An improved iterative dimensionality-reduction parallel detection algorithm is proposed to mitigate the effects of the low diversity gain of the first detective sub-streams and the error propagation in traditional Multiple Input Multiple Output (MIMO) detection algorithm.In each iteration of the algorithm, the first substream is found by exhaustive search while other sub-streams are detected in parallel through ordered successive interference cancellation (OSIC) , and only the estimates of the first sub-stream with the highest diversity order can be obtained at the end of each iteration.Furthermore, between two different iterations, interference cancellation is employed to reduce the dimension of sub-streams.Simulated results indicate that, only with marginal complexity cost, the proposed algorithm helps obtain BER (Bit Error Rate) performance approaching maximum likelihood detection algorithm.Particularly, in a 4×4 QPSK modulation MIMO system, the performance gain of the proposed algorithm over OSIC is 9.3dB at a BER of 10^-3.