Electrical properties are a novel contrast mechanism for quantitative MRI. Conductivity can be used as a biomarker for tumorous tissues. Different analytic Magnetic-Resonance Electrical Properties Tomography (MREPT) methods have been proposed, however, accurate reconstructions require empirical assessment and setting of regularization coefficients per sample. In this work, based on a modified formulation of Convection-Reaction Equation-Based EPT (cr-EPT), the regularization coefficients are learned from the difference between reconstructed conductivity maps and their ground truth, using a convolutional neural network (CNN) model. The CNN model with the modified cr-EPT could achieve conductivity reconstructions with higher accuracy, compared to several analytical models.