In this paper, methods are proposed and validated to determine low and high thresholds to segment out gray matter and white matter for MR images of different pulse sequences of human brain. First, a two-dimensional re...In this paper, methods are proposed and validated to determine low and high thresholds to segment out gray matter and white matter for MR images of different pulse sequences of human brain. First, a two-dimensional reference image is determined to represent the intensity characteristics of the original three-dimensional data. Then a region of interest of the reference image is determined where brain tissues are present. The non-supervised fuzzy c-means clustering is employed to determine: the threshold for obtaining head mask, the low threshold for T2-weighted and PD-weighted images, and the high threshold for T1-weighted, SPGR and FLAIR images. Supervised range-constrained thresholding is employed to determine the low threshold for T1-weighted, SPGR and FLAIR images. Thresholding based on pairs of boundary pixels is proposed to determine the high threshold for T2-and PD-weighted images. Quantification against public data sets with various noise and inhomogeneity levels shows that the proposed methods can yield segmentation robust to noise and intensity inhomogeneity. Qualitatively the proposed methods work well with real clinical data.展开更多
文摘In this paper, methods are proposed and validated to determine low and high thresholds to segment out gray matter and white matter for MR images of different pulse sequences of human brain. First, a two-dimensional reference image is determined to represent the intensity characteristics of the original three-dimensional data. Then a region of interest of the reference image is determined where brain tissues are present. The non-supervised fuzzy c-means clustering is employed to determine: the threshold for obtaining head mask, the low threshold for T2-weighted and PD-weighted images, and the high threshold for T1-weighted, SPGR and FLAIR images. Supervised range-constrained thresholding is employed to determine the low threshold for T1-weighted, SPGR and FLAIR images. Thresholding based on pairs of boundary pixels is proposed to determine the high threshold for T2-and PD-weighted images. Quantification against public data sets with various noise and inhomogeneity levels shows that the proposed methods can yield segmentation robust to noise and intensity inhomogeneity. Qualitatively the proposed methods work well with real clinical data.
