Diffusion-based novel view synthesis for X-ray imaging

X-ray imaging is extensively used in clinical practice due to its low cost, accessibility, and relatively low radiation dose. However, its intrinsic two-dimensional (2D) projection mechanism inevitably leads to severe structural overlap and lesion occlusion in single-view acquisitions, substantially limiting the completeness and reliability of anatomical interpretation.

To overcome the loss of spatial information caused by limited viewing angles, this study proposes a diffusion-based novel view synthesis framework for X-ray imaging that directly generates target-view X-ray images from a small number of reference views, without relying on explicit three-dimensional reconstruction.

Built upon the Stable Diffusion framework, the proposed method introduces a lightweight and physically interpretable angular-difference conditioning mechanism to explicitly model cross-view geometric relationships. By embedding the relative angular offset between reference and target views into the attention-based feature interaction process, the model effectively captures view-dependent structural transformations and synthesizes anatomically consistent novel-view X-ray images.

Extensive experiments on a synthesized multi-view X-ray dataset demonstrate that the proposed approach achieves robust and stable performance under sparse-view input conditions, significantly outperforming existing methods in terms of view generalization and structural preservation. In particular, for target view offsets of $-{30}^{\circ }$ , $+{45}^{\circ }$ , and $-{60}^{\circ }$ , the proposed method achieves PSNR values of $22.98$  dB, $22.99$  dB, and $22.10$  dB, respectively, together with consistently improved SSIM and LPIPS scores.

These results indicate that the proposed framework provides an efficient and flexible solution for supplementing missing X-ray views in clinically constrained imaging scenarios.