3D Reconstruction of X-Ray Skeletal Images

This study presents a comprehensive three-stage deep learning framework for monocular three-dimensional reconstruction of X-ray skeletal images, addressing critical challenges in medical image interpretation and diagnostic accuracy. The proposed pipeline integrates state-of-the-art neural network architectures to transform single-view X-ray images into high-quality 3D visualizations within approximately 45 seconds. The first stage employs the Cbc-SwinIR (Coordinate-based Convolution - Swin Transformer Image Restoration) model for 4× super-resolution enhancement, utilizing shallow feature extraction with channel pruning and deep Swin Transformer blocks to significantly improve image clarity and detail preservation. The second stage implements the MAXIM (Multi-Axis MLP for Image Processing) model to remove halo, scatter, and beam hardening artifacts through a UNet-like architecture with multi-axis gated MLP blocks, followed by precise target segmentation using the Semantic-SAM model with multi-granularity Hungarian matching-based optimization. The final stage utilizes the One-2-3-45 model for monocular 3D reconstruction, combining depth map generation, Delaunay triangulation, and neural signed distance field optimization to generate anatomically faithful three-dimensional meshes. Comprehensive validation across multiple viewpoints demonstrates outstanding performance with average SSIM, PSNR, and MSE, confirming the framework's effectiveness for clinical applications in surgical planning, medical education, and diagnostic decision support while maintaining computational efficiency and non-invasive operation principles.

Keywords: X-ray Image Enhancement, Monocular 3D Reconstruction, Deep Learning Medical Imaging, Semantic Segmentation, Swin Transformer