Classical option pricing models rely on prior assumptions made on the dynamics of the underlying assets and the rationality of the market. While empirical evidence showed that these models may explain the option prices to certain extend, their performance may be poor when the actual situation deviates from the assumptions. Neural network models are capable of learning the underlying relationship through the data without prior assumptions. However, to avoid over-fitting, these models often require massive amount of data, which are typically not available for option pricing problems. In this paper we propose a new model by integrating neural networks as components to a classical option pricing model, thus significantly increasing the model flexibility while requiring only a reasonable amount of data. We further show that the training of the model, also known as the calibration in the field of financial engineering, may be formulated into a simulation optimization problem, and it may be solved in a way that is compatible to the training of neural networks. Preliminary numerical results show that our approach works well.