We introduce for the first time an iterative MR electrical properties tomography reconstruction method by exploiting a cascade of multi-layer convolutional neural networks (CNNs) able to learn spatial priors in an iterative fashion, alternated with physics-based gradient descent direction calculations. This method was tested on 2D simulated human brain data. The presented results demonstrate the feasibility of this methodology to reconstruct conductivity and permittivity maps at 128MHz. Ultimately, this method allows computational advantages compared to standard contrast source inversion electrical properties tomography (CSI-EPT), i.e. faster reconstructions, which will be extremely relevant when moving to 3D reconstructions.