In large-scale MIMO systems, where pilot-contamination effects cannot be avoided, it is essential to explicitly take the imperfect channel state information into account when designing receive and transmit filters. We investigate how receiver design based on the mean square error of the detected data symbols allows us to reduce the impact of pilot-contamination. Asymptotic analysis shows that under mild conditions on the channel covariance matrices, the mean square error goes to zero. The asymptotic results further allow us to derive a low-complexity filter design, which is also asymptotically optimal. Simulation results based on a typical spatial channel model confirm the theoretical considerations.