In MIMO-OFDM systems, a conventional channel estimation technique using comb type training symbol structure has high mean squared error (MSE) at edge subcarriers. To reduce the MSE at those subcarriers, we propose the cyclic comb type training structure and the weighted-averaging method. In the proposed technique, all kinds of comb type training symbols are transmitted cyclically at each transmitter. At the receiver, channel frequency responses estimated using each training symbol are averaged with weight which is obtained from the corresponding MSE. By computer simulations, we have shown that the proposed cyclic comb type training symbols gives higher SNR gain than the conventional training structure.