As is known, eigenvalue based spectrum sensing for multi-antenna cognitive radio systems is a highly preferred method because it requires no priori knowledge concerning the signal to be sensed. The performance of the method depends on the most accurate calculation of the test statistic and the threshold value. Regaining of the threshold values was aimed in this study in order to increase performance in generalized likelihood ratio detection based sensing. While the threshold value was calculated, the probability of false detection (P-fa) and the probability of detection were regained by using a new probability distribution function for Wishart matrices. The simulations were performed under noise uncertainty for MIMO-OFDM systems, and a noticeable performance increase was observed with the new threshold value in comparison with conventional methods.