Automated Classification of Brain Tissue in MRI

Accurate, uniform volumetric measurements of brain structures are of great importance tomorphological studies of the brain in mental health, disease, and developmental research. Automatedmeasurement methods will improve the consistency, accuracy, and availability of such estimates, andwill allow more efficient screening of magnetic resonance (MR) images. In Phase I, ORINCON willdevelop a hierarchical system of fuzzy min-max neural networks (FMMNNs), capable of classifyingregions of MR brain images into four categories (gray matter, white matter, cerebro-spinal fluid (CSF),and hyperintensities). Using proton-density weighted (PDW) and T2 weighted (T2W) images from UCSDSchool of Medicine, sets of "stacked" MR images will first be preprocessed for improved spatialhomogeneity and brain boundary determination. After initial preprocessing, linear combinations of imageswill he formed, yielding images for CSF/Brain and Gray/White matter separation. The hierarchical neuralnets will he trained and tested with these sets of images. Three-dimensional analysis methods willcategorize pixels in specified image regions, which will then be compared to pre-existing estimates ina confusion matrix format. A final report will describe experimental procedures, analyze test results, andrecommend Phase II research directions for automated identification and volume quantification of specificbrain structures.