In order to meet the requirements of 6G communication for environmental adaptability and interference resistance, obtaining accurate Channel State Information (CSI) is of paramount importance. However, traditional communication methods struggle to fulfill these demands, leading to a growing interest in deep learning-based channel estimation solutions among researchers. This paper introduces a solution to the channel estimation problem in OFDM systems, employing a deep learning approach based on the MLP-Mixer block, referred to as CENet. CENet's channel-mixing and token-mixing structures enable better capturing of both temporal and spectral channel characteristics. The proposed channel estimation method consists of two parts: firstly, preliminary channel estimation results are generated using the LS algorithm, and then CENet is employed to further refine these preliminary results. Simulation results demonstrate the superiority of the proposed approach over other deep learning methods. Additionally, this paper extends the method to MIMO scenarios and introduces pruning techniques to reduce redundant parameters in the MLP layers, thereby reducing computational complexity.