In this paper, we propose a denoise-and-forward network coding (DNF-NC) transmission scheme for its applications in two-way relay multiple-input and multiple-output (MIMO) systems. We first consider a scenario with a single pair of source nodes, and minimum mean square error (MMSE) receiver is applied at each node. The global optimal precoding design based on the mean square error (MSE) criterion can be achieved by solving two independent convex optimization problems. To achieve a better tradeoff between performance and complexity, an alternative power optimization approach is proposed using a channel diagonalization technique. Then, we proceed to considering a more challenging bidirectional communication scenario with multiple pairs of source nodes. With intrapair coordination at the sources, we modify DNF-NC by employing a signal alignment technique to combat interpair interference. The numerical results demonstrate that the proposed DNF-NC schemes can significantly improve bit error rate (BER) performance in both two scenarios, and such performance gains can be achieved with relatively low computational complexity.