An explainable graph transformer framework integrating structural reasoning and visual attention for interpretable nail disease diagnosis

Abstract Automated nail-disease classification remains a challenging task as different nail disorders present subtle and overlapping variations in color, texture and morphology. This paper introduces GraphX, an explainable graph-transformer framework that improves the baseline Graph Vision Transformer by incorporating image-specific graph construction, graph-transformer fusion, and attention guidance during training. The Adaptive Graph Construction module constructs a sparse graph per image based on the cosine similarity of spatially and texture enriched patch representations and retains the eight most similar neighbors per patch. The Cross-Attention Fusion module allows reciprocal refinement between graph-based structural embeddings and transformer-based contextual tokens. The Explainability-Aware Attention module learns to match the attention distribution with continuous lesion-prior maps generated from local intensity contrast and color deviation, providing weak spatial guidance without dermatologist-annotated lesion mask. In classification, GraphX achieved 94.97% accuracy and macro-averaged one-vs-rest ROC-AUC of 0.98, which outperformed the baseline models evaluated. Explainability was measured with Grad-CAM visualization, normalized attribution entropy, and deletion–insertion faithfulness testing. The results demonstrate strong internal classification performance and provide preliminary evidence of model’s attention patterns, highlighting the potential of GraphX as a transparent framework for automated nail disease classification.