IntraTomo: Self-supervised Learning-based Tomography via Sinogram Synthesis and Prediction

Abstract

Combines learning-based and model-based approaches for ill-posed CT inverse problems. Two modules: sinogram prediction (density field as continuous differentiable NN function, self-supervised from incomplete/degraded sinogram) and geometry refinement (local & non-local geometrical priors), applied iteratively. Outperforms on limited-angle tomography (45°), sparse view (as few as 8 views), super-resolution (8×).