High-Resolution 3d Ct Synthesis From Bidirectional X-Ray Images Using 3d Diffusion Model

3D Computed Tomography (CT) offers invaluable geometric insights into bone structures, but the high radiation dose and medical cost constraints are significant barriers. Moreover, CT reconstruction demands multiple X-ray projections, necessitating a dedicated scanning system, whereas bidirectional X-rays are already the front-line diagnostic tool in routine practice. Therefore, reconstructing 3D bone structures from bidirectional X-ray data can reduce the need for additional CT scans, provide rapid access to 3D information, and lower medical costs. Recently, diffusion models have emerged as potent tools for generating high-fidelity images. However, high computational costs have limited their utility. Collecting large-scale datasets for training in clinical environments presents another challenge.In this study, we introduce a novel approach to synthesize 3D CT volumes from a bidirectional X-ray projection using a 3D diffusion model. To reduce the computational burden and the need for a large dataset, our 3D diffusion model was trained using patch-wise loss. A conditional score function of our model incorporates 2D bidirectional X-ray images and patch coordinate information to synthesize high-resolution CT. Initial findings indicate that our diffusion model synthesizes 3D CT volumes from a bidirectional X-ray, effectively capturing 3D geometric correlations while enabling single-GPU training and rapid 3D volumetric sampling.