Automated quantification of brain magnetic resonance image hyperintensities using hybrid clustering and knowledge-based methods

Previous computerized methods of hyperintensity identification in brain mag netic resonance images (MRI) either rely heavily on human intervention or o n simple thresholding techniques. Such methods can lead to considerable Var iation in the quantification of brain hyperintensities depending upon image parameters such as contrast. This paper describes an automated, knowledge- guided method of hyperintensity detection in brain MRI that addresses probl ems associated with human subjectivity and thresholding techniques. This me thod, which we call knowledge-guided hyperintensity detection (KGHID), uses encoded knowledge of brain anatomy and MRI characteristics of individual t issues to reclassify pixels from an initial unsupervised segmentation. With this encoded knowledge, KGHID discriminates lesions embedded within the wh ite matter, hyperintense lesions of the basal ganglia and the periventricul ar ring. The method is designed for high sensitivity detection and monitori ng of subtle lesions in patients with neurodegenerative diseases. (C) 1999 John Wiley & Sons, Inc.