecular states from routine H&E images after training with spatial transcriptomic data. However, external validation, standardized benchmarks, and prospective clinical testing were limited. Conclusion: Integration of histopathology with spatial transcriptomics provides a powerful framework for cancer classification by combining morphological and molecular information within their native spatial context. Current evidence demonstrates strong potential for tumor subtyping, spatial-domain detection, histology-based transcriptomic prediction, and tissue classification, but clinical translation remains constrained by small cohorts, platform heterogeneity, limited external validation, high spatial-transcriptomics costs, and lack of standardized reporting. Larger multicenter studies and clinically oriented validation are required before routine diagnostic implementation.