Systematic review of transformers, graph neural networks, and federated learning in oncology: Applications, challenges, and pathways to clinical translation

Background Advanced deep learning models, including transformers, graph neural networks (GNNs), and federated learning (FL), are reshaping oncology by enhancing the analysis of complex, multimodal cancer data for diagnostic and prognostic purposes. Despite their promise, challenges in explainability, generalizability, and clinical readiness hinder widespread adoption. Methods This systematic review aims to evaluate the applications of transformers, GNNs, and FL in cancer diagnosis and prognosis, identify key technical challenges, and outline strategies for clinical translation. A comprehensive search across PubMed, Scopus, IEEE Xplore, Google Scholar, Web of Science, arXiv, and medRxiv identified 1004 articles, with 16 systematic reviews meeting inclusion criteria after rigorous screening. Results Data were extracted on model applications, cancer types, data modalities, and technical challenges, and the quality was assessed using the PROBAST tool. Transformers excel in image-based diagnostics, such as skin cancer detection, and multimodal data integration. GNNs are effective for modeling biological networks, aiding in cancer subtype classification and driver gene prediction. FL enables privacy-preserving, multi-institutional model training for cancers like breast, lung, and prostate. Major challenges include limited explainability, poor generalizability across diverse cohorts, data heterogeneity, and insufficient external validation. Conclusions Transformers, GNNs, and FL offer transformative potential in oncology but face significant hurdles in interpretability and generalizability. Future progress requires interdisciplinary frameworks, transparent benchmarking, and robust validation to ensure clinical trust and impact.