Graph Neural Networks in Medical Imaging: A Systematic Review (Preprint)

Hafsa Akebli, Vincenzo Della Mea, Kevin Roitero

Abstract: Background: Graph Neural Networks (GNNs) have attracted growing interest in medical imaging due to their ability to model relational and non-Euclidean structures arising from anatomical organization and structured spatial dependencies within medical images. In recent years, GNN-based approaches have been applied across multiple imaging modalities and clinical tasks. However, existing reviews are largely limited to specific imaging domains or modalities, while others adopt a broader healthcare perspective in which medical imaging is not the primary focus, resulting in the absence of a systematic, imaging-centered synthesis. Objective: This study aims to systematically review the application of GNNs in medical imaging, with particular focus on graph construction strategies, the GNN architectures employed, the imaging modalities, anatomical targets, and disease contexts investigated, as well as the approaches used to integrate multimodal information within graph-based frameworks. Methods: We conducted a systematic literature review in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. A structured search was performed across IEEE Xplore, ACM Digital Library, and PubMed. After applying predefined inclusion and exclusion criteria, 89 studies were retained for in-depth analysis. Each study was examined and categorized according to imaging modality, biomedical task, anatomical and disease focus, graph construction strategy, GNN architecture, and use of multimodal data. Results: The reviewed literature reveals a rapid growth of GNN applications in medical imaging, with neuroimaging and histopathology emerging as the most extensively studied domains. Graph construction strategies vary widely, ranging from region-of-interest and patch-level representations to population-level similarity graphs, often driven by spatial, functional, or morphological relationships. Across the 89 included studies, GNNs are most commonly applied to disease classification, prognosis and survival prediction tasks, with Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs) being the most frequently adopted architectures. Most studies remain unimodal, but multimodal approaches are also present, often combining multiple imaging modalities or integrating imaging with clinical data within graph-based formulations. Conclusions: This systematic review provides a PRISMA-compliant, imaging-centered synthesis of GNN applications in medical imaging. By analyzing graph construction strategies, GNN architectures, imaging modalities, multimodal settings, anatomical targets, disease contexts, and clinical tasks, the review offers a structured overview of how graph-based models are currently designed and applied across medical imaging domains. The findings reveal a field that is rapidly expanding but methodologically heterogeneous, highlighting the importance of clearer design choices and reporting practices, particularly with respect to graph representations and multimodal integration, to enable meaningful comparison across studies and support future translational research.

External IDs:doi:10.2196/preprints.89873