MSR-Net: Mamba-Based Symmetric Architecture for 3D Spine Reconstruction from 2D Orthogonal X-ray Images

Accurate 3D spine reconstruction is crucial for diagnosing spinal diseases, but traditional X-ray imaging is limited by tissue overlap, which blurs detailed 3D anatomical structures. While CT imaging provides a 3D perspective, it is associated with high radiation doses, significant costs, and the potential distortion of the spine’s natural curvature due to the supine scanning position. Therefore, reconstructing a 3D spine model from 2D X-ray images offers a more cost-effective approach and has garnered significant interest in clinical applications. As a 2D-to-3D reconstruction problem, previous work has integrated information across multiple 2D views in latent space for efficient training. However, this cross-dimensional feature transformation can result in the loss of valuable signal information during latent encoding, reducing the correlation between the output and the corresponding 2D input. This remains a challenge in complex anatomical structures and datasets with significant variability. To address this, we propose MSR-Net, which establishes a symmetric architecture for reconstruction that retains information without requiring prior knowledge by duplicating the 2D views to match the 3D depth.We introduce the GL-Mamba module to perform multi-scale global mapping of X-ray image features after dimensional expansion, while effectively enhancing the resistance of spinal bone local pixels to forgetting. A dynamic discriminator strategy is also introduced to refine model granularity by evaluating global and local information. Our framework demonstrates improved expressiveness in the reconstructed 3D spine models (including spinous processes and bone surface textures) on the training dataset. It also generates highly correlated output results on unseen datasets, showcasing its robust generalization capabilities and clinical potential.