Deep learning-based radiomics and pathomics for decoding tumor microenvironment and predicting immunotherapy outcomes in gastric cancer

Abstract

Although immune checkpoint inhibitors have reshaped the therapeutic landscape of gastric cancer, durable responses are achieved in only a minority of patients. The tumor microenvironment, particularly immune-cell composition and spatial organization, critically modulates cancer progression and the efficacy of immune checkpoint inhibitors. Nevertheless, the lack of clinically accessible approaches for spatially resolving the tumor immune contexture, whether by direct pathological quantification or indirect radiomic inference, has hindered a mechanistic understanding of how it dictates immunotherapy outcomes. Deep learning-based radiomics and pathomics extract high-dimensional phenotypic information from radiological and pathological images, thereby characterizing tumors at macroscopic and microscopic scales. Each modality, however, captures only a partial view of the tumor microenvironment. Multimodal models can integrate complementary imaging features associated with tumor heterogeneity and spatial immune organization, thus improving predictive performance and linking multimodal features to immune-related biological processes. In this review, we summarize recent advances in deep learning-based radiomics, pathomics, and multimodal frameworks for decoding the immune microenvironment and predicting immunotherapy outcomes in gastric cancer. We also discuss their potential to complement established biomarkers and support personalized treatment decisions. Finally, we highlight key challenges, including limited data availability, interinstitutional variability, insufficient external validation, poor interpretability, and barriers to clinical implementation. Addressing these limitations may accelerate the development of robust imaging biomarkers for guiding immunotherapy decisions in gastric cancer.