Algorithms providing accurate estimation of shear modulus and identification of tissue boundary are always desired for magnetic resonance elastography (MRE). In this study, we proposed a model-based neural network (MNN) embedding classic direct inversion (DI) and local frequency estimation (LFE). Convolution layers with Inception-like structure were applied for preprocessing, while postprocessing was implemented with a U-net structure. Additionally, DI and LFE algorithms were encapsulated as layers transforming wave images to shear modulus maps. The network performance was tested using a phantom with an inclusion. Compared with conventional algorithms, MNN provided modulus estimation with 5-fold higher contrast-to-noise ratio with clear boundaries.