Clustering and classification of multispectral magnetic resonance images

N-dimensional clustering and classification algorithms that offer a method for efficiently classifying, segmenting, and visualizing the information contained in multispectral magnetic resonance images are discussed. Novel imaging methods, such as chemical shift imaging, contain spectral information on tissue metabolism. A problem associated with this method of imaging is that the information contained in the spectrum is not easily interpreted, nor is it extendable to the high-resolution proton image. Each of the frequency bands in the chemical shift image can be thought of as a unique feature, or band, of a multiband image. This resulting multiband image is used as input to an algorithm which groups the data into a set of clusters. The cluster image is segmented via unsupervised and supervised classification. This classification defines regions of the image defined by the chemical characteristics of the tissue. The advantage of this approach over current image interpretation schemes is that this method allows one to compile several feature images, revealing relationships between contributing features, which can be visualized in a single image