Automated segmentation of brain tissue and white matter in cryosection images from Chinese visible human dataset
Min Li,Xiao Lin Zheng,Hong Yan Luo,Richard Castillo,Shao Xiang Zhang,Li Wen Tan,Edward Castillo,Thomas Guerrero
DOI: https://doi.org/10.5405/jmbe.1336
IF: 2
2014-01-01
Journal of Medical and Biological Engineering
Abstract:Cryosection images contain fairly rich and original details of the brain anatomy. Accurate and fast segmentation of cryosection images is of great significance for research of the human brain and development of the Visible Human Project. However, most automated approaches in the literature are designed for magnetic resonance imaging or computed tomography data, and they may not be suitable for cryosection images. Cryosection image segmentation is often realized manually or semi-automatically in practice. The present study proposes an automated algorithm for cryosection image segmentation of brain tissue and white matter and evaluates its accuracy using the Chinese Visible Human (CVH) dataset. This method combines a mathematical morphological approach to delineate brain tissue and k-means clustering to uniquely identify white matter. Firstly, the region of brain tissue is detected coarsely using connected component labeling combined with morphological reconstruction. Then, morphological operations are used for final boundary determination to complete the segmentation of brain tissue Finally, k-means clustering is employed to extract white matter based on segmented brain tissue. The algorithm was applied to the CVH dataset to automatically extract the entire brain tissue and white matter in the cerebrum, cerebellum, and brain stem. Additionally, the proposed mathematical morphological approach is compared with the region growing method and the threshold morphological method for brain segmentation, and the k-means clustering method is compared with the fuzzy c-means clustering algorithm and the Gaussian mixture model coupled with the expectation-maximization algorithm for white matter extraction. To evaluate performance, a quantitative analysis was conducted using the dice similarity index, false-positive dice, and false-negative dice for comparison with manually obtained segmentation results produced by anatomy experts. Results indicate that the proposed algorithm is capable of accurate segmentation and substantial agreement with the gold standard.