GraphViT: A graph-integrated vision transformer for accurate and explainable diabetic foot ulcer classification

Diabetic Foot Ulcers (DFUs) present a critical global health challenge, where early and precise multi-class identification of infection and ischemia is vital to preventing limb amputation. While deep learning has advanced automated DFU analysis, existing models often struggle to simultaneously capture global contextual information and the fine-grained, local relational patterns necessary for complex multi-class discrimination. To bridge this gap, we propose GraphViT , a novel hybrid architecture that synergistically integrates the Vision Transformer (ViT) with Graph Neural Networks (GNNs). The novelty of GraphViT lies in its dual-stream reasoning: the ViT backbone extracts high-dimensional global semantic features, while a specialized multi-layer GraphConv-GAT module re-interprets these features through graph-based reasoning. By organizing image embeddings into k-nearest neighbor similarity graphs, the model explicitly propagates information across semantically related regions, regardless of their spatial distance. Extensive benchmarking on the DFUC2021 dataset demonstrates that GraphViT significantly outperforms state-of-the-art CNNs, transformers, and prior hybrid models, achieving an accuracy of 90.36% and an F1-score of 90.27%. Furthermore, we introduce an Explainable AI (XAI) framework using Grad-CAM and LIME, providing clinical transparency by localizing pathological regions with high consistency. Our results confirm that the progressive integration of graph-based relational reasoning provides a superior paradigm for accurate and interpretable diabetic foot ulcer classification.