The work described here is motivated by the need to improve the results of clustering. The usefulness of clustering is enhanced for classificatory problems for which labeled training data may be available, but is expensive to obtain and further, some available data implies some expertise or knowledge about the domain involved. The problem of segmenting magnetic resonance images into regions of interest such as gray matter, white matter and cerebro-spinal fluid is a good example of a domain for which knowledge based clustering is well suited. It is possible to use a supervised algorithm to label regions [1]. Because the data can vary by machine, by day and by patient, however, it is necessary to acquire training data for each image.
Alternatively, if domain knowledge and some relative information about the relation of the classes after clustering in feature space are available, knowledge based clustering can be applied. It will be shown to be an unsupervised technique which uses knowledge to guide the clustering process. The knowledge is encapsulated as rules in an expert system (the Clips expert system tool is used in this work) and will enable clusters (regions) to be labeled. Further, the use of the knowledge and re-clustering focussed regions under the guidance of the knowledge-based system will reduce or even solve, problems of under and over-segmentation. Knowledge of the number of classes in an image, as well as their morphology, location, size, and relative feature space characteristics in an image are some examples of domain knowledge utilized in this work.
Knowledge-based image processing/understanding is an idea which has been worked upon for some time. It is applicable to a large range of problems and applications [2,3,4,5,6]. This paper presents a new multi-paradigm approach that incorporates both knowledge and a clustering algorithm to generate labeled clusters. The approach is called knowledge based clustering with reclustering and allows examples or patterns to be labeled with minimal or no labeled training examples, when appropriate domain knowledge is available.
We have applied the technique to the domains of magnetic resonance (MR) image brain scans and coastal zone color scanner (CZCS) satellite images. To our knowledge, this is the first successful attempt to use domain knowledge, other than labeled examples [7], to automatically guide the otherwise unsupervised clustering process.