BrainGraphNet-ViT model: a hybrid model combining vision transformers and graph convolutional networks for brain tumor diagnosis

The prognosis for brain tumor patients heavily depends on early and precise diagnosis, a critical challenge in global health that aligns with Sustainable Development Goal 3 (SDG 3), which aims to ensure healthy lives and promote well-being for all. Conventional Convolutional Neural Networks (CNNs) often fall short at modeling global contextual information and the intricate anatomical architecture, both of which are crucial for precise tumor characterization in MRI data. To address this limitation, this paper introduces BrainGraphNet-ViT, a novel hybrid framework that synergistically combines a Graph Convolutional Network (GCN)—built upon a Graph Transformer operator—with advanced vision architectures. By integrating models like Vision Transformer (ViT), EfficientNet-B0, ResNet50, and VGG19, the framework leverages both topological relationships and hierarchical visual features. Through a dual-stream design that processes MRI scans via complementary pathways for visual feature extraction and graph-based structural reasoning, our model achieves a comprehensive representation of both local tumor characteristics and global contextual information. Extensive assessment on the Brain Tumor MRI Dataset confirms the exceptional performance of our proposed BrainGraphNet-ViT model, which attained a peak classification accuracy of 99.31%, substantially surpassing other hybrid models and established state-of-the-art methods while delivering consistent results across all tumor classes. By effectively fusing structural reasoning with visual feature extraction, BrainGraphNet-ViT establishes a new paradigm for brain tumor diagnosis, offering enhanced reliability and interpretability that paves the way for sophisticated computer-assisted diagnosis systems in neuro-oncology.