Study On Automatical Segmentation Algorithm For Slice Images Of Digital Human Brain

Digital human is an emerging research field that involves the comprehensive application of modern scientific technology, especially the computer science and anatomy, to explore the digitalized information of human body. As the basis of the three-dimensional reconstruction of organs and tissues, the accurate segmentation of slice images plays a crucial role in the process of the digital human modeling. At present, segmentation of color slice image is mainly realized semi-automatically or manually both at home and abroad. It is obviously time-consuming and laborious to process the slice image dataset with massive data in such a way. Therefore, to develop the accurate and highly automatic algorithms for resolving this issue is of great significance.In this paper, the slice images of human brain from the dataset of first Chinese visible human female were adopted for investigation. The features of slice images and anatomical structures of human brain were analyzed initially. Following that, four types of more automatic algorithms were proposed for the continuous segmentation of brain tissue and white matter. The first one is the algorithm based on mathematic morphology. Initially, the region of cerebrum tissue in each slice image is roughly distinguished through morphological reconstruction. Then its edge is refined by means of dilation and erosion to complete the segmentation. The second one is the algorithm based on the theory of zooming template. All the images are divided into several sections. In each section, one image was selected and its target area recognized in the edge detection method is used as a template for guiding the segmentation of the other slice images. The size of the template is adjusted by means of mathematic morphology operation so that continuous segmentation can be recognized. The third one is the improved region growing algorithm. The basic idea of this algorithm is to apply the theory of both seed points'expansion and morphological reconstruction for image segmentation. In more details, the seed pixels are selected according to neighborhood average and the gray threshold is set to judge which pixels belong to the target area. Furthermore, the seed selection criteria are summed up with its generality verified. The last one is the algorithm based on RGB color-space clustering. Specifically, the color histogram is used to obtain the clustering centers, while Euclidian distance is adopted to combine the similar pixels together. The above algorithms are respectively applied for continuous segmentation of cerebrum tissue, cerebellar and brainstem tissue, cerebral white matter, as well as white matter in cerebellar and brainstem.To evaluate the performance of these segmentation algorithms, on one hand, the automatically and manually segmented results were compared and analyzed qualitatively in terms of vision effects and quantitatively by the use of Dice Similarity Index (DSI). On the other hand, the segmented results were reconstructed in three dimensions. The results have shown that the proposed algorithms are more effective and accurate. Moreover, they can capture details of the complex tissue and reduce manual intervention. Consequently, they are suitable for the rapid and automatic segmentation of massive sequential slice images, and thus create a good condition for building an accurate digital brain model.