A method for detection and removal of random spike noise in magnetic resonance (MR) raw data (k-space data) is described. This method would reduce or eliminate the corduroy-type and higher than usual level artifacts in MR images resulting from random spike noise in k-space data. The method described involves applying a spatially varying threshold to be k-space data. Any data point that has a magnitude greater than that of the threshold value at that location will be replaced by a local complex average of the neighboring data points or some other suitable data replacement scheme.