Traditional brain atlases are mainly based on hand-crafted anatomical structures, ignoring the useful connectivity pattern information. In our work we use diffusion weighted imaging data to incorporate connectivity information into atlas generation. We use FSL to process the data to extract the connectivity matrix. The brain parcellation problem is then formulated as a min-cut problem on a big, sparse graph. Spectral clustering and an original multi-class Hopfield network (MHN) method is applied to solve the problem, each working with a different analytical framework: MHN works in the diffusion space to generate individual parcellations, while spectral clustering works on standard space averaged connectome to generate group level atlases. Group study of brain images with Schizophrenia is conducted, showing significant improvement in accuracy for disease diagnosis using features extracted with the proposed parcellation scheme. Hypothesis test was performed on local structures to explore possible structural causes of the disease.