Stable learning for neural network tomography by using back projection type image

This paper presents a stable learning method of the neural network tomography, in case of asymmetrical few view projection. The neural network collocation method (NNCM) is one of effective reconstruction tools for symmetrical few view tomography. But in cases of asymmetrical few view, the learning process of NNCM tends to be unstable and fails to reconstruct appropriate tomographic images. We solve the unstable learning problem of NNCM by introducing two types of back projection reconstructed images in the initial learning stage of NNCM. The numerical simulation with an assumed tomographic image show the effectiveness of the proposed method.