Applied graph neural networks: Domain-driven insights from medicine to remote sensing

Abstract

Graph Neural Networks have emerged as a leading paradigm for processing data with irregular, graph-based structures. Their capacity to capture intricate interdependencies makes them particularly well-suited for complex domains such as medical image analysis, remote sensing, and privacy-centric computing. This paper presents a comprehensive review of 52 scholarly articles published between 2017 and 2025, focusing on the practical implementation of Graph Neural Networks across five key fields. Among the architectures examined, Graph Convolutional Networks dominate (60 %), followed by Graph Attention Networks (35 %), with the remainder comprising hybrid and domain-tailored models (22 %). These percentages are calculated based on the frequency of model usage across the reviewed 40 studies. Since several studies employed more than one Graph Neural Network architecture, the percentages may overlap and sum to more than 100 %. Although a total of 52 articles were reviewed during the mapping process, only 40 studies met all eligibility and quality criteria and were therefore included in the quantitative analysis. All percentage values reported in the abstract are calculated based solely on these 40 finalized studies. The remaining 12 articles were part of the broader literature survey but were excluded from statistical computation because they did not satisfy the inclusion criteria. To enable robust cross-domain assessment, we introduce an original evaluation framework composed of nine practical dimensions, including metrics like predictive accuracy, model interpretability, computational efficiency, resilience to noise, and support for real-time operation. The analysis highlights the superiority of Graph Convolutional Networks in hyperspectral imagery tasks, while Graph Attention Networks show growing success in detailed medical diagnostics due to their attention mechanisms. Unlike earlier reviews that focus on theoretical progress, this study emphasizes the effectiveness of real-world models under deployment conditions, with a focus on reproducibility, domain constraints, and scalability. We conclude by outlining future research priorities, such as the design of resource-efficient Graph Neural Networks for embedded systems and the creation of unified benchmarks to evaluate graph learning across multiple domains.