COMPUTERIZED BRAIN-TISSUE CLASSIFICATION OF MAGNETIC-RESONANCE IMAGES- A NEW APPROACH TO THE PROBLEM OF PARTIAL VOLUME ARTIFACT

Due to the finite spatial resolution of digital magnetic resonance ima ges of the brain, and the complexity of anatomical interfaces between brain regions of different tissue type, it is inevitable that some vox els will represent a mixture of two or three different tissue types. O utright assignment of such ''bipartial'' or ''tripartial'' voxels to o ne class or another is more problematic and less reliable than assignm ent of ''full-volume'' voxels, wholly representative of a single tissu e type. We have developed a computerized system for brain tissue class ification of dual echo MR data, which uses a polychotomous logistic mo del for discriminant analysis, combined with a Bayes allocation rule i ncorporating differential prior probabilities, and spatial connectivit y tests, to assign each voxel in the image to one of four possible cla sses: gray matter, white matter, cerebrospinal fluid, or unclassified The system supports automated volumetric analysis of segmented images, has low operational overheads, and compares favorably with previous m ultivariate or ''multispectral'' approaches to brain MR image segmenta tion in terms of both validity (bootstrap misclassification rate = 3.3 %) and interoperator reliability (intra-class correlation coefficients for all three tissue classes >0.9). We argue that these improvements in performance stem from better methodological management of the relat ed problems of non-Normality of MR signal intensity values and partial volume artifact. (C) 1995 Academic Press, Inc.