PVT-FractureNet: A Pyramid Vision Transformer Model for Radiographic Bone Fracture Classification

Bone fracture classification from radiographic images plays a critical role in orthopedic diagnosis and treatment planning, yet manual interpretation remains time-consuming and prone to inter-observer variability. Traditional CNN approaches often have difficulty capturing fine-grained fracture characteristics as well as broader skeletal structural patterns, limiting their ability to accurately classify multiple fracture types. To overcome this limitation, this study proposes PVT-FractureNet, a deep learning framework based on the Pyramid Vision Transformer (PVT) architecture designed for holistic radiographic-based categorization of bone fractures. The primary contribution offered by this study is the development of a transformer-driven hierarchical architecture that proficiently combines multi-level feature representations with explainable attention visualization to enhance diagnostic accuracy and interpretability. The model was trained and evaluated on a Kaggle dataset comprising ten fracture categories, including avulsion, comminuted, greenstick, hairline, impacted, longitudinal, oblique, pathological, spiral, and fracture-dislocation. Preprocessing steps included normalization, resizing, and data augmentation, followed by feature extraction and classification using multi-head self-attention and spatial-reduction mechanisms. Experimental results demonstrated that PVT-FractureNet achieved an average accuracy of 89.9%, specificity of 91.1%, and AUC of 0.811, with the highest performance observed in Greenstick and Fracture-Dislocation classes (AUC > 0.90). Grad-CAM and Score-CAM visualizations further showed that the model precisely pinpointed fracture regions that align with clinically meaningful anatomical cues. In conclusion, PVT-FractureNet demonstrates robust generalization capability, clear interpretability, and dependable diagnostic performance, establishing it as a promising framework for automated transformer-based bone fracture classification in clinical radiology.