Using adaptive fuzzy rules for image segmentation

Segmenting magnetic resonance images of the same body region taken at different times is a challenging task. Obtaining reliable data to train a classifier is difficult due the differences among subjects and even differences over time in images acquired from a single subject. Unsupervised clustering can be used to group like tissues into classes. However, clustering does not provide class labels, is time consuming, and may not always provide suitable data partitions. We show how a set of adaptive fuzzy rules can be used to identify many of the voxels from a magnetic resonance image before clustering is done. This allows clustering to be done on a subset of an image with a “good” initialization, which mitigates the time required. The identified voxels can also be used to identify clusters. The fuzzy rule based system followed by a clustering step has been applied to 105, 5 mm thick, magnetic resonance images of the human brain which are taken from 15 different subjects. It is shown that the segmentations produced are approximately 5 times faster than those produced by fuzzy clustering alone and are comparable in the accuracy of the segmentation