• computed tomography
• denoising
• generative models
• image processing
• inverse problems
• machine learning

We combine reconstruction and denoising into a single model that directly maps a noisy sinogram to a denoised reconstruction, resulting in improved reconstruction quality and reduced inference time. Furthermore we develop a generative model capable of generating multiple possible reconstructions and iteratively improving them. The work also includes the development of a fast CUDA library and a detailed model of tomographic noise.