A Latent Diffusion Model With Spatial Attributes for 3D Reconstruction in Biplanar X-ray Images

Computed tomography (CT) imaging provides essential 3D anatomical information for diagnosis and treatment planning, yet its clinical use is often limited by radiation exposure and equipment constraints. Given the accessibility and low radiation of biplanar X-ray imaging, 3D CT reconstruction from two orthogonal projections is an appealing alternative. However, the inherent 2D-to-3D ambiguity under extreme view sparsity makes it an ill-posed inverse problem, often resulting in geometric inconsistencies and degraded anatomical fidelity. To address this challenge, we propose a Spatial Attribute-aware Latent Diffusion Model (SA-LDM) that encodes deterministic X-ray geometry as structured spatial priors to guide 3D reconstruction. A Cross-View Feature Interaction (CVFI) module establishes semantic correspondences between orthogonal projections, while Multi-Scale Attention (MSA) layers extract hierarchical anatomical cues to enhance structural consistency. A Spatial Attribute-aware Feature Aggregation (SAFA) module further integrates spatial geometric attributes for geometry-aware multi-view fusion. Operating in a compact latent space with a Discrete Wavelet Domain (DWD) loss, SA-LDM accelerates convergence and preserves fine anatomical details while maintaining high computational efficiency. Through this design, SA-LDM reconstructs anatomically coherent and detail-preserving 3D volumes from only two orthogonal X-rays, effectively mitigating the geometric inconsistencies and structural distortions inherent to sparse-view reconstruction. Extensive experiments on lung, pelvis, and knee datasets demonstrate state-of-the-art performance across both quantitative and perceptual metrics, indicating that SA-LDM is able to generate anatomically faithful 3D reconstructions under extreme view sparsity.