Enhanced Vertebra-GNN: A Graph-Based Deep Learning Framework for Anatomically Informed Spinal Disease Classification

Accurate classification of spinal pathologies from radiographic images is essential for timely diagnosis and effective treatment planning. In this study, we propose an end-to-end deep learning framework that integrates convolutional neural networks (CNNs) with graph neural networks (GNNs) to model both the appearance and anatomical relationships of vertebrae from lateral spine X-rays. The proposed Enhanced Vertebra-GNN utilizes CNN-extracted regional features from segmented vertebral regions-of-interest and encodes inter-vertebral dependencies using a graph representation of the spine. Three variants of the GNN architecture-GCN, GAT, and Graph Transformer-were explored to optimize relational feature learning. Experimental results demonstrate the superiority of our model over conventional CNN baselines, with substantial gains in F $_1^{}$ -score and consistent accuracy across all classes. Visual interpretability analyses via Grad-CAM and attention mapping further confirm the model's ability to focus on clinically relevant regions. These findings suggest that Enhanced Vertebra-GNN offers a reliable, interpretable, and anatomically aware tool for automated spine assessment.