SUPERVISING ISODATA WITH AN INFORMATION THEORETIC STOPPING RULE

New biomedical imaging modalities, such as Magnetic Resonance Imaging (MRI), provide fertile multidimensional environments for the automatic identification of biological soft tissues, but lack the a priori information required to appropriately train supervised classifiers. Hierarchical cluster analysis techniques can generate this information but they are inefficient on large data sets. We have developed an unsupervised clustering method that is a variant of the well known ISODATA algorithm. We replaced the heuristic rules that control ISODATA with rules that search for the minimum value of an information theoretic criterion. The criteria investigated in this study are Akaike's Information Criterion (AIC) and the Consistent AIC (CAIC). Both measure the global fit of a cluster model to the input data, and the smallest criterion value suggests the best fit. We tested the new method on multivariate Gaussian and real world data, including MR images of normal and diseased tissue in vivo. © 1990.