The precise segmentation of magnetic resonance images (MRI) is an important subject in both medical and computer science communities. With MRI's property of multi-spectrum, we use the information from its PD-,T1-, and T2-weighted images, mapping them into a multi-dimensional intensity space and getting its vector gradient. Through the improvement of the step function, an unsupervised self-organizing map (SOM) neural network is trained dynamically. To improve the effectiveness of segmentation, we develop a semi-supervised training scheme at the edge of image in multi-dimensional intensity space.