In this paper, we propose a novel diffusion model, PWD (Prior information embedding and wavelet feature fusion fast sampling diffusion model), for medical image reconstruction in limited-angle CBCT (LACT). To solve the slow inference speed problem of the existing diffusion model, PWD enables efficient sampling through prior information embedding and wavelet feature fusion. In the training phase, the structural correspondence between the LACT image and the fully sampled image is learned, and in the inference phase, the LACT image is used as prior information to achieve high-quality reconstruction with a small sampling step. Multi-scale feature fusion in the wavelet domain improves the reconstruction of microstructures. Experimental results using clinical dental arch CBCT and apex image datasets show that PWD outperforms the existing methods in PSNR and SSIM indices. In particular, PSNR is improved by more than 1.7 dB and SSIM by more than 10% with only 50 samplings.
