In New Radio (NR) massive multiple input multiple output (MIMO) system, accurate acquisition of downlink channel state information (CSI) at the base station (BS) is of great importance. However, the overhead of CSI feedback becomes larger with the increased number of BS antennas. In order to further reduce feedback overhead, deep learning (DL)-based approaches are recently proposed to solve this problem. In this paper, we propose a DL model using a neural network for implicit CSI feedback mechanism to compare the existing feedback method based on the eType II codebook introduced by the 3rd Generation Partnership Project (3GPP) Rel-16. Simulation results show that DL-based approach has higher performance gains in comparison with the standard method and shows great potential in future MIMO systems.