A proper tomographic reconstruction becomes increasingly challenging when the number of projections is limited. This is also the case when projections are acquired using Lorentz Transmission Electron Microscopy (LTEM) for the reconstruction of the electromagnetic potentials of magnetic nanoparticles. Even the most advanced TEM holder, built specifically for the purpose of tomography, does not allow performing a full 180° tilt series. This leads to a poor reconstruction with sample edge artifacts. Hence, we propose to use neural networks to overcome the issue of limited projections. In this method, we will, first, generate simulated projections using the Aharonov-Bohm magnetic phase shift expression. Next, we will apply an array of filters ranging from low-pass to high-pass cutoff frequency to generate a training set. Subsequently, neural networks will be used to non–linearly weigh each of the filters such that the final scaling suppresses the reconstruction edge artifacts.