Fuzzy classification method taking account of partial volume and image non-uniformity effects

The authors present a method for the unsupervised classification and quantification of tissues in MR images, which accounts for important sources of error: the partial volume effect and image non-uniformity. The general method combines fuzzy unsupervised segmentation techniques with local neighbourhood information. The authors model the image as a fuzzy membership process which allows individual pixels to represent a mixture of tissues. To compensate for non-uniformity, they look at the similarity between the membership values of pixels in local neighbourhoods. The algorithm is an iterative minimisation procedure. It can be applied to single-valued- or multi-channel datasets of two or three spatial dimensions. The authors apply this method to 3D MR images of the brain and compare its performance with other unsupervised segmentation techniques. Their study shows that this method provides accurate classification and quantification of tissues