LCEDiff: Physics-inspired generative modeling in the sinogram domain for sparse-view CT reconstruction

We present a generalized framework for sparse-view CT reconstruction using a generative diffusion model based on the Local Correlation Equation (LCE). In particular, due to the complex data distribution in the sinogram domain, it is difficult to utilize signal models for generative sampling. In this work, we build upon the LCE to propose LCEDiff that models the sinogram distribution using LCE while recovers missing information by using diffusion modeling. Experimental results show that our proposed method not only reconstructs satisfactory images under the sparse-view condition in the sinogram domain, but also improves sampling stability.