Discrete Tomography A Neural Network Approach

Tomography tries to reconstruct an object from a number of projections in multiple directions. Many application domains can be imagined, but we will focus on high throughput applications, and will therefore try to reduce the number of necessary projections, while being able to generate good quality reconstructions. We apply several forms of Neural Networks, an Artificial Intelligence method. These networks are especially suited for solving underdetermined problems, and therefore well suited to our problem. Many different variants of Neural Networks are developed since its introduction; some simple, while others can consist of many nodes in many hidden layers increasing its training complexity. We will here focus on the simpler forms of Neural Networks: a feedforward (multilayer) perceptron. We present some experimental results, which demonstrate the capabilities of reconstructing high quality images using relatively simple Neural Networks.