Multiview Spinal Fracture Detection in Radiographic Projections Using Graph-Based Convolutional Fusion of X-Ray Views

Abstract

This research proposes a high-accuracy graph-based deep learning framework for multiview spinal fracture detection, integrating anatomical graph construction, convolutional encoders, and attention-guided fusion across radiographic projections. Utilizing a hybrid CNN-GCN architecture, the system models vertebral structures as nodes within a spatially coherent graph, enabling topological reasoning and cross-view feature alignment. Evaluated on multiview datasets comprising AP, lateral, and oblique projections, the model achieved vertebral-level detection accuracies up to 96.3%, with F1-scores exceeding 0.903 across diverse fracture types, including compression, burst, and multisegment injuries. Compared to conventional single-view CNN approaches, this framework demonstrated a 5.1%–5.5% improvement in diagnostic accuracy and a 24% increase in sensitivity for subtle or complex fracture patterns. The inclusion of inter-view consistency constraints and multiscale fusion layers enabled robust generalization across varied patient anatomies and projection combinations. With inference times below 200ms per multiview study and end-to-end trainability, the model supports real-time deployment within radiology workflows. Its modular, view-invariant design ensures compatibility with evolving imaging protocols and scalability for large-scale clinical deployment.

Ähnliche Arbeiten