Fast tissue segmentation based on a 4D feature map in characterization of intracranial lesions

The aim of this work was to develop a fast and accurate method for tissue s egmentation in magnetic resonance imaging (MRI) based on a four-dimensional (4D) feature map and compare it with that derived from a 3D feature map. H igh-resolution MRI was performed in 5 normal individuals, in 12 patients wi th brain multiple sclerosis (MS), and 9 patients with malignant brain tumor s. Three inputs (proton-density, TB-weighted fast spin-echo, and T1-weighte d spin-echo MR images) were routinely utilized. As a fourth input, either m agnetization transfer MRT was used or T1-weighted post-contrast MRI tin pat ients only). A modified k-nearest neighbor segmentation algorithm was optim ized for maximum computation speed and highquality segmentation. In that re gard, we a) discarded the redundant seed points; b) discarded the points wi thin 0.5 standard deviation from the cluster center that were nonoverlappin g with other tissue; and c) removed outlying seed points outside 5 times th e standard deviation from the cluster center of each tissue class. After se gmentation, a stack of color-coded segmented images was created. Our new te chnique utilizing all four MRI Inputs provided better segmentation than tha t based on three inputs (P < 0.001 for FAS and P < 0.001 for tumors). The t issues were smoother due to the reduction of statistical noise, and the del ineation of the tissues became sharper. Details that were previously blurre d or invisible now became apparent. In normal persons a detailed depiction of deep gray matter nuclei was obtained, In malignant tumors, up to five ab normal tissue types were identified: 1) solid tumor core, 2) cyst, 3) edema In white matter 4) edema in gray matter, and 5) necrosis. Delineation of M S plaque in different stages of demyelination became much sharper. In concl usion, the proposed methodology warrants further development and clinical e valuation, (C) 1999 Wiley-Liss, Inc.