The objective of this paper is to establish an "Averaged Learning Subspace Method" (ALSM) applicable for classification of multi-spectral MR images. By using the ALSM to process the massive amounts of information in multi-spectral MR images, and classification tissues of brain. The classification result of each tissue has shown by binary image, respectively. The classification results would assist doctor to diagnose more efficiently and more accurately and thus to gain more time for necessary action. In order to further evaluate the performance of ALSM, the high order statistics is adopted assessment and compare with perceptron neural network.