Integrating fuzzy rules into the fast, robust segmentation of magnetic resonance images

Fuzzy clustering algorithms have been used in the automatic segmentation of large multi-dimensional data sets such as 2D magnetic resonance images. Although the clustering algorithms perform well, they are very time consuming. Better performance at a clustering stage is achieved if the data-set can be partially classified before clustering is applied. Further, if the pre-classification is itself fuzzy, the clustering algorithm can also be initialized using the output of the pre-classification stage. We show that the use of fuzzy rules to do this pre-classification is very effective. Good segmentation of the normal human brain into tissues of interest is obtained in about one eighth the time of using clustering alone with the relatively low number of error pixels being comparable to that obtained from fuzzy c-means clustering. Our approach is also robust in the sense that the rules are image independent and hence can accommodate the kind of intensity variations that commonly occur in magnetic resonance images of the human brain.