Parameters Estimation Directly from Sinograms with Neural Networks

Artificial Neural networks (ANNs) have been known to be suitable for various mathematical transformations. With the advent of GPU-based implementations of NNs, they are making a significant impact on the application of artificial intelligence in medicine. In this project, we investigate the NNs usefulness in relation to the reconstruction of Radon projection data. Medical image reconstruction is a type of estimation problem. For example, estimate emission counts or absorption coefficients from noisy Radon projections of an image, also known as a sinogram. We investigate with a simple simulation of how different parameters of an image can be recovered directly from sinograms with NN models. We have estimated object locations in 2D, radii for circular objects, constant attenuation coefficients, and shapes of complex 2D objects when they are used in creating sinograms. We investigated different NN architectures, including a convolutional neural network.