PFDFuse: A physics-feature-guided diffusion model for neutron and X-ray image fusion

Neutron imaging and X-ray imaging are highly complementary for nondestructive testing (NDT). Combining these two modalities enables more comprehensive and reliable inspection of complex industrial samples and is therefore of clear practical relevance. However, most existing fusion methods do not sufficiently exploit the modality-specific differences in imaging physics, which limits the fused images in terms of detail preservation and information representation. Building upon the denoising diffusion model for multimodality image fusion (DDFM), this study introduces targeted modifications to the reverse diffusion sampling process and proposes a physics-feature-guided diffusion fusion approach, termed PFDFuse. Specifically, PFDFuse replaces the original spatially uniform update weight with a pixel-level imaging-physics update weight, which regulates the update magnitude at each reverse diffusion time step. This design encourages the fusion process to rely more on X-ray information around high-atomic-number metallic boundaries, while assigning greater emphasis to neutron information in hydrogen-rich and other low-atomic-number regions. Comparative experiments on real samples demonstrate that the proposed method achieves consistently better overall performance than multiple deep learning-based fusion approaches, including DDFM, indicating its effectiveness in improving neutron–X-ray fusion quality and its potential for practical engineering deployment.