Tomographic Image Reconstruction based on Artificial Neural Network (ANN) techniques

A new approach for tomographic image reconstruction from projections using Artificial Neural Network (ANN) techniques is presented in this work. The design of the proposed reconstruction system is based on simple but efficient network architecture, which best utilizes all available input information. Due to the computational complexity, which grows quadratically with the image size, the training phase of the system is characterized by relatively large CPU times. The trained network, on the contrary, is able to provide all necessary information in a quick and efficient way giving results comparable to other time consuming iterative reconstruction algorithms. The performance of the network studied with a large number of software phantoms is directly compared to other iterative and analytical techniques. For a given image size and projections number, the role of the hidden layers in the network architecture is examined and the quality dependence of the reconstructed image on the size of the geometrical patterns used in the training phase is also investigated. ANN based tomographic image reconstruction can be easily implemented in modern FPGA devices and can serve as a quick initialization method to other complicated and time consuming procedures.