Multi-Sequence MRI Thoracic Spine Compression Fracture Classification Based on Local-Global Feature Selection and Cross-Attention Fusion

Thoracic vertebral compression fracture (VCF) is a common fracture type in clinical practice, which seriously affects the quality of life of patients. The benign and malignant classifications of VCF are of significant diagnostic importance for the selection of clinical treatment options. Magnetic resonance imaging (MRI), as the main diagnostic tool for VCF, has the characteristics of multi-sequence imaging, and the structural and tissue information provided by different sequences complement each other. This paper proposes a multi-sequence MRI feature fusion network CAMSF based on vision transformer (ViT), which solves the problems of limited local feature extraction of ViT and insufficient interaction between multi-sequence image features. This method combines a multi-head self-attention mechanism and a deep convolution module to effectively capture global and local features, and uses a graph pooling-based feature selection method to optimize redundant features. In addition, a fusion method based on the cross-attention mechanism is used to deeply interact with different MRI sequence features to obtain complementary information across sequence images. Experimental results on the thoracic VCF dataset show that the classification accuracy of CAMSF reaches 84.88% and the AUC reaches 83.39%, which is significantly better than the existing ViT-based multimodal image feature fusion method. The ablation experiment further verifies the effectiveness of each module, indicating that CAMSF can effectively fuse multi-sequence MRI features and provide an effective solution for the accurate classification of VCF.