Fast GPU-Driven Model-Based X-Ray CT Image Reconstruction via Alternating Dual Updates

Model-based image reconstruction (MBIR) methods for X-ray CT reconstruction can improve image quality and reduce patient X-ray dose. These methods produce images by solving high-dimensional, statistically motivated numerical optimization problems, but unfortunately the high computational costs of solving these problems have kept MBIR algorithms from reaching ubiquity in the clinic. In this paper, we present an X-ray CT image reconstruction algorithm that uses duality and group coordinate ascent to alternately perform efficient tomography and denoising updates. The algorithm can handle non-smooth regularizers like anisotropic total variation (TV) and stores only two image-sized vectors on the GPU. Preliminary experiments show the algorithm converges very quickly in time.