Simultaneous Segmentation of Multiple Organs in Multi-Dimensional Medical Images

点击翻译按钮获取中文摘要

We developed a digital atlas of human anatomy and a simultaneous segmentation algorithm for multiple organs. The digital atlas of human anatomy is a computational and statistical database including information about human anatomy, pathology and function, such as physiological and motility function. This study focused on twelve organs in three dimensional upper abdominal CT images, namely oesophagus, heart, stomach, liver, gallbladder, pancreas, left and right kidneys, spleen, portal vein, aorta, and inferior vena cava. We built 1) probabilistic atlases of the organs showing the probability of existence, 2) statistical shape model including average and eigen-shapes, and 3) statistical database of features, such as gray values in the CT images. We also developed a simultaneous segmentation algorithm of 12 abdominal organs based on the digital atlas of human anatomy. The algorithm consists of spatial normalization, rough extraction and fine extraction of the organs. A hierarchal spatial normalization process was proposed in this project, which normalized large organs first, then performed normalization for small organs whose spatial variation is large. The rough extraction process was based on maximum a posterior method. The method estimated probabilistic distribution parameters of features using a modified EM algorithm which used the probabilistic atlases as a priori information. We proposed a posterior probability based on features of neighboring voxels. Finally the proposed method performed a multiple level set method which extracted the twelve organs simultaneously based on mutual interaction between neighboring organs. We applied the proposed algorithm to abdominal CT volumes of 17 cases and confirmed that the Jaccard index was 75% on average which is quite high comparing to the previous method.