A Tomographic Reconstruction Method using Coordinate-based Neural Network with Spatial Regularization

Tomographic reconstruction is concerned with computing the cross-sections of an object from a finite number projections. Many conventional methods represent as images on regular grid. In this paper, we study recent coordinate-based neural network for tomographic reconstruction, where inputs spatial coordinate and outputs attenuation coefficient coordinate. This allows continuous representation object. Based network, propose regularization term, to obtain high-quality reconstruction. Experimental results synthetic data show that term improves quality significantly, compared baseline. We also provide ablation different architecture configurations hyper-parameters.