Brain Computer Interfaces (BCIs) based on magnetoencephalography (MEG) signals for Motor Imagery (MI) have the potential to help stroke patients recover neurologically and improve motor function in healthy individuals. While multi-channel MEG data provide high spatio-temporal resolution, not all channels contribute equally to the performance of BCI systems. We have taken the causal interactions between the channels of MEG signals in different trials of MI-BCIs into account. Therefore, a novel MEG channel selection method based on copula transfer entropy is proposed to reduce noise and redundant information. Effective features are extracted using the regularized common spatial pattern approach after channel selection. Finally, a support vector machine classifier with a radial basis function kernel is employed to classify MI tasks involving both hands and both feet movements. The effectiveness of the proposed method is validated on a publicly available MEG-BCI dataset. Experimental results demonstrate a significant improvement in classification accuracy and a substantial reduction in the number of selected channels compared to using all channels, for both intra-session and inter-session classifications. Furthermore, our selection method achieves significantly higher classification accuracy than the state-of-the-art methods in MEG-MI field.